radv: add initial non-conformant radv vulkan driver

This squashes all the radv development up until now into
one for merging.

History can be found:
https://github.com/airlied/mesa/tree/semi-interesting

This requires llvm 3.9 and is in no way considered
a conformant vulkan implementation. It can run a number
of vulkan applications, and supports all GPUs using
the amdgpu kernel driver.

Thanks to Intel for providing anv and spirv->nir,
and Emil Velikov for reviewing build integration.

Parts of this are:
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Acked-by: Edward O'Callaghan <funfunctor@folklore1984.net>

Authors: Bas Nieuwenhuizen and Dave Airlie
Signed-off-by: Dave Airlie <airlied@redhat.com>

63 files changed, 32093 insertions, 8 deletions

diff --git a/configure.ac b/configure.ac
index 99abcd7a5f..c8aa829a8c 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1715,6 +1715,10 @@ if test -n "$with_vulkan_drivers"; then
             HAVE_INTEL_VULKAN=yes;
 
             ;;
+        xradeon)
+            PKG_CHECK_MODULES([AMDGPU], [libdrm_amdgpu >= $LIBDRM_AMDGPU_REQUIRED])
+            HAVE_RADEON_VULKAN=yes;
+	    ;;
         *)
             AC_MSG_ERROR([Vulkan driver '$driver' does not exist])
             ;;
@@ -2198,7 +2202,7 @@ if test "x$enable_gallium_llvm" = xauto; then
     i*86|x86_64|amd64) enable_gallium_llvm=yes;;
     esac
 fi
-if test "x$enable_gallium_llvm" = xyes; then
+if test "x$enable_gallium_llvm" = xyes || test "x$HAVE_RADEON_VULKAN" = xyes; then
     if test -n "$llvm_prefix"; then
         AC_PATH_TOOL([LLVM_CONFIG], [llvm-config], [no], ["$llvm_prefix/bin"])
     else
@@ -2368,10 +2372,7 @@ radeon_llvm_check() {
     else
         amdgpu_llvm_target_name='amdgpu'
     fi
-    if test "x$enable_gallium_llvm" != "xyes"; then
-        AC_MSG_ERROR([--enable-gallium-llvm is required when building $1])
-    fi
-    llvm_check_version_for "3" "6" "0" $1
+    llvm_check_version_for $2 $3 $4 $1
     if test true && $LLVM_CONFIG --targets-built | grep -iqvw $amdgpu_llvm_target_name ; then
         AC_MSG_ERROR([LLVM $amdgpu_llvm_target_name not enabled in your LLVM build.])
     fi
@@ -2382,6 +2383,13 @@ radeon_llvm_check() {
     fi
 }
 
+radeon_gallium_llvm_check() {
+    if test "x$enable_gallium_llvm" != "xyes"; then
+        AC_MSG_ERROR([--enable-gallium-llvm is required when building $1])
+    fi
+    radeon_llvm_check $*
+}
+
 swr_llvm_check() {
     gallium_require_llvm $1
     if test ${LLVM_VERSION_INT} -lt 306; then
@@ -2466,7 +2474,7 @@ if test -n "$with_gallium_drivers"; then
             gallium_require_drm "Gallium R600"
             gallium_require_drm_loader
             if test "x$enable_opencl" = xyes; then
-                radeon_llvm_check "r600g"
+                radeon_gallium_llvm_check "r600g" "3" "6" "0"
                 LLVM_COMPONENTS="${LLVM_COMPONENTS} bitreader asmparser"
             fi
             ;;
@@ -2476,7 +2484,7 @@ if test -n "$with_gallium_drivers"; then
             PKG_CHECK_MODULES([AMDGPU], [libdrm_amdgpu >= $LIBDRM_AMDGPU_REQUIRED])
             gallium_require_drm "radeonsi"
             gallium_require_drm_loader
-            radeon_llvm_check "radeonsi"
+            radeon_gallium_llvm_check "radeonsi" "3" "6" "0"
             require_egl_drm "radeonsi"
             ;;
         xnouveau)
@@ -2541,6 +2549,10 @@ if test -n "$with_gallium_drivers"; then
     done
 fi
 
+if test "x$HAVE_RADEON_VULKAN" != "x0"; then
+    radeon_llvm_check "radv" "3" "9" "0"
+fi
+
 dnl Set LLVM_LIBS - This is done after the driver configuration so
 dnl that drivers can add additional components to LLVM_COMPONENTS.
 dnl Previously, gallium drivers were updating LLVM_LIBS directly
@@ -2632,8 +2644,13 @@ AM_CONDITIONAL(HAVE_R200_DRI, test x$HAVE_R200_DRI = xyes)
 AM_CONDITIONAL(HAVE_RADEON_DRI, test x$HAVE_RADEON_DRI = xyes)
 AM_CONDITIONAL(HAVE_SWRAST_DRI, test x$HAVE_SWRAST_DRI = xyes)
 
+AM_CONDITIONAL(HAVE_RADEON_VULKAN, test "x$HAVE_RADEON_VULKAN" = xyes)
 AM_CONDITIONAL(HAVE_INTEL_VULKAN, test "x$HAVE_INTEL_VULKAN" = xyes)
 
+AM_CONDITIONAL(HAVE_AMD_DRIVERS, test "x$HAVE_GALLIUM_R600" = xyes -o \
+                                      "x$HAVE_GALLIUM_RADEONSI" = xyes -o \
+                                      "x$HAVE_RADEON_VULKAN" = xyes)
+
 AM_CONDITIONAL(HAVE_INTEL_DRIVERS, test "x$HAVE_INTEL_VULKAN" = xyes -o \
                                         "x$HAVE_I965_DRI" = xyes)
 
@@ -2726,6 +2743,8 @@ dnl Substitute the config
 AC_CONFIG_FILES([Makefile
 		src/Makefile
 		src/amd/Makefile
+		src/amd/common/Makefile
+		src/amd/vulkan/Makefile
 		src/compiler/Makefile
 		src/egl/Makefile
 		src/egl/main/egl.pc
diff --git a/src/Makefile.am b/src/Makefile.am
index 91d6a7adb3..17c8798323 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -74,7 +74,7 @@ endif
 # include only conditionally ?
 SUBDIRS += compiler
 
-if HAVE_GALLIUM_RADEON_COMMON
+if HAVE_AMD_DRIVERS
 SUBDIRS += amd
 endif
 
@@ -120,6 +120,12 @@ if HAVE_INTEL_VULKAN
 SUBDIRS += intel/vulkan
 endif
 
+# Requires wayland-drm
+if HAVE_RADEON_VULKAN
+SUBDIRS += amd/common
+SUBDIRS += amd/vulkan
+endif
+
 if HAVE_GALLIUM
 SUBDIRS += gallium
 endif
diff --git a/src/amd/common/Makefile.am b/src/amd/common/Makefile.am
new file mode 100644
index 0000000000..788152d850
--- /dev/null
+++ b/src/amd/common/Makefile.am
@@ -0,0 +1,51 @@
+# Copyright © 2016 Bas Nieuwenhuizen
+#
+# Permission is hereby granted, free of charge, to any person obtaining a
+# copy of this software and associated documentation files (the "Software"),
+# to deal in the Software without restriction, including without limitation
+# the rights to use, copy, modify, merge, publish, distribute, sublicense,
+# and/or sell copies of the Software, and to permit persons to whom the
+# Software is furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice (including the next
+# paragraph) shall be included in all copies or substantial portions of the
+# Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+# IN THE SOFTWARE.
+
+include Makefile.sources
+
+# TODO cleanup these
+AM_CPPFLAGS = \
+	$(VALGRIND_CFLAGS) \
+	$(DEFINES) \
+	-I$(top_srcdir)/include \
+	-I$(top_builddir)/src \
+	-I$(top_srcdir)/src \
+	-I$(top_builddir)/src/compiler \
+	-I$(top_builddir)/src/compiler/nir \
+	-I$(top_srcdir)/src/compiler \
+	-I$(top_srcdir)/src/mapi \
+	-I$(top_srcdir)/src/mesa \
+	-I$(top_srcdir)/src/mesa/drivers/dri/common \
+	-I$(top_srcdir)/src/gallium/auxiliary \
+	-I$(top_srcdir)/src/gallium/include
+
+AM_CFLAGS = $(VISIBILITY_CFLAGS) \
+	$(PTHREAD_CFLAGS) \
+	$(LLVM_CFLAGS) \
+	$(LIBELF_CFLAGS)
+
+AM_CXXFLAGS = \
+	$(VISIBILITY_CXXFLAGS) \
+	$(LLVM_CXXFLAGS)
+
+noinst_LTLIBRARIES = libamd_common.la
+
+libamd_common_la_SOURCES = $(AMD_COMPILER_SOURCES)
diff --git a/src/amd/common/Makefile.sources b/src/amd/common/Makefile.sources
new file mode 100644
index 0000000000..380dba0862
--- /dev/null
+++ b/src/amd/common/Makefile.sources
@@ -0,0 +1,29 @@
+# Copyright © 2016 Bas Nieuwenhuizen
+#
+# Permission is hereby granted, free of charge, to any person obtaining a
+# copy of this software and associated documentation files (the "Software"),
+# to deal in the Software without restriction, including without limitation
+# the rights to use, copy, modify, merge, publish, distribute, sublicense,
+# and/or sell copies of the Software, and to permit persons to whom the
+# Software is furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice (including the next
+# paragraph) shall be included in all copies or substantial portions of the
+# Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+# IN THE SOFTWARE.
+
+AMD_COMPILER_SOURCES := \
+	ac_binary.c \
+	ac_binary.h \
+	ac_llvm_helper.cpp \
+	ac_llvm_util.c \
+	ac_llvm_util.h \
+	ac_nir_to_llvm.c \
+	ac_nir_to_llvm.h
diff --git a/src/amd/common/ac_binary.c b/src/amd/common/ac_binary.c
new file mode 100644
index 0000000000..01cf000d9b
--- /dev/null
+++ b/src/amd/common/ac_binary.c
@@ -0,0 +1,288 @@
+/*
+ * Copyright 2014 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors: Tom Stellard <thomas.stellard@amd.com>
+ *
+ * Based on radeon_elf_util.c.
+ */
+
+#include "ac_binary.h"
+
+#include "util/u_math.h"
+#include "util/u_memory.h"
+
+#include <gelf.h>
+#include <libelf.h>
+#include <stdio.h>
+
+#include <sid.h>
+
+#define SPILLED_SGPRS                                     0x4
+#define SPILLED_VGPRS                                     0x8
+
+static void parse_symbol_table(Elf_Data *symbol_table_data,
+				const GElf_Shdr *symbol_table_header,
+				struct ac_shader_binary *binary)
+{
+	GElf_Sym symbol;
+	unsigned i = 0;
+	unsigned symbol_count =
+		symbol_table_header->sh_size / symbol_table_header->sh_entsize;
+
+	/* We are over allocating this list, because symbol_count gives the
+	 * total number of symbols, and we will only be filling the list
+	 * with offsets of global symbols.  The memory savings from
+	 * allocating the correct size of this list will be small, and
+	 * I don't think it is worth the cost of pre-computing the number
+	 * of global symbols.
+	 */
+	binary->global_symbol_offsets = CALLOC(symbol_count, sizeof(uint64_t));
+
+	while (gelf_getsym(symbol_table_data, i++, &symbol)) {
+		unsigned i;
+		if (GELF_ST_BIND(symbol.st_info) != STB_GLOBAL ||
+		    symbol.st_shndx == 0 /* Undefined symbol */) {
+			continue;
+		}
+
+		binary->global_symbol_offsets[binary->global_symbol_count] =
+					symbol.st_value;
+
+		/* Sort the list using bubble sort.  This list will usually
+		 * be small. */
+		for (i = binary->global_symbol_count; i > 0; --i) {
+			uint64_t lhs = binary->global_symbol_offsets[i - 1];
+			uint64_t rhs = binary->global_symbol_offsets[i];
+			if (lhs < rhs) {
+				break;
+			}
+			binary->global_symbol_offsets[i] = lhs;
+			binary->global_symbol_offsets[i - 1] = rhs;
+		}
+		++binary->global_symbol_count;
+	}
+}
+
+static void parse_relocs(Elf *elf, Elf_Data *relocs, Elf_Data *symbols,
+			unsigned symbol_sh_link,
+			struct ac_shader_binary *binary)
+{
+	unsigned i;
+
+	if (!relocs || !symbols || !binary->reloc_count) {
+		return;
+	}
+	binary->relocs = CALLOC(binary->reloc_count,
+			sizeof(struct ac_shader_reloc));
+	for (i = 0; i < binary->reloc_count; i++) {
+		GElf_Sym symbol;
+		GElf_Rel rel;
+		char *symbol_name;
+		struct ac_shader_reloc *reloc = &binary->relocs[i];
+
+		gelf_getrel(relocs, i, &rel);
+		gelf_getsym(symbols, GELF_R_SYM(rel.r_info), &symbol);
+		symbol_name = elf_strptr(elf, symbol_sh_link, symbol.st_name);
+
+		reloc->offset = rel.r_offset;
+		strncpy(reloc->name, symbol_name, sizeof(reloc->name)-1);
+		reloc->name[sizeof(reloc->name)-1] = 0;
+	}
+}
+
+void ac_elf_read(const char *elf_data, unsigned elf_size,
+		 struct ac_shader_binary *binary)
+{
+	char *elf_buffer;
+	Elf *elf;
+	Elf_Scn *section = NULL;
+	Elf_Data *symbols = NULL, *relocs = NULL;
+	size_t section_str_index;
+	unsigned symbol_sh_link = 0;
+
+	/* One of the libelf implementations
+	 * (http://www.mr511.de/software/english.htm) requires calling
+	 * elf_version() before elf_memory().
+	 */
+	elf_version(EV_CURRENT);
+	elf_buffer = MALLOC(elf_size);
+	memcpy(elf_buffer, elf_data, elf_size);
+
+	elf = elf_memory(elf_buffer, elf_size);
+
+	elf_getshdrstrndx(elf, &section_str_index);
+
+	while ((section = elf_nextscn(elf, section))) {
+		const char *name;
+		Elf_Data *section_data = NULL;
+		GElf_Shdr section_header;
+		if (gelf_getshdr(section, &section_header) != &section_header) {
+			fprintf(stderr, "Failed to read ELF section header\n");
+			return;
+		}
+		name = elf_strptr(elf, section_str_index, section_header.sh_name);
+		if (!strcmp(name, ".text")) {
+			section_data = elf_getdata(section, section_data);
+			binary->code_size = section_data->d_size;
+			binary->code = MALLOC(binary->code_size * sizeof(unsigned char));
+			memcpy(binary->code, section_data->d_buf, binary->code_size);
+		} else if (!strcmp(name, ".AMDGPU.config")) {
+			section_data = elf_getdata(section, section_data);
+			binary->config_size = section_data->d_size;
+			binary->config = MALLOC(binary->config_size * sizeof(unsigned char));
+			memcpy(binary->config, section_data->d_buf, binary->config_size);
+		} else if (!strcmp(name, ".AMDGPU.disasm")) {
+			/* Always read disassembly if it's available. */
+			section_data = elf_getdata(section, section_data);
+			binary->disasm_string = strndup(section_data->d_buf,
+							section_data->d_size);
+		} else if (!strncmp(name, ".rodata", 7)) {
+			section_data = elf_getdata(section, section_data);
+			binary->rodata_size = section_data->d_size;
+			binary->rodata = MALLOC(binary->rodata_size * sizeof(unsigned char));
+			memcpy(binary->rodata, section_data->d_buf, binary->rodata_size);
+		} else if (!strncmp(name, ".symtab", 7)) {
+			symbols = elf_getdata(section, section_data);
+			symbol_sh_link = section_header.sh_link;
+			parse_symbol_table(symbols, &section_header, binary);
+		} else if (!strcmp(name, ".rel.text")) {
+			relocs = elf_getdata(section, section_data);
+			binary->reloc_count = section_header.sh_size /
+					section_header.sh_entsize;
+		}
+	}
+
+	parse_relocs(elf, relocs, symbols, symbol_sh_link, binary);
+
+	if (elf){
+		elf_end(elf);
+	}
+	FREE(elf_buffer);
+
+	/* Cache the config size per symbol */
+	if (binary->global_symbol_count) {
+		binary->config_size_per_symbol =
+			binary->config_size / binary->global_symbol_count;
+	} else {
+		binary->global_symbol_count = 1;
+		binary->config_size_per_symbol = binary->config_size;
+	}
+}
+
+static
+const unsigned char *ac_shader_binary_config_start(
+	const struct ac_shader_binary *binary,
+	uint64_t symbol_offset)
+{
+	unsigned i;
+	for (i = 0; i < binary->global_symbol_count; ++i) {
+		if (binary->global_symbol_offsets[i] == symbol_offset) {
+			unsigned offset = i * binary->config_size_per_symbol;
+			return binary->config + offset;
+		}
+	}
+	return binary->config;
+}
+
+
+static const char *scratch_rsrc_dword0_symbol =
+	"SCRATCH_RSRC_DWORD0";
+
+static const char *scratch_rsrc_dword1_symbol =
+	"SCRATCH_RSRC_DWORD1";
+
+void ac_shader_binary_read_config(struct ac_shader_binary *binary,
+				  struct ac_shader_config *conf,
+				  unsigned symbol_offset)
+{
+	unsigned i;
+	const unsigned char *config =
+		ac_shader_binary_config_start(binary, symbol_offset);
+	bool really_needs_scratch = false;
+
+	/* LLVM adds SGPR spills to the scratch size.
+	 * Find out if we really need the scratch buffer.
+	 */
+	for (i = 0; i < binary->reloc_count; i++) {
+		const struct ac_shader_reloc *reloc = &binary->relocs[i];
+
+		if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name) ||
+		    !strcmp(scratch_rsrc_dword1_symbol, reloc->name)) {
+			really_needs_scratch = true;
+			break;
+		}
+	}
+
+	for (i = 0; i < binary->config_size_per_symbol; i+= 8) {
+		unsigned reg = util_le32_to_cpu(*(uint32_t*)(config + i));
+		unsigned value = util_le32_to_cpu(*(uint32_t*)(config + i + 4));
+		switch (reg) {
+		case R_00B028_SPI_SHADER_PGM_RSRC1_PS:
+		case R_00B128_SPI_SHADER_PGM_RSRC1_VS:
+		case R_00B228_SPI_SHADER_PGM_RSRC1_GS:
+		case R_00B848_COMPUTE_PGM_RSRC1:
+			conf->num_sgprs = MAX2(conf->num_sgprs, (G_00B028_SGPRS(value) + 1) * 8);
+			conf->num_vgprs = MAX2(conf->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4);
+			conf->float_mode =  G_00B028_FLOAT_MODE(value);
+			break;
+		case R_00B02C_SPI_SHADER_PGM_RSRC2_PS:
+			conf->lds_size = MAX2(conf->lds_size, G_00B02C_EXTRA_LDS_SIZE(value));
+			break;
+		case R_00B84C_COMPUTE_PGM_RSRC2:
+			conf->lds_size = MAX2(conf->lds_size, G_00B84C_LDS_SIZE(value));
+			break;
+		case R_0286CC_SPI_PS_INPUT_ENA:
+			conf->spi_ps_input_ena = value;
+			break;
+		case R_0286D0_SPI_PS_INPUT_ADDR:
+			conf->spi_ps_input_addr = value;
+			break;
+		case R_0286E8_SPI_TMPRING_SIZE:
+		case R_00B860_COMPUTE_TMPRING_SIZE:
+			/* WAVESIZE is in units of 256 dwords. */
+			if (really_needs_scratch)
+				conf->scratch_bytes_per_wave =
+					G_00B860_WAVESIZE(value) * 256 * 4;
+			break;
+		case SPILLED_SGPRS:
+			conf->spilled_sgprs = value;
+			break;
+		case SPILLED_VGPRS:
+			conf->spilled_vgprs = value;
+			break;
+		default:
+			{
+				static bool printed;
+
+				if (!printed) {
+					fprintf(stderr, "Warning: LLVM emitted unknown "
+						"config register: 0x%x\n", reg);
+					printed = true;
+				}
+			}
+			break;
+		}
+
+		if (!conf->spi_ps_input_addr)
+			conf->spi_ps_input_addr = conf->spi_ps_input_ena;
+	}
+}
diff --git a/src/amd/common/ac_binary.h b/src/amd/common/ac_binary.h
new file mode 100644
index 0000000000..282f33d22b
--- /dev/null
+++ b/src/amd/common/ac_binary.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright 2014 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors: Tom Stellard <thomas.stellard@amd.com>
+ *
+ */
+
+#pragma once
+
+#include <stdint.h>
+
+struct ac_shader_reloc {
+	char name[32];
+	uint64_t offset;
+};
+
+struct ac_shader_binary {
+	/** Shader code */
+	unsigned char *code;
+	unsigned code_size;
+
+	/** Config/Context register state that accompanies this shader.
+	 * This is a stream of dword pairs.  First dword contains the
+	 * register address, the second dword contains the value.*/
+	unsigned char *config;
+	unsigned config_size;
+
+	/** The number of bytes of config information for each global symbol.
+	 */
+	unsigned config_size_per_symbol;
+
+	/** Constant data accessed by the shader.  This will be uploaded
+	 * into a constant buffer. */
+	unsigned char *rodata;
+	unsigned rodata_size;
+
+	/** List of symbol offsets for the shader */
+	uint64_t *global_symbol_offsets;
+	unsigned global_symbol_count;
+
+	struct ac_shader_reloc *relocs;
+	unsigned reloc_count;
+
+	/** Disassembled shader in a string. */
+	char *disasm_string;
+};
+
+struct ac_shader_config {
+	unsigned num_sgprs;
+	unsigned num_vgprs;
+	unsigned spilled_sgprs;
+	unsigned spilled_vgprs;
+	unsigned lds_size;
+	unsigned spi_ps_input_ena;
+	unsigned spi_ps_input_addr;
+	unsigned float_mode;
+	unsigned scratch_bytes_per_wave;
+};
+
+/*
+ * Parse the elf binary stored in \p elf_data and create a
+ * ac_shader_binary object.
+ */
+void ac_elf_read(const char *elf_data, unsigned elf_size,
+		 struct ac_shader_binary *binary);
+
+void ac_shader_binary_read_config(struct ac_shader_binary *binary,
+				  struct ac_shader_config *conf,
+				  unsigned symbol_offset);
diff --git a/src/amd/common/ac_llvm_helper.cpp b/src/amd/common/ac_llvm_helper.cpp
new file mode 100644
index 0000000000..062f0aad92
--- /dev/null
+++ b/src/amd/common/ac_llvm_helper.cpp
@@ -0,0 +1,46 @@
+/*
+ * Copyright 2014 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ */
+
+/* based on Marek's patch to lp_bld_misc.cpp */
+
+// Workaround http://llvm.org/PR23628
+#if HAVE_LLVM >= 0x0307
+#  pragma push_macro("DEBUG")
+#  undef DEBUG
+#endif
+
+#include "ac_nir_to_llvm.h"
+#include <llvm-c/Core.h>
+#include <llvm/Target/TargetOptions.h>
+#include <llvm/ExecutionEngine/ExecutionEngine.h>
+
+extern "C" void
+ac_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes)
+{
+   llvm::Argument *A = llvm::unwrap<llvm::Argument>(val);
+   llvm::AttrBuilder B;
+   B.addDereferenceableAttr(bytes);
+   A->addAttr(llvm::AttributeSet::get(A->getContext(), A->getArgNo() + 1,  B));
+}
diff --git a/src/amd/common/ac_llvm_util.c b/src/amd/common/ac_llvm_util.c
new file mode 100644
index 0000000000..a8408dd79e
--- /dev/null
+++ b/src/amd/common/ac_llvm_util.c
@@ -0,0 +1,142 @@
+/*
+ * Copyright 2014 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ */
+/* based on pieces from si_pipe.c and radeon_llvm_emit.c */
+#include "ac_llvm_util.h"
+
+#include <llvm-c/Core.h>
+
+#include "c11/threads.h"
+
+#include <assert.h>
+#include <stdio.h>
+
+static void ac_init_llvm_target()
+{
+#if HAVE_LLVM < 0x0307
+	LLVMInitializeR600TargetInfo();
+	LLVMInitializeR600Target();
+	LLVMInitializeR600TargetMC();
+	LLVMInitializeR600AsmPrinter();
+#else
+	LLVMInitializeAMDGPUTargetInfo();
+	LLVMInitializeAMDGPUTarget();
+	LLVMInitializeAMDGPUTargetMC();
+	LLVMInitializeAMDGPUAsmPrinter();
+#endif
+}
+
+static once_flag ac_init_llvm_target_once_flag = ONCE_FLAG_INIT;
+
+static LLVMTargetRef ac_get_llvm_target(const char *triple)
+{
+	LLVMTargetRef target = NULL;
+	char *err_message = NULL;
+
+	call_once(&ac_init_llvm_target_once_flag, ac_init_llvm_target);
+
+	if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
+		fprintf(stderr, "Cannot find target for triple %s ", triple);
+		if (err_message) {
+			fprintf(stderr, "%s\n", err_message);
+		}
+		LLVMDisposeMessage(err_message);
+		return NULL;
+	}
+	return target;
+}
+
+static const char *ac_get_llvm_processor_name(enum radeon_family family)
+{
+	switch (family) {
+	case CHIP_TAHITI:
+		return "tahiti";
+	case CHIP_PITCAIRN:
+		return "pitcairn";
+	case CHIP_VERDE:
+		return "verde";
+	case CHIP_OLAND:
+		return "oland";
+	case CHIP_HAINAN:
+		return "hainan";
+	case CHIP_BONAIRE:
+		return "bonaire";
+	case CHIP_KABINI:
+		return "kabini";
+	case CHIP_KAVERI:
+		return "kaveri";
+	case CHIP_HAWAII:
+		return "hawaii";
+	case CHIP_MULLINS:
+		return "mullins";
+	case CHIP_TONGA:
+		return "tonga";
+	case CHIP_ICELAND:
+		return "iceland";
+	case CHIP_CARRIZO:
+		return "carrizo";
+#if HAVE_LLVM <= 0x0307
+	case CHIP_FIJI:
+		return "tonga";
+	case CHIP_STONEY:
+		return "carrizo";
+#else
+	case CHIP_FIJI:
+		return "fiji";
+	case CHIP_STONEY:
+		return "stoney";
+#endif
+#if HAVE_LLVM <= 0x0308
+	case CHIP_POLARIS10:
+		return "tonga";
+	case CHIP_POLARIS11:
+		return "tonga";
+#else
+	case CHIP_POLARIS10:
+		return "polaris10";
+	case CHIP_POLARIS11:
+		return "polaris11";
+#endif
+	default:
+		return "";
+	}
+}
+
+LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family)
+{
+	assert(family >= CHIP_TAHITI);
+
+	const char *triple = "amdgcn--";
+	LLVMTargetRef target = ac_get_llvm_target(triple);
+	LLVMTargetMachineRef tm = LLVMCreateTargetMachine(
+	                             target,
+	                             triple,
+	                             ac_get_llvm_processor_name(family),
+	                             "+DumpCode,+vgpr-spilling",
+	                             LLVMCodeGenLevelDefault,
+	                             LLVMRelocDefault,
+	                             LLVMCodeModelDefault);
+
+	return tm;
+}
diff --git a/src/amd/common/ac_llvm_util.h b/src/amd/common/ac_llvm_util.h
new file mode 100644
index 0000000000..8357fbf3f3
--- /dev/null
+++ b/src/amd/common/ac_llvm_util.h
@@ -0,0 +1,31 @@
+/*
+ * Copyright 2016 Bas Nieuwenhuizen
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ */
+#pragma once
+
+#include <llvm-c/TargetMachine.h>
+
+#include "amd_family.h"
+
+LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family);
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
new file mode 100644
index 0000000000..e6ff7c8436
--- /dev/null
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -0,0 +1,4547 @@
+/*
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "ac_nir_to_llvm.h"
+#include "ac_binary.h"
+#include "sid.h"
+#include "nir/nir.h"
+#include "../vulkan/radv_descriptor_set.h"
+#include "util/bitscan.h"
+#include <llvm-c/Transforms/Scalar.h>
+
+enum radeon_llvm_calling_convention {
+	RADEON_LLVM_AMDGPU_VS = 87,
+	RADEON_LLVM_AMDGPU_GS = 88,
+	RADEON_LLVM_AMDGPU_PS = 89,
+	RADEON_LLVM_AMDGPU_CS = 90,
+};
+
+#define CONST_ADDR_SPACE 2
+#define LOCAL_ADDR_SPACE 3
+
+#define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
+#define RADEON_LLVM_MAX_OUTPUTS (VARYING_SLOT_VAR31 + 1)
+
+enum desc_type {
+	DESC_IMAGE,
+	DESC_FMASK,
+	DESC_SAMPLER,
+	DESC_BUFFER,
+};
+
+struct nir_to_llvm_context {
+	const struct ac_nir_compiler_options *options;
+	struct ac_shader_variant_info *shader_info;
+
+	LLVMContextRef context;
+	LLVMModuleRef module;
+	LLVMBuilderRef builder;
+	LLVMValueRef main_function;
+
+	struct hash_table *defs;
+	struct hash_table *phis;
+
+	LLVMValueRef descriptor_sets[4];
+	LLVMValueRef push_constants;
+	LLVMValueRef num_work_groups;
+	LLVMValueRef workgroup_ids;
+	LLVMValueRef local_invocation_ids;
+	LLVMValueRef tg_size;
+
+	LLVMValueRef vertex_buffers;
+	LLVMValueRef base_vertex;
+	LLVMValueRef start_instance;
+	LLVMValueRef vertex_id;
+	LLVMValueRef rel_auto_id;
+	LLVMValueRef vs_prim_id;
+	LLVMValueRef instance_id;
+
+	LLVMValueRef prim_mask;
+	LLVMValueRef sample_positions;
+	LLVMValueRef persp_sample, persp_center, persp_centroid;
+	LLVMValueRef linear_sample, linear_center, linear_centroid;
+	LLVMValueRef front_face;
+	LLVMValueRef ancillary;
+	LLVMValueRef frag_pos[4];
+
+	LLVMBasicBlockRef continue_block;
+	LLVMBasicBlockRef break_block;
+
+	LLVMTypeRef i1;
+	LLVMTypeRef i8;
+	LLVMTypeRef i16;
+	LLVMTypeRef i32;
+	LLVMTypeRef i64;
+	LLVMTypeRef v2i32;
+	LLVMTypeRef v3i32;
+	LLVMTypeRef v4i32;
+	LLVMTypeRef v8i32;
+	LLVMTypeRef f32;
+	LLVMTypeRef f16;
+	LLVMTypeRef v2f32;
+	LLVMTypeRef v4f32;
+	LLVMTypeRef v16i8;
+	LLVMTypeRef voidt;
+
+	LLVMValueRef i32zero;
+	LLVMValueRef i32one;
+	LLVMValueRef f32zero;
+	LLVMValueRef f32one;
+	LLVMValueRef v4f32empty;
+
+	unsigned range_md_kind;
+	unsigned uniform_md_kind;
+	unsigned fpmath_md_kind;
+	unsigned invariant_load_md_kind;
+	LLVMValueRef empty_md;
+	LLVMValueRef fpmath_md_2p5_ulp;
+	gl_shader_stage stage;
+
+	LLVMValueRef lds;
+	LLVMValueRef inputs[RADEON_LLVM_MAX_INPUTS * 4];
+	LLVMValueRef outputs[RADEON_LLVM_MAX_OUTPUTS * 4];
+
+	LLVMValueRef shared_memory;
+	uint64_t input_mask;
+	uint64_t output_mask;
+	int num_locals;
+	LLVMValueRef *locals;
+	bool has_ddxy;
+	unsigned num_clips;
+	unsigned num_culls;
+};
+
+struct ac_tex_info {
+	LLVMValueRef args[12];
+	int arg_count;
+	LLVMTypeRef dst_type;
+	bool has_offset;
+};
+
+static LLVMValueRef
+emit_llvm_intrinsic(struct nir_to_llvm_context *ctx, const char *name,
+                    LLVMTypeRef return_type, LLVMValueRef *params,
+                    unsigned param_count, LLVMAttribute attribs);
+static LLVMValueRef get_sampler_desc(struct nir_to_llvm_context *ctx,
+				     nir_deref_var *deref,
+				     enum desc_type desc_type);
+static unsigned radeon_llvm_reg_index_soa(unsigned index, unsigned chan)
+{
+	return (index * 4) + chan;
+}
+
+static unsigned llvm_get_type_size(LLVMTypeRef type)
+{
+	LLVMTypeKind kind = LLVMGetTypeKind(type);
+
+	switch (kind) {
+	case LLVMIntegerTypeKind:
+		return LLVMGetIntTypeWidth(type) / 8;
+	case LLVMFloatTypeKind:
+		return 4;
+	case LLVMPointerTypeKind:
+		return 8;
+	case LLVMVectorTypeKind:
+		return LLVMGetVectorSize(type) *
+		       llvm_get_type_size(LLVMGetElementType(type));
+	default:
+		assert(0);
+		return 0;
+	}
+}
+
+static void set_llvm_calling_convention(LLVMValueRef func,
+                                        gl_shader_stage stage)
+{
+	enum radeon_llvm_calling_convention calling_conv;
+
+	switch (stage) {
+	case MESA_SHADER_VERTEX:
+	case MESA_SHADER_TESS_CTRL:
+	case MESA_SHADER_TESS_EVAL:
+		calling_conv = RADEON_LLVM_AMDGPU_VS;
+		break;
+	case MESA_SHADER_GEOMETRY:
+		calling_conv = RADEON_LLVM_AMDGPU_GS;
+		break;
+	case MESA_SHADER_FRAGMENT:
+		calling_conv = RADEON_LLVM_AMDGPU_PS;
+		break;
+	case MESA_SHADER_COMPUTE:
+		calling_conv = RADEON_LLVM_AMDGPU_CS;
+		break;
+	default:
+		unreachable("Unhandle shader type");
+	}
+
+	LLVMSetFunctionCallConv(func, calling_conv);
+}
+
+static LLVMValueRef
+create_llvm_function(LLVMContextRef ctx, LLVMModuleRef module,
+                     LLVMBuilderRef builder, LLVMTypeRef *return_types,
+                     unsigned num_return_elems, LLVMTypeRef *param_types,
+                     unsigned param_count, unsigned array_params,
+                     unsigned sgpr_params, bool unsafe_math)
+{
+	LLVMTypeRef main_function_type, ret_type;
+	LLVMBasicBlockRef main_function_body;
+
+	if (num_return_elems)
+		ret_type = LLVMStructTypeInContext(ctx, return_types,
+		                                   num_return_elems, true);
+	else
+		ret_type = LLVMVoidTypeInContext(ctx);
+
+	/* Setup the function */
+	main_function_type =
+	    LLVMFunctionType(ret_type, param_types, param_count, 0);
+	LLVMValueRef main_function =
+	    LLVMAddFunction(module, "main", main_function_type);
+	main_function_body =
+	    LLVMAppendBasicBlockInContext(ctx, main_function, "main_body");
+	LLVMPositionBuilderAtEnd(builder, main_function_body);
+
+	LLVMSetFunctionCallConv(main_function, RADEON_LLVM_AMDGPU_CS);
+	for (unsigned i = 0; i < sgpr_params; ++i) {
+		LLVMValueRef P = LLVMGetParam(main_function, i);
+
+		if (i < array_params) {
+			LLVMAddAttribute(P, LLVMByValAttribute);
+			ac_add_attr_dereferenceable(P, UINT64_MAX);
+		}
+		else
+			LLVMAddAttribute(P, LLVMInRegAttribute);
+	}
+
+	if (unsafe_math) {
+		/* These were copied from some LLVM test. */
+		LLVMAddTargetDependentFunctionAttr(main_function,
+						   "less-precise-fpmad",
+						   "true");
+		LLVMAddTargetDependentFunctionAttr(main_function,
+						   "no-infs-fp-math",
+						   "true");
+		LLVMAddTargetDependentFunctionAttr(main_function,
+						   "no-nans-fp-math",
+						   "true");
+		LLVMAddTargetDependentFunctionAttr(main_function,
+						   "unsafe-fp-math",
+						   "true");
+	}
+	return main_function;
+}
+
+static LLVMTypeRef const_array(LLVMTypeRef elem_type, int num_elements)
+{
+	return LLVMPointerType(LLVMArrayType(elem_type, num_elements),
+	                       CONST_ADDR_SPACE);
+}
+
+static LLVMValueRef get_shared_memory_ptr(struct nir_to_llvm_context *ctx,
+					  int idx,
+					  LLVMTypeRef type)
+{
+	LLVMValueRef offset;
+	LLVMValueRef ptr;
+	int addr_space;
+
+	offset = LLVMConstInt(ctx->i32, idx, false);
+
+	ptr = ctx->shared_memory;
+	ptr = LLVMBuildGEP(ctx->builder, ptr, &offset, 1, "");
+	addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
+	ptr = LLVMBuildBitCast(ctx->builder, ptr, LLVMPointerType(type, addr_space), "");
+	return ptr;
+}
+
+static LLVMValueRef to_integer(struct nir_to_llvm_context *ctx, LLVMValueRef v)
+{
+	LLVMTypeRef type = LLVMTypeOf(v);
+	if (type == ctx->f32) {
+		return LLVMBuildBitCast(ctx->builder, v, ctx->i32, "");
+	} else if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
+		LLVMTypeRef elem_type = LLVMGetElementType(type);
+		if (elem_type == ctx->f32) {
+			LLVMTypeRef nt = LLVMVectorType(ctx->i32, LLVMGetVectorSize(type));
+			return LLVMBuildBitCast(ctx->builder, v, nt, "");
+		}
+	}
+	return v;
+}
+
+static LLVMValueRef to_float(struct nir_to_llvm_context *ctx, LLVMValueRef v)
+{
+	LLVMTypeRef type = LLVMTypeOf(v);
+	if (type == ctx->i32) {
+		return LLVMBuildBitCast(ctx->builder, v, ctx->f32, "");
+	} else if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
+		LLVMTypeRef elem_type = LLVMGetElementType(type);
+		if (elem_type == ctx->i32) {
+			LLVMTypeRef nt = LLVMVectorType(ctx->f32, LLVMGetVectorSize(type));
+			return LLVMBuildBitCast(ctx->builder, v, nt, "");
+		}
+	}
+	return v;
+}
+
+static LLVMValueRef build_indexed_load(struct nir_to_llvm_context *ctx,
+				       LLVMValueRef base_ptr, LLVMValueRef index,
+				       bool uniform)
+{
+	LLVMValueRef pointer;
+	LLVMValueRef indices[] = {ctx->i32zero, index};
+
+	pointer = LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
+	if (uniform)
+		LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
+	return LLVMBuildLoad(ctx->builder, pointer, "");
+}
+
+static LLVMValueRef build_indexed_load_const(struct nir_to_llvm_context *ctx,
+					     LLVMValueRef base_ptr, LLVMValueRef index)
+{
+	LLVMValueRef result = build_indexed_load(ctx, base_ptr, index, true);
+	LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
+	return result;
+}
+
+static void create_function(struct nir_to_llvm_context *ctx,
+                            struct nir_shader *nir)
+{
+	LLVMTypeRef arg_types[23];
+	unsigned arg_idx = 0;
+	unsigned array_count = 0;
+	unsigned sgpr_count = 0, user_sgpr_count;
+	unsigned i;
+	for (unsigned i = 0; i < 4; ++i)
+		arg_types[arg_idx++] = const_array(ctx->i8, 1024 * 1024);
+
+	arg_types[arg_idx++] = const_array(ctx->i8, 1024 * 1024);
+
+	array_count = arg_idx;
+	switch (nir->stage) {
+	case MESA_SHADER_COMPUTE:
+		arg_types[arg_idx++] = LLVMVectorType(ctx->i32, 3); /* grid size */
+		user_sgpr_count = arg_idx;
+		arg_types[arg_idx++] = LLVMVectorType(ctx->i32, 3);
+		arg_types[arg_idx++] = ctx->i32;
+		sgpr_count = arg_idx;
+
+		arg_types[arg_idx++] = LLVMVectorType(ctx->i32, 3);
+		break;
+	case MESA_SHADER_VERTEX:
+		arg_types[arg_idx++] = const_array(ctx->v16i8, 16);
+		arg_types[arg_idx++] = ctx->i32; // base vertex
+		arg_types[arg_idx++] = ctx->i32; // start instance
+		user_sgpr_count = sgpr_count = arg_idx;
+		arg_types[arg_idx++] = ctx->i32; // vertex id
+		arg_types[arg_idx++] = ctx->i32; // rel auto id
+		arg_types[arg_idx++] = ctx->i32; // vs prim id
+		arg_types[arg_idx++] = ctx->i32; // instance id
+		break;
+	case MESA_SHADER_FRAGMENT:
+		arg_types[arg_idx++] = const_array(ctx->f32, 32);
+		user_sgpr_count = arg_idx;
+		arg_types[arg_idx++] = ctx->i32; /* prim mask */
+		sgpr_count = arg_idx;
+		arg_types[arg_idx++] = ctx->v2i32; /* persp sample */
+		arg_types[arg_idx++] = ctx->v2i32; /* persp center */
+		arg_types[arg_idx++] = ctx->v2i32; /* persp centroid */
+		arg_types[arg_idx++] = ctx->v3i32; /* persp pull model */
+		arg_types[arg_idx++] = ctx->v2i32; /* linear sample */
+		arg_types[arg_idx++] = ctx->v2i32; /* linear center */
+		arg_types[arg_idx++] = ctx->v2i32; /* linear centroid */
+		arg_types[arg_idx++] = ctx->f32;  /* line stipple tex */
+		arg_types[arg_idx++] = ctx->f32;  /* pos x float */
+		arg_types[arg_idx++] = ctx->f32;  /* pos y float */
+		arg_types[arg_idx++] = ctx->f32;  /* pos z float */
+		arg_types[arg_idx++] = ctx->f32;  /* pos w float */
+		arg_types[arg_idx++] = ctx->i32;  /* front face */
+		arg_types[arg_idx++] = ctx->i32;  /* ancillary */
+		arg_types[arg_idx++] = ctx->f32;  /* sample coverage */
+		arg_types[arg_idx++] = ctx->i32;  /* fixed pt */
+		break;
+	default:
+		unreachable("Shader stage not implemented");
+	}
+
+	ctx->main_function = create_llvm_function(
+	    ctx->context, ctx->module, ctx->builder, NULL, 0, arg_types,
+	    arg_idx, array_count, sgpr_count, ctx->options->unsafe_math);
+	set_llvm_calling_convention(ctx->main_function, nir->stage);
+
+
+	ctx->shader_info->num_input_sgprs = 0;
+	ctx->shader_info->num_input_vgprs = 0;
+
+	for (i = 0; i < user_sgpr_count; i++)
+		ctx->shader_info->num_user_sgprs += llvm_get_type_size(arg_types[i]) / 4;
+
+	ctx->shader_info->num_input_sgprs = ctx->shader_info->num_user_sgprs;
+	for (; i < sgpr_count; i++)
+		ctx->shader_info->num_input_sgprs += llvm_get_type_size(arg_types[i]) / 4;
+
+	if (nir->stage != MESA_SHADER_FRAGMENT)
+		for (; i < arg_idx; ++i)
+			ctx->shader_info->num_input_vgprs += llvm_get_type_size(arg_types[i]) / 4;
+
+	arg_idx = 0;
+	for (unsigned i = 0; i < 4; ++i)
+		ctx->descriptor_sets[i] =
+		    LLVMGetParam(ctx->main_function, arg_idx++);
+
+	ctx->push_constants = LLVMGetParam(ctx->main_function, arg_idx++);
+
+	switch (nir->stage) {
+	case MESA_SHADER_COMPUTE:
+		ctx->num_work_groups =
+		    LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->workgroup_ids =
+		    LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->tg_size =
+		    LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->local_invocation_ids =
+		    LLVMGetParam(ctx->main_function, arg_idx++);
+		break;
+	case MESA_SHADER_VERTEX:
+		ctx->vertex_buffers = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->base_vertex = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->start_instance = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->vertex_id = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->rel_auto_id = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->vs_prim_id = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->instance_id = LLVMGetParam(ctx->main_function, arg_idx++);
+		break;
+	case MESA_SHADER_FRAGMENT:
+		ctx->sample_positions = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->prim_mask = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->persp_sample = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->persp_center = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->persp_centroid = LLVMGetParam(ctx->main_function, arg_idx++);
+		arg_idx++;
+		ctx->linear_sample = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->linear_center = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->linear_centroid = LLVMGetParam(ctx->main_function, arg_idx++);
+		arg_idx++; /* line stipple */
+		ctx->frag_pos[0] = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->frag_pos[1] = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->frag_pos[2] = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->frag_pos[3] = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->front_face = LLVMGetParam(ctx->main_function, arg_idx++);
+		ctx->ancillary = LLVMGetParam(ctx->main_function, arg_idx++);
+		break;
+	default:
+		unreachable("Shader stage not implemented");
+	}
+}
+
+static void setup_types(struct nir_to_llvm_context *ctx)
+{
+	LLVMValueRef args[4];
+
+	ctx->voidt = LLVMVoidTypeInContext(ctx->context);
+	ctx->i1 = LLVMIntTypeInContext(ctx->context, 1);
+	ctx->i8 = LLVMIntTypeInContext(ctx->context, 8);
+	ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
+	ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
+	ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
+	ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
+	ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
+	ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
+	ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
+	ctx->f32 = LLVMFloatTypeInContext(ctx->context);
+	ctx->f16 = LLVMHalfTypeInContext(ctx->context);
+	ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
+	ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
+	ctx->v16i8 = LLVMVectorType(ctx->i8, 16);
+
+	ctx->i32zero = LLVMConstInt(ctx->i32, 0, false);
+	ctx->i32one = LLVMConstInt(ctx->i32, 1, false);
+	ctx->f32zero = LLVMConstReal(ctx->f32, 0.0);
+	ctx->f32one = LLVMConstReal(ctx->f32, 1.0);
+
+	args[0] = ctx->f32zero;
+	args[1] = ctx->f32zero;
+	args[2] = ctx->f32zero;
+	args[3] = ctx->f32one;
+	ctx->v4f32empty = LLVMConstVector(args, 4);
+
+	ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
+						      "range", 5);
+	ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
+							       "invariant.load", 14);
+	ctx->uniform_md_kind =
+	    LLVMGetMDKindIDInContext(ctx->context, "amdgpu.uniform", 14);
+	ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
+
+	ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
+
+	args[0] = LLVMConstReal(ctx->f32, 2.5);
+	ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
+}
+
+static int get_llvm_num_components(LLVMValueRef value)
+{
+	LLVMTypeRef type = LLVMTypeOf(value);
+	unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
+	                              ? LLVMGetVectorSize(type)
+	                              : 1;
+	return num_components;
+}
+
+static LLVMValueRef llvm_extract_elem(struct nir_to_llvm_context *ctx,
+				      LLVMValueRef value,
+				      int index)
+{
+	int count = get_llvm_num_components(value);
+
+	assert(index < count);
+	if (count == 1)
+		return value;
+
+	return LLVMBuildExtractElement(ctx->builder, value,
+				       LLVMConstInt(ctx->i32, index, false), "");
+}
+
+static LLVMValueRef trim_vector(struct nir_to_llvm_context *ctx,
+                                LLVMValueRef value, unsigned count)
+{
+	unsigned num_components = get_llvm_num_components(value);
+	if (count == num_components)
+		return value;
+
+	LLVMValueRef masks[] = {
+	    LLVMConstInt(ctx->i32, 0, false), LLVMConstInt(ctx->i32, 1, false),
+	    LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false)};
+
+	if (count == 1)
+		return LLVMBuildExtractElement(ctx->builder, value, masks[0],
+		                               "");
+
+	LLVMValueRef swizzle = LLVMConstVector(masks, count);
+	return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
+}
+
+static LLVMValueRef
+build_gather_values_extended(struct nir_to_llvm_context *ctx,
+			     LLVMValueRef *values,
+			     unsigned value_count,
+			     unsigned value_stride,
+			     bool load)
+{
+	LLVMBuilderRef builder = ctx->builder;
+	LLVMValueRef vec;
+	unsigned i;
+
+
+	if (value_count == 1) {
+		if (load)
+			return LLVMBuildLoad(builder, values[0], "");
+		return values[0];
+	}
+
+	for (i = 0; i < value_count; i++) {
+		LLVMValueRef value = values[i * value_stride];
+		if (load)
+			value = LLVMBuildLoad(builder, value, "");
+
+		if (!i)
+			vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
+		LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
+		vec = LLVMBuildInsertElement(builder, vec, value, index, "");
+	}
+	return vec;
+}
+
+
+static void
+build_store_values_extended(struct nir_to_llvm_context *ctx,
+			     LLVMValueRef *values,
+			     unsigned value_count,
+			     unsigned value_stride,
+			     LLVMValueRef vec)
+{
+	LLVMBuilderRef builder = ctx->builder;
+	unsigned i;
+
+	if (value_count == 1) {
+		LLVMBuildStore(builder, vec, values[0]);
+		return;
+	}
+
+	for (i = 0; i < value_count; i++) {
+		LLVMValueRef ptr = values[i * value_stride];
+		LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
+		LLVMValueRef value = LLVMBuildExtractElement(builder, vec, index, "");
+		LLVMBuildStore(builder, value, ptr);
+	}
+}
+
+static LLVMValueRef
+build_gather_values(struct nir_to_llvm_context *ctx,
+		    LLVMValueRef *values,
+		    unsigned value_count)
+{
+	return build_gather_values_extended(ctx, values, value_count, 1, false);
+}
+
+static LLVMTypeRef get_def_type(struct nir_to_llvm_context *ctx,
+                                nir_ssa_def *def)
+{
+	LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, def->bit_size);
+	if (def->num_components > 1) {
+		type = LLVMVectorType(type, def->num_components);
+	}
+	return type;
+}
+
+static LLVMValueRef get_src(struct nir_to_llvm_context *ctx, nir_src src)
+{
+	assert(src.is_ssa);
+	struct hash_entry *entry = _mesa_hash_table_search(ctx->defs, src.ssa);
+	return (LLVMValueRef)entry->data;
+}
+
+
+static LLVMBasicBlockRef get_block(struct nir_to_llvm_context *ctx,
+                                   struct nir_block *b)
+{
+	struct hash_entry *entry = _mesa_hash_table_search(ctx->defs, b);
+	return (LLVMBasicBlockRef)entry->data;
+}
+
+static LLVMValueRef get_alu_src(struct nir_to_llvm_context *ctx,
+                                nir_alu_src src,
+                                unsigned num_components)
+{
+	LLVMValueRef value = get_src(ctx, src.src);
+	bool need_swizzle = false;
+
+	assert(value);
+	LLVMTypeRef type = LLVMTypeOf(value);
+	unsigned src_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
+	                              ? LLVMGetVectorSize(type)
+	                              : 1;
+
+	for (unsigned i = 0; i < num_components; ++i) {
+		assert(src.swizzle[i] < src_components);
+		if (src.swizzle[i] != i)
+			need_swizzle = true;
+	}
+
+	if (need_swizzle || num_components != src_components) {
+		LLVMValueRef masks[] = {
+		    LLVMConstInt(ctx->i32, src.swizzle[0], false),
+		    LLVMConstInt(ctx->i32, src.swizzle[1], false),
+		    LLVMConstInt(ctx->i32, src.swizzle[2], false),
+		    LLVMConstInt(ctx->i32, src.swizzle[3], false)};
+
+		if (src_components > 1 && num_components == 1) {
+			value = LLVMBuildExtractElement(ctx->builder, value,
+			                                masks[0], "");
+		} else if (src_components == 1 && num_components > 1) {
+			LLVMValueRef values[] = {value, value, value, value};
+			value = build_gather_values(ctx, values, num_components);
+		} else {
+			LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
+			value = LLVMBuildShuffleVector(ctx->builder, value, value,
+		                                       swizzle, "");
+		}
+	}
+	assert(!src.negate);
+	assert(!src.abs);
+	return value;
+}
+
+static LLVMValueRef emit_int_cmp(struct nir_to_llvm_context *ctx,
+                                 LLVMIntPredicate pred, LLVMValueRef src0,
+                                 LLVMValueRef src1)
+{
+	LLVMValueRef result = LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
+	return LLVMBuildSelect(ctx->builder, result,
+	                       LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
+	                       LLVMConstInt(ctx->i32, 0, false), "");
+}
+
+static LLVMValueRef emit_float_cmp(struct nir_to_llvm_context *ctx,
+                                   LLVMRealPredicate pred, LLVMValueRef src0,
+                                   LLVMValueRef src1)
+{
+	LLVMValueRef result;
+	src0 = to_float(ctx, src0);
+	src1 = to_float(ctx, src1);
+	result = LLVMBuildFCmp(ctx->builder, pred, src0, src1, "");
+	return LLVMBuildSelect(ctx->builder, result,
+	                       LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
+	                       LLVMConstInt(ctx->i32, 0, false), "");
+}
+
+static LLVMValueRef emit_intrin_1f_param(struct nir_to_llvm_context *ctx,
+					 const char *intrin,
+					 LLVMValueRef src0)
+{
+	LLVMValueRef params[] = {
+		to_float(ctx, src0),
+	};
+	return emit_llvm_intrinsic(ctx, intrin, ctx->f32, params, 1, LLVMReadNoneAttribute);
+}
+
+static LLVMValueRef emit_intrin_2f_param(struct nir_to_llvm_context *ctx,
+				       const char *intrin,
+				       LLVMValueRef src0, LLVMValueRef src1)
+{
+	LLVMValueRef params[] = {
+		to_float(ctx, src0),
+		to_float(ctx, src1),
+	};
+	return emit_llvm_intrinsic(ctx, intrin, ctx->f32, params, 2, LLVMReadNoneAttribute);
+}
+
+static LLVMValueRef emit_intrin_3f_param(struct nir_to_llvm_context *ctx,
+					 const char *intrin,
+					 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
+{
+	LLVMValueRef params[] = {
+		to_float(ctx, src0),
+		to_float(ctx, src1),
+		to_float(ctx, src2),
+	};
+	return emit_llvm_intrinsic(ctx, intrin, ctx->f32, params, 3, LLVMReadNoneAttribute);
+}
+
+static LLVMValueRef emit_bcsel(struct nir_to_llvm_context *ctx,
+			       LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
+{
+	LLVMValueRef v = LLVMBuildICmp(ctx->builder, LLVMIntNE, src0,
+				       ctx->i32zero, "");
+	return LLVMBuildSelect(ctx->builder, v, src1, src2, "");
+}
+
+static LLVMValueRef emit_find_lsb(struct nir_to_llvm_context *ctx,
+				  LLVMValueRef src0)
+{
+	LLVMValueRef params[2] = {
+		src0,
+
+		/* The value of 1 means that ffs(x=0) = undef, so LLVM won't
+		 * add special code to check for x=0. The reason is that
+		 * the LLVM behavior for x=0 is different from what we
+		 * need here.
+		 *
+		 * The hardware already implements the correct behavior.
+		 */
+		LLVMConstInt(ctx->i32, 1, false),
+	};
+	return emit_llvm_intrinsic(ctx, "llvm.cttz.i32", ctx->i32, params, 2, LLVMReadNoneAttribute);
+}
+
+static LLVMValueRef emit_ifind_msb(struct nir_to_llvm_context *ctx,
+				   LLVMValueRef src0)
+{
+	LLVMValueRef msb = emit_llvm_intrinsic(ctx, "llvm.AMDGPU.flbit.i32",
+					       ctx->i32, &src0, 1,
+					       LLVMReadNoneAttribute);
+
+	/* The HW returns the last bit index from MSB, but NIR wants
+	 * the index from LSB. Invert it by doing "31 - msb". */
+	msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
+			   msb, "");
+
+	LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
+	LLVMValueRef cond = LLVMBuildOr(ctx->builder,
+					LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+						      src0, ctx->i32zero, ""),
+					LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+						      src0, all_ones, ""), "");
+
+	return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
+}
+
+static LLVMValueRef emit_ufind_msb(struct nir_to_llvm_context *ctx,
+				   LLVMValueRef src0)
+{
+	LLVMValueRef args[2] = {
+		src0,
+		ctx->i32one,
+	};
+	LLVMValueRef msb = emit_llvm_intrinsic(ctx, "llvm.ctlz.i32",
+					       ctx->i32, args, ARRAY_SIZE(args),
+					       LLVMReadNoneAttribute);
+
+	/* The HW returns the last bit index from MSB, but NIR wants
+	 * the index from LSB. Invert it by doing "31 - msb". */
+	msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
+			   msb, "");
+
+	return LLVMBuildSelect(ctx->builder,
+			       LLVMBuildICmp(ctx->builder, LLVMIntEQ, src0,
+					     ctx->i32zero, ""),
+			       LLVMConstInt(ctx->i32, -1, true), msb, "");
+}
+
+static LLVMValueRef emit_minmax_int(struct nir_to_llvm_context *ctx,
+				    LLVMIntPredicate pred,
+				    LLVMValueRef src0, LLVMValueRef src1)
+{
+	return LLVMBuildSelect(ctx->builder,
+			       LLVMBuildICmp(ctx->builder, pred, src0, src1, ""),
+			       src0,
+			       src1, "");
+
+}
+static LLVMValueRef emit_iabs(struct nir_to_llvm_context *ctx,
+			      LLVMValueRef src0)
+{
+	return emit_minmax_int(ctx, LLVMIntSGT, src0,
+			       LLVMBuildNeg(ctx->builder, src0, ""));
+}
+
+static LLVMValueRef emit_fsign(struct nir_to_llvm_context *ctx,
+			       LLVMValueRef src0)
+{
+	LLVMValueRef cmp, val;
+
+	cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, ctx->f32zero, "");
+	val = LLVMBuildSelect(ctx->builder, cmp, ctx->f32one, src0, "");
+	cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, ctx->f32zero, "");
+	val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(ctx->f32, -1.0), "");
+	return val;
+}
+
+static LLVMValueRef emit_isign(struct nir_to_llvm_context *ctx,
+			       LLVMValueRef src0)
+{
+	LLVMValueRef cmp, val;
+
+	cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, ctx->i32zero, "");
+	val = LLVMBuildSelect(ctx->builder, cmp, ctx->i32one, src0, "");
+	cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, ctx->i32zero, "");
+	val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(ctx->i32, -1, true), "");
+	return val;
+}
+
+static LLVMValueRef emit_ffract(struct nir_to_llvm_context *ctx,
+				LLVMValueRef src0)
+{
+	const char *intr = "llvm.floor.f32";
+	LLVMValueRef fsrc0 = to_float(ctx, src0);
+	LLVMValueRef params[] = {
+		fsrc0,
+	};
+	LLVMValueRef floor = emit_llvm_intrinsic(ctx, intr,
+						 ctx->f32, params, 1,
+						 LLVMReadNoneAttribute);
+	return LLVMBuildFSub(ctx->builder, fsrc0, floor, "");
+}
+
+static LLVMValueRef emit_uint_carry(struct nir_to_llvm_context *ctx,
+				    const char *intrin,
+				    LLVMValueRef src0, LLVMValueRef src1)
+{
+	LLVMTypeRef ret_type;
+	LLVMTypeRef types[] = { ctx->i32, ctx->i1 };
+	LLVMValueRef res;
+	LLVMValueRef params[] = { src0, src1 };
+	ret_type = LLVMStructTypeInContext(ctx->context, types,
+					   2, true);
+
+	res = emit_llvm_intrinsic(ctx, intrin, ret_type,
+				  params, 2, LLVMReadNoneAttribute);
+
+	res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
+	res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
+	return res;
+}
+
+static LLVMValueRef emit_b2f(struct nir_to_llvm_context *ctx,
+			     LLVMValueRef src0)
+{
+	return LLVMBuildAnd(ctx->builder, src0, LLVMBuildBitCast(ctx->builder, LLVMConstReal(ctx->f32, 1.0), ctx->i32, ""), "");
+}
+
+static LLVMValueRef emit_umul_high(struct nir_to_llvm_context *ctx,
+				   LLVMValueRef src0, LLVMValueRef src1)
+{
+	LLVMValueRef dst64, result;
+	src0 = LLVMBuildZExt(ctx->builder, src0, ctx->i64, "");
+	src1 = LLVMBuildZExt(ctx->builder, src1, ctx->i64, "");
+
+	dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
+	dst64 = LLVMBuildLShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
+	result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
+	return result;
+}
+
+static LLVMValueRef emit_imul_high(struct nir_to_llvm_context *ctx,
+				   LLVMValueRef src0, LLVMValueRef src1)
+{
+	LLVMValueRef dst64, result;
+	src0 = LLVMBuildSExt(ctx->builder, src0, ctx->i64, "");
+	src1 = LLVMBuildSExt(ctx->builder, src1, ctx->i64, "");
+
+	dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
+	dst64 = LLVMBuildAShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
+	result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
+	return result;
+}
+
+static LLVMValueRef emit_bitfield_extract(struct nir_to_llvm_context *ctx,
+					  const char *intrin,
+					  LLVMValueRef srcs[3])
+{
+	LLVMValueRef result;
+	LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, srcs[2], LLVMConstInt(ctx->i32, 32, false), "");
+	result = emit_llvm_intrinsic(ctx, intrin, ctx->i32, srcs, 3, LLVMReadNoneAttribute);
+
+	result = LLVMBuildSelect(ctx->builder, icond, srcs[0], result, "");
+	return result;
+}
+
+static LLVMValueRef emit_bitfield_insert(struct nir_to_llvm_context *ctx,
+					 LLVMValueRef src0, LLVMValueRef src1,
+					 LLVMValueRef src2, LLVMValueRef src3)
+{
+	LLVMValueRef bfi_args[3], result;
+
+	bfi_args[0] = LLVMBuildShl(ctx->builder,
+				   LLVMBuildSub(ctx->builder,
+						LLVMBuildShl(ctx->builder,
+							     ctx->i32one,
+							     src3, ""),
+						ctx->i32one, ""),
+				   src2, "");
+	bfi_args[1] = LLVMBuildShl(ctx->builder, src1, src2, "");
+	bfi_args[2] = src0;
+
+	LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, src3, LLVMConstInt(ctx->i32, 32, false), "");
+
+	/* Calculate:
+	 *   (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
+	 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
+	 */
+	result = LLVMBuildXor(ctx->builder, bfi_args[2],
+			      LLVMBuildAnd(ctx->builder, bfi_args[0],
+					   LLVMBuildXor(ctx->builder, bfi_args[1], bfi_args[2], ""), ""), "");
+
+	result = LLVMBuildSelect(ctx->builder, icond, src1, result, "");
+	return result;
+}
+
+static LLVMValueRef emit_pack_half_2x16(struct nir_to_llvm_context *ctx,
+					LLVMValueRef src0)
+{
+	LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
+	int i;
+	LLVMValueRef comp[2];
+
+	src0 = to_float(ctx, src0);
+	comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32zero, "");
+	comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32one, "");
+	for (i = 0; i < 2; i++) {
+		comp[i] = LLVMBuildFPTrunc(ctx->builder, comp[i], ctx->f16, "");
+		comp[i] = LLVMBuildBitCast(ctx->builder, comp[i], ctx->i16, "");
+		comp[i] = LLVMBuildZExt(ctx->builder, comp[i], ctx->i32, "");
+	}
+
+	comp[1] = LLVMBuildShl(ctx->builder, comp[1], const16, "");
+	comp[0] = LLVMBuildOr(ctx->builder, comp[0], comp[1], "");
+
+	return comp[0];
+}
+
+static LLVMValueRef emit_unpack_half_2x16(struct nir_to_llvm_context *ctx,
+					  LLVMValueRef src0)
+{
+	LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
+	LLVMValueRef temps[2], result, val;
+	int i;
+
+	for (i = 0; i < 2; i++) {
+		val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
+		val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
+		val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
+		temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
+	}
+
+	result = LLVMBuildInsertElement(ctx->builder, LLVMGetUndef(ctx->v2f32), temps[0],
+					ctx->i32zero, "");
+	result = LLVMBuildInsertElement(ctx->builder, result, temps[1],
+					ctx->i32one, "");
+	return result;
+}
+
+/**
+ * Set range metadata on an instruction.  This can only be used on load and
+ * call instructions.  If you know an instruction can only produce the values
+ * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
+ * \p lo is the minimum value inclusive.
+ * \p hi is the maximum value exclusive.
+ */
+static void set_range_metadata(struct nir_to_llvm_context *ctx,
+			       LLVMValueRef value, unsigned lo, unsigned hi)
+{
+	LLVMValueRef range_md, md_args[2];
+	LLVMTypeRef type = LLVMTypeOf(value);
+	LLVMContextRef context = LLVMGetTypeContext(type);
+
+	md_args[0] = LLVMConstInt(type, lo, false);
+	md_args[1] = LLVMConstInt(type, hi, false);
+	range_md = LLVMMDNodeInContext(context, md_args, 2);
+	LLVMSetMetadata(value, ctx->range_md_kind, range_md);
+}
+
+static LLVMValueRef get_thread_id(struct nir_to_llvm_context *ctx)
+{
+	LLVMValueRef tid;
+	LLVMValueRef tid_args[2];
+	tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
+	tid_args[1] = ctx->i32zero;
+	tid_args[1] = emit_llvm_intrinsic(ctx,
+					  "llvm.amdgcn.mbcnt.lo", ctx->i32,
+					  tid_args, 2, LLVMReadNoneAttribute);
+
+	tid = emit_llvm_intrinsic(ctx,
+				  "llvm.amdgcn.mbcnt.hi", ctx->i32,
+				  tid_args, 2, LLVMReadNoneAttribute);
+	set_range_metadata(ctx, tid, 0, 64);
+	return tid;
+}
+
+/*
+ * SI implements derivatives using the local data store (LDS)
+ * All writes to the LDS happen in all executing threads at
+ * the same time. TID is the Thread ID for the current
+ * thread and is a value between 0 and 63, representing
+ * the thread's position in the wavefront.
+ *
+ * For the pixel shader threads are grouped into quads of four pixels.
+ * The TIDs of the pixels of a quad are:
+ *
+ *  +------+------+
+ *  |4n + 0|4n + 1|
+ *  +------+------+
+ *  |4n + 2|4n + 3|
+ *  +------+------+
+ *
+ * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
+ * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
+ * the current pixel's column, and masking with 0xfffffffe yields the TID
+ * of the left pixel of the current pixel's row.
+ *
+ * Adding 1 yields the TID of the pixel to the right of the left pixel, and
+ * adding 2 yields the TID of the pixel below the top pixel.
+ */
+/* masks for thread ID. */
+#define TID_MASK_TOP_LEFT 0xfffffffc
+#define TID_MASK_TOP      0xfffffffd
+#define TID_MASK_LEFT     0xfffffffe
+static LLVMValueRef emit_ddxy(struct nir_to_llvm_context *ctx,
+			      nir_alu_instr *instr,
+			      LLVMValueRef src0)
+{
+	LLVMValueRef indices[2];
+	LLVMValueRef store_ptr, load_ptr0, load_ptr1;
+	LLVMValueRef tl, trbl, result;
+	LLVMValueRef tl_tid, trbl_tid;
+	LLVMValueRef args[2];
+	unsigned mask;
+	int idx;
+	ctx->has_ddxy = true;
+	if (!ctx->lds)
+		ctx->lds = LLVMAddGlobalInAddressSpace(ctx->module,
+						       LLVMArrayType(ctx->i32, 64),
+						       "ddxy_lds", LOCAL_ADDR_SPACE);
+
+	indices[0] = ctx->i32zero;
+	indices[1] = get_thread_id(ctx);
+	store_ptr = LLVMBuildGEP(ctx->builder, ctx->lds,
+				 indices, 2, "");
+
+	if (instr->op == nir_op_fddx_fine || instr->op == nir_op_fddx)
+		mask = TID_MASK_LEFT;
+	else if (instr->op == nir_op_fddy_fine || instr->op == nir_op_fddy)
+		mask = TID_MASK_TOP;
+	else
+		mask = TID_MASK_TOP_LEFT;
+
+	tl_tid = LLVMBuildAnd(ctx->builder, indices[1],
+			      LLVMConstInt(ctx->i32, mask, false), "");
+	indices[1] = tl_tid;
+	load_ptr0 = LLVMBuildGEP(ctx->builder, ctx->lds,
+				 indices, 2, "");
+
+	/* for DDX we want to next X pixel, DDY next Y pixel. */
+	if (instr->op == nir_op_fddx_fine ||
+	    instr->op == nir_op_fddx_coarse ||
+	    instr->op == nir_op_fddx)
+		idx = 1;
+	else
+		idx = 2;
+
+	trbl_tid = LLVMBuildAdd(ctx->builder, indices[1],
+				LLVMConstInt(ctx->i32, idx, false), "");
+	indices[1] = trbl_tid;
+	load_ptr1 = LLVMBuildGEP(ctx->builder, ctx->lds,
+				 indices, 2, "");
+
+	if (ctx->options->family >= CHIP_TONGA) {
+		args[0] = LLVMBuildMul(ctx->builder, tl_tid,
+				       LLVMConstInt(ctx->i32, 4, false), "");
+		args[1] = src0;
+		tl = emit_llvm_intrinsic(ctx, "llvm.amdgcn.ds.bpermute",
+					 ctx->i32, args, 2,
+					 LLVMReadNoneAttribute);
+
+		args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
+				       LLVMConstInt(ctx->i32, 4, false), "");
+		trbl = emit_llvm_intrinsic(ctx, "llvm.amdgcn.ds.bpermute",
+					   ctx->i32, args, 2,
+					   LLVMReadNoneAttribute);
+	} else {
+		LLVMBuildStore(ctx->builder, src0, store_ptr);
+
+		tl = LLVMBuildLoad(ctx->builder, load_ptr0, "");
+		trbl = LLVMBuildLoad(ctx->builder, load_ptr1, "");
+	}
+	tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
+	trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
+	result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
+	return result;
+}
+
+/*
+ * this takes an I,J coordinate pair,
+ * and works out the X and Y derivatives.
+ * it returns DDX(I), DDX(J), DDY(I), DDY(J).
+ */
+static LLVMValueRef emit_ddxy_interp(
+	struct nir_to_llvm_context *ctx,
+	LLVMValueRef interp_ij)
+{
+	LLVMValueRef indices[2];
+	LLVMValueRef store_ptr, load_ptr_x, load_ptr_y, load_ptr_ddx, load_ptr_ddy, temp, temp2;
+	LLVMValueRef tl, tr, bl, result[4];
+	unsigned c;
+
+	if (!ctx->lds)
+		ctx->lds = LLVMAddGlobalInAddressSpace(ctx->module,
+						       LLVMArrayType(ctx->i32, 64),
+						       "ddxy_lds", LOCAL_ADDR_SPACE);
+
+	indices[0] = ctx->i32zero;
+	indices[1] = get_thread_id(ctx);
+	store_ptr = LLVMBuildGEP(ctx->builder, ctx->lds,
+				 indices, 2, "");
+
+	temp = LLVMBuildAnd(ctx->builder, indices[1],
+			    LLVMConstInt(ctx->i32, TID_MASK_LEFT, false), "");
+
+	temp2 = LLVMBuildAnd(ctx->builder, indices[1],
+			     LLVMConstInt(ctx->i32, TID_MASK_TOP, false), "");
+
+	indices[1] = temp;
+	load_ptr_x = LLVMBuildGEP(ctx->builder, ctx->lds,
+				  indices, 2, "");
+
+	indices[1] = temp2;
+	load_ptr_y = LLVMBuildGEP(ctx->builder, ctx->lds,
+				  indices, 2, "");
+
+	indices[1] = LLVMBuildAdd(ctx->builder, temp,
+				  LLVMConstInt(ctx->i32, 1, false), "");
+	load_ptr_ddx = LLVMBuildGEP(ctx->builder, ctx->lds,
+				   indices, 2, "");
+
+	indices[1] = LLVMBuildAdd(ctx->builder, temp2,
+				  LLVMConstInt(ctx->i32, 2, false), "");
+	load_ptr_ddy = LLVMBuildGEP(ctx->builder, ctx->lds,
+				   indices, 2, "");
+
+	for (c = 0; c < 2; ++c) {
+		LLVMValueRef store_val;
+		LLVMValueRef c_ll = LLVMConstInt(ctx->i32, c, false);
+
+		store_val = LLVMBuildExtractElement(ctx->builder,
+						    interp_ij, c_ll, "");
+		LLVMBuildStore(ctx->builder,
+			       store_val,
+			       store_ptr);
+
+		tl = LLVMBuildLoad(ctx->builder, load_ptr_x, "");
+		tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
+
+		tr = LLVMBuildLoad(ctx->builder, load_ptr_ddx, "");
+		tr = LLVMBuildBitCast(ctx->builder, tr, ctx->f32, "");
+
+		result[c] = LLVMBuildFSub(ctx->builder, tr, tl, "");
+
+		tl = LLVMBuildLoad(ctx->builder, load_ptr_y, "");
+		tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
+
+		bl = LLVMBuildLoad(ctx->builder, load_ptr_ddy, "");
+		bl = LLVMBuildBitCast(ctx->builder, bl, ctx->f32, "");
+
+		result[c + 2] = LLVMBuildFSub(ctx->builder, bl, tl, "");
+	}
+
+	return build_gather_values(ctx, result, 4);
+}
+
+static LLVMValueRef emit_fdiv(struct nir_to_llvm_context *ctx,
+			      LLVMValueRef num,
+			      LLVMValueRef den)
+{
+	LLVMValueRef ret = LLVMBuildFDiv(ctx->builder, num, den, "");
+
+	if (!LLVMIsConstant(ret))
+		LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
+	return ret;
+}
+
+static void visit_alu(struct nir_to_llvm_context *ctx, nir_alu_instr *instr)
+{
+	LLVMValueRef src[4], result = NULL;
+	unsigned num_components = instr->dest.dest.ssa.num_components;
+	unsigned src_components;
+
+	assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
+	switch (instr->op) {
+	case nir_op_vec2:
+	case nir_op_vec3:
+	case nir_op_vec4:
+		src_components = 1;
+		break;
+	case nir_op_pack_half_2x16:
+		src_components = 2;
+		break;
+	case nir_op_unpack_half_2x16:
+		src_components = 1;
+		break;
+	default:
+		src_components = num_components;
+		break;
+	}
+	for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
+		src[i] = get_alu_src(ctx, instr->src[i], src_components);
+
+	switch (instr->op) {
+	case nir_op_fmov:
+	case nir_op_imov:
+		result = src[0];
+		break;
+	case nir_op_fneg:
+	        src[0] = to_float(ctx, src[0]);
+		result = LLVMBuildFNeg(ctx->builder, src[0], "");
+		break;
+	case nir_op_ineg:
+		result = LLVMBuildNeg(ctx->builder, src[0], "");
+		break;
+	case nir_op_inot:
+		result = LLVMBuildNot(ctx->builder, src[0], "");
+		break;
+	case nir_op_iadd:
+		result = LLVMBuildAdd(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_fadd:
+		src[0] = to_float(ctx, src[0]);
+		src[1] = to_float(ctx, src[1]);
+		result = LLVMBuildFAdd(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_fsub:
+		src[0] = to_float(ctx, src[0]);
+		src[1] = to_float(ctx, src[1]);
+		result = LLVMBuildFSub(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_isub:
+		result = LLVMBuildSub(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_imul:
+		result = LLVMBuildMul(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_imod:
+		result = LLVMBuildSRem(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_umod:
+		result = LLVMBuildURem(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_fmod:
+		src[0] = to_float(ctx, src[0]);
+		src[1] = to_float(ctx, src[1]);
+		result = emit_fdiv(ctx, src[0], src[1]);
+		result = emit_intrin_1f_param(ctx, "llvm.floor.f32", result);
+		result = LLVMBuildFMul(ctx->builder, src[1] , result, "");
+		result = LLVMBuildFSub(ctx->builder, src[0], result, "");
+		break;
+	case nir_op_frem:
+		src[0] = to_float(ctx, src[0]);
+		src[1] = to_float(ctx, src[1]);
+		result = LLVMBuildFRem(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_idiv:
+		result = LLVMBuildSDiv(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_udiv:
+		result = LLVMBuildUDiv(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_fmul:
+		src[0] = to_float(ctx, src[0]);
+		src[1] = to_float(ctx, src[1]);
+		result = LLVMBuildFMul(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_fdiv:
+		src[0] = to_float(ctx, src[0]);
+		src[1] = to_float(ctx, src[1]);
+		result = emit_fdiv(ctx, src[0], src[1]);
+		break;
+	case nir_op_frcp:
+		src[0] = to_float(ctx, src[0]);
+		result = emit_fdiv(ctx, ctx->f32one, src[0]);
+		break;
+	case nir_op_iand:
+		result = LLVMBuildAnd(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_ior:
+		result = LLVMBuildOr(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_ixor:
+		result = LLVMBuildXor(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_ishl:
+		result = LLVMBuildShl(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_ishr:
+		result = LLVMBuildAShr(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_ushr:
+		result = LLVMBuildLShr(ctx->builder, src[0], src[1], "");
+		break;
+	case nir_op_ilt:
+		result = emit_int_cmp(ctx, LLVMIntSLT, src[0], src[1]);
+		break;
+	case nir_op_ine:
+		result = emit_int_cmp(ctx, LLVMIntNE, src[0], src[1]);
+		break;
+	case nir_op_ieq:
+		result = emit_int_cmp(ctx, LLVMIntEQ, src[0], src[1]);
+		break;
+	case nir_op_ige:
+		result = emit_int_cmp(ctx, LLVMIntSGE, src[0], src[1]);
+		break;
+	case nir_op_ult:
+		result = emit_int_cmp(ctx, LLVMIntULT, src[0], src[1]);
+		break;
+	case nir_op_uge:
+		result = emit_int_cmp(ctx, LLVMIntUGE, src[0], src[1]);
+		break;
+	case nir_op_feq:
+		result = emit_float_cmp(ctx, LLVMRealUEQ, src[0], src[1]);
+		break;
+	case nir_op_fne:
+		result = emit_float_cmp(ctx, LLVMRealUNE, src[0], src[1]);
+		break;
+	case nir_op_flt:
+		result = emit_float_cmp(ctx, LLVMRealULT, src[0], src[1]);
+		break;
+	case nir_op_fge:
+		result = emit_float_cmp(ctx, LLVMRealUGE, src[0], src[1]);
+		break;
+	case nir_op_fabs:
+		result = emit_intrin_1f_param(ctx, "llvm.fabs.f32", src[0]);
+		break;
+	case nir_op_iabs:
+		result = emit_iabs(ctx, src[0]);
+		break;
+	case nir_op_imax:
+		result = emit_minmax_int(ctx, LLVMIntSGT, src[0], src[1]);
+		break;
+	case nir_op_imin:
+		result = emit_minmax_int(ctx, LLVMIntSLT, src[0], src[1]);
+		break;
+	case nir_op_umax:
+		result = emit_minmax_int(ctx, LLVMIntUGT, src[0], src[1]);
+		break;
+	case nir_op_umin:
+		result = emit_minmax_int(ctx, LLVMIntULT, src[0], src[1]);
+		break;
+	case nir_op_isign:
+		result = emit_isign(ctx, src[0]);
+		break;
+	case nir_op_fsign:
+		src[0] = to_float(ctx, src[0]);
+		result = emit_fsign(ctx, src[0]);
+		break;
+	case nir_op_ffloor:
+		result = emit_intrin_1f_param(ctx, "llvm.floor.f32", src[0]);
+		break;
+	case nir_op_ftrunc:
+		result = emit_intrin_1f_param(ctx, "llvm.trunc.f32", src[0]);
+		break;
+	case nir_op_fceil:
+		result = emit_intrin_1f_param(ctx, "llvm.ceil.f32", src[0]);
+		break;
+	case nir_op_fround_even:
+		result = emit_intrin_1f_param(ctx, "llvm.rint.f32", src[0]);
+		break;
+	case nir_op_ffract:
+		result = emit_ffract(ctx, src[0]);
+		break;
+	case nir_op_fsin:
+		result = emit_intrin_1f_param(ctx, "llvm.sin.f32", src[0]);
+		break;
+	case nir_op_fcos:
+		result = emit_intrin_1f_param(ctx, "llvm.cos.f32", src[0]);
+		break;
+	case nir_op_fsqrt:
+		result = emit_intrin_1f_param(ctx, "llvm.sqrt.f32", src[0]);
+		break;
+	case nir_op_fexp2:
+		result = emit_intrin_1f_param(ctx, "llvm.exp2.f32", src[0]);
+		break;
+	case nir_op_flog2:
+		result = emit_intrin_1f_param(ctx, "llvm.log2.f32", src[0]);
+		break;
+	case nir_op_frsq:
+		result = emit_intrin_1f_param(ctx, "llvm.sqrt.f32", src[0]);
+		result = emit_fdiv(ctx, ctx->f32one, result);
+		break;
+	case nir_op_fpow:
+		result = emit_intrin_2f_param(ctx, "llvm.pow.f32", src[0], src[1]);
+		break;
+	case nir_op_fmax:
+		result = emit_intrin_2f_param(ctx, "llvm.maxnum.f32", src[0], src[1]);
+		break;
+	case nir_op_fmin:
+		result = emit_intrin_2f_param(ctx, "llvm.minnum.f32", src[0], src[1]);
+		break;
+	case nir_op_ffma:
+		result = emit_intrin_3f_param(ctx, "llvm.fma.f32", src[0], src[1], src[2]);
+		break;
+	case nir_op_ibitfield_extract:
+		result = emit_bitfield_extract(ctx, "llvm.AMDGPU.bfe.i32", src);
+		break;
+	case nir_op_ubitfield_extract:
+		result = emit_bitfield_extract(ctx, "llvm.AMDGPU.bfe.u32", src);
+		break;
+	case nir_op_bitfield_insert:
+		result = emit_bitfield_insert(ctx, src[0], src[1], src[2], src[3]);
+		break;
+	case nir_op_bitfield_reverse:
+		result = emit_llvm_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32, src, 1, LLVMReadNoneAttribute);
+		break;
+	case nir_op_bit_count:
+		result = emit_llvm_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32, src, 1, LLVMReadNoneAttribute);
+		break;
+	case nir_op_vec2:
+	case nir_op_vec3:
+	case nir_op_vec4:
+		for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
+			src[i] = to_integer(ctx, src[i]);
+		result = build_gather_values(ctx, src, num_components);
+		break;
+	case nir_op_f2i:
+		src[0] = to_float(ctx, src[0]);
+		result = LLVMBuildFPToSI(ctx->builder, src[0], ctx->i32, "");
+		break;
+	case nir_op_f2u:
+		src[0] = to_float(ctx, src[0]);
+		result = LLVMBuildFPToUI(ctx->builder, src[0], ctx->i32, "");
+		break;
+	case nir_op_i2f:
+		result = LLVMBuildSIToFP(ctx->builder, src[0], ctx->f32, "");
+		break;
+	case nir_op_u2f:
+		result = LLVMBuildUIToFP(ctx->builder, src[0], ctx->f32, "");
+		break;
+	case nir_op_bcsel:
+		result = emit_bcsel(ctx, src[0], src[1], src[2]);
+		break;
+	case nir_op_find_lsb:
+		result = emit_find_lsb(ctx, src[0]);
+		break;
+	case nir_op_ufind_msb:
+		result = emit_ufind_msb(ctx, src[0]);
+		break;
+	case nir_op_ifind_msb:
+		result = emit_ifind_msb(ctx, src[0]);
+		break;
+	case nir_op_uadd_carry:
+		result = emit_uint_carry(ctx, "llvm.uadd.with.overflow.i32", src[0], src[1]);
+		break;
+	case nir_op_usub_borrow:
+		result = emit_uint_carry(ctx, "llvm.usub.with.overflow.i32", src[0], src[1]);
+		break;
+	case nir_op_b2f:
+		result = emit_b2f(ctx, src[0]);
+		break;
+	case nir_op_fquantize2f16:
+		src[0] = to_float(ctx, src[0]);
+		result = LLVMBuildFPTrunc(ctx->builder, src[0], ctx->f16, "");
+		/* need to convert back up to f32 */
+		result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
+		break;
+	case nir_op_umul_high:
+		result = emit_umul_high(ctx, src[0], src[1]);
+		break;
+	case nir_op_imul_high:
+		result = emit_imul_high(ctx, src[0], src[1]);
+		break;
+	case nir_op_pack_half_2x16:
+		result = emit_pack_half_2x16(ctx, src[0]);
+		break;
+	case nir_op_unpack_half_2x16:
+		result = emit_unpack_half_2x16(ctx, src[0]);
+		break;
+	case nir_op_fddx:
+	case nir_op_fddy:
+	case nir_op_fddx_fine:
+	case nir_op_fddy_fine:
+	case nir_op_fddx_coarse:
+	case nir_op_fddy_coarse:
+		result = emit_ddxy(ctx, instr, src[0]);
+		break;
+	default:
+		fprintf(stderr, "Unknown NIR alu instr: ");
+		nir_print_instr(&instr->instr, stderr);
+		fprintf(stderr, "\n");
+		abort();
+	}
+
+	if (result) {
+		assert(instr->dest.dest.is_ssa);
+		result = to_integer(ctx, result);
+		_mesa_hash_table_insert(ctx->defs, &instr->dest.dest.ssa,
+		                        result);
+	}
+}
+
+static void visit_load_const(struct nir_to_llvm_context *ctx,
+                             nir_load_const_instr *instr)
+{
+	LLVMValueRef values[4], value = NULL;
+	LLVMTypeRef element_type =
+	    LLVMIntTypeInContext(ctx->context, instr->def.bit_size);
+
+	for (unsigned i = 0; i < instr->def.num_components; ++i) {
+		switch (instr->def.bit_size) {
+		case 32:
+			values[i] = LLVMConstInt(element_type,
+			                         instr->value.u32[i], false);
+			break;
+		case 64:
+			values[i] = LLVMConstInt(element_type,
+			                         instr->value.u64[i], false);
+			break;
+		default:
+			fprintf(stderr,
+			        "unsupported nir load_const bit_size: %d\n",
+			        instr->def.bit_size);
+			abort();
+		}
+	}
+	if (instr->def.num_components > 1) {
+		value = LLVMConstVector(values, instr->def.num_components);
+	} else
+		value = values[0];
+
+	_mesa_hash_table_insert(ctx->defs, &instr->def, value);
+}
+
+static LLVMValueRef cast_ptr(struct nir_to_llvm_context *ctx, LLVMValueRef ptr,
+                             LLVMTypeRef type)
+{
+	int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
+	return LLVMBuildBitCast(ctx->builder, ptr,
+	                        LLVMPointerType(type, addr_space), "");
+}
+
+static LLVMValueRef
+emit_llvm_intrinsic(struct nir_to_llvm_context *ctx, const char *name,
+                    LLVMTypeRef return_type, LLVMValueRef *params,
+                    unsigned param_count, LLVMAttribute attribs)
+{
+	LLVMValueRef function;
+
+	function = LLVMGetNamedFunction(ctx->module, name);
+	if (!function) {
+		LLVMTypeRef param_types[32], function_type;
+		unsigned i;
+
+		assert(param_count <= 32);
+
+		for (i = 0; i < param_count; ++i) {
+			assert(params[i]);
+			param_types[i] = LLVMTypeOf(params[i]);
+		}
+		function_type =
+		    LLVMFunctionType(return_type, param_types, param_count, 0);
+		function = LLVMAddFunction(ctx->module, name, function_type);
+
+		LLVMSetFunctionCallConv(function, LLVMCCallConv);
+		LLVMSetLinkage(function, LLVMExternalLinkage);
+
+		LLVMAddFunctionAttr(function, attribs | LLVMNoUnwindAttribute);
+	}
+	return LLVMBuildCall(ctx->builder, function, params, param_count, "");
+}
+
+static LLVMValueRef
+get_buffer_size(struct nir_to_llvm_context *ctx, LLVMValueRef descriptor, bool in_elements)
+{
+	LLVMValueRef size =
+		LLVMBuildExtractElement(ctx->builder, descriptor,
+					LLVMConstInt(ctx->i32, 2, false), "");
+
+	/* VI only */
+	if (ctx->options->chip_class >= VI && in_elements) {
+		/* On VI, the descriptor contains the size in bytes,
+		 * but TXQ must return the size in elements.
+		 * The stride is always non-zero for resources using TXQ.
+		 */
+		LLVMValueRef stride =
+			LLVMBuildExtractElement(ctx->builder, descriptor,
+						LLVMConstInt(ctx->i32, 1, false), "");
+		stride = LLVMBuildLShr(ctx->builder, stride,
+				       LLVMConstInt(ctx->i32, 16, false), "");
+		stride = LLVMBuildAnd(ctx->builder, stride,
+				      LLVMConstInt(ctx->i32, 0x3fff, false), "");
+
+		size = LLVMBuildUDiv(ctx->builder, size, stride, "");
+	}
+	return size;
+}
+
+/**
+ * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
+ * intrinsic names).
+ */
+static void build_int_type_name(
+	LLVMTypeRef type,
+	char *buf, unsigned bufsize)
+{
+	assert(bufsize >= 6);
+
+	if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
+		snprintf(buf, bufsize, "v%ui32",
+			 LLVMGetVectorSize(type));
+	else
+		strcpy(buf, "i32");
+}
+
+static LLVMValueRef radv_lower_gather4_integer(struct nir_to_llvm_context *ctx,
+					       struct ac_tex_info *tinfo,
+					       nir_tex_instr *instr,
+					       const char *intr_name,
+					       unsigned coord_vgpr_index)
+{
+	LLVMValueRef coord = tinfo->args[0];
+	LLVMValueRef half_texel[2];
+	int c;
+
+	//TODO Rect
+	{
+		LLVMValueRef txq_args[10];
+		int txq_arg_count = 0;
+		LLVMValueRef size;
+		bool da = instr->is_array || instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE;
+		txq_args[txq_arg_count++] = LLVMConstInt(ctx->i32, 0, false);
+		txq_args[txq_arg_count++] = tinfo->args[1];
+		txq_args[txq_arg_count++] = LLVMConstInt(ctx->i32, 0xf, 0); /* dmask */
+		txq_args[txq_arg_count++] = LLVMConstInt(ctx->i32, 0, 0); /* unorm */
+		txq_args[txq_arg_count++] = LLVMConstInt(ctx->i32, 0, 0); /* r128 */
+		txq_args[txq_arg_count++] = LLVMConstInt(ctx->i32, da ? 1 : 0, 0);
+		txq_args[txq_arg_count++] = LLVMConstInt(ctx->i32, 0, 0); /* glc */
+		txq_args[txq_arg_count++] = LLVMConstInt(ctx->i32, 0, 0); /* slc */
+		txq_args[txq_arg_count++] = LLVMConstInt(ctx->i32, 0, 0); /* tfe */
+		txq_args[txq_arg_count++] = LLVMConstInt(ctx->i32, 0, 0); /* lwe */
+		size = emit_llvm_intrinsic(ctx, "llvm.SI.getresinfo.i32", ctx->v4i32,
+					   txq_args, txq_arg_count,
+					   LLVMReadNoneAttribute);
+
+		for (c = 0; c < 2; c++) {
+			half_texel[c] = LLVMBuildExtractElement(ctx->builder, size,
+								ctx->i32zero, "");
+			half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
+			half_texel[c] = emit_fdiv(ctx, ctx->f32one, half_texel[c]);
+			half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
+						      LLVMConstReal(ctx->f32, -0.5), "");
+		}
+	}
+
+	for (c = 0; c < 2; c++) {
+		LLVMValueRef tmp;
+		LLVMValueRef index = LLVMConstInt(ctx->i32, coord_vgpr_index + c, 0);
+		tmp = LLVMBuildExtractElement(ctx->builder, coord, index, "");
+		tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
+		tmp = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
+		tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
+		coord = LLVMBuildInsertElement(ctx->builder, coord, tmp, index, "");
+	}
+
+	tinfo->args[0] = coord;
+	return emit_llvm_intrinsic(ctx, intr_name, tinfo->dst_type, tinfo->args, tinfo->arg_count,
+				   LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
+
+}
+
+static LLVMValueRef build_tex_intrinsic(struct nir_to_llvm_context *ctx,
+					nir_tex_instr *instr,
+					struct ac_tex_info *tinfo)
+{
+	const char *name = "llvm.SI.image.sample";
+	const char *infix = "";
+	char intr_name[127];
+	char type[64];
+	bool is_shadow = instr->is_shadow;
+	bool has_offset = tinfo->has_offset;
+	switch (instr->op) {
+	case nir_texop_txf:
+	case nir_texop_txf_ms:
+	case nir_texop_samples_identical:
+		name = instr->sampler_dim == GLSL_SAMPLER_DIM_MS ? "llvm.SI.image.load" :
+		       instr->sampler_dim == GLSL_SAMPLER_DIM_BUF ? "llvm.SI.vs.load.input" :
+			"llvm.SI.image.load.mip";
+		is_shadow = false;
+		has_offset = false;
+		break;
+	case nir_texop_txb:
+		infix = ".b";
+		break;
+	case nir_texop_txl:
+		infix = ".l";
+		break;
+	case nir_texop_txs:
+		name = "llvm.SI.getresinfo";
+		break;
+	case nir_texop_query_levels:
+		name = "llvm.SI.getresinfo";
+		break;
+	case nir_texop_tex:
+		if (ctx->stage != MESA_SHADER_FRAGMENT)
+			infix = ".lz";
+		break;
+	case nir_texop_txd:
+		infix = ".d";
+		break;
+	case nir_texop_tg4:
+		name = "llvm.SI.gather4";
+		infix = ".lz";
+		break;
+	case nir_texop_lod:
+		name = "llvm.SI.getlod";
+		is_shadow = false;
+		has_offset = false;
+		break;
+	default:
+		break;
+	}
+
+	build_int_type_name(LLVMTypeOf(tinfo->args[0]), type, sizeof(type));
+	sprintf(intr_name, "%s%s%s%s.%s", name, is_shadow ? ".c" : "", infix,
+		has_offset ? ".o" : "", type);
+
+	if (instr->op == nir_texop_tg4) {
+		enum glsl_base_type stype = glsl_get_sampler_result_type(instr->texture->var->type);
+		if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
+			return radv_lower_gather4_integer(ctx, tinfo, instr, intr_name,
+							  (int)has_offset + (int)is_shadow);
+		}
+	}
+	return emit_llvm_intrinsic(ctx, intr_name, tinfo->dst_type, tinfo->args, tinfo->arg_count,
+				   LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
+
+}
+
+static LLVMValueRef visit_vulkan_resource_index(struct nir_to_llvm_context *ctx,
+                                                nir_intrinsic_instr *instr)
+{
+	LLVMValueRef index = get_src(ctx, instr->src[0]);
+	unsigned desc_set = nir_intrinsic_desc_set(instr);
+	unsigned binding = nir_intrinsic_binding(instr);
+	LLVMValueRef desc_ptr = ctx->descriptor_sets[desc_set];
+	struct radv_descriptor_set_layout *layout = ctx->options->layout->set[desc_set].layout;
+	unsigned base_offset = layout->binding[binding].offset;
+	LLVMValueRef offset, stride;
+
+	if (layout->binding[binding].type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC ||
+	    layout->binding[binding].type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC) {
+		desc_ptr = ctx->push_constants;
+		base_offset = ctx->options->layout->push_constant_size;
+		base_offset +=  16 * layout->binding[binding].dynamic_offset_offset;
+		stride = LLVMConstInt(ctx->i32, 16, false);
+	} else
+		stride = LLVMConstInt(ctx->i32, layout->binding[binding].size, false);
+
+	offset = LLVMConstInt(ctx->i32, base_offset, false);
+	index = LLVMBuildMul(ctx->builder, index, stride, "");
+	offset = LLVMBuildAdd(ctx->builder, offset, index, "");
+
+	LLVMValueRef indices[] = {ctx->i32zero, offset};
+	desc_ptr = LLVMBuildGEP(ctx->builder, desc_ptr, indices, 2, "");
+	desc_ptr = cast_ptr(ctx, desc_ptr, ctx->v4i32);
+	LLVMSetMetadata(desc_ptr, ctx->uniform_md_kind, ctx->empty_md);
+
+	return LLVMBuildLoad(ctx->builder, desc_ptr, "");
+}
+
+static LLVMValueRef visit_load_push_constant(struct nir_to_llvm_context *ctx,
+                                             nir_intrinsic_instr *instr)
+{
+	LLVMValueRef ptr;
+
+	LLVMValueRef indices[] = {ctx->i32zero, get_src(ctx, instr->src[0])};
+	ptr = LLVMBuildGEP(ctx->builder, ctx->push_constants, indices, 2, "");
+	ptr = cast_ptr(ctx, ptr, get_def_type(ctx, &instr->dest.ssa));
+
+	return LLVMBuildLoad(ctx->builder, ptr, "");
+}
+
+static LLVMValueRef visit_get_buffer_size(struct nir_to_llvm_context *ctx,
+                                          nir_intrinsic_instr *instr)
+{
+	LLVMValueRef desc = get_src(ctx, instr->src[0]);
+
+	return get_buffer_size(ctx, desc, false);
+}
+static void visit_store_ssbo(struct nir_to_llvm_context *ctx,
+                             nir_intrinsic_instr *instr)
+{
+	const char *store_name;
+	LLVMTypeRef data_type = ctx->f32;
+	unsigned writemask = nir_intrinsic_write_mask(instr);
+	LLVMValueRef base_data, base_offset;
+	LLVMValueRef params[6];
+
+	if (ctx->stage == MESA_SHADER_FRAGMENT)
+		ctx->shader_info->fs.writes_memory = true;
+
+	params[1] = get_src(ctx, instr->src[1]);
+	params[2] = LLVMConstInt(ctx->i32, 0, false); /* vindex */
+	params[4] = LLVMConstInt(ctx->i1, 0, false);  /* glc */
+	params[5] = LLVMConstInt(ctx->i1, 0, false);  /* slc */
+
+	if (instr->num_components > 1)
+		data_type = LLVMVectorType(ctx->f32, instr->num_components);
+
+	base_data = to_float(ctx, get_src(ctx, instr->src[0]));
+	base_data = trim_vector(ctx, base_data, instr->num_components);
+	base_data = LLVMBuildBitCast(ctx->builder, base_data,
+				     data_type, "");
+	base_offset = get_src(ctx, instr->src[2]);      /* voffset */
+	while (writemask) {
+		int start, count;
+		LLVMValueRef data;
+		LLVMValueRef offset;
+		LLVMValueRef tmp;
+		u_bit_scan_consecutive_range(&writemask, &start, &count);
+
+		/* Due to an LLVM limitation, split 3-element writes
+		 * into a 2-element and a 1-element write. */
+		if (count == 3) {
+			writemask |= 1 << (start + 2);
+			count = 2;
+		}
+
+		if (count == 4) {
+			store_name = "llvm.amdgcn.buffer.store.v4f32";
+			data = base_data;
+		} else if (count == 2) {
+			tmp = LLVMBuildExtractElement(ctx->builder,
+						      base_data, LLVMConstInt(ctx->i32, start, false), "");
+			data = LLVMBuildInsertElement(ctx->builder, LLVMGetUndef(ctx->v2f32), tmp,
+						      ctx->i32zero, "");
+
+			tmp = LLVMBuildExtractElement(ctx->builder,
+						      base_data, LLVMConstInt(ctx->i32, start + 1, false), "");
+			data = LLVMBuildInsertElement(ctx->builder, data, tmp,
+						      ctx->i32one, "");
+			store_name = "llvm.amdgcn.buffer.store.v2f32";
+
+		} else {
+			assert(count == 1);
+			if (get_llvm_num_components(base_data) > 1)
+				data = LLVMBuildExtractElement(ctx->builder, base_data,
+							       LLVMConstInt(ctx->i32, start, false), "");
+			else
+				data = base_data;
+			store_name = "llvm.amdgcn.buffer.store.f32";
+		}
+
+		offset = base_offset;
+		if (start != 0) {
+			offset = LLVMBuildAdd(ctx->builder, offset, LLVMConstInt(ctx->i32, start * 4, false), "");
+		}
+		params[0] = data;
+		params[3] = offset;
+		emit_llvm_intrinsic(ctx, store_name,
+				    LLVMVoidTypeInContext(ctx->context), params, 6, 0);
+	}
+}
+
+static LLVMValueRef visit_atomic_ssbo(struct nir_to_llvm_context *ctx,
+                                      nir_intrinsic_instr *instr)
+{
+	const char *name;
+	LLVMValueRef params[5];
+	int arg_count = 0;
+	if (ctx->stage == MESA_SHADER_FRAGMENT)
+		ctx->shader_info->fs.writes_memory = true;
+
+	if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
+		params[arg_count++] = get_src(ctx, instr->src[3]);
+	}
+	params[arg_count++] = get_src(ctx, instr->src[2]);
+	params[arg_count++] = get_src(ctx, instr->src[0]);
+	params[arg_count++] = LLVMConstInt(ctx->i32, 0, false); /* vindex */
+	params[arg_count++] = get_src(ctx, instr->src[1]);      /* voffset */
+	params[arg_count++] = LLVMConstInt(ctx->i1, 0, false);  /* slc */
+
+	switch (instr->intrinsic) {
+	case nir_intrinsic_ssbo_atomic_add:
+		name = "llvm.amdgcn.buffer.atomic.add";
+		break;
+	case nir_intrinsic_ssbo_atomic_imin:
+		name = "llvm.amdgcn.buffer.atomic.smin";
+		break;
+	case nir_intrinsic_ssbo_atomic_umin:
+		name = "llvm.amdgcn.buffer.atomic.umin";
+		break;
+	case nir_intrinsic_ssbo_atomic_imax:
+		name = "llvm.amdgcn.buffer.atomic.smax";
+		break;
+	case nir_intrinsic_ssbo_atomic_umax:
+		name = "llvm.amdgcn.buffer.atomic.umax";
+		break;
+	case nir_intrinsic_ssbo_atomic_and:
+		name = "llvm.amdgcn.buffer.atomic.and";
+		break;
+	case nir_intrinsic_ssbo_atomic_or:
+		name = "llvm.amdgcn.buffer.atomic.or";
+		break;
+	case nir_intrinsic_ssbo_atomic_xor:
+		name = "llvm.amdgcn.buffer.atomic.xor";
+		break;
+	case nir_intrinsic_ssbo_atomic_exchange:
+		name = "llvm.amdgcn.buffer.atomic.swap";
+		break;
+	case nir_intrinsic_ssbo_atomic_comp_swap:
+		name = "llvm.amdgcn.buffer.atomic.cmpswap";
+		break;
+	default:
+		abort();
+	}
+
+	return emit_llvm_intrinsic(ctx, name, ctx->i32, params, arg_count, 0);
+}
+
+static LLVMValueRef visit_load_buffer(struct nir_to_llvm_context *ctx,
+                                      nir_intrinsic_instr *instr)
+{
+	const char *load_name;
+	LLVMTypeRef data_type = ctx->f32;
+	if (instr->num_components == 3)
+		data_type = LLVMVectorType(ctx->f32, 4);
+	else if (instr->num_components > 1)
+		data_type = LLVMVectorType(ctx->f32, instr->num_components);
+
+	if (instr->num_components == 4 || instr->num_components == 3)
+		load_name = "llvm.amdgcn.buffer.load.v4f32";
+	else if (instr->num_components == 2)
+		load_name = "llvm.amdgcn.buffer.load.v2f32";
+	else if (instr->num_components == 1)
+		load_name = "llvm.amdgcn.buffer.load.f32";
+	else
+		abort();
+
+	LLVMValueRef params[] = {
+	    get_src(ctx, instr->src[0]),
+	    LLVMConstInt(ctx->i32, 0, false),
+	    get_src(ctx, instr->src[1]),
+	    LLVMConstInt(ctx->i1, 0, false),
+	    LLVMConstInt(ctx->i1, 0, false),
+	};
+
+	LLVMValueRef ret =
+	    emit_llvm_intrinsic(ctx, load_name, data_type, params, 5, 0);
+
+	if (instr->num_components == 3)
+		ret = trim_vector(ctx, ret, 3);
+
+	return LLVMBuildBitCast(ctx->builder, ret,
+	                        get_def_type(ctx, &instr->dest.ssa), "");
+}
+
+static void
+radv_get_deref_offset(struct nir_to_llvm_context *ctx, nir_deref *tail,
+                      bool vs_in, unsigned *const_out, LLVMValueRef *indir_out)
+{
+	unsigned const_offset = 0;
+	LLVMValueRef offset = NULL;
+
+
+	while (tail->child != NULL) {
+		const struct glsl_type *parent_type = tail->type;
+		tail = tail->child;
+
+		if (tail->deref_type == nir_deref_type_array) {
+			nir_deref_array *deref_array = nir_deref_as_array(tail);
+			LLVMValueRef index, stride, local_offset;
+			unsigned size = glsl_count_attribute_slots(tail->type, vs_in);
+
+			const_offset += size * deref_array->base_offset;
+			if (deref_array->deref_array_type == nir_deref_array_type_direct)
+				continue;
+
+			assert(deref_array->deref_array_type == nir_deref_array_type_indirect);
+			index = get_src(ctx, deref_array->indirect);
+			stride = LLVMConstInt(ctx->i32, size, 0);
+			local_offset = LLVMBuildMul(ctx->builder, stride, index, "");
+
+			if (offset)
+				offset = LLVMBuildAdd(ctx->builder, offset, local_offset, "");
+			else
+				offset = local_offset;
+		} else if (tail->deref_type == nir_deref_type_struct) {
+			nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
+
+			for (unsigned i = 0; i < deref_struct->index; i++) {
+				const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
+				const_offset += glsl_count_attribute_slots(ft, vs_in);
+			}
+		} else
+			unreachable("unsupported deref type");
+
+	}
+
+	if (const_offset && offset)
+		offset = LLVMBuildAdd(ctx->builder, offset,
+				      LLVMConstInt(ctx->i32, const_offset, 0),
+				      "");
+
+	*const_out = const_offset;
+	*indir_out = offset;
+}
+
+static LLVMValueRef visit_load_var(struct nir_to_llvm_context *ctx,
+				   nir_intrinsic_instr *instr)
+{
+	LLVMValueRef values[4];
+	int idx = instr->variables[0]->var->data.driver_location;
+	int ve = instr->dest.ssa.num_components;
+	LLVMValueRef indir_index;
+	unsigned const_index;
+	switch (instr->variables[0]->var->data.mode) {
+	case nir_var_shader_in:
+		radv_get_deref_offset(ctx, &instr->variables[0]->deref,
+				      ctx->stage == MESA_SHADER_VERTEX,
+				      &const_index, &indir_index);
+		for (unsigned chan = 0; chan < ve; chan++) {
+			if (indir_index) {
+				unsigned count = glsl_count_attribute_slots(
+						instr->variables[0]->var->type,
+						ctx->stage == MESA_SHADER_VERTEX);
+				LLVMValueRef tmp_vec = build_gather_values_extended(
+						ctx, ctx->inputs + idx + chan, count,
+						4, false);
+
+				values[chan] = LLVMBuildExtractElement(ctx->builder,
+								       tmp_vec,
+								       indir_index, "");
+			} else
+				values[chan] = ctx->inputs[idx + chan + const_index * 4];
+		}
+		return to_integer(ctx, build_gather_values(ctx, values, ve));
+		break;
+	case nir_var_local:
+		radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
+				      &const_index, &indir_index);
+		for (unsigned chan = 0; chan < ve; chan++) {
+			if (indir_index) {
+				unsigned count = glsl_count_attribute_slots(
+					instr->variables[0]->var->type, false);
+				LLVMValueRef tmp_vec = build_gather_values_extended(
+						ctx, ctx->locals + idx + chan, count,
+						4, true);
+
+				values[chan] = LLVMBuildExtractElement(ctx->builder,
+								       tmp_vec,
+								       indir_index, "");
+			} else {
+				values[chan] = LLVMBuildLoad(ctx->builder, ctx->locals[idx + chan + const_index * 4], "");
+			}
+		}
+		return to_integer(ctx, build_gather_values(ctx, values, ve));
+	case nir_var_shader_out:
+		radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
+				      &const_index, &indir_index);
+		for (unsigned chan = 0; chan < ve; chan++) {
+			if (indir_index) {
+				unsigned count = glsl_count_attribute_slots(
+						instr->variables[0]->var->type, false);
+				LLVMValueRef tmp_vec = build_gather_values_extended(
+						ctx, ctx->outputs + idx + chan, count,
+						4, true);
+
+				values[chan] = LLVMBuildExtractElement(ctx->builder,
+								       tmp_vec,
+								       indir_index, "");
+			} else {
+			values[chan] = LLVMBuildLoad(ctx->builder,
+						     ctx->outputs[idx + chan + const_index * 4],
+						     "");
+			}
+		}
+		return to_integer(ctx, build_gather_values(ctx, values, ve));
+	case nir_var_shared: {
+		radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
+				      &const_index, &indir_index);
+		LLVMValueRef ptr = get_shared_memory_ptr(ctx, idx, ctx->i32);
+		LLVMValueRef derived_ptr;
+		LLVMValueRef index = ctx->i32zero;
+		if (indir_index)
+			index = LLVMBuildAdd(ctx->builder, index, indir_index, "");
+		derived_ptr = LLVMBuildGEP(ctx->builder, ptr, &index, 1, "");
+
+		return to_integer(ctx, LLVMBuildLoad(ctx->builder, derived_ptr, ""));
+		break;
+	}
+	default:
+		break;
+	}
+	return NULL;
+}
+
+static void
+visit_store_var(struct nir_to_llvm_context *ctx,
+				   nir_intrinsic_instr *instr)
+{
+	LLVMValueRef temp_ptr, value;
+	int idx = instr->variables[0]->var->data.driver_location;
+	LLVMValueRef src = to_float(ctx, get_src(ctx, instr->src[0]));
+	int writemask = instr->const_index[0];
+	LLVMValueRef indir_index;
+	unsigned const_index;
+	switch (instr->variables[0]->var->data.mode) {
+	case nir_var_shader_out:
+		radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
+				      &const_index, &indir_index);
+		for (unsigned chan = 0; chan < 4; chan++) {
+			int stride = 4;
+			if (!(writemask & (1 << chan)))
+				continue;
+			if (get_llvm_num_components(src) == 1)
+				value = src;
+			else
+				value = LLVMBuildExtractElement(ctx->builder, src,
+								LLVMConstInt(ctx->i32,
+									     chan, false),
+								"");
+
+			if (instr->variables[0]->var->data.location == VARYING_SLOT_CLIP_DIST0 ||
+			    instr->variables[0]->var->data.location == VARYING_SLOT_CULL_DIST0)
+				stride = 1;
+			if (indir_index) {
+				unsigned count = glsl_count_attribute_slots(
+						instr->variables[0]->var->type, false);
+				LLVMValueRef tmp_vec = build_gather_values_extended(
+						ctx, ctx->outputs + idx + chan, count,
+						stride, true);
+
+				if (get_llvm_num_components(tmp_vec) > 1) {
+					tmp_vec = LLVMBuildInsertElement(ctx->builder, tmp_vec,
+									 value, indir_index, "");
+				} else
+					tmp_vec = value;
+				build_store_values_extended(ctx, ctx->outputs + idx + chan,
+							    count, stride, tmp_vec);
+
+			} else {
+				temp_ptr = ctx->outputs[idx + chan + const_index * stride];
+
+				LLVMBuildStore(ctx->builder, value, temp_ptr);
+			}
+		}
+		break;
+	case nir_var_local:
+		radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
+				      &const_index, &indir_index);
+		for (unsigned chan = 0; chan < 4; chan++) {
+			if (!(writemask & (1 << chan)))
+				continue;
+
+			if (get_llvm_num_components(src) == 1)
+				value = src;
+			else
+				value = LLVMBuildExtractElement(ctx->builder, src,
+								LLVMConstInt(ctx->i32, chan, false), "");
+			if (indir_index) {
+				unsigned count = glsl_count_attribute_slots(
+					instr->variables[0]->var->type, false);
+				LLVMValueRef tmp_vec = build_gather_values_extended(
+					ctx, ctx->locals + idx + chan, count,
+					4, true);
+
+				tmp_vec = LLVMBuildInsertElement(ctx->builder, tmp_vec,
+								 value, indir_index, "");
+				build_store_values_extended(ctx, ctx->locals + idx + chan,
+							    count, 4, tmp_vec);
+			} else {
+				temp_ptr = ctx->locals[idx + chan + const_index * 4];
+
+				LLVMBuildStore(ctx->builder, value, temp_ptr);
+			}
+		}
+		break;
+	case nir_var_shared: {
+		LLVMValueRef ptr;
+		radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
+				      &const_index, &indir_index);
+
+		ptr = get_shared_memory_ptr(ctx, idx, ctx->i32);
+		LLVMValueRef index = ctx->i32zero;
+		LLVMValueRef derived_ptr;
+
+		if (indir_index)
+			index = LLVMBuildAdd(ctx->builder, index, indir_index, "");
+		derived_ptr = LLVMBuildGEP(ctx->builder, ptr, &index, 1, "");
+		LLVMBuildStore(ctx->builder,
+			       to_integer(ctx, src), derived_ptr);
+		break;
+	}
+	default:
+		break;
+	}
+}
+
+static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
+{
+	switch (dim) {
+	case GLSL_SAMPLER_DIM_BUF:
+		return 1;
+	case GLSL_SAMPLER_DIM_1D:
+		return array ? 2 : 1;
+	case GLSL_SAMPLER_DIM_2D:
+		return array ? 3 : 2;
+	case GLSL_SAMPLER_DIM_3D:
+	case GLSL_SAMPLER_DIM_CUBE:
+		return 3;
+	case GLSL_SAMPLER_DIM_RECT:
+	case GLSL_SAMPLER_DIM_SUBPASS:
+		return 2;
+	default:
+		break;
+	}
+	return 0;
+}
+
+static LLVMValueRef get_image_coords(struct nir_to_llvm_context *ctx,
+				     nir_intrinsic_instr *instr, bool add_frag_pos)
+{
+	const struct glsl_type *type = instr->variables[0]->var->type;
+	if(instr->variables[0]->deref.child)
+		type = instr->variables[0]->deref.child->type;
+
+	LLVMValueRef src0 = get_src(ctx, instr->src[0]);
+	LLVMValueRef coords[4];
+	LLVMValueRef masks[] = {
+		LLVMConstInt(ctx->i32, 0, false), LLVMConstInt(ctx->i32, 1, false),
+		LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false),
+	};
+	LLVMValueRef res;
+	int count;
+	count = image_type_to_components_count(glsl_get_sampler_dim(type),
+					       glsl_sampler_type_is_array(type));
+
+	if (count == 1) {
+		if (instr->src[0].ssa->num_components)
+			res = LLVMBuildExtractElement(ctx->builder, src0, masks[0], "");
+		else
+			res = src0;
+	} else {
+		int chan;
+		for (chan = 0; chan < count; ++chan) {
+			coords[chan] = LLVMBuildExtractElement(ctx->builder, src0, masks[chan], "");
+		}
+
+		if (add_frag_pos) {
+			for (chan = 0; chan < count; ++chan)
+				coords[chan] = LLVMBuildAdd(ctx->builder, coords[chan], LLVMBuildFPToUI(ctx->builder, ctx->frag_pos[chan], ctx->i32, ""), "");
+		}
+		if (count == 3) {
+			coords[3] = LLVMGetUndef(ctx->i32);
+			count = 4;
+		}
+		res = build_gather_values(ctx, coords, count);
+	}
+	return res;
+}
+
+static LLVMValueRef visit_image_load(struct nir_to_llvm_context *ctx,
+				     nir_intrinsic_instr *instr)
+{
+	LLVMValueRef params[7];
+	LLVMValueRef res;
+	char intrinsic_name[32];
+	char coords_type[8];
+	const nir_variable *var = instr->variables[0]->var;
+	const struct glsl_type *type = var->type;
+	if(instr->variables[0]->deref.child)
+		type = instr->variables[0]->deref.child->type;
+
+	type = glsl_without_array(type);
+	if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
+		params[0] = get_sampler_desc(ctx, instr->variables[0], DESC_BUFFER);
+		params[1] = LLVMBuildExtractElement(ctx->builder, get_src(ctx, instr->src[0]),
+						    LLVMConstInt(ctx->i32, 0, false), ""); /* vindex */
+		params[2] = LLVMConstInt(ctx->i32, 0, false); /* voffset */
+		params[3] = LLVMConstInt(ctx->i1, 0, false);  /* glc */
+		params[4] = LLVMConstInt(ctx->i1, 0, false);  /* slc */
+		res = emit_llvm_intrinsic(ctx, "llvm.amdgcn.buffer.load.format.v4f32", ctx->v4f32,
+					  params, 5, 0);
+
+		res = trim_vector(ctx, res, instr->dest.ssa.num_components);
+		res = to_integer(ctx, res);
+	} else {
+		bool da = glsl_sampler_type_is_array(type) ||
+		          glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE;
+		bool add_frag_pos = glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_SUBPASS;
+
+		params[0] = get_image_coords(ctx, instr, add_frag_pos);
+		params[1] = get_sampler_desc(ctx, instr->variables[0], DESC_IMAGE);
+		params[2] = LLVMConstInt(ctx->i32, 15, false); /* dmask */
+		params[3] = LLVMConstInt(ctx->i1, 0, false);  /* r128 */
+		params[4] = da ? ctx->i32one : ctx->i32zero; /* da */
+		params[5] = LLVMConstInt(ctx->i1, 0, false);  /* glc */
+		params[6] = LLVMConstInt(ctx->i1, 0, false);  /* slc */
+
+		build_int_type_name(LLVMTypeOf(params[0]),
+				coords_type, sizeof(coords_type));
+
+		snprintf(intrinsic_name, sizeof(intrinsic_name),
+			"llvm.amdgcn.image.load.%s", coords_type);
+		res = emit_llvm_intrinsic(ctx, intrinsic_name, ctx->v4f32,
+					params, 7, LLVMReadOnlyAttribute);
+	}
+	return to_integer(ctx, res);
+}
+
+static void visit_image_store(struct nir_to_llvm_context *ctx,
+			      nir_intrinsic_instr *instr)
+{
+	LLVMValueRef params[8];
+	char intrinsic_name[32];
+	char coords_type[8];
+	const nir_variable *var = instr->variables[0]->var;
+	LLVMValueRef i1false = LLVMConstInt(ctx->i1, 0, 0);
+	LLVMValueRef i1true = LLVMConstInt(ctx->i1, 1, 0);
+	const struct glsl_type *type = glsl_without_array(var->type);
+
+	if (ctx->stage == MESA_SHADER_FRAGMENT)
+		ctx->shader_info->fs.writes_memory = true;
+
+	if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
+		params[0] = to_float(ctx, get_src(ctx, instr->src[2])); /* data */
+		params[1] = get_sampler_desc(ctx, instr->variables[0], DESC_BUFFER);
+		params[2] = LLVMBuildExtractElement(ctx->builder, get_src(ctx, instr->src[0]),
+						    LLVMConstInt(ctx->i32, 0, false), ""); /* vindex */
+		params[3] = LLVMConstInt(ctx->i32, 0, false); /* voffset */
+		params[4] = i1false;  /* glc */
+		params[5] = i1false;  /* slc */
+		emit_llvm_intrinsic(ctx, "llvm.amdgcn.buffer.store.format.v4f32", ctx->voidt,
+				    params, 6, 0);
+	} else {
+		bool da = glsl_sampler_type_is_array(type) ||
+		          glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE;
+
+		params[0] = get_src(ctx, instr->src[2]); /* coords */
+		params[1] = get_image_coords(ctx, instr, false);
+		params[2] = get_sampler_desc(ctx, instr->variables[0], DESC_IMAGE);
+		params[3] = LLVMConstInt(ctx->i32, 15, false); /* dmask */
+		params[4] = i1false;  /* r128 */
+		params[5] = da ? i1true : i1false; /* da */
+		params[6] = i1false;  /* glc */
+		params[7] = i1false;  /* slc */
+
+		build_int_type_name(LLVMTypeOf(params[1]),
+				    coords_type, sizeof(coords_type));
+
+		snprintf(intrinsic_name, sizeof(intrinsic_name),
+			 "llvm.amdgcn.image.store.%s", coords_type);
+		emit_llvm_intrinsic(ctx, intrinsic_name, ctx->voidt,
+				    params, 8, 0);
+	}
+
+}
+
+static LLVMValueRef visit_image_atomic(struct nir_to_llvm_context *ctx,
+                                       nir_intrinsic_instr *instr)
+{
+	LLVMValueRef params[6];
+	int param_count = 0;
+	const nir_variable *var = instr->variables[0]->var;
+	LLVMValueRef i1false = LLVMConstInt(ctx->i1, 0, 0);
+	LLVMValueRef i1true = LLVMConstInt(ctx->i1, 1, 0);
+	const char *base_name = "llvm.amdgcn.image.atomic";
+	const char *atomic_name;
+	LLVMValueRef coords;
+	char intrinsic_name[32], coords_type[8];
+	const struct glsl_type *type = glsl_without_array(var->type);
+
+	if (ctx->stage == MESA_SHADER_FRAGMENT)
+		ctx->shader_info->fs.writes_memory = true;
+
+	params[param_count++] = get_src(ctx, instr->src[2]);
+	if (instr->intrinsic == nir_intrinsic_image_atomic_comp_swap)
+		params[param_count++] = get_src(ctx, instr->src[3]);
+
+	if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
+		params[param_count++] = get_sampler_desc(ctx, instr->variables[0], DESC_BUFFER);
+		coords = params[param_count++] = LLVMBuildExtractElement(ctx->builder, get_src(ctx, instr->src[0]),
+									LLVMConstInt(ctx->i32, 0, false), ""); /* vindex */
+		params[param_count++] = ctx->i32zero; /* voffset */
+		params[param_count++] = i1false;  /* glc */
+		params[param_count++] = i1false;  /* slc */
+	} else {
+		bool da = glsl_sampler_type_is_array(type) ||
+		          glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE;
+
+		coords = params[param_count++] = get_image_coords(ctx, instr, false);
+		params[param_count++] = get_sampler_desc(ctx, instr->variables[0], DESC_IMAGE);
+		params[param_count++] = i1false; /* r128 */
+		params[param_count++] = da ? i1true : i1false;      /* da */
+		params[param_count++] = i1false;  /* slc */
+	}
+
+	switch (instr->intrinsic) {
+	case nir_intrinsic_image_atomic_add:
+		atomic_name = "add";
+		break;
+	case nir_intrinsic_image_atomic_min:
+		atomic_name = "smin";
+		break;
+	case nir_intrinsic_image_atomic_max:
+		atomic_name = "smax";
+		break;
+	case nir_intrinsic_image_atomic_and:
+		atomic_name = "and";
+		break;
+	case nir_intrinsic_image_atomic_or:
+		atomic_name = "or";
+		break;
+	case nir_intrinsic_image_atomic_xor:
+		atomic_name = "xor";
+		break;
+	case nir_intrinsic_image_atomic_exchange:
+		atomic_name = "swap";
+		break;
+	case nir_intrinsic_image_atomic_comp_swap:
+		atomic_name = "cmpswap";
+		break;
+	default:
+		abort();
+	}
+	build_int_type_name(LLVMTypeOf(coords),
+			    coords_type, sizeof(coords_type));
+
+	snprintf(intrinsic_name, sizeof(intrinsic_name),
+			 "%s.%s.%s", base_name, atomic_name, coords_type);
+	return emit_llvm_intrinsic(ctx, intrinsic_name, ctx->i32, params, param_count, 0);
+}
+
+static LLVMValueRef visit_image_size(struct nir_to_llvm_context *ctx,
+				     nir_intrinsic_instr *instr)
+{
+	LLVMValueRef res;
+	LLVMValueRef params[10];
+	const nir_variable *var = instr->variables[0]->var;
+	const struct glsl_type *type = instr->variables[0]->var->type;
+	bool da = glsl_sampler_type_is_array(var->type) ||
+	          glsl_get_sampler_dim(var->type) == GLSL_SAMPLER_DIM_CUBE;
+	if(instr->variables[0]->deref.child)
+		type = instr->variables[0]->deref.child->type;
+
+	if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF)
+		return get_buffer_size(ctx, get_sampler_desc(ctx, instr->variables[0], DESC_BUFFER), true);
+	params[0] = ctx->i32zero;
+	params[1] = get_sampler_desc(ctx, instr->variables[0], DESC_IMAGE);
+	params[2] = LLVMConstInt(ctx->i32, 15, false);
+	params[3] = ctx->i32zero;
+	params[4] = ctx->i32zero;
+	params[5] = da ? ctx->i32one : ctx->i32zero;
+	params[6] = ctx->i32zero;
+	params[7] = ctx->i32zero;
+	params[8] = ctx->i32zero;
+	params[9] = ctx->i32zero;
+
+	res = emit_llvm_intrinsic(ctx, "llvm.SI.getresinfo.i32", ctx->v4i32,
+				  params, 10, LLVMReadNoneAttribute);
+
+	if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE &&
+	    glsl_sampler_type_is_array(type)) {
+		LLVMValueRef two = LLVMConstInt(ctx->i32, 2, false);
+		LLVMValueRef six = LLVMConstInt(ctx->i32, 6, false);
+		LLVMValueRef z = LLVMBuildExtractElement(ctx->builder, res, two, "");
+		z = LLVMBuildSDiv(ctx->builder, z, six, "");
+		res = LLVMBuildInsertElement(ctx->builder, res, z, two, "");
+	}
+	return res;
+}
+
+static void emit_waitcnt(struct nir_to_llvm_context *ctx)
+{
+	LLVMValueRef args[1] = {
+		LLVMConstInt(ctx->i32, 0xf70, false),
+	};
+	emit_llvm_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
+			    ctx->voidt, args, 1, 0);
+}
+
+static void emit_barrier(struct nir_to_llvm_context *ctx)
+{
+	// TODO tess
+	emit_llvm_intrinsic(ctx, "llvm.amdgcn.s.barrier",
+			    ctx->voidt, NULL, 0, 0);
+}
+
+static LLVMValueRef
+visit_load_local_invocation_index(struct nir_to_llvm_context *ctx)
+{
+	LLVMValueRef result;
+	LLVMValueRef thread_id = get_thread_id(ctx);
+	result = LLVMBuildAnd(ctx->builder, ctx->tg_size,
+			      LLVMConstInt(ctx->i32, 0xfc0, false), "");
+
+	return LLVMBuildAdd(ctx->builder, result, thread_id, "");
+}
+
+static LLVMValueRef visit_var_atomic(struct nir_to_llvm_context *ctx,
+				     nir_intrinsic_instr *instr)
+{
+	LLVMValueRef ptr, result;
+	int idx = instr->variables[0]->var->data.driver_location;
+	LLVMValueRef src = get_src(ctx, instr->src[0]);
+	ptr = get_shared_memory_ptr(ctx, idx, ctx->i32);
+
+	if (instr->intrinsic == nir_intrinsic_var_atomic_comp_swap) {
+		LLVMValueRef src1 = get_src(ctx, instr->src[1]);
+		result = LLVMBuildAtomicCmpXchg(ctx->builder,
+						ptr, src, src1,
+						LLVMAtomicOrderingSequentiallyConsistent,
+						LLVMAtomicOrderingSequentiallyConsistent,
+						false);
+	} else {
+		LLVMAtomicRMWBinOp op;
+		switch (instr->intrinsic) {
+		case nir_intrinsic_var_atomic_add:
+			op = LLVMAtomicRMWBinOpAdd;
+			break;
+		case nir_intrinsic_var_atomic_umin:
+			op = LLVMAtomicRMWBinOpUMin;
+			break;
+		case nir_intrinsic_var_atomic_umax:
+			op = LLVMAtomicRMWBinOpUMax;
+			break;
+		case nir_intrinsic_var_atomic_imin:
+			op = LLVMAtomicRMWBinOpMin;
+			break;
+		case nir_intrinsic_var_atomic_imax:
+			op = LLVMAtomicRMWBinOpMax;
+			break;
+		case nir_intrinsic_var_atomic_and:
+			op = LLVMAtomicRMWBinOpAnd;
+			break;
+		case nir_intrinsic_var_atomic_or:
+			op = LLVMAtomicRMWBinOpOr;
+			break;
+		case nir_intrinsic_var_atomic_xor:
+			op = LLVMAtomicRMWBinOpXor;
+			break;
+		case nir_intrinsic_var_atomic_exchange:
+			op = LLVMAtomicRMWBinOpXchg;
+			break;
+		default:
+			return NULL;
+		}
+
+		result = LLVMBuildAtomicRMW(ctx->builder, op, ptr, to_integer(ctx, src),
+					    LLVMAtomicOrderingSequentiallyConsistent,
+					    false);
+	}
+	return result;
+}
+
+#define INTERP_CENTER 0
+#define INTERP_CENTROID 1
+#define INTERP_SAMPLE 2
+
+static LLVMValueRef lookup_interp_param(struct nir_to_llvm_context *ctx,
+					enum glsl_interp_mode interp, unsigned location)
+{
+	switch (interp) {
+	case INTERP_MODE_FLAT:
+	default:
+		return NULL;
+	case INTERP_MODE_SMOOTH:
+	case INTERP_MODE_NONE:
+		if (location == INTERP_CENTER)
+			return ctx->persp_center;
+		else if (location == INTERP_CENTROID)
+			return ctx->persp_centroid;
+		else if (location == INTERP_SAMPLE)
+			return ctx->persp_sample;
+		break;
+	case INTERP_MODE_NOPERSPECTIVE:
+		if (location == INTERP_CENTER)
+			return ctx->linear_center;
+		else if (location == INTERP_CENTROID)
+			return ctx->linear_centroid;
+		else if (location == INTERP_SAMPLE)
+			return ctx->linear_sample;
+		break;
+	}
+	return NULL;
+}
+
+static LLVMValueRef load_sample_position(struct nir_to_llvm_context *ctx,
+					 LLVMValueRef sample_id)
+{
+	/* offset = sample_id * 8  (8 = 2 floats containing samplepos.xy) */
+	LLVMValueRef offset0 = LLVMBuildMul(ctx->builder, sample_id, LLVMConstInt(ctx->i32, 8, false), "");
+	LLVMValueRef offset1 = LLVMBuildAdd(ctx->builder, offset0, LLVMConstInt(ctx->i32, 4, false), "");
+	LLVMValueRef result[2];
+
+	result[0] = build_indexed_load_const(ctx, ctx->sample_positions, offset0);
+	result[1] = build_indexed_load_const(ctx, ctx->sample_positions, offset1);
+
+	return build_gather_values(ctx, result, 2);
+}
+
+static LLVMValueRef visit_interp(struct nir_to_llvm_context *ctx,
+				 nir_intrinsic_instr *instr)
+{
+	LLVMValueRef result[2];
+	LLVMValueRef interp_param, attr_number;
+	unsigned location;
+	unsigned chan;
+	LLVMValueRef src_c0, src_c1;
+	const char *intr_name;
+	LLVMValueRef src0;
+	int input_index = instr->variables[0]->var->data.location - VARYING_SLOT_VAR0;
+	switch (instr->intrinsic) {
+	case nir_intrinsic_interp_var_at_centroid:
+		location = INTERP_CENTROID;
+		break;
+	case nir_intrinsic_interp_var_at_sample:
+	case nir_intrinsic_interp_var_at_offset:
+		location = INTERP_SAMPLE;
+		src0 = get_src(ctx, instr->src[0]);
+		break;
+	default:
+		break;
+	}
+
+	if (instr->intrinsic == nir_intrinsic_interp_var_at_offset) {
+		src_c0 = to_float(ctx, LLVMBuildExtractElement(ctx->builder, src0, ctx->i32zero, ""));
+		src_c1 = to_float(ctx, LLVMBuildExtractElement(ctx->builder, src0, ctx->i32one, ""));
+	} else if (instr->intrinsic == nir_intrinsic_interp_var_at_sample) {
+		LLVMValueRef sample_position;
+		LLVMValueRef halfval = LLVMConstReal(ctx->f32, 0.5f);
+
+		/* fetch sample ID */
+		sample_position = load_sample_position(ctx, src0);
+
+		src_c0 = LLVMBuildExtractElement(ctx->builder, sample_position, ctx->i32zero, "");
+		src_c0 = LLVMBuildFSub(ctx->builder, src_c0, halfval, "");
+		src_c1 = LLVMBuildExtractElement(ctx->builder, sample_position, ctx->i32one, "");
+		src_c1 = LLVMBuildFSub(ctx->builder, src_c1, halfval, "");
+	}
+	interp_param = lookup_interp_param(ctx, instr->variables[0]->var->data.interpolation, location);
+	attr_number = LLVMConstInt(ctx->i32, input_index, false);
+
+	if (location == INTERP_SAMPLE) {
+		LLVMValueRef ij_out[2];
+		LLVMValueRef ddxy_out = emit_ddxy_interp(ctx, interp_param);
+
+		/*
+		 * take the I then J parameters, and the DDX/Y for it, and
+		 * calculate the IJ inputs for the interpolator.
+		 * temp1 = ddx * offset/sample.x + I;
+		 * interp_param.I = ddy * offset/sample.y + temp1;
+		 * temp1 = ddx * offset/sample.x + J;
+		 * interp_param.J = ddy * offset/sample.y + temp1;
+		 */
+		for (unsigned i = 0; i < 2; i++) {
+			LLVMValueRef ix_ll = LLVMConstInt(ctx->i32, i, false);
+			LLVMValueRef iy_ll = LLVMConstInt(ctx->i32, i + 2, false);
+			LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->builder,
+								      ddxy_out, ix_ll, "");
+			LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->builder,
+								      ddxy_out, iy_ll, "");
+			LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->builder,
+									 interp_param, ix_ll, "");
+			LLVMValueRef temp1, temp2;
+
+			interp_el = LLVMBuildBitCast(ctx->builder, interp_el,
+						     ctx->f32, "");
+
+			temp1 = LLVMBuildFMul(ctx->builder, ddx_el, src_c0, "");
+			temp1 = LLVMBuildFAdd(ctx->builder, temp1, interp_el, "");
+
+			temp2 = LLVMBuildFMul(ctx->builder, ddy_el, src_c1, "");
+			temp2 = LLVMBuildFAdd(ctx->builder, temp2, temp1, "");
+
+			ij_out[i] = LLVMBuildBitCast(ctx->builder,
+						     temp2, ctx->i32, "");
+		}
+		interp_param = build_gather_values(ctx, ij_out, 2);
+
+	}
+	intr_name = interp_param ? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
+	for (chan = 0; chan < 2; chan++) {
+		LLVMValueRef args[4];
+		LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, false);
+
+		args[0] = llvm_chan;
+		args[1] = attr_number;
+		args[2] = ctx->prim_mask;
+		args[3] = interp_param;
+		result[chan] = emit_llvm_intrinsic(ctx, intr_name,
+						   ctx->f32, args, args[3] ? 4 : 3,
+						   LLVMReadNoneAttribute);
+	}
+	return build_gather_values(ctx, result, 2);
+}
+
+static void visit_intrinsic(struct nir_to_llvm_context *ctx,
+                            nir_intrinsic_instr *instr)
+{
+	LLVMValueRef result = NULL;
+
+	switch (instr->intrinsic) {
+	case nir_intrinsic_load_work_group_id: {
+		result = ctx->workgroup_ids;
+		break;
+	}
+	case nir_intrinsic_load_base_vertex: {
+		result = ctx->base_vertex;
+		break;
+	}
+	case nir_intrinsic_load_vertex_id_zero_base: {
+		result = ctx->vertex_id;
+		break;
+	}
+	case nir_intrinsic_load_local_invocation_id: {
+		result = ctx->local_invocation_ids;
+		break;
+	}
+	case nir_intrinsic_load_base_instance:
+		result = ctx->start_instance;
+		break;
+	case nir_intrinsic_load_sample_id:
+		result = ctx->ancillary;
+		break;
+	case nir_intrinsic_load_front_face:
+		result = ctx->front_face;
+		break;
+	case nir_intrinsic_load_instance_id:
+		result = ctx->instance_id;
+		ctx->shader_info->vs.vgpr_comp_cnt = MAX2(3,
+		                            ctx->shader_info->vs.vgpr_comp_cnt);
+		break;
+	case nir_intrinsic_load_num_work_groups:
+		result = ctx->num_work_groups;
+		break;
+	case nir_intrinsic_load_local_invocation_index:
+		result = visit_load_local_invocation_index(ctx);
+		break;
+	case nir_intrinsic_load_push_constant:
+		result = visit_load_push_constant(ctx, instr);
+		break;
+	case nir_intrinsic_vulkan_resource_index:
+		result = visit_vulkan_resource_index(ctx, instr);
+		break;
+	case nir_intrinsic_store_ssbo:
+		visit_store_ssbo(ctx, instr);
+		break;
+	case nir_intrinsic_load_ssbo:
+		result = visit_load_buffer(ctx, instr);
+		break;
+	case nir_intrinsic_ssbo_atomic_add:
+	case nir_intrinsic_ssbo_atomic_imin:
+	case nir_intrinsic_ssbo_atomic_umin:
+	case nir_intrinsic_ssbo_atomic_imax:
+	case nir_intrinsic_ssbo_atomic_umax:
+	case nir_intrinsic_ssbo_atomic_and:
+	case nir_intrinsic_ssbo_atomic_or:
+	case nir_intrinsic_ssbo_atomic_xor:
+	case nir_intrinsic_ssbo_atomic_exchange:
+	case nir_intrinsic_ssbo_atomic_comp_swap:
+		result = visit_atomic_ssbo(ctx, instr);
+		break;
+	case nir_intrinsic_load_ubo:
+		result = visit_load_buffer(ctx, instr);
+		break;
+	case nir_intrinsic_get_buffer_size:
+		result = visit_get_buffer_size(ctx, instr);
+		break;
+	case nir_intrinsic_load_var:
+		result = visit_load_var(ctx, instr);
+		break;
+	case nir_intrinsic_store_var:
+		visit_store_var(ctx, instr);
+		break;
+	case nir_intrinsic_image_load:
+		result = visit_image_load(ctx, instr);
+		break;
+	case nir_intrinsic_image_store:
+		visit_image_store(ctx, instr);
+		break;
+	case nir_intrinsic_image_atomic_add:
+	case nir_intrinsic_image_atomic_min:
+	case nir_intrinsic_image_atomic_max:
+	case nir_intrinsic_image_atomic_and:
+	case nir_intrinsic_image_atomic_or:
+	case nir_intrinsic_image_atomic_xor:
+	case nir_intrinsic_image_atomic_exchange:
+	case nir_intrinsic_image_atomic_comp_swap:
+		result = visit_image_atomic(ctx, instr);
+		break;
+	case nir_intrinsic_image_size:
+		result = visit_image_size(ctx, instr);
+		break;
+	case nir_intrinsic_discard:
+		ctx->shader_info->fs.can_discard = true;
+		emit_llvm_intrinsic(ctx, "llvm.AMDGPU.kilp",
+				    LLVMVoidTypeInContext(ctx->context),
+				    NULL, 0, 0);
+		break;
+	case nir_intrinsic_memory_barrier:
+		emit_waitcnt(ctx);
+		break;
+	case nir_intrinsic_barrier:
+		emit_barrier(ctx);
+		break;
+	case nir_intrinsic_var_atomic_add:
+	case nir_intrinsic_var_atomic_imin:
+	case nir_intrinsic_var_atomic_umin:
+	case nir_intrinsic_var_atomic_imax:
+	case nir_intrinsic_var_atomic_umax:
+	case nir_intrinsic_var_atomic_and:
+	case nir_intrinsic_var_atomic_or:
+	case nir_intrinsic_var_atomic_xor:
+	case nir_intrinsic_var_atomic_exchange:
+	case nir_intrinsic_var_atomic_comp_swap:
+		result = visit_var_atomic(ctx, instr);
+		break;
+	case nir_intrinsic_interp_var_at_centroid:
+	case nir_intrinsic_interp_var_at_sample:
+	case nir_intrinsic_interp_var_at_offset:
+		result = visit_interp(ctx, instr);
+		break;
+	default:
+		fprintf(stderr, "Unknown intrinsic: ");
+		nir_print_instr(&instr->instr, stderr);
+		fprintf(stderr, "\n");
+		break;
+	}
+	if (result) {
+		_mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
+	}
+}
+
+static LLVMValueRef get_sampler_desc(struct nir_to_llvm_context *ctx,
+					  nir_deref_var *deref,
+					  enum desc_type desc_type)
+{
+	unsigned desc_set = deref->var->data.descriptor_set;
+	LLVMValueRef list = ctx->descriptor_sets[desc_set];
+	struct radv_descriptor_set_layout *layout = ctx->options->layout->set[desc_set].layout;
+	struct radv_descriptor_set_binding_layout *binding = layout->binding + deref->var->data.binding;
+	unsigned offset = binding->offset;
+	unsigned stride = binding->size;
+	unsigned type_size;
+	LLVMBuilderRef builder = ctx->builder;
+	LLVMTypeRef type;
+	LLVMValueRef indices[2];
+	LLVMValueRef index = NULL;
+
+	assert(deref->var->data.binding < layout->binding_count);
+
+	switch (desc_type) {
+	case DESC_IMAGE:
+		type = ctx->v8i32;
+		type_size = 32;
+		break;
+	case DESC_FMASK:
+		type = ctx->v8i32;
+		offset += 32;
+		type_size = 32;
+		break;
+	case DESC_SAMPLER:
+		type = ctx->v4i32;
+		if (binding->type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
+			offset += 64;
+
+		type_size = 16;
+		break;
+	case DESC_BUFFER:
+		type = ctx->v4i32;
+		type_size = 16;
+		break;
+	}
+
+	if (deref->deref.child) {
+		nir_deref_array *child = (nir_deref_array*)deref->deref.child;
+
+		assert(child->deref_array_type != nir_deref_array_type_wildcard);
+		offset += child->base_offset * stride;
+		if (child->deref_array_type == nir_deref_array_type_indirect) {
+			index = get_src(ctx, child->indirect);
+		}
+	}
+
+	assert(stride % type_size == 0);
+
+	if (!index)
+		index = ctx->i32zero;
+
+	index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, stride / type_size, 0), "");
+	indices[0] = ctx->i32zero;
+	indices[1] = LLVMConstInt(ctx->i32, offset, 0);
+	list = LLVMBuildGEP(builder, list, indices, 2, "");
+	list = LLVMBuildPointerCast(builder, list, const_array(type, 0), "");
+
+	return build_indexed_load_const(ctx, list, index);
+}
+
+static void set_tex_fetch_args(struct nir_to_llvm_context *ctx,
+			       struct ac_tex_info *tinfo,
+			       nir_tex_instr *instr,
+			       nir_texop op,
+			       LLVMValueRef res_ptr, LLVMValueRef samp_ptr,
+			       LLVMValueRef *param, unsigned count,
+			       unsigned dmask)
+{
+	int num_args;
+	unsigned is_rect = 0;
+	bool da = instr->is_array || instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE;
+
+	if (op == nir_texop_lod)
+		da = false;
+	/* Pad to power of two vector */
+	while (count < util_next_power_of_two(count))
+		param[count++] = LLVMGetUndef(ctx->i32);
+
+	if (count > 1)
+		tinfo->args[0] = build_gather_values(ctx, param, count);
+	else
+		tinfo->args[0] = param[0];
+
+	tinfo->args[1] = res_ptr;
+	num_args = 2;
+
+	if (op == nir_texop_txf ||
+	    op == nir_texop_txf_ms ||
+	    op == nir_texop_query_levels ||
+	    op == nir_texop_texture_samples ||
+	    op == nir_texop_txs)
+		tinfo->dst_type = ctx->v4i32;
+	else {
+		tinfo->dst_type = ctx->v4f32;
+		tinfo->args[num_args++] = samp_ptr;
+	}
+
+	if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF && op == nir_texop_txf) {
+		tinfo->args[0] = res_ptr;
+		tinfo->args[1] = LLVMConstInt(ctx->i32, 0, false);
+		tinfo->args[2] = param[0];
+		tinfo->arg_count = 3;
+		return;
+	}
+
+	tinfo->args[num_args++] = LLVMConstInt(ctx->i32, dmask, 0);
+	tinfo->args[num_args++] = LLVMConstInt(ctx->i32, is_rect, 0); /* unorm */
+	tinfo->args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* r128 */
+	tinfo->args[num_args++] = LLVMConstInt(ctx->i32, da ? 1 : 0, 0);
+	tinfo->args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* glc */
+	tinfo->args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* slc */
+	tinfo->args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* tfe */
+	tinfo->args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* lwe */
+
+	tinfo->arg_count = num_args;
+}
+
+static void tex_fetch_ptrs(struct nir_to_llvm_context *ctx,
+			   nir_tex_instr *instr,
+			   LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr,
+			   LLVMValueRef *fmask_ptr)
+{
+	if (instr->sampler_dim  == GLSL_SAMPLER_DIM_BUF)
+		*res_ptr = get_sampler_desc(ctx, instr->texture, DESC_BUFFER);
+	else
+		*res_ptr = get_sampler_desc(ctx, instr->texture, DESC_IMAGE);
+	if (samp_ptr) {
+		if (instr->sampler)
+			*samp_ptr = get_sampler_desc(ctx, instr->sampler, DESC_SAMPLER);
+		else
+			*samp_ptr = get_sampler_desc(ctx, instr->texture, DESC_SAMPLER);
+	}
+	if (fmask_ptr && !instr->sampler && instr->op == nir_texop_txf_ms)
+		*fmask_ptr = get_sampler_desc(ctx, instr->texture, DESC_FMASK);
+}
+
+static LLVMValueRef build_cube_intrinsic(struct nir_to_llvm_context *ctx,
+					 LLVMValueRef *in)
+{
+
+	LLVMValueRef v, cube_vec;
+
+	if (1) {
+		LLVMTypeRef f32 = LLVMTypeOf(in[0]);
+		LLVMValueRef out[4];
+
+		out[0] = emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubetc",
+					     f32, in, 3, LLVMReadNoneAttribute);
+		out[1] = emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubesc",
+					     f32, in, 3, LLVMReadNoneAttribute);
+		out[2] = emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubema",
+					     f32, in, 3, LLVMReadNoneAttribute);
+		out[3] = emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubeid",
+					     f32, in, 3, LLVMReadNoneAttribute);
+
+		return build_gather_values(ctx, out, 4);
+	} else {
+		LLVMValueRef c[4];
+		c[0] = in[0];
+		c[1] = in[1];
+		c[2] = in[2];
+		c[3] = LLVMGetUndef(LLVMTypeOf(in[0]));
+		cube_vec = build_gather_values(ctx, c, 4);
+		v = emit_llvm_intrinsic(ctx, "llvm.AMDGPU.cube", LLVMTypeOf(cube_vec),
+					&cube_vec, 1, LLVMReadNoneAttribute);
+	}
+	return v;
+}
+
+static void cube_to_2d_coords(struct nir_to_llvm_context *ctx,
+			      LLVMValueRef *in, LLVMValueRef *out)
+{
+	LLVMValueRef coords[4];
+	LLVMValueRef mad_args[3];
+	LLVMValueRef v;
+	LLVMValueRef tmp;
+	int i;
+
+	v = build_cube_intrinsic(ctx, in);
+	for (i = 0; i < 4; i++)
+		coords[i] = LLVMBuildExtractElement(ctx->builder, v,
+						    LLVMConstInt(ctx->i32, i, false), "");
+
+	coords[2] = emit_llvm_intrinsic(ctx, "llvm.fabs.f32", ctx->f32,
+					&coords[2], 1, LLVMReadNoneAttribute);
+	coords[2] = emit_fdiv(ctx, ctx->f32one, coords[2]);
+
+	mad_args[1] = coords[2];
+	mad_args[2] = LLVMConstReal(ctx->f32, 1.5);
+	mad_args[0] = coords[0];
+
+	/* emit MAD */
+	tmp = LLVMBuildFMul(ctx->builder, mad_args[0], mad_args[1], "");
+	coords[0] = LLVMBuildFAdd(ctx->builder, tmp, mad_args[2], "");
+
+	mad_args[0] = coords[1];
+
+	/* emit MAD */
+	tmp = LLVMBuildFMul(ctx->builder, mad_args[0], mad_args[1], "");
+	coords[1] = LLVMBuildFAdd(ctx->builder, tmp, mad_args[2], "");
+
+	/* apply xyz = yxw swizzle to cooords */
+	out[0] = coords[1];
+	out[1] = coords[0];
+	out[2] = coords[3];
+}
+
+static void emit_prepare_cube_coords(struct nir_to_llvm_context *ctx,
+				     LLVMValueRef *coords_arg, int num_coords,
+				     bool is_deriv,
+				     bool is_array, LLVMValueRef *derivs_arg)
+{
+	LLVMValueRef coords[4];
+	int i;
+	cube_to_2d_coords(ctx, coords_arg, coords);
+
+	if (is_deriv && derivs_arg) {
+		LLVMValueRef derivs[4];
+		int axis;
+
+		/* Convert cube derivatives to 2D derivatives. */
+		for (axis = 0; axis < 2; axis++) {
+			LLVMValueRef shifted_cube_coords[4], shifted_coords[4];
+
+			/* Shift the cube coordinates by the derivatives to get
+			 * the cube coordinates of the "neighboring pixel".
+			 */
+			for (i = 0; i < 3; i++)
+				shifted_cube_coords[i] =
+					LLVMBuildFAdd(ctx->builder, coords_arg[i],
+						      derivs_arg[axis*3+i], "");
+			shifted_cube_coords[3] = LLVMGetUndef(ctx->f32);
+
+			/* Project the shifted cube coordinates onto the face. */
+			cube_to_2d_coords(ctx, shifted_cube_coords,
+					  shifted_coords);
+
+			/* Subtract both sets of 2D coordinates to get 2D derivatives.
+			 * This won't work if the shifted coordinates ended up
+			 * in a different face.
+			 */
+			for (i = 0; i < 2; i++)
+				derivs[axis * 2 + i] =
+					LLVMBuildFSub(ctx->builder, shifted_coords[i],
+						      coords[i], "");
+		}
+
+		memcpy(derivs_arg, derivs, sizeof(derivs));
+	}
+
+	if (is_array) {
+		/* for cube arrays coord.z = coord.w(array_index) * 8 + face */
+		/* coords_arg.w component - array_index for cube arrays */
+		LLVMValueRef tmp = LLVMBuildFMul(ctx->builder, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), "");
+		coords[2] = LLVMBuildFAdd(ctx->builder, tmp, coords[2], "");
+	}
+
+	memcpy(coords_arg, coords, sizeof(coords));
+}
+
+static void visit_tex(struct nir_to_llvm_context *ctx, nir_tex_instr *instr)
+{
+	LLVMValueRef result = NULL;
+	struct ac_tex_info tinfo = { 0 };
+	unsigned dmask = 0xf;
+	LLVMValueRef address[16];
+	LLVMValueRef coords[5];
+	LLVMValueRef coord = NULL, lod = NULL, comparitor = NULL, bias, offsets = NULL;
+	LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL, sample_index = NULL;
+	LLVMValueRef ddx = NULL, ddy = NULL;
+	LLVMValueRef derivs[6];
+	unsigned chan, count = 0;
+	unsigned const_src = 0, num_deriv_comp = 0;
+
+	tex_fetch_ptrs(ctx, instr, &res_ptr, &samp_ptr, &fmask_ptr);
+
+	for (unsigned i = 0; i < instr->num_srcs; i++) {
+		switch (instr->src[i].src_type) {
+		case nir_tex_src_coord:
+			coord = get_src(ctx, instr->src[i].src);
+			break;
+		case nir_tex_src_projector:
+			break;
+		case nir_tex_src_comparitor:
+			comparitor = get_src(ctx, instr->src[i].src);
+			break;
+		case nir_tex_src_offset:
+			offsets = get_src(ctx, instr->src[i].src);
+			const_src = i;
+			break;
+		case nir_tex_src_bias:
+			bias = get_src(ctx, instr->src[i].src);
+			break;
+		case nir_tex_src_lod:
+			lod = get_src(ctx, instr->src[i].src);
+			break;
+		case nir_tex_src_ms_index:
+			sample_index = get_src(ctx, instr->src[i].src);
+			break;
+		case nir_tex_src_ms_mcs:
+			break;
+		case nir_tex_src_ddx:
+			ddx = get_src(ctx, instr->src[i].src);
+			num_deriv_comp = instr->src[i].src.ssa->num_components;
+			break;
+		case nir_tex_src_ddy:
+			ddy = get_src(ctx, instr->src[i].src);
+			break;
+		case nir_tex_src_texture_offset:
+		case nir_tex_src_sampler_offset:
+		case nir_tex_src_plane:
+		default:
+			break;
+		}
+	}
+
+	if (instr->op == nir_texop_texture_samples) {
+		LLVMValueRef res, samples;
+		res = LLVMBuildBitCast(ctx->builder, res_ptr, ctx->v8i32, "");
+		samples = LLVMBuildExtractElement(ctx->builder, res,
+						  LLVMConstInt(ctx->i32, 3, false), "");
+		samples = LLVMBuildLShr(ctx->builder, samples,
+					LLVMConstInt(ctx->i32, 16, false), "");
+		samples = LLVMBuildAnd(ctx->builder, samples,
+				       LLVMConstInt(ctx->i32, 0xf, false), "");
+		samples = LLVMBuildShl(ctx->builder, ctx->i32one,
+				       samples, "");
+
+		result = samples;
+		goto write_result;
+	}
+
+	if (coord)
+		for (chan = 0; chan < instr->coord_components; chan++)
+			coords[chan] = llvm_extract_elem(ctx, coord, chan);
+
+	if (offsets && instr->op != nir_texop_txf) {
+		LLVMValueRef offset[3], pack;
+		for (chan = 0; chan < 3; ++chan)
+			offset[chan] = ctx->i32zero;
+
+		tinfo.has_offset = true;
+		for (chan = 0; chan < get_llvm_num_components(offsets); chan++) {
+			offset[chan] = llvm_extract_elem(ctx, offsets, chan);
+			offset[chan] = LLVMBuildAnd(ctx->builder, offset[chan],
+						    LLVMConstInt(ctx->i32, 0x3f, false), "");
+			if (chan)
+				offset[chan] = LLVMBuildShl(ctx->builder, offset[chan],
+							    LLVMConstInt(ctx->i32, chan * 8, false), "");
+		}
+		pack = LLVMBuildOr(ctx->builder, offset[0], offset[1], "");
+		pack = LLVMBuildOr(ctx->builder, pack, offset[2], "");
+		address[count++] = pack;
+
+	}
+	/* pack LOD bias value */
+	if (instr->op == nir_texop_txb && bias) {
+		address[count++] = bias;
+	}
+
+	/* Pack depth comparison value */
+	if (instr->is_shadow && comparitor) {
+		address[count++] = llvm_extract_elem(ctx, comparitor, 0);
+	}
+
+	/* pack derivatives */
+	if (ddx || ddy) {
+		switch (instr->sampler_dim) {
+		case GLSL_SAMPLER_DIM_3D:
+		case GLSL_SAMPLER_DIM_CUBE:
+			num_deriv_comp = 3;
+			break;
+		case GLSL_SAMPLER_DIM_2D:
+		default:
+			num_deriv_comp = 2;
+			break;
+		case GLSL_SAMPLER_DIM_1D:
+			num_deriv_comp = 1;
+			break;
+		}
+
+		for (unsigned i = 0; i < num_deriv_comp; i++) {
+			derivs[i * 2] = to_float(ctx, llvm_extract_elem(ctx, ddx, i));
+			derivs[i * 2 + 1] = to_float(ctx, llvm_extract_elem(ctx, ddy, i));
+		}
+	}
+
+	if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE && coord) {
+		for (chan = 0; chan < instr->coord_components; chan++)
+			coords[chan] = to_float(ctx, coords[chan]);
+		if (instr->coord_components == 3)
+			coords[3] = LLVMGetUndef(ctx->f32);
+		emit_prepare_cube_coords(ctx, coords, instr->coord_components, instr->op == nir_texop_txd, instr->is_array, derivs);
+		if (num_deriv_comp)
+			num_deriv_comp--;
+	}
+
+	if (ddx || ddy) {
+		for (unsigned i = 0; i < num_deriv_comp * 2; i++)
+			address[count++] = derivs[i];
+	}
+
+	/* Pack texture coordinates */
+	if (coord) {
+		address[count++] = coords[0];
+		if (instr->coord_components > 1)
+			address[count++] = coords[1];
+		if (instr->coord_components > 2) {
+			/* This seems like a bit of a hack - but it passes Vulkan CTS with it */
+			if (instr->sampler_dim != GLSL_SAMPLER_DIM_3D && instr->op != nir_texop_txf) {
+				coords[2] = to_float(ctx, coords[2]);
+				coords[2] = emit_llvm_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &coords[2],
+								1, 0);
+				coords[2] = to_integer(ctx, coords[2]);
+			}
+			address[count++] = coords[2];
+		}
+	}
+
+	/* Pack LOD */
+	if ((instr->op == nir_texop_txl || instr->op == nir_texop_txf) && lod) {
+		address[count++] = lod;
+	} else if (instr->op == nir_texop_txf_ms && sample_index) {
+		address[count++] = sample_index;
+	} else if(instr->op == nir_texop_txs) {
+		count = 0;
+		address[count++] = lod;
+	}
+
+	for (chan = 0; chan < count; chan++) {
+		address[chan] = LLVMBuildBitCast(ctx->builder,
+						 address[chan], ctx->i32, "");
+	}
+
+	if (instr->op == nir_texop_samples_identical) {
+		LLVMValueRef txf_address[4];
+		struct ac_tex_info txf_info = { 0 };
+		unsigned txf_count = count;
+		memcpy(txf_address, address, sizeof(txf_address));
+
+		if (!instr->is_array)
+			txf_address[2] = ctx->i32zero;
+		txf_address[3] = ctx->i32zero;
+
+		set_tex_fetch_args(ctx, &txf_info, instr, nir_texop_txf,
+				   res_ptr, samp_ptr,
+				   txf_address, txf_count, 0xf);
+
+		result = build_tex_intrinsic(ctx, instr, &txf_info);
+		goto write_result;
+	}
+
+	/* TODO sample FMASK magic */
+	if (instr->sampler_dim == GLSL_SAMPLER_DIM_MS) {
+		LLVMValueRef txf_address[4];
+		struct ac_tex_info txf_info = { 0 };
+		unsigned txf_count = count;
+		memcpy(txf_address, address, sizeof(txf_address));
+
+		if (!instr->is_array)
+			txf_address[2] = ctx->i32zero;
+		txf_address[3] = ctx->i32zero;
+
+		set_tex_fetch_args(ctx, &txf_info, instr, nir_texop_txf,
+				   res_ptr, samp_ptr,
+				   txf_address, txf_count, 0xf);
+
+		result = build_tex_intrinsic(ctx, instr, &txf_info);
+		LLVMValueRef four = LLVMConstInt(ctx->i32, 4, false);
+		LLVMValueRef F = LLVMConstInt(ctx->i32, 0xf, false);
+
+		LLVMValueRef fmask = LLVMBuildExtractElement(ctx->builder,
+							     result,
+							     ctx->i32zero, "");
+
+		unsigned sample_chan = instr->is_array ? 3 : 2;
+
+		LLVMValueRef sample_index4 =
+			LLVMBuildMul(ctx->builder, address[sample_chan], four, "");
+		LLVMValueRef shifted_fmask =
+			LLVMBuildLShr(ctx->builder, fmask, sample_index4, "");
+		LLVMValueRef final_sample =
+			LLVMBuildAnd(ctx->builder, shifted_fmask, F, "");
+
+		/* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
+		 * resource descriptor is 0 (invalid),
+		 */
+		LLVMValueRef fmask_desc =
+			LLVMBuildBitCast(ctx->builder, fmask_ptr,
+					 ctx->v8i32, "");
+
+		LLVMValueRef fmask_word1 =
+			LLVMBuildExtractElement(ctx->builder, fmask_desc,
+						ctx->i32one, "");
+
+		LLVMValueRef word1_is_nonzero =
+			LLVMBuildICmp(ctx->builder, LLVMIntNE,
+				      fmask_word1, ctx->i32zero, "");
+
+		/* Replace the MSAA sample index. */
+		address[sample_chan] =
+			LLVMBuildSelect(ctx->builder, word1_is_nonzero,
+					final_sample, address[sample_chan], "");
+	}
+
+	if (offsets && instr->op == nir_texop_txf) {
+		nir_const_value *const_offset =
+			nir_src_as_const_value(instr->src[const_src].src);
+
+		assert(const_offset);
+		if (instr->coord_components > 2)
+			address[2] = LLVMBuildAdd(ctx->builder,
+						  address[2], LLVMConstInt(ctx->i32, const_offset->i32[2], false), "");
+		if (instr->coord_components > 1)
+			address[1] = LLVMBuildAdd(ctx->builder,
+						  address[1], LLVMConstInt(ctx->i32, const_offset->i32[1], false), "");
+		address[0] = LLVMBuildAdd(ctx->builder,
+					  address[0], LLVMConstInt(ctx->i32, const_offset->i32[0], false), "");
+
+	}
+
+	/* TODO TG4 support */
+	if (instr->op == nir_texop_tg4) {
+		if (instr->is_shadow)
+			dmask = 1;
+		else
+			dmask = 1 << instr->component;
+	}
+	set_tex_fetch_args(ctx, &tinfo, instr, instr->op,
+			   res_ptr, samp_ptr, address, count, dmask);
+
+	result = build_tex_intrinsic(ctx, instr, &tinfo);
+
+	if (instr->op == nir_texop_query_levels)
+		result = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, 3, false), "");
+	else if (instr->op == nir_texop_txs &&
+		 instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
+		 instr->is_array) {
+		LLVMValueRef two = LLVMConstInt(ctx->i32, 2, false);
+		LLVMValueRef six = LLVMConstInt(ctx->i32, 6, false);
+		LLVMValueRef z = LLVMBuildExtractElement(ctx->builder, result, two, "");
+		z = LLVMBuildSDiv(ctx->builder, z, six, "");
+		result = LLVMBuildInsertElement(ctx->builder, result, z, two, "");
+	}
+
+write_result:
+	if (result) {
+		assert(instr->dest.is_ssa);
+		result = to_integer(ctx, result);
+		_mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
+	}
+}
+
+
+static void visit_phi(struct nir_to_llvm_context *ctx, nir_phi_instr *instr)
+{
+	LLVMTypeRef type = get_def_type(ctx, &instr->dest.ssa);
+	LLVMValueRef result = LLVMBuildPhi(ctx->builder, type, "");
+
+	_mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
+	_mesa_hash_table_insert(ctx->phis, instr, result);
+}
+
+static void visit_post_phi(struct nir_to_llvm_context *ctx,
+                           nir_phi_instr *instr,
+                           LLVMValueRef llvm_phi)
+{
+	nir_foreach_phi_src(src, instr) {
+		LLVMBasicBlockRef block = get_block(ctx, src->pred);
+		LLVMValueRef llvm_src = get_src(ctx, src->src);
+
+		LLVMAddIncoming(llvm_phi, &llvm_src, &block, 1);
+	}
+}
+
+static void phi_post_pass(struct nir_to_llvm_context *ctx)
+{
+	struct hash_entry *entry;
+	hash_table_foreach(ctx->phis, entry) {
+		visit_post_phi(ctx, (nir_phi_instr*)entry->key,
+		               (LLVMValueRef)entry->data);
+	}
+}
+
+
+static void visit_ssa_undef(struct nir_to_llvm_context *ctx,
+			    nir_ssa_undef_instr *instr)
+{
+	unsigned num_components = instr->def.num_components;
+	LLVMValueRef undef;
+
+	if (num_components == 1)
+		undef = LLVMGetUndef(ctx->i32);
+	else {
+		undef = LLVMGetUndef(LLVMVectorType(ctx->i32, num_components));
+	}
+	_mesa_hash_table_insert(ctx->defs, &instr->def, undef);
+}
+
+static void visit_jump(struct nir_to_llvm_context *ctx,
+		       nir_jump_instr *instr)
+{
+	switch (instr->type) {
+	case nir_jump_break:
+		LLVMBuildBr(ctx->builder, ctx->break_block);
+		LLVMClearInsertionPosition(ctx->builder);
+		break;
+	case nir_jump_continue:
+		LLVMBuildBr(ctx->builder, ctx->continue_block);
+		LLVMClearInsertionPosition(ctx->builder);
+		break;
+	default:
+		fprintf(stderr, "Unknown NIR jump instr: ");
+		nir_print_instr(&instr->instr, stderr);
+		fprintf(stderr, "\n");
+		abort();
+	}
+}
+
+static void visit_cf_list(struct nir_to_llvm_context *ctx,
+                          struct exec_list *list);
+
+static void visit_block(struct nir_to_llvm_context *ctx, nir_block *block)
+{
+	LLVMBasicBlockRef llvm_block = LLVMGetInsertBlock(ctx->builder);
+	nir_foreach_instr(instr, block)
+	{
+		switch (instr->type) {
+		case nir_instr_type_alu:
+			visit_alu(ctx, nir_instr_as_alu(instr));
+			break;
+		case nir_instr_type_load_const:
+			visit_load_const(ctx, nir_instr_as_load_const(instr));
+			break;
+		case nir_instr_type_intrinsic:
+			visit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
+			break;
+		case nir_instr_type_tex:
+			visit_tex(ctx, nir_instr_as_tex(instr));
+			break;
+		case nir_instr_type_phi:
+			visit_phi(ctx, nir_instr_as_phi(instr));
+			break;
+		case nir_instr_type_ssa_undef:
+			visit_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
+			break;
+		case nir_instr_type_jump:
+			visit_jump(ctx, nir_instr_as_jump(instr));
+			break;
+		default:
+			fprintf(stderr, "Unknown NIR instr type: ");
+			nir_print_instr(instr, stderr);
+			fprintf(stderr, "\n");
+			abort();
+		}
+	}
+
+	_mesa_hash_table_insert(ctx->defs, block, llvm_block);
+}
+
+static void visit_if(struct nir_to_llvm_context *ctx, nir_if *if_stmt)
+{
+	LLVMValueRef value = get_src(ctx, if_stmt->condition);
+
+	LLVMBasicBlockRef merge_block =
+	    LLVMAppendBasicBlockInContext(ctx->context, ctx->main_function, "");
+	LLVMBasicBlockRef if_block =
+	    LLVMAppendBasicBlockInContext(ctx->context, ctx->main_function, "");
+	LLVMBasicBlockRef else_block = merge_block;
+	if (!exec_list_is_empty(&if_stmt->else_list))
+		else_block = LLVMAppendBasicBlockInContext(
+		    ctx->context, ctx->main_function, "");
+
+	LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE, value,
+	                                  LLVMConstInt(ctx->i32, 0, false), "");
+	LLVMBuildCondBr(ctx->builder, cond, if_block, else_block);
+
+	LLVMPositionBuilderAtEnd(ctx->builder, if_block);
+	visit_cf_list(ctx, &if_stmt->then_list);
+	if (LLVMGetInsertBlock(ctx->builder))
+		LLVMBuildBr(ctx->builder, merge_block);
+
+	if (!exec_list_is_empty(&if_stmt->else_list)) {
+		LLVMPositionBuilderAtEnd(ctx->builder, else_block);
+		visit_cf_list(ctx, &if_stmt->else_list);
+		if (LLVMGetInsertBlock(ctx->builder))
+			LLVMBuildBr(ctx->builder, merge_block);
+	}
+
+	LLVMPositionBuilderAtEnd(ctx->builder, merge_block);
+}
+
+static void visit_loop(struct nir_to_llvm_context *ctx, nir_loop *loop)
+{
+	LLVMBasicBlockRef continue_parent = ctx->continue_block;
+	LLVMBasicBlockRef break_parent = ctx->break_block;
+
+	ctx->continue_block =
+	    LLVMAppendBasicBlockInContext(ctx->context, ctx->main_function, "");
+	ctx->break_block =
+	    LLVMAppendBasicBlockInContext(ctx->context, ctx->main_function, "");
+
+	LLVMBuildBr(ctx->builder, ctx->continue_block);
+	LLVMPositionBuilderAtEnd(ctx->builder, ctx->continue_block);
+	visit_cf_list(ctx, &loop->body);
+
+	if (LLVMGetInsertBlock(ctx->builder))
+		LLVMBuildBr(ctx->builder, ctx->continue_block);
+	LLVMPositionBuilderAtEnd(ctx->builder, ctx->break_block);
+
+	ctx->continue_block = continue_parent;
+	ctx->break_block = break_parent;
+}
+
+static void visit_cf_list(struct nir_to_llvm_context *ctx,
+                          struct exec_list *list)
+{
+	foreach_list_typed(nir_cf_node, node, node, list)
+	{
+		switch (node->type) {
+		case nir_cf_node_block:
+			visit_block(ctx, nir_cf_node_as_block(node));
+			break;
+
+		case nir_cf_node_if:
+			visit_if(ctx, nir_cf_node_as_if(node));
+			break;
+
+		case nir_cf_node_loop:
+			visit_loop(ctx, nir_cf_node_as_loop(node));
+			break;
+
+		default:
+			assert(0);
+		}
+	}
+}
+
+static void
+handle_vs_input_decl(struct nir_to_llvm_context *ctx,
+		     struct nir_variable *variable)
+{
+	LLVMValueRef t_list_ptr = ctx->vertex_buffers;
+	LLVMValueRef t_offset;
+	LLVMValueRef t_list;
+	LLVMValueRef args[3];
+	LLVMValueRef input;
+	LLVMValueRef buffer_index;
+	int index = variable->data.location - VERT_ATTRIB_GENERIC0;
+	int idx = variable->data.location;
+	unsigned attrib_count = glsl_count_attribute_slots(variable->type, true);
+
+	variable->data.driver_location = idx * 4;
+
+	if (ctx->options->key.vs.instance_rate_inputs & (1u << index)) {
+		buffer_index = LLVMBuildAdd(ctx->builder, ctx->instance_id,
+					    ctx->start_instance, "");
+		ctx->shader_info->vs.vgpr_comp_cnt = MAX2(3,
+		                            ctx->shader_info->vs.vgpr_comp_cnt);
+	} else
+		buffer_index = LLVMBuildAdd(ctx->builder, ctx->vertex_id,
+					    ctx->base_vertex, "");
+
+	for (unsigned i = 0; i < attrib_count; ++i, ++idx) {
+		t_offset = LLVMConstInt(ctx->i32, index + i, false);
+
+		t_list = build_indexed_load_const(ctx, t_list_ptr, t_offset);
+		args[0] = t_list;
+		args[1] = LLVMConstInt(ctx->i32, 0, false);
+		args[2] = buffer_index;
+		input = emit_llvm_intrinsic(ctx,
+			"llvm.SI.vs.load.input", ctx->v4f32, args, 3,
+			LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
+
+		for (unsigned chan = 0; chan < 4; chan++) {
+			LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, false);
+			ctx->inputs[radeon_llvm_reg_index_soa(idx, chan)] =
+				to_integer(ctx, LLVMBuildExtractElement(ctx->builder,
+							input, llvm_chan, ""));
+		}
+	}
+}
+
+
+static void interp_fs_input(struct nir_to_llvm_context *ctx,
+			    unsigned attr,
+			    LLVMValueRef interp_param,
+			    LLVMValueRef prim_mask,
+			    LLVMValueRef result[4])
+{
+	const char *intr_name;
+	LLVMValueRef attr_number;
+	unsigned chan;
+
+	attr_number = LLVMConstInt(ctx->i32, attr, false);
+
+	/* fs.constant returns the param from the middle vertex, so it's not
+	 * really useful for flat shading. It's meant to be used for custom
+	 * interpolation (but the intrinsic can't fetch from the other two
+	 * vertices).
+	 *
+	 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
+	 * to do the right thing. The only reason we use fs.constant is that
+	 * fs.interp cannot be used on integers, because they can be equal
+	 * to NaN.
+	 */
+	intr_name = interp_param ? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
+
+	for (chan = 0; chan < 4; chan++) {
+		LLVMValueRef args[4];
+		LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, false);
+
+		args[0] = llvm_chan;
+		args[1] = attr_number;
+		args[2] = prim_mask;
+		args[3] = interp_param;
+		result[chan] = emit_llvm_intrinsic(ctx, intr_name,
+						   ctx->f32, args, args[3] ? 4 : 3,
+						  LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
+	}
+}
+
+static void
+handle_fs_input_decl(struct nir_to_llvm_context *ctx,
+		     struct nir_variable *variable)
+{
+	int idx = variable->data.location;
+	unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
+	LLVMValueRef interp;
+
+	variable->data.driver_location = idx * 4;
+	ctx->input_mask |= ((1ull << attrib_count) - 1) << variable->data.location;
+
+	if (glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_FLOAT)
+		interp = lookup_interp_param(ctx, variable->data.interpolation, INTERP_CENTER);
+	else
+		interp = NULL;
+
+	for (unsigned i = 0; i < attrib_count; ++i)
+		ctx->inputs[radeon_llvm_reg_index_soa(idx + i, 0)] = interp;
+
+}
+
+static void
+handle_shader_input_decl(struct nir_to_llvm_context *ctx,
+			 struct nir_variable *variable)
+{
+	switch (ctx->stage) {
+	case MESA_SHADER_VERTEX:
+		handle_vs_input_decl(ctx, variable);
+		break;
+	case MESA_SHADER_FRAGMENT:
+		handle_fs_input_decl(ctx, variable);
+		break;
+	default:
+		break;
+	}
+
+}
+
+static void
+handle_fs_inputs_pre(struct nir_to_llvm_context *ctx,
+		     struct nir_shader *nir)
+{
+	unsigned index = 0;
+	for (unsigned i = 0; i < RADEON_LLVM_MAX_INPUTS; ++i) {
+		LLVMValueRef interp_param;
+		LLVMValueRef *inputs = ctx->inputs +radeon_llvm_reg_index_soa(i, 0);
+
+		if (!(ctx->input_mask & (1ull << i)))
+			continue;
+
+		if (i >= VARYING_SLOT_VAR0 || i == VARYING_SLOT_PNTC) {
+			interp_param = *inputs;
+			interp_fs_input(ctx, index, interp_param, ctx->prim_mask,
+					inputs);
+
+			if (!interp_param)
+				ctx->shader_info->fs.flat_shaded_mask |= 1u << index;
+			++index;
+		} else if (i == VARYING_SLOT_POS) {
+			for(int i = 0; i < 3; ++i)
+				inputs[i] = ctx->frag_pos[i];
+
+			inputs[3] = emit_fdiv(ctx, ctx->f32one, ctx->frag_pos[3]);
+		}
+	}
+	ctx->shader_info->fs.num_interp = index;
+	if (ctx->input_mask & (1 << VARYING_SLOT_PNTC))
+		ctx->shader_info->fs.has_pcoord = true;
+	ctx->shader_info->fs.input_mask = ctx->input_mask >> VARYING_SLOT_VAR0;
+}
+
+static LLVMValueRef
+ac_build_alloca(struct nir_to_llvm_context *ctx,
+                LLVMTypeRef type,
+                const char *name)
+{
+	LLVMBuilderRef builder = ctx->builder;
+	LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
+	LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
+	LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
+	LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
+	LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ctx->context);
+	LLVMValueRef res;
+
+	if (first_instr) {
+		LLVMPositionBuilderBefore(first_builder, first_instr);
+	} else {
+		LLVMPositionBuilderAtEnd(first_builder, first_block);
+	}
+
+	res = LLVMBuildAlloca(first_builder, type, name);
+	LLVMBuildStore(builder, LLVMConstNull(type), res);
+
+	LLVMDisposeBuilder(first_builder);
+
+	return res;
+}
+
+static LLVMValueRef si_build_alloca_undef(struct nir_to_llvm_context *ctx,
+					  LLVMTypeRef type,
+					  const char *name)
+{
+	LLVMValueRef ptr = ac_build_alloca(ctx, type, name);
+	LLVMBuildStore(ctx->builder, LLVMGetUndef(type), ptr);
+	return ptr;
+}
+
+static void
+handle_shader_output_decl(struct nir_to_llvm_context *ctx,
+			  struct nir_variable *variable)
+{
+	int idx = variable->data.location;
+	unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
+
+	variable->data.driver_location = idx * 4;
+
+	if (ctx->stage == MESA_SHADER_VERTEX) {
+
+		if (idx == VARYING_SLOT_CLIP_DIST0 ||
+		    idx == VARYING_SLOT_CULL_DIST0) {
+			int length = glsl_get_length(variable->type);
+			if (idx == VARYING_SLOT_CLIP_DIST0) {
+				ctx->shader_info->vs.clip_dist_mask = (1 << length) - 1;
+				ctx->num_clips = length;
+			} else if (idx == VARYING_SLOT_CULL_DIST0) {
+				ctx->shader_info->vs.cull_dist_mask = (1 << length) - 1;
+				ctx->num_culls = length;
+			}
+			if (length > 4)
+				attrib_count = 2;
+			else
+				attrib_count = 1;
+		}
+	}
+
+	for (unsigned i = 0; i < attrib_count; ++i) {
+		for (unsigned chan = 0; chan < 4; chan++) {
+			ctx->outputs[radeon_llvm_reg_index_soa(idx + i, chan)] =
+		                       si_build_alloca_undef(ctx, ctx->f32, "");
+		}
+	}
+	ctx->output_mask |= ((1ull << attrib_count) - 1) << variable->data.location;
+}
+
+static void
+setup_locals(struct nir_to_llvm_context *ctx,
+	     struct nir_function *func)
+{
+	int i, j;
+	ctx->num_locals = 0;
+	nir_foreach_variable(variable, &func->impl->locals) {
+		unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
+		variable->data.driver_location = ctx->num_locals * 4;
+		ctx->num_locals += attrib_count;
+	}
+	ctx->locals = malloc(4 * ctx->num_locals * sizeof(LLVMValueRef));
+	if (!ctx->locals)
+	    return;
+
+	for (i = 0; i < ctx->num_locals; i++) {
+		for (j = 0; j < 4; j++) {
+			ctx->locals[i * 4 + j] =
+				si_build_alloca_undef(ctx, ctx->f32, "temp");
+		}
+	}
+}
+
+static LLVMValueRef
+emit_float_saturate(struct nir_to_llvm_context *ctx, LLVMValueRef v, float lo, float hi)
+{
+	v = to_float(ctx, v);
+	v = emit_intrin_2f_param(ctx, "llvm.maxnum.f32", v, LLVMConstReal(ctx->f32, lo));
+	return emit_intrin_2f_param(ctx, "llvm.minnum.f32", v, LLVMConstReal(ctx->f32, hi));
+}
+
+
+static LLVMValueRef emit_pack_int16(struct nir_to_llvm_context *ctx,
+					LLVMValueRef src0, LLVMValueRef src1)
+{
+	LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
+	LLVMValueRef comp[2];
+
+	comp[0] = LLVMBuildAnd(ctx->builder, src0, LLVMConstInt(ctx-> i32, 65535, 0), "");
+	comp[1] = LLVMBuildAnd(ctx->builder, src1, LLVMConstInt(ctx-> i32, 65535, 0), "");
+	comp[1] = LLVMBuildShl(ctx->builder, comp[1], const16, "");
+	return LLVMBuildOr(ctx->builder, comp[0], comp[1], "");
+}
+
+/* Initialize arguments for the shader export intrinsic */
+static void
+si_llvm_init_export_args(struct nir_to_llvm_context *ctx,
+			 LLVMValueRef *values,
+			 unsigned target,
+			 LLVMValueRef *args)
+{
+	/* Default is 0xf. Adjusted below depending on the format. */
+	args[0] = LLVMConstInt(ctx->i32, target != V_008DFC_SQ_EXP_NULL ? 0xf : 0, false);
+	/* Specify whether the EXEC mask represents the valid mask */
+	args[1] = LLVMConstInt(ctx->i32, 0, false);
+
+	/* Specify whether this is the last export */
+	args[2] = LLVMConstInt(ctx->i32, 0, false);
+	/* Specify the target we are exporting */
+	args[3] = LLVMConstInt(ctx->i32, target, false);
+
+	args[4] = LLVMConstInt(ctx->i32, 0, false); /* COMPR flag */
+	args[5] = LLVMGetUndef(ctx->f32);
+	args[6] = LLVMGetUndef(ctx->f32);
+	args[7] = LLVMGetUndef(ctx->f32);
+	args[8] = LLVMGetUndef(ctx->f32);
+
+	if (!values)
+		return;
+
+	if (ctx->stage == MESA_SHADER_FRAGMENT && target >= V_008DFC_SQ_EXP_MRT) {
+		LLVMValueRef val[4];
+		unsigned index = target - V_008DFC_SQ_EXP_MRT;
+		unsigned col_format = (ctx->options->key.fs.col_format >> (4 * index)) & 0xf;
+		bool is_int8 = (ctx->options->key.fs.is_int8 >> index) & 1;
+
+		switch(col_format) {
+		case V_028714_SPI_SHADER_ZERO:
+			args[0] = LLVMConstInt(ctx->i32, 0x0, 0);
+			args[3] = LLVMConstInt(ctx->i32, V_008DFC_SQ_EXP_NULL, 0);
+			break;
+
+		case V_028714_SPI_SHADER_32_R:
+			args[0] = LLVMConstInt(ctx->i32, 0x1, 0);
+			args[5] = values[0];
+			break;
+
+		case V_028714_SPI_SHADER_32_GR:
+			args[0] = LLVMConstInt(ctx->i32, 0x3, 0);
+			args[5] = values[0];
+			args[6] = values[1];
+			break;
+
+		case V_028714_SPI_SHADER_32_AR:
+			args[0] = LLVMConstInt(ctx->i32, 0x9, 0);
+			args[5] = values[0];
+			args[8] = values[3];
+			break;
+
+		case V_028714_SPI_SHADER_FP16_ABGR:
+			args[4] = ctx->i32one;
+
+			for (unsigned chan = 0; chan < 2; chan++) {
+				LLVMValueRef pack_args[2] = {
+					values[2 * chan],
+					values[2 * chan + 1]
+				};
+				LLVMValueRef packed;
+
+				packed = emit_llvm_intrinsic(ctx, "llvm.SI.packf16",
+							     ctx->i32, pack_args, 2,
+							     LLVMReadNoneAttribute);
+				args[chan + 5] = packed;
+			}
+			break;
+
+		case V_028714_SPI_SHADER_UNORM16_ABGR:
+			for (unsigned chan = 0; chan < 4; chan++) {
+				val[chan] = emit_float_saturate(ctx, values[chan], 0, 1);
+				val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
+							LLVMConstReal(ctx->f32, 65535), "");
+				val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
+							LLVMConstReal(ctx->f32, 0.5), "");
+				val[chan] = LLVMBuildFPToUI(ctx->builder, val[chan],
+							ctx->i32, "");
+			}
+
+			args[4] = ctx->i32one;
+			args[5] = emit_pack_int16(ctx, val[0], val[1]);
+			args[6] = emit_pack_int16(ctx, val[2], val[3]);
+			break;
+
+		case V_028714_SPI_SHADER_SNORM16_ABGR:
+			for (unsigned chan = 0; chan < 4; chan++) {
+				val[chan] = emit_float_saturate(ctx, values[chan], -1, 1);
+				val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
+							LLVMConstReal(ctx->f32, 32767), "");
+
+				/* If positive, add 0.5, else add -0.5. */
+				val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
+						LLVMBuildSelect(ctx->builder,
+							LLVMBuildFCmp(ctx->builder, LLVMRealOGE,
+								val[chan], ctx->f32zero, ""),
+							LLVMConstReal(ctx->f32, 0.5),
+							LLVMConstReal(ctx->f32, -0.5), ""), "");
+				val[chan] = LLVMBuildFPToSI(ctx->builder, val[chan], ctx->i32, "");
+			}
+
+			args[4] = ctx->i32one;
+			args[5] = emit_pack_int16(ctx, val[0], val[1]);
+			args[6] = emit_pack_int16(ctx, val[2], val[3]);
+			break;
+
+		case V_028714_SPI_SHADER_UINT16_ABGR: {
+			LLVMValueRef max = LLVMConstInt(ctx->i32, is_int8 ? 255 : 65535, 0);
+
+			for (unsigned chan = 0; chan < 4; chan++) {
+				val[chan] = to_integer(ctx, values[chan]);
+				val[chan] = emit_minmax_int(ctx, LLVMIntULT, val[chan], max);
+			}
+
+			args[4] = ctx->i32one;
+			args[5] = emit_pack_int16(ctx, val[0], val[1]);
+			args[6] = emit_pack_int16(ctx, val[2], val[3]);
+			break;
+		}
+
+		case V_028714_SPI_SHADER_SINT16_ABGR: {
+			LLVMValueRef max = LLVMConstInt(ctx->i32, is_int8 ? 127 : 32767, 0);
+			LLVMValueRef min = LLVMConstInt(ctx->i32, is_int8 ? -128 : -32768, 0);
+
+			/* Clamp. */
+			for (unsigned chan = 0; chan < 4; chan++) {
+				val[chan] = to_integer(ctx, values[chan]);
+				val[chan] = emit_minmax_int(ctx, LLVMIntSLT, val[chan], max);
+				val[chan] = emit_minmax_int(ctx, LLVMIntSGT, val[chan], min);
+			}
+
+			args[4] = ctx->i32one;
+			args[5] = emit_pack_int16(ctx, val[0], val[1]);
+			args[6] = emit_pack_int16(ctx, val[2], val[3]);
+			break;
+		}
+
+		default:
+		case V_028714_SPI_SHADER_32_ABGR:
+			memcpy(&args[5], values, sizeof(values[0]) * 4);
+			break;
+		}
+	} else
+		memcpy(&args[5], values, sizeof(values[0]) * 4);
+
+	for (unsigned i = 5; i < 9; ++i)
+		args[i] = to_float(ctx, args[i]);
+}
+
+static void
+handle_vs_outputs_post(struct nir_to_llvm_context *ctx,
+		      struct nir_shader *nir)
+{
+	uint32_t param_count = 0;
+	unsigned target;
+	unsigned pos_idx, num_pos_exports = 0;
+	LLVMValueRef args[9];
+	LLVMValueRef pos_args[4][9] = { { 0 } };
+	LLVMValueRef psize_value = 0;
+	int i;
+	const uint64_t clip_mask = ctx->output_mask & ((1ull << VARYING_SLOT_CLIP_DIST0) |
+						       (1ull << VARYING_SLOT_CLIP_DIST1) |
+						       (1ull << VARYING_SLOT_CULL_DIST0) |
+						       (1ull << VARYING_SLOT_CULL_DIST1));
+
+	if (clip_mask) {
+		LLVMValueRef slots[8];
+		unsigned j;
+
+		if (ctx->shader_info->vs.cull_dist_mask)
+			ctx->shader_info->vs.cull_dist_mask <<= ctx->num_clips;
+
+		i = VARYING_SLOT_CLIP_DIST0;
+		for (j = 0; j < ctx->num_clips; j++)
+			slots[j] = to_float(ctx, LLVMBuildLoad(ctx->builder,
+							       ctx->outputs[radeon_llvm_reg_index_soa(i, j)], ""));
+		i = VARYING_SLOT_CULL_DIST0;
+		for (j = 0; j < ctx->num_culls; j++)
+			slots[ctx->num_clips + j] = to_float(ctx, LLVMBuildLoad(ctx->builder,
+									   ctx->outputs[radeon_llvm_reg_index_soa(i, j)], ""));
+
+		for (i = ctx->num_clips + ctx->num_culls; i < 8; i++)
+			slots[i] = LLVMGetUndef(ctx->f32);
+
+		if (ctx->num_clips + ctx->num_culls > 4) {
+			target = V_008DFC_SQ_EXP_POS + 3;
+			si_llvm_init_export_args(ctx, &slots[4], target, args);
+			memcpy(pos_args[target - V_008DFC_SQ_EXP_POS],
+			       args, sizeof(args));
+		}
+
+		target = V_008DFC_SQ_EXP_POS + 2;
+		si_llvm_init_export_args(ctx, &slots[0], target, args);
+		memcpy(pos_args[target - V_008DFC_SQ_EXP_POS],
+		       args, sizeof(args));
+
+	}
+
+	for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
+		LLVMValueRef values[4];
+		if (!(ctx->output_mask & (1ull << i)))
+			continue;
+
+		for (unsigned j = 0; j < 4; j++)
+			values[j] = to_float(ctx, LLVMBuildLoad(ctx->builder,
+					      ctx->outputs[radeon_llvm_reg_index_soa(i, j)], ""));
+
+		if (i == VARYING_SLOT_POS) {
+			target = V_008DFC_SQ_EXP_POS;
+		} else if (i == VARYING_SLOT_CLIP_DIST0 ||
+			   i == VARYING_SLOT_CLIP_DIST1 ||
+			   i == VARYING_SLOT_CULL_DIST0 ||
+			   i == VARYING_SLOT_CULL_DIST1) {
+			continue;
+		} else if (i == VARYING_SLOT_PSIZ) {
+			ctx->shader_info->vs.writes_pointsize = true;
+			psize_value = values[0];
+			continue;
+		} else if (i >= VARYING_SLOT_VAR0) {
+			ctx->shader_info->vs.export_mask |= 1u << (i - VARYING_SLOT_VAR0);
+			target = V_008DFC_SQ_EXP_PARAM + param_count;
+			param_count++;
+		}
+
+		si_llvm_init_export_args(ctx, values, target, args);
+
+		if (target >= V_008DFC_SQ_EXP_POS &&
+		    target <= (V_008DFC_SQ_EXP_POS + 3)) {
+			memcpy(pos_args[target - V_008DFC_SQ_EXP_POS],
+			       args, sizeof(args));
+		} else {
+			emit_llvm_intrinsic(ctx,
+					    "llvm.SI.export",
+					    LLVMVoidTypeInContext(ctx->context),
+					    args, 9, 0);
+		}
+	}
+
+	/* We need to add the position output manually if it's missing. */
+	if (!pos_args[0][0]) {
+		pos_args[0][0] = LLVMConstInt(ctx->i32, 0xf, false);
+		pos_args[0][1] = ctx->i32zero; /* EXEC mask */
+		pos_args[0][2] = ctx->i32zero; /* last export? */
+		pos_args[0][3] = LLVMConstInt(ctx->i32, V_008DFC_SQ_EXP_POS, false);
+		pos_args[0][4] = ctx->i32zero; /* COMPR flag */
+		pos_args[0][5] = ctx->f32zero; /* X */
+		pos_args[0][6] = ctx->f32zero; /* Y */
+		pos_args[0][7] = ctx->f32zero; /* Z */
+		pos_args[0][8] = ctx->f32one;  /* W */
+	}
+
+	if (ctx->shader_info->vs.writes_pointsize == true) {
+		pos_args[1][0] = LLVMConstInt(ctx->i32, (ctx->shader_info->vs.writes_pointsize == true), false); /* writemask */
+		pos_args[1][1] = ctx->i32zero;  /* EXEC mask */
+		pos_args[1][2] = ctx->i32zero;  /* last export? */
+		pos_args[1][3] = LLVMConstInt(ctx->i32, V_008DFC_SQ_EXP_POS + 1, false);
+		pos_args[1][4] = ctx->i32zero;  /* COMPR flag */
+		pos_args[1][5] = ctx->f32zero; /* X */
+		pos_args[1][6] = ctx->f32zero; /* Y */
+		pos_args[1][7] = ctx->f32zero; /* Z */
+		pos_args[1][8] = ctx->f32zero;  /* W */
+
+		if (ctx->shader_info->vs.writes_pointsize == true)
+			pos_args[1][5] = psize_value;
+	}
+	for (i = 0; i < 4; i++) {
+		if (pos_args[i][0])
+			num_pos_exports++;
+	}
+
+	pos_idx = 0;
+	for (i = 0; i < 4; i++) {
+		if (!pos_args[i][0])
+			continue;
+
+		/* Specify the target we are exporting */
+		pos_args[i][3] = LLVMConstInt(ctx->i32, V_008DFC_SQ_EXP_POS + pos_idx++, false);
+		if (pos_idx == num_pos_exports)
+			pos_args[i][2] = ctx->i32one;
+		emit_llvm_intrinsic(ctx,
+				    "llvm.SI.export",
+				    LLVMVoidTypeInContext(ctx->context),
+				    pos_args[i], 9, 0);
+	}
+
+	ctx->shader_info->vs.pos_exports = num_pos_exports;
+	ctx->shader_info->vs.param_exports = param_count;
+}
+
+static void
+si_export_mrt_color(struct nir_to_llvm_context *ctx,
+		    LLVMValueRef *color, unsigned param, bool is_last)
+{
+	LLVMValueRef args[9];
+	/* Export */
+	si_llvm_init_export_args(ctx, color, param,
+				 args);
+
+	if (is_last) {
+		args[1] = ctx->i32one; /* whether the EXEC mask is valid */
+		args[2] = ctx->i32one; /* DONE bit */
+	} else if (args[0] == ctx->i32zero)
+		return; /* unnecessary NULL export */
+
+	emit_llvm_intrinsic(ctx, "llvm.SI.export",
+			    ctx->voidt, args, 9, 0);
+}
+
+static void
+si_export_mrt_z(struct nir_to_llvm_context *ctx,
+		LLVMValueRef depth, LLVMValueRef stencil,
+		LLVMValueRef samplemask)
+{
+	LLVMValueRef args[9];
+	unsigned mask = 0;
+	args[1] = ctx->i32one; /* whether the EXEC mask is valid */
+	args[2] = ctx->i32one; /* DONE bit */
+	/* Specify the target we are exporting */
+	args[3] = LLVMConstInt(ctx->i32, V_008DFC_SQ_EXP_MRTZ, false);
+
+	args[4] = ctx->i32zero; /* COMP flag */
+	args[5] = LLVMGetUndef(ctx->f32); /* R, depth */
+	args[6] = LLVMGetUndef(ctx->f32); /* G, stencil test val[0:7], stencil op val[8:15] */
+	args[7] = LLVMGetUndef(ctx->f32); /* B, sample mask */
+	args[8] = LLVMGetUndef(ctx->f32); /* A, alpha to mask */
+
+	if (depth) {
+		args[5] = depth;
+		mask |= 0x1;
+	}
+
+	if (stencil) {
+		args[6] = stencil;
+		mask |= 0x2;
+	}
+
+	if (samplemask) {
+		args[7] = samplemask;
+		mask |= 0x04;
+	}
+
+	/* SI (except OLAND) has a bug that it only looks
+	 * at the X writemask component. */
+	if (ctx->options->chip_class == SI &&
+	    ctx->options->family != CHIP_OLAND)
+		mask |= 0x01;
+
+	args[0] = LLVMConstInt(ctx->i32, mask, false);
+	emit_llvm_intrinsic(ctx, "llvm.SI.export",
+			    ctx->voidt, args, 9, 0);
+}
+
+static void
+handle_fs_outputs_post(struct nir_to_llvm_context *ctx,
+		       struct nir_shader *nir)
+{
+	unsigned index = 0;
+	LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
+
+	for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
+		LLVMValueRef values[4];
+		bool last;
+		if (!(ctx->output_mask & (1ull << i)))
+			continue;
+
+		last = ctx->output_mask <= ((1ull << (i + 1)) - 1);
+
+		if (i == FRAG_RESULT_DEPTH) {
+			ctx->shader_info->fs.writes_z = true;
+			depth = to_float(ctx, LLVMBuildLoad(ctx->builder,
+							    ctx->outputs[radeon_llvm_reg_index_soa(i, 0)], ""));
+		} else if (i == FRAG_RESULT_STENCIL) {
+			ctx->shader_info->fs.writes_stencil = true;
+			stencil = to_float(ctx, LLVMBuildLoad(ctx->builder,
+							      ctx->outputs[radeon_llvm_reg_index_soa(i, 0)], ""));
+		} else {
+			for (unsigned j = 0; j < 4; j++)
+				values[j] = to_float(ctx, LLVMBuildLoad(ctx->builder,
+									ctx->outputs[radeon_llvm_reg_index_soa(i, j)], ""));
+
+			si_export_mrt_color(ctx, values, V_008DFC_SQ_EXP_MRT + index, last);
+			index++;
+		}
+	}
+
+	if (depth || stencil)
+		si_export_mrt_z(ctx, depth, stencil, samplemask);
+	else if (!index)
+		si_export_mrt_color(ctx, NULL, V_008DFC_SQ_EXP_NULL, true);
+
+	ctx->shader_info->fs.output_mask = index ? ((1ull << index) - 1) : 0;
+}
+
+static void
+handle_shader_outputs_post(struct nir_to_llvm_context *ctx,
+			   struct nir_shader *nir)
+{
+	switch (ctx->stage) {
+	case MESA_SHADER_VERTEX:
+		handle_vs_outputs_post(ctx, nir);
+		break;
+	case MESA_SHADER_FRAGMENT:
+		handle_fs_outputs_post(ctx, nir);
+		break;
+	default:
+		break;
+	}
+}
+
+static void
+handle_shared_compute_var(struct nir_to_llvm_context *ctx,
+			  struct nir_variable *variable, uint32_t *offset, int idx)
+{
+	unsigned size = glsl_count_attribute_slots(variable->type, false);
+	variable->data.driver_location = *offset;
+	*offset += size;
+}
+
+static void ac_llvm_finalize_module(struct nir_to_llvm_context * ctx)
+{
+	LLVMPassManagerRef passmgr;
+	/* Create the pass manager */
+	passmgr = LLVMCreateFunctionPassManagerForModule(
+							ctx->module);
+
+	/* This pass should eliminate all the load and store instructions */
+	LLVMAddPromoteMemoryToRegisterPass(passmgr);
+
+	/* Add some optimization passes */
+	LLVMAddScalarReplAggregatesPass(passmgr);
+	LLVMAddLICMPass(passmgr);
+	LLVMAddAggressiveDCEPass(passmgr);
+	LLVMAddCFGSimplificationPass(passmgr);
+	LLVMAddInstructionCombiningPass(passmgr);
+
+	/* Run the pass */
+	LLVMInitializeFunctionPassManager(passmgr);
+	LLVMRunFunctionPassManager(passmgr, ctx->main_function);
+	LLVMFinalizeFunctionPassManager(passmgr);
+
+	LLVMDisposeBuilder(ctx->builder);
+	LLVMDisposePassManager(passmgr);
+}
+
+static
+LLVMModuleRef ac_translate_nir_to_llvm(LLVMTargetMachineRef tm,
+                                       struct nir_shader *nir,
+                                       struct ac_shader_variant_info *shader_info,
+                                       const struct ac_nir_compiler_options *options)
+{
+	struct nir_to_llvm_context ctx = {0};
+	struct nir_function *func;
+	ctx.options = options;
+	ctx.shader_info = shader_info;
+	ctx.context = LLVMContextCreate();
+	ctx.module = LLVMModuleCreateWithNameInContext("shader", ctx.context);
+
+	memset(shader_info, 0, sizeof(*shader_info));
+
+	LLVMSetTarget(ctx.module, "amdgcn--");
+	setup_types(&ctx);
+
+	ctx.builder = LLVMCreateBuilderInContext(ctx.context);
+	ctx.stage = nir->stage;
+
+	create_function(&ctx, nir);
+
+	if (nir->stage == MESA_SHADER_COMPUTE) {
+		int num_shared = 0;
+		nir_foreach_variable(variable, &nir->shared)
+			num_shared++;
+		if (num_shared) {
+			int idx = 0;
+			uint32_t shared_size = 0;
+			LLVMValueRef var;
+			LLVMTypeRef i8p = LLVMPointerType(ctx.i8, LOCAL_ADDR_SPACE);
+			nir_foreach_variable(variable, &nir->shared) {
+				handle_shared_compute_var(&ctx, variable, &shared_size, idx);
+				idx++;
+			}
+
+			shared_size *= 4;
+			var = LLVMAddGlobalInAddressSpace(ctx.module,
+							  LLVMArrayType(ctx.i8, shared_size),
+							  "compute_lds",
+							  LOCAL_ADDR_SPACE);
+			LLVMSetAlignment(var, 4);
+			ctx.shared_memory = LLVMBuildBitCast(ctx.builder, var, i8p, "");
+		}
+	}
+
+	nir_foreach_variable(variable, &nir->inputs)
+		handle_shader_input_decl(&ctx, variable);
+
+	if (nir->stage == MESA_SHADER_FRAGMENT)
+		handle_fs_inputs_pre(&ctx, nir);
+
+	nir_foreach_variable(variable, &nir->outputs)
+		handle_shader_output_decl(&ctx, variable);
+
+	ctx.defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+	                                   _mesa_key_pointer_equal);
+	ctx.phis = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+	                                   _mesa_key_pointer_equal);
+
+	func = (struct nir_function *)exec_list_get_head(&nir->functions);
+
+	setup_locals(&ctx, func);
+
+	visit_cf_list(&ctx, &func->impl->body);
+	phi_post_pass(&ctx);
+
+	handle_shader_outputs_post(&ctx, nir);
+	LLVMBuildRetVoid(ctx.builder);
+
+	ac_llvm_finalize_module(&ctx);
+	free(ctx.locals);
+	ralloc_free(ctx.defs);
+	ralloc_free(ctx.phis);
+
+	return ctx.module;
+}
+
+static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di, void *context)
+{
+	unsigned *retval = (unsigned *)context;
+	LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
+	char *description = LLVMGetDiagInfoDescription(di);
+
+	if (severity == LLVMDSError) {
+		*retval = 1;
+		fprintf(stderr, "LLVM triggered Diagnostic Handler: %s\n",
+		        description);
+	}
+
+	LLVMDisposeMessage(description);
+}
+
+static unsigned ac_llvm_compile(LLVMModuleRef M,
+                                struct ac_shader_binary *binary,
+                                LLVMTargetMachineRef tm)
+{
+	unsigned retval = 0;
+	char *err;
+	LLVMContextRef llvm_ctx;
+	LLVMMemoryBufferRef out_buffer;
+	unsigned buffer_size;
+	const char *buffer_data;
+	LLVMBool mem_err;
+
+	/* Setup Diagnostic Handler*/
+	llvm_ctx = LLVMGetModuleContext(M);
+
+	LLVMContextSetDiagnosticHandler(llvm_ctx, ac_diagnostic_handler,
+	                                &retval);
+
+	/* Compile IR*/
+	mem_err = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile,
+	                                              &err, &out_buffer);
+
+	/* Process Errors/Warnings */
+	if (mem_err) {
+		fprintf(stderr, "%s: %s", __FUNCTION__, err);
+		free(err);
+		retval = 1;
+		goto out;
+	}
+
+	/* Extract Shader Code*/
+	buffer_size = LLVMGetBufferSize(out_buffer);
+	buffer_data = LLVMGetBufferStart(out_buffer);
+
+	ac_elf_read(buffer_data, buffer_size, binary);
+
+	/* Clean up */
+	LLVMDisposeMemoryBuffer(out_buffer);
+
+out:
+	return retval;
+}
+
+void ac_compile_nir_shader(LLVMTargetMachineRef tm,
+                           struct ac_shader_binary *binary,
+                           struct ac_shader_config *config,
+                           struct ac_shader_variant_info *shader_info,
+                           struct nir_shader *nir,
+                           const struct ac_nir_compiler_options *options,
+			   bool dump_shader)
+{
+
+	LLVMModuleRef llvm_module = ac_translate_nir_to_llvm(tm, nir, shader_info,
+	                                                     options);
+	if (dump_shader)
+		LLVMDumpModule(llvm_module);
+
+	memset(binary, 0, sizeof(*binary));
+	int v = ac_llvm_compile(llvm_module, binary, tm);
+	if (v) {
+		fprintf(stderr, "compile failed\n");
+	}
+
+	if (dump_shader)
+		fprintf(stderr, "disasm:\n%s\n", binary->disasm_string);
+
+	ac_shader_binary_read_config(binary, config, 0);
+
+	LLVMContextRef ctx = LLVMGetModuleContext(llvm_module);
+	LLVMDisposeModule(llvm_module);
+	LLVMContextDispose(ctx);
+
+	if (nir->stage == MESA_SHADER_FRAGMENT) {
+		shader_info->num_input_vgprs = 0;
+		if (G_0286CC_PERSP_SAMPLE_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 2;
+		if (G_0286CC_PERSP_CENTER_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 2;
+		if (G_0286CC_PERSP_CENTROID_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 2;
+		if (G_0286CC_PERSP_PULL_MODEL_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 3;
+		if (G_0286CC_LINEAR_SAMPLE_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 2;
+		if (G_0286CC_LINEAR_CENTER_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 2;
+		if (G_0286CC_LINEAR_CENTROID_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 2;
+		if (G_0286CC_LINE_STIPPLE_TEX_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 1;
+		if (G_0286CC_POS_X_FLOAT_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 1;
+		if (G_0286CC_POS_Y_FLOAT_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 1;
+		if (G_0286CC_POS_Z_FLOAT_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 1;
+		if (G_0286CC_POS_W_FLOAT_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 1;
+		if (G_0286CC_FRONT_FACE_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 1;
+		if (G_0286CC_ANCILLARY_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 1;
+		if (G_0286CC_SAMPLE_COVERAGE_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 1;
+		if (G_0286CC_POS_FIXED_PT_ENA(config->spi_ps_input_addr))
+			shader_info->num_input_vgprs += 1;
+	}
+	config->num_vgprs = MAX2(config->num_vgprs, shader_info->num_input_vgprs);
+
+	/* +3 for scratch wave offset and VCC */
+	config->num_sgprs = MAX2(config->num_sgprs,
+	                         shader_info->num_input_sgprs + 3);
+	if (nir->stage == MESA_SHADER_COMPUTE) {
+		for (int i = 0; i < 3; ++i)
+			shader_info->cs.block_size[i] = nir->info.cs.local_size[i];
+	}
+
+	if (nir->stage == MESA_SHADER_FRAGMENT)
+		shader_info->fs.early_fragment_test = nir->info.fs.early_fragment_tests;
+}
diff --git a/src/amd/common/ac_nir_to_llvm.h b/src/amd/common/ac_nir_to_llvm.h
new file mode 100644
index 0000000000..550fe84507
--- /dev/null
+++ b/src/amd/common/ac_nir_to_llvm.h
@@ -0,0 +1,102 @@
+/*
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#pragma once
+
+#include <stdbool.h>
+#include "llvm-c/Core.h"
+#include "llvm-c/TargetMachine.h"
+#include "amd_family.h"
+
+struct ac_shader_binary;
+struct ac_shader_config;
+struct nir_shader;
+struct radv_pipeline_layout;
+
+
+struct ac_vs_variant_key {
+	uint32_t instance_rate_inputs;
+};
+
+struct ac_fs_variant_key {
+	uint32_t col_format;
+	uint32_t is_int8;
+};
+
+union ac_shader_variant_key {
+	struct ac_vs_variant_key vs;
+	struct ac_fs_variant_key fs;
+};
+
+struct ac_nir_compiler_options {
+	struct radv_pipeline_layout *layout;
+	union ac_shader_variant_key key;
+	bool unsafe_math;
+	enum radeon_family family;
+	enum chip_class chip_class;
+};
+
+struct ac_shader_variant_info {
+	unsigned num_user_sgprs;
+	unsigned num_input_sgprs;
+	unsigned num_input_vgprs;
+	union {
+		struct {
+			unsigned param_exports;
+			unsigned pos_exports;
+			unsigned vgpr_comp_cnt;
+			uint32_t export_mask;
+			bool writes_pointsize;
+			uint8_t clip_dist_mask;
+			uint8_t cull_dist_mask;
+		} vs;
+		struct {
+			unsigned num_interp;
+			uint32_t input_mask;
+			unsigned output_mask;
+			uint32_t flat_shaded_mask;
+			bool has_pcoord;
+			bool can_discard;
+			bool writes_z;
+			bool writes_stencil;
+			bool early_fragment_test;
+			bool writes_memory;
+		} fs;
+		struct {
+			unsigned block_size[3];
+		} cs;
+	};
+};
+
+void ac_compile_nir_shader(LLVMTargetMachineRef tm,
+                           struct ac_shader_binary *binary,
+                           struct ac_shader_config *config,
+                           struct ac_shader_variant_info *shader_info,
+                           struct nir_shader *nir,
+                           const struct ac_nir_compiler_options *options,
+			   bool dump_shader);
+
+#ifdef __cplusplus
+extern "C"
+#endif
+void ac_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes);
diff --git a/src/amd/vulkan/.gitignore b/src/amd/vulkan/.gitignore
new file mode 100644
index 0000000000..e55e353d8f
--- /dev/null
+++ b/src/amd/vulkan/.gitignore
@@ -0,0 +1,6 @@
+# Generated source files
+/radv_entrypoints.c
+/radv_entrypoints.h
+/radv_timestamp.h
+/dev_icd.json
+/vk_format_table.c
diff --git a/src/amd/vulkan/Makefile.am b/src/amd/vulkan/Makefile.am
new file mode 100644
index 0000000000..387ba4dda8
--- /dev/null
+++ b/src/amd/vulkan/Makefile.am
@@ -0,0 +1,165 @@
+# Copyright © 2016 Red Hat
+#
+# Permission is hereby granted, free of charge, to any person obtaining a
+# copy of this software and associated documentation files (the "Software"),
+# to deal in the Software without restriction, including without limitation
+# the rights to use, copy, modify, merge, publish, distribute, sublicense,
+# and/or sell copies of the Software, and to permit persons to whom the
+# Software is furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice (including the next
+# paragraph) shall be included in all copies or substantial portions of the
+# Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+# IN THE SOFTWARE.
+
+include Makefile.sources
+
+vulkan_includedir = $(includedir)/vulkan
+
+vulkan_include_HEADERS = \
+	$(top_srcdir)/include/vulkan/vk_platform.h \
+	$(top_srcdir)/include/vulkan/vulkan.h
+
+lib_LTLIBRARIES = libvulkan_radeon.la
+
+# The gallium includes are for the util/u_math.h include from main/macros.h
+
+AM_CPPFLAGS = \
+	$(AMDGPU_CFLAGS) \
+	$(VALGRIND_CFLAGS) \
+	$(DEFINES) \
+	-I$(top_srcdir)/include \
+	-I$(top_builddir)/src \
+	-I$(top_srcdir)/src \
+	-I$(top_srcdir)/src/amd \
+	-I$(top_srcdir)/src/amd/common \
+	-I$(top_builddir)/src/compiler \
+	-I$(top_builddir)/src/compiler/nir \
+	-I$(top_srcdir)/src/compiler \
+	-I$(top_srcdir)/src/mapi \
+	-I$(top_srcdir)/src/mesa \
+	-I$(top_srcdir)/src/mesa/drivers/dri/common \
+	-I$(top_srcdir)/src/gallium/auxiliary \
+	-I$(top_srcdir)/src/gallium/include
+
+AM_CFLAGS = $(VISIBILITY_FLAGS) \
+	$(PTHREAD_CFLAGS) \
+	$(LLVM_CFLAGS)
+
+VULKAN_SOURCES = \
+	$(VULKAN_GENERATED_FILES) \
+	$(VULKAN_FILES)
+
+VULKAN_LIB_DEPS = $(AMDGPU_LIBS)
+
+
+if HAVE_PLATFORM_X11
+AM_CPPFLAGS += \
+	$(XCB_DRI3_CFLAGS) \
+	-DVK_USE_PLATFORM_XCB_KHR \
+	-DVK_USE_PLATFORM_XLIB_KHR
+
+VULKAN_SOURCES += $(VULKAN_WSI_X11_FILES)
+
+# FIXME: Use pkg-config for X11-xcb ldflags.
+VULKAN_LIB_DEPS += $(XCB_DRI3_LIBS) -lX11-xcb
+endif
+
+
+if HAVE_PLATFORM_WAYLAND
+AM_CPPFLAGS += \
+	-I$(top_builddir)/src/egl/wayland/wayland-drm \
+	-I$(top_srcdir)/src/egl/wayland/wayland-drm \
+	$(WAYLAND_CFLAGS) \
+	-DVK_USE_PLATFORM_WAYLAND_KHR
+
+VULKAN_SOURCES += $(VULKAN_WSI_WAYLAND_FILES)
+
+VULKAN_LIB_DEPS += \
+	$(top_builddir)/src/egl/wayland/wayland-drm/libwayland-drm.la \
+	$(WAYLAND_LIBS)
+endif
+
+noinst_LTLIBRARIES = libvulkan_common.la
+libvulkan_common_la_SOURCES = $(VULKAN_SOURCES)
+
+VULKAN_LIB_DEPS += \
+	libvulkan_common.la \
+	$(top_builddir)/src/amd/common/libamd_common.la \
+	$(top_builddir)/src/compiler/nir/libnir.la \
+	$(top_builddir)/src/util/libmesautil.la \
+	$(LLVM_LIBS) \
+	$(LIBELF_LIBS) \
+	$(PTHREAD_LIBS) \
+	$(LIBDRM_LIBS) \
+	$(PTHREAD_LIBS) \
+	$(DLOPEN_LIBS) \
+	-lm
+
+nodist_EXTRA_libvulkan_radeon_la_SOURCES = dummy.cpp
+libvulkan_radeon_la_SOURCES = $(VULKAN_GEM_FILES)
+
+radv_entrypoints.h : radv_entrypoints_gen.py $(vulkan_include_HEADERS)
+	$(AM_V_GEN) cat $(vulkan_include_HEADERS) |\
+	$(PYTHON2) $(srcdir)/radv_entrypoints_gen.py header > $@
+
+radv_entrypoints.c : radv_entrypoints_gen.py $(vulkan_include_HEADERS)
+	$(AM_V_GEN) cat $(vulkan_include_HEADERS) |\
+	$(PYTHON2) $(srcdir)/radv_entrypoints_gen.py code > $@
+
+.PHONY: radv_timestamp.h
+
+radv_timestamp.h:
+	@echo "Updating radv_timestamp.h"
+	$(AM_V_GEN) echo "#define RADV_TIMESTAMP \"$(TIMESTAMP_CMD)\"" > $@
+
+vk_format_table.c: vk_format_table.py \
+		   vk_format_parse.py \
+                   vk_format_layout.csv
+	$(PYTHON2) $(srcdir)/vk_format_table.py $(srcdir)/vk_format_layout.csv > $@
+
+BUILT_SOURCES = $(VULKAN_GENERATED_FILES)
+CLEANFILES = $(BUILT_SOURCES) dev_icd.json radv_timestamp.h
+EXTRA_DIST = \
+	$(top_srcdir)/include/vulkan/vk_icd.h \
+	radv_entrypoints_gen.py \
+	dev_icd.json.in \
+	radeon_icd.json
+
+libvulkan_radeon_la_LIBADD = $(VULKAN_LIB_DEPS) $(top_builddir)/src/amd/addrlib/libamdgpu_addrlib.la
+
+libvulkan_radeon_la_LDFLAGS = \
+	-shared \
+	-module \
+	-no-undefined \
+	-avoid-version \
+	$(BSYMBOLIC) \
+	$(LLVM_LDFLAGS) \
+	$(GC_SECTIONS) \
+	$(LD_NO_UNDEFINED)
+
+
+icdconfdir = @VULKAN_ICD_INSTALL_DIR@
+icdconf_DATA = radeon_icd.json
+# The following is used for development purposes, by setting VK_ICD_FILENAMES.
+noinst_DATA = dev_icd.json
+
+dev_icd.json : dev_icd.json.in
+	$(AM_V_GEN) $(SED) \
+		-e "s#@build_libdir@#${abs_top_builddir}/${LIB_DIR}#" \
+		< $(srcdir)/dev_icd.json.in > $@
+
+include $(top_srcdir)/install-lib-links.mk
+
+noinst_HEADERS =
+
+LDADD = \
+	$(PTHREAD_LIBS) -lm -lstdc++
+
diff --git a/src/amd/vulkan/Makefile.sources b/src/amd/vulkan/Makefile.sources
new file mode 100644
index 0000000000..97fd0b686f
--- /dev/null
+++ b/src/amd/vulkan/Makefile.sources
@@ -0,0 +1,67 @@
+# Copyright © 2016 Red Hat
+#
+# Permission is hereby granted, free of charge, to any person obtaining a
+# copy of this software and associated documentation files (the "Software"),
+# to deal in the Software without restriction, including without limitation
+# the rights to use, copy, modify, merge, publish, distribute, sublicense,
+# and/or sell copies of the Software, and to permit persons to whom the
+# Software is furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice (including the next
+# paragraph) shall be included in all copies or substantial portions of the
+# Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+# IN THE SOFTWARE.
+
+RADV_WS_AMDGPU_FILES := \
+	winsys/amdgpu/radv_amdgpu_bo.c \
+	winsys/amdgpu/radv_amdgpu_cs.c \
+	winsys/amdgpu/radv_amdgpu_surface.c \
+	winsys/amdgpu/radv_amdgpu_winsys.c \
+	winsys/amdgpu/radv_amdgpu_winsys.h
+
+VULKAN_FILES := \
+	radv_cmd_buffer.c \
+	radv_device.c \
+	radv_descriptor_set.c \
+	radv_formats.c \
+	radv_image.c \
+	radv_meta.c \
+	radv_meta.h \
+	radv_meta_blit.c \
+	radv_meta_blit2d.c \
+	radv_meta_buffer.c \
+	radv_meta_bufimage.c \
+	radv_meta_clear.c \
+	radv_meta_copy.c \
+	radv_meta_decompress.c \
+	radv_meta_fast_clear.c \
+	radv_meta_resolve.c \
+	radv_meta_resolve_cs.c \
+	radv_pass.c \
+	radv_pipeline.c \
+	radv_pipeline_cache.c \
+	radv_query.c \
+	radv_util.c \
+	radv_wsi.c \
+	si_cmd_buffer.c \
+	vk_format_table.c \
+	$(RADV_WS_AMDGPU_FILES)
+
+VULKAN_WSI_WAYLAND_FILES := \
+	radv_wsi_wayland.c
+
+VULKAN_WSI_X11_FILES := \
+	radv_wsi_x11.c
+
+VULKAN_GENERATED_FILES := \
+	radv_entrypoints.c \
+	radv_entrypoints.h \
+	radv_timestamp.h
+
diff --git a/src/amd/vulkan/dev_icd.json.in b/src/amd/vulkan/dev_icd.json.in
new file mode 100644
index 0000000000..fd8c7e3d74
--- /dev/null
+++ b/src/amd/vulkan/dev_icd.json.in
@@ -0,0 +1,7 @@
+{
+    "file_format_version": "1.0.0",
+    "ICD": {
+        "library_path": "@build_libdir@/libvulkan_radeon.so",
+        "abi_versions": "1.0.3"
+    }
+}
diff --git a/src/amd/vulkan/radeon_icd.json b/src/amd/vulkan/radeon_icd.json
new file mode 100644
index 0000000000..2085bd6788
--- /dev/null
+++ b/src/amd/vulkan/radeon_icd.json
@@ -0,0 +1,7 @@
+{
+    "file_format_version": "1.0.0",
+    "ICD": {
+        "library_path": "libvulkan_radeon.so",
+        "abi_versions": "1.0.3"
+    }
+}
diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
new file mode 100644
index 0000000000..e3e9e32632
--- /dev/null
+++ b/src/amd/vulkan/radv_cmd_buffer.c
@@ -0,0 +1,2413 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based in part on anv driver which is:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "radv_private.h"
+#include "radv_radeon_winsys.h"
+#include "radv_cs.h"
+#include "sid.h"
+#include "vk_format.h"
+#include "radv_meta.h"
+
+static void radv_handle_image_transition(struct radv_cmd_buffer *cmd_buffer,
+					 struct radv_image *image,
+					 VkImageLayout src_layout,
+					 VkImageLayout dst_layout,
+					 VkImageSubresourceRange range,
+					 VkImageAspectFlags pending_clears);
+
+const struct radv_dynamic_state default_dynamic_state = {
+	.viewport = {
+		.count = 0,
+	},
+	.scissor = {
+		.count = 0,
+	},
+	.line_width = 1.0f,
+	.depth_bias = {
+		.bias = 0.0f,
+		.clamp = 0.0f,
+		.slope = 0.0f,
+	},
+	.blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
+	.depth_bounds = {
+		.min = 0.0f,
+		.max = 1.0f,
+	},
+	.stencil_compare_mask = {
+		.front = ~0u,
+		.back = ~0u,
+	},
+	.stencil_write_mask = {
+		.front = ~0u,
+		.back = ~0u,
+	},
+	.stencil_reference = {
+		.front = 0u,
+		.back = 0u,
+	},
+};
+
+void
+radv_dynamic_state_copy(struct radv_dynamic_state *dest,
+			const struct radv_dynamic_state *src,
+			uint32_t copy_mask)
+{
+	if (copy_mask & (1 << VK_DYNAMIC_STATE_VIEWPORT)) {
+		dest->viewport.count = src->viewport.count;
+		typed_memcpy(dest->viewport.viewports, src->viewport.viewports,
+			     src->viewport.count);
+	}
+
+	if (copy_mask & (1 << VK_DYNAMIC_STATE_SCISSOR)) {
+		dest->scissor.count = src->scissor.count;
+		typed_memcpy(dest->scissor.scissors, src->scissor.scissors,
+			     src->scissor.count);
+	}
+
+	if (copy_mask & (1 << VK_DYNAMIC_STATE_LINE_WIDTH))
+		dest->line_width = src->line_width;
+
+	if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BIAS))
+		dest->depth_bias = src->depth_bias;
+
+	if (copy_mask & (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS))
+		typed_memcpy(dest->blend_constants, src->blend_constants, 4);
+
+	if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS))
+		dest->depth_bounds = src->depth_bounds;
+
+	if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK))
+		dest->stencil_compare_mask = src->stencil_compare_mask;
+
+	if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK))
+		dest->stencil_write_mask = src->stencil_write_mask;
+
+	if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE))
+		dest->stencil_reference = src->stencil_reference;
+}
+
+static VkResult radv_create_cmd_buffer(
+	struct radv_device *                         device,
+	struct radv_cmd_pool *                       pool,
+	VkCommandBufferLevel                        level,
+	VkCommandBuffer*                            pCommandBuffer)
+{
+	struct radv_cmd_buffer *cmd_buffer;
+	VkResult result;
+
+	cmd_buffer = radv_alloc(&pool->alloc, sizeof(*cmd_buffer), 8,
+				VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (cmd_buffer == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	memset(cmd_buffer, 0, sizeof(*cmd_buffer));
+	cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+	cmd_buffer->device = device;
+	cmd_buffer->pool = pool;
+	cmd_buffer->level = level;
+
+	if (pool) {
+		list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
+	} else {
+		/* Init the pool_link so we can safefly call list_del when we destroy
+		 * the command buffer
+		 */
+		list_inithead(&cmd_buffer->pool_link);
+	}
+
+	cmd_buffer->cs = device->ws->cs_create(device->ws, RING_GFX);
+	if (!cmd_buffer->cs) {
+		result = VK_ERROR_OUT_OF_HOST_MEMORY;
+		goto fail;
+	}
+
+	*pCommandBuffer = radv_cmd_buffer_to_handle(cmd_buffer);
+
+	cmd_buffer->upload.offset = 0;
+	cmd_buffer->upload.size = 0;
+	list_inithead(&cmd_buffer->upload.list);
+
+	return VK_SUCCESS;
+
+fail:
+	radv_free(&cmd_buffer->pool->alloc, cmd_buffer);
+
+	return result;
+}
+
+static bool
+radv_cmd_buffer_resize_upload_buf(struct radv_cmd_buffer *cmd_buffer,
+				  uint64_t min_needed)
+{
+	uint64_t new_size;
+	struct radeon_winsys_bo *bo;
+	struct radv_cmd_buffer_upload *upload;
+	struct radv_device *device = cmd_buffer->device;
+
+	new_size = MAX2(min_needed, 16 * 1024);
+	new_size = MAX2(new_size, 2 * cmd_buffer->upload.size);
+
+	bo = device->ws->buffer_create(device->ws,
+				       new_size, 4096,
+				       RADEON_DOMAIN_GTT,
+				       RADEON_FLAG_CPU_ACCESS);
+
+	if (!bo) {
+		cmd_buffer->record_fail = true;
+		return false;
+	}
+
+	device->ws->cs_add_buffer(cmd_buffer->cs, bo, 8);
+	if (cmd_buffer->upload.upload_bo) {
+		upload = malloc(sizeof(*upload));
+
+		if (!upload) {
+			cmd_buffer->record_fail = true;
+			device->ws->buffer_destroy(bo);
+			return false;
+		}
+
+		memcpy(upload, &cmd_buffer->upload, sizeof(*upload));
+		list_add(&upload->list, &cmd_buffer->upload.list);
+	}
+
+	cmd_buffer->upload.upload_bo = bo;
+	cmd_buffer->upload.size = new_size;
+	cmd_buffer->upload.offset = 0;
+	cmd_buffer->upload.map = device->ws->buffer_map(cmd_buffer->upload.upload_bo);
+
+	if (!cmd_buffer->upload.map) {
+		cmd_buffer->record_fail = true;
+		return false;
+	}
+
+	return true;
+}
+
+bool
+radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
+			     unsigned size,
+			     unsigned alignment,
+			     unsigned *out_offset,
+			     void **ptr)
+{
+	uint64_t offset = align(cmd_buffer->upload.offset, alignment);
+	if (offset + size > cmd_buffer->upload.size) {
+		if (!radv_cmd_buffer_resize_upload_buf(cmd_buffer, size))
+			return false;
+		offset = 0;
+	}
+
+	*out_offset = offset;
+	*ptr = cmd_buffer->upload.map + offset;
+
+	cmd_buffer->upload.offset = offset + size;
+	return true;
+}
+
+bool
+radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
+			    unsigned size, unsigned alignment,
+			    const void *data, unsigned *out_offset)
+{
+	uint8_t *ptr;
+
+	if (!radv_cmd_buffer_upload_alloc(cmd_buffer, size, alignment,
+					  out_offset, (void **)&ptr))
+		return false;
+
+	if (ptr)
+		memcpy(ptr, data, size);
+
+	return true;
+}
+
+static void
+radv_emit_graphics_blend_state(struct radv_cmd_buffer *cmd_buffer,
+			       struct radv_pipeline *pipeline)
+{
+	radeon_set_context_reg_seq(cmd_buffer->cs, R_028780_CB_BLEND0_CONTROL, 8);
+	radeon_emit_array(cmd_buffer->cs, pipeline->graphics.blend.cb_blend_control,
+			  8);
+	radeon_set_context_reg(cmd_buffer->cs, R_028808_CB_COLOR_CONTROL, pipeline->graphics.blend.cb_color_control);
+	radeon_set_context_reg(cmd_buffer->cs, R_028B70_DB_ALPHA_TO_MASK, pipeline->graphics.blend.db_alpha_to_mask);
+}
+
+static void
+radv_emit_graphics_depth_stencil_state(struct radv_cmd_buffer *cmd_buffer,
+				       struct radv_pipeline *pipeline)
+{
+	struct radv_depth_stencil_state *ds = &pipeline->graphics.ds;
+	radeon_set_context_reg(cmd_buffer->cs, R_028800_DB_DEPTH_CONTROL, ds->db_depth_control);
+	radeon_set_context_reg(cmd_buffer->cs, R_02842C_DB_STENCIL_CONTROL, ds->db_stencil_control);
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028000_DB_RENDER_CONTROL, ds->db_render_control);
+	radeon_set_context_reg(cmd_buffer->cs, R_028010_DB_RENDER_OVERRIDE2, ds->db_render_override2);
+}
+
+/* 12.4 fixed-point */
+static unsigned radv_pack_float_12p4(float x)
+{
+	return x <= 0    ? 0 :
+	       x >= 4096 ? 0xffff : x * 16;
+}
+
+static void
+radv_update_multisample_state(struct radv_cmd_buffer *cmd_buffer,
+			      struct radv_pipeline *pipeline)
+{
+	int num_samples = pipeline->graphics.ms.num_samples;
+	struct radv_multisample_state *ms = &pipeline->graphics.ms;
+	struct radv_pipeline *old_pipeline = cmd_buffer->state.emitted_pipeline;
+
+	radeon_set_context_reg_seq(cmd_buffer->cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
+	radeon_emit(cmd_buffer->cs, ms->pa_sc_aa_mask[0]);
+	radeon_emit(cmd_buffer->cs, ms->pa_sc_aa_mask[1]);
+
+	if (old_pipeline && num_samples == old_pipeline->graphics.ms.num_samples)
+		return;
+
+	radeon_set_context_reg_seq(cmd_buffer->cs, CM_R_028BDC_PA_SC_LINE_CNTL, 2);
+	radeon_emit(cmd_buffer->cs, ms->pa_sc_line_cntl);
+	radeon_emit(cmd_buffer->cs, ms->pa_sc_aa_config);
+
+	radeon_set_context_reg(cmd_buffer->cs, CM_R_028804_DB_EQAA, ms->db_eqaa);
+	radeon_set_context_reg(cmd_buffer->cs, EG_R_028A4C_PA_SC_MODE_CNTL_1, ms->pa_sc_mode_cntl_1);
+
+	radv_cayman_emit_msaa_sample_locs(cmd_buffer->cs, num_samples);
+
+	uint32_t samples_offset;
+	void *samples_ptr;
+	void *src;
+	radv_cmd_buffer_upload_alloc(cmd_buffer, num_samples * 4 * 2, 256, &samples_offset,
+				     &samples_ptr);
+	switch (num_samples) {
+	case 1:
+		src = cmd_buffer->device->sample_locations_1x;
+		break;
+	case 2:
+		src = cmd_buffer->device->sample_locations_2x;
+		break;
+	case 4:
+		src = cmd_buffer->device->sample_locations_4x;
+		break;
+	case 8:
+		src = cmd_buffer->device->sample_locations_8x;
+		break;
+	case 16:
+		src = cmd_buffer->device->sample_locations_16x;
+		break;
+	}
+	memcpy(samples_ptr, src, num_samples * 4 * 2);
+
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(cmd_buffer->upload.upload_bo);
+	va += samples_offset;
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B030_SPI_SHADER_USER_DATA_PS_0 + 10 * 4, 2);
+	radeon_emit(cmd_buffer->cs, va);
+	radeon_emit(cmd_buffer->cs, va >> 32);
+}
+
+static void
+radv_emit_graphics_raster_state(struct radv_cmd_buffer *cmd_buffer,
+				struct radv_pipeline *pipeline)
+{
+	struct radv_raster_state *raster = &pipeline->graphics.raster;
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028810_PA_CL_CLIP_CNTL,
+			       raster->pa_cl_clip_cntl);
+
+	radeon_set_context_reg(cmd_buffer->cs, R_0286D4_SPI_INTERP_CONTROL_0,
+			       raster->spi_interp_control);
+
+	radeon_set_context_reg_seq(cmd_buffer->cs, R_028A00_PA_SU_POINT_SIZE, 2);
+	radeon_emit(cmd_buffer->cs, 0);
+	radeon_emit(cmd_buffer->cs, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
+		    S_028A04_MAX_SIZE(radv_pack_float_12p4(8192/2))); /* R_028A04_PA_SU_POINT_MINMAX */
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028BE4_PA_SU_VTX_CNTL,
+			       raster->pa_su_vtx_cntl);
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028814_PA_SU_SC_MODE_CNTL,
+			       raster->pa_su_sc_mode_cntl);
+}
+
+static void
+radv_emit_vertex_shader(struct radv_cmd_buffer *cmd_buffer,
+			struct radv_pipeline *pipeline)
+{
+	struct radeon_winsys *ws = cmd_buffer->device->ws;
+	struct radv_shader_variant *vs;
+	uint64_t va;
+	unsigned export_count;
+	unsigned clip_dist_mask, cull_dist_mask, total_mask;
+
+	assert (pipeline->shaders[MESA_SHADER_VERTEX]);
+
+	vs = pipeline->shaders[MESA_SHADER_VERTEX];
+	va = ws->buffer_get_va(vs->bo);
+	ws->cs_add_buffer(cmd_buffer->cs, vs->bo, 8);
+
+	clip_dist_mask = vs->info.vs.clip_dist_mask;
+	cull_dist_mask = vs->info.vs.cull_dist_mask;
+	total_mask = clip_dist_mask | cull_dist_mask;
+	radeon_set_context_reg(cmd_buffer->cs, R_028A40_VGT_GS_MODE, 0);
+	radeon_set_context_reg(cmd_buffer->cs, R_028A84_VGT_PRIMITIVEID_EN, 0);
+
+	export_count = MAX2(1, vs->info.vs.param_exports);
+	radeon_set_context_reg(cmd_buffer->cs, R_0286C4_SPI_VS_OUT_CONFIG,
+			       S_0286C4_VS_EXPORT_COUNT(export_count - 1));
+	radeon_set_context_reg(cmd_buffer->cs, R_02870C_SPI_SHADER_POS_FORMAT,
+			       S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
+			       S_02870C_POS1_EXPORT_FORMAT(vs->info.vs.pos_exports > 1 ?
+							   V_02870C_SPI_SHADER_4COMP :
+							   V_02870C_SPI_SHADER_NONE) |
+			       S_02870C_POS2_EXPORT_FORMAT(vs->info.vs.pos_exports > 2 ?
+							   V_02870C_SPI_SHADER_4COMP :
+							   V_02870C_SPI_SHADER_NONE) |
+			       S_02870C_POS3_EXPORT_FORMAT(vs->info.vs.pos_exports > 3 ?
+							   V_02870C_SPI_SHADER_4COMP :
+							   V_02870C_SPI_SHADER_NONE));
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B120_SPI_SHADER_PGM_LO_VS, 4);
+	radeon_emit(cmd_buffer->cs, va >> 8);
+	radeon_emit(cmd_buffer->cs, va >> 40);
+	radeon_emit(cmd_buffer->cs, vs->rsrc1);
+	radeon_emit(cmd_buffer->cs, vs->rsrc2);
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028818_PA_CL_VTE_CNTL,
+			       S_028818_VTX_W0_FMT(1) |
+			       S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
+			       S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
+			       S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
+
+	radeon_set_context_reg(cmd_buffer->cs, R_02881C_PA_CL_VS_OUT_CNTL,
+			       S_02881C_USE_VTX_POINT_SIZE(vs->info.vs.writes_pointsize) |
+			       S_02881C_VS_OUT_MISC_VEC_ENA(vs->info.vs.writes_pointsize) |
+			       S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0f) != 0) |
+			       S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xf0) != 0) |
+			       pipeline->graphics.raster.pa_cl_vs_out_cntl |
+			       cull_dist_mask << 8 |
+			       clip_dist_mask);
+
+}
+
+
+
+static void
+radv_emit_fragment_shader(struct radv_cmd_buffer *cmd_buffer,
+			  struct radv_pipeline *pipeline)
+{
+	struct radeon_winsys *ws = cmd_buffer->device->ws;
+	struct radv_shader_variant *ps, *vs;
+	uint64_t va;
+	unsigned spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
+	struct radv_blend_state *blend = &pipeline->graphics.blend;
+	unsigned ps_offset = 0;
+	unsigned z_order;
+	assert (pipeline->shaders[MESA_SHADER_FRAGMENT]);
+
+	ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
+	vs = pipeline->shaders[MESA_SHADER_VERTEX];
+	va = ws->buffer_get_va(ps->bo);
+	ws->cs_add_buffer(cmd_buffer->cs, ps->bo, 8);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B020_SPI_SHADER_PGM_LO_PS, 4);
+	radeon_emit(cmd_buffer->cs, va >> 8);
+	radeon_emit(cmd_buffer->cs, va >> 40);
+	radeon_emit(cmd_buffer->cs, ps->rsrc1);
+	radeon_emit(cmd_buffer->cs, ps->rsrc2);
+
+	if (ps->info.fs.early_fragment_test || !ps->info.fs.writes_memory)
+		z_order = V_02880C_EARLY_Z_THEN_LATE_Z;
+	else
+		z_order = V_02880C_LATE_Z;
+
+
+	radeon_set_context_reg(cmd_buffer->cs, R_02880C_DB_SHADER_CONTROL,
+			       S_02880C_Z_EXPORT_ENABLE(ps->info.fs.writes_z) |
+			       S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps->info.fs.writes_stencil) |
+			       S_02880C_KILL_ENABLE(!!ps->info.fs.can_discard) |
+			       S_02880C_Z_ORDER(z_order) |
+			       S_02880C_DEPTH_BEFORE_SHADER(ps->info.fs.early_fragment_test) |
+			       S_02880C_EXEC_ON_HIER_FAIL(ps->info.fs.writes_memory) |
+			       S_02880C_EXEC_ON_NOOP(ps->info.fs.writes_memory));
+
+	radeon_set_context_reg(cmd_buffer->cs, R_0286CC_SPI_PS_INPUT_ENA,
+			       ps->config.spi_ps_input_ena);
+
+	radeon_set_context_reg(cmd_buffer->cs, R_0286D0_SPI_PS_INPUT_ADDR,
+			       ps->config.spi_ps_input_addr);
+
+	spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
+	radeon_set_context_reg(cmd_buffer->cs, R_0286D8_SPI_PS_IN_CONTROL,
+			       S_0286D8_NUM_INTERP(ps->info.fs.num_interp));
+
+	radeon_set_context_reg(cmd_buffer->cs, R_0286E0_SPI_BARYC_CNTL, spi_baryc_cntl);
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028710_SPI_SHADER_Z_FORMAT,
+			       ps->info.fs.writes_stencil ? V_028710_SPI_SHADER_32_GR :
+			       ps->info.fs.writes_z ? V_028710_SPI_SHADER_32_R :
+			       V_028710_SPI_SHADER_ZERO);
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028714_SPI_SHADER_COL_FORMAT, blend->spi_shader_col_format);
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028238_CB_TARGET_MASK, blend->cb_target_mask);
+	radeon_set_context_reg(cmd_buffer->cs, R_02823C_CB_SHADER_MASK, blend->cb_shader_mask);
+
+	if (ps->info.fs.has_pcoord) {
+		unsigned val;
+		val = S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20);
+		radeon_set_context_reg(cmd_buffer->cs, R_028644_SPI_PS_INPUT_CNTL_0 + 4 * ps_offset, val);
+		ps_offset = 1;
+	}
+
+	for (unsigned i = 0; i < 32 && (1u << i) <= ps->info.fs.input_mask; ++i) {
+		unsigned vs_offset, flat_shade;
+		unsigned val;
+
+		if (!(ps->info.fs.input_mask & (1u << i)))
+			continue;
+
+
+		if (!(vs->info.vs.export_mask & (1u << i))) {
+			radeon_set_context_reg(cmd_buffer->cs, R_028644_SPI_PS_INPUT_CNTL_0 + 4 * ps_offset,
+					       S_028644_OFFSET(0x20));
+			++ps_offset;
+			continue;
+		}
+
+		vs_offset = util_bitcount(vs->info.vs.export_mask & ((1u << i) - 1));
+		flat_shade = !!(ps->info.fs.flat_shaded_mask & (1u << ps_offset));
+
+		val = S_028644_OFFSET(vs_offset) | S_028644_FLAT_SHADE(flat_shade);
+		radeon_set_context_reg(cmd_buffer->cs, R_028644_SPI_PS_INPUT_CNTL_0 + 4 * ps_offset, val);
+		++ps_offset;
+	}
+}
+
+static void
+radv_emit_graphics_pipeline(struct radv_cmd_buffer *cmd_buffer,
+			    struct radv_pipeline *pipeline)
+{
+	if (!pipeline || cmd_buffer->state.emitted_pipeline == pipeline)
+		return;
+
+	radv_emit_graphics_depth_stencil_state(cmd_buffer, pipeline);
+	radv_emit_graphics_blend_state(cmd_buffer, pipeline);
+	radv_emit_graphics_raster_state(cmd_buffer, pipeline);
+	radv_update_multisample_state(cmd_buffer, pipeline);
+	radv_emit_vertex_shader(cmd_buffer, pipeline);
+	radv_emit_fragment_shader(cmd_buffer, pipeline);
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN,
+			       pipeline->graphics.prim_restart_enable);
+
+	cmd_buffer->state.emitted_pipeline = pipeline;
+}
+
+static void
+radv_emit_viewport(struct radv_cmd_buffer *cmd_buffer)
+{
+	si_write_viewport(cmd_buffer->cs, 0, cmd_buffer->state.dynamic.viewport.count,
+			  cmd_buffer->state.dynamic.viewport.viewports);
+}
+
+static void
+radv_emit_scissor(struct radv_cmd_buffer *cmd_buffer)
+{
+	uint32_t count = cmd_buffer->state.dynamic.scissor.count;
+	si_write_scissors(cmd_buffer->cs, 0, count,
+			  cmd_buffer->state.dynamic.scissor.scissors);
+	radeon_set_context_reg(cmd_buffer->cs, R_028A48_PA_SC_MODE_CNTL_0,
+			       cmd_buffer->state.pipeline->graphics.ms.pa_sc_mode_cntl_0 | S_028A48_VPORT_SCISSOR_ENABLE(count ? 1 : 0));
+}
+
+static void
+radv_emit_fb_color_state(struct radv_cmd_buffer *cmd_buffer,
+			 int index,
+			 struct radv_color_buffer_info *cb)
+{
+	bool is_vi = cmd_buffer->device->instance->physicalDevice.rad_info.chip_class >= VI;
+	radeon_set_context_reg_seq(cmd_buffer->cs, R_028C60_CB_COLOR0_BASE + index * 0x3c, 11);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_base);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_pitch);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_slice);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_view);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_info);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_attrib);
+	radeon_emit(cmd_buffer->cs, cb->cb_dcc_control);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_cmask);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_cmask_slice);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_fmask);
+	radeon_emit(cmd_buffer->cs, cb->cb_color_fmask_slice);
+
+	if (is_vi) { /* DCC BASE */
+		radeon_set_context_reg(cmd_buffer->cs, R_028C94_CB_COLOR0_DCC_BASE + index * 0x3c, cb->cb_dcc_base);
+	}
+}
+
+static void
+radv_emit_fb_ds_state(struct radv_cmd_buffer *cmd_buffer,
+		      struct radv_ds_buffer_info *ds,
+		      struct radv_image *image,
+		      VkImageLayout layout)
+{
+	uint32_t db_z_info = ds->db_z_info;
+
+	if (!radv_layout_has_htile(image, layout))
+		db_z_info &= C_028040_TILE_SURFACE_ENABLE;
+
+	if (!radv_layout_can_expclear(image, layout))
+		db_z_info &= C_028040_ALLOW_EXPCLEAR & C_028044_ALLOW_EXPCLEAR;
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028008_DB_DEPTH_VIEW, ds->db_depth_view);
+	radeon_set_context_reg(cmd_buffer->cs, R_028014_DB_HTILE_DATA_BASE, ds->db_htile_data_base);
+
+	radeon_set_context_reg_seq(cmd_buffer->cs, R_02803C_DB_DEPTH_INFO, 9);
+	radeon_emit(cmd_buffer->cs, ds->db_depth_info);	/* R_02803C_DB_DEPTH_INFO */
+	radeon_emit(cmd_buffer->cs, db_z_info);			/* R_028040_DB_Z_INFO */
+	radeon_emit(cmd_buffer->cs, ds->db_stencil_info);	/* R_028044_DB_STENCIL_INFO */
+	radeon_emit(cmd_buffer->cs, ds->db_z_read_base);	/* R_028048_DB_Z_READ_BASE */
+	radeon_emit(cmd_buffer->cs, ds->db_stencil_read_base);	/* R_02804C_DB_STENCIL_READ_BASE */
+	radeon_emit(cmd_buffer->cs, ds->db_z_write_base);	/* R_028050_DB_Z_WRITE_BASE */
+	radeon_emit(cmd_buffer->cs, ds->db_stencil_write_base);	/* R_028054_DB_STENCIL_WRITE_BASE */
+	radeon_emit(cmd_buffer->cs, ds->db_depth_size);	/* R_028058_DB_DEPTH_SIZE */
+	radeon_emit(cmd_buffer->cs, ds->db_depth_slice);	/* R_02805C_DB_DEPTH_SLICE */
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028ABC_DB_HTILE_SURFACE, ds->db_htile_surface);
+	radeon_set_context_reg(cmd_buffer->cs, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
+			       ds->pa_su_poly_offset_db_fmt_cntl);
+}
+
+/*
+ * To hw resolve multisample images both src and dst need to have the same
+ * micro tiling mode. However we don't always know in advance when creating
+ * the images. This function gets called if we have a resolve attachment,
+ * and tests if the attachment image has the same tiling mode, then it
+ * checks if the generated framebuffer data has the same tiling mode, and
+ * updates it if not.
+ */
+static void radv_set_optimal_micro_tile_mode(struct radv_device *device,
+					     struct radv_attachment_info *att,
+					     uint32_t micro_tile_mode)
+{
+	struct radv_image *image = att->attachment->image;
+	uint32_t tile_mode_index;
+	if (image->surface.nsamples <= 1)
+		return;
+
+	if (image->surface.micro_tile_mode != micro_tile_mode) {
+		radv_image_set_optimal_micro_tile_mode(device, image, micro_tile_mode);
+	}
+
+	if (att->cb.micro_tile_mode != micro_tile_mode) {
+		tile_mode_index = image->surface.tiling_index[0];
+
+		att->cb.cb_color_attrib &= C_028C74_TILE_MODE_INDEX;
+		att->cb.cb_color_attrib |= S_028C74_TILE_MODE_INDEX(tile_mode_index);
+		att->cb.micro_tile_mode = micro_tile_mode;
+	}
+}
+
+void
+radv_set_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
+			  struct radv_image *image,
+			  VkClearDepthStencilValue ds_clear_value,
+			  VkImageAspectFlags aspects)
+{
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(image->bo);
+	va += image->offset + image->clear_value_offset;
+	unsigned reg_offset = 0, reg_count = 0;
+
+	if (!image->htile.size || !aspects)
+		return;
+
+	if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
+		++reg_count;
+	} else {
+		++reg_offset;
+		va += 4;
+	}
+	if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
+		++reg_count;
+
+	cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, image->bo, 8);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_WRITE_DATA, 2 + reg_count, 0));
+	radeon_emit(cmd_buffer->cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
+				    S_370_WR_CONFIRM(1) |
+				    S_370_ENGINE_SEL(V_370_PFP));
+	radeon_emit(cmd_buffer->cs, va);
+	radeon_emit(cmd_buffer->cs, va >> 32);
+	if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT)
+		radeon_emit(cmd_buffer->cs, ds_clear_value.stencil);
+	if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
+		radeon_emit(cmd_buffer->cs, fui(ds_clear_value.depth));
+
+	radeon_set_context_reg_seq(cmd_buffer->cs, R_028028_DB_STENCIL_CLEAR + 4 * reg_offset, reg_count);
+	if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT)
+		radeon_emit(cmd_buffer->cs, ds_clear_value.stencil); /* R_028028_DB_STENCIL_CLEAR */
+	if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
+		radeon_emit(cmd_buffer->cs, fui(ds_clear_value.depth)); /* R_02802C_DB_DEPTH_CLEAR */
+}
+
+static void
+radv_load_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
+			   struct radv_image *image)
+{
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(image->bo);
+	va += image->offset + image->clear_value_offset;
+
+	if (!image->htile.size)
+		return;
+
+	cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, image->bo, 8);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_COPY_DATA, 4, 0));
+	radeon_emit(cmd_buffer->cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
+				    COPY_DATA_DST_SEL(COPY_DATA_REG) |
+				    COPY_DATA_COUNT_SEL);
+	radeon_emit(cmd_buffer->cs, va);
+	radeon_emit(cmd_buffer->cs, va >> 32);
+	radeon_emit(cmd_buffer->cs, R_028028_DB_STENCIL_CLEAR >> 2);
+	radeon_emit(cmd_buffer->cs, 0);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
+	radeon_emit(cmd_buffer->cs, 0);
+}
+
+void
+radv_set_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
+			  struct radv_image *image,
+			  int idx,
+			  uint32_t color_values[2])
+{
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(image->bo);
+	va += image->offset + image->clear_value_offset;
+
+	if (!image->cmask.size && !image->surface.dcc_size)
+		return;
+
+	cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, image->bo, 8);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_WRITE_DATA, 4, 0));
+	radeon_emit(cmd_buffer->cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
+				    S_370_WR_CONFIRM(1) |
+				    S_370_ENGINE_SEL(V_370_PFP));
+	radeon_emit(cmd_buffer->cs, va);
+	radeon_emit(cmd_buffer->cs, va >> 32);
+	radeon_emit(cmd_buffer->cs, color_values[0]);
+	radeon_emit(cmd_buffer->cs, color_values[1]);
+
+	radeon_set_context_reg_seq(cmd_buffer->cs, R_028C8C_CB_COLOR0_CLEAR_WORD0 + idx * 0x3c, 2);
+	radeon_emit(cmd_buffer->cs, color_values[0]);
+	radeon_emit(cmd_buffer->cs, color_values[1]);
+}
+
+static void
+radv_load_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
+			   struct radv_image *image,
+			   int idx)
+{
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(image->bo);
+	va += image->offset + image->clear_value_offset;
+
+	if (!image->cmask.size && !image->surface.dcc_size)
+		return;
+
+	uint32_t reg = R_028C8C_CB_COLOR0_CLEAR_WORD0 + idx * 0x3c;
+	cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, image->bo, 8);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_COPY_DATA, 4, 0));
+	radeon_emit(cmd_buffer->cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
+				    COPY_DATA_DST_SEL(COPY_DATA_REG) |
+				    COPY_DATA_COUNT_SEL);
+	radeon_emit(cmd_buffer->cs, va);
+	radeon_emit(cmd_buffer->cs, va >> 32);
+	radeon_emit(cmd_buffer->cs, reg >> 2);
+	radeon_emit(cmd_buffer->cs, 0);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
+	radeon_emit(cmd_buffer->cs, 0);
+}
+
+void
+radv_emit_framebuffer_state(struct radv_cmd_buffer *cmd_buffer)
+{
+	int i;
+	struct radv_framebuffer *framebuffer = cmd_buffer->state.framebuffer;
+	const struct radv_subpass *subpass = cmd_buffer->state.subpass;
+	int dst_resolve_micro_tile_mode = -1;
+
+	if (subpass->has_resolve) {
+		uint32_t a = subpass->resolve_attachments[0].attachment;
+		const struct radv_image *image = framebuffer->attachments[a].attachment->image;
+		dst_resolve_micro_tile_mode = image->surface.micro_tile_mode;
+	}
+	for (i = 0; i < subpass->color_count; ++i) {
+		int idx = subpass->color_attachments[i].attachment;
+		struct radv_attachment_info *att = &framebuffer->attachments[idx];
+
+		if (dst_resolve_micro_tile_mode != -1) {
+			radv_set_optimal_micro_tile_mode(cmd_buffer->device,
+							 att, dst_resolve_micro_tile_mode);
+		}
+		cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, att->attachment->bo, 8);
+
+		assert(att->attachment->aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT);
+		radv_emit_fb_color_state(cmd_buffer, i, &att->cb);
+
+		radv_load_color_clear_regs(cmd_buffer, att->attachment->image, i);
+	}
+
+	for (i = subpass->color_count; i < 8; i++)
+		radeon_set_context_reg(cmd_buffer->cs, R_028C70_CB_COLOR0_INFO + i * 0x3C,
+				       S_028C70_FORMAT(V_028C70_COLOR_INVALID));
+
+	if(subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
+		int idx = subpass->depth_stencil_attachment.attachment;
+		VkImageLayout layout = subpass->depth_stencil_attachment.layout;
+		struct radv_attachment_info *att = &framebuffer->attachments[idx];
+		struct radv_image *image = att->attachment->image;
+		cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, att->attachment->bo, 8);
+
+		radv_emit_fb_ds_state(cmd_buffer, &att->ds, image, layout);
+
+		if (att->ds.offset_scale != cmd_buffer->state.offset_scale) {
+			cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
+			cmd_buffer->state.offset_scale = att->ds.offset_scale;
+		}
+		radv_load_depth_clear_regs(cmd_buffer, image);
+	} else {
+		radeon_set_context_reg_seq(cmd_buffer->cs, R_028040_DB_Z_INFO, 2);
+		radeon_emit(cmd_buffer->cs, S_028040_FORMAT(V_028040_Z_INVALID)); /* R_028040_DB_Z_INFO */
+		radeon_emit(cmd_buffer->cs, S_028044_FORMAT(V_028044_STENCIL_INVALID)); /* R_028044_DB_STENCIL_INFO */
+	}
+	radeon_set_context_reg(cmd_buffer->cs, R_028208_PA_SC_WINDOW_SCISSOR_BR,
+			       S_028208_BR_X(framebuffer->width) |
+			       S_028208_BR_Y(framebuffer->height));
+}
+
+void radv_set_db_count_control(struct radv_cmd_buffer *cmd_buffer)
+{
+	uint32_t db_count_control;
+
+	if(!cmd_buffer->state.active_occlusion_queries) {
+		if (cmd_buffer->device->instance->physicalDevice.rad_info.chip_class >= CIK) {
+			db_count_control = 0;
+		} else {
+			db_count_control = S_028004_ZPASS_INCREMENT_DISABLE(1);
+		}
+	} else {
+		if (cmd_buffer->device->instance->physicalDevice.rad_info.chip_class >= CIK) {
+			db_count_control = S_028004_PERFECT_ZPASS_COUNTS(1) |
+				S_028004_SAMPLE_RATE(0) | /* TODO: set this to the number of samples of the current framebuffer */
+				S_028004_ZPASS_ENABLE(1) |
+				S_028004_SLICE_EVEN_ENABLE(1) |
+				S_028004_SLICE_ODD_ENABLE(1);
+		} else {
+			db_count_control = S_028004_PERFECT_ZPASS_COUNTS(1) |
+				S_028004_SAMPLE_RATE(0); /* TODO: set this to the number of samples of the current framebuffer */
+		}
+	}
+
+	radeon_set_context_reg(cmd_buffer->cs, R_028004_DB_COUNT_CONTROL, db_count_control);
+}
+
+static void
+radv_cmd_buffer_flush_dynamic_state(struct radv_cmd_buffer *cmd_buffer)
+{
+	struct radv_dynamic_state *d = &cmd_buffer->state.dynamic;
+
+	if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH) {
+		unsigned width = cmd_buffer->state.dynamic.line_width * 8;
+		radeon_set_context_reg(cmd_buffer->cs, R_028A08_PA_SU_LINE_CNTL,
+				       S_028A08_WIDTH(CLAMP(width, 0, 0xFFF)));
+	}
+
+	if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS) {
+		radeon_set_context_reg_seq(cmd_buffer->cs, R_028414_CB_BLEND_RED, 4);
+		radeon_emit_array(cmd_buffer->cs, (uint32_t*)d->blend_constants, 4);
+	}
+
+	if (cmd_buffer->state.dirty & (RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE |
+				       RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK |
+				       RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK)) {
+		radeon_set_context_reg_seq(cmd_buffer->cs, R_028430_DB_STENCILREFMASK, 2);
+		radeon_emit(cmd_buffer->cs, S_028430_STENCILTESTVAL(d->stencil_reference.front) |
+			    S_028430_STENCILMASK(d->stencil_compare_mask.front) |
+			    S_028430_STENCILWRITEMASK(d->stencil_write_mask.front) |
+			    S_028430_STENCILOPVAL(1));
+		radeon_emit(cmd_buffer->cs, S_028434_STENCILTESTVAL_BF(d->stencil_reference.back) |
+			    S_028434_STENCILMASK_BF(d->stencil_compare_mask.back) |
+			    S_028434_STENCILWRITEMASK_BF(d->stencil_write_mask.back) |
+			    S_028434_STENCILOPVAL_BF(1));
+	}
+
+	if (cmd_buffer->state.dirty & (RADV_CMD_DIRTY_PIPELINE |
+				       RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS)) {
+		radeon_set_context_reg(cmd_buffer->cs, R_028020_DB_DEPTH_BOUNDS_MIN, fui(d->depth_bounds.min));
+		radeon_set_context_reg(cmd_buffer->cs, R_028024_DB_DEPTH_BOUNDS_MAX, fui(d->depth_bounds.max));
+	}
+
+	if (cmd_buffer->state.dirty & (RADV_CMD_DIRTY_PIPELINE |
+				       RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS)) {
+		struct radv_raster_state *raster = &cmd_buffer->state.pipeline->graphics.raster;
+		unsigned slope = fui(d->depth_bias.slope * 16.0f);
+		unsigned bias = fui(d->depth_bias.bias * cmd_buffer->state.offset_scale);
+
+		if (G_028814_POLY_OFFSET_FRONT_ENABLE(raster->pa_su_sc_mode_cntl)) {
+			radeon_set_context_reg_seq(cmd_buffer->cs, R_028B7C_PA_SU_POLY_OFFSET_CLAMP, 5);
+			radeon_emit(cmd_buffer->cs, fui(d->depth_bias.clamp)); /* CLAMP */
+			radeon_emit(cmd_buffer->cs, slope); /* FRONT SCALE */
+			radeon_emit(cmd_buffer->cs, bias); /* FRONT OFFSET */
+			radeon_emit(cmd_buffer->cs, slope); /* BACK SCALE */
+			radeon_emit(cmd_buffer->cs, bias); /* BACK OFFSET */
+		}
+	}
+
+	cmd_buffer->state.dirty = 0;
+}
+
+static void
+radv_flush_constants(struct radv_cmd_buffer *cmd_buffer,
+		     struct radv_pipeline_layout *layout,
+		     VkShaderStageFlags stages) {
+	unsigned offset;
+	void *ptr;
+	uint64_t va;
+
+	stages &= cmd_buffer->push_constant_stages;
+	if (!stages || !layout || (!layout->push_constant_size && !layout->dynamic_offset_count))
+		return;
+
+	radv_cmd_buffer_upload_alloc(cmd_buffer, layout->push_constant_size +
+				     16 * layout->dynamic_offset_count,
+				     256, &offset, &ptr);
+
+	memcpy(ptr, cmd_buffer->push_constants, layout->push_constant_size);
+	memcpy((char*)ptr + layout->push_constant_size, cmd_buffer->dynamic_buffers,
+	       16 * layout->dynamic_offset_count);
+
+	va = cmd_buffer->device->ws->buffer_get_va(cmd_buffer->upload.upload_bo);
+	va += offset;
+
+	if (stages & VK_SHADER_STAGE_VERTEX_BIT) {
+		radeon_set_sh_reg_seq(cmd_buffer->cs,
+				      R_00B130_SPI_SHADER_USER_DATA_VS_0 + 8 * 4, 2);
+		radeon_emit(cmd_buffer->cs, va);
+		radeon_emit(cmd_buffer->cs, va >> 32);
+	}
+
+	if (stages & VK_SHADER_STAGE_FRAGMENT_BIT) {
+		radeon_set_sh_reg_seq(cmd_buffer->cs,
+				      R_00B030_SPI_SHADER_USER_DATA_PS_0 + 8 * 4, 2);
+		radeon_emit(cmd_buffer->cs, va);
+		radeon_emit(cmd_buffer->cs, va >> 32);
+	}
+
+	if (stages & VK_SHADER_STAGE_COMPUTE_BIT) {
+		radeon_set_sh_reg_seq(cmd_buffer->cs,
+				      R_00B900_COMPUTE_USER_DATA_0 + 8 * 4, 2);
+		radeon_emit(cmd_buffer->cs, va);
+		radeon_emit(cmd_buffer->cs, va >> 32);
+	}
+
+	cmd_buffer->push_constant_stages &= ~stages;
+}
+
+static void
+radv_cmd_buffer_flush_state(struct radv_cmd_buffer *cmd_buffer)
+{
+	struct radv_pipeline *pipeline = cmd_buffer->state.pipeline;
+	struct radv_device *device = cmd_buffer->device;
+	uint32_t ia_multi_vgt_param;
+	uint32_t ls_hs_config = 0;
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs,
+					      4096);
+
+	if ((cmd_buffer->state.vertex_descriptors_dirty || cmd_buffer->state.vb_dirty) &&
+	    cmd_buffer->state.pipeline->num_vertex_attribs) {
+		unsigned vb_offset;
+		void *vb_ptr;
+		uint32_t i = 0;
+		uint32_t num_attribs = cmd_buffer->state.pipeline->num_vertex_attribs;
+		uint64_t va;
+
+		/* allocate some descriptor state for vertex buffers */
+		radv_cmd_buffer_upload_alloc(cmd_buffer, num_attribs * 16, 256,
+					     &vb_offset, &vb_ptr);
+
+		for (i = 0; i < num_attribs; i++) {
+			uint32_t *desc = &((uint32_t *)vb_ptr)[i * 4];
+			uint32_t offset;
+			int vb = cmd_buffer->state.pipeline->va_binding[i];
+			struct radv_buffer *buffer = cmd_buffer->state.vertex_bindings[vb].buffer;
+			uint32_t stride = cmd_buffer->state.pipeline->binding_stride[vb];
+
+			device->ws->cs_add_buffer(cmd_buffer->cs, buffer->bo, 8);
+			va = device->ws->buffer_get_va(buffer->bo);
+
+			offset = cmd_buffer->state.vertex_bindings[vb].offset + cmd_buffer->state.pipeline->va_offset[i];
+			va += offset + buffer->offset;
+			desc[0] = va;
+			desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) | S_008F04_STRIDE(stride);
+			if (cmd_buffer->device->instance->physicalDevice.rad_info.chip_class <= CIK && stride)
+				desc[2] = (buffer->size - offset - cmd_buffer->state.pipeline->va_format_size[i]) / stride + 1;
+			else
+				desc[2] = buffer->size - offset;
+			desc[3] = cmd_buffer->state.pipeline->va_rsrc_word3[i];
+		}
+
+		va = device->ws->buffer_get_va(cmd_buffer->upload.upload_bo);
+		va += vb_offset;
+		radeon_set_sh_reg_seq(cmd_buffer->cs,
+				      R_00B130_SPI_SHADER_USER_DATA_VS_0 + 10 * 4, 2);
+		radeon_emit(cmd_buffer->cs, va);
+		radeon_emit(cmd_buffer->cs, va >> 32);
+
+	}
+
+	cmd_buffer->state.vertex_descriptors_dirty = false;
+	cmd_buffer->state.vb_dirty = 0;
+	if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_PIPELINE)
+		radv_emit_graphics_pipeline(cmd_buffer, pipeline);
+
+	if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_RENDER_TARGETS)
+		radv_emit_framebuffer_state(cmd_buffer);
+
+	if (cmd_buffer->state.dirty & (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT))
+		radv_emit_viewport(cmd_buffer);
+
+	if (cmd_buffer->state.dirty & (RADV_CMD_DIRTY_DYNAMIC_SCISSOR))
+		radv_emit_scissor(cmd_buffer);
+
+	if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_PIPELINE) {
+		radeon_set_context_reg(cmd_buffer->cs, R_028B54_VGT_SHADER_STAGES_EN, 0);
+		ia_multi_vgt_param = si_get_ia_multi_vgt_param(cmd_buffer);
+
+		if (cmd_buffer->device->instance->physicalDevice.rad_info.chip_class >= CIK) {
+			radeon_set_context_reg_idx(cmd_buffer->cs, R_028AA8_IA_MULTI_VGT_PARAM, 1, ia_multi_vgt_param);
+			radeon_set_context_reg_idx(cmd_buffer->cs, R_028B58_VGT_LS_HS_CONFIG, 2, ls_hs_config);
+			radeon_set_uconfig_reg_idx(cmd_buffer->cs, R_030908_VGT_PRIMITIVE_TYPE, 1, cmd_buffer->state.pipeline->graphics.prim);
+		} else {
+			radeon_set_config_reg(cmd_buffer->cs, R_008958_VGT_PRIMITIVE_TYPE, cmd_buffer->state.pipeline->graphics.prim);
+			radeon_set_context_reg(cmd_buffer->cs, R_028AA8_IA_MULTI_VGT_PARAM, ia_multi_vgt_param);
+			radeon_set_context_reg(cmd_buffer->cs, R_028B58_VGT_LS_HS_CONFIG, ls_hs_config);
+		}
+		radeon_set_context_reg(cmd_buffer->cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, cmd_buffer->state.pipeline->graphics.gs_out);
+	}
+
+	radv_cmd_buffer_flush_dynamic_state(cmd_buffer);
+
+	radv_flush_constants(cmd_buffer, cmd_buffer->state.pipeline->layout,
+			     VK_SHADER_STAGE_ALL_GRAPHICS);
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+
+	si_emit_cache_flush(cmd_buffer);
+}
+
+static void radv_stage_flush(struct radv_cmd_buffer *cmd_buffer,
+			     VkPipelineStageFlags src_stage_mask)
+{
+	if (src_stage_mask & (VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
+	                      VK_PIPELINE_STAGE_TRANSFER_BIT |
+	                      VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT |
+	                      VK_PIPELINE_STAGE_ALL_COMMANDS_BIT)) {
+		cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH;
+	}
+
+	if (src_stage_mask & (VK_PIPELINE_STAGE_TESSELLATION_CONTROL_SHADER_BIT |
+			      VK_PIPELINE_STAGE_TESSELLATION_EVALUATION_SHADER_BIT |
+			      VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT |
+			      VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
+			      VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
+			      VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
+			      VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT |
+			      VK_PIPELINE_STAGE_TRANSFER_BIT |
+			      VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT |
+			      VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT |
+			      VK_PIPELINE_STAGE_ALL_COMMANDS_BIT)) {
+		cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_PS_PARTIAL_FLUSH;
+	} else if (src_stage_mask & (VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT |
+	                             VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT |
+	                             VK_PIPELINE_STAGE_VERTEX_INPUT_BIT |
+	                             VK_PIPELINE_STAGE_VERTEX_SHADER_BIT)) {
+		cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_VS_PARTIAL_FLUSH;
+	}
+}
+
+static void radv_subpass_barrier(struct radv_cmd_buffer *cmd_buffer, const struct radv_subpass_barrier *barrier)
+{
+	radv_stage_flush(cmd_buffer, barrier->src_stage_mask);
+
+	/* TODO: actual cache flushes */
+}
+
+static void radv_handle_subpass_image_transition(struct radv_cmd_buffer *cmd_buffer,
+						 VkAttachmentReference att)
+{
+	unsigned idx = att.attachment;
+	struct radv_image_view *view = cmd_buffer->state.framebuffer->attachments[idx].attachment;
+	VkImageSubresourceRange range;
+	range.aspectMask = 0;
+	range.baseMipLevel = view->base_mip;
+	range.levelCount = 1;
+	range.baseArrayLayer = view->base_layer;
+	range.layerCount = cmd_buffer->state.framebuffer->layers;
+
+	radv_handle_image_transition(cmd_buffer,
+				     view->image,
+				     cmd_buffer->state.attachments[idx].current_layout,
+				     att.layout, range,
+				     cmd_buffer->state.attachments[idx].pending_clear_aspects);
+
+	cmd_buffer->state.attachments[idx].current_layout = att.layout;
+
+
+}
+
+void
+radv_cmd_buffer_set_subpass(struct radv_cmd_buffer *cmd_buffer,
+			    const struct radv_subpass *subpass, bool transitions)
+{
+	if (transitions) {
+		radv_subpass_barrier(cmd_buffer, &subpass->start_barrier);
+
+		for (unsigned i = 0; i < subpass->color_count; ++i) {
+			radv_handle_subpass_image_transition(cmd_buffer,
+							subpass->color_attachments[i]);
+		}
+
+		for (unsigned i = 0; i < subpass->input_count; ++i) {
+			radv_handle_subpass_image_transition(cmd_buffer,
+							subpass->input_attachments[i]);
+		}
+
+		if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
+			radv_handle_subpass_image_transition(cmd_buffer,
+							subpass->depth_stencil_attachment);
+		}
+	}
+
+	cmd_buffer->state.subpass = subpass;
+
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_RENDER_TARGETS;
+}
+
+static void
+radv_cmd_state_setup_attachments(struct radv_cmd_buffer *cmd_buffer,
+				 struct radv_render_pass *pass,
+				 const VkRenderPassBeginInfo *info)
+{
+	struct radv_cmd_state *state = &cmd_buffer->state;
+
+	if (pass->attachment_count == 0) {
+		state->attachments = NULL;
+		return;
+	}
+
+	state->attachments = radv_alloc(&cmd_buffer->pool->alloc,
+					pass->attachment_count *
+					sizeof(state->attachments[0]),
+					8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (state->attachments == NULL) {
+		/* FIXME: Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */
+		abort();
+	}
+
+	for (uint32_t i = 0; i < pass->attachment_count; ++i) {
+		struct radv_render_pass_attachment *att = &pass->attachments[i];
+		VkImageAspectFlags att_aspects = vk_format_aspects(att->format);
+		VkImageAspectFlags clear_aspects = 0;
+
+		if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
+			/* color attachment */
+			if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+				clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
+			}
+		} else {
+			/* depthstencil attachment */
+			if ((att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) &&
+			    att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+				clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
+			}
+			if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
+			    att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+				clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
+			}
+		}
+
+		state->attachments[i].pending_clear_aspects = clear_aspects;
+		if (clear_aspects && info) {
+			assert(info->clearValueCount > i);
+			state->attachments[i].clear_value = info->pClearValues[i];
+		}
+
+		state->attachments[i].current_layout = att->initial_layout;
+	}
+}
+
+VkResult radv_AllocateCommandBuffers(
+	VkDevice _device,
+	const VkCommandBufferAllocateInfo *pAllocateInfo,
+	VkCommandBuffer *pCommandBuffers)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_cmd_pool, pool, pAllocateInfo->commandPool);
+
+	VkResult result = VK_SUCCESS;
+	uint32_t i;
+
+	for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
+		result = radv_create_cmd_buffer(device, pool, pAllocateInfo->level,
+						&pCommandBuffers[i]);
+		if (result != VK_SUCCESS)
+			break;
+	}
+
+	if (result != VK_SUCCESS)
+		radv_FreeCommandBuffers(_device, pAllocateInfo->commandPool,
+					i, pCommandBuffers);
+
+	return result;
+}
+
+static void
+radv_cmd_buffer_destroy(struct radv_cmd_buffer *cmd_buffer)
+{
+	list_del(&cmd_buffer->pool_link);
+
+	list_for_each_entry_safe(struct radv_cmd_buffer_upload, up,
+				 &cmd_buffer->upload.list, list) {
+		cmd_buffer->device->ws->buffer_destroy(up->upload_bo);
+		list_del(&up->list);
+		free(up);
+	}
+
+	if (cmd_buffer->upload.upload_bo)
+		cmd_buffer->device->ws->buffer_destroy(cmd_buffer->upload.upload_bo);
+	cmd_buffer->device->ws->cs_destroy(cmd_buffer->cs);
+	radv_free(&cmd_buffer->pool->alloc, cmd_buffer);
+}
+
+void radv_FreeCommandBuffers(
+	VkDevice device,
+	VkCommandPool commandPool,
+	uint32_t commandBufferCount,
+	const VkCommandBuffer *pCommandBuffers)
+{
+	for (uint32_t i = 0; i < commandBufferCount; i++) {
+		RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
+
+		if (cmd_buffer)
+			radv_cmd_buffer_destroy(cmd_buffer);
+	}
+}
+
+static void  radv_reset_cmd_buffer(struct radv_cmd_buffer *cmd_buffer)
+{
+
+	cmd_buffer->device->ws->cs_reset(cmd_buffer->cs);
+
+	list_for_each_entry_safe(struct radv_cmd_buffer_upload, up,
+				 &cmd_buffer->upload.list, list) {
+		cmd_buffer->device->ws->buffer_destroy(up->upload_bo);
+		list_del(&up->list);
+		free(up);
+	}
+
+	if (cmd_buffer->upload.upload_bo)
+		cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs,
+						      cmd_buffer->upload.upload_bo, 8);
+	cmd_buffer->upload.offset = 0;
+
+	cmd_buffer->record_fail = false;
+}
+
+VkResult radv_ResetCommandBuffer(
+	VkCommandBuffer commandBuffer,
+	VkCommandBufferResetFlags flags)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	radv_reset_cmd_buffer(cmd_buffer);
+	return VK_SUCCESS;
+}
+
+VkResult radv_BeginCommandBuffer(
+	VkCommandBuffer commandBuffer,
+	const VkCommandBufferBeginInfo *pBeginInfo)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	radv_reset_cmd_buffer(cmd_buffer);
+
+	memset(&cmd_buffer->state, 0, sizeof(cmd_buffer->state));
+
+	/* setup initial configuration into command buffer */
+	if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
+		/* Flush read caches at the beginning of CS not flushed by the kernel. */
+		cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_INV_ICACHE |
+			RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
+			RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
+			RADV_CMD_FLAG_INV_VMEM_L1 |
+			RADV_CMD_FLAG_INV_SMEM_L1 |
+			RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER |
+			RADV_CMD_FLAG_INV_GLOBAL_L2;
+		si_init_config(&cmd_buffer->device->instance->physicalDevice, cmd_buffer);
+		radv_set_db_count_control(cmd_buffer);
+		si_emit_cache_flush(cmd_buffer);
+	}
+
+	if (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
+		cmd_buffer->state.framebuffer = radv_framebuffer_from_handle(pBeginInfo->pInheritanceInfo->framebuffer);
+		cmd_buffer->state.pass = radv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
+
+		struct radv_subpass *subpass =
+			&cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
+
+		radv_cmd_state_setup_attachments(cmd_buffer, cmd_buffer->state.pass, NULL);
+		radv_cmd_buffer_set_subpass(cmd_buffer, subpass, false);
+	}
+
+	return VK_SUCCESS;
+}
+
+void radv_CmdBindVertexBuffers(
+	VkCommandBuffer                             commandBuffer,
+	uint32_t                                    firstBinding,
+	uint32_t                                    bindingCount,
+	const VkBuffer*                             pBuffers,
+	const VkDeviceSize*                         pOffsets)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	struct radv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
+
+	/* We have to defer setting up vertex buffer since we need the buffer
+	 * stride from the pipeline. */
+
+	assert(firstBinding + bindingCount < MAX_VBS);
+	for (uint32_t i = 0; i < bindingCount; i++) {
+		vb[firstBinding + i].buffer = radv_buffer_from_handle(pBuffers[i]);
+		vb[firstBinding + i].offset = pOffsets[i];
+		cmd_buffer->state.vb_dirty |= 1 << (firstBinding + i);
+	}
+}
+
+void radv_CmdBindIndexBuffer(
+	VkCommandBuffer                             commandBuffer,
+	VkBuffer buffer,
+	VkDeviceSize offset,
+	VkIndexType indexType)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	cmd_buffer->state.index_buffer = radv_buffer_from_handle(buffer);
+	cmd_buffer->state.index_offset = offset;
+	cmd_buffer->state.index_type = indexType; /* vk matches hw */
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_INDEX_BUFFER;
+	cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, cmd_buffer->state.index_buffer->bo, 8);
+}
+
+
+void radv_bind_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
+			      struct radv_descriptor_set *set,
+			      unsigned idx)
+{
+	struct radeon_winsys *ws = cmd_buffer->device->ws;
+
+	cmd_buffer->state.descriptors[idx] = set;
+
+	if (!set)
+		return;
+
+	for (unsigned j = 0; j < set->layout->buffer_count; ++j)
+		if (set->descriptors[j])
+			ws->cs_add_buffer(cmd_buffer->cs, set->descriptors[j], 7);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs,
+				R_00B030_SPI_SHADER_USER_DATA_PS_0 + 8 * idx, 2);
+	radeon_emit(cmd_buffer->cs, set->va);
+	radeon_emit(cmd_buffer->cs, set->va >> 32);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs,
+				R_00B130_SPI_SHADER_USER_DATA_VS_0 + 8 * idx, 2);
+	radeon_emit(cmd_buffer->cs, set->va);
+	radeon_emit(cmd_buffer->cs, set->va >> 32);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs,
+				R_00B900_COMPUTE_USER_DATA_0 + 8 * idx, 2);
+	radeon_emit(cmd_buffer->cs, set->va);
+	radeon_emit(cmd_buffer->cs, set->va >> 32);
+
+	if(set->bo)
+		ws->cs_add_buffer(cmd_buffer->cs, set->bo, 8);
+}
+
+void radv_CmdBindDescriptorSets(
+	VkCommandBuffer                             commandBuffer,
+	VkPipelineBindPoint                         pipelineBindPoint,
+	VkPipelineLayout                            _layout,
+	uint32_t                                    firstSet,
+	uint32_t                                    descriptorSetCount,
+	const VkDescriptorSet*                      pDescriptorSets,
+	uint32_t                                    dynamicOffsetCount,
+	const uint32_t*                             pDynamicOffsets)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_pipeline_layout, layout, _layout);
+	unsigned dyn_idx = 0;
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs,
+					      MAX_SETS * 4 * 6);
+
+	for (unsigned i = 0; i < descriptorSetCount; ++i) {
+		unsigned idx = i + firstSet;
+		RADV_FROM_HANDLE(radv_descriptor_set, set, pDescriptorSets[i]);
+		radv_bind_descriptor_set(cmd_buffer, set, idx);
+
+		for(unsigned j = 0; j < set->layout->dynamic_offset_count; ++j, ++dyn_idx) {
+			unsigned idx = j + layout->set[i].dynamic_offset_start;
+			uint32_t *dst = cmd_buffer->dynamic_buffers + idx * 4;
+			assert(dyn_idx < dynamicOffsetCount);
+
+			struct radv_descriptor_range *range = set->dynamic_descriptors + j;
+			uint64_t va = range->va + pDynamicOffsets[dyn_idx];
+			dst[0] = va;
+			dst[1] = S_008F04_BASE_ADDRESS_HI(va >> 32);
+			dst[2] = range->size;
+			dst[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+			         S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+			         S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+			         S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+			         S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+			         S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
+			cmd_buffer->push_constant_stages |=
+			                     set->layout->dynamic_shader_stages;
+		}
+	}
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+void radv_CmdPushConstants(VkCommandBuffer commandBuffer,
+			   VkPipelineLayout layout,
+			   VkShaderStageFlags stageFlags,
+			   uint32_t offset,
+			   uint32_t size,
+			   const void* pValues)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	memcpy(cmd_buffer->push_constants + offset, pValues, size);
+	cmd_buffer->push_constant_stages |= stageFlags;
+}
+
+VkResult radv_EndCommandBuffer(
+	VkCommandBuffer                             commandBuffer)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	si_emit_cache_flush(cmd_buffer);
+	if (!cmd_buffer->device->ws->cs_finalize(cmd_buffer->cs) ||
+	    cmd_buffer->record_fail)
+		return VK_ERROR_OUT_OF_DEVICE_MEMORY;
+	return VK_SUCCESS;
+}
+
+static void
+radv_emit_compute_pipeline(struct radv_cmd_buffer *cmd_buffer)
+{
+	struct radeon_winsys *ws = cmd_buffer->device->ws;
+	struct radv_shader_variant *compute_shader;
+	struct radv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
+	uint64_t va;
+
+	if (!pipeline || pipeline == cmd_buffer->state.emitted_compute_pipeline)
+		return;
+
+	cmd_buffer->state.emitted_compute_pipeline = pipeline;
+
+	compute_shader = pipeline->shaders[MESA_SHADER_COMPUTE];
+	va = ws->buffer_get_va(compute_shader->bo);
+
+	ws->cs_add_buffer(cmd_buffer->cs, compute_shader->bo, 8);
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 16);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B830_COMPUTE_PGM_LO, 2);
+	radeon_emit(cmd_buffer->cs, va >> 8);
+	radeon_emit(cmd_buffer->cs, va >> 40);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
+	radeon_emit(cmd_buffer->cs, compute_shader->rsrc1);
+	radeon_emit(cmd_buffer->cs, compute_shader->rsrc2);
+
+	/* change these once we have scratch support */
+	radeon_set_sh_reg(cmd_buffer->cs, R_00B860_COMPUTE_TMPRING_SIZE,
+			  S_00B860_WAVES(32) | S_00B860_WAVESIZE(0));
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
+	radeon_emit(cmd_buffer->cs,
+		    S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[0]));
+	radeon_emit(cmd_buffer->cs,
+		    S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[1]));
+	radeon_emit(cmd_buffer->cs,
+		    S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[2]));
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+
+void radv_CmdBindPipeline(
+	VkCommandBuffer                             commandBuffer,
+	VkPipelineBindPoint                         pipelineBindPoint,
+	VkPipeline                                  _pipeline)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
+
+	switch (pipelineBindPoint) {
+	case VK_PIPELINE_BIND_POINT_COMPUTE:
+		cmd_buffer->state.compute_pipeline = pipeline;
+		cmd_buffer->push_constant_stages |= VK_SHADER_STAGE_COMPUTE_BIT;
+		break;
+	case VK_PIPELINE_BIND_POINT_GRAPHICS:
+		cmd_buffer->state.pipeline = pipeline;
+		cmd_buffer->state.vertex_descriptors_dirty = true;
+		cmd_buffer->state.dirty |= RADV_CMD_DIRTY_PIPELINE;
+		cmd_buffer->push_constant_stages |= pipeline->active_stages;
+
+		/* Apply the dynamic state from the pipeline */
+		cmd_buffer->state.dirty |= pipeline->dynamic_state_mask;
+		radv_dynamic_state_copy(&cmd_buffer->state.dynamic,
+					&pipeline->dynamic_state,
+					pipeline->dynamic_state_mask);
+		break;
+	default:
+		assert(!"invalid bind point");
+		break;
+	}
+}
+
+void radv_CmdSetViewport(
+	VkCommandBuffer                             commandBuffer,
+	uint32_t                                    firstViewport,
+	uint32_t                                    viewportCount,
+	const VkViewport*                           pViewports)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	const uint32_t total_count = firstViewport + viewportCount;
+	if (cmd_buffer->state.dynamic.viewport.count < total_count)
+		cmd_buffer->state.dynamic.viewport.count = total_count;
+
+	memcpy(cmd_buffer->state.dynamic.viewport.viewports + firstViewport,
+	       pViewports, viewportCount * sizeof(*pViewports));
+
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_VIEWPORT;
+}
+
+void radv_CmdSetScissor(
+	VkCommandBuffer                             commandBuffer,
+	uint32_t                                    firstScissor,
+	uint32_t                                    scissorCount,
+	const VkRect2D*                             pScissors)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	const uint32_t total_count = firstScissor + scissorCount;
+	if (cmd_buffer->state.dynamic.scissor.count < total_count)
+		cmd_buffer->state.dynamic.scissor.count = total_count;
+
+	memcpy(cmd_buffer->state.dynamic.scissor.scissors + firstScissor,
+	       pScissors, scissorCount * sizeof(*pScissors));
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_SCISSOR;
+}
+
+void radv_CmdSetLineWidth(
+	VkCommandBuffer                             commandBuffer,
+	float                                       lineWidth)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	cmd_buffer->state.dynamic.line_width = lineWidth;
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
+}
+
+void radv_CmdSetDepthBias(
+	VkCommandBuffer                             commandBuffer,
+	float                                       depthBiasConstantFactor,
+	float                                       depthBiasClamp,
+	float                                       depthBiasSlopeFactor)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	cmd_buffer->state.dynamic.depth_bias.bias = depthBiasConstantFactor;
+	cmd_buffer->state.dynamic.depth_bias.clamp = depthBiasClamp;
+	cmd_buffer->state.dynamic.depth_bias.slope = depthBiasSlopeFactor;
+
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
+}
+
+void radv_CmdSetBlendConstants(
+	VkCommandBuffer                             commandBuffer,
+	const float                                 blendConstants[4])
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	memcpy(cmd_buffer->state.dynamic.blend_constants,
+	       blendConstants, sizeof(float) * 4);
+
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
+}
+
+void radv_CmdSetDepthBounds(
+	VkCommandBuffer                             commandBuffer,
+	float                                       minDepthBounds,
+	float                                       maxDepthBounds)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	cmd_buffer->state.dynamic.depth_bounds.min = minDepthBounds;
+	cmd_buffer->state.dynamic.depth_bounds.max = maxDepthBounds;
+
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS;
+}
+
+void radv_CmdSetStencilCompareMask(
+	VkCommandBuffer                             commandBuffer,
+	VkStencilFaceFlags                          faceMask,
+	uint32_t                                    compareMask)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
+		cmd_buffer->state.dynamic.stencil_compare_mask.front = compareMask;
+	if (faceMask & VK_STENCIL_FACE_BACK_BIT)
+		cmd_buffer->state.dynamic.stencil_compare_mask.back = compareMask;
+
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
+}
+
+void radv_CmdSetStencilWriteMask(
+	VkCommandBuffer                             commandBuffer,
+	VkStencilFaceFlags                          faceMask,
+	uint32_t                                    writeMask)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
+		cmd_buffer->state.dynamic.stencil_write_mask.front = writeMask;
+	if (faceMask & VK_STENCIL_FACE_BACK_BIT)
+		cmd_buffer->state.dynamic.stencil_write_mask.back = writeMask;
+
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
+}
+
+void radv_CmdSetStencilReference(
+	VkCommandBuffer                             commandBuffer,
+	VkStencilFaceFlags                          faceMask,
+	uint32_t                                    reference)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
+		cmd_buffer->state.dynamic.stencil_reference.front = reference;
+	if (faceMask & VK_STENCIL_FACE_BACK_BIT)
+		cmd_buffer->state.dynamic.stencil_reference.back = reference;
+
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
+}
+
+
+void radv_CmdExecuteCommands(
+	VkCommandBuffer                             commandBuffer,
+	uint32_t                                    commandBufferCount,
+	const VkCommandBuffer*                      pCmdBuffers)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, primary, commandBuffer);
+
+	for (uint32_t i = 0; i < commandBufferCount; i++) {
+		RADV_FROM_HANDLE(radv_cmd_buffer, secondary, pCmdBuffers[i]);
+
+		primary->device->ws->cs_execute_secondary(primary->cs, secondary->cs);
+	}
+
+	/* if we execute secondary we need to re-emit out pipelines */
+	if (commandBufferCount) {
+		primary->state.emitted_pipeline = NULL;
+		primary->state.dirty |= RADV_CMD_DIRTY_PIPELINE;
+		primary->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_ALL;
+	}
+}
+
+VkResult radv_CreateCommandPool(
+	VkDevice                                    _device,
+	const VkCommandPoolCreateInfo*              pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkCommandPool*                              pCmdPool)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_cmd_pool *pool;
+
+	pool = radv_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
+			   VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (pool == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	if (pAllocator)
+		pool->alloc = *pAllocator;
+	else
+		pool->alloc = device->alloc;
+
+	list_inithead(&pool->cmd_buffers);
+
+	*pCmdPool = radv_cmd_pool_to_handle(pool);
+
+	return VK_SUCCESS;
+
+}
+
+void radv_DestroyCommandPool(
+	VkDevice                                    _device,
+	VkCommandPool                               commandPool,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_cmd_pool, pool, commandPool);
+
+	if (!pool)
+		return;
+
+	list_for_each_entry_safe(struct radv_cmd_buffer, cmd_buffer,
+				 &pool->cmd_buffers, pool_link) {
+		radv_cmd_buffer_destroy(cmd_buffer);
+	}
+
+	radv_free2(&device->alloc, pAllocator, pool);
+}
+
+VkResult radv_ResetCommandPool(
+	VkDevice                                    device,
+	VkCommandPool                               commandPool,
+	VkCommandPoolResetFlags                     flags)
+{
+	RADV_FROM_HANDLE(radv_cmd_pool, pool, commandPool);
+
+	list_for_each_entry(struct radv_cmd_buffer, cmd_buffer,
+			    &pool->cmd_buffers, pool_link) {
+		radv_reset_cmd_buffer(cmd_buffer);
+	}
+
+	return VK_SUCCESS;
+}
+
+void radv_CmdBeginRenderPass(
+	VkCommandBuffer                             commandBuffer,
+	const VkRenderPassBeginInfo*                pRenderPassBegin,
+	VkSubpassContents                           contents)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_render_pass, pass, pRenderPassBegin->renderPass);
+	RADV_FROM_HANDLE(radv_framebuffer, framebuffer, pRenderPassBegin->framebuffer);
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs,
+					      2048);
+
+	cmd_buffer->state.framebuffer = framebuffer;
+	cmd_buffer->state.pass = pass;
+	cmd_buffer->state.render_area = pRenderPassBegin->renderArea;
+	radv_cmd_state_setup_attachments(cmd_buffer, pass, pRenderPassBegin);
+
+	si_emit_cache_flush(cmd_buffer);
+
+	radv_cmd_buffer_set_subpass(cmd_buffer, pass->subpasses, true);
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+
+	radv_cmd_buffer_clear_subpass(cmd_buffer);
+}
+
+void radv_CmdNextSubpass(
+    VkCommandBuffer                             commandBuffer,
+    VkSubpassContents                           contents)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	si_emit_cache_flush(cmd_buffer);
+	radv_cmd_buffer_resolve_subpass(cmd_buffer);
+
+	radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs,
+					      2048);
+
+	radv_cmd_buffer_set_subpass(cmd_buffer, cmd_buffer->state.subpass + 1, true);
+	radv_cmd_buffer_clear_subpass(cmd_buffer);
+}
+
+void radv_CmdDraw(
+	VkCommandBuffer                             commandBuffer,
+	uint32_t                                    vertexCount,
+	uint32_t                                    instanceCount,
+	uint32_t                                    firstVertex,
+	uint32_t                                    firstInstance)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	radv_cmd_buffer_flush_state(cmd_buffer);
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 9);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B130_SPI_SHADER_USER_DATA_VS_0 + 12 * 4, 2);
+	radeon_emit(cmd_buffer->cs, firstVertex);
+	radeon_emit(cmd_buffer->cs, firstInstance);
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_NUM_INSTANCES, 0, 0));
+	radeon_emit(cmd_buffer->cs, instanceCount);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_DRAW_INDEX_AUTO, 1, 0));
+	radeon_emit(cmd_buffer->cs, vertexCount);
+	radeon_emit(cmd_buffer->cs, V_0287F0_DI_SRC_SEL_AUTO_INDEX |
+		    S_0287F0_USE_OPAQUE(0));
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+static void radv_emit_primitive_reset_index(struct radv_cmd_buffer *cmd_buffer)
+{
+	uint32_t primitive_reset_index = cmd_buffer->state.last_primitive_reset_index ? 0xffffffffu : 0xffffu;
+
+	if (cmd_buffer->state.pipeline->graphics.prim_restart_enable &&
+	    primitive_reset_index != cmd_buffer->state.last_primitive_reset_index) {
+		cmd_buffer->state.last_primitive_reset_index = primitive_reset_index;
+		radeon_set_context_reg(cmd_buffer->cs, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX,
+				       primitive_reset_index);
+	}
+}
+
+void radv_CmdDrawIndexed(
+	VkCommandBuffer                             commandBuffer,
+	uint32_t                                    indexCount,
+	uint32_t                                    instanceCount,
+	uint32_t                                    firstIndex,
+	int32_t                                     vertexOffset,
+	uint32_t                                    firstInstance)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	int index_size = cmd_buffer->state.index_type ? 4 : 2;
+	uint32_t index_max_size = (cmd_buffer->state.index_buffer->size - cmd_buffer->state.index_offset) / index_size;
+	uint64_t index_va;
+
+	radv_cmd_buffer_flush_state(cmd_buffer);
+	radv_emit_primitive_reset_index(cmd_buffer);
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 14);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_INDEX_TYPE, 0, 0));
+	radeon_emit(cmd_buffer->cs, cmd_buffer->state.index_type);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B130_SPI_SHADER_USER_DATA_VS_0 + 12 * 4, 2);
+	radeon_emit(cmd_buffer->cs, vertexOffset);
+	radeon_emit(cmd_buffer->cs, firstInstance);
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_NUM_INSTANCES, 0, 0));
+	radeon_emit(cmd_buffer->cs, instanceCount);
+
+	index_va = cmd_buffer->device->ws->buffer_get_va(cmd_buffer->state.index_buffer->bo);
+	index_va += firstIndex * index_size + cmd_buffer->state.index_buffer->offset + cmd_buffer->state.index_offset;
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_DRAW_INDEX_2, 4, false));
+	radeon_emit(cmd_buffer->cs, index_max_size);
+	radeon_emit(cmd_buffer->cs, index_va);
+	radeon_emit(cmd_buffer->cs, (index_va >> 32UL) & 0xFF);
+	radeon_emit(cmd_buffer->cs, indexCount);
+	radeon_emit(cmd_buffer->cs, V_0287F0_DI_SRC_SEL_DMA);
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+static void
+radv_emit_indirect_draw(struct radv_cmd_buffer *cmd_buffer,
+			VkBuffer _buffer,
+			VkDeviceSize offset,
+			uint32_t draw_count,
+			uint32_t stride,
+			bool indexed)
+{
+	RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+	unsigned di_src_sel = indexed ? V_0287F0_DI_SRC_SEL_DMA
+					    : V_0287F0_DI_SRC_SEL_AUTO_INDEX;
+	uint64_t indirect_va = cmd_buffer->device->ws->buffer_get_va(buffer->bo);
+	indirect_va += offset + buffer->offset;
+
+	if (!draw_count)
+		return;
+
+	cmd_buffer->device->ws->cs_add_buffer(cs, buffer->bo, 8);
+
+	radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0));
+	radeon_emit(cs, 1);
+	radeon_emit(cs, indirect_va);
+	radeon_emit(cs, indirect_va >> 32);
+
+	radeon_emit(cs, PKT3(indexed ? PKT3_DRAW_INDEX_INDIRECT_MULTI :
+				       PKT3_DRAW_INDIRECT_MULTI,
+			     8, false));
+	radeon_emit(cs, 0);
+	radeon_emit(cs, (R_00B160_SPI_SHADER_USER_DATA_VS_12 - SI_SH_REG_OFFSET) >> 2);
+	radeon_emit(cs, (R_00B164_SPI_SHADER_USER_DATA_VS_13 - SI_SH_REG_OFFSET) >> 2);
+	radeon_emit(cs, 0); /* draw_index */
+	radeon_emit(cs, draw_count); /* count */
+	radeon_emit(cs, 0); /* count_addr -- disabled */
+	radeon_emit(cs, 0);
+	radeon_emit(cs, stride); /* stride */
+	radeon_emit(cs, di_src_sel);
+}
+
+void radv_CmdDrawIndirect(
+	VkCommandBuffer                             commandBuffer,
+	VkBuffer                                    _buffer,
+	VkDeviceSize                                offset,
+	uint32_t                                    drawCount,
+	uint32_t                                    stride)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	radv_cmd_buffer_flush_state(cmd_buffer);
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 14);
+
+	radv_emit_indirect_draw(cmd_buffer, _buffer, offset, drawCount, stride, false);
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+void radv_CmdDrawIndexedIndirect(
+	VkCommandBuffer                             commandBuffer,
+	VkBuffer                                    _buffer,
+	VkDeviceSize                                offset,
+	uint32_t                                    drawCount,
+	uint32_t                                    stride)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	int index_size = cmd_buffer->state.index_type ? 4 : 2;
+	uint32_t index_max_size = (cmd_buffer->state.index_buffer->size - cmd_buffer->state.index_offset) / index_size;
+	uint64_t index_va;
+	radv_cmd_buffer_flush_state(cmd_buffer);
+	radv_emit_primitive_reset_index(cmd_buffer);
+
+	index_va = cmd_buffer->device->ws->buffer_get_va(cmd_buffer->state.index_buffer->bo);
+	index_va += cmd_buffer->state.index_buffer->offset + cmd_buffer->state.index_offset;
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 21);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_INDEX_TYPE, 0, 0));
+	radeon_emit(cmd_buffer->cs, cmd_buffer->state.index_type);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_INDEX_BASE, 1, 0));
+	radeon_emit(cmd_buffer->cs, index_va);
+	radeon_emit(cmd_buffer->cs, index_va >> 32);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_INDEX_BUFFER_SIZE, 0, 0));
+	radeon_emit(cmd_buffer->cs, index_max_size);
+
+	radv_emit_indirect_draw(cmd_buffer, _buffer, offset, drawCount, stride, true);
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+void radv_CmdDispatch(
+	VkCommandBuffer                             commandBuffer,
+	uint32_t                                    x,
+	uint32_t                                    y,
+	uint32_t                                    z)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	radv_emit_compute_pipeline(cmd_buffer);
+	radv_flush_constants(cmd_buffer, cmd_buffer->state.compute_pipeline->layout,
+			     VK_SHADER_STAGE_COMPUTE_BIT);
+	si_emit_cache_flush(cmd_buffer);
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 10);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B900_COMPUTE_USER_DATA_0 + 10 * 4, 3);
+	radeon_emit(cmd_buffer->cs, x);
+	radeon_emit(cmd_buffer->cs, y);
+	radeon_emit(cmd_buffer->cs, z);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_DISPATCH_DIRECT, 3, 0) |
+		    PKT3_SHADER_TYPE_S(1));
+	radeon_emit(cmd_buffer->cs, x);
+	radeon_emit(cmd_buffer->cs, y);
+	radeon_emit(cmd_buffer->cs, z);
+	radeon_emit(cmd_buffer->cs, 1);
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+void radv_CmdDispatchIndirect(
+	VkCommandBuffer                             commandBuffer,
+	VkBuffer                                    _buffer,
+	VkDeviceSize                                offset)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(buffer->bo);
+	va += buffer->offset + offset;
+
+	cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, buffer->bo, 8);
+
+	radv_emit_compute_pipeline(cmd_buffer);
+	radv_flush_constants(cmd_buffer, cmd_buffer->state.compute_pipeline->layout,
+			     VK_SHADER_STAGE_COMPUTE_BIT);
+	si_emit_cache_flush(cmd_buffer);
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 25);
+
+	for (unsigned i = 0; i < 3; ++i) {
+		radeon_emit(cmd_buffer->cs, PKT3(PKT3_COPY_DATA, 4, 0));
+		radeon_emit(cmd_buffer->cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
+					    COPY_DATA_DST_SEL(COPY_DATA_REG));
+		radeon_emit(cmd_buffer->cs, (va +  4 * i));
+		radeon_emit(cmd_buffer->cs, (va + 4 * i) >> 32);
+		radeon_emit(cmd_buffer->cs, (R_00B928_COMPUTE_USER_DATA_10 >> 2) + i);
+		radeon_emit(cmd_buffer->cs, 0);
+	}
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_SET_BASE, 2, 0) |
+				    PKT3_SHADER_TYPE_S(1));
+	radeon_emit(cmd_buffer->cs, 1);
+	radeon_emit(cmd_buffer->cs, va);
+	radeon_emit(cmd_buffer->cs, va >> 32);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_DISPATCH_INDIRECT, 1, 0) |
+				    PKT3_SHADER_TYPE_S(1));
+	radeon_emit(cmd_buffer->cs, 0);
+	radeon_emit(cmd_buffer->cs, 1);
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+void radv_unaligned_dispatch(
+	struct radv_cmd_buffer                      *cmd_buffer,
+	uint32_t                                    x,
+	uint32_t                                    y,
+	uint32_t                                    z)
+{
+	struct radv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
+	struct radv_shader_variant *compute_shader = pipeline->shaders[MESA_SHADER_COMPUTE];
+	uint32_t blocks[3], remainder[3];
+
+	blocks[0] = round_up_u32(x, compute_shader->info.cs.block_size[0]);
+	blocks[1] = round_up_u32(y, compute_shader->info.cs.block_size[1]);
+	blocks[2] = round_up_u32(z, compute_shader->info.cs.block_size[2]);
+
+	/* If aligned, these should be an entire block size, not 0 */
+	remainder[0] = x + compute_shader->info.cs.block_size[0] - align_u32_npot(x, compute_shader->info.cs.block_size[0]);
+	remainder[1] = y + compute_shader->info.cs.block_size[1] - align_u32_npot(y, compute_shader->info.cs.block_size[1]);
+	remainder[2] = z + compute_shader->info.cs.block_size[2] - align_u32_npot(z, compute_shader->info.cs.block_size[2]);
+
+	radv_emit_compute_pipeline(cmd_buffer);
+	radv_flush_constants(cmd_buffer, cmd_buffer->state.compute_pipeline->layout,
+			     VK_SHADER_STAGE_COMPUTE_BIT);
+	si_emit_cache_flush(cmd_buffer);
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 15);
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
+	radeon_emit(cmd_buffer->cs,
+		    S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[0]) |
+		    S_00B81C_NUM_THREAD_PARTIAL(remainder[0]));
+	radeon_emit(cmd_buffer->cs,
+		    S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[1]) |
+		    S_00B81C_NUM_THREAD_PARTIAL(remainder[1]));
+	radeon_emit(cmd_buffer->cs,
+		    S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[2]) |
+		    S_00B81C_NUM_THREAD_PARTIAL(remainder[2]));
+
+	radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B900_COMPUTE_USER_DATA_0 + 10 * 4, 3);
+	radeon_emit(cmd_buffer->cs, blocks[0]);
+	radeon_emit(cmd_buffer->cs, blocks[1]);
+	radeon_emit(cmd_buffer->cs, blocks[2]);
+
+	radeon_emit(cmd_buffer->cs, PKT3(PKT3_DISPATCH_DIRECT, 3, 0) |
+		    PKT3_SHADER_TYPE_S(1));
+	radeon_emit(cmd_buffer->cs, blocks[0]);
+	radeon_emit(cmd_buffer->cs, blocks[1]);
+	radeon_emit(cmd_buffer->cs, blocks[2]);
+	radeon_emit(cmd_buffer->cs, S_00B800_COMPUTE_SHADER_EN(1) |
+	                            S_00B800_PARTIAL_TG_EN(1));
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+void radv_CmdEndRenderPass(
+	VkCommandBuffer                             commandBuffer)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+
+	radv_subpass_barrier(cmd_buffer, &cmd_buffer->state.pass->end_barrier);
+
+	si_emit_cache_flush(cmd_buffer);
+	radv_cmd_buffer_resolve_subpass(cmd_buffer);
+
+	for (unsigned i = 0; i < cmd_buffer->state.framebuffer->attachment_count; ++i) {
+		VkImageLayout layout = cmd_buffer->state.pass->attachments[i].final_layout;
+		radv_handle_subpass_image_transition(cmd_buffer,
+		                      (VkAttachmentReference){i, layout});
+	}
+
+	radv_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
+
+	cmd_buffer->state.pass = NULL;
+	cmd_buffer->state.subpass = NULL;
+	cmd_buffer->state.attachments = NULL;
+	cmd_buffer->state.framebuffer = NULL;
+}
+
+
+static void radv_initialize_htile(struct radv_cmd_buffer *cmd_buffer,
+				  struct radv_image *image)
+{
+
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
+	                                RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
+
+	radv_fill_buffer(cmd_buffer, image->bo, image->offset + image->htile.offset,
+			 image->htile.size, 0xffffffff);
+
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB_META |
+	                                RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
+	                                RADV_CMD_FLAG_INV_VMEM_L1 |
+	                                RADV_CMD_FLAG_INV_GLOBAL_L2;
+}
+
+static void radv_handle_depth_image_transition(struct radv_cmd_buffer *cmd_buffer,
+					       struct radv_image *image,
+					       VkImageLayout src_layout,
+					       VkImageLayout dst_layout,
+					       VkImageSubresourceRange range,
+					       VkImageAspectFlags pending_clears)
+{
+	if (dst_layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL &&
+	    (pending_clears & vk_format_aspects(image->vk_format)) == vk_format_aspects(image->vk_format) &&
+	    cmd_buffer->state.render_area.offset.x == 0 && cmd_buffer->state.render_area.offset.y == 0 &&
+	    cmd_buffer->state.render_area.extent.width == image->extent.width &&
+	    cmd_buffer->state.render_area.extent.height == image->extent.height) {
+		/* The clear will initialize htile. */
+		return;
+	} else if (src_layout == VK_IMAGE_LAYOUT_UNDEFINED &&
+	           radv_layout_has_htile(image, dst_layout)) {
+		/* TODO: merge with the clear if applicable */
+		radv_initialize_htile(cmd_buffer, image);
+	} else if (!radv_layout_has_htile(image, src_layout) &&
+	           radv_layout_has_htile(image, dst_layout)) {
+		radv_initialize_htile(cmd_buffer, image);
+	} else if ((radv_layout_has_htile(image, src_layout) &&
+	            !radv_layout_has_htile(image, dst_layout)) ||
+	           (radv_layout_is_htile_compressed(image, src_layout) &&
+	            !radv_layout_is_htile_compressed(image, dst_layout))) {
+
+		range.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT;
+		range.baseMipLevel = 0;
+		range.levelCount = 1;
+
+		radv_decompress_depth_image_inplace(cmd_buffer, image, &range);
+	}
+}
+
+void radv_initialise_cmask(struct radv_cmd_buffer *cmd_buffer,
+			   struct radv_image *image, uint32_t value)
+{
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+	                                RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
+
+	radv_fill_buffer(cmd_buffer, image->bo, image->offset + image->cmask.offset,
+			 image->cmask.size, value);
+
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
+	                                RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
+	                                RADV_CMD_FLAG_INV_VMEM_L1 |
+	                                RADV_CMD_FLAG_INV_GLOBAL_L2;
+}
+
+static void radv_handle_cmask_image_transition(struct radv_cmd_buffer *cmd_buffer,
+					       struct radv_image *image,
+					       VkImageLayout src_layout,
+					       VkImageLayout dst_layout,
+					       VkImageSubresourceRange range,
+					       VkImageAspectFlags pending_clears)
+{
+	if (src_layout == VK_IMAGE_LAYOUT_UNDEFINED) {
+		if (image->fmask.size)
+			radv_initialise_cmask(cmd_buffer, image, 0xccccccccu);
+		else
+			radv_initialise_cmask(cmd_buffer, image, 0xffffffffu);
+	} else if (radv_layout_has_cmask(image, src_layout) &&
+		   !radv_layout_has_cmask(image, dst_layout)) {
+		radv_fast_clear_flush_image_inplace(cmd_buffer, image);
+	}
+}
+
+void radv_initialize_dcc(struct radv_cmd_buffer *cmd_buffer,
+			 struct radv_image *image, uint32_t value)
+{
+
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+	                                RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
+
+	radv_fill_buffer(cmd_buffer, image->bo, image->offset + image->dcc_offset,
+			 image->surface.dcc_size, value);
+
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+	                                RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
+	                                RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
+	                                RADV_CMD_FLAG_INV_VMEM_L1 |
+	                                RADV_CMD_FLAG_INV_GLOBAL_L2;
+}
+
+static void radv_handle_dcc_image_transition(struct radv_cmd_buffer *cmd_buffer,
+					     struct radv_image *image,
+					     VkImageLayout src_layout,
+					     VkImageLayout dst_layout,
+					     VkImageSubresourceRange range,
+					     VkImageAspectFlags pending_clears)
+{
+	if (src_layout == VK_IMAGE_LAYOUT_UNDEFINED) {
+		radv_initialize_dcc(cmd_buffer, image, 0x20202020u);
+	} else if(src_layout == VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL &&
+		  dst_layout != VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL) {
+		radv_fast_clear_flush_image_inplace(cmd_buffer, image);
+	}
+}
+
+static void radv_handle_image_transition(struct radv_cmd_buffer *cmd_buffer,
+					 struct radv_image *image,
+					 VkImageLayout src_layout,
+					 VkImageLayout dst_layout,
+					 VkImageSubresourceRange range,
+					 VkImageAspectFlags pending_clears)
+{
+	if (image->htile.size)
+		radv_handle_depth_image_transition(cmd_buffer, image, src_layout,
+						   dst_layout, range, pending_clears);
+
+	if (image->cmask.size)
+		radv_handle_cmask_image_transition(cmd_buffer, image, src_layout,
+						   dst_layout, range, pending_clears);
+
+	if (image->surface.dcc_size)
+		radv_handle_dcc_image_transition(cmd_buffer, image, src_layout,
+						 dst_layout, range, pending_clears);
+}
+
+void radv_CmdPipelineBarrier(
+	VkCommandBuffer                             commandBuffer,
+	VkPipelineStageFlags                        srcStageMask,
+	VkPipelineStageFlags                        destStageMask,
+	VkBool32                                    byRegion,
+	uint32_t                                    memoryBarrierCount,
+	const VkMemoryBarrier*                      pMemoryBarriers,
+	uint32_t                                    bufferMemoryBarrierCount,
+	const VkBufferMemoryBarrier*                pBufferMemoryBarriers,
+	uint32_t                                    imageMemoryBarrierCount,
+	const VkImageMemoryBarrier*                 pImageMemoryBarriers)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	VkAccessFlags src_flags = 0;
+	VkAccessFlags dst_flags = 0;
+	uint32_t b;
+	for (uint32_t i = 0; i < memoryBarrierCount; i++) {
+		src_flags |= pMemoryBarriers[i].srcAccessMask;
+		dst_flags |= pMemoryBarriers[i].dstAccessMask;
+	}
+
+	for (uint32_t i = 0; i < bufferMemoryBarrierCount; i++) {
+		src_flags |= pBufferMemoryBarriers[i].srcAccessMask;
+		dst_flags |= pBufferMemoryBarriers[i].dstAccessMask;
+	}
+
+	for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
+		RADV_FROM_HANDLE(radv_image, image, pImageMemoryBarriers[i].image);
+		src_flags |= pImageMemoryBarriers[i].srcAccessMask;
+		dst_flags |= pImageMemoryBarriers[i].dstAccessMask;
+
+		radv_handle_image_transition(cmd_buffer, image,
+					     pImageMemoryBarriers[i].oldLayout,
+					     pImageMemoryBarriers[i].newLayout,
+					     pImageMemoryBarriers[i].subresourceRange,
+					     0);
+	}
+
+	enum radv_cmd_flush_bits flush_bits = 0;
+
+	for_each_bit(b, src_flags) {
+		switch ((VkAccessFlagBits)(1 << b)) {
+		case VK_ACCESS_SHADER_WRITE_BIT:
+			flush_bits |= RADV_CMD_FLAG_INV_GLOBAL_L2;
+			break;
+		case VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT:
+			flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB;
+			break;
+		case VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT:
+			flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB;
+			break;
+		case VK_ACCESS_TRANSFER_WRITE_BIT:
+			flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB;
+			break;
+		default:
+			break;
+		}
+	}
+
+	for_each_bit(b, dst_flags) {
+		switch ((VkAccessFlagBits)(1 << b)) {
+		case VK_ACCESS_INDIRECT_COMMAND_READ_BIT:
+		case VK_ACCESS_INDEX_READ_BIT:
+		case VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT:
+		case VK_ACCESS_UNIFORM_READ_BIT:
+			flush_bits |= RADV_CMD_FLAG_INV_VMEM_L1;
+			break;
+		case VK_ACCESS_SHADER_READ_BIT:
+			flush_bits |= RADV_CMD_FLAG_INV_GLOBAL_L2;
+			break;
+		case VK_ACCESS_COLOR_ATTACHMENT_READ_BIT:
+		case VK_ACCESS_TRANSFER_READ_BIT:
+			flush_bits |= RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER | RADV_CMD_FLAG_INV_GLOBAL_L2;
+		default:
+			break;
+		}
+	}
+
+	flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
+		RADV_CMD_FLAG_PS_PARTIAL_FLUSH;
+
+	cmd_buffer->state.flush_bits |= flush_bits;
+}
+
+
+static void write_event(struct radv_cmd_buffer *cmd_buffer,
+			struct radv_event *event,
+			VkPipelineStageFlags stageMask,
+			unsigned value)
+{
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(event->bo);
+
+	cmd_buffer->device->ws->cs_add_buffer(cs, event->bo, 8);
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 12);
+
+	/* TODO: this is overkill. Probably should figure something out from
+	 * the stage mask. */
+
+	if (cmd_buffer->device->instance->physicalDevice.rad_info.chip_class == CIK) {
+		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_BOTTOM_OF_PIPE_TS) |
+				EVENT_INDEX(5));
+		radeon_emit(cs, va);
+		radeon_emit(cs, (va >> 32) | EOP_DATA_SEL(1));
+		radeon_emit(cs, 2);
+		radeon_emit(cs, 0);
+	}
+
+	radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+	radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_BOTTOM_OF_PIPE_TS) |
+			EVENT_INDEX(5));
+	radeon_emit(cs, va);
+	radeon_emit(cs, (va >> 32) | EOP_DATA_SEL(1));
+	radeon_emit(cs, value);
+	radeon_emit(cs, 0);
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
+
+void radv_CmdSetEvent(VkCommandBuffer commandBuffer,
+		      VkEvent _event,
+		      VkPipelineStageFlags stageMask)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_event, event, _event);
+
+	write_event(cmd_buffer, event, stageMask, 1);
+}
+
+void radv_CmdResetEvent(VkCommandBuffer commandBuffer,
+			VkEvent _event,
+			VkPipelineStageFlags stageMask)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_event, event, _event);
+
+	write_event(cmd_buffer, event, stageMask, 0);
+}
+
+void radv_CmdWaitEvents(VkCommandBuffer commandBuffer,
+			uint32_t eventCount,
+			const VkEvent* pEvents,
+			VkPipelineStageFlags srcStageMask,
+			VkPipelineStageFlags dstStageMask,
+			uint32_t memoryBarrierCount,
+			const VkMemoryBarrier* pMemoryBarriers,
+			uint32_t bufferMemoryBarrierCount,
+			const VkBufferMemoryBarrier* pBufferMemoryBarriers,
+			uint32_t imageMemoryBarrierCount,
+			const VkImageMemoryBarrier* pImageMemoryBarriers)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+
+	for (unsigned i = 0; i < eventCount; ++i) {
+		RADV_FROM_HANDLE(radv_event, event, pEvents[i]);
+		uint64_t va = cmd_buffer->device->ws->buffer_get_va(event->bo);
+
+		cmd_buffer->device->ws->cs_add_buffer(cs, event->bo, 8);
+
+		unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 7);
+
+		radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, 0));
+		radeon_emit(cs, WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_MEM_SPACE(1));
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+		radeon_emit(cs, 1); /* reference value */
+		radeon_emit(cs, 0xffffffff); /* mask */
+		radeon_emit(cs, 4); /* poll interval */
+
+		assert(cmd_buffer->cs->cdw <= cdw_max);
+	}
+
+
+	for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
+		RADV_FROM_HANDLE(radv_image, image, pImageMemoryBarriers[i].image);
+
+		radv_handle_image_transition(cmd_buffer, image,
+					     pImageMemoryBarriers[i].oldLayout,
+					     pImageMemoryBarriers[i].newLayout,
+					     pImageMemoryBarriers[i].subresourceRange,
+					     0);
+	}
+
+	/* TODO: figure out how to do memory barriers without waiting */
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER |
+					RADV_CMD_FLAG_INV_GLOBAL_L2 |
+					RADV_CMD_FLAG_INV_VMEM_L1 |
+					RADV_CMD_FLAG_INV_SMEM_L1;
+}
diff --git a/src/amd/vulkan/radv_cs.h b/src/amd/vulkan/radv_cs.h
new file mode 100644
index 0000000000..6481df1357
--- /dev/null
+++ b/src/amd/vulkan/radv_cs.h
@@ -0,0 +1,117 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#pragma once
+
+#include <string.h>
+#include <stdint.h>
+#include <assert.h>
+#include "r600d_common.h"
+
+static inline unsigned radeon_check_space(struct radeon_winsys *ws,
+                                      struct radeon_winsys_cs *cs,
+                                      unsigned needed)
+{
+        if (cs->max_dw - cs->cdw < needed)
+                ws->cs_grow(cs, needed);
+        return cs->cdw + needed;
+}
+
+static inline void radeon_set_config_reg_seq(struct radeon_winsys_cs *cs, unsigned reg, unsigned num)
+{
+        assert(reg < R600_CONTEXT_REG_OFFSET);
+        assert(cs->cdw + 2 + num <= cs->max_dw);
+        radeon_emit(cs, PKT3(PKT3_SET_CONFIG_REG, num, 0));
+        radeon_emit(cs, (reg - R600_CONFIG_REG_OFFSET) >> 2);
+}
+
+static inline void radeon_set_config_reg(struct radeon_winsys_cs *cs, unsigned reg, unsigned value)
+{
+        radeon_set_config_reg_seq(cs, reg, 1);
+        radeon_emit(cs, value);
+}
+
+static inline void radeon_set_context_reg_seq(struct radeon_winsys_cs *cs, unsigned reg, unsigned num)
+{
+        assert(reg >= R600_CONTEXT_REG_OFFSET);
+        assert(cs->cdw + 2 + num <= cs->max_dw);
+        radeon_emit(cs, PKT3(PKT3_SET_CONTEXT_REG, num, 0));
+        radeon_emit(cs, (reg - R600_CONTEXT_REG_OFFSET) >> 2);
+}
+
+static inline void radeon_set_context_reg(struct radeon_winsys_cs *cs, unsigned reg, unsigned value)
+{
+        radeon_set_context_reg_seq(cs, reg, 1);
+        radeon_emit(cs, value);
+}
+
+
+static inline void radeon_set_context_reg_idx(struct radeon_winsys_cs *cs,
+					      unsigned reg, unsigned idx,
+					      unsigned value)
+{
+	assert(reg >= R600_CONTEXT_REG_OFFSET);
+	assert(cs->cdw + 3 <= cs->max_dw);
+	radeon_emit(cs, PKT3(PKT3_SET_CONTEXT_REG, 1, 0));
+	radeon_emit(cs, (reg - R600_CONTEXT_REG_OFFSET) >> 2 | (idx << 28));
+	radeon_emit(cs, value);
+}
+
+static inline void radeon_set_sh_reg_seq(struct radeon_winsys_cs *cs, unsigned reg, unsigned num)
+{
+	assert(reg >= SI_SH_REG_OFFSET && reg < SI_SH_REG_END);
+	assert(cs->cdw + 2 + num <= cs->max_dw);
+	radeon_emit(cs, PKT3(PKT3_SET_SH_REG, num, 0));
+	radeon_emit(cs, (reg - SI_SH_REG_OFFSET) >> 2);
+}
+
+static inline void radeon_set_sh_reg(struct radeon_winsys_cs *cs, unsigned reg, unsigned value)
+{
+	radeon_set_sh_reg_seq(cs, reg, 1);
+	radeon_emit(cs, value);
+}
+
+static inline void radeon_set_uconfig_reg_seq(struct radeon_winsys_cs *cs, unsigned reg, unsigned num)
+{
+	assert(reg >= CIK_UCONFIG_REG_OFFSET && reg < CIK_UCONFIG_REG_END);
+	assert(cs->cdw + 2 + num <= cs->max_dw);
+	radeon_emit(cs, PKT3(PKT3_SET_UCONFIG_REG, num, 0));
+	radeon_emit(cs, (reg - CIK_UCONFIG_REG_OFFSET) >> 2);
+}
+
+static inline void radeon_set_uconfig_reg(struct radeon_winsys_cs *cs, unsigned reg, unsigned value)
+{
+	radeon_set_uconfig_reg_seq(cs, reg, 1);
+	radeon_emit(cs, value);
+}
+
+static inline void radeon_set_uconfig_reg_idx(struct radeon_winsys_cs *cs,
+					      unsigned reg, unsigned idx,
+					      unsigned value)
+{
+	assert(reg >= CIK_UCONFIG_REG_OFFSET && reg < CIK_UCONFIG_REG_END);
+	assert(cs->cdw + 3 <= cs->max_dw);
+	radeon_emit(cs, PKT3(PKT3_SET_UCONFIG_REG, 1, 0));
+	radeon_emit(cs, (reg - CIK_UCONFIG_REG_OFFSET) >> 2 | (idx << 28));
+	radeon_emit(cs, value);
+}
diff --git a/src/amd/vulkan/radv_descriptor_set.c b/src/amd/vulkan/radv_descriptor_set.c
new file mode 100644
index 0000000000..d1d2b1f4ce
--- /dev/null
+++ b/src/amd/vulkan/radv_descriptor_set.c
@@ -0,0 +1,716 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#include <assert.h>
+#include <stdbool.h>
+#include <string.h>
+#include <unistd.h>
+#include <fcntl.h>
+
+#include "util/mesa-sha1.h"
+#include "radv_private.h"
+#include "sid.h"
+
+VkResult radv_CreateDescriptorSetLayout(
+	VkDevice                                    _device,
+	const VkDescriptorSetLayoutCreateInfo*      pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkDescriptorSetLayout*                      pSetLayout)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_descriptor_set_layout *set_layout;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO);
+
+	uint32_t max_binding = 0;
+	uint32_t immutable_sampler_count = 0;
+	for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
+		max_binding = MAX(max_binding, pCreateInfo->pBindings[j].binding);
+		if (pCreateInfo->pBindings[j].pImmutableSamplers)
+			immutable_sampler_count += pCreateInfo->pBindings[j].descriptorCount;
+	}
+
+	size_t size = sizeof(struct radv_descriptor_set_layout) +
+		(max_binding + 1) * sizeof(set_layout->binding[0]) +
+		immutable_sampler_count * sizeof(struct radv_sampler *);
+
+	set_layout = radv_alloc2(&device->alloc, pAllocator, size, 8,
+				 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (!set_layout)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	/* We just allocate all the samplers at the end of the struct */
+	struct radv_sampler **samplers =
+		(struct radv_sampler **)&set_layout->binding[max_binding + 1];
+
+	set_layout->binding_count = max_binding + 1;
+	set_layout->shader_stages = 0;
+	set_layout->size = 0;
+
+	memset(set_layout->binding, 0, size - sizeof(struct radv_descriptor_set_layout));
+
+	uint32_t buffer_count = 0;
+	uint32_t dynamic_offset_count = 0;
+
+	for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
+		const VkDescriptorSetLayoutBinding *binding = &pCreateInfo->pBindings[j];
+		uint32_t b = binding->binding;
+		uint32_t alignment;
+
+		switch (binding->descriptorType) {
+		case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
+		case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
+			set_layout->binding[b].dynamic_offset_count = 1;
+			set_layout->dynamic_shader_stages |= binding->stageFlags;
+			set_layout->binding[b].size = 0;
+			set_layout->binding[b].buffer_count = 1;
+			alignment = 1;
+			break;
+		case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+		case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+		case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
+		case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
+			set_layout->binding[b].size = 16;
+			set_layout->binding[b].buffer_count = 1;
+			alignment = 16;
+			break;
+		case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
+		case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
+		case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
+			/* main descriptor + fmask descriptor */
+			set_layout->binding[b].size = 64;
+			set_layout->binding[b].buffer_count = 1;
+			alignment = 32;
+			break;
+		case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+			/* main descriptor + fmask descriptor + sampler */
+			set_layout->binding[b].size = 96;
+			set_layout->binding[b].buffer_count = 1;
+			alignment = 32;
+			break;
+		case VK_DESCRIPTOR_TYPE_SAMPLER:
+			set_layout->binding[b].size = 16;
+			alignment = 16;
+			break;
+		default:
+			break;
+		}
+
+		set_layout->size = align(set_layout->size, alignment);
+		assert(binding->descriptorCount > 0);
+		set_layout->binding[b].type = binding->descriptorType;
+		set_layout->binding[b].array_size = binding->descriptorCount;
+		set_layout->binding[b].offset = set_layout->size;
+		set_layout->binding[b].buffer_offset = buffer_count;
+		set_layout->binding[b].dynamic_offset_offset = dynamic_offset_count;
+
+		set_layout->size += binding->descriptorCount * set_layout->binding[b].size;
+		buffer_count += binding->descriptorCount * set_layout->binding[b].buffer_count;
+		dynamic_offset_count += binding->descriptorCount *
+			set_layout->binding[b].dynamic_offset_count;
+
+
+		if (binding->pImmutableSamplers) {
+			set_layout->binding[b].immutable_samplers = samplers;
+			samplers += binding->descriptorCount;
+
+			for (uint32_t i = 0; i < binding->descriptorCount; i++)
+				set_layout->binding[b].immutable_samplers[i] =
+					radv_sampler_from_handle(binding->pImmutableSamplers[i]);
+		} else {
+			set_layout->binding[b].immutable_samplers = NULL;
+		}
+
+		set_layout->shader_stages |= binding->stageFlags;
+	}
+
+	set_layout->buffer_count = buffer_count;
+	set_layout->dynamic_offset_count = dynamic_offset_count;
+
+	*pSetLayout = radv_descriptor_set_layout_to_handle(set_layout);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroyDescriptorSetLayout(
+	VkDevice                                    _device,
+	VkDescriptorSetLayout                       _set_layout,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_descriptor_set_layout, set_layout, _set_layout);
+
+	if (!set_layout)
+		return;
+
+	radv_free2(&device->alloc, pAllocator, set_layout);
+}
+
+/*
+ * Pipeline layouts.  These have nothing to do with the pipeline.  They are
+ * just muttiple descriptor set layouts pasted together
+ */
+
+VkResult radv_CreatePipelineLayout(
+	VkDevice                                    _device,
+	const VkPipelineLayoutCreateInfo*           pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkPipelineLayout*                           pPipelineLayout)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_pipeline_layout *layout;
+	struct mesa_sha1 *ctx;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO);
+
+	layout = radv_alloc2(&device->alloc, pAllocator, sizeof(*layout), 8,
+			     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (layout == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	layout->num_sets = pCreateInfo->setLayoutCount;
+
+	unsigned dynamic_offset_count = 0;
+
+
+	ctx = _mesa_sha1_init();
+	for (uint32_t set = 0; set < pCreateInfo->setLayoutCount; set++) {
+		RADV_FROM_HANDLE(radv_descriptor_set_layout, set_layout,
+				 pCreateInfo->pSetLayouts[set]);
+		layout->set[set].layout = set_layout;
+
+		layout->set[set].dynamic_offset_start = dynamic_offset_count;
+		for (uint32_t b = 0; b < set_layout->binding_count; b++) {
+			dynamic_offset_count += set_layout->binding[b].array_size * set_layout->binding[b].dynamic_offset_count;
+		}
+		_mesa_sha1_update(ctx, set_layout->binding,
+				  sizeof(set_layout->binding[0]) * set_layout->binding_count);
+	}
+
+	layout->dynamic_offset_count = dynamic_offset_count;
+	layout->push_constant_size = 0;
+	for (unsigned i = 0; i < pCreateInfo->pushConstantRangeCount; ++i) {
+		const VkPushConstantRange *range = pCreateInfo->pPushConstantRanges + i;
+		layout->push_constant_size = MAX2(layout->push_constant_size,
+						  range->offset + range->size);
+	}
+
+	layout->push_constant_size = align(layout->push_constant_size, 16);
+	_mesa_sha1_update(ctx, &layout->push_constant_size,
+			  sizeof(layout->push_constant_size));
+	_mesa_sha1_final(ctx, layout->sha1);
+	*pPipelineLayout = radv_pipeline_layout_to_handle(layout);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroyPipelineLayout(
+	VkDevice                                    _device,
+	VkPipelineLayout                            _pipelineLayout,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_pipeline_layout, pipeline_layout, _pipelineLayout);
+
+	if (!pipeline_layout)
+		return;
+	radv_free2(&device->alloc, pAllocator, pipeline_layout);
+}
+
+#define EMPTY 1
+
+static VkResult
+radv_descriptor_set_create(struct radv_device *device,
+			   struct radv_descriptor_pool *pool,
+			   struct radv_cmd_buffer *cmd_buffer,
+			   const struct radv_descriptor_set_layout *layout,
+			   struct radv_descriptor_set **out_set)
+{
+	struct radv_descriptor_set *set;
+	unsigned mem_size = sizeof(struct radv_descriptor_set) +
+		sizeof(struct radeon_winsys_bo *) * layout->buffer_count;
+	set = radv_alloc2(&device->alloc, NULL, mem_size, 8,
+			  VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+	if (!set)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	memset(set, 0, mem_size);
+
+	if (layout->dynamic_offset_count) {
+		unsigned size = sizeof(struct radv_descriptor_range) *
+		                layout->dynamic_offset_count;
+		set->dynamic_descriptors = radv_alloc2(&device->alloc, NULL, size, 8,
+			                               VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+		if (!set->dynamic_descriptors) {
+			radv_free2(&device->alloc, NULL, set);
+			return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		}
+	}
+
+	set->layout = layout;
+	if (layout->size) {
+		uint32_t layout_size = align_u32(layout->size, 32);
+		set->size = layout->size;
+		if (!cmd_buffer) {
+			if (pool->current_offset + layout_size <= pool->size) {
+				set->bo = pool->bo;
+				set->mapped_ptr = (uint32_t*)(pool->mapped_ptr + pool->current_offset);
+				set->va = device->ws->buffer_get_va(set->bo) + pool->current_offset;
+				pool->current_offset += layout_size;
+
+			} else {
+				int entry = pool->free_list, prev_entry = -1;
+				uint32_t offset;
+				while (entry >= 0) {
+					if (pool->free_nodes[entry].size >= layout_size) {
+						if (prev_entry >= 0)
+							pool->free_nodes[prev_entry].next = pool->free_nodes[entry].next;
+						else
+							pool->free_list = pool->free_nodes[entry].next;
+						break;
+					}
+					prev_entry = entry;
+					entry = pool->free_nodes[entry].next;
+				}
+
+				if (entry < 0) {
+					radv_free2(&device->alloc, NULL, set);
+					return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
+				}
+				offset = pool->free_nodes[entry].offset;
+				pool->free_nodes[entry].next = pool->full_list;
+				pool->full_list = entry;
+
+				set->bo = pool->bo;
+				set->mapped_ptr = (uint32_t*)(pool->mapped_ptr + offset);
+				set->va = device->ws->buffer_get_va(set->bo) + offset;
+			}
+		} else {
+			unsigned bo_offset;
+			if (!radv_cmd_buffer_upload_alloc(cmd_buffer, set->size, 32,
+							  &bo_offset,
+							  (void**)&set->mapped_ptr)) {
+				radv_free2(&device->alloc, NULL, set->dynamic_descriptors);
+				radv_free2(&device->alloc, NULL, set);
+				return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+			}
+
+			set->va = device->ws->buffer_get_va(cmd_buffer->upload.upload_bo);
+			set->va += bo_offset;
+		}
+	}
+
+	if (pool)
+		list_add(&set->descriptor_pool, &pool->descriptor_sets);
+	else
+		list_inithead(&set->descriptor_pool);
+
+	for (unsigned i = 0; i < layout->binding_count; ++i) {
+		if (!layout->binding[i].immutable_samplers)
+			continue;
+
+		unsigned offset = layout->binding[i].offset / 4;
+		if (layout->binding[i].type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
+			offset += 16;
+
+		for (unsigned j = 0; j < layout->binding[i].array_size; ++j) {
+			struct radv_sampler* sampler = layout->binding[i].immutable_samplers[j];
+
+			memcpy(set->mapped_ptr + offset, &sampler->state, 16);
+			offset += layout->binding[i].size / 4;
+		}
+
+	}
+	*out_set = set;
+	return VK_SUCCESS;
+}
+
+static void
+radv_descriptor_set_destroy(struct radv_device *device,
+			    struct radv_descriptor_pool *pool,
+			    struct radv_descriptor_set *set,
+			    bool free_bo)
+{
+	if (free_bo && set->size) {
+		assert(pool->full_list >= 0);
+		int next = pool->free_nodes[pool->full_list].next;
+		pool->free_nodes[pool->full_list].next = pool->free_list;
+		pool->free_nodes[pool->full_list].offset = (uint8_t*)set->mapped_ptr - pool->mapped_ptr;
+		pool->free_nodes[pool->full_list].size = align_u32(set->size, 32);
+		pool->free_list = pool->full_list;
+		pool->full_list = next;
+	}
+	if (set->dynamic_descriptors)
+		radv_free2(&device->alloc, NULL, set->dynamic_descriptors);
+	if (!list_empty(&set->descriptor_pool))
+		list_del(&set->descriptor_pool);
+	radv_free2(&device->alloc, NULL, set);
+}
+
+VkResult
+radv_temp_descriptor_set_create(struct radv_device *device,
+				struct radv_cmd_buffer *cmd_buffer,
+				VkDescriptorSetLayout _layout,
+				VkDescriptorSet *_set)
+{
+	RADV_FROM_HANDLE(radv_descriptor_set_layout, layout, _layout);
+	struct radv_descriptor_set *set;
+	VkResult ret;
+
+	ret = radv_descriptor_set_create(device, NULL, cmd_buffer, layout, &set);
+	*_set = radv_descriptor_set_to_handle(set);
+	return ret;
+}
+
+void
+radv_temp_descriptor_set_destroy(struct radv_device *device,
+				 VkDescriptorSet _set)
+{
+	RADV_FROM_HANDLE(radv_descriptor_set, set, _set);
+
+	radv_descriptor_set_destroy(device, NULL, set, false);
+}
+
+VkResult radv_CreateDescriptorPool(
+	VkDevice                                    _device,
+	const VkDescriptorPoolCreateInfo*           pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkDescriptorPool*                           pDescriptorPool)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_descriptor_pool *pool;
+	unsigned max_sets = pCreateInfo->maxSets * 2;
+	int size = sizeof(struct radv_descriptor_pool) +
+	           max_sets * sizeof(struct radv_descriptor_pool_free_node);
+	uint64_t bo_size = 0;
+	pool = radv_alloc2(&device->alloc, pAllocator, size, 8,
+			   VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (!pool)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	memset(pool, 0, sizeof(*pool));
+
+	pool->free_list = -1;
+	pool->full_list = 0;
+	pool->free_nodes[max_sets - 1].next = -1;
+	pool->max_sets = max_sets;
+
+	for (int i = 0; i  + 1 < max_sets; ++i)
+		pool->free_nodes[i].next = i + 1;
+
+	for (unsigned i = 0; i < pCreateInfo->poolSizeCount; ++i) {
+		switch(pCreateInfo->pPoolSizes[i].type) {
+		case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
+		case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
+			break;
+		case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+		case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+		case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
+		case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
+		case VK_DESCRIPTOR_TYPE_SAMPLER:
+			/* 32 as we may need to align for images */
+			bo_size += 32 * pCreateInfo->pPoolSizes[i].descriptorCount;
+			break;
+		case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
+		case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
+		case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
+			bo_size += 64 * pCreateInfo->pPoolSizes[i].descriptorCount;
+			break;
+		case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+			bo_size += 96 * pCreateInfo->pPoolSizes[i].descriptorCount;
+			break;
+		default:
+			unreachable("unknown descriptor type\n");
+			break;
+		}
+	}
+
+	if (bo_size) {
+		pool->bo = device->ws->buffer_create(device->ws, bo_size,
+							32, RADEON_DOMAIN_VRAM, 0);
+		pool->mapped_ptr = (uint8_t*)device->ws->buffer_map(pool->bo);
+	}
+	pool->size = bo_size;
+
+	list_inithead(&pool->descriptor_sets);
+	*pDescriptorPool = radv_descriptor_pool_to_handle(pool);
+	return VK_SUCCESS;
+}
+
+void radv_DestroyDescriptorPool(
+	VkDevice                                    _device,
+	VkDescriptorPool                            _pool,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_descriptor_pool, pool, _pool);
+
+	if (!pool)
+		return;
+
+	list_for_each_entry_safe(struct radv_descriptor_set, set,
+				 &pool->descriptor_sets, descriptor_pool) {
+		radv_descriptor_set_destroy(device, pool, set, false);
+	}
+
+	if (pool->bo)
+		device->ws->buffer_destroy(pool->bo);
+	radv_free2(&device->alloc, pAllocator, pool);
+}
+
+VkResult radv_ResetDescriptorPool(
+	VkDevice                                    _device,
+	VkDescriptorPool                            descriptorPool,
+	VkDescriptorPoolResetFlags                  flags)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_descriptor_pool, pool, descriptorPool);
+
+	list_for_each_entry_safe(struct radv_descriptor_set, set,
+				 &pool->descriptor_sets, descriptor_pool) {
+		radv_descriptor_set_destroy(device, pool, set, false);
+	}
+
+	pool->current_offset = 0;
+	pool->free_list = -1;
+	pool->full_list = 0;
+	pool->free_nodes[pool->max_sets - 1].next = -1;
+
+	for (int i = 0; i  + 1 < pool->max_sets; ++i)
+		pool->free_nodes[i].next = i + 1;
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_AllocateDescriptorSets(
+	VkDevice                                    _device,
+	const VkDescriptorSetAllocateInfo*          pAllocateInfo,
+	VkDescriptorSet*                            pDescriptorSets)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_descriptor_pool, pool, pAllocateInfo->descriptorPool);
+
+	VkResult result = VK_SUCCESS;
+	uint32_t i;
+	struct radv_descriptor_set *set;
+
+	/* allocate a set of buffers for each shader to contain descriptors */
+	for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) {
+		RADV_FROM_HANDLE(radv_descriptor_set_layout, layout,
+				 pAllocateInfo->pSetLayouts[i]);
+
+		result = radv_descriptor_set_create(device, pool, NULL, layout, &set);
+		if (result != VK_SUCCESS)
+			break;
+
+		pDescriptorSets[i] = radv_descriptor_set_to_handle(set);
+	}
+
+	if (result != VK_SUCCESS)
+		radv_FreeDescriptorSets(_device, pAllocateInfo->descriptorPool,
+					i, pDescriptorSets);
+	return result;
+}
+
+VkResult radv_FreeDescriptorSets(
+	VkDevice                                    _device,
+	VkDescriptorPool                            descriptorPool,
+	uint32_t                                    count,
+	const VkDescriptorSet*                      pDescriptorSets)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_descriptor_pool, pool, descriptorPool);
+
+	for (uint32_t i = 0; i < count; i++) {
+		RADV_FROM_HANDLE(radv_descriptor_set, set, pDescriptorSets[i]);
+
+		if (set)
+			radv_descriptor_set_destroy(device, pool, set, true);
+	}
+	return VK_SUCCESS;
+}
+
+static void write_texel_buffer_descriptor(struct radv_device *device,
+					  unsigned *dst,
+					  struct radeon_winsys_bo **buffer_list,
+					  const VkBufferView _buffer_view)
+{
+	RADV_FROM_HANDLE(radv_buffer_view, buffer_view, _buffer_view);
+
+	memcpy(dst, buffer_view->state, 4 * 4);
+	*buffer_list = buffer_view->bo;
+}
+
+static void write_buffer_descriptor(struct radv_device *device,
+                                    unsigned *dst,
+                                    struct radeon_winsys_bo **buffer_list,
+                                    const VkDescriptorBufferInfo *buffer_info)
+{
+	RADV_FROM_HANDLE(radv_buffer, buffer, buffer_info->buffer);
+	uint64_t va = device->ws->buffer_get_va(buffer->bo);
+	uint32_t range = buffer_info->range;
+
+	if (buffer_info->range == VK_WHOLE_SIZE)
+		range = buffer->size - buffer_info->offset;
+
+	va += buffer_info->offset + buffer->offset;
+	dst[0] = va;
+	dst[1] = S_008F04_BASE_ADDRESS_HI(va >> 32);
+	dst[2] = range;
+	dst[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+		S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+		S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+		S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+		S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+		S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
+
+	*buffer_list = buffer->bo;
+}
+
+static void write_dynamic_buffer_descriptor(struct radv_device *device,
+                                            struct radv_descriptor_range *range,
+                                            struct radeon_winsys_bo **buffer_list,
+                                            const VkDescriptorBufferInfo *buffer_info)
+{
+	RADV_FROM_HANDLE(radv_buffer, buffer, buffer_info->buffer);
+	uint64_t va = device->ws->buffer_get_va(buffer->bo);
+	unsigned size = buffer_info->range;
+
+	if (buffer_info->range == VK_WHOLE_SIZE)
+		size = buffer->size - buffer_info->offset;
+
+	va += buffer_info->offset + buffer->offset;
+	range->va = va;
+	range->size = size;
+
+	*buffer_list = buffer->bo;
+}
+
+static void
+write_image_descriptor(struct radv_device *device,
+		       unsigned *dst,
+		       struct radeon_winsys_bo **buffer_list,
+		       const VkDescriptorImageInfo *image_info)
+{
+	RADV_FROM_HANDLE(radv_image_view, iview, image_info->imageView);
+	memcpy(dst, iview->descriptor, 8 * 4);
+	memcpy(dst + 8, iview->fmask_descriptor, 8 * 4);
+	*buffer_list = iview->bo;
+}
+
+static void
+write_combined_image_sampler_descriptor(struct radv_device *device,
+					unsigned *dst,
+					struct radeon_winsys_bo **buffer_list,
+					const VkDescriptorImageInfo *image_info,
+					bool has_sampler)
+{
+	RADV_FROM_HANDLE(radv_sampler, sampler, image_info->sampler);
+
+	write_image_descriptor(device, dst, buffer_list, image_info);
+	/* copy over sampler state */
+	if (has_sampler)
+		memcpy(dst + 16, sampler->state, 16);
+}
+
+static void
+write_sampler_descriptor(struct radv_device *device,
+					unsigned *dst,
+					const VkDescriptorImageInfo *image_info)
+{
+	RADV_FROM_HANDLE(radv_sampler, sampler, image_info->sampler);
+
+	memcpy(dst, sampler->state, 16);
+}
+
+void radv_UpdateDescriptorSets(
+	VkDevice                                    _device,
+	uint32_t                                    descriptorWriteCount,
+	const VkWriteDescriptorSet*                 pDescriptorWrites,
+	uint32_t                                    descriptorCopyCount,
+	const VkCopyDescriptorSet*                  pDescriptorCopies)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	uint32_t i, j;
+	for (i = 0; i < descriptorWriteCount; i++) {
+		const VkWriteDescriptorSet *writeset = &pDescriptorWrites[i];
+		RADV_FROM_HANDLE(radv_descriptor_set, set, writeset->dstSet);
+		const struct radv_descriptor_set_binding_layout *binding_layout =
+			set->layout->binding + writeset->dstBinding;
+		uint32_t *ptr = set->mapped_ptr;
+		struct radeon_winsys_bo **buffer_list =  set->descriptors;
+
+		ptr += binding_layout->offset / 4;
+		ptr += binding_layout->size * writeset->dstArrayElement / 4;
+		buffer_list += binding_layout->buffer_offset;
+		buffer_list += binding_layout->buffer_count * writeset->dstArrayElement;
+		for (j = 0; j < writeset->descriptorCount; ++j) {
+			switch(writeset->descriptorType) {
+			case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
+			case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
+				unsigned idx = writeset->dstArrayElement + j;
+				idx += binding_layout->dynamic_offset_offset;
+				write_dynamic_buffer_descriptor(device, set->dynamic_descriptors + idx,
+								buffer_list, writeset->pBufferInfo + j);
+				break;
+			}
+			case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+			case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+				write_buffer_descriptor(device, ptr, buffer_list,
+							writeset->pBufferInfo + j);
+				break;
+			case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
+			case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
+				write_texel_buffer_descriptor(device, ptr, buffer_list,
+							      writeset->pTexelBufferView[j]);
+				break;
+			case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
+			case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
+			case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
+				write_image_descriptor(device, ptr, buffer_list,
+						       writeset->pImageInfo + j);
+				break;
+			case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+				write_combined_image_sampler_descriptor(device, ptr, buffer_list,
+									writeset->pImageInfo + j,
+									!binding_layout->immutable_samplers);
+				break;
+			case VK_DESCRIPTOR_TYPE_SAMPLER:
+				assert(!binding_layout->immutable_samplers);
+				write_sampler_descriptor(device, ptr,
+							 writeset->pImageInfo + j);
+				break;
+			default:
+				unreachable("unimplemented descriptor type");
+				break;
+			}
+			ptr += binding_layout->size / 4;
+			buffer_list += binding_layout->buffer_count;
+		}
+
+	}
+	if (descriptorCopyCount)
+		radv_finishme("copy descriptors");
+}
diff --git a/src/amd/vulkan/radv_descriptor_set.h b/src/amd/vulkan/radv_descriptor_set.h
new file mode 100644
index 0000000000..cad1ccf157
--- /dev/null
+++ b/src/amd/vulkan/radv_descriptor_set.h
@@ -0,0 +1,81 @@
+/*
+ * Copyright © 2016 Bas Nieuwenhuizen
+  *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#pragma once
+
+#include <vulkan/vulkan.h>
+
+#define MAX_SETS         8
+
+struct radv_descriptor_set_binding_layout {
+   VkDescriptorType type;
+
+   /* Number of array elements in this binding */
+   uint16_t array_size;
+
+   uint16_t offset;
+   uint16_t buffer_offset;
+   uint16_t dynamic_offset_offset;
+
+   /* redundant with the type, each for a single array element */
+   uint16_t size;
+   uint16_t buffer_count;
+   uint16_t dynamic_offset_count;
+
+   /* Immutable samplers (or NULL if no immutable samplers) */
+   struct radv_sampler **immutable_samplers;
+};
+
+struct radv_descriptor_set_layout {
+   /* Number of bindings in this descriptor set */
+   uint16_t binding_count;
+
+   /* Total size of the descriptor set with room for all array entries */
+   uint16_t size;
+
+   /* Shader stages affected by this descriptor set */
+   uint16_t shader_stages;
+   uint16_t dynamic_shader_stages;
+
+   /* Number of buffers in this descriptor set */
+   uint16_t buffer_count;
+
+   /* Number of dynamic offsets used by this descriptor set */
+   uint16_t dynamic_offset_count;
+
+   /* Bindings in this descriptor set */
+   struct radv_descriptor_set_binding_layout binding[0];
+};
+
+struct radv_pipeline_layout {
+   struct {
+      struct radv_descriptor_set_layout *layout;
+      uint32_t size;
+      uint32_t dynamic_offset_start;
+   } set[MAX_SETS];
+
+   uint32_t num_sets;
+   uint32_t push_constant_size;
+   uint32_t dynamic_offset_count;
+
+   unsigned char sha1[20];
+};
diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
new file mode 100644
index 0000000000..e9e00eb967
--- /dev/null
+++ b/src/amd/vulkan/radv_device.c
@@ -0,0 +1,1782 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based in part on anv driver which is:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <stdbool.h>
+#include <string.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include "radv_private.h"
+#include "util/strtod.h"
+
+#include <amdgpu.h>
+#include <amdgpu_drm.h>
+#include "amdgpu_id.h"
+#include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
+#include "ac_llvm_util.h"
+#include "vk_format.h"
+#include "sid.h"
+#include "radv_timestamp.h"
+#include "util/debug.h"
+struct radv_dispatch_table dtable;
+
+struct radv_fence {
+	struct radeon_winsys_fence *fence;
+	bool submitted;
+	bool signalled;
+};
+
+static VkResult
+radv_physical_device_init(struct radv_physical_device *device,
+			  struct radv_instance *instance,
+			  const char *path)
+{
+	VkResult result;
+	int fd;
+
+	fd = open(path, O_RDWR | O_CLOEXEC);
+	if (fd < 0)
+		return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
+				 "failed to open %s: %m", path);
+
+	device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+	device->instance = instance;
+	assert(strlen(path) < ARRAY_SIZE(device->path));
+	strncpy(device->path, path, ARRAY_SIZE(device->path));
+
+	device->ws = radv_amdgpu_winsys_create(fd);
+	if (!device->ws) {
+		result = VK_ERROR_INCOMPATIBLE_DRIVER;
+		goto fail;
+	}
+	device->ws->query_info(device->ws, &device->rad_info);
+	result = radv_init_wsi(device);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	fprintf(stderr, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
+	device->name = device->rad_info.name;
+	return VK_SUCCESS;
+
+fail:
+	close(fd);
+	return result;
+}
+
+static void
+radv_physical_device_finish(struct radv_physical_device *device)
+{
+	radv_finish_wsi(device);
+	device->ws->destroy(device->ws);
+}
+
+static const VkExtensionProperties global_extensions[] = {
+	{
+		.extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
+		.specVersion = 25,
+	},
+#ifdef VK_USE_PLATFORM_XCB_KHR
+	{
+		.extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
+		.specVersion = 5,
+	},
+#endif
+#ifdef VK_USE_PLATFORM_WAYLAND_KHR
+	{
+		.extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
+		.specVersion = 4,
+	},
+#endif
+};
+
+static const VkExtensionProperties device_extensions[] = {
+	{
+		.extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
+		.specVersion = 67,
+	},
+};
+
+static void *
+default_alloc_func(void *pUserData, size_t size, size_t align,
+                   VkSystemAllocationScope allocationScope)
+{
+	return malloc(size);
+}
+
+static void *
+default_realloc_func(void *pUserData, void *pOriginal, size_t size,
+                     size_t align, VkSystemAllocationScope allocationScope)
+{
+	return realloc(pOriginal, size);
+}
+
+static void
+default_free_func(void *pUserData, void *pMemory)
+{
+	free(pMemory);
+}
+
+static const VkAllocationCallbacks default_alloc = {
+	.pUserData = NULL,
+	.pfnAllocation = default_alloc_func,
+	.pfnReallocation = default_realloc_func,
+	.pfnFree = default_free_func,
+};
+
+VkResult radv_CreateInstance(
+	const VkInstanceCreateInfo*                 pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkInstance*                                 pInstance)
+{
+	struct radv_instance *instance;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
+
+	uint32_t client_version;
+	if (pCreateInfo->pApplicationInfo &&
+	    pCreateInfo->pApplicationInfo->apiVersion != 0) {
+		client_version = pCreateInfo->pApplicationInfo->apiVersion;
+	} else {
+		client_version = VK_MAKE_VERSION(1, 0, 0);
+	}
+
+	if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
+	    client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
+		return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
+				 "Client requested version %d.%d.%d",
+				 VK_VERSION_MAJOR(client_version),
+				 VK_VERSION_MINOR(client_version),
+				 VK_VERSION_PATCH(client_version));
+	}
+
+	for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
+		bool found = false;
+		for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
+			if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
+				   global_extensions[j].extensionName) == 0) {
+				found = true;
+				break;
+			}
+		}
+		if (!found)
+			return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
+	}
+
+	instance = radv_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
+			       VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
+	if (!instance)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+
+	if (pAllocator)
+		instance->alloc = *pAllocator;
+	else
+		instance->alloc = default_alloc;
+
+	instance->apiVersion = client_version;
+	instance->physicalDeviceCount = -1;
+
+	_mesa_locale_init();
+
+	VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
+
+	*pInstance = radv_instance_to_handle(instance);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroyInstance(
+	VkInstance                                  _instance,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_instance, instance, _instance);
+
+	if (instance->physicalDeviceCount > 0) {
+		/* We support at most one physical device. */
+		assert(instance->physicalDeviceCount == 1);
+		radv_physical_device_finish(&instance->physicalDevice);
+	}
+
+	VG(VALGRIND_DESTROY_MEMPOOL(instance));
+
+	_mesa_locale_fini();
+
+	radv_free(&instance->alloc, instance);
+}
+
+VkResult radv_EnumeratePhysicalDevices(
+	VkInstance                                  _instance,
+	uint32_t*                                   pPhysicalDeviceCount,
+	VkPhysicalDevice*                           pPhysicalDevices)
+{
+	RADV_FROM_HANDLE(radv_instance, instance, _instance);
+	VkResult result;
+
+	if (instance->physicalDeviceCount < 0) {
+		char path[20];
+		for (unsigned i = 0; i < 8; i++) {
+			snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
+			result = radv_physical_device_init(&instance->physicalDevice,
+							   instance, path);
+			if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
+				break;
+		}
+
+		if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
+			instance->physicalDeviceCount = 0;
+		} else if (result == VK_SUCCESS) {
+			instance->physicalDeviceCount = 1;
+		} else {
+			return result;
+		}
+	}
+
+	/* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
+	 * otherwise it's an inout parameter.
+	 *
+	 * The Vulkan spec (git aaed022) says:
+	 *
+	 *    pPhysicalDeviceCount is a pointer to an unsigned integer variable
+	 *    that is initialized with the number of devices the application is
+	 *    prepared to receive handles to. pname:pPhysicalDevices is pointer to
+	 *    an array of at least this many VkPhysicalDevice handles [...].
+	 *
+	 *    Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
+	 *    overwrites the contents of the variable pointed to by
+	 *    pPhysicalDeviceCount with the number of physical devices in in the
+	 *    instance; otherwise, vkEnumeratePhysicalDevices overwrites
+	 *    pPhysicalDeviceCount with the number of physical handles written to
+	 *    pPhysicalDevices.
+	 */
+	if (!pPhysicalDevices) {
+		*pPhysicalDeviceCount = instance->physicalDeviceCount;
+	} else if (*pPhysicalDeviceCount >= 1) {
+		pPhysicalDevices[0] = radv_physical_device_to_handle(&instance->physicalDevice);
+		*pPhysicalDeviceCount = 1;
+	} else {
+		*pPhysicalDeviceCount = 0;
+	}
+
+	return VK_SUCCESS;
+}
+
+void radv_GetPhysicalDeviceFeatures(
+	VkPhysicalDevice                            physicalDevice,
+	VkPhysicalDeviceFeatures*                   pFeatures)
+{
+	//   RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
+
+	memset(pFeatures, 0, sizeof(*pFeatures));
+
+	*pFeatures = (VkPhysicalDeviceFeatures) {
+		.robustBufferAccess                       = true,
+		.fullDrawIndexUint32                      = true,
+		.imageCubeArray                           = true,
+		.independentBlend                         = true,
+		.geometryShader                           = false,
+		.tessellationShader                       = false,
+		.sampleRateShading                        = false,
+		.dualSrcBlend                             = true,
+		.logicOp                                  = true,
+		.multiDrawIndirect                        = true,
+		.drawIndirectFirstInstance                = true,
+		.depthClamp                               = true,
+		.depthBiasClamp                           = true,
+		.fillModeNonSolid                         = true,
+		.depthBounds                              = true,
+		.wideLines                                = true,
+		.largePoints                              = true,
+		.alphaToOne                               = true,
+		.multiViewport                            = false,
+		.samplerAnisotropy                        = false, /* FINISHME */
+		.textureCompressionETC2                   = false,
+		.textureCompressionASTC_LDR               = false,
+		.textureCompressionBC                     = true,
+		.occlusionQueryPrecise                    = true,
+		.pipelineStatisticsQuery                  = false,
+		.vertexPipelineStoresAndAtomics           = true,
+		.fragmentStoresAndAtomics                 = true,
+		.shaderTessellationAndGeometryPointSize   = true,
+		.shaderImageGatherExtended                = false,
+		.shaderStorageImageExtendedFormats        = false,
+		.shaderStorageImageMultisample            = false,
+		.shaderUniformBufferArrayDynamicIndexing  = true,
+		.shaderSampledImageArrayDynamicIndexing   = true,
+		.shaderStorageBufferArrayDynamicIndexing  = true,
+		.shaderStorageImageArrayDynamicIndexing   = true,
+		.shaderStorageImageReadWithoutFormat      = false,
+		.shaderStorageImageWriteWithoutFormat     = true,
+		.shaderClipDistance                       = true,
+		.shaderCullDistance                       = true,
+		.shaderFloat64                            = false,
+		.shaderInt64                              = false,
+		.shaderInt16                              = false,
+		.alphaToOne                               = true,
+		.variableMultisampleRate                  = false,
+		.inheritedQueries                         = false,
+	};
+}
+
+void
+radv_device_get_cache_uuid(void *uuid)
+{
+	memset(uuid, 0, VK_UUID_SIZE);
+	snprintf(uuid, VK_UUID_SIZE, "radv-%s", RADV_TIMESTAMP);
+}
+
+void radv_GetPhysicalDeviceProperties(
+	VkPhysicalDevice                            physicalDevice,
+	VkPhysicalDeviceProperties*                 pProperties)
+{
+	RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
+	VkSampleCountFlags sample_counts = 0xf;
+	VkPhysicalDeviceLimits limits = {
+		.maxImageDimension1D                      = (1 << 14),
+		.maxImageDimension2D                      = (1 << 14),
+		.maxImageDimension3D                      = (1 << 11),
+		.maxImageDimensionCube                    = (1 << 14),
+		.maxImageArrayLayers                      = (1 << 11),
+		.maxTexelBufferElements                   = 128 * 1024 * 1024,
+		.maxUniformBufferRange                    = UINT32_MAX,
+		.maxStorageBufferRange                    = UINT32_MAX,
+		.maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
+		.maxMemoryAllocationCount                 = UINT32_MAX,
+		.maxSamplerAllocationCount                = 64 * 1024,
+		.bufferImageGranularity                   = 64, /* A cache line */
+		.sparseAddressSpaceSize                   = 0,
+		.maxBoundDescriptorSets                   = MAX_SETS,
+		.maxPerStageDescriptorSamplers            = 64,
+		.maxPerStageDescriptorUniformBuffers      = 64,
+		.maxPerStageDescriptorStorageBuffers      = 64,
+		.maxPerStageDescriptorSampledImages       = 64,
+		.maxPerStageDescriptorStorageImages       = 64,
+		.maxPerStageDescriptorInputAttachments    = 64,
+		.maxPerStageResources                     = 128,
+		.maxDescriptorSetSamplers                 = 256,
+		.maxDescriptorSetUniformBuffers           = 256,
+		.maxDescriptorSetUniformBuffersDynamic    = 256,
+		.maxDescriptorSetStorageBuffers           = 256,
+		.maxDescriptorSetStorageBuffersDynamic    = 256,
+		.maxDescriptorSetSampledImages            = 256,
+		.maxDescriptorSetStorageImages            = 256,
+		.maxDescriptorSetInputAttachments         = 256,
+		.maxVertexInputAttributes                 = 32,
+		.maxVertexInputBindings                   = 32,
+		.maxVertexInputAttributeOffset            = 2047,
+		.maxVertexInputBindingStride              = 2048,
+		.maxVertexOutputComponents                = 128,
+		.maxTessellationGenerationLevel           = 0,
+		.maxTessellationPatchSize                 = 0,
+		.maxTessellationControlPerVertexInputComponents = 0,
+		.maxTessellationControlPerVertexOutputComponents = 0,
+		.maxTessellationControlPerPatchOutputComponents = 0,
+		.maxTessellationControlTotalOutputComponents = 0,
+		.maxTessellationEvaluationInputComponents = 0,
+		.maxTessellationEvaluationOutputComponents = 0,
+		.maxGeometryShaderInvocations             = 32,
+		.maxGeometryInputComponents               = 64,
+		.maxGeometryOutputComponents              = 128,
+		.maxGeometryOutputVertices                = 256,
+		.maxGeometryTotalOutputComponents         = 1024,
+		.maxFragmentInputComponents               = 128,
+		.maxFragmentOutputAttachments             = 8,
+		.maxFragmentDualSrcAttachments            = 2,
+		.maxFragmentCombinedOutputResources       = 8,
+		.maxComputeSharedMemorySize               = 32768,
+		.maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
+		.maxComputeWorkGroupInvocations           = 16 * 1024,
+		.maxComputeWorkGroupSize = {
+			16 * 1024/*devinfo->max_cs_threads*/,
+			16 * 1024,
+			16 * 1024
+		},
+		.subPixelPrecisionBits                    = 4 /* FIXME */,
+		.subTexelPrecisionBits                    = 4 /* FIXME */,
+		.mipmapPrecisionBits                      = 4 /* FIXME */,
+		.maxDrawIndexedIndexValue                 = UINT32_MAX,
+		.maxDrawIndirectCount                     = UINT32_MAX,
+		.maxSamplerLodBias                        = 16,
+		.maxSamplerAnisotropy                     = 16,
+		.maxViewports                             = MAX_VIEWPORTS,
+		.maxViewportDimensions                    = { (1 << 14), (1 << 14) },
+		.viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
+		.viewportSubPixelBits                     = 13, /* We take a float? */
+		.minMemoryMapAlignment                    = 4096, /* A page */
+		.minTexelBufferOffsetAlignment            = 1,
+		.minUniformBufferOffsetAlignment          = 4,
+		.minStorageBufferOffsetAlignment          = 4,
+		.minTexelOffset                           = -8,
+		.maxTexelOffset                           = 7,
+		.minTexelGatherOffset                     = -8,
+		.maxTexelGatherOffset                     = 7,
+		.minInterpolationOffset                   = 0, /* FIXME */
+		.maxInterpolationOffset                   = 0, /* FIXME */
+		.subPixelInterpolationOffsetBits          = 0, /* FIXME */
+		.maxFramebufferWidth                      = (1 << 14),
+		.maxFramebufferHeight                     = (1 << 14),
+		.maxFramebufferLayers                     = (1 << 10),
+		.framebufferColorSampleCounts             = sample_counts,
+		.framebufferDepthSampleCounts             = sample_counts,
+		.framebufferStencilSampleCounts           = sample_counts,
+		.framebufferNoAttachmentsSampleCounts     = sample_counts,
+		.maxColorAttachments                      = MAX_RTS,
+		.sampledImageColorSampleCounts            = sample_counts,
+		.sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
+		.sampledImageDepthSampleCounts            = sample_counts,
+		.sampledImageStencilSampleCounts          = sample_counts,
+		.storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
+		.maxSampleMaskWords                       = 1,
+		.timestampComputeAndGraphics              = false,
+		.timestampPeriod                          = 100000.0 / pdevice->rad_info.clock_crystal_freq,
+		.maxClipDistances                         = 8,
+		.maxCullDistances                         = 8,
+		.maxCombinedClipAndCullDistances          = 8,
+		.discreteQueuePriorities                  = 1,
+		.pointSizeRange                           = { 0.125, 255.875 },
+		.lineWidthRange                           = { 0.0, 7.9921875 },
+		.pointSizeGranularity                     = (1.0 / 8.0),
+		.lineWidthGranularity                     = (1.0 / 128.0),
+		.strictLines                              = false, /* FINISHME */
+		.standardSampleLocations                  = true,
+		.optimalBufferCopyOffsetAlignment         = 128,
+		.optimalBufferCopyRowPitchAlignment       = 128,
+		.nonCoherentAtomSize                      = 64,
+	};
+
+	*pProperties = (VkPhysicalDeviceProperties) {
+		.apiVersion = VK_MAKE_VERSION(1, 0, 5),
+		.driverVersion = 1,
+		.vendorID = 0x1002,
+		.deviceID = pdevice->rad_info.pci_id,
+		.deviceType = VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU,
+		.limits = limits,
+		.sparseProperties = {0}, /* Broadwell doesn't do sparse. */
+	};
+
+	strcpy(pProperties->deviceName, pdevice->name);
+	radv_device_get_cache_uuid(pProperties->pipelineCacheUUID);
+}
+
+void radv_GetPhysicalDeviceQueueFamilyProperties(
+	VkPhysicalDevice                            physicalDevice,
+	uint32_t*                                   pCount,
+	VkQueueFamilyProperties*                    pQueueFamilyProperties)
+{
+	if (pQueueFamilyProperties == NULL) {
+		*pCount = 1;
+		return;
+	}
+	assert(*pCount >= 1);
+
+	*pQueueFamilyProperties = (VkQueueFamilyProperties) {
+		.queueFlags = VK_QUEUE_GRAPHICS_BIT |
+		VK_QUEUE_COMPUTE_BIT |
+		VK_QUEUE_TRANSFER_BIT,
+		.queueCount = 1,
+		.timestampValidBits = 64,
+		.minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
+	};
+}
+
+void radv_GetPhysicalDeviceMemoryProperties(
+	VkPhysicalDevice                            physicalDevice,
+	VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
+{
+	RADV_FROM_HANDLE(radv_physical_device, physical_device, physicalDevice);
+
+	pMemoryProperties->memoryTypeCount = 3;
+	pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
+		.propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
+		.heapIndex = 0,
+	};
+	pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
+		.propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
+		VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+		VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
+		.heapIndex = 0,
+	};
+	pMemoryProperties->memoryTypes[2] = (VkMemoryType) {
+		.propertyFlags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT|
+		VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
+		VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
+		.heapIndex = 1,
+	};
+
+	pMemoryProperties->memoryHeapCount = 2;
+	pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
+		.size = physical_device->rad_info.vram_size,
+		.flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
+	};
+	pMemoryProperties->memoryHeaps[1] = (VkMemoryHeap) {
+		.size = physical_device->rad_info.gart_size,
+		.flags = 0,
+	};
+}
+
+static VkResult
+radv_queue_init(struct radv_device *device, struct radv_queue *queue)
+{
+	queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+	queue->device = device;
+
+	return VK_SUCCESS;
+}
+
+static void
+radv_queue_finish(struct radv_queue *queue)
+{
+}
+
+VkResult radv_CreateDevice(
+	VkPhysicalDevice                            physicalDevice,
+	const VkDeviceCreateInfo*                   pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkDevice*                                   pDevice)
+{
+	RADV_FROM_HANDLE(radv_physical_device, physical_device, physicalDevice);
+	VkResult result;
+	struct radv_device *device;
+
+	for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
+		bool found = false;
+		for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
+			if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
+				   device_extensions[j].extensionName) == 0) {
+				found = true;
+				break;
+			}
+		}
+		if (!found)
+			return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
+	}
+
+	device = radv_alloc2(&physical_device->instance->alloc, pAllocator,
+			     sizeof(*device), 8,
+			     VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+	if (!device)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+	device->instance = physical_device->instance;
+
+	device->ws = physical_device->ws;
+	if (pAllocator)
+		device->alloc = *pAllocator;
+	else
+		device->alloc = physical_device->instance->alloc;
+
+	device->hw_ctx = device->ws->ctx_create(device->ws);
+	if (!device->hw_ctx) {
+		result = VK_ERROR_OUT_OF_HOST_MEMORY;
+		goto fail_free;
+	}
+
+	radv_queue_init(device, &device->queue);
+
+	result = radv_device_init_meta(device);
+	if (result != VK_SUCCESS) {
+		device->ws->ctx_destroy(device->hw_ctx);
+		goto fail_free;
+	}
+	device->allow_fast_clears = env_var_as_boolean("RADV_FAST_CLEARS", false);
+	device->allow_dcc = !env_var_as_boolean("RADV_DCC_DISABLE", false);
+
+	if (device->allow_fast_clears && device->allow_dcc)
+		radv_finishme("DCC fast clears have not been tested\n");
+
+	radv_device_init_msaa(device);
+	device->empty_cs = device->ws->cs_create(device->ws, RING_GFX);
+	radeon_emit(device->empty_cs, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
+	radeon_emit(device->empty_cs, CONTEXT_CONTROL_LOAD_ENABLE(1));
+	radeon_emit(device->empty_cs, CONTEXT_CONTROL_SHADOW_ENABLE(1));
+	device->ws->cs_finalize(device->empty_cs);
+	*pDevice = radv_device_to_handle(device);
+	return VK_SUCCESS;
+fail_free:
+	radv_free(&device->alloc, device);
+	return result;
+}
+
+void radv_DestroyDevice(
+	VkDevice                                    _device,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+
+	device->ws->ctx_destroy(device->hw_ctx);
+	radv_queue_finish(&device->queue);
+	radv_device_finish_meta(device);
+
+	radv_free(&device->alloc, device);
+}
+
+VkResult radv_EnumerateInstanceExtensionProperties(
+	const char*                                 pLayerName,
+	uint32_t*                                   pPropertyCount,
+	VkExtensionProperties*                      pProperties)
+{
+	unsigned i;
+	if (pProperties == NULL) {
+		*pPropertyCount = ARRAY_SIZE(global_extensions);
+		return VK_SUCCESS;
+	}
+
+	for (i = 0; i < *pPropertyCount; i++)
+		memcpy(&pProperties[i], &global_extensions[i], sizeof(VkExtensionProperties));
+
+	*pPropertyCount = i;
+	if (i < ARRAY_SIZE(global_extensions))
+		return VK_INCOMPLETE;
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_EnumerateDeviceExtensionProperties(
+	VkPhysicalDevice                            physicalDevice,
+	const char*                                 pLayerName,
+	uint32_t*                                   pPropertyCount,
+	VkExtensionProperties*                      pProperties)
+{
+	unsigned i;
+
+	if (pProperties == NULL) {
+		*pPropertyCount = ARRAY_SIZE(device_extensions);
+		return VK_SUCCESS;
+	}
+
+	for (i = 0; i < *pPropertyCount; i++)
+		memcpy(&pProperties[i], &device_extensions[i], sizeof(VkExtensionProperties));
+
+	*pPropertyCount = i;
+	if (i < ARRAY_SIZE(device_extensions))
+		return VK_INCOMPLETE;
+	return VK_SUCCESS;
+}
+
+VkResult radv_EnumerateInstanceLayerProperties(
+	uint32_t*                                   pPropertyCount,
+	VkLayerProperties*                          pProperties)
+{
+	if (pProperties == NULL) {
+		*pPropertyCount = 0;
+		return VK_SUCCESS;
+	}
+
+	/* None supported at this time */
+	return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
+}
+
+VkResult radv_EnumerateDeviceLayerProperties(
+	VkPhysicalDevice                            physicalDevice,
+	uint32_t*                                   pPropertyCount,
+	VkLayerProperties*                          pProperties)
+{
+	if (pProperties == NULL) {
+		*pPropertyCount = 0;
+		return VK_SUCCESS;
+	}
+
+	/* None supported at this time */
+	return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
+}
+
+void radv_GetDeviceQueue(
+	VkDevice                                    _device,
+	uint32_t                                    queueNodeIndex,
+	uint32_t                                    queueIndex,
+	VkQueue*                                    pQueue)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+
+	assert(queueIndex == 0);
+
+	*pQueue = radv_queue_to_handle(&device->queue);
+}
+
+VkResult radv_QueueSubmit(
+	VkQueue                                     _queue,
+	uint32_t                                    submitCount,
+	const VkSubmitInfo*                         pSubmits,
+	VkFence                                     _fence)
+{
+	RADV_FROM_HANDLE(radv_queue, queue, _queue);
+	RADV_FROM_HANDLE(radv_fence, fence, _fence);
+	struct radeon_winsys_fence *base_fence = fence ? fence->fence : NULL;
+	struct radeon_winsys_ctx *ctx = queue->device->hw_ctx;
+	int ret;
+
+	for (uint32_t i = 0; i < submitCount; i++) {
+		struct radeon_winsys_cs **cs_array;
+		bool can_patch = true;
+
+		if (!pSubmits[i].commandBufferCount)
+			continue;
+
+		cs_array = malloc(sizeof(struct radeon_winsys_cs *) *
+					        pSubmits[i].commandBufferCount);
+
+		for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
+			RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer,
+					 pSubmits[i].pCommandBuffers[j]);
+			assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
+
+			cs_array[j] = cmd_buffer->cs;
+			if ((cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT))
+				can_patch = false;
+		}
+		ret = queue->device->ws->cs_submit(ctx, cs_array,
+						   pSubmits[i].commandBufferCount,
+						   can_patch, base_fence);
+		if (ret)
+			radv_loge("failed to submit CS %d\n", i);
+		free(cs_array);
+	}
+
+	if (fence) {
+		if (!submitCount)
+			ret = queue->device->ws->cs_submit(ctx, &queue->device->empty_cs,
+							   1, false, base_fence);
+
+		fence->submitted = true;
+	}
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_QueueWaitIdle(
+	VkQueue                                     _queue)
+{
+	RADV_FROM_HANDLE(radv_queue, queue, _queue);
+
+	queue->device->ws->ctx_wait_idle(queue->device->hw_ctx);
+	return VK_SUCCESS;
+}
+
+VkResult radv_DeviceWaitIdle(
+	VkDevice                                    _device)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+
+	device->ws->ctx_wait_idle(device->hw_ctx);
+	return VK_SUCCESS;
+}
+
+PFN_vkVoidFunction radv_GetInstanceProcAddr(
+	VkInstance                                  instance,
+	const char*                                 pName)
+{
+	return radv_lookup_entrypoint(pName);
+}
+
+/* The loader wants us to expose a second GetInstanceProcAddr function
+ * to work around certain LD_PRELOAD issues seen in apps.
+ */
+PUBLIC
+VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
+	VkInstance                                  instance,
+	const char*                                 pName);
+
+PUBLIC
+VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
+	VkInstance                                  instance,
+	const char*                                 pName)
+{
+	return radv_GetInstanceProcAddr(instance, pName);
+}
+
+PFN_vkVoidFunction radv_GetDeviceProcAddr(
+	VkDevice                                    device,
+	const char*                                 pName)
+{
+	return radv_lookup_entrypoint(pName);
+}
+
+VkResult radv_AllocateMemory(
+	VkDevice                                    _device,
+	const VkMemoryAllocateInfo*                 pAllocateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkDeviceMemory*                             pMem)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_device_memory *mem;
+	VkResult result;
+	enum radeon_bo_domain domain;
+	uint32_t flags = 0;
+	assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
+
+	if (pAllocateInfo->allocationSize == 0) {
+		/* Apparently, this is allowed */
+		*pMem = VK_NULL_HANDLE;
+		return VK_SUCCESS;
+	}
+
+	mem = radv_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
+			  VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (mem == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
+	if (pAllocateInfo->memoryTypeIndex == 2)
+		domain = RADEON_DOMAIN_GTT;
+	else
+		domain = RADEON_DOMAIN_VRAM;
+
+	if (pAllocateInfo->memoryTypeIndex == 0)
+		flags |= RADEON_FLAG_NO_CPU_ACCESS;
+	else
+		flags |= RADEON_FLAG_CPU_ACCESS;
+	mem->bo = device->ws->buffer_create(device->ws, alloc_size, 32768,
+					       domain, flags);
+
+	if (!mem->bo) {
+		result = VK_ERROR_OUT_OF_DEVICE_MEMORY;
+		goto fail;
+	}
+	mem->type_index = pAllocateInfo->memoryTypeIndex;
+
+	*pMem = radv_device_memory_to_handle(mem);
+
+	return VK_SUCCESS;
+
+fail:
+	radv_free2(&device->alloc, pAllocator, mem);
+
+	return result;
+}
+
+void radv_FreeMemory(
+	VkDevice                                    _device,
+	VkDeviceMemory                              _mem,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_device_memory, mem, _mem);
+
+	if (mem == NULL)
+		return;
+
+	device->ws->buffer_destroy(mem->bo);
+	mem->bo = NULL;
+
+	radv_free2(&device->alloc, pAllocator, mem);
+}
+
+VkResult radv_MapMemory(
+	VkDevice                                    _device,
+	VkDeviceMemory                              _memory,
+	VkDeviceSize                                offset,
+	VkDeviceSize                                size,
+	VkMemoryMapFlags                            flags,
+	void**                                      ppData)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_device_memory, mem, _memory);
+
+	if (mem == NULL) {
+		*ppData = NULL;
+		return VK_SUCCESS;
+	}
+
+	*ppData = device->ws->buffer_map(mem->bo);
+	if (*ppData) {
+		*ppData += offset;
+		return VK_SUCCESS;
+	}
+
+	return VK_ERROR_MEMORY_MAP_FAILED;
+}
+
+void radv_UnmapMemory(
+	VkDevice                                    _device,
+	VkDeviceMemory                              _memory)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_device_memory, mem, _memory);
+
+	if (mem == NULL)
+		return;
+
+	device->ws->buffer_unmap(mem->bo);
+}
+
+VkResult radv_FlushMappedMemoryRanges(
+	VkDevice                                    _device,
+	uint32_t                                    memoryRangeCount,
+	const VkMappedMemoryRange*                  pMemoryRanges)
+{
+	return VK_SUCCESS;
+}
+
+VkResult radv_InvalidateMappedMemoryRanges(
+	VkDevice                                    _device,
+	uint32_t                                    memoryRangeCount,
+	const VkMappedMemoryRange*                  pMemoryRanges)
+{
+	return VK_SUCCESS;
+}
+
+void radv_GetBufferMemoryRequirements(
+	VkDevice                                    device,
+	VkBuffer                                    _buffer,
+	VkMemoryRequirements*                       pMemoryRequirements)
+{
+	RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
+
+	/* The Vulkan spec (git aaed022) says:
+	 *
+	 *    memoryTypeBits is a bitfield and contains one bit set for every
+	 *    supported memory type for the resource. The bit `1<<i` is set if and
+	 *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
+	 *    structure for the physical device is supported.
+	 *
+	 * We support exactly one memory type.
+	 */
+	pMemoryRequirements->memoryTypeBits = 0x7;
+
+	pMemoryRequirements->size = buffer->size;
+	pMemoryRequirements->alignment = 16;
+}
+
+void radv_GetImageMemoryRequirements(
+	VkDevice                                    device,
+	VkImage                                     _image,
+	VkMemoryRequirements*                       pMemoryRequirements)
+{
+	RADV_FROM_HANDLE(radv_image, image, _image);
+
+	/* The Vulkan spec (git aaed022) says:
+	 *
+	 *    memoryTypeBits is a bitfield and contains one bit set for every
+	 *    supported memory type for the resource. The bit `1<<i` is set if and
+	 *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
+	 *    structure for the physical device is supported.
+	 *
+	 * We support exactly one memory type.
+	 */
+	pMemoryRequirements->memoryTypeBits = 0x7;
+
+	pMemoryRequirements->size = image->size;
+	pMemoryRequirements->alignment = image->alignment;
+}
+
+void radv_GetImageSparseMemoryRequirements(
+	VkDevice                                    device,
+	VkImage                                     image,
+	uint32_t*                                   pSparseMemoryRequirementCount,
+	VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
+{
+	stub();
+}
+
+void radv_GetDeviceMemoryCommitment(
+	VkDevice                                    device,
+	VkDeviceMemory                              memory,
+	VkDeviceSize*                               pCommittedMemoryInBytes)
+{
+	*pCommittedMemoryInBytes = 0;
+}
+
+VkResult radv_BindBufferMemory(
+	VkDevice                                    device,
+	VkBuffer                                    _buffer,
+	VkDeviceMemory                              _memory,
+	VkDeviceSize                                memoryOffset)
+{
+	RADV_FROM_HANDLE(radv_device_memory, mem, _memory);
+	RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
+
+	if (mem) {
+		buffer->bo = mem->bo;
+		buffer->offset = memoryOffset;
+	} else {
+		buffer->bo = NULL;
+		buffer->offset = 0;
+	}
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_BindImageMemory(
+	VkDevice                                    device,
+	VkImage                                     _image,
+	VkDeviceMemory                              _memory,
+	VkDeviceSize                                memoryOffset)
+{
+	RADV_FROM_HANDLE(radv_device_memory, mem, _memory);
+	RADV_FROM_HANDLE(radv_image, image, _image);
+
+	if (mem) {
+		image->bo = mem->bo;
+		image->offset = memoryOffset;
+	} else {
+		image->bo = NULL;
+		image->offset = 0;
+	}
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_QueueBindSparse(
+	VkQueue                                     queue,
+	uint32_t                                    bindInfoCount,
+	const VkBindSparseInfo*                     pBindInfo,
+	VkFence                                     fence)
+{
+	stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
+}
+
+VkResult radv_CreateFence(
+	VkDevice                                    _device,
+	const VkFenceCreateInfo*                    pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkFence*                                    pFence)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_fence *fence = radv_alloc2(&device->alloc, pAllocator,
+					       sizeof(*fence), 8,
+					       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+	if (!fence)
+		return VK_ERROR_OUT_OF_HOST_MEMORY;
+
+	memset(fence, 0, sizeof(*fence));
+	fence->submitted = false;
+	fence->signalled = !!(pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT);
+	fence->fence = device->ws->create_fence();
+
+
+	*pFence = radv_fence_to_handle(fence);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroyFence(
+	VkDevice                                    _device,
+	VkFence                                     _fence,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_fence, fence, _fence);
+
+	if (!fence)
+		return;
+	device->ws->destroy_fence(fence->fence);
+	radv_free2(&device->alloc, pAllocator, fence);
+}
+
+static uint64_t radv_get_absolute_timeout(uint64_t timeout)
+{
+	uint64_t current_time;
+	struct timespec tv;
+
+	clock_gettime(CLOCK_MONOTONIC, &tv);
+	current_time = tv.tv_nsec + tv.tv_sec*1000000000ull;
+
+	timeout = MIN2(UINT64_MAX - current_time, timeout);
+
+	return current_time + timeout;
+}
+
+VkResult radv_WaitForFences(
+	VkDevice                                    _device,
+	uint32_t                                    fenceCount,
+	const VkFence*                              pFences,
+	VkBool32                                    waitAll,
+	uint64_t                                    timeout)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	timeout = radv_get_absolute_timeout(timeout);
+
+	if (!waitAll && fenceCount > 1) {
+		fprintf(stderr, "radv: WaitForFences without waitAll not implemented yet\n");
+	}
+
+	for (uint32_t i = 0; i < fenceCount; ++i) {
+		RADV_FROM_HANDLE(radv_fence, fence, pFences[i]);
+		bool expired = false;
+
+		if (!fence->submitted)
+			return VK_TIMEOUT;
+
+		if (fence->signalled)
+			continue;
+
+		expired = device->ws->fence_wait(device->ws, fence->fence, true, timeout);
+		if (!expired)
+			return VK_TIMEOUT;
+
+		fence->signalled = true;
+	}
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_ResetFences(VkDevice device,
+			  uint32_t fenceCount,
+			  const VkFence *pFences)
+{
+	for (unsigned i = 0; i < fenceCount; ++i) {
+		RADV_FROM_HANDLE(radv_fence, fence, pFences[i]);
+		fence->submitted = fence->signalled = false;
+	}
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_GetFenceStatus(VkDevice _device, VkFence _fence)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_fence, fence, _fence);
+
+	if (!fence->submitted)
+		return VK_NOT_READY;
+
+	if (!device->ws->fence_wait(device->ws, fence->fence, false, 0))
+		return VK_NOT_READY;
+
+	return VK_SUCCESS;
+}
+
+
+// Queue semaphore functions
+
+VkResult radv_CreateSemaphore(
+	VkDevice                                    device,
+	const VkSemaphoreCreateInfo*                pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkSemaphore*                                pSemaphore)
+{
+	/* The DRM execbuffer ioctl always execute in-oder, even between different
+	 * rings. As such, there's nothing to do for the user space semaphore.
+	 */
+
+	*pSemaphore = (VkSemaphore)1;
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroySemaphore(
+	VkDevice                                    device,
+	VkSemaphore                                 semaphore,
+	const VkAllocationCallbacks*                pAllocator)
+{
+}
+
+VkResult radv_CreateEvent(
+	VkDevice                                    _device,
+	const VkEventCreateInfo*                    pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkEvent*                                    pEvent)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_event *event = radv_alloc2(&device->alloc, pAllocator,
+					       sizeof(*event), 8,
+					       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+	if (!event)
+		return VK_ERROR_OUT_OF_HOST_MEMORY;
+
+	event->bo = device->ws->buffer_create(device->ws, 8, 8,
+					      RADEON_DOMAIN_GTT,
+					      RADEON_FLAG_CPU_ACCESS);
+	if (!event->bo) {
+		radv_free2(&device->alloc, pAllocator, event);
+		return VK_ERROR_OUT_OF_DEVICE_MEMORY;
+	}
+
+	event->map = (uint64_t*)device->ws->buffer_map(event->bo);
+
+	*pEvent = radv_event_to_handle(event);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroyEvent(
+	VkDevice                                    _device,
+	VkEvent                                     _event,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_event, event, _event);
+
+	if (!event)
+		return;
+	device->ws->buffer_destroy(event->bo);
+	radv_free2(&device->alloc, pAllocator, event);
+}
+
+VkResult radv_GetEventStatus(
+	VkDevice                                    _device,
+	VkEvent                                     _event)
+{
+	RADV_FROM_HANDLE(radv_event, event, _event);
+
+	if (*event->map == 1)
+		return VK_EVENT_SET;
+	return VK_EVENT_RESET;
+}
+
+VkResult radv_SetEvent(
+	VkDevice                                    _device,
+	VkEvent                                     _event)
+{
+	RADV_FROM_HANDLE(radv_event, event, _event);
+	*event->map = 1;
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_ResetEvent(
+    VkDevice                                    _device,
+    VkEvent                                     _event)
+{
+	RADV_FROM_HANDLE(radv_event, event, _event);
+	*event->map = 0;
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_CreateBuffer(
+	VkDevice                                    _device,
+	const VkBufferCreateInfo*                   pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkBuffer*                                   pBuffer)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_buffer *buffer;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
+
+	buffer = radv_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
+			     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (buffer == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	buffer->size = pCreateInfo->size;
+	buffer->usage = pCreateInfo->usage;
+	buffer->bo = NULL;
+	buffer->offset = 0;
+
+	*pBuffer = radv_buffer_to_handle(buffer);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroyBuffer(
+	VkDevice                                    _device,
+	VkBuffer                                    _buffer,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
+
+	if (!buffer)
+		return;
+
+	radv_free2(&device->alloc, pAllocator, buffer);
+}
+
+static inline unsigned
+si_tile_mode_index(const struct radv_image *image, unsigned level, bool stencil)
+{
+	if (stencil)
+		return image->surface.stencil_tiling_index[level];
+	else
+		return image->surface.tiling_index[level];
+}
+
+static void
+radv_initialise_color_surface(struct radv_device *device,
+			      struct radv_color_buffer_info *cb,
+			      struct radv_image_view *iview)
+{
+	const struct vk_format_description *desc;
+	unsigned ntype, format, swap, endian;
+	unsigned blend_clamp = 0, blend_bypass = 0;
+	unsigned pitch_tile_max, slice_tile_max, tile_mode_index;
+	uint64_t va;
+	const struct radeon_surf *surf = &iview->image->surface;
+	const struct radeon_surf_level *level_info = &surf->level[iview->base_mip];
+
+	desc = vk_format_description(iview->vk_format);
+
+	memset(cb, 0, sizeof(*cb));
+
+	va = device->ws->buffer_get_va(iview->bo) + iview->image->offset;
+	va += level_info->offset;
+	cb->cb_color_base = va >> 8;
+
+	/* CMASK variables */
+	va = device->ws->buffer_get_va(iview->bo) + iview->image->offset;
+	va += iview->image->cmask.offset;
+	cb->cb_color_cmask = va >> 8;
+	cb->cb_color_cmask_slice = iview->image->cmask.slice_tile_max;
+
+	va = device->ws->buffer_get_va(iview->bo) + iview->image->offset;
+	va += iview->image->dcc_offset;
+	cb->cb_dcc_base = va >> 8;
+
+	cb->cb_color_view = S_028C6C_SLICE_START(iview->base_layer) |
+		S_028C6C_SLICE_MAX(iview->base_layer + iview->extent.depth - 1);
+
+	cb->micro_tile_mode = iview->image->surface.micro_tile_mode;
+	pitch_tile_max = level_info->nblk_x / 8 - 1;
+	slice_tile_max = (level_info->nblk_x * level_info->nblk_y) / 64 - 1;
+	tile_mode_index = si_tile_mode_index(iview->image, iview->base_mip, false);
+
+	cb->cb_color_pitch = S_028C64_TILE_MAX(pitch_tile_max);
+	cb->cb_color_slice = S_028C68_TILE_MAX(slice_tile_max);
+
+	/* Intensity is implemented as Red, so treat it that way. */
+	cb->cb_color_attrib = S_028C74_FORCE_DST_ALPHA_1(desc->swizzle[3] == VK_SWIZZLE_1) |
+		S_028C74_TILE_MODE_INDEX(tile_mode_index);
+
+	if (iview->image->samples > 1) {
+		unsigned log_samples = util_logbase2(iview->image->samples);
+
+		cb->cb_color_attrib |= S_028C74_NUM_SAMPLES(log_samples) |
+			S_028C74_NUM_FRAGMENTS(log_samples);
+	}
+
+	if (iview->image->fmask.size) {
+		va = device->ws->buffer_get_va(iview->bo) + iview->image->offset + iview->image->fmask.offset;
+		if (device->instance->physicalDevice.rad_info.chip_class >= CIK)
+			cb->cb_color_pitch |= S_028C64_FMASK_TILE_MAX(iview->image->fmask.pitch_in_pixels / 8 - 1);
+		cb->cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(iview->image->fmask.tile_mode_index);
+		cb->cb_color_fmask = va >> 8;
+		cb->cb_color_fmask_slice = S_028C88_TILE_MAX(iview->image->fmask.slice_tile_max);
+	} else {
+		/* This must be set for fast clear to work without FMASK. */
+		if (device->instance->physicalDevice.rad_info.chip_class >= CIK)
+			cb->cb_color_pitch |= S_028C64_FMASK_TILE_MAX(pitch_tile_max);
+		cb->cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index);
+		cb->cb_color_fmask = cb->cb_color_base;
+		cb->cb_color_fmask_slice = S_028C88_TILE_MAX(slice_tile_max);
+	}
+
+	ntype = radv_translate_color_numformat(iview->vk_format,
+					       desc,
+					       vk_format_get_first_non_void_channel(iview->vk_format));
+	format = radv_translate_colorformat(iview->vk_format);
+	if (format == V_028C70_COLOR_INVALID || ntype == ~0u)
+		radv_finishme("Illegal color\n");
+	swap = radv_translate_colorswap(iview->vk_format, FALSE);
+	endian = radv_colorformat_endian_swap(format);
+
+	/* blend clamp should be set for all NORM/SRGB types */
+	if (ntype == V_028C70_NUMBER_UNORM ||
+	    ntype == V_028C70_NUMBER_SNORM ||
+	    ntype == V_028C70_NUMBER_SRGB)
+		blend_clamp = 1;
+
+	/* set blend bypass according to docs if SINT/UINT or
+	   8/24 COLOR variants */
+	if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT ||
+	    format == V_028C70_COLOR_8_24 || format == V_028C70_COLOR_24_8 ||
+	    format == V_028C70_COLOR_X24_8_32_FLOAT) {
+		blend_clamp = 0;
+		blend_bypass = 1;
+	}
+#if 0
+	if ((ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT) &&
+	    (format == V_028C70_COLOR_8 ||
+	     format == V_028C70_COLOR_8_8 ||
+	     format == V_028C70_COLOR_8_8_8_8))
+		->color_is_int8 = true;
+#endif
+	cb->cb_color_info = S_028C70_FORMAT(format) |
+		S_028C70_COMP_SWAP(swap) |
+		S_028C70_BLEND_CLAMP(blend_clamp) |
+		S_028C70_BLEND_BYPASS(blend_bypass) |
+		S_028C70_SIMPLE_FLOAT(1) |
+		S_028C70_ROUND_MODE(ntype != V_028C70_NUMBER_UNORM &&
+				    ntype != V_028C70_NUMBER_SNORM &&
+				    ntype != V_028C70_NUMBER_SRGB &&
+				    format != V_028C70_COLOR_8_24 &&
+				    format != V_028C70_COLOR_24_8) |
+		S_028C70_NUMBER_TYPE(ntype) |
+		S_028C70_ENDIAN(endian);
+	if (iview->image->samples > 1)
+		if (iview->image->fmask.size)
+			cb->cb_color_info |= S_028C70_COMPRESSION(1);
+
+	if (iview->image->cmask.size && device->allow_fast_clears)
+		cb->cb_color_info |= S_028C70_FAST_CLEAR(1);
+
+	if (iview->image->surface.dcc_size && level_info->dcc_enabled)
+		cb->cb_color_info |= S_028C70_DCC_ENABLE(1);
+
+	if (device->instance->physicalDevice.rad_info.chip_class >= VI) {
+		unsigned max_uncompressed_block_size = 2;
+		if (iview->image->samples > 1) {
+			if (iview->image->surface.bpe == 1)
+				max_uncompressed_block_size = 0;
+			else if (iview->image->surface.bpe == 2)
+				max_uncompressed_block_size = 1;
+		}
+
+		cb->cb_dcc_control = S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size) |
+			S_028C78_INDEPENDENT_64B_BLOCKS(1);
+	}
+
+	/* This must be set for fast clear to work without FMASK. */
+	if (!iview->image->fmask.size &&
+	    device->instance->physicalDevice.rad_info.chip_class == SI) {
+		unsigned bankh = util_logbase2(iview->image->surface.bankh);
+		cb->cb_color_attrib |= S_028C74_FMASK_BANK_HEIGHT(bankh);
+	}
+}
+
+static void
+radv_initialise_ds_surface(struct radv_device *device,
+			   struct radv_ds_buffer_info *ds,
+			   struct radv_image_view *iview)
+{
+	unsigned level = iview->base_mip;
+	unsigned format;
+	uint64_t va, s_offs, z_offs;
+	const struct radeon_surf_level *level_info = &iview->image->surface.level[level];
+	memset(ds, 0, sizeof(*ds));
+	switch (iview->vk_format) {
+	case VK_FORMAT_D24_UNORM_S8_UINT:
+	case VK_FORMAT_X8_D24_UNORM_PACK32:
+		ds->pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
+		ds->offset_scale = 2.0f;
+		break;
+	case VK_FORMAT_D16_UNORM:
+	case VK_FORMAT_D16_UNORM_S8_UINT:
+		ds->pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
+		ds->offset_scale = 4.0f;
+		break;
+	case VK_FORMAT_D32_SFLOAT:
+	case VK_FORMAT_D32_SFLOAT_S8_UINT:
+		ds->pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
+			S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
+		ds->offset_scale = 1.0f;
+		break;
+	default:
+		break;
+	}
+
+	format = radv_translate_dbformat(iview->vk_format);
+	if (format == V_028040_Z_INVALID) {
+		fprintf(stderr, "Invalid DB format: %d, disabling DB.\n", iview->vk_format);
+	}
+
+	va = device->ws->buffer_get_va(iview->bo) + iview->image->offset;
+	s_offs = z_offs = va;
+	z_offs += iview->image->surface.level[level].offset;
+	s_offs += iview->image->surface.stencil_level[level].offset;
+
+	ds->db_depth_view = S_028008_SLICE_START(iview->base_layer) |
+		S_028008_SLICE_MAX(iview->base_layer + iview->extent.depth - 1);
+	ds->db_depth_info = S_02803C_ADDR5_SWIZZLE_MASK(1);
+	ds->db_z_info = S_028040_FORMAT(format) | S_028040_ZRANGE_PRECISION(1);
+
+	if (iview->image->samples > 1)
+		ds->db_z_info |= S_028040_NUM_SAMPLES(util_logbase2(iview->image->samples));
+
+	if (iview->image->surface.flags & RADEON_SURF_SBUFFER)
+		ds->db_stencil_info = S_028044_FORMAT(V_028044_STENCIL_8);
+	else
+		ds->db_stencil_info = S_028044_FORMAT(V_028044_STENCIL_INVALID);
+
+	if (device->instance->physicalDevice.rad_info.chip_class >= CIK) {
+		struct radeon_info *info = &device->instance->physicalDevice.rad_info;
+		unsigned tiling_index = iview->image->surface.tiling_index[level];
+		unsigned stencil_index = iview->image->surface.stencil_tiling_index[level];
+		unsigned macro_index = iview->image->surface.macro_tile_index;
+		unsigned tile_mode = info->si_tile_mode_array[tiling_index];
+		unsigned stencil_tile_mode = info->si_tile_mode_array[stencil_index];
+		unsigned macro_mode = info->cik_macrotile_mode_array[macro_index];
+
+		ds->db_depth_info |=
+			S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode)) |
+			S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode)) |
+			S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode)) |
+			S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode)) |
+			S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode)) |
+			S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode));
+		ds->db_z_info |= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode));
+		ds->db_stencil_info |= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode));
+	} else {
+		unsigned tile_mode_index = si_tile_mode_index(iview->image, level, false);
+		ds->db_z_info |= S_028040_TILE_MODE_INDEX(tile_mode_index);
+		tile_mode_index = si_tile_mode_index(iview->image, level, true);
+		ds->db_stencil_info |= S_028044_TILE_MODE_INDEX(tile_mode_index);
+	}
+
+	if (iview->image->htile.size && !level) {
+		ds->db_z_info |= S_028040_TILE_SURFACE_ENABLE(1) |
+			S_028040_ALLOW_EXPCLEAR(1);
+
+		if (iview->image->surface.flags & RADEON_SURF_SBUFFER) {
+			/* Workaround: For a not yet understood reason, the
+			 * combination of MSAA, fast stencil clear and stencil
+			 * decompress messes with subsequent stencil buffer
+			 * uses. Problem was reproduced on Verde, Bonaire,
+			 * Tonga, and Carrizo.
+			 *
+			 * Disabling EXPCLEAR works around the problem.
+			 *
+			 * Check piglit's arb_texture_multisample-stencil-clear
+			 * test if you want to try changing this.
+			 */
+			if (iview->image->samples <= 1)
+				ds->db_stencil_info |= S_028044_ALLOW_EXPCLEAR(1);
+		} else
+			/* Use all of the htile_buffer for depth if there's no stencil. */
+			ds->db_stencil_info |= S_028044_TILE_STENCIL_DISABLE(1);
+
+		va = device->ws->buffer_get_va(iview->bo) + iview->image->offset +
+		     iview->image->htile.offset;
+		ds->db_htile_data_base = va >> 8;
+		ds->db_htile_surface = S_028ABC_FULL_CACHE(1);
+	} else {
+		ds->db_htile_data_base = 0;
+		ds->db_htile_surface = 0;
+	}
+
+	ds->db_z_read_base = ds->db_z_write_base = z_offs >> 8;
+	ds->db_stencil_read_base = ds->db_stencil_write_base = s_offs >> 8;
+
+	ds->db_depth_size = S_028058_PITCH_TILE_MAX((level_info->nblk_x / 8) - 1) |
+		S_028058_HEIGHT_TILE_MAX((level_info->nblk_y / 8) - 1);
+	ds->db_depth_slice = S_02805C_SLICE_TILE_MAX((level_info->nblk_x * level_info->nblk_y) / 64 - 1);
+}
+
+VkResult radv_CreateFramebuffer(
+	VkDevice                                    _device,
+	const VkFramebufferCreateInfo*              pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkFramebuffer*                              pFramebuffer)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_framebuffer *framebuffer;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
+
+	size_t size = sizeof(*framebuffer) +
+		sizeof(struct radv_attachment_info) * pCreateInfo->attachmentCount;
+	framebuffer = radv_alloc2(&device->alloc, pAllocator, size, 8,
+				  VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (framebuffer == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	framebuffer->attachment_count = pCreateInfo->attachmentCount;
+	for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
+		VkImageView _iview = pCreateInfo->pAttachments[i];
+		struct radv_image_view *iview = radv_image_view_from_handle(_iview);
+		framebuffer->attachments[i].attachment = iview;
+		if (iview->aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT) {
+			radv_initialise_color_surface(device, &framebuffer->attachments[i].cb, iview);
+		} else if (iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)) {
+			radv_initialise_ds_surface(device, &framebuffer->attachments[i].ds, iview);
+		}
+	}
+
+	framebuffer->width = pCreateInfo->width;
+	framebuffer->height = pCreateInfo->height;
+	framebuffer->layers = pCreateInfo->layers;
+
+	*pFramebuffer = radv_framebuffer_to_handle(framebuffer);
+	return VK_SUCCESS;
+}
+
+void radv_DestroyFramebuffer(
+	VkDevice                                    _device,
+	VkFramebuffer                               _fb,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_framebuffer, fb, _fb);
+
+	if (!fb)
+		return;
+	radv_free2(&device->alloc, pAllocator, fb);
+}
+
+static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode)
+{
+	switch (address_mode) {
+	case VK_SAMPLER_ADDRESS_MODE_REPEAT:
+		return V_008F30_SQ_TEX_WRAP;
+	case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT:
+		return V_008F30_SQ_TEX_MIRROR;
+	case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE:
+		return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL;
+	case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER:
+		return V_008F30_SQ_TEX_CLAMP_BORDER;
+	case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE:
+		return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL;
+	default:
+		unreachable("illegal tex wrap mode");
+		break;
+	}
+}
+
+static unsigned
+radv_tex_compare(VkCompareOp op)
+{
+	switch (op) {
+	case VK_COMPARE_OP_NEVER:
+		return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
+	case VK_COMPARE_OP_LESS:
+		return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS;
+	case VK_COMPARE_OP_EQUAL:
+		return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL;
+	case VK_COMPARE_OP_LESS_OR_EQUAL:
+		return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL;
+	case VK_COMPARE_OP_GREATER:
+		return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER;
+	case VK_COMPARE_OP_NOT_EQUAL:
+		return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL;
+	case VK_COMPARE_OP_GREATER_OR_EQUAL:
+		return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL;
+	case VK_COMPARE_OP_ALWAYS:
+		return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS;
+	default:
+		unreachable("illegal compare mode");
+		break;
+	}
+}
+
+static unsigned
+radv_tex_filter(VkFilter filter, unsigned max_ansio)
+{
+	switch (filter) {
+	case VK_FILTER_NEAREST:
+		return (max_ansio > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT :
+			V_008F38_SQ_TEX_XY_FILTER_POINT);
+	case VK_FILTER_LINEAR:
+		return (max_ansio > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR :
+			V_008F38_SQ_TEX_XY_FILTER_BILINEAR);
+	case VK_FILTER_CUBIC_IMG:
+	default:
+		fprintf(stderr, "illegal texture filter");
+		return 0;
+	}
+}
+
+static unsigned
+radv_tex_mipfilter(VkSamplerMipmapMode mode)
+{
+	switch (mode) {
+	case VK_SAMPLER_MIPMAP_MODE_NEAREST:
+		return V_008F38_SQ_TEX_Z_FILTER_POINT;
+	case VK_SAMPLER_MIPMAP_MODE_LINEAR:
+		return V_008F38_SQ_TEX_Z_FILTER_LINEAR;
+	default:
+		return V_008F38_SQ_TEX_Z_FILTER_NONE;
+	}
+}
+
+static unsigned
+radv_tex_bordercolor(VkBorderColor bcolor)
+{
+	switch (bcolor) {
+	case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK:
+	case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK:
+		return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK;
+	case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK:
+	case VK_BORDER_COLOR_INT_OPAQUE_BLACK:
+		return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK;
+	case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE:
+	case VK_BORDER_COLOR_INT_OPAQUE_WHITE:
+		return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE;
+	default:
+		break;
+	}
+	return 0;
+}
+
+static void
+radv_init_sampler(struct radv_device *device,
+		  struct radv_sampler *sampler,
+		  const VkSamplerCreateInfo *pCreateInfo)
+{
+	uint32_t max_aniso = 0;
+	uint32_t max_aniso_ratio = 0;//TODO
+	bool is_vi;
+	is_vi = (device->instance->physicalDevice.rad_info.chip_class >= VI);
+
+	sampler->state[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo->addressModeU)) |
+			     S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo->addressModeV)) |
+			     S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo->addressModeW)) |
+			     S_008F30_MAX_ANISO_RATIO(max_aniso_ratio) |
+			     S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo->compareOp)) |
+			     S_008F30_FORCE_UNNORMALIZED(pCreateInfo->unnormalizedCoordinates ? 1 : 0) |
+			     S_008F30_DISABLE_CUBE_WRAP(0) |
+			     S_008F30_COMPAT_MODE(is_vi));
+	sampler->state[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo->minLod, 0, 15), 8)) |
+			     S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo->maxLod, 0, 15), 8)));
+	sampler->state[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo->mipLodBias, -16, 16), 8)) |
+			     S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo->magFilter, max_aniso)) |
+			     S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo->minFilter, max_aniso)) |
+			     S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo->mipmapMode)) |
+			     S_008F38_MIP_POINT_PRECLAMP(1) |
+			     S_008F38_DISABLE_LSB_CEIL(1) |
+			     S_008F38_FILTER_PREC_FIX(1) |
+			     S_008F38_ANISO_OVERRIDE(is_vi));
+	sampler->state[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
+			     S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo->borderColor)));
+}
+
+VkResult radv_CreateSampler(
+	VkDevice                                    _device,
+	const VkSamplerCreateInfo*                  pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkSampler*                                  pSampler)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_sampler *sampler;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
+
+	sampler = radv_alloc2(&device->alloc, pAllocator, sizeof(*sampler), 8,
+			      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (!sampler)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	radv_init_sampler(device, sampler, pCreateInfo);
+	*pSampler = radv_sampler_to_handle(sampler);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroySampler(
+	VkDevice                                    _device,
+	VkSampler                                   _sampler,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_sampler, sampler, _sampler);
+
+	if (!sampler)
+		return;
+	radv_free2(&device->alloc, pAllocator, sampler);
+}
diff --git a/src/amd/vulkan/radv_device_info.h b/src/amd/vulkan/radv_device_info.h
new file mode 100644
index 0000000000..c38bd2a99e
--- /dev/null
+++ b/src/amd/vulkan/radv_device_info.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright © 2016 Red Hat.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#pragma once
+
+#include <amdgpu.h>
+
+#include "radv_radeon_winsys.h"
+struct radv_device_info {
+	uint32_t pci_id;
+	enum chip_class chip_class;
+};
diff --git a/src/amd/vulkan/radv_entrypoints_gen.py b/src/amd/vulkan/radv_entrypoints_gen.py
new file mode 100644
index 0000000000..e8ef8a4e9e
--- /dev/null
+++ b/src/amd/vulkan/radv_entrypoints_gen.py
@@ -0,0 +1,351 @@
+# coding=utf-8
+#
+# Copyright © 2015 Intel Corporation
+#
+# Permission is hereby granted, free of charge, to any person obtaining a
+# copy of this software and associated documentation files (the "Software"),
+# to deal in the Software without restriction, including without limitation
+# the rights to use, copy, modify, merge, publish, distribute, sublicense,
+# and/or sell copies of the Software, and to permit persons to whom the
+# Software is furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice (including the next
+# paragraph) shall be included in all copies or substantial portions of the
+# Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+# IN THE SOFTWARE.
+#
+
+import fileinput, re, sys
+
+# Each function typedef in the vulkan.h header is all on one line and matches
+# this regepx. We hope that won't change.
+
+p = re.compile('typedef ([^ ]*) *\((?:VKAPI_PTR)? *\*PFN_vk([^(]*)\)(.*);')
+
+entrypoints = []
+
+# We generate a static hash table for entry point lookup
+# (vkGetProcAddress). We use a linear congruential generator for our hash
+# function and a power-of-two size table. The prime numbers are determined
+# experimentally.
+
+none = 0xffff
+hash_size = 256
+u32_mask = 2**32 - 1
+hash_mask = hash_size - 1
+
+prime_factor = 5024183
+prime_step = 19
+
+def hash(name):
+    h = 0;
+    for c in name:
+        h = (h * prime_factor + ord(c)) & u32_mask
+
+    return h
+
+def get_platform_guard_macro(name):
+    if "Xlib" in name:
+        return "VK_USE_PLATFORM_XLIB_KHR"
+    elif "Xcb" in name:
+        return "VK_USE_PLATFORM_XCB_KHR"
+    elif "Wayland" in name:
+        return "VK_USE_PLATFORM_WAYLAND_KHR"
+    elif "Mir" in name:
+        return "VK_USE_PLATFORM_MIR_KHR"
+    elif "Android" in name:
+        return "VK_USE_PLATFORM_ANDROID_KHR"
+    elif "Win32" in name:
+        return "VK_USE_PLATFORM_WIN32_KHR"
+    else:
+        return None
+
+def print_guard_start(name):
+    guard = get_platform_guard_macro(name)
+    if guard is not None:
+        print "#ifdef {0}".format(guard)
+
+def print_guard_end(name):
+    guard = get_platform_guard_macro(name)
+    if guard is not None:
+        print "#endif // {0}".format(guard)
+
+opt_header = False
+opt_code = False
+
+if (sys.argv[1] == "header"):
+    opt_header = True
+    sys.argv.pop()
+elif (sys.argv[1] == "code"):
+    opt_code = True
+    sys.argv.pop()
+
+# Parse the entry points in the header
+
+i = 0
+for line in fileinput.input():
+    m  = p.match(line)
+    if (m):
+        if m.group(2) == 'VoidFunction':
+            continue
+        fullname = "vk" + m.group(2)
+        h = hash(fullname)
+        entrypoints.append((m.group(1), m.group(2), m.group(3), i, h))
+        i = i + 1
+
+# For outputting entrypoints.h we generate a radv_EntryPoint() prototype
+# per entry point.
+
+if opt_header:
+    print "/* This file generated from vk_gen.py, don't edit directly. */\n"
+
+    print "struct radv_dispatch_table {"
+    print "   union {"
+    print "      void *entrypoints[%d];" % len(entrypoints)
+    print "      struct {"
+
+    for type, name, args, num, h in entrypoints:
+        guard = get_platform_guard_macro(name)
+        if guard is not None:
+            print "#ifdef {0}".format(guard)
+            print "         PFN_vk{0} {0};".format(name)
+            print "#else"
+            print "         void *{0};".format(name)
+            print "#endif"
+        else:
+            print "         PFN_vk{0} {0};".format(name)
+    print "      };\n"
+    print "   };\n"
+    print "};\n"
+
+    print "void radv_set_dispatch_devinfo(const struct radv_device_info *info);\n"
+
+    for type, name, args, num, h in entrypoints:
+        print_guard_start(name)
+        print "%s radv_%s%s;" % (type, name, args)
+        print "%s vi_%s%s;" % (type, name, args)
+        print "%s cik_%s%s;" % (type, name, args)
+        print "%s si_%s%s;" % (type, name, args)
+        print "%s radv_validate_%s%s;" % (type, name, args)
+        print_guard_end(name)
+    exit()
+
+
+
+print """/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+/* DO NOT EDIT! This is a generated file. */
+
+#include "radv_private.h"
+
+struct radv_entrypoint {
+   uint32_t name;
+   uint32_t hash;
+};
+
+/* We use a big string constant to avoid lots of reloctions from the entry
+ * point table to lots of little strings. The entries in the entry point table
+ * store the index into this big string.
+ */
+
+static const char strings[] ="""
+
+offsets = []
+i = 0;
+for type, name, args, num, h in entrypoints:
+    print "   \"vk%s\\0\"" % name
+    offsets.append(i)
+    i += 2 + len(name) + 1
+print "   ;"
+
+# Now generate the table of all entry points and their validation functions
+
+print "\nstatic const struct radv_entrypoint entrypoints[] = {"
+for type, name, args, num, h in entrypoints:
+    print "   { %5d, 0x%08x }," % (offsets[num], h)
+print "};\n"
+
+print """
+
+/* Weak aliases for all potential implementations. These will resolve to
+ * NULL if they're not defined, which lets the resolve_entrypoint() function
+ * either pick the correct entry point.
+ */
+"""
+
+for layer in [ "radv", "validate", "si", "cik", "vi" ]:
+    for type, name, args, num, h in entrypoints:
+        print_guard_start(name)
+        print "%s %s_%s%s __attribute__ ((weak));" % (type, layer, name, args)
+        print_guard_end(name)
+    print "\nconst struct radv_dispatch_table %s_layer = {" % layer
+    for type, name, args, num, h in entrypoints:
+        print_guard_start(name)
+        print "   .%s = %s_%s," % (name, layer, name)
+        print_guard_end(name)
+    print "};\n"
+
+print """
+#ifdef DEBUG
+static bool enable_validate = true;
+#else
+static bool enable_validate = false;
+#endif
+
+/* We can't use symbols that need resolving (like, oh, getenv) in the resolve
+ * function. This means that we have to determine whether or not to use the
+ * validation layer sometime before that. The constructor function attribute asks
+ * the dynamic linker to invoke determine_validate() at dlopen() time which
+ * works.
+ */
+static void __attribute__ ((constructor))
+determine_validate(void)
+{
+   const char *s = getenv("ANV_VALIDATE");
+
+   if (s)
+      enable_validate = atoi(s);
+}
+
+static const struct radv_device_info *dispatch_devinfo;
+
+void
+radv_set_dispatch_devinfo(const struct radv_device_info *devinfo)
+{
+   dispatch_devinfo = devinfo;
+}
+
+void * __attribute__ ((noinline))
+radv_resolve_entrypoint(uint32_t index)
+{
+   if (enable_validate && validate_layer.entrypoints[index])
+      return validate_layer.entrypoints[index];
+
+   if (dispatch_devinfo == NULL) {
+      return radv_layer.entrypoints[index];
+   }
+
+   switch (dispatch_devinfo->chip_class) {
+   case VI:
+      if (vi_layer.entrypoints[index])
+         return vi_layer.entrypoints[index];
+      /* fall through */
+   case CIK:
+      if (cik_layer.entrypoints[index])
+         return cik_layer.entrypoints[index];
+      /* fall through */
+   case SI:
+      if (si_layer.entrypoints[index])
+         return si_layer.entrypoints[index];
+      /* fall through */
+   case 0:
+      return radv_layer.entrypoints[index];
+   default:
+      unreachable("unsupported gen\\n");
+   }
+}
+"""
+
+# Now generate the hash table used for entry point look up.  This is a
+# uint16_t table of entry point indices. We use 0xffff to indicate an entry
+# in the hash table is empty.
+
+map = [none for f in xrange(hash_size)]
+collisions = [0 for f in xrange(10)]
+for type, name, args, num, h in entrypoints:
+    level = 0
+    while map[h & hash_mask] != none:
+        h = h + prime_step
+        level = level + 1
+    if level > 9:
+        collisions[9] += 1
+    else:
+        collisions[level] += 1
+    map[h & hash_mask] = num
+
+print "/* Hash table stats:"
+print " * size %d entries" % hash_size
+print " * collisions  entries"
+for i in xrange(10):
+    if (i == 9):
+        plus = "+"
+    else:
+        plus = " "
+
+    print " *     %2d%s     %4d" % (i, plus, collisions[i])
+print " */\n"
+
+print "#define none 0x%04x\n" % none
+
+print "static const uint16_t map[] = {"
+for i in xrange(0, hash_size, 8):
+    print "   ",
+    for j in xrange(i, i + 8):
+        if map[j] & 0xffff == 0xffff:
+            print "  none,",
+        else:
+            print "0x%04x," % (map[j] & 0xffff),
+    print
+
+print "};"    
+
+# Finally we generate the hash table lookup function.  The hash function and
+# linear probing algorithm matches the hash table generated above.
+
+print """
+void *
+radv_lookup_entrypoint(const char *name)
+{
+   static const uint32_t prime_factor = %d;
+   static const uint32_t prime_step = %d;
+   const struct radv_entrypoint *e;
+   uint32_t hash, h, i;
+   const char *p;
+
+   hash = 0;
+   for (p = name; *p; p++)
+      hash = hash * prime_factor + *p;
+
+   h = hash;
+   do {
+      i = map[h & %d];
+      if (i == none)
+         return NULL;
+      e = &entrypoints[i];
+      h += prime_step;
+   } while (e->hash != hash);
+
+   if (strcmp(name, strings + e->name) != 0)
+      return NULL;
+
+   return radv_resolve_entrypoint(i);
+}
+""" % (prime_factor, prime_step, hash_mask)
diff --git a/src/amd/vulkan/radv_formats.c b/src/amd/vulkan/radv_formats.c
new file mode 100644
index 0000000000..90c140c18b
--- /dev/null
+++ b/src/amd/vulkan/radv_formats.c
@@ -0,0 +1,1085 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "radv_private.h"
+
+#include "vk_format.h"
+#include "sid.h"
+#include "r600d_common.h"
+
+#include "util/u_half.h"
+#include "util/format_srgb.h"
+
+uint32_t radv_translate_buffer_dataformat(const struct vk_format_description *desc,
+					  int first_non_void)
+{
+	unsigned type;
+	int i;
+
+	if (desc->format == VK_FORMAT_B10G11R11_UFLOAT_PACK32)
+		return V_008F0C_BUF_DATA_FORMAT_10_11_11;
+
+	if (first_non_void < 0)
+		return V_008F0C_BUF_DATA_FORMAT_INVALID;
+	type = desc->channel[first_non_void].type;
+
+	if (type == VK_FORMAT_TYPE_FIXED)
+		return V_008F0C_BUF_DATA_FORMAT_INVALID;
+	if (desc->nr_channels == 4 &&
+	    desc->channel[0].size == 10 &&
+	    desc->channel[1].size == 10 &&
+	    desc->channel[2].size == 10 &&
+	    desc->channel[3].size == 2)
+		return V_008F0C_BUF_DATA_FORMAT_2_10_10_10;
+
+	/* See whether the components are of the same size. */
+	for (i = 0; i < desc->nr_channels; i++) {
+		if (desc->channel[first_non_void].size != desc->channel[i].size)
+			return V_008F0C_BUF_DATA_FORMAT_INVALID;
+	}
+
+	switch (desc->channel[first_non_void].size) {
+	case 8:
+		switch (desc->nr_channels) {
+		case 1:
+			return V_008F0C_BUF_DATA_FORMAT_8;
+		case 2:
+			return V_008F0C_BUF_DATA_FORMAT_8_8;
+		case 4:
+			return V_008F0C_BUF_DATA_FORMAT_8_8_8_8;
+		}
+		break;
+	case 16:
+		switch (desc->nr_channels) {
+		case 1:
+			return V_008F0C_BUF_DATA_FORMAT_16;
+		case 2:
+			return V_008F0C_BUF_DATA_FORMAT_16_16;
+		case 4:
+			return V_008F0C_BUF_DATA_FORMAT_16_16_16_16;
+		}
+		break;
+	case 32:
+		/* From the Southern Islands ISA documentation about MTBUF:
+		 * 'Memory reads of data in memory that is 32 or 64 bits do not
+		 * undergo any format conversion.'
+		 */
+		if (type != VK_FORMAT_TYPE_FLOAT &&
+		    !desc->channel[first_non_void].pure_integer)
+			return V_008F0C_BUF_DATA_FORMAT_INVALID;
+
+		switch (desc->nr_channels) {
+		case 1:
+			return V_008F0C_BUF_DATA_FORMAT_32;
+		case 2:
+			return V_008F0C_BUF_DATA_FORMAT_32_32;
+		case 3:
+			return V_008F0C_BUF_DATA_FORMAT_32_32_32;
+		case 4:
+			return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
+		}
+		break;
+	}
+
+	return V_008F0C_BUF_DATA_FORMAT_INVALID;
+}
+
+uint32_t radv_translate_buffer_numformat(const struct vk_format_description *desc,
+					 int first_non_void)
+{
+	if (desc->format == VK_FORMAT_B10G11R11_UFLOAT_PACK32)
+		return V_008F0C_BUF_NUM_FORMAT_FLOAT;
+
+	if (first_non_void < 0)
+		return ~0;
+
+	switch (desc->channel[first_non_void].type) {
+	case VK_FORMAT_TYPE_SIGNED:
+		if (desc->channel[first_non_void].normalized)
+			return V_008F0C_BUF_NUM_FORMAT_SNORM;
+		else if (desc->channel[first_non_void].pure_integer)
+			return V_008F0C_BUF_NUM_FORMAT_SINT;
+		else
+			return V_008F0C_BUF_NUM_FORMAT_SSCALED;
+		break;
+	case VK_FORMAT_TYPE_UNSIGNED:
+		if (desc->channel[first_non_void].normalized)
+			return V_008F0C_BUF_NUM_FORMAT_UNORM;
+		else if (desc->channel[first_non_void].pure_integer)
+			return V_008F0C_BUF_NUM_FORMAT_UINT;
+		else
+			return V_008F0C_BUF_NUM_FORMAT_USCALED;
+		break;
+	case VK_FORMAT_TYPE_FLOAT:
+	default:
+		return V_008F0C_BUF_NUM_FORMAT_FLOAT;
+	}
+}
+
+uint32_t radv_translate_tex_dataformat(VkFormat format,
+				       const struct vk_format_description *desc,
+				       int first_non_void)
+{
+	bool uniform = true;
+	int i;
+
+	if (!desc)
+		return ~0;
+	/* Colorspace (return non-RGB formats directly). */
+	switch (desc->colorspace) {
+		/* Depth stencil formats */
+	case VK_FORMAT_COLORSPACE_ZS:
+		switch (format) {
+		case VK_FORMAT_D16_UNORM:
+			return V_008F14_IMG_DATA_FORMAT_16;
+		case VK_FORMAT_D24_UNORM_S8_UINT:
+			return V_008F14_IMG_DATA_FORMAT_8_24;
+		case VK_FORMAT_S8_UINT:
+			return V_008F14_IMG_DATA_FORMAT_8;
+		case VK_FORMAT_D32_SFLOAT:
+			return V_008F14_IMG_DATA_FORMAT_32;
+		case VK_FORMAT_D32_SFLOAT_S8_UINT:
+			return V_008F14_IMG_DATA_FORMAT_X24_8_32;
+		default:
+			goto out_unknown;
+		}
+
+	case VK_FORMAT_COLORSPACE_YUV:
+		goto out_unknown; /* TODO */
+
+	case VK_FORMAT_COLORSPACE_SRGB:
+		if (desc->nr_channels != 4 && desc->nr_channels != 1)
+			goto out_unknown;
+		break;
+
+	default:
+		break;
+	}
+
+	if (desc->layout == VK_FORMAT_LAYOUT_RGTC) {
+		switch(format) {
+		case VK_FORMAT_BC4_UNORM_BLOCK:
+		case VK_FORMAT_BC4_SNORM_BLOCK:
+			return V_008F14_IMG_DATA_FORMAT_BC4;
+		case VK_FORMAT_BC5_UNORM_BLOCK:
+		case VK_FORMAT_BC5_SNORM_BLOCK:
+			return V_008F14_IMG_DATA_FORMAT_BC5;
+		default:
+			break;
+		}
+	}
+
+	if (desc->layout == VK_FORMAT_LAYOUT_S3TC) {
+		switch(format) {
+		case VK_FORMAT_BC1_RGB_UNORM_BLOCK:
+		case VK_FORMAT_BC1_RGB_SRGB_BLOCK:
+		case VK_FORMAT_BC1_RGBA_UNORM_BLOCK:
+		case VK_FORMAT_BC1_RGBA_SRGB_BLOCK:
+			return V_008F14_IMG_DATA_FORMAT_BC1;
+		case VK_FORMAT_BC2_UNORM_BLOCK:
+		case VK_FORMAT_BC2_SRGB_BLOCK:
+			return V_008F14_IMG_DATA_FORMAT_BC2;
+		case VK_FORMAT_BC3_UNORM_BLOCK:
+		case VK_FORMAT_BC3_SRGB_BLOCK:
+			return V_008F14_IMG_DATA_FORMAT_BC3;
+		default:
+			break;
+		}
+	}
+
+	if (desc->layout == VK_FORMAT_LAYOUT_BPTC) {
+		switch(format) {
+		case VK_FORMAT_BC6H_UFLOAT_BLOCK:
+		case VK_FORMAT_BC6H_SFLOAT_BLOCK:
+			return V_008F14_IMG_DATA_FORMAT_BC6;
+		case VK_FORMAT_BC7_UNORM_BLOCK:
+		case VK_FORMAT_BC7_SRGB_BLOCK:
+			return V_008F14_IMG_DATA_FORMAT_BC7;
+		default:
+			break;
+		}
+	}
+
+	if (format == VK_FORMAT_E5B9G9R9_UFLOAT_PACK32) {
+		return V_008F14_IMG_DATA_FORMAT_5_9_9_9;
+	} else if (format == VK_FORMAT_B10G11R11_UFLOAT_PACK32) {
+		return V_008F14_IMG_DATA_FORMAT_10_11_11;
+	}
+
+	/* R8G8Bx_SNORM - TODO CxV8U8 */
+
+	/* hw cannot support mixed formats (except depth/stencil, since only
+	 * depth is read).*/
+	if (desc->is_mixed && desc->colorspace != VK_FORMAT_COLORSPACE_ZS)
+		goto out_unknown;
+
+	/* See whether the components are of the same size. */
+	for (i = 1; i < desc->nr_channels; i++) {
+		uniform = uniform && desc->channel[0].size == desc->channel[i].size;
+	}
+
+	/* Non-uniform formats. */
+	if (!uniform) {
+		switch(desc->nr_channels) {
+		case 3:
+			if (desc->channel[0].size == 5 &&
+			    desc->channel[1].size == 6 &&
+			    desc->channel[2].size == 5) {
+				return V_008F14_IMG_DATA_FORMAT_5_6_5;
+			}
+			goto out_unknown;
+		case 4:
+			if (desc->channel[0].size == 5 &&
+			    desc->channel[1].size == 5 &&
+			    desc->channel[2].size == 5 &&
+			    desc->channel[3].size == 1) {
+				return V_008F14_IMG_DATA_FORMAT_1_5_5_5;
+			}
+			if (desc->channel[0].size == 1 &&
+			    desc->channel[1].size == 5 &&
+			    desc->channel[2].size == 5 &&
+			    desc->channel[3].size == 5) {
+				return V_008F14_IMG_DATA_FORMAT_5_5_5_1;
+			}
+			if (desc->channel[0].size == 10 &&
+			    desc->channel[1].size == 10 &&
+			    desc->channel[2].size == 10 &&
+			    desc->channel[3].size == 2) {
+				/* Closed VK driver does this also no 2/10/10/10 snorm */
+				if (desc->channel[0].type == VK_FORMAT_TYPE_SIGNED &&
+				    desc->channel[0].normalized)
+					goto out_unknown;
+				return V_008F14_IMG_DATA_FORMAT_2_10_10_10;
+			}
+			goto out_unknown;
+		}
+		goto out_unknown;
+	}
+
+	if (first_non_void < 0 || first_non_void > 3)
+		goto out_unknown;
+
+	/* uniform formats */
+	switch (desc->channel[first_non_void].size) {
+	case 4:
+		switch (desc->nr_channels) {
+#if 0 /* Not supported for render targets */
+		case 2:
+			return V_008F14_IMG_DATA_FORMAT_4_4;
+#endif
+		case 4:
+			return V_008F14_IMG_DATA_FORMAT_4_4_4_4;
+		}
+		break;
+	case 8:
+		switch (desc->nr_channels) {
+		case 1:
+			return V_008F14_IMG_DATA_FORMAT_8;
+		case 2:
+			return V_008F14_IMG_DATA_FORMAT_8_8;
+		case 4:
+			return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
+		}
+		break;
+	case 16:
+		switch (desc->nr_channels) {
+		case 1:
+			return V_008F14_IMG_DATA_FORMAT_16;
+		case 2:
+			return V_008F14_IMG_DATA_FORMAT_16_16;
+		case 4:
+			return V_008F14_IMG_DATA_FORMAT_16_16_16_16;
+		}
+		break;
+	case 32:
+		switch (desc->nr_channels) {
+		case 1:
+			return V_008F14_IMG_DATA_FORMAT_32;
+		case 2:
+			return V_008F14_IMG_DATA_FORMAT_32_32;
+#if 0 /* Not supported for render targets */
+		case 3:
+			return V_008F14_IMG_DATA_FORMAT_32_32_32;
+#endif
+		case 4:
+			return V_008F14_IMG_DATA_FORMAT_32_32_32_32;
+		}
+	}
+
+out_unknown:
+	/* R600_ERR("Unable to handle texformat %d %s\n", format, vk_format_name(format)); */
+	return ~0;
+}
+
+uint32_t radv_translate_tex_numformat(VkFormat format,
+				      const struct vk_format_description *desc,
+				      int first_non_void)
+{
+	switch (format) {
+	case VK_FORMAT_D24_UNORM_S8_UINT:
+		return V_008F14_IMG_NUM_FORMAT_UNORM;
+	default:
+		if (first_non_void < 0) {
+			if (vk_format_is_compressed(format)) {
+				switch (format) {
+				case VK_FORMAT_BC1_RGB_SRGB_BLOCK:
+				case VK_FORMAT_BC1_RGBA_SRGB_BLOCK:
+				case VK_FORMAT_BC2_SRGB_BLOCK:
+				case VK_FORMAT_BC3_SRGB_BLOCK:
+				case VK_FORMAT_BC7_SRGB_BLOCK:
+					return V_008F14_IMG_NUM_FORMAT_SRGB;
+				case VK_FORMAT_BC4_SNORM_BLOCK:
+				case VK_FORMAT_BC5_SNORM_BLOCK:
+			        case VK_FORMAT_BC6H_SFLOAT_BLOCK:
+					return V_008F14_IMG_NUM_FORMAT_SNORM;
+				default:
+					return V_008F14_IMG_NUM_FORMAT_UNORM;
+				}
+			} else if (desc->layout == VK_FORMAT_LAYOUT_SUBSAMPLED) {
+				return V_008F14_IMG_NUM_FORMAT_UNORM;
+			} else {
+				return V_008F14_IMG_NUM_FORMAT_FLOAT;
+			}
+		} else if (desc->colorspace == VK_FORMAT_COLORSPACE_SRGB) {
+			return V_008F14_IMG_NUM_FORMAT_SRGB;
+		} else {
+			switch (desc->channel[first_non_void].type) {
+			case VK_FORMAT_TYPE_FLOAT:
+				return V_008F14_IMG_NUM_FORMAT_FLOAT;
+			case VK_FORMAT_TYPE_SIGNED:
+				if (desc->channel[first_non_void].normalized)
+					return V_008F14_IMG_NUM_FORMAT_SNORM;
+				else if (desc->channel[first_non_void].pure_integer)
+					return V_008F14_IMG_NUM_FORMAT_SINT;
+				else
+					return V_008F14_IMG_NUM_FORMAT_SSCALED;
+			case VK_FORMAT_TYPE_UNSIGNED:
+				if (desc->channel[first_non_void].normalized)
+					return V_008F14_IMG_NUM_FORMAT_UNORM;
+				else if (desc->channel[first_non_void].pure_integer)
+					return V_008F14_IMG_NUM_FORMAT_UINT;
+				else
+					return V_008F14_IMG_NUM_FORMAT_USCALED;
+			default:
+				return V_008F14_IMG_NUM_FORMAT_UNORM;
+			}
+		}
+	}
+}
+
+uint32_t radv_translate_color_numformat(VkFormat format,
+					const struct vk_format_description *desc,
+					int first_non_void)
+{
+	unsigned ntype;
+	if (first_non_void == 4 || desc->channel[first_non_void].type == VK_FORMAT_TYPE_FLOAT)
+		ntype = V_028C70_NUMBER_FLOAT;
+	else {
+		ntype = V_028C70_NUMBER_UNORM;
+		if (desc->colorspace == VK_FORMAT_COLORSPACE_SRGB)
+			ntype = V_028C70_NUMBER_SRGB;
+		else if (desc->channel[first_non_void].type == VK_FORMAT_TYPE_SIGNED) {
+			if (desc->channel[first_non_void].pure_integer) {
+				ntype = V_028C70_NUMBER_SINT;
+			} else if (desc->channel[first_non_void].normalized) {
+				ntype = V_028C70_NUMBER_SNORM;
+			} else
+				ntype = ~0u;
+		} else if (desc->channel[first_non_void].type == VK_FORMAT_TYPE_UNSIGNED) {
+			if (desc->channel[first_non_void].pure_integer) {
+				ntype = V_028C70_NUMBER_UINT;
+			} else if (desc->channel[first_non_void].normalized) {
+				ntype = V_028C70_NUMBER_UNORM;
+			} else
+				ntype = ~0u;
+		}
+	}
+	return ntype;
+}
+
+static bool radv_is_sampler_format_supported(VkFormat format, bool *linear_sampling)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+	uint32_t num_format;
+	if (!desc || format == VK_FORMAT_UNDEFINED)
+		return false;
+	num_format = radv_translate_tex_numformat(format, desc,
+						  vk_format_get_first_non_void_channel(format));
+
+	if (num_format == V_008F14_IMG_NUM_FORMAT_USCALED ||
+	    num_format == V_008F14_IMG_NUM_FORMAT_SSCALED)
+		return false;
+
+	if (num_format == V_008F14_IMG_NUM_FORMAT_UNORM ||
+	    num_format == V_008F14_IMG_NUM_FORMAT_SNORM ||
+	    num_format == V_008F14_IMG_NUM_FORMAT_FLOAT ||
+	    num_format == V_008F14_IMG_NUM_FORMAT_SRGB)
+		*linear_sampling = true;
+	else
+		*linear_sampling = false;
+	return radv_translate_tex_dataformat(format, vk_format_description(format),
+					     vk_format_get_first_non_void_channel(format)) != ~0U;
+}
+
+
+static bool radv_is_storage_image_format_supported(struct radv_physical_device *physical_device,
+						   VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+	unsigned data_format, num_format;
+	if (!desc || format == VK_FORMAT_UNDEFINED)
+		return false;
+
+	data_format = radv_translate_tex_dataformat(format, desc,
+						    vk_format_get_first_non_void_channel(format));
+	num_format = radv_translate_tex_numformat(format, desc,
+						  vk_format_get_first_non_void_channel(format));
+
+	if(data_format == ~0 || num_format == ~0)
+		return false;
+
+	/* Extracted from the GCN3 ISA document. */
+	switch(num_format) {
+	case V_008F14_IMG_NUM_FORMAT_UNORM:
+	case V_008F14_IMG_NUM_FORMAT_SNORM:
+	case V_008F14_IMG_NUM_FORMAT_UINT:
+	case V_008F14_IMG_NUM_FORMAT_SINT:
+	case V_008F14_IMG_NUM_FORMAT_FLOAT:
+		break;
+	default:
+		return false;
+	}
+
+	switch(data_format) {
+	case V_008F14_IMG_DATA_FORMAT_8:
+	case V_008F14_IMG_DATA_FORMAT_16:
+	case V_008F14_IMG_DATA_FORMAT_8_8:
+	case V_008F14_IMG_DATA_FORMAT_32:
+	case V_008F14_IMG_DATA_FORMAT_16_16:
+	case V_008F14_IMG_DATA_FORMAT_10_11_11:
+	case V_008F14_IMG_DATA_FORMAT_11_11_10:
+	case V_008F14_IMG_DATA_FORMAT_10_10_10_2:
+	case V_008F14_IMG_DATA_FORMAT_2_10_10_10:
+	case V_008F14_IMG_DATA_FORMAT_8_8_8_8:
+	case V_008F14_IMG_DATA_FORMAT_32_32:
+	case V_008F14_IMG_DATA_FORMAT_16_16_16_16:
+	case V_008F14_IMG_DATA_FORMAT_32_32_32_32:
+	case V_008F14_IMG_DATA_FORMAT_5_6_5:
+	case V_008F14_IMG_DATA_FORMAT_1_5_5_5:
+	case V_008F14_IMG_DATA_FORMAT_5_5_5_1:
+	case V_008F14_IMG_DATA_FORMAT_4_4_4_4:
+		/* TODO: FMASK formats. */
+		return true;
+	default:
+		return false;
+	}
+}
+
+static bool radv_is_buffer_format_supported(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+	unsigned data_format, num_format;
+	if (!desc || format == VK_FORMAT_UNDEFINED)
+		return false;
+
+	data_format = radv_translate_buffer_dataformat(desc,
+						       vk_format_get_first_non_void_channel(format));
+	num_format = radv_translate_buffer_numformat(desc,
+						     vk_format_get_first_non_void_channel(format));
+
+	return data_format != V_008F0C_BUF_DATA_FORMAT_INVALID &&
+		num_format != ~0;
+}
+
+bool radv_is_colorbuffer_format_supported(VkFormat format, bool *blendable)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+	uint32_t color_format = radv_translate_colorformat(format);
+	uint32_t color_swap = radv_translate_colorswap(format, false);
+	uint32_t color_num_format = radv_translate_color_numformat(format,
+								   desc,
+								   vk_format_get_first_non_void_channel(format));
+
+	if (color_num_format == V_028C70_NUMBER_UINT || color_num_format == V_028C70_NUMBER_SINT ||
+	    color_format == V_028C70_COLOR_8_24 || color_format == V_028C70_COLOR_24_8 ||
+	    color_format == V_028C70_COLOR_X24_8_32_FLOAT) {
+		*blendable = false;
+	} else
+		*blendable = true;
+	return color_format != V_028C70_COLOR_INVALID &&
+		color_swap != ~0U &&
+		color_num_format != ~0;
+}
+
+static bool radv_is_zs_format_supported(VkFormat format)
+{
+	return radv_translate_dbformat(format) != V_028040_Z_INVALID;
+}
+
+static void
+radv_physical_device_get_format_properties(struct radv_physical_device *physical_device,
+					   VkFormat format,
+					   VkFormatProperties *out_properties)
+{
+	VkFormatFeatureFlags linear = 0, tiled = 0, buffer = 0;
+	const struct vk_format_description *desc = vk_format_description(format);
+	bool blendable;
+	if (!desc) {
+		out_properties->linearTilingFeatures = linear;
+		out_properties->optimalTilingFeatures = tiled;
+		out_properties->bufferFeatures = buffer;
+		return;
+	}
+
+	if (radv_is_storage_image_format_supported(physical_device, format)) {
+		tiled |= VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT;
+		linear |= VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT;
+	}
+
+	if (radv_is_buffer_format_supported(format)) {
+		buffer |= VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT |
+			VK_FORMAT_FEATURE_UNIFORM_TEXEL_BUFFER_BIT |
+			VK_FORMAT_FEATURE_STORAGE_TEXEL_BUFFER_BIT;
+	}
+
+	if (vk_format_is_depth_or_stencil(format)) {
+		if (radv_is_zs_format_supported(format))
+			tiled |= VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT;
+		tiled |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT;
+		tiled |= VK_FORMAT_FEATURE_BLIT_SRC_BIT |
+			VK_FORMAT_FEATURE_BLIT_DST_BIT;
+	} else {
+		bool linear_sampling;
+		if (radv_is_sampler_format_supported(format, &linear_sampling)) {
+			linear |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT |
+				VK_FORMAT_FEATURE_BLIT_SRC_BIT;
+			tiled |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT |
+				VK_FORMAT_FEATURE_BLIT_SRC_BIT;
+			if (linear_sampling) {
+				linear |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT;
+				tiled |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT;
+			}
+		}
+		if (radv_is_colorbuffer_format_supported(format, &blendable)) {
+			linear |= VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
+			tiled |= VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
+			if (blendable) {
+				linear |= VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BLEND_BIT;
+				tiled |= VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BLEND_BIT;
+			}
+		}
+	}
+
+	if (format == VK_FORMAT_R32_UINT || format == VK_FORMAT_R32_SINT) {
+		buffer |= VK_FORMAT_FEATURE_STORAGE_TEXEL_BUFFER_ATOMIC_BIT;
+		linear |= VK_FORMAT_FEATURE_STORAGE_IMAGE_ATOMIC_BIT;
+		tiled |= VK_FORMAT_FEATURE_STORAGE_IMAGE_ATOMIC_BIT;
+	}
+
+	out_properties->linearTilingFeatures = linear;
+	out_properties->optimalTilingFeatures = tiled;
+	out_properties->bufferFeatures = buffer;
+}
+
+uint32_t radv_translate_colorformat(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+
+#define HAS_SIZE(x,y,z,w)						\
+	(desc->channel[0].size == (x) && desc->channel[1].size == (y) && \
+         desc->channel[2].size == (z) && desc->channel[3].size == (w))
+
+	if (format == VK_FORMAT_B10G11R11_UFLOAT_PACK32) /* isn't plain */
+		return V_028C70_COLOR_10_11_11;
+
+	if (desc->layout != VK_FORMAT_LAYOUT_PLAIN)
+		return V_028C70_COLOR_INVALID;
+
+	/* hw cannot support mixed formats (except depth/stencil, since
+	 * stencil is not written to). */
+	if (desc->is_mixed && desc->colorspace != VK_FORMAT_COLORSPACE_ZS)
+		return V_028C70_COLOR_INVALID;
+
+	switch (desc->nr_channels) {
+	case 1:
+		switch (desc->channel[0].size) {
+		case 8:
+			return V_028C70_COLOR_8;
+		case 16:
+			return V_028C70_COLOR_16;
+		case 32:
+			return V_028C70_COLOR_32;
+		}
+		break;
+	case 2:
+		if (desc->channel[0].size == desc->channel[1].size) {
+			switch (desc->channel[0].size) {
+			case 8:
+				return V_028C70_COLOR_8_8;
+			case 16:
+				return V_028C70_COLOR_16_16;
+			case 32:
+				return V_028C70_COLOR_32_32;
+			}
+		} else if (HAS_SIZE(8,24,0,0)) {
+			return V_028C70_COLOR_24_8;
+		} else if (HAS_SIZE(24,8,0,0)) {
+			return V_028C70_COLOR_8_24;
+		}
+		break;
+	case 3:
+		if (HAS_SIZE(5,6,5,0)) {
+			return V_028C70_COLOR_5_6_5;
+		} else if (HAS_SIZE(32,8,24,0)) {
+			return V_028C70_COLOR_X24_8_32_FLOAT;
+		}
+		break;
+	case 4:
+		if (desc->channel[0].size == desc->channel[1].size &&
+		    desc->channel[0].size == desc->channel[2].size &&
+		    desc->channel[0].size == desc->channel[3].size) {
+			switch (desc->channel[0].size) {
+			case 4:
+				return V_028C70_COLOR_4_4_4_4;
+			case 8:
+				return V_028C70_COLOR_8_8_8_8;
+			case 16:
+				return V_028C70_COLOR_16_16_16_16;
+			case 32:
+				return V_028C70_COLOR_32_32_32_32;
+			}
+		} else if (HAS_SIZE(5,5,5,1)) {
+			return V_028C70_COLOR_1_5_5_5;
+		} else if (HAS_SIZE(1,5,5,5)) {
+			return V_028C70_COLOR_5_5_5_1;
+		} else if (HAS_SIZE(10,10,10,2)) {
+			return V_028C70_COLOR_2_10_10_10;
+		}
+		break;
+	}
+	return V_028C70_COLOR_INVALID;
+}
+
+uint32_t radv_colorformat_endian_swap(uint32_t colorformat)
+{
+	if (0/*SI_BIG_ENDIAN*/) {
+		switch(colorformat) {
+			/* 8-bit buffers. */
+		case V_028C70_COLOR_8:
+			return V_028C70_ENDIAN_NONE;
+
+			/* 16-bit buffers. */
+		case V_028C70_COLOR_5_6_5:
+		case V_028C70_COLOR_1_5_5_5:
+		case V_028C70_COLOR_4_4_4_4:
+		case V_028C70_COLOR_16:
+		case V_028C70_COLOR_8_8:
+			return V_028C70_ENDIAN_8IN16;
+
+			/* 32-bit buffers. */
+		case V_028C70_COLOR_8_8_8_8:
+		case V_028C70_COLOR_2_10_10_10:
+		case V_028C70_COLOR_8_24:
+		case V_028C70_COLOR_24_8:
+		case V_028C70_COLOR_16_16:
+			return V_028C70_ENDIAN_8IN32;
+
+			/* 64-bit buffers. */
+		case V_028C70_COLOR_16_16_16_16:
+			return V_028C70_ENDIAN_8IN16;
+
+		case V_028C70_COLOR_32_32:
+			return V_028C70_ENDIAN_8IN32;
+
+			/* 128-bit buffers. */
+		case V_028C70_COLOR_32_32_32_32:
+			return V_028C70_ENDIAN_8IN32;
+		default:
+			return V_028C70_ENDIAN_NONE; /* Unsupported. */
+		}
+	} else {
+		return V_028C70_ENDIAN_NONE;
+	}
+}
+
+uint32_t radv_translate_dbformat(VkFormat format)
+{
+	switch (format) {
+	case VK_FORMAT_D16_UNORM:
+	case VK_FORMAT_D16_UNORM_S8_UINT:
+		return V_028040_Z_16;
+	case VK_FORMAT_X8_D24_UNORM_PACK32:
+	case VK_FORMAT_D24_UNORM_S8_UINT:
+		return V_028040_Z_24; /* deprecated on SI */
+	case VK_FORMAT_D32_SFLOAT:
+	case VK_FORMAT_D32_SFLOAT_S8_UINT:
+		return V_028040_Z_32_FLOAT;
+	default:
+		return V_028040_Z_INVALID;
+	}
+}
+
+unsigned radv_translate_colorswap(VkFormat format, bool do_endian_swap)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+
+#define HAS_SWIZZLE(chan,swz) (desc->swizzle[chan] == VK_SWIZZLE_##swz)
+
+	if (format == VK_FORMAT_B10G11R11_UFLOAT_PACK32)
+		return V_0280A0_SWAP_STD;
+
+	if (desc->layout != VK_FORMAT_LAYOUT_PLAIN)
+		return ~0U;
+
+	switch (desc->nr_channels) {
+	case 1:
+		if (HAS_SWIZZLE(0,X))
+			return V_0280A0_SWAP_STD; /* X___ */
+		else if (HAS_SWIZZLE(3,X))
+			return V_0280A0_SWAP_ALT_REV; /* ___X */
+		break;
+	case 2:
+		if ((HAS_SWIZZLE(0,X) && HAS_SWIZZLE(1,Y)) ||
+		    (HAS_SWIZZLE(0,X) && HAS_SWIZZLE(1,NONE)) ||
+		    (HAS_SWIZZLE(0,NONE) && HAS_SWIZZLE(1,Y)))
+			return V_0280A0_SWAP_STD; /* XY__ */
+		else if ((HAS_SWIZZLE(0,Y) && HAS_SWIZZLE(1,X)) ||
+			 (HAS_SWIZZLE(0,Y) && HAS_SWIZZLE(1,NONE)) ||
+		         (HAS_SWIZZLE(0,NONE) && HAS_SWIZZLE(1,X)))
+			/* YX__ */
+			return (do_endian_swap ? V_0280A0_SWAP_STD : V_0280A0_SWAP_STD_REV);
+		else if (HAS_SWIZZLE(0,X) && HAS_SWIZZLE(3,Y))
+			return V_0280A0_SWAP_ALT; /* X__Y */
+		else if (HAS_SWIZZLE(0,Y) && HAS_SWIZZLE(3,X))
+			return V_0280A0_SWAP_ALT_REV; /* Y__X */
+		break;
+	case 3:
+		if (HAS_SWIZZLE(0,X))
+			return (do_endian_swap ? V_0280A0_SWAP_STD_REV : V_0280A0_SWAP_STD);
+		else if (HAS_SWIZZLE(0,Z))
+			return V_0280A0_SWAP_STD_REV; /* ZYX */
+		break;
+	case 4:
+		/* check the middle channels, the 1st and 4th channel can be NONE */
+		if (HAS_SWIZZLE(1,Y) && HAS_SWIZZLE(2,Z)) {
+			return V_0280A0_SWAP_STD; /* XYZW */
+		} else if (HAS_SWIZZLE(1,Z) && HAS_SWIZZLE(2,Y)) {
+			return V_0280A0_SWAP_STD_REV; /* WZYX */
+		} else if (HAS_SWIZZLE(1,Y) && HAS_SWIZZLE(2,X)) {
+			return V_0280A0_SWAP_ALT; /* ZYXW */
+		} else if (HAS_SWIZZLE(1,Z) && HAS_SWIZZLE(2,W)) {
+			/* YZWX */
+			if (desc->is_array)
+				return V_0280A0_SWAP_ALT_REV;
+			else
+				return (do_endian_swap ? V_0280A0_SWAP_ALT : V_0280A0_SWAP_ALT_REV);
+		}
+		break;
+	}
+	return ~0U;
+}
+
+bool radv_format_pack_clear_color(VkFormat format,
+				  uint32_t clear_vals[2],
+				  VkClearColorValue *value)
+{
+	uint8_t r, g, b, a;
+	const struct vk_format_description *desc = vk_format_description(format);
+
+	if (vk_format_get_component_bits(format, VK_FORMAT_COLORSPACE_RGB, 0) <= 8) {
+		if (desc->colorspace == VK_FORMAT_COLORSPACE_RGB) {
+			r = float_to_ubyte(value->float32[0]);
+			g = float_to_ubyte(value->float32[1]);
+			b = float_to_ubyte(value->float32[2]);
+			a = float_to_ubyte(value->float32[3]);
+		} else if (desc->colorspace == VK_FORMAT_COLORSPACE_SRGB) {
+			r = util_format_linear_float_to_srgb_8unorm(value->float32[0]);
+			g = util_format_linear_float_to_srgb_8unorm(value->float32[1]);
+			b = util_format_linear_float_to_srgb_8unorm(value->float32[2]);
+			a = float_to_ubyte(value->float32[3]);
+		}
+	}
+	switch (format) {
+	case VK_FORMAT_R8_UNORM:
+	case VK_FORMAT_R8_SRGB:
+		clear_vals[0] = r;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R8G8_UNORM:
+	case VK_FORMAT_R8G8_SRGB:
+		clear_vals[0] = r | g << 8;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R8G8B8A8_SRGB:
+	case VK_FORMAT_R8G8B8A8_UNORM:
+		clear_vals[0] = r | g << 8 | b << 16 | a << 24;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_B8G8R8A8_SRGB:
+	case VK_FORMAT_B8G8R8A8_UNORM:
+		clear_vals[0] = b | g << 8 | r << 16 | a << 24;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_A8B8G8R8_UNORM_PACK32:
+	case VK_FORMAT_A8B8G8R8_SRGB_PACK32:
+		clear_vals[0] = r | g << 8 | b << 16 | a << 24;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R8_UINT:
+		clear_vals[0] = value->uint32[0] & 0xff;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R16_UINT:
+		clear_vals[0] = value->uint32[0] & 0xffff;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R8G8_UINT:
+		clear_vals[0] = value->uint32[0] & 0xff;
+		clear_vals[0] |= (value->uint32[1] & 0xff) << 8;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R8G8B8A8_UINT:
+		clear_vals[0] = value->uint32[0] & 0xff;
+		clear_vals[0] |= (value->uint32[1] & 0xff) << 8;
+		clear_vals[0] |= (value->uint32[2] & 0xff) << 16;
+		clear_vals[0] |= (value->uint32[3] & 0xff) << 24;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_A8B8G8R8_UINT_PACK32:
+		clear_vals[0] = value->uint32[0] & 0xff;
+		clear_vals[0] |= (value->uint32[1] & 0xff) << 8;
+		clear_vals[0] |= (value->uint32[2] & 0xff) << 16;
+		clear_vals[0] |= (value->uint32[3] & 0xff) << 24;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R16G16_UINT:
+		clear_vals[0] = value->uint32[0] & 0xffff;
+		clear_vals[0] |= (value->uint32[1] & 0xffff) << 16;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R16G16B16A16_UINT:
+		clear_vals[0] = value->uint32[0] & 0xffff;
+		clear_vals[0] |= (value->uint32[1] & 0xffff) << 16;
+		clear_vals[1] = value->uint32[2] & 0xffff;
+		clear_vals[1] |= (value->uint32[3] & 0xffff) << 16;
+		break;
+	case VK_FORMAT_R32_UINT:
+		clear_vals[0] = value->uint32[0];
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R32G32_UINT:
+		clear_vals[0] = value->uint32[0];
+		clear_vals[1] = value->uint32[1];
+		break;
+	case VK_FORMAT_R32_SINT:
+		clear_vals[0] = value->int32[0];
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R16_SFLOAT:
+		clear_vals[0] = util_float_to_half(value->float32[0]);
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R16G16_SFLOAT:
+		clear_vals[0] = util_float_to_half(value->float32[0]);
+		clear_vals[0] |= (uint32_t)util_float_to_half(value->float32[1]) << 16;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R16G16B16A16_SFLOAT:
+		clear_vals[0] = util_float_to_half(value->float32[0]);
+		clear_vals[0] |= (uint32_t)util_float_to_half(value->float32[1]) << 16;
+		clear_vals[1] = util_float_to_half(value->float32[2]);
+		clear_vals[1] |= (uint32_t)util_float_to_half(value->float32[3]) << 16;
+		break;
+	case VK_FORMAT_R16_UNORM:
+		clear_vals[0] = ((uint16_t)util_iround(CLAMP(value->float32[0], 0.0f, 1.0f) * 0xffff)) & 0xffff;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R16G16_UNORM:
+		clear_vals[0] = ((uint16_t)util_iround(CLAMP(value->float32[0], 0.0f, 1.0f) * 0xffff)) & 0xffff;
+		clear_vals[0] |= ((uint16_t)util_iround(CLAMP(value->float32[1], 0.0f, 1.0f) * 0xffff)) << 16;
+		clear_vals[1] = 0;
+		break;
+	case VK_FORMAT_R16G16B16A16_UNORM:
+		clear_vals[0] = ((uint16_t)util_iround(CLAMP(value->float32[0], 0.0f, 1.0f) * 0xffff)) & 0xffff;
+		clear_vals[0] |= ((uint16_t)util_iround(CLAMP(value->float32[1], 0.0f, 1.0f) * 0xffff)) << 16;
+		clear_vals[1] = ((uint16_t)util_iround(CLAMP(value->float32[2], 0.0f, 1.0f) * 0xffff)) & 0xffff;
+		clear_vals[1] |= ((uint16_t)util_iround(CLAMP(value->float32[3], 0.0f, 1.0f) * 0xffff)) << 16;
+		break;
+	case VK_FORMAT_A2B10G10R10_UNORM_PACK32:
+		/* TODO */
+		return false;
+	case VK_FORMAT_R32G32_SFLOAT:
+		clear_vals[0] = fui(value->float32[0]);
+		clear_vals[1] = fui(value->float32[1]);
+		break;
+	case VK_FORMAT_R32_SFLOAT:
+		clear_vals[1] = 0;
+		clear_vals[0] = fui(value->float32[0]);
+		break;
+	default:
+		fprintf(stderr, "failed to fast clear %d\n", format);
+		return false;
+	}
+	return true;
+}
+
+void radv_GetPhysicalDeviceFormatProperties(
+	VkPhysicalDevice                            physicalDevice,
+	VkFormat                                    format,
+	VkFormatProperties*                         pFormatProperties)
+{
+	RADV_FROM_HANDLE(radv_physical_device, physical_device, physicalDevice);
+
+	radv_physical_device_get_format_properties(physical_device,
+						   format,
+						   pFormatProperties);
+}
+
+VkResult radv_GetPhysicalDeviceImageFormatProperties(
+	VkPhysicalDevice                            physicalDevice,
+	VkFormat                                    format,
+	VkImageType                                 type,
+	VkImageTiling                               tiling,
+	VkImageUsageFlags                           usage,
+	VkImageCreateFlags                          createFlags,
+	VkImageFormatProperties*                    pImageFormatProperties)
+{
+	RADV_FROM_HANDLE(radv_physical_device, physical_device, physicalDevice);
+	VkFormatProperties format_props;
+	VkFormatFeatureFlags format_feature_flags;
+	VkExtent3D maxExtent;
+	uint32_t maxMipLevels;
+	uint32_t maxArraySize;
+	VkSampleCountFlags sampleCounts = VK_SAMPLE_COUNT_1_BIT;
+
+	radv_physical_device_get_format_properties(physical_device, format,
+						   &format_props);
+	if (tiling == VK_IMAGE_TILING_LINEAR) {
+		format_feature_flags = format_props.linearTilingFeatures;
+	} else if (tiling == VK_IMAGE_TILING_OPTIMAL) {
+		format_feature_flags = format_props.optimalTilingFeatures;
+	} else {
+		unreachable("bad VkImageTiling");
+	}
+
+	if (format_feature_flags == 0)
+		goto unsupported;
+
+	switch (type) {
+	default:
+		unreachable("bad vkimage type\n");
+	case VK_IMAGE_TYPE_1D:
+		maxExtent.width = 16384;
+		maxExtent.height = 1;
+		maxExtent.depth = 1;
+		maxMipLevels = 15; /* log2(maxWidth) + 1 */
+		maxArraySize = 2048;
+		break;
+	case VK_IMAGE_TYPE_2D:
+		maxExtent.width = 16384;
+		maxExtent.height = 16384;
+		maxExtent.depth = 1;
+		maxMipLevels = 15; /* log2(maxWidth) + 1 */
+		maxArraySize = 2048;
+		break;
+	case VK_IMAGE_TYPE_3D:
+		maxExtent.width = 2048;
+		maxExtent.height = 2048;
+		maxExtent.depth = 2048;
+		maxMipLevels = 12; /* log2(maxWidth) + 1 */
+		maxArraySize = 1;
+		break;
+	}
+
+	if (tiling == VK_IMAGE_TILING_OPTIMAL &&
+	    type == VK_IMAGE_TYPE_2D &&
+	    (format_feature_flags & (VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT |
+				     VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT)) &&
+	    !(createFlags & VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT) &&
+	    !(usage & VK_IMAGE_USAGE_STORAGE_BIT)) {
+		sampleCounts |= VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT | VK_SAMPLE_COUNT_8_BIT;
+	}
+
+	if (usage & VK_IMAGE_USAGE_SAMPLED_BIT) {
+		if (!(format_feature_flags & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT)) {
+			goto unsupported;
+		}
+	}
+
+	if (usage & VK_IMAGE_USAGE_STORAGE_BIT) {
+		if (!(format_feature_flags & VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT)) {
+			goto unsupported;
+		}
+	}
+
+	if (usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
+		if (!(format_feature_flags & VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT)) {
+			goto unsupported;
+		}
+	}
+
+	if (usage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
+		if (!(format_feature_flags & VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT)) {
+			goto unsupported;
+		}
+	}
+
+	*pImageFormatProperties = (VkImageFormatProperties) {
+		.maxExtent = maxExtent,
+		.maxMipLevels = maxMipLevels,
+		.maxArrayLayers = maxArraySize,
+		.sampleCounts = sampleCounts,
+
+		/* FINISHME: Accurately calculate
+		 * VkImageFormatProperties::maxResourceSize.
+		 */
+		.maxResourceSize = UINT32_MAX,
+	};
+
+	return VK_SUCCESS;
+unsupported:
+	*pImageFormatProperties = (VkImageFormatProperties) {
+		.maxExtent = { 0, 0, 0 },
+		.maxMipLevels = 0,
+		.maxArrayLayers = 0,
+		.sampleCounts = 0,
+		.maxResourceSize = 0,
+	};
+
+	return VK_ERROR_FORMAT_NOT_SUPPORTED;
+}
+
+void radv_GetPhysicalDeviceSparseImageFormatProperties(
+	VkPhysicalDevice                            physicalDevice,
+	VkFormat                                    format,
+	VkImageType                                 type,
+	uint32_t                                    samples,
+	VkImageUsageFlags                           usage,
+	VkImageTiling                               tiling,
+	uint32_t*                                   pNumProperties,
+	VkSparseImageFormatProperties*              pProperties)
+{
+	/* Sparse images are not yet supported. */
+	*pNumProperties = 0;
+}
diff --git a/src/amd/vulkan/radv_image.c b/src/amd/vulkan/radv_image.c
new file mode 100644
index 0000000000..2223c89cf1
--- /dev/null
+++ b/src/amd/vulkan/radv_image.c
@@ -0,0 +1,1030 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based in part on anv driver which is:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "radv_private.h"
+#include "vk_format.h"
+#include "radv_radeon_winsys.h"
+#include "sid.h"
+#include "util/debug.h"
+static unsigned
+radv_choose_tiling(struct radv_device *Device,
+		   const struct radv_image_create_info *create_info)
+{
+	const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
+
+	if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR) {
+		assert(pCreateInfo->samples <= 1);
+		return RADEON_SURF_MODE_LINEAR_ALIGNED;
+	}
+
+	/* MSAA resources must be 2D tiled. */
+	if (pCreateInfo->samples > 1)
+		return RADEON_SURF_MODE_2D;
+
+	return RADEON_SURF_MODE_2D;
+}
+static int
+radv_init_surface(struct radv_device *device,
+		  struct radeon_surf *surface,
+		  const struct radv_image_create_info *create_info)
+{
+	const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
+	unsigned array_mode = radv_choose_tiling(device, create_info);
+	const struct vk_format_description *desc =
+		vk_format_description(pCreateInfo->format);
+	bool is_depth, is_stencil, blendable;
+
+	is_depth = vk_format_has_depth(desc);
+	is_stencil = vk_format_has_stencil(desc);
+	surface->npix_x = pCreateInfo->extent.width;
+	surface->npix_y = pCreateInfo->extent.height;
+	surface->npix_z = pCreateInfo->extent.depth;
+
+	surface->blk_w = vk_format_get_blockwidth(pCreateInfo->format);
+	surface->blk_h = vk_format_get_blockheight(pCreateInfo->format);
+	surface->blk_d = 1;
+	surface->array_size = pCreateInfo->arrayLayers;
+	surface->last_level = pCreateInfo->mipLevels - 1;
+
+	surface->bpe = vk_format_get_blocksize(pCreateInfo->format);
+	/* align byte per element on dword */
+	if (surface->bpe == 3) {
+		surface->bpe = 4;
+	}
+	surface->nsamples = pCreateInfo->samples ? pCreateInfo->samples : 1;
+	surface->flags = RADEON_SURF_SET(array_mode, MODE);
+
+	switch (pCreateInfo->imageType){
+	case VK_IMAGE_TYPE_1D:
+		if (pCreateInfo->arrayLayers > 1)
+			surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D_ARRAY, TYPE);
+		else
+			surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D, TYPE);
+		break;
+	case VK_IMAGE_TYPE_2D:
+		if (pCreateInfo->arrayLayers > 1)
+			surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D_ARRAY, TYPE);
+		else
+			surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D, TYPE);
+		break;
+	case VK_IMAGE_TYPE_3D:
+		surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_3D, TYPE);
+		break;
+	default:
+		unreachable("unhandled image type");
+	}
+
+	if (is_depth) {
+		surface->flags |= RADEON_SURF_ZBUFFER;
+	}
+
+	if (is_stencil)
+		surface->flags |= RADEON_SURF_SBUFFER |
+			RADEON_SURF_HAS_SBUFFER_MIPTREE;
+
+	surface->flags |= RADEON_SURF_HAS_TILE_MODE_INDEX;
+
+	if ((pCreateInfo->usage & (VK_IMAGE_USAGE_TRANSFER_SRC_BIT |
+	                           VK_IMAGE_USAGE_STORAGE_BIT)) ||
+	    (pCreateInfo->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) ||
+            (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR) ||
+            device->instance->physicalDevice.rad_info.chip_class < VI ||
+            create_info->scanout || !device->allow_dcc ||
+            !radv_is_colorbuffer_format_supported(pCreateInfo->format, &blendable))
+		surface->flags |= RADEON_SURF_DISABLE_DCC;
+	if (create_info->scanout)
+		surface->flags |= RADEON_SURF_SCANOUT;
+	return 0;
+}
+#define ATI_VENDOR_ID 0x1002
+static uint32_t si_get_bo_metadata_word1(struct radv_device *device)
+{
+	return (ATI_VENDOR_ID << 16) | device->instance->physicalDevice.rad_info.pci_id;
+}
+
+static inline unsigned
+si_tile_mode_index(const struct radv_image *image, unsigned level, bool stencil)
+{
+	if (stencil)
+		return image->surface.stencil_tiling_index[level];
+	else
+		return image->surface.tiling_index[level];
+}
+
+static unsigned radv_map_swizzle(unsigned swizzle)
+{
+	switch (swizzle) {
+	case VK_SWIZZLE_Y:
+		return V_008F0C_SQ_SEL_Y;
+	case VK_SWIZZLE_Z:
+		return V_008F0C_SQ_SEL_Z;
+	case VK_SWIZZLE_W:
+		return V_008F0C_SQ_SEL_W;
+	case VK_SWIZZLE_0:
+		return V_008F0C_SQ_SEL_0;
+	case VK_SWIZZLE_1:
+		return V_008F0C_SQ_SEL_1;
+	default: /* VK_SWIZZLE_X */
+		return V_008F0C_SQ_SEL_X;
+	}
+}
+
+static void
+radv_make_buffer_descriptor(struct radv_device *device,
+			    struct radv_buffer *buffer,
+			    VkFormat vk_format,
+			    unsigned offset,
+			    unsigned range,
+			    uint32_t *state)
+{
+	const struct vk_format_description *desc;
+	unsigned stride;
+	uint64_t gpu_address = device->ws->buffer_get_va(buffer->bo);
+	uint64_t va = gpu_address + buffer->offset;
+	unsigned num_format, data_format;
+	int first_non_void;
+	desc = vk_format_description(vk_format);
+	first_non_void = vk_format_get_first_non_void_channel(vk_format);
+	stride = desc->block.bits / 8;
+
+	num_format = radv_translate_buffer_numformat(desc, first_non_void);
+	data_format = radv_translate_buffer_dataformat(desc, first_non_void);
+
+	va += offset;
+	state[0] = va;
+	state[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) |
+		S_008F04_STRIDE(stride);
+	state[2] = range;
+	state[3] = S_008F0C_DST_SEL_X(radv_map_swizzle(desc->swizzle[0])) |
+		   S_008F0C_DST_SEL_Y(radv_map_swizzle(desc->swizzle[1])) |
+		   S_008F0C_DST_SEL_Z(radv_map_swizzle(desc->swizzle[2])) |
+		   S_008F0C_DST_SEL_W(radv_map_swizzle(desc->swizzle[3])) |
+		   S_008F0C_NUM_FORMAT(num_format) |
+		   S_008F0C_DATA_FORMAT(data_format);
+}
+
+static void
+si_set_mutable_tex_desc_fields(struct radv_device *device,
+			       struct radv_image *image,
+			       const struct radeon_surf_level *base_level_info,
+			       unsigned base_level, unsigned first_level,
+			       unsigned block_width, bool is_stencil,
+			       uint32_t *state)
+{
+	uint64_t gpu_address = device->ws->buffer_get_va(image->bo) + image->offset;
+	uint64_t va = gpu_address + base_level_info->offset;
+	unsigned pitch = base_level_info->nblk_x * block_width;
+
+	state[1] &= C_008F14_BASE_ADDRESS_HI;
+	state[3] &= C_008F1C_TILING_INDEX;
+	state[4] &= C_008F20_PITCH;
+	state[6] &= C_008F28_COMPRESSION_EN;
+
+	assert(!(va & 255));
+
+	state[0] = va >> 8;
+	state[1] |= S_008F14_BASE_ADDRESS_HI(va >> 40);
+	state[3] |= S_008F1C_TILING_INDEX(si_tile_mode_index(image, base_level,
+							     is_stencil));
+	state[4] |= S_008F20_PITCH(pitch - 1);
+
+	if (image->surface.dcc_size && image->surface.level[first_level].dcc_enabled) {
+		state[6] |= S_008F28_COMPRESSION_EN(1);
+		state[7] = (gpu_address +
+			    image->dcc_offset +
+			    base_level_info->dcc_offset) >> 8;
+	}
+}
+
+static unsigned radv_tex_dim(VkImageType image_type, VkImageViewType view_type,
+			     unsigned nr_layers, unsigned nr_samples, bool is_storage_image)
+{
+	if (view_type == VK_IMAGE_VIEW_TYPE_CUBE || view_type == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY)
+		return is_storage_image ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY : V_008F1C_SQ_RSRC_IMG_CUBE;
+	switch (image_type) {
+	case VK_IMAGE_TYPE_1D:
+		return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_1D_ARRAY : V_008F1C_SQ_RSRC_IMG_1D;
+	case VK_IMAGE_TYPE_2D:
+		if (nr_samples > 1)
+			return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY : V_008F1C_SQ_RSRC_IMG_2D_MSAA;
+		else
+			return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY : V_008F1C_SQ_RSRC_IMG_2D;
+	case VK_IMAGE_TYPE_3D:
+		if (view_type == VK_IMAGE_VIEW_TYPE_3D)
+			return V_008F1C_SQ_RSRC_IMG_3D;
+		else
+			return V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
+	default:
+		unreachable("illegale image type");
+	}
+}
+/**
+ * Build the sampler view descriptor for a texture.
+ */
+static void
+si_make_texture_descriptor(struct radv_device *device,
+			   struct radv_image *image,
+			   bool sampler,
+			   VkImageViewType view_type,
+			   VkFormat vk_format,
+			   const VkComponentMapping *mapping,
+			   unsigned first_level, unsigned last_level,
+			   unsigned first_layer, unsigned last_layer,
+			   unsigned width, unsigned height, unsigned depth,
+			   uint32_t *state,
+			   uint32_t *fmask_state)
+{
+	const struct vk_format_description *desc;
+	enum vk_swizzle swizzle[4];
+	int first_non_void;
+	unsigned num_format, data_format, type;
+
+	desc = vk_format_description(vk_format);
+
+	if (desc->colorspace == VK_FORMAT_COLORSPACE_ZS) {
+		const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
+		const unsigned char swizzle_yyyy[4] = {1, 1, 1, 1};
+
+		switch (vk_format) {
+		case VK_FORMAT_X8_D24_UNORM_PACK32:
+		case VK_FORMAT_D24_UNORM_S8_UINT:
+		case VK_FORMAT_D32_SFLOAT_S8_UINT:
+			vk_format_compose_swizzles(mapping, swizzle_yyyy, swizzle);
+			break;
+		default:
+			vk_format_compose_swizzles(mapping, swizzle_xxxx, swizzle);
+		}
+	} else {
+		vk_format_compose_swizzles(mapping, desc->swizzle, swizzle);
+	}
+
+	first_non_void = vk_format_get_first_non_void_channel(vk_format);
+
+	num_format = radv_translate_tex_numformat(vk_format, desc, first_non_void);
+	if (num_format == ~0) {
+		num_format = 0;
+	}
+
+	data_format = radv_translate_tex_dataformat(vk_format, desc, first_non_void);
+	if (data_format == ~0) {
+		data_format = 0;
+	}
+
+	type = radv_tex_dim(image->type, view_type, image->array_size, image->samples,
+			    (image->usage & VK_IMAGE_USAGE_STORAGE_BIT));
+	if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
+	        height = 1;
+		depth = image->array_size;
+	} else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY ||
+		   type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
+		if (view_type != VK_IMAGE_VIEW_TYPE_3D)
+			depth = image->array_size;
+	} else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
+		depth = image->array_size / 6;
+
+	state[0] = 0;
+	state[1] = (S_008F14_DATA_FORMAT(data_format) |
+		    S_008F14_NUM_FORMAT(num_format));
+	state[2] = (S_008F18_WIDTH(width - 1) |
+		    S_008F18_HEIGHT(height - 1));
+	state[3] = (S_008F1C_DST_SEL_X(radv_map_swizzle(swizzle[0])) |
+		    S_008F1C_DST_SEL_Y(radv_map_swizzle(swizzle[1])) |
+		    S_008F1C_DST_SEL_Z(radv_map_swizzle(swizzle[2])) |
+		    S_008F1C_DST_SEL_W(radv_map_swizzle(swizzle[3])) |
+		    S_008F1C_BASE_LEVEL(image->samples > 1 ?
+					0 : first_level) |
+		    S_008F1C_LAST_LEVEL(image->samples > 1 ?
+					util_logbase2(image->samples) :
+					last_level) |
+		    S_008F1C_POW2_PAD(image->levels > 1) |
+		    S_008F1C_TYPE(type));
+	state[4] = S_008F20_DEPTH(depth - 1);
+	state[5] = (S_008F24_BASE_ARRAY(first_layer) |
+		    S_008F24_LAST_ARRAY(last_layer));
+	state[6] = 0;
+	state[7] = 0;
+
+	if (image->dcc_offset) {
+		unsigned swap = radv_translate_colorswap(vk_format, FALSE);
+
+		state[6] = S_008F28_ALPHA_IS_ON_MSB(swap <= 1);
+	} else {
+		/* The last dword is unused by hw. The shader uses it to clear
+		 * bits in the first dword of sampler state.
+		 */
+		if (device->instance->physicalDevice.rad_info.chip_class <= CIK && image->samples <= 1) {
+			if (first_level == last_level)
+				state[7] = C_008F30_MAX_ANISO_RATIO;
+			else
+				state[7] = 0xffffffff;
+		}
+	}
+
+	/* Initialize the sampler view for FMASK. */
+	if (image->fmask.size) {
+		uint32_t fmask_format;
+		uint64_t gpu_address = device->ws->buffer_get_va(image->bo);
+		uint64_t va;
+
+		va = gpu_address + image->offset + image->fmask.offset;
+
+		switch (image->samples) {
+		case 2:
+			fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2;
+			break;
+		case 4:
+			fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4;
+			break;
+		case 8:
+			fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8;
+			break;
+		default:
+			assert(0);
+			fmask_format = V_008F14_IMG_DATA_FORMAT_INVALID;
+		}
+
+		fmask_state[0] = va >> 8;
+		fmask_state[1] = S_008F14_BASE_ADDRESS_HI(va >> 40) |
+			S_008F14_DATA_FORMAT(fmask_format) |
+			S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_UINT);
+		fmask_state[2] = S_008F18_WIDTH(width - 1) |
+			S_008F18_HEIGHT(height - 1);
+		fmask_state[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X) |
+			S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
+			S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
+			S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
+			S_008F1C_TILING_INDEX(image->fmask.tile_mode_index) |
+			S_008F1C_TYPE(radv_tex_dim(image->type, view_type, 1, 0, false));
+		fmask_state[4] = S_008F20_DEPTH(depth - 1) |
+			S_008F20_PITCH(image->fmask.pitch_in_pixels - 1);
+		fmask_state[5] = S_008F24_BASE_ARRAY(first_layer) |
+			S_008F24_LAST_ARRAY(last_layer);
+		fmask_state[6] = 0;
+		fmask_state[7] = 0;
+	}
+}
+
+static void
+radv_query_opaque_metadata(struct radv_device *device,
+			   struct radv_image *image,
+			   struct radeon_bo_metadata *md)
+{
+	static const VkComponentMapping fixedmapping;
+	uint32_t desc[8], i;
+
+	/* Metadata image format format version 1:
+	 * [0] = 1 (metadata format identifier)
+	 * [1] = (VENDOR_ID << 16) | PCI_ID
+	 * [2:9] = image descriptor for the whole resource
+	 *         [2] is always 0, because the base address is cleared
+	 *         [9] is the DCC offset bits [39:8] from the beginning of
+	 *             the buffer
+	 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
+	 */
+	md->metadata[0] = 1; /* metadata image format version 1 */
+
+	/* TILE_MODE_INDEX is ambiguous without a PCI ID. */
+	md->metadata[1] = si_get_bo_metadata_word1(device);
+
+
+	si_make_texture_descriptor(device, image, true,
+				   (VkImageViewType)image->type, image->vk_format,
+				   &fixedmapping, 0, image->levels - 1, 0,
+				   image->array_size,
+				   image->extent.width, image->extent.height,
+				   image->extent.depth,
+				   desc, NULL);
+
+	si_set_mutable_tex_desc_fields(device, image, &image->surface.level[0], 0, 0,
+				       image->surface.blk_w, false, desc);
+
+	/* Clear the base address and set the relative DCC offset. */
+	desc[0] = 0;
+	desc[1] &= C_008F14_BASE_ADDRESS_HI;
+	desc[7] = image->dcc_offset >> 8;
+
+	/* Dwords [2:9] contain the image descriptor. */
+	memcpy(&md->metadata[2], desc, sizeof(desc));
+
+	/* Dwords [10:..] contain the mipmap level offsets. */
+	for (i = 0; i <= image->levels - 1; i++)
+		md->metadata[10+i] = image->surface.level[i].offset >> 8;
+
+	md->size_metadata = (11 + image->levels - 1) * 4;
+}
+
+void
+radv_init_metadata(struct radv_device *device,
+		   struct radv_image *image,
+		   struct radeon_bo_metadata *metadata)
+{
+	struct radeon_surf *surface = &image->surface;
+
+	memset(metadata, 0, sizeof(*metadata));
+	metadata->microtile = surface->level[0].mode >= RADEON_SURF_MODE_1D ?
+		RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
+	metadata->macrotile = surface->level[0].mode >= RADEON_SURF_MODE_2D ?
+		RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
+	metadata->pipe_config = surface->pipe_config;
+	metadata->bankw = surface->bankw;
+	metadata->bankh = surface->bankh;
+	metadata->tile_split = surface->tile_split;
+	metadata->mtilea = surface->mtilea;
+	metadata->num_banks = surface->num_banks;
+	metadata->stride = surface->level[0].pitch_bytes;
+	metadata->scanout = (surface->flags & RADEON_SURF_SCANOUT) != 0;
+
+	radv_query_opaque_metadata(device, image, metadata);
+}
+
+/* The number of samples can be specified independently of the texture. */
+static void
+radv_image_get_fmask_info(struct radv_device *device,
+			  struct radv_image *image,
+			  unsigned nr_samples,
+			  struct radv_fmask_info *out)
+{
+	/* FMASK is allocated like an ordinary texture. */
+	struct radeon_surf fmask = image->surface;
+
+	memset(out, 0, sizeof(*out));
+
+	fmask.bo_alignment = 0;
+	fmask.bo_size = 0;
+	fmask.nsamples = 1;
+	fmask.flags |= RADEON_SURF_FMASK;
+
+	/* Force 2D tiling if it wasn't set. This may occur when creating
+	 * FMASK for MSAA resolve on R6xx. On R6xx, the single-sample
+	 * destination buffer must have an FMASK too. */
+	fmask.flags = RADEON_SURF_CLR(fmask.flags, MODE);
+	fmask.flags |= RADEON_SURF_SET(RADEON_SURF_MODE_2D, MODE);
+
+	fmask.flags |= RADEON_SURF_HAS_TILE_MODE_INDEX;
+
+	switch (nr_samples) {
+	case 2:
+	case 4:
+		fmask.bpe = 1;
+		break;
+	case 8:
+		fmask.bpe = 4;
+		break;
+	default:
+		return;
+	}
+
+	device->ws->surface_init(device->ws, &fmask);
+	assert(fmask.level[0].mode == RADEON_SURF_MODE_2D);
+
+	out->slice_tile_max = (fmask.level[0].nblk_x * fmask.level[0].nblk_y) / 64;
+	if (out->slice_tile_max)
+		out->slice_tile_max -= 1;
+
+	out->tile_mode_index = fmask.tiling_index[0];
+	out->pitch_in_pixels = fmask.level[0].nblk_x;
+	out->bank_height = fmask.bankh;
+	out->alignment = MAX2(256, fmask.bo_alignment);
+	out->size = fmask.bo_size;
+}
+
+static void
+radv_image_alloc_fmask(struct radv_device *device,
+		       struct radv_image *image)
+{
+	radv_image_get_fmask_info(device, image, image->samples, &image->fmask);
+
+	image->fmask.offset = align64(image->size, image->fmask.alignment);
+	image->size = image->fmask.offset + image->fmask.size;
+}
+
+static void
+radv_image_get_cmask_info(struct radv_device *device,
+			  struct radv_image *image,
+			  struct radv_cmask_info *out)
+{
+	unsigned pipe_interleave_bytes = device->instance->physicalDevice.rad_info.pipe_interleave_bytes;
+	unsigned num_pipes = device->instance->physicalDevice.rad_info.num_tile_pipes;
+	unsigned cl_width, cl_height;
+
+	switch (num_pipes) {
+	case 2:
+		cl_width = 32;
+		cl_height = 16;
+		break;
+	case 4:
+		cl_width = 32;
+		cl_height = 32;
+		break;
+	case 8:
+		cl_width = 64;
+		cl_height = 32;
+		break;
+	case 16: /* Hawaii */
+		cl_width = 64;
+		cl_height = 64;
+		break;
+	default:
+		assert(0);
+		return;
+	}
+
+	unsigned base_align = num_pipes * pipe_interleave_bytes;
+
+	unsigned width = align(image->surface.npix_x, cl_width*8);
+	unsigned height = align(image->surface.npix_y, cl_height*8);
+	unsigned slice_elements = (width * height) / (8*8);
+
+	/* Each element of CMASK is a nibble. */
+	unsigned slice_bytes = slice_elements / 2;
+
+	out->pitch = width;
+	out->height = height;
+	out->xalign = cl_width * 8;
+	out->yalign = cl_height * 8;
+	out->slice_tile_max = (width * height) / (128*128);
+	if (out->slice_tile_max)
+		out->slice_tile_max -= 1;
+
+	out->alignment = MAX2(256, base_align);
+	out->size = (image->type == VK_IMAGE_TYPE_3D ? image->extent.depth : image->array_size) *
+		    align(slice_bytes, base_align);
+}
+
+static void
+radv_image_alloc_cmask(struct radv_device *device,
+		       struct radv_image *image)
+{
+	radv_image_get_cmask_info(device, image, &image->cmask);
+
+	image->cmask.offset = align64(image->size, image->cmask.alignment);
+	/* + 8 for storing the clear values */
+	image->clear_value_offset = image->cmask.offset + image->cmask.size;
+	image->size = image->cmask.offset + image->cmask.size + 8;
+}
+
+static void
+radv_image_alloc_dcc(struct radv_device *device,
+		       struct radv_image *image)
+{
+	image->dcc_offset = align64(image->size, image->surface.dcc_alignment);
+	/* + 8 for storing the clear values */
+	image->clear_value_offset = image->dcc_offset + image->surface.dcc_size;
+	image->size = image->dcc_offset + image->surface.dcc_size + 8;
+}
+
+static unsigned
+radv_image_get_htile_size(struct radv_device *device,
+			  struct radv_image *image)
+{
+	unsigned cl_width, cl_height, width, height;
+	unsigned slice_elements, slice_bytes, base_align;
+	unsigned num_pipes = device->instance->physicalDevice.rad_info.num_tile_pipes;
+	unsigned pipe_interleave_bytes = device->instance->physicalDevice.rad_info.pipe_interleave_bytes;
+
+	/* Overalign HTILE on P2 configs to work around GPU hangs in
+	 * piglit/depthstencil-render-miplevels 585.
+	 *
+	 * This has been confirmed to help Kabini & Stoney, where the hangs
+	 * are always reproducible. I think I have seen the test hang
+	 * on Carrizo too, though it was very rare there.
+	 */
+	if (device->instance->physicalDevice.rad_info.chip_class >= CIK && num_pipes < 4)
+		num_pipes = 4;
+
+	switch (num_pipes) {
+	case 1:
+		cl_width = 32;
+		cl_height = 16;
+		break;
+	case 2:
+		cl_width = 32;
+		cl_height = 32;
+		break;
+	case 4:
+		cl_width = 64;
+		cl_height = 32;
+		break;
+	case 8:
+		cl_width = 64;
+		cl_height = 64;
+		break;
+	case 16:
+		cl_width = 128;
+		cl_height = 64;
+		break;
+	default:
+		assert(0);
+		return 0;
+	}
+
+	width = align(image->surface.npix_x, cl_width * 8);
+	height = align(image->surface.npix_y, cl_height * 8);
+
+	slice_elements = (width * height) / (8 * 8);
+	slice_bytes = slice_elements * 4;
+
+	base_align = num_pipes * pipe_interleave_bytes;
+
+	image->htile.pitch = width;
+	image->htile.height = height;
+	image->htile.xalign = cl_width * 8;
+	image->htile.yalign = cl_height * 8;
+
+	return image->array_size *
+		align(slice_bytes, base_align);
+}
+
+static void
+radv_image_alloc_htile(struct radv_device *device,
+		       struct radv_image *image)
+{
+	if (env_var_as_boolean("RADV_HIZ_DISABLE", false))
+		return;
+
+	image->htile.size = radv_image_get_htile_size(device, image);
+
+	if (!image->htile.size)
+		return;
+
+	image->htile.offset = align64(image->size, 32768);
+
+	/* + 8 for storing the clear values */
+	image->clear_value_offset = image->htile.offset + image->htile.size;
+	image->size = image->htile.offset + image->htile.size + 8;
+	image->alignment = align64(image->alignment, 32768);
+}
+
+VkResult
+radv_image_create(VkDevice _device,
+		  const struct radv_image_create_info *create_info,
+		  const VkAllocationCallbacks* alloc,
+		  VkImage *pImage)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
+	struct radv_image *image = NULL;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO);
+
+	radv_assert(pCreateInfo->mipLevels > 0);
+	radv_assert(pCreateInfo->arrayLayers > 0);
+	radv_assert(pCreateInfo->samples > 0);
+	radv_assert(pCreateInfo->extent.width > 0);
+	radv_assert(pCreateInfo->extent.height > 0);
+	radv_assert(pCreateInfo->extent.depth > 0);
+
+	image = radv_alloc2(&device->alloc, alloc, sizeof(*image), 8,
+			    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (!image)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	memset(image, 0, sizeof(*image));
+	image->type = pCreateInfo->imageType;
+	image->extent = pCreateInfo->extent;
+	image->vk_format = pCreateInfo->format;
+	image->levels = pCreateInfo->mipLevels;
+	image->array_size = pCreateInfo->arrayLayers;
+	image->samples = pCreateInfo->samples;
+	image->tiling = pCreateInfo->tiling;
+	image->usage = pCreateInfo->usage;
+	radv_init_surface(device, &image->surface, create_info);
+
+	device->ws->surface_init(device->ws, &image->surface);
+
+	image->size = image->surface.bo_size;
+	image->alignment = image->surface.bo_alignment;
+
+	if ((pCreateInfo->usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) &&
+	    image->surface.dcc_size)
+		radv_image_alloc_dcc(device, image);
+	else
+		image->surface.dcc_size = 0;
+
+	if ((pCreateInfo->usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) &&
+	    pCreateInfo->mipLevels == 1 &&
+	    !image->surface.dcc_size && image->extent.depth == 1)
+		radv_image_alloc_cmask(device, image);
+	if (image->samples > 1 && vk_format_is_color(pCreateInfo->format)) {
+		radv_image_alloc_fmask(device, image);
+	} else if (vk_format_is_depth(pCreateInfo->format)) {
+
+		radv_image_alloc_htile(device, image);
+	}
+
+
+	if (create_info->stride && create_info->stride != image->surface.level[0].pitch_bytes) {
+		image->surface.level[0].nblk_x = create_info->stride / image->surface.bpe;
+		image->surface.level[0].pitch_bytes = create_info->stride;
+		image->surface.level[0].slice_size = create_info->stride * image->surface.level[0].nblk_y;
+	}
+	*pImage = radv_image_to_handle(image);
+
+	return VK_SUCCESS;
+}
+
+void
+radv_image_view_init(struct radv_image_view *iview,
+		     struct radv_device *device,
+		     const VkImageViewCreateInfo* pCreateInfo,
+		     struct radv_cmd_buffer *cmd_buffer,
+		     VkImageUsageFlags usage_mask)
+{
+	RADV_FROM_HANDLE(radv_image, image, pCreateInfo->image);
+	const VkImageSubresourceRange *range = &pCreateInfo->subresourceRange;
+	bool is_stencil = false;
+	switch (image->type) {
+	default:
+		unreachable("bad VkImageType");
+	case VK_IMAGE_TYPE_1D:
+	case VK_IMAGE_TYPE_2D:
+		assert(range->baseArrayLayer + radv_get_layerCount(image, range) - 1 <= image->array_size);
+		break;
+	case VK_IMAGE_TYPE_3D:
+		assert(range->baseArrayLayer + radv_get_layerCount(image, range) - 1
+		       <= radv_minify(image->extent.depth, range->baseMipLevel));
+		break;
+	}
+	iview->image = image;
+	iview->bo = image->bo;
+	iview->type = pCreateInfo->viewType;
+	iview->vk_format = pCreateInfo->format;
+	iview->aspect_mask = pCreateInfo->subresourceRange.aspectMask;
+
+	if (iview->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT)
+		is_stencil = true;
+	iview->extent = (VkExtent3D) {
+		.width  = radv_minify(image->extent.width , range->baseMipLevel),
+		.height = radv_minify(image->extent.height, range->baseMipLevel),
+		.depth  = radv_minify(image->extent.depth , range->baseMipLevel),
+	};
+
+	iview->extent.width = round_up_u32(iview->extent.width * vk_format_get_blockwidth(iview->vk_format),
+					   vk_format_get_blockwidth(image->vk_format));
+	iview->extent.height = round_up_u32(iview->extent.height * vk_format_get_blockheight(iview->vk_format),
+					    vk_format_get_blockheight(image->vk_format));
+
+	iview->base_layer = range->baseArrayLayer;
+	iview->layer_count = radv_get_layerCount(image, range);
+	iview->base_mip = range->baseMipLevel;
+
+	si_make_texture_descriptor(device, image, false,
+				   iview->type,
+				   pCreateInfo->format,
+				   &pCreateInfo->components,
+				   0, radv_get_levelCount(image, range) - 1,
+				   range->baseArrayLayer,
+				   range->baseArrayLayer + radv_get_layerCount(image, range) - 1,
+				   iview->extent.width,
+				   iview->extent.height,
+				   iview->extent.depth,
+				   iview->descriptor,
+				   iview->fmask_descriptor);
+	si_set_mutable_tex_desc_fields(device, image,
+				       is_stencil ? &image->surface.stencil_level[range->baseMipLevel] : &image->surface.level[range->baseMipLevel], range->baseMipLevel,
+				       range->baseMipLevel,
+				       image->surface.blk_w, is_stencil, iview->descriptor);
+}
+
+void radv_image_set_optimal_micro_tile_mode(struct radv_device *device,
+					    struct radv_image *image, uint32_t micro_tile_mode)
+{
+	/* These magic numbers were copied from addrlib. It doesn't use any
+	 * definitions for them either. They are all 2D_TILED_THIN1 modes with
+	 * different bpp and micro tile mode.
+	 */
+	if (device->instance->physicalDevice.rad_info.chip_class >= CIK) {
+		switch (micro_tile_mode) {
+		case 0: /* displayable */
+			image->surface.tiling_index[0] = 10;
+			break;
+		case 1: /* thin */
+			image->surface.tiling_index[0] = 14;
+			break;
+		case 3: /* rotated */
+			image->surface.tiling_index[0] = 28;
+			break;
+		default: /* depth, thick */
+			assert(!"unexpected micro mode");
+			return;
+		}
+	} else { /* SI */
+		switch (micro_tile_mode) {
+		case 0: /* displayable */
+			switch (image->surface.bpe) {
+			case 8:
+                            image->surface.tiling_index[0] = 10;
+                            break;
+			case 16:
+                            image->surface.tiling_index[0] = 11;
+                            break;
+		default: /* 32, 64 */
+                            image->surface.tiling_index[0] = 12;
+                            break;
+			}
+			break;
+		case 1: /* thin */
+			switch (image->surface.bpe) {
+			case 8:
+                                image->surface.tiling_index[0] = 14;
+                                break;
+			case 16:
+                                image->surface.tiling_index[0] = 15;
+                                break;
+			case 32:
+                                image->surface.tiling_index[0] = 16;
+                                break;
+			default: /* 64, 128 */
+                                image->surface.tiling_index[0] = 17;
+                                break;
+			}
+			break;
+		default: /* depth, thick */
+			assert(!"unexpected micro mode");
+			return;
+		}
+	}
+
+	image->surface.micro_tile_mode = micro_tile_mode;
+}
+
+bool radv_layout_has_htile(const struct radv_image *image,
+                           VkImageLayout layout)
+{
+	return (layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL ||
+		layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
+}
+
+bool radv_layout_is_htile_compressed(const struct radv_image *image,
+                                     VkImageLayout layout)
+{
+	return layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
+}
+
+bool radv_layout_can_expclear(const struct radv_image *image,
+                              VkImageLayout layout)
+{
+	return (layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL ||
+		layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
+}
+
+bool radv_layout_has_cmask(const struct radv_image *image,
+			   VkImageLayout layout)
+{
+	return (layout == VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL ||
+		layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
+}
+
+VkResult
+radv_CreateImage(VkDevice device,
+		 const VkImageCreateInfo *pCreateInfo,
+		 const VkAllocationCallbacks *pAllocator,
+		 VkImage *pImage)
+{
+	return radv_image_create(device,
+				 &(struct radv_image_create_info) {
+					 .vk_info = pCreateInfo,
+						 .scanout = false,
+						 },
+				 pAllocator,
+				 pImage);
+}
+
+void
+radv_DestroyImage(VkDevice _device, VkImage _image,
+		  const VkAllocationCallbacks *pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+
+	if (!_image)
+		return;
+
+	radv_free2(&device->alloc, pAllocator, radv_image_from_handle(_image));
+}
+
+void radv_GetImageSubresourceLayout(
+	VkDevice                                    device,
+	VkImage                                     _image,
+	const VkImageSubresource*                   pSubresource,
+	VkSubresourceLayout*                        pLayout)
+{
+	RADV_FROM_HANDLE(radv_image, image, _image);
+	int level = pSubresource->mipLevel;
+	int layer = pSubresource->arrayLayer;
+
+	pLayout->offset = image->surface.level[level].offset + image->surface.level[level].slice_size * layer;
+	pLayout->rowPitch = image->surface.level[level].pitch_bytes;
+	pLayout->arrayPitch = image->surface.level[level].slice_size;
+	pLayout->depthPitch = image->surface.level[level].slice_size;
+	pLayout->size = image->surface.level[level].slice_size;
+	if (image->type == VK_IMAGE_TYPE_3D)
+		pLayout->size *= image->surface.level[level].nblk_z;
+}
+
+
+VkResult
+radv_CreateImageView(VkDevice _device,
+		     const VkImageViewCreateInfo *pCreateInfo,
+		     const VkAllocationCallbacks *pAllocator,
+		     VkImageView *pView)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_image_view *view;
+
+	view = radv_alloc2(&device->alloc, pAllocator, sizeof(*view), 8,
+			   VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (view == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	radv_image_view_init(view, device, pCreateInfo, NULL, ~0);
+
+	*pView = radv_image_view_to_handle(view);
+
+	return VK_SUCCESS;
+}
+
+void
+radv_DestroyImageView(VkDevice _device, VkImageView _iview,
+		      const VkAllocationCallbacks *pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_image_view, iview, _iview);
+
+	if (!iview)
+		return;
+	radv_free2(&device->alloc, pAllocator, iview);
+}
+
+void radv_buffer_view_init(struct radv_buffer_view *view,
+			   struct radv_device *device,
+			   const VkBufferViewCreateInfo* pCreateInfo,
+			   struct radv_cmd_buffer *cmd_buffer)
+{
+	RADV_FROM_HANDLE(radv_buffer, buffer, pCreateInfo->buffer);
+
+	view->bo = buffer->bo;
+	view->range = pCreateInfo->range == VK_WHOLE_SIZE ?
+		buffer->size - pCreateInfo->offset : pCreateInfo->range;
+	view->vk_format = pCreateInfo->format;
+
+	radv_make_buffer_descriptor(device, buffer, view->vk_format,
+				    pCreateInfo->offset, view->range, view->state);
+}
+
+VkResult
+radv_CreateBufferView(VkDevice _device,
+		      const VkBufferViewCreateInfo *pCreateInfo,
+		      const VkAllocationCallbacks *pAllocator,
+		      VkBufferView *pView)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_buffer_view *view;
+
+	view = radv_alloc2(&device->alloc, pAllocator, sizeof(*view), 8,
+			   VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (!view)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	radv_buffer_view_init(view, device, pCreateInfo, NULL);
+
+	*pView = radv_buffer_view_to_handle(view);
+
+	return VK_SUCCESS;
+}
+
+void
+radv_DestroyBufferView(VkDevice _device, VkBufferView bufferView,
+		       const VkAllocationCallbacks *pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_buffer_view, view, bufferView);
+
+	if (!view)
+		return;
+
+	radv_free2(&device->alloc, pAllocator, view);
+}
diff --git a/src/amd/vulkan/radv_meta.c b/src/amd/vulkan/radv_meta.c
new file mode 100644
index 0000000000..04fa247dd3
--- /dev/null
+++ b/src/amd/vulkan/radv_meta.c
@@ -0,0 +1,388 @@
+/*
+ * Copyright © 2016 Red Hat
+ * based on intel anv code:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "radv_meta.h"
+
+#include <fcntl.h>
+#include <limits.h>
+#include <pwd.h>
+#include <sys/stat.h>
+
+void
+radv_meta_save(struct radv_meta_saved_state *state,
+	       const struct radv_cmd_buffer *cmd_buffer,
+	       uint32_t dynamic_mask)
+{
+	state->old_pipeline = cmd_buffer->state.pipeline;
+	state->old_descriptor_set0 = cmd_buffer->state.descriptors[0];
+	memcpy(state->old_vertex_bindings, cmd_buffer->state.vertex_bindings,
+	       sizeof(state->old_vertex_bindings));
+
+	state->dynamic_mask = dynamic_mask;
+	radv_dynamic_state_copy(&state->dynamic, &cmd_buffer->state.dynamic,
+				dynamic_mask);
+
+	memcpy(state->push_constants, cmd_buffer->push_constants, MAX_PUSH_CONSTANTS_SIZE);
+}
+
+void
+radv_meta_restore(const struct radv_meta_saved_state *state,
+		  struct radv_cmd_buffer *cmd_buffer)
+{
+	cmd_buffer->state.pipeline = state->old_pipeline;
+	radv_bind_descriptor_set(cmd_buffer, state->old_descriptor_set0, 0);
+	memcpy(cmd_buffer->state.vertex_bindings, state->old_vertex_bindings,
+	       sizeof(state->old_vertex_bindings));
+
+	cmd_buffer->state.vb_dirty |= (1 << RADV_META_VERTEX_BINDING_COUNT) - 1;
+	cmd_buffer->state.dirty |= RADV_CMD_DIRTY_PIPELINE;
+
+	radv_dynamic_state_copy(&cmd_buffer->state.dynamic, &state->dynamic,
+				state->dynamic_mask);
+	cmd_buffer->state.dirty |= state->dynamic_mask;
+
+	memcpy(cmd_buffer->push_constants, state->push_constants, MAX_PUSH_CONSTANTS_SIZE);
+	cmd_buffer->push_constant_stages |= VK_SHADER_STAGE_ALL_GRAPHICS | VK_SHADER_STAGE_COMPUTE_BIT;
+}
+
+void
+radv_meta_save_pass(struct radv_meta_saved_pass_state *state,
+                    const struct radv_cmd_buffer *cmd_buffer)
+{
+	state->pass = cmd_buffer->state.pass;
+	state->subpass = cmd_buffer->state.subpass;
+	state->framebuffer = cmd_buffer->state.framebuffer;
+	state->attachments = cmd_buffer->state.attachments;
+	state->render_area = cmd_buffer->state.render_area;
+}
+
+void
+radv_meta_restore_pass(const struct radv_meta_saved_pass_state *state,
+                       struct radv_cmd_buffer *cmd_buffer)
+{
+	cmd_buffer->state.pass = state->pass;
+	cmd_buffer->state.subpass = state->subpass;
+	cmd_buffer->state.framebuffer = state->framebuffer;
+	cmd_buffer->state.attachments = state->attachments;
+	cmd_buffer->state.render_area = state->render_area;
+	if (state->subpass)
+		radv_emit_framebuffer_state(cmd_buffer);
+}
+
+void
+radv_meta_save_compute(struct radv_meta_saved_compute_state *state,
+                       const struct radv_cmd_buffer *cmd_buffer,
+                       unsigned push_constant_size)
+{
+	state->old_pipeline = cmd_buffer->state.compute_pipeline;
+	state->old_descriptor_set0 = cmd_buffer->state.descriptors[0];
+
+	if (push_constant_size)
+		memcpy(state->push_constants, cmd_buffer->push_constants, push_constant_size);
+}
+
+void
+radv_meta_restore_compute(const struct radv_meta_saved_compute_state *state,
+                          struct radv_cmd_buffer *cmd_buffer,
+                          unsigned push_constant_size)
+{
+	radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer), VK_PIPELINE_BIND_POINT_COMPUTE,
+			     radv_pipeline_to_handle(state->old_pipeline));
+	radv_bind_descriptor_set(cmd_buffer, state->old_descriptor_set0, 0);
+
+	if (push_constant_size) {
+		memcpy(cmd_buffer->push_constants, state->push_constants, push_constant_size);
+		cmd_buffer->push_constant_stages |= VK_SHADER_STAGE_COMPUTE_BIT;
+	}
+}
+
+VkImageViewType
+radv_meta_get_view_type(const struct radv_image *image)
+{
+	switch (image->type) {
+	case VK_IMAGE_TYPE_1D: return VK_IMAGE_VIEW_TYPE_1D;
+	case VK_IMAGE_TYPE_2D: return VK_IMAGE_VIEW_TYPE_2D;
+	case VK_IMAGE_TYPE_3D: return VK_IMAGE_VIEW_TYPE_3D;
+	default:
+		unreachable("bad VkImageViewType");
+	}
+}
+
+/**
+ * When creating a destination VkImageView, this function provides the needed
+ * VkImageViewCreateInfo::subresourceRange::baseArrayLayer.
+ */
+uint32_t
+radv_meta_get_iview_layer(const struct radv_image *dest_image,
+			  const VkImageSubresourceLayers *dest_subresource,
+			  const VkOffset3D *dest_offset)
+{
+	switch (dest_image->type) {
+	case VK_IMAGE_TYPE_1D:
+	case VK_IMAGE_TYPE_2D:
+		return dest_subresource->baseArrayLayer;
+	case VK_IMAGE_TYPE_3D:
+		/* HACK: Vulkan does not allow attaching a 3D image to a framebuffer,
+		 * but meta does it anyway. When doing so, we translate the
+		 * destination's z offset into an array offset.
+		 */
+		return dest_offset->z;
+	default:
+		assert(!"bad VkImageType");
+		return 0;
+	}
+}
+
+static void *
+meta_alloc(void* _device, size_t size, size_t alignment,
+           VkSystemAllocationScope allocationScope)
+{
+	struct radv_device *device = _device;
+	return device->alloc.pfnAllocation(device->alloc.pUserData, size, alignment,
+					   VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+}
+
+static void *
+meta_realloc(void* _device, void *original, size_t size, size_t alignment,
+             VkSystemAllocationScope allocationScope)
+{
+	struct radv_device *device = _device;
+	return device->alloc.pfnReallocation(device->alloc.pUserData, original,
+					     size, alignment,
+					     VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+}
+
+static void
+meta_free(void* _device, void *data)
+{
+	struct radv_device *device = _device;
+	return device->alloc.pfnFree(device->alloc.pUserData, data);
+}
+
+static bool
+radv_builtin_cache_path(char *path)
+{
+	char *xdg_cache_home = getenv("XDG_CACHE_HOME");
+	const char *suffix = "/radv_builtin_shaders";
+	const char *suffix2 = "/.cache/radv_builtin_shaders";
+	struct passwd pwd, *result;
+	char path2[PATH_MAX + 1]; /* PATH_MAX is not a real max,but suffices here. */
+
+	if (xdg_cache_home) {
+
+		if (strlen(xdg_cache_home) + strlen(suffix) > PATH_MAX)
+			return false;
+
+		strcpy(path, xdg_cache_home);
+		strcat(path, suffix);
+		return true;
+	}
+
+	getpwuid_r(getuid(), &pwd, path2, PATH_MAX - strlen(suffix2), &result);
+	if (!result)
+		return false;
+
+	strcpy(path, pwd.pw_dir);
+	strcat(path, "/.cache");
+	mkdir(path, 0755);
+
+	strcat(path, suffix);
+	return true;
+}
+
+static void
+radv_load_meta_pipeline(struct radv_device *device)
+{
+	char path[PATH_MAX + 1];
+	struct stat st;
+	void *data = NULL;
+
+	if (!radv_builtin_cache_path(path))
+		return;
+
+	int fd = open(path, O_RDONLY);
+	if (fd < 0)
+		return;
+	if (fstat(fd, &st))
+		goto fail;
+	data = malloc(st.st_size);
+	if (!data)
+		goto fail;
+	if(read(fd, data, st.st_size) == -1)
+		goto fail;
+
+	radv_pipeline_cache_load(&device->meta_state.cache, data, st.st_size);
+fail:
+	free(data);
+	close(fd);
+}
+
+static void
+radv_store_meta_pipeline(struct radv_device *device)
+{
+	char path[PATH_MAX + 1], path2[PATH_MAX + 7];
+	size_t size;
+	void *data = NULL;
+
+	if (!device->meta_state.cache.modified)
+		return;
+
+	if (radv_GetPipelineCacheData(radv_device_to_handle(device),
+				      radv_pipeline_cache_to_handle(&device->meta_state.cache),
+				      &size, NULL))
+		return;
+
+	if (!radv_builtin_cache_path(path))
+		return;
+
+	strcpy(path2, path);
+	strcat(path2, "XXXXXX");
+	int fd = mkstemp(path2);//open(path, O_WRONLY | O_CREAT, 0600);
+	if (fd < 0)
+		return;
+	data = malloc(size);
+	if (!data)
+		goto fail;
+
+	if (radv_GetPipelineCacheData(radv_device_to_handle(device),
+				      radv_pipeline_cache_to_handle(&device->meta_state.cache),
+				      &size, data))
+		goto fail;
+	if(write(fd, data, size) == -1)
+		goto fail;
+
+	rename(path2, path);
+fail:
+	free(data);
+	close(fd);
+	unlink(path2);
+}
+
+VkResult
+radv_device_init_meta(struct radv_device *device)
+{
+	VkResult result;
+
+	device->meta_state.alloc = (VkAllocationCallbacks) {
+		.pUserData = device,
+		.pfnAllocation = meta_alloc,
+		.pfnReallocation = meta_realloc,
+		.pfnFree = meta_free,
+	};
+
+	device->meta_state.cache.alloc = device->meta_state.alloc;
+	radv_pipeline_cache_init(&device->meta_state.cache, device);
+	radv_load_meta_pipeline(device);
+
+	result = radv_device_init_meta_clear_state(device);
+	if (result != VK_SUCCESS)
+		goto fail_clear;
+
+	result = radv_device_init_meta_resolve_state(device);
+	if (result != VK_SUCCESS)
+		goto fail_resolve;
+
+	result = radv_device_init_meta_blit_state(device);
+	if (result != VK_SUCCESS)
+		goto fail_blit;
+
+	result = radv_device_init_meta_blit2d_state(device);
+	if (result != VK_SUCCESS)
+		goto fail_blit2d;
+
+	result = radv_device_init_meta_bufimage_state(device);
+	if (result != VK_SUCCESS)
+		goto fail_bufimage;
+
+	result = radv_device_init_meta_depth_decomp_state(device);
+	if (result != VK_SUCCESS)
+		goto fail_depth_decomp;
+
+	result = radv_device_init_meta_buffer_state(device);
+	if (result != VK_SUCCESS)
+		goto fail_buffer;
+
+	result = radv_device_init_meta_fast_clear_flush_state(device);
+	if (result != VK_SUCCESS)
+		goto fail_fast_clear;
+
+	result = radv_device_init_meta_resolve_compute_state(device);
+	if (result != VK_SUCCESS)
+		goto fail_resolve_compute;
+	return VK_SUCCESS;
+
+fail_resolve_compute:
+	radv_device_finish_meta_fast_clear_flush_state(device);
+fail_fast_clear:
+	radv_device_finish_meta_buffer_state(device);
+fail_buffer:
+	radv_device_finish_meta_depth_decomp_state(device);
+fail_depth_decomp:
+	radv_device_finish_meta_bufimage_state(device);
+fail_bufimage:
+	radv_device_finish_meta_blit2d_state(device);
+fail_blit2d:
+	radv_device_finish_meta_blit_state(device);
+fail_blit:
+	radv_device_finish_meta_resolve_state(device);
+fail_resolve:
+	radv_device_finish_meta_clear_state(device);
+fail_clear:
+	radv_pipeline_cache_finish(&device->meta_state.cache);
+	return result;
+}
+
+void
+radv_device_finish_meta(struct radv_device *device)
+{
+	radv_device_finish_meta_clear_state(device);
+	radv_device_finish_meta_resolve_state(device);
+	radv_device_finish_meta_blit_state(device);
+	radv_device_finish_meta_blit2d_state(device);
+	radv_device_finish_meta_bufimage_state(device);
+	radv_device_finish_meta_depth_decomp_state(device);
+	radv_device_finish_meta_buffer_state(device);
+	radv_device_finish_meta_fast_clear_flush_state(device);
+	radv_device_finish_meta_resolve_compute_state(device);
+
+	radv_store_meta_pipeline(device);
+	radv_pipeline_cache_finish(&device->meta_state.cache);
+}
+
+/*
+ * The most common meta operations all want to have the viewport
+ * reset and any scissors disabled. The rest of the dynamic state
+ * should have no effect.
+ */
+void
+radv_meta_save_graphics_reset_vport_scissor(struct radv_meta_saved_state *saved_state,
+					    struct radv_cmd_buffer *cmd_buffer)
+{
+	uint32_t dirty_state = (1 << VK_DYNAMIC_STATE_VIEWPORT) | (1 << VK_DYNAMIC_STATE_SCISSOR);
+	radv_meta_save(saved_state, cmd_buffer, dirty_state);
+	cmd_buffer->state.dynamic.viewport.count = 0;
+	cmd_buffer->state.dynamic.scissor.count = 0;
+	cmd_buffer->state.dirty |= dirty_state;
+}
diff --git a/src/amd/vulkan/radv_meta.h b/src/amd/vulkan/radv_meta.h
new file mode 100644
index 0000000000..ae63a308b2
--- /dev/null
+++ b/src/amd/vulkan/radv_meta.h
@@ -0,0 +1,190 @@
+/*
+ * Copyright © 2016 Red Hat
+ * based on intel anv code:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#pragma once
+
+#include "radv_private.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RADV_META_VERTEX_BINDING_COUNT 2
+
+struct radv_meta_saved_state {
+	struct radv_vertex_binding old_vertex_bindings[RADV_META_VERTEX_BINDING_COUNT];
+	struct radv_descriptor_set *old_descriptor_set0;
+	struct radv_pipeline *old_pipeline;
+
+	/**
+	 * Bitmask of (1 << VK_DYNAMIC_STATE_*). Defines the set of saved dynamic
+	 * state.
+	 */
+	uint32_t dynamic_mask;
+	struct radv_dynamic_state dynamic;
+
+	char push_constants[128];
+};
+
+struct radv_meta_saved_pass_state {
+	struct radv_render_pass *pass;
+	const struct radv_subpass *subpass;
+	struct radv_attachment_state *attachments;
+	struct radv_framebuffer *framebuffer;
+	VkRect2D render_area;
+};
+
+struct radv_meta_saved_compute_state {
+	struct radv_descriptor_set *old_descriptor_set0;
+	struct radv_pipeline *old_pipeline;
+
+	char push_constants[128];
+};
+
+VkResult radv_device_init_meta_clear_state(struct radv_device *device);
+void radv_device_finish_meta_clear_state(struct radv_device *device);
+
+VkResult radv_device_init_meta_resolve_state(struct radv_device *device);
+void radv_device_finish_meta_resolve_state(struct radv_device *device);
+
+VkResult radv_device_init_meta_depth_decomp_state(struct radv_device *device);
+void radv_device_finish_meta_depth_decomp_state(struct radv_device *device);
+
+VkResult radv_device_init_meta_fast_clear_flush_state(struct radv_device *device);
+void radv_device_finish_meta_fast_clear_flush_state(struct radv_device *device);
+
+VkResult radv_device_init_meta_blit_state(struct radv_device *device);
+void radv_device_finish_meta_blit_state(struct radv_device *device);
+
+VkResult radv_device_init_meta_blit2d_state(struct radv_device *device);
+void radv_device_finish_meta_blit2d_state(struct radv_device *device);
+
+VkResult radv_device_init_meta_buffer_state(struct radv_device *device);
+void radv_device_finish_meta_buffer_state(struct radv_device *device);
+
+VkResult radv_device_init_meta_resolve_compute_state(struct radv_device *device);
+void radv_device_finish_meta_resolve_compute_state(struct radv_device *device);
+void radv_meta_save(struct radv_meta_saved_state *state,
+		    const struct radv_cmd_buffer *cmd_buffer,
+		    uint32_t dynamic_mask);
+
+void radv_meta_restore(const struct radv_meta_saved_state *state,
+		       struct radv_cmd_buffer *cmd_buffer);
+
+void radv_meta_save_pass(struct radv_meta_saved_pass_state *state,
+			 const struct radv_cmd_buffer *cmd_buffer);
+
+void radv_meta_restore_pass(const struct radv_meta_saved_pass_state *state,
+			    struct radv_cmd_buffer *cmd_buffer);
+
+void radv_meta_save_compute(struct radv_meta_saved_compute_state *state,
+			    const struct radv_cmd_buffer *cmd_buffer,
+			    unsigned push_constant_size);
+
+void radv_meta_restore_compute(const struct radv_meta_saved_compute_state *state,
+			       struct radv_cmd_buffer *cmd_buffer,
+			       unsigned push_constant_size);
+
+VkImageViewType radv_meta_get_view_type(const struct radv_image *image);
+
+uint32_t radv_meta_get_iview_layer(const struct radv_image *dest_image,
+				   const VkImageSubresourceLayers *dest_subresource,
+				   const VkOffset3D *dest_offset);
+
+struct radv_meta_blit2d_surf {
+	/** The size of an element in bytes. */
+	uint8_t bs;
+	VkFormat format;
+
+	struct radv_image *image;
+	unsigned level;
+	unsigned layer;
+	VkImageAspectFlags aspect_mask;
+};
+
+struct radv_meta_blit2d_buffer {
+	struct radv_buffer *buffer;
+	uint32_t offset;
+	uint32_t pitch;
+	uint8_t bs;
+	VkFormat format;
+};
+
+struct radv_meta_blit2d_rect {
+	uint32_t src_x, src_y;
+	uint32_t dst_x, dst_y;
+	uint32_t width, height;
+};
+
+void radv_meta_begin_blit2d(struct radv_cmd_buffer *cmd_buffer,
+			    struct radv_meta_saved_state *save);
+
+void radv_meta_blit2d(struct radv_cmd_buffer *cmd_buffer,
+		      struct radv_meta_blit2d_surf *src_img,
+		      struct radv_meta_blit2d_buffer *src_buf,
+		      struct radv_meta_blit2d_surf *dst,
+		      unsigned num_rects,
+		      struct radv_meta_blit2d_rect *rects);
+
+void radv_meta_end_blit2d(struct radv_cmd_buffer *cmd_buffer,
+			  struct radv_meta_saved_state *save);
+
+
+VkResult radv_device_init_meta_bufimage_state(struct radv_device *device);
+void radv_device_finish_meta_bufimage_state(struct radv_device *device);
+void radv_meta_begin_bufimage(struct radv_cmd_buffer *cmd_buffer,
+			      struct radv_meta_saved_compute_state *save);
+void radv_meta_end_bufimage(struct radv_cmd_buffer *cmd_buffer,
+			    struct radv_meta_saved_compute_state *save);
+
+void radv_meta_image_to_buffer(struct radv_cmd_buffer *cmd_buffer,
+			       struct radv_meta_blit2d_surf *src,
+			       struct radv_meta_blit2d_buffer *dst,
+			       unsigned num_rects,
+			       struct radv_meta_blit2d_rect *rects);
+
+void radv_decompress_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
+					 struct radv_image *image,
+					 VkImageSubresourceRange *subresourceRange);
+void radv_resummarize_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
+					  struct radv_image *image,
+					  VkImageSubresourceRange *subresourceRange);
+void radv_fast_clear_flush_image_inplace(struct radv_cmd_buffer *cmd_buffer,
+					 struct radv_image *image);
+
+void radv_meta_save_graphics_reset_vport_scissor(struct radv_meta_saved_state *saved_state,
+						 struct radv_cmd_buffer *cmd_buffer);
+
+void radv_meta_resolve_compute_image(struct radv_cmd_buffer *cmd_buffer,
+				     struct radv_image *src_image,
+				     VkImageLayout src_image_layout,
+				     struct radv_image *dest_image,
+				     VkImageLayout dest_image_layout,
+				     uint32_t region_count,
+				     const VkImageResolve *regions);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/amd/vulkan/radv_meta_blit.c b/src/amd/vulkan/radv_meta_blit.c
new file mode 100644
index 0000000000..497e42f8c0
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_blit.c
@@ -0,0 +1,1270 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "radv_meta.h"
+#include "nir/nir_builder.h"
+
+struct blit_region {
+	VkOffset3D src_offset;
+	VkExtent3D src_extent;
+	VkOffset3D dest_offset;
+	VkExtent3D dest_extent;
+};
+
+static nir_shader *
+build_nir_vertex_shader(void)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, "meta_blit_vs");
+
+	nir_variable *pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						   vec4, "a_pos");
+	pos_in->data.location = VERT_ATTRIB_GENERIC0;
+	nir_variable *pos_out = nir_variable_create(b.shader, nir_var_shader_out,
+						    vec4, "gl_Position");
+	pos_out->data.location = VARYING_SLOT_POS;
+	nir_copy_var(&b, pos_out, pos_in);
+
+	nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						       vec4, "a_tex_pos");
+	tex_pos_in->data.location = VERT_ATTRIB_GENERIC1;
+	nir_variable *tex_pos_out = nir_variable_create(b.shader, nir_var_shader_out,
+							vec4, "v_tex_pos");
+	tex_pos_out->data.location = VARYING_SLOT_VAR0;
+	tex_pos_out->data.interpolation = INTERP_MODE_SMOOTH;
+	nir_copy_var(&b, tex_pos_out, tex_pos_in);
+
+	return b.shader;
+}
+
+static nir_shader *
+build_nir_copy_fragment_shader(enum glsl_sampler_dim tex_dim)
+{
+	char shader_name[64];
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+
+	sprintf(shader_name, "meta_blit_fs.%d", tex_dim);
+	b.shader->info.name = ralloc_strdup(b.shader, shader_name);
+
+	nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						       vec4, "v_tex_pos");
+	tex_pos_in->data.location = VARYING_SLOT_VAR0;
+
+	/* Swizzle the array index which comes in as Z coordinate into the right
+	 * position.
+	 */
+	unsigned swz[] = { 0, (tex_dim == GLSL_SAMPLER_DIM_1D ? 2 : 1), 2 };
+	nir_ssa_def *const tex_pos =
+		nir_swizzle(&b, nir_load_var(&b, tex_pos_in), swz,
+			    (tex_dim == GLSL_SAMPLER_DIM_1D ? 2 : 3), false);
+
+	const struct glsl_type *sampler_type =
+		glsl_sampler_type(tex_dim, false, tex_dim != GLSL_SAMPLER_DIM_3D,
+				  glsl_get_base_type(vec4));
+	nir_variable *sampler = nir_variable_create(b.shader, nir_var_uniform,
+						    sampler_type, "s_tex");
+	sampler->data.descriptor_set = 0;
+	sampler->data.binding = 0;
+
+	nir_tex_instr *tex = nir_tex_instr_create(b.shader, 1);
+	tex->sampler_dim = tex_dim;
+	tex->op = nir_texop_tex;
+	tex->src[0].src_type = nir_tex_src_coord;
+	tex->src[0].src = nir_src_for_ssa(tex_pos);
+	tex->dest_type = nir_type_float; /* TODO */
+	tex->is_array = glsl_sampler_type_is_array(sampler_type);
+	tex->coord_components = tex_pos->num_components;
+	tex->texture = nir_deref_var_create(tex, sampler);
+	tex->sampler = nir_deref_var_create(tex, sampler);
+
+	nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
+	nir_builder_instr_insert(&b, &tex->instr);
+
+	nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
+						      vec4, "f_color");
+	color_out->data.location = FRAG_RESULT_DATA0;
+	nir_store_var(&b, color_out, &tex->dest.ssa, 0xf);
+
+	return b.shader;
+}
+
+static nir_shader *
+build_nir_copy_fragment_shader_depth(enum glsl_sampler_dim tex_dim)
+{
+	char shader_name[64];
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+
+	sprintf(shader_name, "meta_blit_depth_fs.%d", tex_dim);
+	b.shader->info.name = ralloc_strdup(b.shader, shader_name);
+
+	nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						       vec4, "v_tex_pos");
+	tex_pos_in->data.location = VARYING_SLOT_VAR0;
+
+	/* Swizzle the array index which comes in as Z coordinate into the right
+	 * position.
+	 */
+	unsigned swz[] = { 0, (tex_dim == GLSL_SAMPLER_DIM_1D ? 2 : 1), 2 };
+	nir_ssa_def *const tex_pos =
+		nir_swizzle(&b, nir_load_var(&b, tex_pos_in), swz,
+			    (tex_dim == GLSL_SAMPLER_DIM_1D ? 2 : 3), false);
+
+	const struct glsl_type *sampler_type =
+		glsl_sampler_type(tex_dim, false, tex_dim != GLSL_SAMPLER_DIM_3D,
+				  glsl_get_base_type(vec4));
+	nir_variable *sampler = nir_variable_create(b.shader, nir_var_uniform,
+						    sampler_type, "s_tex");
+	sampler->data.descriptor_set = 0;
+	sampler->data.binding = 0;
+
+	nir_tex_instr *tex = nir_tex_instr_create(b.shader, 1);
+	tex->sampler_dim = tex_dim;
+	tex->op = nir_texop_tex;
+	tex->src[0].src_type = nir_tex_src_coord;
+	tex->src[0].src = nir_src_for_ssa(tex_pos);
+	tex->dest_type = nir_type_float; /* TODO */
+	tex->is_array = glsl_sampler_type_is_array(sampler_type);
+	tex->coord_components = tex_pos->num_components;
+	tex->texture = nir_deref_var_create(tex, sampler);
+	tex->sampler = nir_deref_var_create(tex, sampler);
+
+	nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
+	nir_builder_instr_insert(&b, &tex->instr);
+
+	nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
+						      vec4, "f_color");
+	color_out->data.location = FRAG_RESULT_DEPTH;
+	nir_store_var(&b, color_out, &tex->dest.ssa, 0x1);
+
+	return b.shader;
+}
+
+static nir_shader *
+build_nir_copy_fragment_shader_stencil(enum glsl_sampler_dim tex_dim)
+{
+	char shader_name[64];
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+
+	sprintf(shader_name, "meta_blit_stencil_fs.%d", tex_dim);
+	b.shader->info.name = ralloc_strdup(b.shader, shader_name);
+
+	nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						       vec4, "v_tex_pos");
+	tex_pos_in->data.location = VARYING_SLOT_VAR0;
+
+	/* Swizzle the array index which comes in as Z coordinate into the right
+	 * position.
+	 */
+	unsigned swz[] = { 0, (tex_dim == GLSL_SAMPLER_DIM_1D ? 2 : 1), 2 };
+	nir_ssa_def *const tex_pos =
+		nir_swizzle(&b, nir_load_var(&b, tex_pos_in), swz,
+			    (tex_dim == GLSL_SAMPLER_DIM_1D ? 2 : 3), false);
+
+	const struct glsl_type *sampler_type =
+		glsl_sampler_type(tex_dim, false, tex_dim != GLSL_SAMPLER_DIM_3D,
+				  glsl_get_base_type(vec4));
+	nir_variable *sampler = nir_variable_create(b.shader, nir_var_uniform,
+						    sampler_type, "s_tex");
+	sampler->data.descriptor_set = 0;
+	sampler->data.binding = 0;
+
+	nir_tex_instr *tex = nir_tex_instr_create(b.shader, 1);
+	tex->sampler_dim = tex_dim;
+	tex->op = nir_texop_tex;
+	tex->src[0].src_type = nir_tex_src_coord;
+	tex->src[0].src = nir_src_for_ssa(tex_pos);
+	tex->dest_type = nir_type_float; /* TODO */
+	tex->is_array = glsl_sampler_type_is_array(sampler_type);
+	tex->coord_components = tex_pos->num_components;
+	tex->texture = nir_deref_var_create(tex, sampler);
+	tex->sampler = nir_deref_var_create(tex, sampler);
+
+	nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
+	nir_builder_instr_insert(&b, &tex->instr);
+
+	nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
+						      vec4, "f_color");
+	color_out->data.location = FRAG_RESULT_STENCIL;
+	nir_store_var(&b, color_out, &tex->dest.ssa, 0x1);
+
+	return b.shader;
+}
+
+static void
+meta_emit_blit(struct radv_cmd_buffer *cmd_buffer,
+               struct radv_image *src_image,
+               struct radv_image_view *src_iview,
+               VkOffset3D src_offset,
+               VkExtent3D src_extent,
+               struct radv_image *dest_image,
+               struct radv_image_view *dest_iview,
+               VkOffset3D dest_offset,
+               VkExtent3D dest_extent,
+               VkFilter blit_filter)
+{
+	struct radv_device *device = cmd_buffer->device;
+	unsigned offset = 0;
+	struct blit_vb_data {
+		float pos[2];
+		float tex_coord[3];
+	} vb_data[3];
+
+	assert(src_image->samples == dest_image->samples);
+	unsigned vb_size = 3 * sizeof(*vb_data);
+	vb_data[0] = (struct blit_vb_data) {
+		.pos = {
+			dest_offset.x,
+			dest_offset.y,
+		},
+		.tex_coord = {
+			(float)(src_offset.x) / (float)src_iview->extent.width,
+			(float)(src_offset.y) / (float)src_iview->extent.height,
+			(float)src_offset.z / (float)src_iview->extent.depth,
+		},
+	};
+
+	vb_data[1] = (struct blit_vb_data) {
+		.pos = {
+			dest_offset.x,
+			dest_offset.y + dest_extent.height,
+		},
+		.tex_coord = {
+			(float)src_offset.x / (float)src_iview->extent.width,
+			(float)(src_offset.y + src_extent.height) /
+			(float)src_iview->extent.height,
+			(float)src_offset.z / (float)src_iview->extent.depth,
+		},
+	};
+
+	vb_data[2] = (struct blit_vb_data) {
+		.pos = {
+			dest_offset.x + dest_extent.width,
+			dest_offset.y,
+		},
+		.tex_coord = {
+			(float)(src_offset.x + src_extent.width) / (float)src_iview->extent.width,
+			(float)src_offset.y / (float)src_iview->extent.height,
+			(float)src_offset.z / (float)src_iview->extent.depth,
+		},
+	};
+	radv_cmd_buffer_upload_data(cmd_buffer, vb_size, 16, vb_data, &offset);
+
+	struct radv_buffer vertex_buffer = {
+		.device = device,
+		.size = vb_size,
+		.bo = cmd_buffer->upload.upload_bo,
+		.offset = offset,
+	};
+
+	radv_CmdBindVertexBuffers(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1,
+				  (VkBuffer[]) {
+						  radv_buffer_to_handle(&vertex_buffer)
+						  },
+				  (VkDeviceSize[]) {
+					  0,
+						  });
+
+	VkSampler sampler;
+	RADV_CALL(CreateSampler)(radv_device_to_handle(device),
+				 &(VkSamplerCreateInfo) {
+					 .sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO,
+						 .magFilter = blit_filter,
+						 .minFilter = blit_filter,
+						 .addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE,
+						 .addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE,
+						 .addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE,
+						 }, &cmd_buffer->pool->alloc, &sampler);
+
+	VkDescriptorSet set;
+	radv_temp_descriptor_set_create(cmd_buffer->device, cmd_buffer,
+					        device->meta_state.blit.ds_layout,
+					        &set);
+
+	radv_UpdateDescriptorSets(radv_device_to_handle(device),
+				  1, /* writeCount */
+				  (VkWriteDescriptorSet[]) {
+					  {
+						  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+							  .dstSet = set,
+							  .dstBinding = 0,
+							  .dstArrayElement = 0,
+							  .descriptorCount = 1,
+							  .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+							  .pImageInfo = (VkDescriptorImageInfo[]) {
+							  {
+								  .sampler = sampler,
+								  .imageView = radv_image_view_to_handle(src_iview),
+								  .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
+							  },
+						  }
+					  }
+				  }, 0, NULL);
+
+	VkFramebuffer fb;
+	radv_CreateFramebuffer(radv_device_to_handle(device),
+			       &(VkFramebufferCreateInfo) {
+				       .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
+					       .attachmentCount = 1,
+					       .pAttachments = (VkImageView[]) {
+					       radv_image_view_to_handle(dest_iview),
+				       },
+				       .width = dest_iview->extent.width,
+				       .height = dest_iview->extent.height,
+				       .layers = 1,
+				}, &cmd_buffer->pool->alloc, &fb);
+	VkPipeline pipeline;
+	switch (src_iview->aspect_mask) {
+	case VK_IMAGE_ASPECT_COLOR_BIT: {
+		unsigned fs_key = radv_format_meta_fs_key(dest_image->vk_format);
+
+		RADV_CALL(CmdBeginRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer),
+					      &(VkRenderPassBeginInfo) {
+						      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+							      .renderPass = device->meta_state.blit.render_pass[fs_key],
+							      .framebuffer = fb,
+							      .renderArea = {
+							      .offset = { dest_offset.x, dest_offset.y },
+							      .extent = { dest_extent.width, dest_extent.height },
+						      },
+							      .clearValueCount = 0,
+								       .pClearValues = NULL,
+						       }, VK_SUBPASS_CONTENTS_INLINE);
+		switch (src_image->type) {
+		case VK_IMAGE_TYPE_1D:
+			pipeline = device->meta_state.blit.pipeline_1d_src[fs_key];
+			break;
+		case VK_IMAGE_TYPE_2D:
+			pipeline = device->meta_state.blit.pipeline_2d_src[fs_key];
+			break;
+		case VK_IMAGE_TYPE_3D:
+			pipeline = device->meta_state.blit.pipeline_3d_src[fs_key];
+			break;
+		default:
+			unreachable(!"bad VkImageType");
+		}
+		break;
+	}
+	case VK_IMAGE_ASPECT_DEPTH_BIT:
+		RADV_CALL(CmdBeginRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer),
+					      &(VkRenderPassBeginInfo) {
+						      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+							      .renderPass = device->meta_state.blit.depth_only_rp,
+							      .framebuffer = fb,
+							      .renderArea = {
+							      .offset = { dest_offset.x, dest_offset.y },
+							      .extent = { dest_extent.width, dest_extent.height },
+						      },
+							      .clearValueCount = 0,
+								       .pClearValues = NULL,
+						       }, VK_SUBPASS_CONTENTS_INLINE);
+		switch (src_image->type) {
+		case VK_IMAGE_TYPE_1D:
+			pipeline = device->meta_state.blit.depth_only_1d_pipeline;
+			break;
+		case VK_IMAGE_TYPE_2D:
+			pipeline = device->meta_state.blit.depth_only_2d_pipeline;
+			break;
+		case VK_IMAGE_TYPE_3D:
+			pipeline = device->meta_state.blit.depth_only_3d_pipeline;
+			break;
+		default:
+			unreachable(!"bad VkImageType");
+		}
+		break;
+	case VK_IMAGE_ASPECT_STENCIL_BIT:
+		RADV_CALL(CmdBeginRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer),
+					      &(VkRenderPassBeginInfo) {
+						      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+							      .renderPass = device->meta_state.blit.stencil_only_rp,
+							      .framebuffer = fb,
+							      .renderArea = {
+							      .offset = { dest_offset.x, dest_offset.y },
+							      .extent = { dest_extent.width, dest_extent.height },
+						      },
+							      .clearValueCount = 0,
+								       .pClearValues = NULL,
+						       }, VK_SUBPASS_CONTENTS_INLINE);
+		switch (src_image->type) {
+		case VK_IMAGE_TYPE_1D:
+			pipeline = device->meta_state.blit.stencil_only_1d_pipeline;
+			break;
+		case VK_IMAGE_TYPE_2D:
+			pipeline = device->meta_state.blit.stencil_only_2d_pipeline;
+			break;
+		case VK_IMAGE_TYPE_3D:
+			pipeline = device->meta_state.blit.stencil_only_3d_pipeline;
+			break;
+		default:
+			unreachable(!"bad VkImageType");
+		}
+		break;
+	default:
+		unreachable(!"bad VkImageType");
+	}
+
+	if (cmd_buffer->state.pipeline != radv_pipeline_from_handle(pipeline)) {
+		radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+				     VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+	}
+
+	radv_CmdBindDescriptorSets(radv_cmd_buffer_to_handle(cmd_buffer),
+				   VK_PIPELINE_BIND_POINT_GRAPHICS,
+				   device->meta_state.blit.pipeline_layout, 0, 1,
+				   &set, 0, NULL);
+
+	RADV_CALL(CmdDraw)(radv_cmd_buffer_to_handle(cmd_buffer), 3, 1, 0, 0);
+
+	RADV_CALL(CmdEndRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer));
+
+	/* At the point where we emit the draw call, all data from the
+	 * descriptor sets, etc. has been used.  We are free to delete it.
+	 */
+	/* TODO: above comment is not valid for at least descriptor sets/pools,
+	 * as we may not free them till after execution finishes. Check others. */
+
+	radv_temp_descriptor_set_destroy(cmd_buffer->device, set);
+	radv_DestroySampler(radv_device_to_handle(device), sampler,
+			    &cmd_buffer->pool->alloc);
+	radv_DestroyFramebuffer(radv_device_to_handle(device), fb,
+				&cmd_buffer->pool->alloc);
+}
+
+void radv_CmdBlitImage(
+	VkCommandBuffer                             commandBuffer,
+	VkImage                                     srcImage,
+	VkImageLayout                               srcImageLayout,
+	VkImage                                     destImage,
+	VkImageLayout                               destImageLayout,
+	uint32_t                                    regionCount,
+	const VkImageBlit*                          pRegions,
+	VkFilter                                    filter)
+
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_image, src_image, srcImage);
+	RADV_FROM_HANDLE(radv_image, dest_image, destImage);
+	struct radv_meta_saved_state saved_state;
+
+	/* From the Vulkan 1.0 spec:
+	 *
+	 *    vkCmdBlitImage must not be used for multisampled source or
+	 *    destination images. Use vkCmdResolveImage for this purpose.
+	 */
+	assert(src_image->samples == 1);
+	assert(dest_image->samples == 1);
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	for (unsigned r = 0; r < regionCount; r++) {
+		struct radv_image_view src_iview;
+		radv_image_view_init(&src_iview, cmd_buffer->device,
+				     &(VkImageViewCreateInfo) {
+					     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+						     .image = srcImage,
+						     .viewType = radv_meta_get_view_type(src_image),
+						     .format = src_image->vk_format,
+						     .subresourceRange = {
+						     .aspectMask = pRegions[r].srcSubresource.aspectMask,
+						     .baseMipLevel = pRegions[r].srcSubresource.mipLevel,
+						     .levelCount = 1,
+						     .baseArrayLayer = pRegions[r].srcSubresource.baseArrayLayer,
+						     .layerCount = 1
+					     },
+						     },
+				     cmd_buffer, VK_IMAGE_USAGE_SAMPLED_BIT);
+
+		if (pRegions[r].dstOffsets[1].x < pRegions[r].dstOffsets[0].x ||
+		    pRegions[r].dstOffsets[1].y < pRegions[r].dstOffsets[0].y ||
+		    pRegions[r].srcOffsets[1].x < pRegions[r].srcOffsets[0].x ||
+		    pRegions[r].srcOffsets[1].y < pRegions[r].srcOffsets[0].y)
+			radv_finishme("FINISHME: Allow flipping in blits");
+
+		const VkExtent3D dest_extent = {
+			.width = pRegions[r].dstOffsets[1].x - pRegions[r].dstOffsets[0].x,
+			.height = pRegions[r].dstOffsets[1].y - pRegions[r].dstOffsets[0].y,
+			.depth = 1,
+		};
+
+		const VkExtent3D src_extent = {
+			.width = pRegions[r].srcOffsets[1].x - pRegions[r].srcOffsets[0].x,
+			.height = pRegions[r].srcOffsets[1].y - pRegions[r].srcOffsets[0].y,
+			.depth = pRegions[r].srcOffsets[1].z - pRegions[r].srcOffsets[0].z,
+		};
+
+
+		if (pRegions[r].srcSubresource.layerCount > 1)
+			radv_finishme("FINISHME: copy multiple array layers");
+
+		struct radv_image_view dest_iview;
+		unsigned usage;
+		if (pRegions[r].dstSubresource.aspectMask == VK_IMAGE_ASPECT_COLOR_BIT)
+			usage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
+		else
+			usage = VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
+
+		for (unsigned i = pRegions[r].dstOffsets[0].z; i < pRegions[r].dstOffsets[1].z; i++) {
+
+			const VkOffset3D dest_offset = {
+				.x = pRegions[r].dstOffsets[0].x,
+				.y = pRegions[r].dstOffsets[0].y,
+				.z = i,
+			};
+			VkOffset3D src_offset = {
+				.x = pRegions[r].srcOffsets[0].x,
+				.y = pRegions[r].srcOffsets[0].y,
+				.z = i,
+			};
+			const uint32_t dest_array_slice =
+				radv_meta_get_iview_layer(dest_image, &pRegions[r].dstSubresource,
+							  &dest_offset);
+
+			radv_image_view_init(&dest_iview, cmd_buffer->device,
+					     &(VkImageViewCreateInfo) {
+						     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+							     .image = destImage,
+							     .viewType = radv_meta_get_view_type(dest_image),
+							     .format = dest_image->vk_format,
+							     .subresourceRange = {
+							     .aspectMask = pRegions[r].dstSubresource.aspectMask,
+							     .baseMipLevel = pRegions[r].dstSubresource.mipLevel,
+							     .levelCount = 1,
+							     .baseArrayLayer = dest_array_slice,
+							     .layerCount = 1
+						     },
+					     },
+					     cmd_buffer, usage);
+			meta_emit_blit(cmd_buffer,
+				       src_image, &src_iview,
+				       src_offset, src_extent,
+				       dest_image, &dest_iview,
+				       dest_offset, dest_extent,
+				       filter);
+		}
+	}
+
+	radv_meta_restore(&saved_state, cmd_buffer);
+}
+
+void
+radv_device_finish_meta_blit_state(struct radv_device *device)
+{
+	for (unsigned i = 0; i < NUM_META_FS_KEYS; ++i) {
+		if (device->meta_state.blit.render_pass[i])
+			radv_DestroyRenderPass(radv_device_to_handle(device),
+					       device->meta_state.blit.render_pass[i],
+					       &device->meta_state.alloc);
+		if (device->meta_state.blit.pipeline_1d_src[i])
+			radv_DestroyPipeline(radv_device_to_handle(device),
+					     device->meta_state.blit.pipeline_1d_src[i],
+					     &device->meta_state.alloc);
+		if (device->meta_state.blit.pipeline_2d_src[i])
+			radv_DestroyPipeline(radv_device_to_handle(device),
+					     device->meta_state.blit.pipeline_2d_src[i],
+					     &device->meta_state.alloc);
+		if (device->meta_state.blit.pipeline_3d_src[i])
+			radv_DestroyPipeline(radv_device_to_handle(device),
+					     device->meta_state.blit.pipeline_3d_src[i],
+					     &device->meta_state.alloc);
+	}
+
+	if (device->meta_state.blit.depth_only_rp)
+		radv_DestroyRenderPass(radv_device_to_handle(device),
+				       device->meta_state.blit.depth_only_rp,
+				       &device->meta_state.alloc);
+	if (device->meta_state.blit.depth_only_1d_pipeline)
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.blit.depth_only_1d_pipeline,
+				     &device->meta_state.alloc);
+	if (device->meta_state.blit.depth_only_2d_pipeline)
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.blit.depth_only_2d_pipeline,
+				     &device->meta_state.alloc);
+	if (device->meta_state.blit.depth_only_3d_pipeline)
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.blit.depth_only_3d_pipeline,
+				     &device->meta_state.alloc);
+	if (device->meta_state.blit.stencil_only_rp)
+		radv_DestroyRenderPass(radv_device_to_handle(device),
+				       device->meta_state.blit.stencil_only_rp,
+				       &device->meta_state.alloc);
+	if (device->meta_state.blit.stencil_only_1d_pipeline)
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.blit.stencil_only_1d_pipeline,
+				     &device->meta_state.alloc);
+	if (device->meta_state.blit.stencil_only_2d_pipeline)
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.blit.stencil_only_2d_pipeline,
+				     &device->meta_state.alloc);
+	if (device->meta_state.blit.stencil_only_3d_pipeline)
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.blit.stencil_only_3d_pipeline,
+				     &device->meta_state.alloc);
+	if (device->meta_state.blit.pipeline_layout)
+		radv_DestroyPipelineLayout(radv_device_to_handle(device),
+					   device->meta_state.blit.pipeline_layout,
+					   &device->meta_state.alloc);
+	if (device->meta_state.blit.ds_layout)
+		radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
+						device->meta_state.blit.ds_layout,
+						&device->meta_state.alloc);
+}
+
+static VkFormat pipeline_formats[] = {
+   VK_FORMAT_R8G8B8A8_UNORM,
+   VK_FORMAT_R8G8B8A8_UINT,
+   VK_FORMAT_R8G8B8A8_SINT,
+   VK_FORMAT_R16G16B16A16_UNORM,
+   VK_FORMAT_R16G16B16A16_SNORM,
+   VK_FORMAT_R16G16B16A16_UINT,
+   VK_FORMAT_R16G16B16A16_SINT,
+   VK_FORMAT_R32_SFLOAT,
+   VK_FORMAT_R32G32_SFLOAT,
+   VK_FORMAT_R32G32B32A32_SFLOAT
+};
+
+static VkResult
+radv_device_init_meta_blit_color(struct radv_device *device,
+				 struct radv_shader_module *vs)
+{
+	struct radv_shader_module fs_1d = {0}, fs_2d = {0}, fs_3d = {0};
+	VkResult result;
+
+	fs_1d.nir = build_nir_copy_fragment_shader(GLSL_SAMPLER_DIM_1D);
+	fs_2d.nir = build_nir_copy_fragment_shader(GLSL_SAMPLER_DIM_2D);
+	fs_3d.nir = build_nir_copy_fragment_shader(GLSL_SAMPLER_DIM_3D);
+
+	for (unsigned i = 0; i < ARRAY_SIZE(pipeline_formats); ++i) {
+		unsigned key = radv_format_meta_fs_key(pipeline_formats[i]);
+		result = radv_CreateRenderPass(radv_device_to_handle(device),
+					&(VkRenderPassCreateInfo) {
+						.sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+							.attachmentCount = 1,
+							.pAttachments = &(VkAttachmentDescription) {
+							.format = pipeline_formats[i],
+							.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+							.storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+							.initialLayout = VK_IMAGE_LAYOUT_GENERAL,
+							.finalLayout = VK_IMAGE_LAYOUT_GENERAL,
+						},
+							.subpassCount = 1,
+									.pSubpasses = &(VkSubpassDescription) {
+							.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+							.inputAttachmentCount = 0,
+							.colorAttachmentCount = 1,
+							.pColorAttachments = &(VkAttachmentReference) {
+								.attachment = 0,
+								.layout = VK_IMAGE_LAYOUT_GENERAL,
+							},
+							.pResolveAttachments = NULL,
+							.pDepthStencilAttachment = &(VkAttachmentReference) {
+								.attachment = VK_ATTACHMENT_UNUSED,
+								.layout = VK_IMAGE_LAYOUT_GENERAL,
+							},
+							.preserveAttachmentCount = 1,
+							.pPreserveAttachments = (uint32_t[]) { 0 },
+						},
+						.dependencyCount = 0,
+					}, &device->meta_state.alloc, &device->meta_state.blit.render_pass[key]);
+		if (result != VK_SUCCESS)
+			goto fail;
+
+		VkPipelineVertexInputStateCreateInfo vi_create_info = {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+			.vertexBindingDescriptionCount = 1,
+			.pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
+				{
+					.binding = 0,
+					.stride = 5 * sizeof(float),
+					.inputRate = VK_VERTEX_INPUT_RATE_VERTEX
+				},
+			},
+			.vertexAttributeDescriptionCount = 2,
+			.pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
+				{
+					/* Position */
+					.location = 0,
+					.binding = 0,
+					.format = VK_FORMAT_R32G32_SFLOAT,
+					.offset = 0
+				},
+				{
+					/* Texture Coordinate */
+					.location = 1,
+					.binding = 0,
+					.format = VK_FORMAT_R32G32B32_SFLOAT,
+					.offset = 8
+				}
+			}
+		};
+
+		VkPipelineShaderStageCreateInfo pipeline_shader_stages[] = {
+			{
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+				.stage = VK_SHADER_STAGE_VERTEX_BIT,
+				.module = radv_shader_module_to_handle(vs),
+				.pName = "main",
+				.pSpecializationInfo = NULL
+			}, {
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+				.stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+				.module = VK_NULL_HANDLE, /* TEMPLATE VALUE! FILL ME IN! */
+				.pName = "main",
+				.pSpecializationInfo = NULL
+			},
+		};
+
+		const VkGraphicsPipelineCreateInfo vk_pipeline_info = {
+			.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+			.stageCount = ARRAY_SIZE(pipeline_shader_stages),
+			.pStages = pipeline_shader_stages,
+			.pVertexInputState = &vi_create_info,
+			.pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+				.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+				.primitiveRestartEnable = false,
+			},
+			.pViewportState = &(VkPipelineViewportStateCreateInfo) {
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+				.viewportCount = 0,
+				.scissorCount = 0,
+			},
+			.pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+				.rasterizerDiscardEnable = false,
+				.polygonMode = VK_POLYGON_MODE_FILL,
+				.cullMode = VK_CULL_MODE_NONE,
+				.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE
+			},
+			.pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+				.rasterizationSamples = 1,
+				.sampleShadingEnable = false,
+				.pSampleMask = (VkSampleMask[]) { UINT32_MAX },
+			},
+			.pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+				.attachmentCount = 1,
+				.pAttachments = (VkPipelineColorBlendAttachmentState []) {
+					{ .colorWriteMask =
+					VK_COLOR_COMPONENT_A_BIT |
+					VK_COLOR_COMPONENT_R_BIT |
+					VK_COLOR_COMPONENT_G_BIT |
+					VK_COLOR_COMPONENT_B_BIT },
+				}
+			},
+			.pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+				.dynamicStateCount = 2,
+				.pDynamicStates = (VkDynamicState[]) {
+					VK_DYNAMIC_STATE_LINE_WIDTH,
+					VK_DYNAMIC_STATE_BLEND_CONSTANTS,
+				},
+			},
+			.flags = 0,
+			.layout = device->meta_state.blit.pipeline_layout,
+			.renderPass = device->meta_state.blit.render_pass[key],
+			.subpass = 0,
+		};
+
+		const struct radv_graphics_pipeline_create_info radv_pipeline_info = {
+			.use_rectlist = true
+		};
+
+		pipeline_shader_stages[1].module = radv_shader_module_to_handle(&fs_1d);
+		result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+						radv_pipeline_cache_to_handle(&device->meta_state.cache),
+						&vk_pipeline_info, &radv_pipeline_info,
+						&device->meta_state.alloc, &device->meta_state.blit.pipeline_1d_src[key]);
+		if (result != VK_SUCCESS)
+			goto fail;
+
+		pipeline_shader_stages[1].module = radv_shader_module_to_handle(&fs_2d);
+		result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+						radv_pipeline_cache_to_handle(&device->meta_state.cache),
+						&vk_pipeline_info, &radv_pipeline_info,
+						&device->meta_state.alloc, &device->meta_state.blit.pipeline_2d_src[key]);
+		if (result != VK_SUCCESS)
+			goto fail;
+
+		pipeline_shader_stages[1].module = radv_shader_module_to_handle(&fs_3d);
+		result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+						radv_pipeline_cache_to_handle(&device->meta_state.cache),
+						&vk_pipeline_info, &radv_pipeline_info,
+						&device->meta_state.alloc, &device->meta_state.blit.pipeline_3d_src[key]);
+		if (result != VK_SUCCESS)
+			goto fail;
+
+	}
+
+	result = VK_SUCCESS;
+fail:
+	ralloc_free(fs_1d.nir);
+	ralloc_free(fs_2d.nir);
+	ralloc_free(fs_3d.nir);
+	return result;
+}
+
+static VkResult
+radv_device_init_meta_blit_depth(struct radv_device *device,
+				 struct radv_shader_module *vs)
+{
+	struct radv_shader_module fs_1d = {0}, fs_2d = {0}, fs_3d = {0};
+	VkResult result;
+
+	fs_1d.nir = build_nir_copy_fragment_shader_depth(GLSL_SAMPLER_DIM_1D);
+	fs_2d.nir = build_nir_copy_fragment_shader_depth(GLSL_SAMPLER_DIM_2D);
+	fs_3d.nir = build_nir_copy_fragment_shader_depth(GLSL_SAMPLER_DIM_3D);
+
+	result = radv_CreateRenderPass(radv_device_to_handle(device),
+				       &(VkRenderPassCreateInfo) {
+					       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+						       .attachmentCount = 1,
+						       .pAttachments = &(VkAttachmentDescription) {
+						       .format = 0,
+						       .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+						       .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+						       .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
+						       .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
+					       },
+						       .subpassCount = 1,
+						.pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 0,
+						       .pColorAttachments = NULL,
+						       .pResolveAttachments = NULL,
+						       .pDepthStencilAttachment = &(VkAttachmentReference) {
+							       .attachment = 0,
+							       .layout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+						       .preserveAttachmentCount = 1,
+						       .pPreserveAttachments = (uint32_t[]) { 0 },
+					       },
+						.dependencyCount = 0,
+					 }, &device->meta_state.alloc, &device->meta_state.blit.depth_only_rp);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	VkPipelineVertexInputStateCreateInfo vi_create_info = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+		.vertexBindingDescriptionCount = 1,
+		.pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
+			{
+				.binding = 0,
+				.stride = 5 * sizeof(float),
+				.inputRate = VK_VERTEX_INPUT_RATE_VERTEX
+			},
+		},
+		.vertexAttributeDescriptionCount = 2,
+		.pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
+			{
+				/* Position */
+				.location = 0,
+				.binding = 0,
+				.format = VK_FORMAT_R32G32_SFLOAT,
+				.offset = 0
+			},
+			{
+				/* Texture Coordinate */
+				.location = 1,
+				.binding = 0,
+				.format = VK_FORMAT_R32G32B32_SFLOAT,
+				.offset = 8
+			}
+		}
+	};
+
+	VkPipelineShaderStageCreateInfo pipeline_shader_stages[] = {
+		{
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_VERTEX_BIT,
+			.module = radv_shader_module_to_handle(vs),
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		}, {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+			.module = VK_NULL_HANDLE, /* TEMPLATE VALUE! FILL ME IN! */
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		},
+	};
+
+	const VkGraphicsPipelineCreateInfo vk_pipeline_info = {
+		.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+		.stageCount = ARRAY_SIZE(pipeline_shader_stages),
+		.pStages = pipeline_shader_stages,
+		.pVertexInputState = &vi_create_info,
+		.pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+			.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+			.primitiveRestartEnable = false,
+		},
+		.pViewportState = &(VkPipelineViewportStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+			.viewportCount = 0,
+			.scissorCount = 0,
+		},
+		.pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+			.rasterizerDiscardEnable = false,
+			.polygonMode = VK_POLYGON_MODE_FILL,
+			.cullMode = VK_CULL_MODE_NONE,
+			.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE
+		},
+		.pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+			.rasterizationSamples = 1,
+			.sampleShadingEnable = false,
+			.pSampleMask = (VkSampleMask[]) { UINT32_MAX },
+		},
+		.pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+			.attachmentCount = 0,
+			.pAttachments = NULL,
+		},
+		.pDepthStencilState = &(VkPipelineDepthStencilStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
+			.depthTestEnable = true,
+			.depthWriteEnable = true,
+			.depthCompareOp = VK_COMPARE_OP_ALWAYS,
+		},
+		.pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+			.dynamicStateCount = 7,
+			.pDynamicStates = (VkDynamicState[]) {
+				VK_DYNAMIC_STATE_LINE_WIDTH,
+				VK_DYNAMIC_STATE_DEPTH_BIAS,
+				VK_DYNAMIC_STATE_BLEND_CONSTANTS,
+				VK_DYNAMIC_STATE_DEPTH_BOUNDS,
+				VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK,
+				VK_DYNAMIC_STATE_STENCIL_WRITE_MASK,
+				VK_DYNAMIC_STATE_STENCIL_REFERENCE,
+			},
+		},
+		.flags = 0,
+		.layout = device->meta_state.blit.pipeline_layout,
+		.renderPass = device->meta_state.blit.depth_only_rp,
+		.subpass = 0,
+	};
+
+	const struct radv_graphics_pipeline_create_info radv_pipeline_info = {
+		.use_rectlist = true
+	};
+
+	pipeline_shader_stages[1].module = radv_shader_module_to_handle(&fs_1d);
+	result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &vk_pipeline_info, &radv_pipeline_info,
+					       &device->meta_state.alloc, &device->meta_state.blit.depth_only_1d_pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	pipeline_shader_stages[1].module = radv_shader_module_to_handle(&fs_2d);
+	result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &vk_pipeline_info, &radv_pipeline_info,
+					       &device->meta_state.alloc, &device->meta_state.blit.depth_only_2d_pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	pipeline_shader_stages[1].module = radv_shader_module_to_handle(&fs_3d);
+	result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &vk_pipeline_info, &radv_pipeline_info,
+					       &device->meta_state.alloc, &device->meta_state.blit.depth_only_3d_pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+fail:
+	ralloc_free(fs_1d.nir);
+	ralloc_free(fs_2d.nir);
+	ralloc_free(fs_3d.nir);
+	return result;
+}
+
+static VkResult
+radv_device_init_meta_blit_stencil(struct radv_device *device,
+				   struct radv_shader_module *vs)
+{
+	struct radv_shader_module fs_1d = {0}, fs_2d = {0}, fs_3d = {0};
+	VkResult result;
+
+	fs_1d.nir = build_nir_copy_fragment_shader_stencil(GLSL_SAMPLER_DIM_1D);
+	fs_2d.nir = build_nir_copy_fragment_shader_stencil(GLSL_SAMPLER_DIM_2D);
+	fs_3d.nir = build_nir_copy_fragment_shader_stencil(GLSL_SAMPLER_DIM_3D);
+
+	result = radv_CreateRenderPass(radv_device_to_handle(device),
+				       &(VkRenderPassCreateInfo) {
+					       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+						       .attachmentCount = 1,
+						       .pAttachments = &(VkAttachmentDescription) {
+						       .format = 0,
+						       .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+						       .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+						       .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
+						       .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
+					       },
+						       .subpassCount = 1,
+						.pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 0,
+						       .pColorAttachments = NULL,
+						       .pResolveAttachments = NULL,
+						       .pDepthStencilAttachment = &(VkAttachmentReference) {
+							       .attachment = 0,
+							       .layout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+						       .preserveAttachmentCount = 1,
+						       .pPreserveAttachments = (uint32_t[]) { 0 },
+					       },
+						.dependencyCount = 0,
+					 }, &device->meta_state.alloc, &device->meta_state.blit.stencil_only_rp);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	VkPipelineVertexInputStateCreateInfo vi_create_info = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+		.vertexBindingDescriptionCount = 1,
+		.pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
+			{
+				.binding = 0,
+				.stride = 5 * sizeof(float),
+				.inputRate = VK_VERTEX_INPUT_RATE_VERTEX
+			},
+		},
+		.vertexAttributeDescriptionCount = 2,
+		.pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
+			{
+				/* Position */
+				.location = 0,
+				.binding = 0,
+				.format = VK_FORMAT_R32G32_SFLOAT,
+				.offset = 0
+			},
+			{
+				/* Texture Coordinate */
+				.location = 1,
+				.binding = 0,
+				.format = VK_FORMAT_R32G32B32_SFLOAT,
+				.offset = 8
+			}
+		}
+	};
+
+	VkPipelineShaderStageCreateInfo pipeline_shader_stages[] = {
+		{
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_VERTEX_BIT,
+			.module = radv_shader_module_to_handle(vs),
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		}, {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+			.module = VK_NULL_HANDLE, /* TEMPLATE VALUE! FILL ME IN! */
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		},
+	};
+
+	const VkGraphicsPipelineCreateInfo vk_pipeline_info = {
+		.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+		.stageCount = ARRAY_SIZE(pipeline_shader_stages),
+		.pStages = pipeline_shader_stages,
+		.pVertexInputState = &vi_create_info,
+		.pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+			.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+			.primitiveRestartEnable = false,
+		},
+		.pViewportState = &(VkPipelineViewportStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+			.viewportCount = 0,
+			.scissorCount = 0,
+		},
+		.pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+			.rasterizerDiscardEnable = false,
+			.polygonMode = VK_POLYGON_MODE_FILL,
+			.cullMode = VK_CULL_MODE_NONE,
+			.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE
+		},
+		.pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+			.rasterizationSamples = 1,
+			.sampleShadingEnable = false,
+			.pSampleMask = (VkSampleMask[]) { UINT32_MAX },
+		},
+		.pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+			.attachmentCount = 0,
+			.pAttachments = NULL,
+		},
+		.pDepthStencilState = &(VkPipelineDepthStencilStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
+			.depthTestEnable = false,
+			.depthWriteEnable = false,
+			.stencilTestEnable = true,
+			.front = {
+				.failOp = VK_STENCIL_OP_REPLACE,
+				.passOp = VK_STENCIL_OP_REPLACE,
+				.depthFailOp = VK_STENCIL_OP_REPLACE,
+				.compareOp = VK_COMPARE_OP_ALWAYS,
+				.compareMask = 0xff,
+				.writeMask = 0xff,
+				.reference = 0
+			},
+			.back = {
+				.failOp = VK_STENCIL_OP_REPLACE,
+				.passOp = VK_STENCIL_OP_REPLACE,
+				.depthFailOp = VK_STENCIL_OP_REPLACE,
+				.compareOp = VK_COMPARE_OP_ALWAYS,
+				.compareMask = 0xff,
+				.writeMask = 0xff,
+				.reference = 0
+			},
+			.depthCompareOp = VK_COMPARE_OP_ALWAYS,
+		},
+
+		.pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+			.dynamicStateCount = 4,
+			.pDynamicStates = (VkDynamicState[]) {
+				VK_DYNAMIC_STATE_LINE_WIDTH,
+				VK_DYNAMIC_STATE_DEPTH_BIAS,
+				VK_DYNAMIC_STATE_BLEND_CONSTANTS,
+				VK_DYNAMIC_STATE_DEPTH_BOUNDS,
+			},
+		},
+		.flags = 0,
+		.layout = device->meta_state.blit.pipeline_layout,
+		.renderPass = device->meta_state.blit.stencil_only_rp,
+		.subpass = 0,
+	};
+
+	const struct radv_graphics_pipeline_create_info radv_pipeline_info = {
+		.use_rectlist = true
+	};
+
+	pipeline_shader_stages[1].module = radv_shader_module_to_handle(&fs_1d);
+	result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &vk_pipeline_info, &radv_pipeline_info,
+					       &device->meta_state.alloc, &device->meta_state.blit.stencil_only_1d_pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	pipeline_shader_stages[1].module = radv_shader_module_to_handle(&fs_2d);
+	result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &vk_pipeline_info, &radv_pipeline_info,
+					       &device->meta_state.alloc, &device->meta_state.blit.stencil_only_2d_pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	pipeline_shader_stages[1].module = radv_shader_module_to_handle(&fs_3d);
+	result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &vk_pipeline_info, &radv_pipeline_info,
+					       &device->meta_state.alloc, &device->meta_state.blit.stencil_only_3d_pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+fail:
+	ralloc_free(fs_1d.nir);
+	ralloc_free(fs_2d.nir);
+	ralloc_free(fs_3d.nir);
+	return result;
+}
+
+VkResult
+radv_device_init_meta_blit_state(struct radv_device *device)
+{
+	VkResult result;
+	struct radv_shader_module vs = {0};
+	zero(device->meta_state.blit);
+
+	VkDescriptorSetLayoutCreateInfo ds_layout_info = {
+		.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+		.bindingCount = 1,
+		.pBindings = (VkDescriptorSetLayoutBinding[]) {
+			{
+				.binding = 0,
+				.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+				.descriptorCount = 1,
+				.stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT,
+				.pImmutableSamplers = NULL
+			},
+		}
+	};
+	result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
+						&ds_layout_info,
+						&device->meta_state.alloc,
+						&device->meta_state.blit.ds_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	result = radv_CreatePipelineLayout(radv_device_to_handle(device),
+					   &(VkPipelineLayoutCreateInfo) {
+						   .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+							   .setLayoutCount = 1,
+							   .pSetLayouts = &device->meta_state.blit.ds_layout,
+							   },
+					   &device->meta_state.alloc, &device->meta_state.blit.pipeline_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	vs.nir = build_nir_vertex_shader();
+
+	result = radv_device_init_meta_blit_color(device, &vs);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	result = radv_device_init_meta_blit_depth(device, &vs);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	result = radv_device_init_meta_blit_stencil(device, &vs);
+	if (result != VK_SUCCESS)
+		goto fail;
+	return VK_SUCCESS;
+
+fail:
+	ralloc_free(vs.nir);
+	radv_device_finish_meta_blit_state(device);
+	return result;
+}
diff --git a/src/amd/vulkan/radv_meta_blit2d.c b/src/amd/vulkan/radv_meta_blit2d.c
new file mode 100644
index 0000000000..ba426e1e9a
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_blit2d.c
@@ -0,0 +1,1282 @@
+/*
+ * Copyright © 2016 Red Hat
+ *
+ * based on anv driver:
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "radv_meta.h"
+#include "nir/nir_builder.h"
+
+enum blit2d_dst_type {
+	/* We can bind this destination as a "normal" render target and render
+	 * to it just like you would anywhere else.
+	 */
+	BLIT2D_DST_TYPE_NORMAL,
+
+	/* The destination has a 3-channel RGB format.  Since we can't render to
+	 * non-power-of-two textures, we have to bind it as a red texture and
+	 * select the correct component for the given red pixel in the shader.
+	 */
+	BLIT2D_DST_TYPE_RGB,
+
+	BLIT2D_NUM_DST_TYPES,
+};
+
+
+enum blit2d_src_type {
+	BLIT2D_SRC_TYPE_IMAGE,
+	BLIT2D_SRC_TYPE_BUFFER,
+	BLIT2D_NUM_SRC_TYPES,
+};
+
+static void
+create_iview(struct radv_cmd_buffer *cmd_buffer,
+             struct radv_meta_blit2d_surf *surf,
+             VkImageUsageFlags usage,
+             struct radv_image_view *iview, VkFormat depth_format)
+{
+	VkFormat format;
+
+	if (depth_format)
+		format = depth_format;
+	else
+		format = surf->format;
+
+	radv_image_view_init(iview, cmd_buffer->device,
+			     &(VkImageViewCreateInfo) {
+				     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+					     .image = radv_image_to_handle(surf->image),
+					     .viewType = VK_IMAGE_VIEW_TYPE_2D,
+					     .format = format,
+					     .subresourceRange = {
+					     .aspectMask = surf->aspect_mask,
+					     .baseMipLevel = surf->level,
+					     .levelCount = 1,
+					     .baseArrayLayer = surf->layer,
+					     .layerCount = 1
+				     },
+					     }, cmd_buffer, usage);
+}
+
+static void
+create_bview(struct radv_cmd_buffer *cmd_buffer,
+	     struct radv_meta_blit2d_buffer *src,
+	     struct radv_buffer_view *bview, VkFormat depth_format)
+{
+	VkFormat format;
+
+	if (depth_format)
+		format = depth_format;
+	else
+		format = src->format;
+	radv_buffer_view_init(bview, cmd_buffer->device,
+			      &(VkBufferViewCreateInfo) {
+				      .sType = VK_STRUCTURE_TYPE_BUFFER_VIEW_CREATE_INFO,
+				      .flags = 0,
+				      .buffer = radv_buffer_to_handle(src->buffer),
+				      .format = format,
+				      .offset = src->offset,
+				      .range = VK_WHOLE_SIZE,
+			      }, cmd_buffer);
+
+}
+
+struct blit2d_src_temps {
+	struct radv_image_view iview;
+
+	VkDescriptorSet set;
+	struct radv_buffer_view bview;
+};
+
+static void
+blit2d_bind_src(struct radv_cmd_buffer *cmd_buffer,
+                struct radv_meta_blit2d_surf *src_img,
+                struct radv_meta_blit2d_buffer *src_buf,
+                struct radv_meta_blit2d_rect *rect,
+                struct blit2d_src_temps *tmp,
+                enum blit2d_src_type src_type, VkFormat depth_format)
+{
+	struct radv_device *device = cmd_buffer->device;
+	VkDevice vk_device = radv_device_to_handle(cmd_buffer->device);
+
+	if (src_type == BLIT2D_SRC_TYPE_BUFFER) {
+		create_bview(cmd_buffer, src_buf, &tmp->bview, depth_format);
+
+		radv_temp_descriptor_set_create(cmd_buffer->device, cmd_buffer,
+					        device->meta_state.blit2d.ds_layouts[src_type],
+					        &tmp->set);
+
+		radv_UpdateDescriptorSets(vk_device,
+					  1, /* writeCount */
+					  (VkWriteDescriptorSet[]) {
+						  {
+							  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+							  .dstSet = tmp->set,
+							  .dstBinding = 0,
+							  .dstArrayElement = 0,
+							  .descriptorCount = 1,
+							  .descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER,
+							  .pTexelBufferView = (VkBufferView[])  { radv_buffer_view_to_handle(&tmp->bview) }
+						  }
+					  }, 0, NULL);
+
+		radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
+				      device->meta_state.blit2d.p_layouts[src_type],
+				      VK_SHADER_STAGE_FRAGMENT_BIT, 0, 4,
+				      &src_buf->pitch);
+	} else {
+		create_iview(cmd_buffer, src_img, VK_IMAGE_USAGE_SAMPLED_BIT, &tmp->iview,
+			     depth_format);
+
+		radv_temp_descriptor_set_create(cmd_buffer->device, cmd_buffer,
+					        device->meta_state.blit2d.ds_layouts[src_type],
+					        &tmp->set);
+
+		radv_UpdateDescriptorSets(vk_device,
+					  1, /* writeCount */
+					  (VkWriteDescriptorSet[]) {
+						  {
+							  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+								  .dstSet = tmp->set,
+								  .dstBinding = 0,
+								  .dstArrayElement = 0,
+								  .descriptorCount = 1,
+								  .descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
+								  .pImageInfo = (VkDescriptorImageInfo[]) {
+								  {
+									  .sampler = NULL,
+									  .imageView = radv_image_view_to_handle(&tmp->iview),
+									  .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
+								  },
+							  }
+						  }
+					  }, 0, NULL);
+
+	}
+
+	radv_CmdBindDescriptorSets(radv_cmd_buffer_to_handle(cmd_buffer),
+				   VK_PIPELINE_BIND_POINT_GRAPHICS,
+				   device->meta_state.blit2d.p_layouts[src_type], 0, 1,
+				   &tmp->set, 0, NULL);
+}
+
+static void
+blit2d_unbind_src(struct radv_cmd_buffer *cmd_buffer,
+                  struct blit2d_src_temps *tmp,
+                  enum blit2d_src_type src_type)
+{
+	radv_temp_descriptor_set_destroy(cmd_buffer->device, tmp->set);
+}
+
+struct blit2d_dst_temps {
+	VkImage image;
+	struct radv_image_view iview;
+	VkFramebuffer fb;
+};
+
+static void
+blit2d_bind_dst(struct radv_cmd_buffer *cmd_buffer,
+                struct radv_meta_blit2d_surf *dst,
+                uint32_t width,
+                uint32_t height,
+		VkFormat depth_format,
+                struct blit2d_dst_temps *tmp)
+{
+	VkImageUsageFlagBits bits;
+
+	if (dst->aspect_mask == VK_IMAGE_ASPECT_COLOR_BIT)
+		bits = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
+	else
+		bits = VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
+
+	create_iview(cmd_buffer, dst, bits,
+		     &tmp->iview, depth_format);
+
+	radv_CreateFramebuffer(radv_device_to_handle(cmd_buffer->device),
+			       &(VkFramebufferCreateInfo) {
+				       .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
+					       .attachmentCount = 1,
+					       .pAttachments = (VkImageView[]) {
+					       radv_image_view_to_handle(&tmp->iview),
+				       },
+				       .width = width,
+				       .height = height,
+				       .layers = 1
+				}, &cmd_buffer->pool->alloc, &tmp->fb);
+}
+
+static void
+blit2d_unbind_dst(struct radv_cmd_buffer *cmd_buffer,
+                  struct blit2d_dst_temps *tmp)
+{
+	VkDevice vk_device = radv_device_to_handle(cmd_buffer->device);
+	radv_DestroyFramebuffer(vk_device, tmp->fb, &cmd_buffer->pool->alloc);
+}
+
+static void
+bind_pipeline(struct radv_cmd_buffer *cmd_buffer,
+              enum blit2d_src_type src_type, unsigned fs_key)
+{
+	VkPipeline pipeline =
+		cmd_buffer->device->meta_state.blit2d.pipelines[src_type][fs_key];
+
+	if (cmd_buffer->state.pipeline != radv_pipeline_from_handle(pipeline)) {
+		radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+				     VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+	}
+}
+
+static void
+bind_depth_pipeline(struct radv_cmd_buffer *cmd_buffer,
+		    enum blit2d_src_type src_type)
+{
+	VkPipeline pipeline =
+		cmd_buffer->device->meta_state.blit2d.depth_only_pipeline[src_type];
+
+	if (cmd_buffer->state.pipeline != radv_pipeline_from_handle(pipeline)) {
+		radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+				     VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+	}
+}
+
+static void
+bind_stencil_pipeline(struct radv_cmd_buffer *cmd_buffer,
+		      enum blit2d_src_type src_type)
+{
+	VkPipeline pipeline =
+		cmd_buffer->device->meta_state.blit2d.stencil_only_pipeline[src_type];
+
+	if (cmd_buffer->state.pipeline != radv_pipeline_from_handle(pipeline)) {
+		radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+				     VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+	}
+}
+
+static void
+radv_meta_blit2d_normal_dst(struct radv_cmd_buffer *cmd_buffer,
+			    struct radv_meta_blit2d_surf *src_img,
+			    struct radv_meta_blit2d_buffer *src_buf,
+			    struct radv_meta_blit2d_surf *dst,
+			    unsigned num_rects,
+			    struct radv_meta_blit2d_rect *rects, enum blit2d_src_type src_type)
+{
+	struct radv_device *device = cmd_buffer->device;
+
+	for (unsigned r = 0; r < num_rects; ++r) {
+		VkFormat depth_format = 0;
+		if (dst->aspect_mask != VK_IMAGE_ASPECT_COLOR_BIT)
+			depth_format = dst->image->vk_format;
+		struct blit2d_src_temps src_temps;
+		blit2d_bind_src(cmd_buffer, src_img, src_buf, &rects[r], &src_temps, src_type, depth_format);
+
+		uint32_t offset = 0;
+		struct blit2d_dst_temps dst_temps;
+		blit2d_bind_dst(cmd_buffer, dst, rects[r].dst_x + rects[r].width,
+				rects[r].dst_y + rects[r].height, depth_format, &dst_temps);
+
+		struct blit_vb_data {
+			float pos[2];
+			float tex_coord[2];
+		} vb_data[3];
+
+		unsigned vb_size = 3 * sizeof(*vb_data);
+
+		vb_data[0] = (struct blit_vb_data) {
+			.pos = {
+				rects[r].dst_x,
+				rects[r].dst_y,
+			},
+			.tex_coord = {
+				rects[r].src_x,
+				rects[r].src_y,
+			},
+		};
+
+		vb_data[1] = (struct blit_vb_data) {
+			.pos = {
+				rects[r].dst_x,
+				rects[r].dst_y + rects[r].height,
+			},
+			.tex_coord = {
+				rects[r].src_x,
+				rects[r].src_y + rects[r].height,
+			},
+		};
+
+		vb_data[2] = (struct blit_vb_data) {
+			.pos = {
+				rects[r].dst_x + rects[r].width,
+				rects[r].dst_y,
+			},
+			.tex_coord = {
+				rects[r].src_x + rects[r].width,
+				rects[r].src_y,
+			},
+		};
+
+
+		radv_cmd_buffer_upload_data(cmd_buffer, vb_size, 16, vb_data, &offset);
+
+		struct radv_buffer vertex_buffer = {
+			.device = device,
+			.size = vb_size,
+			.bo = cmd_buffer->upload.upload_bo,
+			.offset = offset,
+		};
+
+		radv_CmdBindVertexBuffers(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1,
+					  (VkBuffer[]) {
+						  radv_buffer_to_handle(&vertex_buffer),
+							  },
+					  (VkDeviceSize[]) {
+						  0,
+							  });
+
+
+		if (dst->aspect_mask == VK_IMAGE_ASPECT_COLOR_BIT) {
+			unsigned fs_key = radv_format_meta_fs_key(dst_temps.iview.vk_format);
+
+			RADV_CALL(CmdBeginRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer),
+						      &(VkRenderPassBeginInfo) {
+							      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+								      .renderPass = device->meta_state.blit2d.render_passes[fs_key],
+								      .framebuffer = dst_temps.fb,
+								      .renderArea = {
+								      .offset = { rects[r].dst_x, rects[r].dst_y, },
+								      .extent = { rects[r].width, rects[r].height },
+							      },
+								      .clearValueCount = 0,
+									       .pClearValues = NULL,
+									       }, VK_SUBPASS_CONTENTS_INLINE);
+
+
+			bind_pipeline(cmd_buffer, src_type, fs_key);
+		} else if (dst->aspect_mask == VK_IMAGE_ASPECT_DEPTH_BIT) {
+			RADV_CALL(CmdBeginRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer),
+						      &(VkRenderPassBeginInfo) {
+							      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+								      .renderPass = device->meta_state.blit2d.depth_only_rp,
+								      .framebuffer = dst_temps.fb,
+								      .renderArea = {
+								      .offset = { rects[r].dst_x, rects[r].dst_y, },
+								      .extent = { rects[r].width, rects[r].height },
+							      },
+								      .clearValueCount = 0,
+									       .pClearValues = NULL,
+									       }, VK_SUBPASS_CONTENTS_INLINE);
+
+
+			bind_depth_pipeline(cmd_buffer, src_type);
+
+		} else if (dst->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT) {
+			RADV_CALL(CmdBeginRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer),
+						      &(VkRenderPassBeginInfo) {
+							      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+								      .renderPass = device->meta_state.blit2d.stencil_only_rp,
+								      .framebuffer = dst_temps.fb,
+								      .renderArea = {
+								      .offset = { rects[r].dst_x, rects[r].dst_y, },
+								      .extent = { rects[r].width, rects[r].height },
+							      },
+								      .clearValueCount = 0,
+									       .pClearValues = NULL,
+									       }, VK_SUBPASS_CONTENTS_INLINE);
+
+
+			bind_stencil_pipeline(cmd_buffer, src_type);
+		}
+
+		RADV_CALL(CmdDraw)(radv_cmd_buffer_to_handle(cmd_buffer), 3, 1, 0, 0);
+		RADV_CALL(CmdEndRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer));
+
+		/* At the point where we emit the draw call, all data from the
+		 * descriptor sets, etc. has been used.  We are free to delete it.
+		 */
+		blit2d_unbind_src(cmd_buffer, &src_temps, src_type);
+		blit2d_unbind_dst(cmd_buffer, &dst_temps);
+	}
+}
+
+void
+radv_meta_blit2d(struct radv_cmd_buffer *cmd_buffer,
+		 struct radv_meta_blit2d_surf *src_img,
+		 struct radv_meta_blit2d_buffer *src_buf,
+		 struct radv_meta_blit2d_surf *dst,
+		 unsigned num_rects,
+		 struct radv_meta_blit2d_rect *rects)
+{
+	enum blit2d_src_type src_type = src_buf ? BLIT2D_SRC_TYPE_BUFFER :
+						  BLIT2D_SRC_TYPE_IMAGE;
+	radv_meta_blit2d_normal_dst(cmd_buffer, src_img, src_buf, dst,
+				    num_rects, rects, src_type);
+}
+
+static nir_shader *
+build_nir_vertex_shader(void)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	const struct glsl_type *vec2 = glsl_vector_type(GLSL_TYPE_FLOAT, 2);
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, "meta_blit_vs");
+
+	nir_variable *pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						   vec4, "a_pos");
+	pos_in->data.location = VERT_ATTRIB_GENERIC0;
+	nir_variable *pos_out = nir_variable_create(b.shader, nir_var_shader_out,
+						    vec4, "gl_Position");
+	pos_out->data.location = VARYING_SLOT_POS;
+	nir_copy_var(&b, pos_out, pos_in);
+
+	nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						       vec2, "a_tex_pos");
+	tex_pos_in->data.location = VERT_ATTRIB_GENERIC1;
+	nir_variable *tex_pos_out = nir_variable_create(b.shader, nir_var_shader_out,
+							vec2, "v_tex_pos");
+	tex_pos_out->data.location = VARYING_SLOT_VAR0;
+	tex_pos_out->data.interpolation = INTERP_MODE_SMOOTH;
+	nir_copy_var(&b, tex_pos_out, tex_pos_in);
+
+	return b.shader;
+}
+
+typedef nir_ssa_def* (*texel_fetch_build_func)(struct nir_builder *,
+                                               struct radv_device *,
+                                               nir_ssa_def *);
+
+static nir_ssa_def *
+build_nir_texel_fetch(struct nir_builder *b, struct radv_device *device,
+                      nir_ssa_def *tex_pos)
+{
+	const struct glsl_type *sampler_type =
+		glsl_sampler_type(GLSL_SAMPLER_DIM_2D, false, false, GLSL_TYPE_UINT);
+	nir_variable *sampler = nir_variable_create(b->shader, nir_var_uniform,
+						    sampler_type, "s_tex");
+	sampler->data.descriptor_set = 0;
+	sampler->data.binding = 0;
+
+	nir_tex_instr *tex = nir_tex_instr_create(b->shader, 2);
+	tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
+	tex->op = nir_texop_txf;
+	tex->src[0].src_type = nir_tex_src_coord;
+	tex->src[0].src = nir_src_for_ssa(tex_pos);
+	tex->src[1].src_type = nir_tex_src_lod;
+	tex->src[1].src = nir_src_for_ssa(nir_imm_int(b, 0));
+	tex->dest_type = nir_type_uint;
+	tex->is_array = false;
+	tex->coord_components = 2;
+	tex->texture = nir_deref_var_create(tex, sampler);
+	tex->sampler = NULL;
+
+	nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
+	nir_builder_instr_insert(b, &tex->instr);
+
+	return &tex->dest.ssa;
+}
+
+
+static nir_ssa_def *
+build_nir_buffer_fetch(struct nir_builder *b, struct radv_device *device,
+                      nir_ssa_def *tex_pos)
+{
+	const struct glsl_type *sampler_type =
+		glsl_sampler_type(GLSL_SAMPLER_DIM_BUF, false, false, GLSL_TYPE_UINT);
+	nir_variable *sampler = nir_variable_create(b->shader, nir_var_uniform,
+						    sampler_type, "s_tex");
+	sampler->data.descriptor_set = 0;
+	sampler->data.binding = 0;
+
+	nir_intrinsic_instr *width = nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_push_constant);
+	width->src[0] = nir_src_for_ssa(nir_imm_int(b, 0));
+	width->num_components = 1;
+	nir_ssa_dest_init(&width->instr, &width->dest, 1, 32, "width");
+	nir_builder_instr_insert(b, &width->instr);
+
+	nir_ssa_def *pos_x = nir_channel(b, tex_pos, 0);
+	nir_ssa_def *pos_y = nir_channel(b, tex_pos, 1);
+	pos_y = nir_imul(b, pos_y, &width->dest.ssa);
+	pos_x = nir_iadd(b, pos_x, pos_y);
+	//pos_x = nir_iadd(b, pos_x, nir_imm_int(b, 100000));
+
+	nir_tex_instr *tex = nir_tex_instr_create(b->shader, 1);
+	tex->sampler_dim = GLSL_SAMPLER_DIM_BUF;
+	tex->op = nir_texop_txf;
+	tex->src[0].src_type = nir_tex_src_coord;
+	tex->src[0].src = nir_src_for_ssa(pos_x);
+	tex->dest_type = nir_type_uint;
+	tex->is_array = false;
+	tex->coord_components = 1;
+	tex->texture = nir_deref_var_create(tex, sampler);
+	tex->sampler = NULL;
+
+	nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
+	nir_builder_instr_insert(b, &tex->instr);
+
+	return &tex->dest.ssa;
+}
+
+static const VkPipelineVertexInputStateCreateInfo normal_vi_create_info = {
+	.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+	.vertexBindingDescriptionCount = 1,
+	.pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
+		{
+			.binding = 0,
+			.stride = 4 * sizeof(float),
+			.inputRate = VK_VERTEX_INPUT_RATE_VERTEX
+		},
+	},
+	.vertexAttributeDescriptionCount = 2,
+	.pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
+		{
+			/* Position */
+			.location = 0,
+			.binding = 0,
+			.format = VK_FORMAT_R32G32_SFLOAT,
+			.offset = 0
+		},
+		{
+			/* Texture Coordinate */
+			.location = 1,
+			.binding = 0,
+			.format = VK_FORMAT_R32G32_SFLOAT,
+			.offset = 8
+		},
+	},
+};
+
+static nir_shader *
+build_nir_copy_fragment_shader(struct radv_device *device,
+                               texel_fetch_build_func txf_func, const char* name)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	const struct glsl_type *vec2 = glsl_vector_type(GLSL_TYPE_FLOAT, 2);
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, name);
+
+	nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						       vec2, "v_tex_pos");
+	tex_pos_in->data.location = VARYING_SLOT_VAR0;
+
+	nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
+						      vec4, "f_color");
+	color_out->data.location = FRAG_RESULT_DATA0;
+
+	nir_ssa_def *pos_int = nir_f2i(&b, nir_load_var(&b, tex_pos_in));
+	unsigned swiz[4] = { 0, 1 };
+	nir_ssa_def *tex_pos = nir_swizzle(&b, pos_int, swiz, 2, false);
+
+	nir_ssa_def *color = txf_func(&b, device, tex_pos);
+	nir_store_var(&b, color_out, color, 0xf);
+
+	return b.shader;
+}
+
+static nir_shader *
+build_nir_copy_fragment_shader_depth(struct radv_device *device,
+				     texel_fetch_build_func txf_func, const char* name)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	const struct glsl_type *vec2 = glsl_vector_type(GLSL_TYPE_FLOAT, 2);
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, name);
+
+	nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						       vec2, "v_tex_pos");
+	tex_pos_in->data.location = VARYING_SLOT_VAR0;
+
+	nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
+						      vec4, "f_color");
+	color_out->data.location = FRAG_RESULT_DEPTH;
+
+	nir_ssa_def *pos_int = nir_f2i(&b, nir_load_var(&b, tex_pos_in));
+	unsigned swiz[4] = { 0, 1 };
+	nir_ssa_def *tex_pos = nir_swizzle(&b, pos_int, swiz, 2, false);
+
+	nir_ssa_def *color = txf_func(&b, device, tex_pos);
+	nir_store_var(&b, color_out, color, 0x1);
+
+	return b.shader;
+}
+
+static nir_shader *
+build_nir_copy_fragment_shader_stencil(struct radv_device *device,
+				       texel_fetch_build_func txf_func, const char* name)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	const struct glsl_type *vec2 = glsl_vector_type(GLSL_TYPE_FLOAT, 2);
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, name);
+
+	nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+						       vec2, "v_tex_pos");
+	tex_pos_in->data.location = VARYING_SLOT_VAR0;
+
+	nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
+						      vec4, "f_color");
+	color_out->data.location = FRAG_RESULT_STENCIL;
+
+	nir_ssa_def *pos_int = nir_f2i(&b, nir_load_var(&b, tex_pos_in));
+	unsigned swiz[4] = { 0, 1 };
+	nir_ssa_def *tex_pos = nir_swizzle(&b, pos_int, swiz, 2, false);
+
+	nir_ssa_def *color = txf_func(&b, device, tex_pos);
+	nir_store_var(&b, color_out, color, 0x1);
+
+	return b.shader;
+}
+
+void
+radv_device_finish_meta_blit2d_state(struct radv_device *device)
+{
+	for(unsigned j = 0; j < NUM_META_FS_KEYS; ++j) {
+		if (device->meta_state.blit2d.render_passes[j]) {
+			radv_DestroyRenderPass(radv_device_to_handle(device),
+					       device->meta_state.blit2d.render_passes[j],
+					       &device->meta_state.alloc);
+		}
+	}
+
+	radv_DestroyRenderPass(radv_device_to_handle(device),
+			       device->meta_state.blit2d.depth_only_rp,
+			       &device->meta_state.alloc);
+	radv_DestroyRenderPass(radv_device_to_handle(device),
+			       device->meta_state.blit2d.stencil_only_rp,
+			       &device->meta_state.alloc);
+
+	for (unsigned src = 0; src < BLIT2D_NUM_SRC_TYPES; src++) {
+		if (device->meta_state.blit2d.p_layouts[src]) {
+			radv_DestroyPipelineLayout(radv_device_to_handle(device),
+						device->meta_state.blit2d.p_layouts[src],
+						&device->meta_state.alloc);
+		}
+
+		if (device->meta_state.blit2d.ds_layouts[src]) {
+			radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
+							device->meta_state.blit2d.ds_layouts[src],
+							&device->meta_state.alloc);
+		}
+
+		for (unsigned j = 0; j < NUM_META_FS_KEYS; ++j) {
+			if (device->meta_state.blit2d.pipelines[src][j]) {
+				radv_DestroyPipeline(radv_device_to_handle(device),
+						     device->meta_state.blit2d.pipelines[src][j],
+						     &device->meta_state.alloc);
+			}
+		}
+
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.blit2d.depth_only_pipeline[src],
+				     &device->meta_state.alloc);
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.blit2d.stencil_only_pipeline[src],
+				     &device->meta_state.alloc);
+	}
+}
+
+static VkResult
+blit2d_init_color_pipeline(struct radv_device *device,
+			   enum blit2d_src_type src_type,
+			   VkFormat format)
+{
+	VkResult result;
+	unsigned fs_key = radv_format_meta_fs_key(format);
+	const char *name;
+
+	texel_fetch_build_func src_func;
+	switch(src_type) {
+	case BLIT2D_SRC_TYPE_IMAGE:
+		src_func = build_nir_texel_fetch;
+		name = "meta_blit2d_image_fs";
+		break;
+	case BLIT2D_SRC_TYPE_BUFFER:
+		src_func = build_nir_buffer_fetch;
+		name = "meta_blit2d_buffer_fs";
+		break;
+	default:
+		unreachable("unknown blit src type\n");
+		break;
+	}
+
+	const VkPipelineVertexInputStateCreateInfo *vi_create_info;
+	struct radv_shader_module fs = { .nir = NULL };
+
+
+	fs.nir = build_nir_copy_fragment_shader(device, src_func, name);
+	vi_create_info = &normal_vi_create_info;
+
+	struct radv_shader_module vs = {
+		.nir = build_nir_vertex_shader(),
+	};
+
+	VkPipelineShaderStageCreateInfo pipeline_shader_stages[] = {
+		{
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_VERTEX_BIT,
+			.module = radv_shader_module_to_handle(&vs),
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		}, {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+			.module = radv_shader_module_to_handle(&fs),
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		},
+	};
+
+	if (!device->meta_state.blit2d.render_passes[fs_key]) {
+		result = radv_CreateRenderPass(radv_device_to_handle(device),
+					       &(VkRenderPassCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+						       .attachmentCount = 1,
+						       .pAttachments = &(VkAttachmentDescription) {
+						       .format = format,
+						       .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+						       .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+						       .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
+						       .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+					       .subpassCount = 1,
+					       .pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 1,
+						       .pColorAttachments = &(VkAttachmentReference) {
+							       .attachment = 0,
+							       .layout = VK_IMAGE_LAYOUT_GENERAL,
+							},
+					       .pResolveAttachments = NULL,
+					       .pDepthStencilAttachment = &(VkAttachmentReference) {
+						       .attachment = VK_ATTACHMENT_UNUSED,
+						       .layout = VK_IMAGE_LAYOUT_GENERAL,
+					       },
+					       .preserveAttachmentCount = 1,
+					       .pPreserveAttachments = (uint32_t[]) { 0 },
+					       },
+					.dependencyCount = 0,
+				 }, &device->meta_state.alloc, &device->meta_state.blit2d.render_passes[fs_key]);
+	}
+
+	const VkGraphicsPipelineCreateInfo vk_pipeline_info = {
+		.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+		.stageCount = ARRAY_SIZE(pipeline_shader_stages),
+		.pStages = pipeline_shader_stages,
+		.pVertexInputState = vi_create_info,
+		.pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+			.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+			.primitiveRestartEnable = false,
+		},
+		.pViewportState = &(VkPipelineViewportStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+			.viewportCount = 0,
+			.scissorCount = 0,
+		},
+		.pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+			.rasterizerDiscardEnable = false,
+			.polygonMode = VK_POLYGON_MODE_FILL,
+			.cullMode = VK_CULL_MODE_NONE,
+			.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE
+		},
+		.pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+			.rasterizationSamples = 1,
+			.sampleShadingEnable = false,
+			.pSampleMask = (VkSampleMask[]) { UINT32_MAX },
+		},
+		.pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+			.attachmentCount = 1,
+			.pAttachments = (VkPipelineColorBlendAttachmentState []) {
+				{ .colorWriteMask =
+				  VK_COLOR_COMPONENT_A_BIT |
+				  VK_COLOR_COMPONENT_R_BIT |
+				  VK_COLOR_COMPONENT_G_BIT |
+				  VK_COLOR_COMPONENT_B_BIT },
+			}
+		},
+		.pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+			.dynamicStateCount = 7,
+			.pDynamicStates = (VkDynamicState[]) {
+				VK_DYNAMIC_STATE_LINE_WIDTH,
+				VK_DYNAMIC_STATE_DEPTH_BIAS,
+				VK_DYNAMIC_STATE_BLEND_CONSTANTS,
+				VK_DYNAMIC_STATE_DEPTH_BOUNDS,
+				VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK,
+				VK_DYNAMIC_STATE_STENCIL_WRITE_MASK,
+				VK_DYNAMIC_STATE_STENCIL_REFERENCE,
+			},
+		},
+		.flags = 0,
+		.layout = device->meta_state.blit2d.p_layouts[src_type],
+		.renderPass = device->meta_state.blit2d.render_passes[fs_key],
+		.subpass = 0,
+	};
+
+	const struct radv_graphics_pipeline_create_info radv_pipeline_info = {
+		.use_rectlist = true
+	};
+
+	result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &vk_pipeline_info, &radv_pipeline_info,
+					       &device->meta_state.alloc,
+					       &device->meta_state.blit2d.pipelines[src_type][fs_key]);
+
+
+	ralloc_free(vs.nir);
+	ralloc_free(fs.nir);
+
+	return result;
+}
+
+static VkResult
+blit2d_init_depth_only_pipeline(struct radv_device *device,
+				enum blit2d_src_type src_type)
+{
+	VkResult result;
+	const char *name;
+
+	texel_fetch_build_func src_func;
+	switch(src_type) {
+	case BLIT2D_SRC_TYPE_IMAGE:
+		src_func = build_nir_texel_fetch;
+		name = "meta_blit2d_depth_image_fs";
+		break;
+	case BLIT2D_SRC_TYPE_BUFFER:
+		src_func = build_nir_buffer_fetch;
+		name = "meta_blit2d_depth_buffer_fs";
+		break;
+	default:
+		unreachable("unknown blit src type\n");
+		break;
+	}
+
+	const VkPipelineVertexInputStateCreateInfo *vi_create_info;
+	struct radv_shader_module fs = { .nir = NULL };
+
+	fs.nir = build_nir_copy_fragment_shader_depth(device, src_func, name);
+	vi_create_info = &normal_vi_create_info;
+
+	struct radv_shader_module vs = {
+		.nir = build_nir_vertex_shader(),
+	};
+
+	VkPipelineShaderStageCreateInfo pipeline_shader_stages[] = {
+		{
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_VERTEX_BIT,
+			.module = radv_shader_module_to_handle(&vs),
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		}, {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+			.module = radv_shader_module_to_handle(&fs),
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		},
+	};
+
+	if (!device->meta_state.blit2d.depth_only_rp) {
+		result = radv_CreateRenderPass(radv_device_to_handle(device),
+					       &(VkRenderPassCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+							       .attachmentCount = 1,
+							       .pAttachments = &(VkAttachmentDescription) {
+							       .format = 0,
+							       .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+							       .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+							       .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
+							       .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+						       .subpassCount = 1,
+						       .pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 0,
+						       .pColorAttachments = NULL,
+						       .pResolveAttachments = NULL,
+						       .pDepthStencilAttachment = &(VkAttachmentReference) {
+							       .attachment = 0,
+							       .layout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+						       .preserveAttachmentCount = 1,
+						       .pPreserveAttachments = (uint32_t[]) { 0 },
+					       },
+								.dependencyCount = 0,
+						 }, &device->meta_state.alloc, &device->meta_state.blit2d.depth_only_rp);
+	}
+
+	const VkGraphicsPipelineCreateInfo vk_pipeline_info = {
+		.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+		.stageCount = ARRAY_SIZE(pipeline_shader_stages),
+		.pStages = pipeline_shader_stages,
+		.pVertexInputState = vi_create_info,
+		.pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+			.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+			.primitiveRestartEnable = false,
+		},
+		.pViewportState = &(VkPipelineViewportStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+			.viewportCount = 0,
+			.scissorCount = 0,
+		},
+		.pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+			.rasterizerDiscardEnable = false,
+			.polygonMode = VK_POLYGON_MODE_FILL,
+			.cullMode = VK_CULL_MODE_NONE,
+			.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE
+		},
+		.pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+			.rasterizationSamples = 1,
+			.sampleShadingEnable = false,
+			.pSampleMask = (VkSampleMask[]) { UINT32_MAX },
+		},
+		.pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+			.attachmentCount = 0,
+			.pAttachments = NULL,
+		},
+		.pDepthStencilState = &(VkPipelineDepthStencilStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
+			.depthTestEnable = true,
+			.depthWriteEnable = true,
+			.depthCompareOp = VK_COMPARE_OP_ALWAYS,
+		},
+		.pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+			.dynamicStateCount = 7,
+			.pDynamicStates = (VkDynamicState[]) {
+				VK_DYNAMIC_STATE_LINE_WIDTH,
+				VK_DYNAMIC_STATE_DEPTH_BIAS,
+				VK_DYNAMIC_STATE_BLEND_CONSTANTS,
+				VK_DYNAMIC_STATE_DEPTH_BOUNDS,
+				VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK,
+				VK_DYNAMIC_STATE_STENCIL_WRITE_MASK,
+				VK_DYNAMIC_STATE_STENCIL_REFERENCE,
+			},
+		},
+		.flags = 0,
+		.layout = device->meta_state.blit2d.p_layouts[src_type],
+		.renderPass = device->meta_state.blit2d.depth_only_rp,
+		.subpass = 0,
+	};
+
+	const struct radv_graphics_pipeline_create_info radv_pipeline_info = {
+		.use_rectlist = true
+	};
+
+	result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &vk_pipeline_info, &radv_pipeline_info,
+					       &device->meta_state.alloc,
+					       &device->meta_state.blit2d.depth_only_pipeline[src_type]);
+
+
+	ralloc_free(vs.nir);
+	ralloc_free(fs.nir);
+
+	return result;
+}
+
+static VkResult
+blit2d_init_stencil_only_pipeline(struct radv_device *device,
+				  enum blit2d_src_type src_type)
+{
+	VkResult result;
+	const char *name;
+
+	texel_fetch_build_func src_func;
+	switch(src_type) {
+	case BLIT2D_SRC_TYPE_IMAGE:
+		src_func = build_nir_texel_fetch;
+		name = "meta_blit2d_stencil_image_fs";
+		break;
+	case BLIT2D_SRC_TYPE_BUFFER:
+		src_func = build_nir_buffer_fetch;
+		name = "meta_blit2d_stencil_buffer_fs";
+		break;
+	default:
+		unreachable("unknown blit src type\n");
+		break;
+	}
+
+	const VkPipelineVertexInputStateCreateInfo *vi_create_info;
+	struct radv_shader_module fs = { .nir = NULL };
+
+	fs.nir = build_nir_copy_fragment_shader_stencil(device, src_func, name);
+	vi_create_info = &normal_vi_create_info;
+
+	struct radv_shader_module vs = {
+		.nir = build_nir_vertex_shader(),
+	};
+
+	VkPipelineShaderStageCreateInfo pipeline_shader_stages[] = {
+		{
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_VERTEX_BIT,
+			.module = radv_shader_module_to_handle(&vs),
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		}, {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+			.module = radv_shader_module_to_handle(&fs),
+			.pName = "main",
+			.pSpecializationInfo = NULL
+		},
+	};
+
+	if (!device->meta_state.blit2d.stencil_only_rp) {
+		result = radv_CreateRenderPass(radv_device_to_handle(device),
+					       &(VkRenderPassCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+							       .attachmentCount = 1,
+							       .pAttachments = &(VkAttachmentDescription) {
+							       .format = 0,
+							       .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+							       .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+							       .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
+							       .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+						       .subpassCount = 1,
+						       .pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 0,
+						       .pColorAttachments = NULL,
+						       .pResolveAttachments = NULL,
+						       .pDepthStencilAttachment = &(VkAttachmentReference) {
+							       .attachment = 0,
+							       .layout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+						       .preserveAttachmentCount = 1,
+						       .pPreserveAttachments = (uint32_t[]) { 0 },
+					       },
+								.dependencyCount = 0,
+						 }, &device->meta_state.alloc, &device->meta_state.blit2d.stencil_only_rp);
+	}
+
+	const VkGraphicsPipelineCreateInfo vk_pipeline_info = {
+		.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+		.stageCount = ARRAY_SIZE(pipeline_shader_stages),
+		.pStages = pipeline_shader_stages,
+		.pVertexInputState = vi_create_info,
+		.pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+			.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+			.primitiveRestartEnable = false,
+		},
+		.pViewportState = &(VkPipelineViewportStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+			.viewportCount = 0,
+			.scissorCount = 0,
+		},
+		.pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+			.rasterizerDiscardEnable = false,
+			.polygonMode = VK_POLYGON_MODE_FILL,
+			.cullMode = VK_CULL_MODE_NONE,
+			.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE
+		},
+		.pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+			.rasterizationSamples = 1,
+			.sampleShadingEnable = false,
+			.pSampleMask = (VkSampleMask[]) { UINT32_MAX },
+		},
+		.pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+			.attachmentCount = 0,
+			.pAttachments = NULL,
+		},
+		.pDepthStencilState = &(VkPipelineDepthStencilStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
+			.depthTestEnable = false,
+			.depthWriteEnable = false,
+			.stencilTestEnable = true,
+			.front = {
+				.failOp = VK_STENCIL_OP_REPLACE,
+				.passOp = VK_STENCIL_OP_REPLACE,
+				.depthFailOp = VK_STENCIL_OP_REPLACE,
+				.compareOp = VK_COMPARE_OP_ALWAYS,
+				.compareMask = 0xff,
+				.writeMask = 0xff,
+				.reference = 0
+			},
+			.back = {
+				.failOp = VK_STENCIL_OP_REPLACE,
+				.passOp = VK_STENCIL_OP_REPLACE,
+				.depthFailOp = VK_STENCIL_OP_REPLACE,
+				.compareOp = VK_COMPARE_OP_ALWAYS,
+				.compareMask = 0xff,
+				.writeMask = 0xff,
+				.reference = 0
+			},
+			.depthCompareOp = VK_COMPARE_OP_ALWAYS,
+		},
+		.pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+			.dynamicStateCount = 4,
+			.pDynamicStates = (VkDynamicState[]) {
+				VK_DYNAMIC_STATE_LINE_WIDTH,
+				VK_DYNAMIC_STATE_DEPTH_BIAS,
+				VK_DYNAMIC_STATE_BLEND_CONSTANTS,
+				VK_DYNAMIC_STATE_DEPTH_BOUNDS,
+			},
+		},
+		.flags = 0,
+		.layout = device->meta_state.blit2d.p_layouts[src_type],
+		.renderPass = device->meta_state.blit2d.stencil_only_rp,
+		.subpass = 0,
+	};
+
+	const struct radv_graphics_pipeline_create_info radv_pipeline_info = {
+		.use_rectlist = true
+	};
+
+	result = radv_graphics_pipeline_create(radv_device_to_handle(device),
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &vk_pipeline_info, &radv_pipeline_info,
+					       &device->meta_state.alloc,
+					       &device->meta_state.blit2d.stencil_only_pipeline[src_type]);
+
+
+	ralloc_free(vs.nir);
+	ralloc_free(fs.nir);
+
+	return result;
+}
+
+static VkFormat pipeline_formats[] = {
+   VK_FORMAT_R8G8B8A8_UNORM,
+   VK_FORMAT_R8G8B8A8_UINT,
+   VK_FORMAT_R8G8B8A8_SINT,
+   VK_FORMAT_R16G16B16A16_UNORM,
+   VK_FORMAT_R16G16B16A16_SNORM,
+   VK_FORMAT_R16G16B16A16_UINT,
+   VK_FORMAT_R16G16B16A16_SINT,
+   VK_FORMAT_R32_SFLOAT,
+   VK_FORMAT_R32G32_SFLOAT,
+   VK_FORMAT_R32G32B32A32_SFLOAT
+};
+
+VkResult
+radv_device_init_meta_blit2d_state(struct radv_device *device)
+{
+	VkResult result;
+
+	zero(device->meta_state.blit2d);
+
+	result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
+						&(VkDescriptorSetLayoutCreateInfo) {
+							.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+								.bindingCount = 1,
+								.pBindings = (VkDescriptorSetLayoutBinding[]) {
+								{
+									.binding = 0,
+									.descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
+									.descriptorCount = 1,
+									.stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT,
+									.pImmutableSamplers = NULL
+								},
+							}
+						}, &device->meta_state.alloc, &device->meta_state.blit2d.ds_layouts[BLIT2D_SRC_TYPE_IMAGE]);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	result = radv_CreatePipelineLayout(radv_device_to_handle(device),
+					   &(VkPipelineLayoutCreateInfo) {
+						   .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+							   .setLayoutCount = 1,
+							   .pSetLayouts = &device->meta_state.blit2d.ds_layouts[BLIT2D_SRC_TYPE_IMAGE],
+							   },
+					   &device->meta_state.alloc, &device->meta_state.blit2d.p_layouts[BLIT2D_SRC_TYPE_IMAGE]);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
+						&(VkDescriptorSetLayoutCreateInfo) {
+							.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+								.bindingCount = 1,
+								.pBindings = (VkDescriptorSetLayoutBinding[]) {
+								{
+									.binding = 0,
+									.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER,
+									.descriptorCount = 1,
+									.stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT,
+									.pImmutableSamplers = NULL
+								},
+							}
+						}, &device->meta_state.alloc, &device->meta_state.blit2d.ds_layouts[BLIT2D_SRC_TYPE_BUFFER]);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	const VkPushConstantRange push_constant_range = {VK_SHADER_STAGE_FRAGMENT_BIT, 0, 4};
+	result = radv_CreatePipelineLayout(radv_device_to_handle(device),
+					   &(VkPipelineLayoutCreateInfo) {
+						   .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+						   .setLayoutCount = 1,
+						   .pSetLayouts = &device->meta_state.blit2d.ds_layouts[BLIT2D_SRC_TYPE_BUFFER],
+						   .pushConstantRangeCount = 1,
+						   .pPushConstantRanges = &push_constant_range,
+					   },
+					   &device->meta_state.alloc, &device->meta_state.blit2d.p_layouts[BLIT2D_SRC_TYPE_BUFFER]);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	for (unsigned src = 0; src < BLIT2D_NUM_SRC_TYPES; src++) {
+		for (unsigned j = 0; j < ARRAY_SIZE(pipeline_formats); ++j) {
+			result = blit2d_init_color_pipeline(device, src, pipeline_formats[j]);
+			if (result != VK_SUCCESS)
+				goto fail;
+		}
+
+		result = blit2d_init_depth_only_pipeline(device, src);
+		if (result != VK_SUCCESS)
+			goto fail;
+
+		result = blit2d_init_stencil_only_pipeline(device, src);
+		if (result != VK_SUCCESS)
+			goto fail;
+	}
+
+	return VK_SUCCESS;
+
+fail:
+	radv_device_finish_meta_blit2d_state(device);
+	return result;
+}
diff --git a/src/amd/vulkan/radv_meta_buffer.c b/src/amd/vulkan/radv_meta_buffer.c
new file mode 100644
index 0000000000..adea25e02f
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_buffer.c
@@ -0,0 +1,543 @@
+#include "radv_meta.h"
+#include "nir/nir_builder.h"
+
+#include "sid.h"
+#include "radv_cs.h"
+
+static nir_shader *
+build_buffer_fill_shader(struct radv_device *dev)
+{
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, "meta_buffer_fill");
+	b.shader->info.cs.local_size[0] = 64;
+	b.shader->info.cs.local_size[1] = 1;
+	b.shader->info.cs.local_size[2] = 1;
+
+	nir_ssa_def *invoc_id = nir_load_system_value(&b, nir_intrinsic_load_local_invocation_id, 0);
+	nir_ssa_def *wg_id = nir_load_system_value(&b, nir_intrinsic_load_work_group_id, 0);
+	nir_ssa_def *block_size = nir_imm_ivec4(&b,
+						b.shader->info.cs.local_size[0],
+						b.shader->info.cs.local_size[1],
+						b.shader->info.cs.local_size[2], 0);
+
+	nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
+
+	nir_ssa_def *offset = nir_imul(&b, global_id, nir_imm_int(&b, 16));
+	offset = nir_swizzle(&b, offset, (unsigned[]) {0, 0, 0, 0}, 1, false);
+
+	nir_intrinsic_instr *dst_buf = nir_intrinsic_instr_create(b.shader,
+	                                                          nir_intrinsic_vulkan_resource_index);
+	dst_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
+	nir_intrinsic_set_desc_set(dst_buf, 0);
+	nir_intrinsic_set_binding(dst_buf, 0);
+	nir_ssa_dest_init(&dst_buf->instr, &dst_buf->dest, 1, 32, NULL);
+	nir_builder_instr_insert(&b, &dst_buf->instr);
+
+	nir_intrinsic_instr *load = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
+	load->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
+	load->num_components = 1;
+	nir_ssa_dest_init(&load->instr, &load->dest, 1, 32, "fill_value");
+	nir_builder_instr_insert(&b, &load->instr);
+
+	nir_ssa_def *swizzled_load = nir_swizzle(&b, &load->dest.ssa, (unsigned[]) { 0, 0, 0, 0}, 4, false);
+
+	nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
+	store->src[0] = nir_src_for_ssa(swizzled_load);
+	store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
+	store->src[2] = nir_src_for_ssa(offset);
+	nir_intrinsic_set_write_mask(store, 0xf);
+	store->num_components = 4;
+	nir_builder_instr_insert(&b, &store->instr);
+
+	return b.shader;
+}
+
+static nir_shader *
+build_buffer_copy_shader(struct radv_device *dev)
+{
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, "meta_buffer_copy");
+	b.shader->info.cs.local_size[0] = 64;
+	b.shader->info.cs.local_size[1] = 1;
+	b.shader->info.cs.local_size[2] = 1;
+
+	nir_ssa_def *invoc_id = nir_load_system_value(&b, nir_intrinsic_load_local_invocation_id, 0);
+	nir_ssa_def *wg_id = nir_load_system_value(&b, nir_intrinsic_load_work_group_id, 0);
+	nir_ssa_def *block_size = nir_imm_ivec4(&b,
+						b.shader->info.cs.local_size[0],
+						b.shader->info.cs.local_size[1],
+						b.shader->info.cs.local_size[2], 0);
+
+	nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
+
+	nir_ssa_def *offset = nir_imul(&b, global_id, nir_imm_int(&b, 16));
+	offset = nir_swizzle(&b, offset, (unsigned[]) {0, 0, 0, 0}, 1, false);
+
+	nir_intrinsic_instr *dst_buf = nir_intrinsic_instr_create(b.shader,
+	                                                          nir_intrinsic_vulkan_resource_index);
+	dst_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
+	nir_intrinsic_set_desc_set(dst_buf, 0);
+	nir_intrinsic_set_binding(dst_buf, 0);
+	nir_ssa_dest_init(&dst_buf->instr, &dst_buf->dest, 1, 32, NULL);
+	nir_builder_instr_insert(&b, &dst_buf->instr);
+
+	nir_intrinsic_instr *src_buf = nir_intrinsic_instr_create(b.shader,
+	                                                          nir_intrinsic_vulkan_resource_index);
+	src_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
+	nir_intrinsic_set_desc_set(src_buf, 0);
+	nir_intrinsic_set_binding(src_buf, 1);
+	nir_ssa_dest_init(&src_buf->instr, &src_buf->dest, 1, 32, NULL);
+	nir_builder_instr_insert(&b, &src_buf->instr);
+
+	nir_intrinsic_instr *load = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
+	load->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
+	load->src[1] = nir_src_for_ssa(offset);
+	nir_ssa_dest_init(&load->instr, &load->dest, 4, 32, NULL);
+	load->num_components = 4;
+	nir_builder_instr_insert(&b, &load->instr);
+
+	nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
+	store->src[0] = nir_src_for_ssa(&load->dest.ssa);
+	store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
+	store->src[2] = nir_src_for_ssa(offset);
+	nir_intrinsic_set_write_mask(store, 0xf);
+	store->num_components = 4;
+	nir_builder_instr_insert(&b, &store->instr);
+
+	return b.shader;
+}
+
+
+
+VkResult radv_device_init_meta_buffer_state(struct radv_device *device)
+{
+	VkResult result;
+	struct radv_shader_module fill_cs = { .nir = NULL };
+	struct radv_shader_module copy_cs = { .nir = NULL };
+
+	zero(device->meta_state.buffer);
+
+	fill_cs.nir = build_buffer_fill_shader(device);
+	copy_cs.nir = build_buffer_copy_shader(device);
+
+	VkDescriptorSetLayoutCreateInfo fill_ds_create_info = {
+		.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+		.bindingCount = 1,
+		.pBindings = (VkDescriptorSetLayoutBinding[]) {
+			{
+				.binding = 0,
+				.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+				.descriptorCount = 1,
+				.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+				.pImmutableSamplers = NULL
+			},
+		}
+	};
+
+	result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
+						&fill_ds_create_info,
+						&device->meta_state.alloc,
+						&device->meta_state.buffer.fill_ds_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	VkDescriptorSetLayoutCreateInfo copy_ds_create_info = {
+		.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+		.bindingCount = 2,
+		.pBindings = (VkDescriptorSetLayoutBinding[]) {
+			{
+				.binding = 0,
+				.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+				.descriptorCount = 1,
+				.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+				.pImmutableSamplers = NULL
+			},
+			{
+				.binding = 1,
+				.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+				.descriptorCount = 1,
+				.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+				.pImmutableSamplers = NULL
+			},
+		}
+	};
+
+	result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
+						&copy_ds_create_info,
+						&device->meta_state.alloc,
+						&device->meta_state.buffer.copy_ds_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+
+	VkPipelineLayoutCreateInfo fill_pl_create_info = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+		.setLayoutCount = 1,
+		.pSetLayouts = &device->meta_state.buffer.fill_ds_layout,
+		.pushConstantRangeCount = 1,
+		.pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 4},
+	};
+
+	result = radv_CreatePipelineLayout(radv_device_to_handle(device),
+					  &fill_pl_create_info,
+					  &device->meta_state.alloc,
+					  &device->meta_state.buffer.fill_p_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	VkPipelineLayoutCreateInfo copy_pl_create_info = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+		.setLayoutCount = 1,
+		.pSetLayouts = &device->meta_state.buffer.copy_ds_layout,
+		.pushConstantRangeCount = 0,
+	};
+
+	result = radv_CreatePipelineLayout(radv_device_to_handle(device),
+					  &copy_pl_create_info,
+					  &device->meta_state.alloc,
+					  &device->meta_state.buffer.copy_p_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	VkPipelineShaderStageCreateInfo fill_pipeline_shader_stage = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+		.stage = VK_SHADER_STAGE_COMPUTE_BIT,
+		.module = radv_shader_module_to_handle(&fill_cs),
+		.pName = "main",
+		.pSpecializationInfo = NULL,
+	};
+
+	VkComputePipelineCreateInfo fill_vk_pipeline_info = {
+		.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+		.stage = fill_pipeline_shader_stage,
+		.flags = 0,
+		.layout = device->meta_state.buffer.fill_p_layout,
+	};
+
+	result = radv_CreateComputePipelines(radv_device_to_handle(device),
+					     radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					     1, &fill_vk_pipeline_info, NULL,
+					     &device->meta_state.buffer.fill_pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	VkPipelineShaderStageCreateInfo copy_pipeline_shader_stage = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+		.stage = VK_SHADER_STAGE_COMPUTE_BIT,
+		.module = radv_shader_module_to_handle(&copy_cs),
+		.pName = "main",
+		.pSpecializationInfo = NULL,
+	};
+
+	VkComputePipelineCreateInfo copy_vk_pipeline_info = {
+		.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+		.stage = copy_pipeline_shader_stage,
+		.flags = 0,
+		.layout = device->meta_state.buffer.copy_p_layout,
+	};
+
+	result = radv_CreateComputePipelines(radv_device_to_handle(device),
+					     radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					     1, &copy_vk_pipeline_info, NULL,
+					     &device->meta_state.buffer.copy_pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	ralloc_free(fill_cs.nir);
+	ralloc_free(copy_cs.nir);
+	return VK_SUCCESS;
+fail:
+	radv_device_finish_meta_buffer_state(device);
+	ralloc_free(fill_cs.nir);
+	ralloc_free(copy_cs.nir);
+	return result;
+}
+
+void radv_device_finish_meta_buffer_state(struct radv_device *device)
+{
+	if (device->meta_state.buffer.copy_pipeline)
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.buffer.copy_pipeline,
+				     &device->meta_state.alloc);
+
+	if (device->meta_state.buffer.fill_pipeline)
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.buffer.fill_pipeline,
+				     &device->meta_state.alloc);
+
+	if (device->meta_state.buffer.copy_p_layout)
+		radv_DestroyPipelineLayout(radv_device_to_handle(device),
+					   device->meta_state.buffer.copy_p_layout,
+					   &device->meta_state.alloc);
+
+	if (device->meta_state.buffer.fill_p_layout)
+		radv_DestroyPipelineLayout(radv_device_to_handle(device),
+					   device->meta_state.buffer.fill_p_layout,
+					   &device->meta_state.alloc);
+
+	if (device->meta_state.buffer.copy_ds_layout)
+		radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
+						device->meta_state.buffer.copy_ds_layout,
+						&device->meta_state.alloc);
+
+	if (device->meta_state.buffer.fill_ds_layout)
+		radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
+						device->meta_state.buffer.fill_ds_layout,
+						&device->meta_state.alloc);
+}
+
+static void fill_buffer_shader(struct radv_cmd_buffer *cmd_buffer,
+			       struct radeon_winsys_bo *bo,
+			       uint64_t offset, uint64_t size, uint32_t value)
+{
+	struct radv_device *device = cmd_buffer->device;
+	uint64_t block_count = round_up_u64(size, 1024);
+	struct radv_meta_saved_compute_state saved_state;
+	VkDescriptorSet ds;
+
+	radv_meta_save_compute(&saved_state, cmd_buffer, 4);
+
+	radv_temp_descriptor_set_create(device, cmd_buffer,
+					device->meta_state.buffer.fill_ds_layout,
+					&ds);
+
+	struct radv_buffer dst_buffer = {
+		.bo = bo,
+		.offset = offset,
+		.size = size
+	};
+
+	radv_UpdateDescriptorSets(radv_device_to_handle(device),
+				  1, /* writeCount */
+				  (VkWriteDescriptorSet[]) {
+					  {
+						  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+						  .dstSet = ds,
+						  .dstBinding = 0,
+						  .dstArrayElement = 0,
+						  .descriptorCount = 1,
+						  .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+						  .pBufferInfo = &(VkDescriptorBufferInfo) {
+							.buffer = radv_buffer_to_handle(&dst_buffer),
+							.offset = 0,
+							.range = size
+						  }
+					  }
+				  }, 0, NULL);
+
+	radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+			     VK_PIPELINE_BIND_POINT_COMPUTE,
+			     device->meta_state.buffer.fill_pipeline);
+
+	radv_CmdBindDescriptorSets(radv_cmd_buffer_to_handle(cmd_buffer),
+				   VK_PIPELINE_BIND_POINT_COMPUTE,
+				   device->meta_state.buffer.fill_p_layout, 0, 1,
+				   &ds, 0, NULL);
+
+	radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
+			      device->meta_state.buffer.fill_p_layout,
+			      VK_SHADER_STAGE_COMPUTE_BIT, 0, 4,
+			      &value);
+
+	radv_CmdDispatch(radv_cmd_buffer_to_handle(cmd_buffer), block_count, 1, 1);
+
+	radv_temp_descriptor_set_destroy(device, ds);
+
+	radv_meta_restore_compute(&saved_state, cmd_buffer, 4);
+}
+
+static void copy_buffer_shader(struct radv_cmd_buffer *cmd_buffer,
+			       struct radeon_winsys_bo *src_bo,
+			       struct radeon_winsys_bo *dst_bo,
+			       uint64_t src_offset, uint64_t dst_offset,
+			       uint64_t size)
+{
+	struct radv_device *device = cmd_buffer->device;
+	uint64_t block_count = round_up_u64(size, 1024);
+	struct radv_meta_saved_compute_state saved_state;
+	VkDescriptorSet ds;
+
+	radv_meta_save_compute(&saved_state, cmd_buffer, 0);
+
+	radv_temp_descriptor_set_create(device, cmd_buffer,
+					device->meta_state.buffer.copy_ds_layout,
+					&ds);
+
+	struct radv_buffer dst_buffer = {
+		.bo = dst_bo,
+		.offset = dst_offset,
+		.size = size
+	};
+
+	struct radv_buffer src_buffer = {
+		.bo = src_bo,
+		.offset = src_offset,
+		.size = size
+	};
+
+	radv_UpdateDescriptorSets(radv_device_to_handle(device),
+				  2, /* writeCount */
+				  (VkWriteDescriptorSet[]) {
+					  {
+						  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+						  .dstSet = ds,
+						  .dstBinding = 0,
+						  .dstArrayElement = 0,
+						  .descriptorCount = 1,
+						  .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+						  .pBufferInfo = &(VkDescriptorBufferInfo) {
+							.buffer = radv_buffer_to_handle(&dst_buffer),
+							.offset = 0,
+							.range = size
+						  }
+					  },
+					  {
+						  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+						  .dstSet = ds,
+						  .dstBinding = 1,
+						  .dstArrayElement = 0,
+						  .descriptorCount = 1,
+						  .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+						  .pBufferInfo = &(VkDescriptorBufferInfo) {
+							.buffer = radv_buffer_to_handle(&src_buffer),
+							.offset = 0,
+							.range = size
+						  }
+					  }
+				  }, 0, NULL);
+
+	radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+			     VK_PIPELINE_BIND_POINT_COMPUTE,
+			     device->meta_state.buffer.copy_pipeline);
+
+	radv_CmdBindDescriptorSets(radv_cmd_buffer_to_handle(cmd_buffer),
+				   VK_PIPELINE_BIND_POINT_COMPUTE,
+				   device->meta_state.buffer.copy_p_layout, 0, 1,
+				   &ds, 0, NULL);
+
+
+	radv_CmdDispatch(radv_cmd_buffer_to_handle(cmd_buffer), block_count, 1, 1);
+
+	radv_temp_descriptor_set_destroy(device, ds);
+
+	radv_meta_restore_compute(&saved_state, cmd_buffer, 0);
+}
+
+
+void radv_fill_buffer(struct radv_cmd_buffer *cmd_buffer,
+		      struct radeon_winsys_bo *bo,
+		      uint64_t offset, uint64_t size, uint32_t value)
+{
+	assert(!(offset & 3));
+	assert(!(size & 3));
+
+	if (size >= 4096)
+		fill_buffer_shader(cmd_buffer, bo, offset, size, value);
+	else if (size) {
+		uint64_t va = cmd_buffer->device->ws->buffer_get_va(bo);
+		va += offset;
+		cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, bo, 8);
+		si_cp_dma_clear_buffer(cmd_buffer, va, size, value);
+	}
+}
+
+static
+void radv_copy_buffer(struct radv_cmd_buffer *cmd_buffer,
+		      struct radeon_winsys_bo *src_bo,
+		      struct radeon_winsys_bo *dst_bo,
+		      uint64_t src_offset, uint64_t dst_offset,
+		      uint64_t size)
+{
+	if (size >= 4096 && !(size & 3) && !(src_offset & 3) && !(dst_offset & 3))
+		copy_buffer_shader(cmd_buffer, src_bo, dst_bo,
+				   src_offset, dst_offset, size);
+	else if (size) {
+		uint64_t src_va = cmd_buffer->device->ws->buffer_get_va(src_bo);
+		uint64_t dst_va = cmd_buffer->device->ws->buffer_get_va(dst_bo);
+		src_va += src_offset;
+		dst_va += dst_offset;
+
+		cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, src_bo, 8);
+		cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, dst_bo, 8);
+
+		si_cp_dma_buffer_copy(cmd_buffer, src_va, dst_va, size);
+	}
+}
+
+void radv_CmdFillBuffer(
+    VkCommandBuffer                             commandBuffer,
+    VkBuffer                                    dstBuffer,
+    VkDeviceSize                                dstOffset,
+    VkDeviceSize                                fillSize,
+    uint32_t                                    data)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_buffer, dst_buffer, dstBuffer);
+
+	if (fillSize == VK_WHOLE_SIZE)
+		fillSize = (dst_buffer->size - dstOffset) & ~3ull;
+
+	radv_fill_buffer(cmd_buffer, dst_buffer->bo, dst_buffer->offset + dstOffset,
+			 fillSize, data);
+}
+
+void radv_CmdCopyBuffer(
+	VkCommandBuffer                             commandBuffer,
+	VkBuffer                                    srcBuffer,
+	VkBuffer                                    destBuffer,
+	uint32_t                                    regionCount,
+	const VkBufferCopy*                         pRegions)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_buffer, src_buffer, srcBuffer);
+	RADV_FROM_HANDLE(radv_buffer, dest_buffer, destBuffer);
+
+	for (unsigned r = 0; r < regionCount; r++) {
+		uint64_t src_offset = src_buffer->offset + pRegions[r].srcOffset;
+		uint64_t dest_offset = dest_buffer->offset + pRegions[r].dstOffset;
+		uint64_t copy_size = pRegions[r].size;
+
+		radv_copy_buffer(cmd_buffer, src_buffer->bo, dest_buffer->bo,
+				 src_offset, dest_offset, copy_size);
+	}
+}
+
+void radv_CmdUpdateBuffer(
+	VkCommandBuffer                             commandBuffer,
+	VkBuffer                                    dstBuffer,
+	VkDeviceSize                                dstOffset,
+	VkDeviceSize                                dataSize,
+	const uint32_t*                             pData)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_buffer, dst_buffer, dstBuffer);
+	uint64_t words = dataSize / 4;
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(dst_buffer->bo);
+	va += dstOffset + dst_buffer->offset;
+
+	assert(!(dataSize & 3));
+	assert(!(va & 3));
+
+	if (dataSize < 4096) {
+		cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, dst_buffer->bo, 8);
+
+		radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, words + 4);
+
+		radeon_emit(cmd_buffer->cs, PKT3(PKT3_WRITE_DATA, 2 + words, 0));
+		radeon_emit(cmd_buffer->cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
+		                            S_370_WR_CONFIRM(1) |
+		                            S_370_ENGINE_SEL(V_370_ME));
+		radeon_emit(cmd_buffer->cs, va);
+		radeon_emit(cmd_buffer->cs, va >> 32);
+		radeon_emit_array(cmd_buffer->cs, pData, words);
+	} else {
+		uint32_t buf_offset;
+		radv_cmd_buffer_upload_data(cmd_buffer, dataSize, 32, pData, &buf_offset);
+		radv_copy_buffer(cmd_buffer, cmd_buffer->upload.upload_bo, dst_buffer->bo,
+				 buf_offset, dstOffset + dst_buffer->offset, dataSize);
+	}
+}
diff --git a/src/amd/vulkan/radv_meta_bufimage.c b/src/amd/vulkan/radv_meta_bufimage.c
new file mode 100644
index 0000000000..287ab3f257
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_bufimage.c
@@ -0,0 +1,396 @@
+#include "radv_meta.h"
+#include "nir/nir_builder.h"
+
+static nir_shader *
+build_nir_itob_compute_shader(struct radv_device *dev)
+{
+	nir_builder b;
+	const struct glsl_type *sampler_type = glsl_sampler_type(GLSL_SAMPLER_DIM_2D,
+								 false,
+								 false,
+								 GLSL_TYPE_FLOAT);
+	const struct glsl_type *img_type = glsl_sampler_type(GLSL_SAMPLER_DIM_BUF,
+							     false,
+							     false,
+							     GLSL_TYPE_FLOAT);
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, "meta_itob_cs");
+	b.shader->info.cs.local_size[0] = 16;
+	b.shader->info.cs.local_size[1] = 16;
+	b.shader->info.cs.local_size[2] = 1;
+	nir_variable *input_img = nir_variable_create(b.shader, nir_var_uniform,
+						      sampler_type, "s_tex");
+	input_img->data.descriptor_set = 0;
+	input_img->data.binding = 0;
+
+	nir_variable *output_img = nir_variable_create(b.shader, nir_var_uniform,
+						       img_type, "out_img");
+	output_img->data.descriptor_set = 0;
+	output_img->data.binding = 1;
+
+	nir_ssa_def *invoc_id = nir_load_system_value(&b, nir_intrinsic_load_local_invocation_id, 0);
+	nir_ssa_def *wg_id = nir_load_system_value(&b, nir_intrinsic_load_work_group_id, 0);
+	nir_ssa_def *block_size = nir_imm_ivec4(&b,
+						b.shader->info.cs.local_size[0],
+						b.shader->info.cs.local_size[1],
+						b.shader->info.cs.local_size[2], 0);
+
+	nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
+
+
+
+	nir_intrinsic_instr *offset = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
+	offset->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
+	offset->num_components = 2;
+	nir_ssa_dest_init(&offset->instr, &offset->dest, 2, 32, "offset");
+	nir_builder_instr_insert(&b, &offset->instr);
+
+	nir_intrinsic_instr *stride = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
+	stride->src[0] = nir_src_for_ssa(nir_imm_int(&b, 8));
+	stride->num_components = 1;
+	nir_ssa_dest_init(&stride->instr, &stride->dest, 1, 32, "stride");
+	nir_builder_instr_insert(&b, &stride->instr);
+
+	nir_ssa_def *img_coord = nir_iadd(&b, global_id, &offset->dest.ssa);
+
+	nir_tex_instr *tex = nir_tex_instr_create(b.shader, 2);
+	tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
+	tex->op = nir_texop_txf;
+	tex->src[0].src_type = nir_tex_src_coord;
+	tex->src[0].src = nir_src_for_ssa(img_coord);
+	tex->src[1].src_type = nir_tex_src_lod;
+	tex->src[1].src = nir_src_for_ssa(nir_imm_int(&b, 0));
+	tex->dest_type = nir_type_float;
+	tex->is_array = false;
+	tex->coord_components = 2;
+	tex->texture = nir_deref_var_create(tex, input_img);
+	tex->sampler = NULL;
+
+	nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
+	nir_builder_instr_insert(&b, &tex->instr);
+
+	nir_ssa_def *pos_x = nir_channel(&b, global_id, 0);
+	nir_ssa_def *pos_y = nir_channel(&b, global_id, 1);
+
+	nir_ssa_def *tmp = nir_imul(&b, pos_y, &stride->dest.ssa);
+	tmp = nir_iadd(&b, tmp, pos_x);
+
+	nir_ssa_def *coord = nir_vec4(&b, tmp, tmp, tmp, tmp);
+
+	nir_ssa_def *outval = &tex->dest.ssa;
+	nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_image_store);
+	store->src[0] = nir_src_for_ssa(coord);
+	store->src[1] = nir_src_for_ssa(nir_ssa_undef(&b, 1, 32));
+	store->src[2] = nir_src_for_ssa(outval);
+	store->variables[0] = nir_deref_var_create(store, output_img);
+
+	nir_builder_instr_insert(&b, &store->instr);
+	return b.shader;
+}
+
+/* Image to buffer - don't write use image accessors */
+static VkResult
+radv_device_init_meta_itob_state(struct radv_device *device)
+{
+	VkResult result;
+	struct radv_shader_module cs = { .nir = NULL };
+
+	zero(device->meta_state.itob);
+
+	cs.nir = build_nir_itob_compute_shader(device);
+
+	/*
+	 * two descriptors one for the image being sampled
+	 * one for the buffer being written.
+	 */
+	VkDescriptorSetLayoutCreateInfo ds_create_info = {
+		.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+		.bindingCount = 2,
+		.pBindings = (VkDescriptorSetLayoutBinding[]) {
+			{
+				.binding = 0,
+				.descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
+				.descriptorCount = 1,
+				.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+				.pImmutableSamplers = NULL
+			},
+			{
+				.binding = 1,
+				.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER,
+				.descriptorCount = 1,
+				.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+				.pImmutableSamplers = NULL
+			},
+		}
+	};
+
+	result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
+						&ds_create_info,
+						&device->meta_state.alloc,
+						&device->meta_state.itob.img_ds_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+
+	VkPipelineLayoutCreateInfo pl_create_info = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+		.setLayoutCount = 1,
+		.pSetLayouts = &device->meta_state.itob.img_ds_layout,
+		.pushConstantRangeCount = 1,
+		.pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 12},
+	};
+
+	result = radv_CreatePipelineLayout(radv_device_to_handle(device),
+					  &pl_create_info,
+					  &device->meta_state.alloc,
+					  &device->meta_state.itob.img_p_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	/* compute shader */
+
+	VkPipelineShaderStageCreateInfo pipeline_shader_stage = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+		.stage = VK_SHADER_STAGE_COMPUTE_BIT,
+		.module = radv_shader_module_to_handle(&cs),
+		.pName = "main",
+		.pSpecializationInfo = NULL,
+	};
+
+	VkComputePipelineCreateInfo vk_pipeline_info = {
+		.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+		.stage = pipeline_shader_stage,
+		.flags = 0,
+		.layout = device->meta_state.itob.img_p_layout,
+	};
+
+	result = radv_CreateComputePipelines(radv_device_to_handle(device),
+					     radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					     1, &vk_pipeline_info, NULL,
+					     &device->meta_state.itob.pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	ralloc_free(cs.nir);
+	return VK_SUCCESS;
+fail:
+	ralloc_free(cs.nir);
+	return result;
+}
+
+static void
+radv_device_finish_meta_itob_state(struct radv_device *device)
+{
+	if (device->meta_state.itob.img_p_layout) {
+		radv_DestroyPipelineLayout(radv_device_to_handle(device),
+					   device->meta_state.itob.img_p_layout,
+					   &device->meta_state.alloc);
+	}
+	if (device->meta_state.itob.img_ds_layout) {
+		radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
+						device->meta_state.itob.img_ds_layout,
+						&device->meta_state.alloc);
+	}
+	if (device->meta_state.itob.pipeline) {
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     device->meta_state.itob.pipeline,
+				     &device->meta_state.alloc);
+	}
+}
+
+void
+radv_device_finish_meta_bufimage_state(struct radv_device *device)
+{
+	radv_device_finish_meta_itob_state(device);
+}
+
+VkResult
+radv_device_init_meta_bufimage_state(struct radv_device *device)
+{
+	VkResult result;
+
+	result = radv_device_init_meta_itob_state(device);
+	if (result != VK_SUCCESS)
+		return result;
+	return VK_SUCCESS;
+}
+
+void
+radv_meta_begin_bufimage(struct radv_cmd_buffer *cmd_buffer,
+			 struct radv_meta_saved_compute_state *save)
+{
+	radv_meta_save_compute(save, cmd_buffer, 12);
+}
+
+void
+radv_meta_end_bufimage(struct radv_cmd_buffer *cmd_buffer,
+		       struct radv_meta_saved_compute_state *save)
+{
+	radv_meta_restore_compute(save, cmd_buffer, 12);
+}
+
+static void
+create_iview(struct radv_cmd_buffer *cmd_buffer,
+             struct radv_meta_blit2d_surf *surf,
+             VkImageUsageFlags usage,
+             struct radv_image_view *iview)
+{
+
+	radv_image_view_init(iview, cmd_buffer->device,
+			     &(VkImageViewCreateInfo) {
+				     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+					     .image = radv_image_to_handle(surf->image),
+					     .viewType = VK_IMAGE_VIEW_TYPE_2D,
+					     .format = surf->format,
+					     .subresourceRange = {
+					     .aspectMask = surf->aspect_mask,
+					     .baseMipLevel = surf->level,
+					     .levelCount = 1,
+					     .baseArrayLayer = surf->layer,
+					     .layerCount = 1
+				     },
+					     }, cmd_buffer, usage);
+}
+
+static void
+create_bview(struct radv_cmd_buffer *cmd_buffer,
+	     struct radv_buffer *buffer,
+	     unsigned offset,
+	     VkFormat format,
+	     struct radv_buffer_view *bview)
+{
+	radv_buffer_view_init(bview, cmd_buffer->device,
+			      &(VkBufferViewCreateInfo) {
+				      .sType = VK_STRUCTURE_TYPE_BUFFER_VIEW_CREATE_INFO,
+				      .flags = 0,
+				      .buffer = radv_buffer_to_handle(buffer),
+				      .format = format,
+				      .offset = offset,
+				      .range = VK_WHOLE_SIZE,
+			      }, cmd_buffer);
+
+}
+
+struct itob_temps {
+	struct radv_image_view src_iview;
+
+	struct radv_buffer_view dst_bview;
+	VkDescriptorSet set;
+};
+
+static void
+itob_bind_src_image(struct radv_cmd_buffer *cmd_buffer,
+		   struct radv_meta_blit2d_surf *src,
+		   struct radv_meta_blit2d_rect *rect,
+		   struct itob_temps *tmp)
+{
+	create_iview(cmd_buffer, src, VK_IMAGE_USAGE_SAMPLED_BIT, &tmp->src_iview);
+}
+
+static void
+itob_bind_dst_buffer(struct radv_cmd_buffer *cmd_buffer,
+		     struct radv_meta_blit2d_buffer *dst,
+		     struct radv_meta_blit2d_rect *rect,
+		     struct itob_temps *tmp)
+{
+	create_bview(cmd_buffer, dst->buffer, dst->offset, dst->format, &tmp->dst_bview);
+}
+
+static void
+itob_bind_descriptors(struct radv_cmd_buffer *cmd_buffer,
+		      struct itob_temps *tmp)
+{
+	struct radv_device *device = cmd_buffer->device;
+	VkDevice vk_device = radv_device_to_handle(cmd_buffer->device);
+
+	radv_temp_descriptor_set_create(device, cmd_buffer,
+					device->meta_state.itob.img_ds_layout,
+					&tmp->set);
+
+	radv_UpdateDescriptorSets(vk_device,
+				  2, /* writeCount */
+				  (VkWriteDescriptorSet[]) {
+					  {
+						  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+						  .dstSet = tmp->set,
+						  .dstBinding = 0,
+						  .dstArrayElement = 0,
+						  .descriptorCount = 1,
+						  .descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
+						  .pImageInfo = (VkDescriptorImageInfo[]) {
+							  {
+								  .sampler = NULL,
+								  .imageView = radv_image_view_to_handle(&tmp->src_iview),
+								  .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
+							  },
+						  }
+					  },
+					  {
+						  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+						  .dstSet = tmp->set,
+						  .dstBinding = 1,
+						  .dstArrayElement = 0,
+						  .descriptorCount = 1,
+						  .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER,
+						  .pTexelBufferView = (VkBufferView[])  { radv_buffer_view_to_handle(&tmp->dst_bview) },
+					  }
+				  }, 0, NULL);
+
+	radv_CmdBindDescriptorSets(radv_cmd_buffer_to_handle(cmd_buffer),
+				   VK_PIPELINE_BIND_POINT_COMPUTE,
+				   device->meta_state.itob.img_p_layout, 0, 1,
+				   &tmp->set, 0, NULL);
+}
+
+static void
+itob_unbind_src_image(struct radv_cmd_buffer *cmd_buffer,
+		      struct itob_temps *temps)
+{
+}
+
+static void
+bind_pipeline(struct radv_cmd_buffer *cmd_buffer)
+{
+	VkPipeline pipeline =
+		cmd_buffer->device->meta_state.itob.pipeline;
+
+	if (cmd_buffer->state.compute_pipeline != radv_pipeline_from_handle(pipeline)) {
+		radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+				     VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
+	}
+}
+
+void
+radv_meta_image_to_buffer(struct radv_cmd_buffer *cmd_buffer,
+			  struct radv_meta_blit2d_surf *src,
+			  struct radv_meta_blit2d_buffer *dst,
+			  unsigned num_rects,
+			  struct radv_meta_blit2d_rect *rects)
+{
+	struct radv_device *device = cmd_buffer->device;
+
+	for (unsigned r = 0; r < num_rects; ++r) {
+		struct itob_temps temps;
+
+		itob_bind_src_image(cmd_buffer, src, &rects[r], &temps);
+		itob_bind_dst_buffer(cmd_buffer, dst, &rects[r], &temps);
+		itob_bind_descriptors(cmd_buffer, &temps);
+
+		bind_pipeline(cmd_buffer);
+
+		unsigned push_constants[3] = {
+			rects[r].src_x,
+			rects[r].src_y,
+			dst->pitch
+		};
+		radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
+				      device->meta_state.itob.img_p_layout,
+				      VK_SHADER_STAGE_COMPUTE_BIT, 0, 12,
+				      push_constants);
+
+		radv_unaligned_dispatch(cmd_buffer, rects[r].width, rects[r].height, 1);
+		radv_temp_descriptor_set_destroy(cmd_buffer->device, temps.set);
+		itob_unbind_src_image(cmd_buffer, &temps);
+	}
+
+}
diff --git a/src/amd/vulkan/radv_meta_clear.c b/src/amd/vulkan/radv_meta_clear.c
new file mode 100644
index 0000000000..1c1b8c6b1e
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_clear.c
@@ -0,0 +1,1192 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "radv_meta.h"
+#include "radv_private.h"
+#include "nir/nir_builder.h"
+
+#include "util/format_rgb9e5.h"
+#include "vk_format.h"
+/** Vertex attributes for color clears.  */
+struct color_clear_vattrs {
+	float position[2];
+	VkClearColorValue color;
+};
+
+/** Vertex attributes for depthstencil clears.  */
+struct depthstencil_clear_vattrs {
+	float position[2];
+	float depth_clear;
+};
+
+enum {
+	DEPTH_CLEAR_SLOW,
+	DEPTH_CLEAR_FAST_EXPCLEAR,
+	DEPTH_CLEAR_FAST_NO_EXPCLEAR
+};
+
+static void
+build_color_shaders(struct nir_shader **out_vs,
+                    struct nir_shader **out_fs,
+                    uint32_t frag_output)
+{
+	nir_builder vs_b;
+	nir_builder fs_b;
+
+	nir_builder_init_simple_shader(&vs_b, NULL, MESA_SHADER_VERTEX, NULL);
+	nir_builder_init_simple_shader(&fs_b, NULL, MESA_SHADER_FRAGMENT, NULL);
+
+	vs_b.shader->info.name = ralloc_strdup(vs_b.shader, "meta_clear_color_vs");
+	fs_b.shader->info.name = ralloc_strdup(fs_b.shader, "meta_clear_color_fs");
+
+	const struct glsl_type *position_type = glsl_vec4_type();
+	const struct glsl_type *color_type = glsl_vec4_type();
+
+	nir_variable *vs_in_pos =
+		nir_variable_create(vs_b.shader, nir_var_shader_in, position_type,
+				    "a_position");
+	vs_in_pos->data.location = VERT_ATTRIB_GENERIC0;
+
+	nir_variable *vs_out_pos =
+		nir_variable_create(vs_b.shader, nir_var_shader_out, position_type,
+				    "gl_Position");
+	vs_out_pos->data.location = VARYING_SLOT_POS;
+
+	nir_variable *vs_in_color =
+		nir_variable_create(vs_b.shader, nir_var_shader_in, color_type,
+				    "a_color");
+	vs_in_color->data.location = VERT_ATTRIB_GENERIC1;
+
+	nir_variable *vs_out_color =
+		nir_variable_create(vs_b.shader, nir_var_shader_out, color_type,
+				    "v_color");
+	vs_out_color->data.location = VARYING_SLOT_VAR0;
+	vs_out_color->data.interpolation = INTERP_MODE_FLAT;
+
+	nir_variable *fs_in_color =
+		nir_variable_create(fs_b.shader, nir_var_shader_in, color_type,
+				    "v_color");
+	fs_in_color->data.location = vs_out_color->data.location;
+	fs_in_color->data.interpolation = vs_out_color->data.interpolation;
+
+	nir_variable *fs_out_color =
+		nir_variable_create(fs_b.shader, nir_var_shader_out, color_type,
+				    "f_color");
+	fs_out_color->data.location = FRAG_RESULT_DATA0 + frag_output;
+
+	nir_copy_var(&vs_b, vs_out_pos, vs_in_pos);
+	nir_copy_var(&vs_b, vs_out_color, vs_in_color);
+	nir_copy_var(&fs_b, fs_out_color, fs_in_color);
+
+	*out_vs = vs_b.shader;
+	*out_fs = fs_b.shader;
+}
+
+static VkResult
+create_pipeline(struct radv_device *device,
+		struct radv_render_pass *render_pass,
+		uint32_t samples,
+                struct nir_shader *vs_nir,
+                struct nir_shader *fs_nir,
+                const VkPipelineVertexInputStateCreateInfo *vi_state,
+                const VkPipelineDepthStencilStateCreateInfo *ds_state,
+                const VkPipelineColorBlendStateCreateInfo *cb_state,
+		const struct radv_graphics_pipeline_create_info *extra,
+                const VkAllocationCallbacks *alloc,
+                struct radv_pipeline **pipeline)
+{
+	VkDevice device_h = radv_device_to_handle(device);
+	VkResult result;
+
+	struct radv_shader_module vs_m = { .nir = vs_nir };
+	struct radv_shader_module fs_m = { .nir = fs_nir };
+
+	VkPipeline pipeline_h = VK_NULL_HANDLE;
+	result = radv_graphics_pipeline_create(device_h,
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &(VkGraphicsPipelineCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+							       .stageCount = fs_nir ? 2 : 1,
+							       .pStages = (VkPipelineShaderStageCreateInfo[]) {
+							       {
+								       .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+								       .stage = VK_SHADER_STAGE_VERTEX_BIT,
+								       .module = radv_shader_module_to_handle(&vs_m),
+								       .pName = "main",
+							       },
+							       {
+								       .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+								       .stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+								       .module = radv_shader_module_to_handle(&fs_m),
+								       .pName = "main",
+							       },
+						       },
+							       .pVertexInputState = vi_state,
+									.pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+							       .sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+							       .topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+							       .primitiveRestartEnable = false,
+						       },
+									.pViewportState = &(VkPipelineViewportStateCreateInfo) {
+							       .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+							       .viewportCount = 0,
+							       .scissorCount = 0,
+						       },
+										 .pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+							       .sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+							       .rasterizerDiscardEnable = false,
+							       .polygonMode = VK_POLYGON_MODE_FILL,
+							       .cullMode = VK_CULL_MODE_NONE,
+							       .frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE,
+							       .depthBiasEnable = false,
+						       },
+											  .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+							       .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+							       .rasterizationSamples = samples,
+							       .sampleShadingEnable = false,
+							       .pSampleMask = NULL,
+							       .alphaToCoverageEnable = false,
+							       .alphaToOneEnable = false,
+						       },
+												   .pDepthStencilState = ds_state,
+													    .pColorBlendState = cb_state,
+													    .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+							       /* The meta clear pipeline declares all state as dynamic.
+								* As a consequence, vkCmdBindPipeline writes no dynamic state
+								* to the cmd buffer. Therefore, at the end of the meta clear,
+								* we need only restore dynamic state was vkCmdSet.
+								*/
+							       .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+							       .dynamicStateCount = 6,
+							       .pDynamicStates = (VkDynamicState[]) {
+								       /* Everything except stencil write mask */
+								       VK_DYNAMIC_STATE_LINE_WIDTH,
+								       VK_DYNAMIC_STATE_DEPTH_BIAS,
+								       VK_DYNAMIC_STATE_BLEND_CONSTANTS,
+								       VK_DYNAMIC_STATE_DEPTH_BOUNDS,
+								       VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK,
+								       VK_DYNAMIC_STATE_STENCIL_REFERENCE,
+							       },
+						       },
+													    .flags = 0,
+														     .renderPass = radv_render_pass_to_handle(render_pass),
+														     .subpass = 0,
+														     },
+					       extra,
+					       alloc,
+					       &pipeline_h);
+
+	ralloc_free(vs_nir);
+	ralloc_free(fs_nir);
+
+	*pipeline = radv_pipeline_from_handle(pipeline_h);
+
+	return result;
+}
+
+static VkResult
+create_color_pipeline(struct radv_device *device,
+                      VkFormat vk_format,
+		      uint32_t samples,
+                      uint32_t frag_output,
+                      struct radv_pipeline **pipeline,
+		      VkRenderPass *pass)
+{
+	struct nir_shader *vs_nir;
+	struct nir_shader *fs_nir;
+	VkResult result;
+	build_color_shaders(&vs_nir, &fs_nir, frag_output);
+
+	const VkPipelineVertexInputStateCreateInfo vi_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+		.vertexBindingDescriptionCount = 1,
+		.pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
+			{
+				.binding = 0,
+				.stride = sizeof(struct color_clear_vattrs),
+				.inputRate = VK_VERTEX_INPUT_RATE_VERTEX
+			},
+		},
+		.vertexAttributeDescriptionCount = 2,
+		.pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
+			{
+				/* Position */
+				.location = 0,
+				.binding = 0,
+				.format = VK_FORMAT_R32G32_SFLOAT,
+				.offset = offsetof(struct color_clear_vattrs, position),
+			},
+			{
+				/* Color */
+				.location = 1,
+				.binding = 0,
+				.format = VK_FORMAT_R32G32B32A32_SFLOAT,
+				.offset = offsetof(struct color_clear_vattrs, color),
+			},
+		},
+	};
+
+	const VkPipelineDepthStencilStateCreateInfo ds_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
+		.depthTestEnable = false,
+		.depthWriteEnable = false,
+		.depthBoundsTestEnable = false,
+		.stencilTestEnable = false,
+	};
+
+	VkPipelineColorBlendAttachmentState blend_attachment_state[MAX_RTS] = { 0 };
+	blend_attachment_state[frag_output] = (VkPipelineColorBlendAttachmentState) {
+		.blendEnable = false,
+		.colorWriteMask = VK_COLOR_COMPONENT_A_BIT |
+		VK_COLOR_COMPONENT_R_BIT |
+		VK_COLOR_COMPONENT_G_BIT |
+		VK_COLOR_COMPONENT_B_BIT,
+	};
+
+	const VkPipelineColorBlendStateCreateInfo cb_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+		.logicOpEnable = false,
+		.attachmentCount = MAX_RTS,
+		.pAttachments = blend_attachment_state
+	};
+
+	result = radv_CreateRenderPass(radv_device_to_handle(device),
+				       &(VkRenderPassCreateInfo) {
+					       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+						       .attachmentCount = 1,
+						       .pAttachments = &(VkAttachmentDescription) {
+						       .format = vk_format,
+						       .samples = samples,
+						       .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+						       .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+						       .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
+						       .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
+					       },
+						       .subpassCount = 1,
+								.pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 1,
+						       .pColorAttachments = &(VkAttachmentReference) {
+							       .attachment = 0,
+							       .layout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+						       .pResolveAttachments = NULL,
+						       .pDepthStencilAttachment = &(VkAttachmentReference) {
+							       .attachment = VK_ATTACHMENT_UNUSED,
+							       .layout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+						       .preserveAttachmentCount = 1,
+						       .pPreserveAttachments = (uint32_t[]) { 0 },
+					       },
+								.dependencyCount = 0,
+									 }, &device->meta_state.alloc, pass);
+
+	if (result != VK_SUCCESS)
+		return result;
+	struct radv_graphics_pipeline_create_info extra = {
+		.use_rectlist = true,
+	};
+	result = create_pipeline(device, radv_render_pass_from_handle(*pass),
+				 samples, vs_nir, fs_nir, &vi_state, &ds_state, &cb_state,
+				 &extra, &device->meta_state.alloc, pipeline);
+
+	return result;
+}
+
+static void
+destroy_pipeline(struct radv_device *device, struct radv_pipeline *pipeline)
+{
+	if (!pipeline)
+		return;
+
+	RADV_CALL(DestroyPipeline)(radv_device_to_handle(device),
+				   radv_pipeline_to_handle(pipeline),
+				   &device->meta_state.alloc);
+
+}
+
+static void
+destroy_render_pass(struct radv_device *device, VkRenderPass renderpass)
+{
+	RADV_CALL(DestroyRenderPass)(radv_device_to_handle(device), renderpass,
+				     &device->meta_state.alloc);
+}
+
+void
+radv_device_finish_meta_clear_state(struct radv_device *device)
+{
+	struct radv_meta_state *state = &device->meta_state;
+
+	for (uint32_t i = 0; i < ARRAY_SIZE(state->clear); ++i) {
+		for (uint32_t j = 0; j < ARRAY_SIZE(state->clear[i].color_pipelines); ++j) {
+			destroy_pipeline(device, state->clear[i].color_pipelines[j]);
+			destroy_render_pass(device, state->clear[i].render_pass[j]);
+		}
+
+		for (uint32_t j = 0; j < NUM_DEPTH_CLEAR_PIPELINES; j++) {
+			destroy_pipeline(device, state->clear[i].depth_only_pipeline[j]);
+			destroy_render_pass(device, state->clear[i].depth_only_rp[j]);
+			destroy_pipeline(device, state->clear[i].stencil_only_pipeline[j]);
+			destroy_render_pass(device, state->clear[i].stencil_only_rp[j]);
+			destroy_pipeline(device, state->clear[i].depthstencil_pipeline[j]);
+			destroy_render_pass(device, state->clear[i].depthstencil_rp[j]);
+		}
+	}
+
+}
+
+static void
+emit_color_clear(struct radv_cmd_buffer *cmd_buffer,
+                 const VkClearAttachment *clear_att,
+                 const VkClearRect *clear_rect)
+{
+	struct radv_device *device = cmd_buffer->device;
+	const struct radv_subpass *subpass = cmd_buffer->state.subpass;
+	const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
+	const uint32_t subpass_att = clear_att->colorAttachment;
+	const uint32_t pass_att = subpass->color_attachments[subpass_att].attachment;
+	const struct radv_image_view *iview = fb->attachments[pass_att].attachment;
+	const uint32_t samples = iview->image->samples;
+	const uint32_t samples_log2 = ffs(samples) - 1;
+	unsigned fs_key = radv_format_meta_fs_key(iview->vk_format);
+	struct radv_pipeline *pipeline;
+	VkClearColorValue clear_value = clear_att->clearValue.color;
+	VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
+	VkPipeline pipeline_h;
+	uint32_t offset;
+
+	if (fs_key == -1) {
+		radv_finishme("color clears incomplete");
+		return;
+	}
+	pipeline = device->meta_state.clear[samples_log2].color_pipelines[fs_key];
+	pipeline_h = radv_pipeline_to_handle(pipeline);
+
+	if (!pipeline) {
+		radv_finishme("color clears incomplete");
+		return;
+	}
+	assert(samples_log2 < ARRAY_SIZE(device->meta_state.clear));
+	assert(pipeline);
+	assert(clear_att->aspectMask == VK_IMAGE_ASPECT_COLOR_BIT);
+	assert(clear_att->colorAttachment < subpass->color_count);
+
+	const struct color_clear_vattrs vertex_data[3] = {
+		{
+			.position = {
+				clear_rect->rect.offset.x,
+				clear_rect->rect.offset.y,
+			},
+			.color = clear_value,
+		},
+		{
+			.position = {
+				clear_rect->rect.offset.x,
+				clear_rect->rect.offset.y + clear_rect->rect.extent.height,
+			},
+			.color = clear_value,
+		},
+		{
+			.position = {
+				clear_rect->rect.offset.x + clear_rect->rect.extent.width,
+				clear_rect->rect.offset.y,
+			},
+			.color = clear_value,
+		},
+	};
+
+	struct radv_subpass clear_subpass = {
+		.color_count = 1,
+		.color_attachments = (VkAttachmentReference[]) {
+			subpass->color_attachments[clear_att->colorAttachment]
+		},
+		.depth_stencil_attachment = (VkAttachmentReference) { VK_ATTACHMENT_UNUSED, VK_IMAGE_LAYOUT_UNDEFINED }
+	};
+
+	radv_cmd_buffer_set_subpass(cmd_buffer, &clear_subpass, false);
+
+	radv_cmd_buffer_upload_data(cmd_buffer, sizeof(vertex_data), 16, vertex_data, &offset);
+	struct radv_buffer vertex_buffer = {
+		.device = device,
+		.size = sizeof(vertex_data),
+		.bo = cmd_buffer->upload.upload_bo,
+		.offset = offset,
+	};
+
+
+	RADV_CALL(CmdBindVertexBuffers)(cmd_buffer_h, 0, 1,
+					(VkBuffer[]) { radv_buffer_to_handle(&vertex_buffer) },
+					(VkDeviceSize[]) { 0 });
+
+	if (cmd_buffer->state.pipeline != pipeline) {
+		RADV_CALL(CmdBindPipeline)(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
+					   pipeline_h);
+	}
+
+	RADV_CALL(CmdDraw)(cmd_buffer_h, 3, 1, 0, 0);
+
+	radv_cmd_buffer_set_subpass(cmd_buffer, subpass, false);
+}
+
+
+static void
+build_depthstencil_shader(struct nir_shader **out_vs, struct nir_shader **out_fs)
+{
+	nir_builder vs_b, fs_b;
+
+	nir_builder_init_simple_shader(&vs_b, NULL, MESA_SHADER_VERTEX, NULL);
+	nir_builder_init_simple_shader(&fs_b, NULL, MESA_SHADER_FRAGMENT, NULL);
+
+	vs_b.shader->info.name = ralloc_strdup(vs_b.shader, "meta_clear_depthstencil_vs");
+	fs_b.shader->info.name = ralloc_strdup(fs_b.shader, "meta_clear_depthstencil_fs");
+	const struct glsl_type *position_type = glsl_vec4_type();
+
+	nir_variable *vs_in_pos =
+		nir_variable_create(vs_b.shader, nir_var_shader_in, position_type,
+				    "a_position");
+	vs_in_pos->data.location = VERT_ATTRIB_GENERIC0;
+
+	nir_variable *vs_out_pos =
+		nir_variable_create(vs_b.shader, nir_var_shader_out, position_type,
+				    "gl_Position");
+	vs_out_pos->data.location = VARYING_SLOT_POS;
+
+	nir_copy_var(&vs_b, vs_out_pos, vs_in_pos);
+
+	*out_vs = vs_b.shader;
+	*out_fs = fs_b.shader;
+}
+
+static VkResult
+create_depthstencil_pipeline(struct radv_device *device,
+                             VkImageAspectFlags aspects,
+			     uint32_t samples,
+			     int index,
+                             struct radv_pipeline **pipeline,
+			     VkRenderPass *render_pass)
+{
+	struct nir_shader *vs_nir, *fs_nir;
+	VkResult result;
+	build_depthstencil_shader(&vs_nir, &fs_nir);
+
+	const VkPipelineVertexInputStateCreateInfo vi_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+		.vertexBindingDescriptionCount = 1,
+		.pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
+			{
+				.binding = 0,
+				.stride = sizeof(struct depthstencil_clear_vattrs),
+				.inputRate = VK_VERTEX_INPUT_RATE_VERTEX
+			},
+		},
+		.vertexAttributeDescriptionCount = 1,
+		.pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
+			{
+				/* Position */
+				.location = 0,
+				.binding = 0,
+				.format = VK_FORMAT_R32G32B32_SFLOAT,
+				.offset = offsetof(struct depthstencil_clear_vattrs, position),
+			},
+		},
+	};
+
+	const VkPipelineDepthStencilStateCreateInfo ds_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
+		.depthTestEnable = (aspects & VK_IMAGE_ASPECT_DEPTH_BIT),
+		.depthCompareOp = VK_COMPARE_OP_ALWAYS,
+		.depthWriteEnable = (aspects & VK_IMAGE_ASPECT_DEPTH_BIT),
+		.depthBoundsTestEnable = false,
+		.stencilTestEnable = (aspects & VK_IMAGE_ASPECT_STENCIL_BIT),
+		.front = {
+			.passOp = VK_STENCIL_OP_REPLACE,
+			.compareOp = VK_COMPARE_OP_ALWAYS,
+			.writeMask = UINT32_MAX,
+			.reference = 0, /* dynamic */
+		},
+		.back = { 0 /* dont care */ },
+	};
+
+	const VkPipelineColorBlendStateCreateInfo cb_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+		.logicOpEnable = false,
+		.attachmentCount = 0,
+		.pAttachments = NULL,
+	};
+
+	result = radv_CreateRenderPass(radv_device_to_handle(device),
+				       &(VkRenderPassCreateInfo) {
+					       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+						       .attachmentCount = 1,
+						       .pAttachments = &(VkAttachmentDescription) {
+						       .format = VK_FORMAT_UNDEFINED,
+						       .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+						       .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+						       .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
+						       .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
+					       },
+						       .subpassCount = 1,
+								.pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 0,
+						       .pColorAttachments = NULL,
+						       .pResolveAttachments = NULL,
+						       .pDepthStencilAttachment = &(VkAttachmentReference) {
+							       .attachment = 0,
+							       .layout = VK_IMAGE_LAYOUT_GENERAL,
+						       },
+						       .preserveAttachmentCount = 1,
+						       .pPreserveAttachments = (uint32_t[]) { 0 },
+					       },
+								.dependencyCount = 0,
+									 }, &device->meta_state.alloc, render_pass);
+	if (result != VK_SUCCESS)
+		return result;
+
+	struct radv_graphics_pipeline_create_info extra = {
+		.use_rectlist = true,
+	};
+
+	if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
+		extra.db_depth_clear = index == DEPTH_CLEAR_SLOW ? false : true;
+		extra.db_depth_disable_expclear = index == DEPTH_CLEAR_FAST_NO_EXPCLEAR ? true : false;
+	}
+	if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
+		extra.db_stencil_clear = index == DEPTH_CLEAR_SLOW ? false : true;
+		extra.db_stencil_disable_expclear = index == DEPTH_CLEAR_FAST_NO_EXPCLEAR ? true : false;
+	}
+	result = create_pipeline(device, radv_render_pass_from_handle(*render_pass),
+				 samples, vs_nir, fs_nir, &vi_state, &ds_state, &cb_state,
+				 &extra, &device->meta_state.alloc, pipeline);
+	return result;
+}
+
+static bool depth_view_can_fast_clear(const struct radv_image_view *iview,
+				      VkImageLayout layout,
+				      const VkClearRect *clear_rect)
+{
+	if (clear_rect->rect.offset.x || clear_rect->rect.offset.y ||
+	    clear_rect->rect.extent.width != iview->extent.width ||
+	    clear_rect->rect.extent.height != iview->extent.height)
+		return false;
+	if (iview->image->htile.size &&
+	    iview->base_mip == 0 &&
+	    iview->base_layer == 0 &&
+	    radv_layout_can_expclear(iview->image, layout) &&
+	    memcmp(&iview->extent, &iview->image->extent, sizeof(iview->extent)) == 0)
+		return true;
+	return false;
+}
+
+static struct radv_pipeline *
+pick_depthstencil_pipeline(struct radv_meta_state *meta_state,
+			   const struct radv_image_view *iview,
+			   int samples_log2,
+			   VkImageAspectFlags aspects,
+			   VkImageLayout layout,
+			   const VkClearRect *clear_rect,
+			   VkClearDepthStencilValue clear_value)
+{
+	bool fast = depth_view_can_fast_clear(iview, layout, clear_rect);
+	int index = DEPTH_CLEAR_SLOW;
+
+	if (fast) {
+		/* we don't know the previous clear values, so we always have
+		 * the NO_EXPCLEAR path */
+		index = DEPTH_CLEAR_FAST_NO_EXPCLEAR;
+	}
+
+	switch (aspects) {
+	case VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT:
+		return meta_state->clear[samples_log2].depthstencil_pipeline[index];
+	case VK_IMAGE_ASPECT_DEPTH_BIT:
+		return meta_state->clear[samples_log2].depth_only_pipeline[index];
+	case VK_IMAGE_ASPECT_STENCIL_BIT:
+		return meta_state->clear[samples_log2].stencil_only_pipeline[index];
+	}
+	unreachable("expected depth or stencil aspect");
+}
+
+static void
+emit_depthstencil_clear(struct radv_cmd_buffer *cmd_buffer,
+                        const VkClearAttachment *clear_att,
+                        const VkClearRect *clear_rect)
+{
+	struct radv_device *device = cmd_buffer->device;
+	struct radv_meta_state *meta_state = &device->meta_state;
+	const struct radv_subpass *subpass = cmd_buffer->state.subpass;
+	const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
+	const uint32_t pass_att = subpass->depth_stencil_attachment.attachment;
+	VkClearDepthStencilValue clear_value = clear_att->clearValue.depthStencil;
+	VkImageAspectFlags aspects = clear_att->aspectMask;
+	const struct radv_image_view *iview = fb->attachments[pass_att].attachment;
+	const uint32_t samples = iview->image->samples;
+	const uint32_t samples_log2 = ffs(samples) - 1;
+	VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
+	uint32_t offset;
+
+	assert(aspects == VK_IMAGE_ASPECT_DEPTH_BIT ||
+	       aspects == VK_IMAGE_ASPECT_STENCIL_BIT ||
+	       aspects == (VK_IMAGE_ASPECT_DEPTH_BIT |
+			   VK_IMAGE_ASPECT_STENCIL_BIT));
+	assert(pass_att != VK_ATTACHMENT_UNUSED);
+
+	const struct depthstencil_clear_vattrs vertex_data[3] = {
+		{
+			.position = {
+				clear_rect->rect.offset.x,
+				clear_rect->rect.offset.y,
+			},
+			.depth_clear = clear_value.depth,
+		},
+		{
+			.position = {
+				clear_rect->rect.offset.x,
+				clear_rect->rect.offset.y + clear_rect->rect.extent.height,
+			},
+			.depth_clear = clear_value.depth,
+		},
+		{
+			.position = {
+				clear_rect->rect.offset.x + clear_rect->rect.extent.width,
+				clear_rect->rect.offset.y,
+			},
+			.depth_clear = clear_value.depth,
+		},
+	};
+
+	radv_cmd_buffer_upload_data(cmd_buffer, sizeof(vertex_data), 16, vertex_data, &offset);
+	struct radv_buffer vertex_buffer = {
+		.device = device,
+		.size = sizeof(vertex_data),
+		.bo = cmd_buffer->upload.upload_bo,
+		.offset = offset,
+	};
+
+	if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
+		RADV_CALL(CmdSetStencilReference)(cmd_buffer_h, VK_STENCIL_FACE_FRONT_BIT,
+						  clear_value.stencil);
+	}
+
+	RADV_CALL(CmdBindVertexBuffers)(cmd_buffer_h, 0, 1,
+					(VkBuffer[]) { radv_buffer_to_handle(&vertex_buffer) },
+					(VkDeviceSize[]) { 0 });
+
+	struct radv_pipeline *pipeline = pick_depthstencil_pipeline(meta_state,
+								    iview,
+								    samples_log2,
+								    aspects,
+								    subpass->depth_stencil_attachment.layout,
+								    clear_rect,
+								    clear_value);
+	if (cmd_buffer->state.pipeline != pipeline) {
+		RADV_CALL(CmdBindPipeline)(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
+					   radv_pipeline_to_handle(pipeline));
+	}
+
+	if (depth_view_can_fast_clear(iview, subpass->depth_stencil_attachment.layout, clear_rect))
+		radv_set_depth_clear_regs(cmd_buffer, iview->image, clear_value, aspects);
+
+	RADV_CALL(CmdDraw)(cmd_buffer_h, 3, 1, 0, 0);
+}
+
+
+static VkFormat pipeline_formats[] = {
+	VK_FORMAT_R8G8B8A8_UNORM,
+	VK_FORMAT_R8G8B8A8_UINT,
+	VK_FORMAT_R8G8B8A8_SINT,
+	VK_FORMAT_R16G16B16A16_UNORM,
+	VK_FORMAT_R16G16B16A16_SNORM,
+	VK_FORMAT_R16G16B16A16_UINT,
+	VK_FORMAT_R16G16B16A16_SINT,
+	VK_FORMAT_R32_SFLOAT,
+	VK_FORMAT_R32G32_SFLOAT,
+	VK_FORMAT_R32G32B32A32_SFLOAT
+};
+
+VkResult
+radv_device_init_meta_clear_state(struct radv_device *device)
+{
+	VkResult res;
+	struct radv_meta_state *state = &device->meta_state;
+
+	memset(&device->meta_state.clear, 0, sizeof(device->meta_state.clear));
+
+	for (uint32_t i = 0; i < ARRAY_SIZE(state->clear); ++i) {
+		uint32_t samples = 1 << i;
+		for (uint32_t j = 0; j < ARRAY_SIZE(pipeline_formats); ++j) {
+			VkFormat format = pipeline_formats[j];
+			unsigned fs_key = radv_format_meta_fs_key(format);
+			assert(!state->clear[i].color_pipelines[fs_key]);
+			res = create_color_pipeline(device, format, samples, 0, &state->clear[i].color_pipelines[fs_key],
+						    &state->clear[i].render_pass[fs_key]);
+			if (res != VK_SUCCESS)
+				goto fail;
+
+		}
+
+		for (uint32_t j = 0; j < NUM_DEPTH_CLEAR_PIPELINES; j++) {
+			res = create_depthstencil_pipeline(device,
+							   VK_IMAGE_ASPECT_DEPTH_BIT,
+							   samples,
+							   j,
+							   &state->clear[i].depth_only_pipeline[j],
+							   &state->clear[i].depth_only_rp[j]);
+			if (res != VK_SUCCESS)
+				goto fail;
+
+			res = create_depthstencil_pipeline(device,
+							   VK_IMAGE_ASPECT_STENCIL_BIT,
+							   samples,
+							   j,
+							   &state->clear[i].stencil_only_pipeline[j],
+							   &state->clear[i].stencil_only_rp[j]);
+			if (res != VK_SUCCESS)
+				goto fail;
+
+			res = create_depthstencil_pipeline(device,
+							   VK_IMAGE_ASPECT_DEPTH_BIT |
+							   VK_IMAGE_ASPECT_STENCIL_BIT,
+							   samples,
+							   j,
+							   &state->clear[i].depthstencil_pipeline[j],
+							   &state->clear[i].depthstencil_rp[j]);
+			if (res != VK_SUCCESS)
+				goto fail;
+		}
+	}
+	return VK_SUCCESS;
+
+fail:
+	radv_device_finish_meta_clear_state(device);
+	return res;
+}
+
+static bool
+emit_fast_color_clear(struct radv_cmd_buffer *cmd_buffer,
+		      const VkClearAttachment *clear_att,
+		      const VkClearRect *clear_rect)
+{
+	const struct radv_subpass *subpass = cmd_buffer->state.subpass;
+	const uint32_t subpass_att = clear_att->colorAttachment;
+	const uint32_t pass_att = subpass->color_attachments[subpass_att].attachment;
+	VkImageLayout image_layout = subpass->color_attachments[subpass_att].layout;
+	const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
+	const struct radv_image_view *iview = fb->attachments[pass_att].attachment;
+	VkClearColorValue clear_value = clear_att->clearValue.color;
+	uint32_t clear_color[2];
+	bool ret;
+
+	if (!iview->image->cmask.size && !iview->image->surface.dcc_size)
+		return false;
+
+	if (!cmd_buffer->device->allow_fast_clears)
+		return false;
+
+	if (!radv_layout_has_cmask(iview->image, image_layout))
+		goto fail;
+	if (vk_format_get_blocksizebits(iview->image->vk_format) > 64)
+		goto fail;
+
+	/* don't fast clear 3D */
+	if (iview->image->type == VK_IMAGE_TYPE_3D)
+		goto fail;
+
+	/* all layers are bound */
+	if (iview->base_layer > 0)
+		goto fail;
+	if (iview->image->array_size != iview->layer_count)
+		goto fail;
+
+	if (iview->image->levels > 1)
+		goto fail;
+
+	if (iview->image->surface.level[0].mode < RADEON_SURF_MODE_1D)
+		goto fail;
+
+	if (memcmp(&iview->extent, &iview->image->extent, sizeof(iview->extent)))
+		goto fail;
+
+	if (clear_rect->rect.offset.x || clear_rect->rect.offset.y ||
+	    clear_rect->rect.extent.width != iview->image->extent.width ||
+	    clear_rect->rect.extent.height != iview->image->extent.height)
+		goto fail;
+
+	if (clear_rect->baseArrayLayer != 0)
+		goto fail;
+	if (clear_rect->layerCount != iview->image->array_size)
+		goto fail;
+
+	/* DCC */
+	ret = radv_format_pack_clear_color(iview->image->vk_format,
+					   clear_color, &clear_value);
+	if (ret == false)
+		goto fail;
+
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+	                                RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
+	si_emit_cache_flush(cmd_buffer);
+	/* clear cmask buffer */
+	if (iview->image->surface.dcc_size) {
+		radv_fill_buffer(cmd_buffer, iview->image->bo,
+				 iview->image->offset + iview->image->dcc_offset,
+				 iview->image->surface.dcc_size, 0x20202020);
+	} else {
+		radv_fill_buffer(cmd_buffer, iview->image->bo,
+				 iview->image->offset + iview->image->cmask.offset,
+				 iview->image->cmask.size, 0);
+	}
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
+	                                RADV_CMD_FLAG_INV_VMEM_L1 |
+	                                RADV_CMD_FLAG_INV_GLOBAL_L2;
+
+	radv_set_color_clear_regs(cmd_buffer, iview->image, subpass_att, clear_color);
+
+	return true;
+fail:
+	return false;
+}
+
+/**
+ * The parameters mean that same as those in vkCmdClearAttachments.
+ */
+static void
+emit_clear(struct radv_cmd_buffer *cmd_buffer,
+           const VkClearAttachment *clear_att,
+           const VkClearRect *clear_rect)
+{
+	if (clear_att->aspectMask & VK_IMAGE_ASPECT_COLOR_BIT) {
+
+		if (!emit_fast_color_clear(cmd_buffer, clear_att, clear_rect))
+			emit_color_clear(cmd_buffer, clear_att, clear_rect);
+	} else {
+		assert(clear_att->aspectMask & (VK_IMAGE_ASPECT_DEPTH_BIT |
+						VK_IMAGE_ASPECT_STENCIL_BIT));
+		emit_depthstencil_clear(cmd_buffer, clear_att, clear_rect);
+	}
+}
+
+static bool
+subpass_needs_clear(const struct radv_cmd_buffer *cmd_buffer)
+{
+	const struct radv_cmd_state *cmd_state = &cmd_buffer->state;
+	uint32_t ds;
+
+	if (!cmd_state->subpass)
+		return false;
+	ds = cmd_state->subpass->depth_stencil_attachment.attachment;
+	for (uint32_t i = 0; i < cmd_state->subpass->color_count; ++i) {
+		uint32_t a = cmd_state->subpass->color_attachments[i].attachment;
+		if (cmd_state->attachments[a].pending_clear_aspects) {
+			return true;
+		}
+	}
+
+	if (ds != VK_ATTACHMENT_UNUSED &&
+	    cmd_state->attachments[ds].pending_clear_aspects) {
+		return true;
+	}
+
+	return false;
+}
+
+/**
+ * Emit any pending attachment clears for the current subpass.
+ *
+ * @see radv_attachment_state::pending_clear_aspects
+ */
+void
+radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer *cmd_buffer)
+{
+	struct radv_cmd_state *cmd_state = &cmd_buffer->state;
+	struct radv_meta_saved_state saved_state;
+
+	if (!subpass_needs_clear(cmd_buffer))
+		return;
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	if (cmd_state->framebuffer->layers > 1)
+		radv_finishme("clearing multi-layer framebuffer");
+
+	VkClearRect clear_rect = {
+		.rect = cmd_state->render_area,
+		.baseArrayLayer = 0,
+		.layerCount = 1, /* FINISHME: clear multi-layer framebuffer */
+	};
+
+	for (uint32_t i = 0; i < cmd_state->subpass->color_count; ++i) {
+		uint32_t a = cmd_state->subpass->color_attachments[i].attachment;
+
+		if (!cmd_state->attachments[a].pending_clear_aspects)
+			continue;
+
+		assert(cmd_state->attachments[a].pending_clear_aspects ==
+		       VK_IMAGE_ASPECT_COLOR_BIT);
+
+		VkClearAttachment clear_att = {
+			.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+			.colorAttachment = i, /* Use attachment index relative to subpass */
+			.clearValue = cmd_state->attachments[a].clear_value,
+		};
+
+		emit_clear(cmd_buffer, &clear_att, &clear_rect);
+		cmd_state->attachments[a].pending_clear_aspects = 0;
+	}
+
+	uint32_t ds = cmd_state->subpass->depth_stencil_attachment.attachment;
+
+	if (ds != VK_ATTACHMENT_UNUSED) {
+
+		if (cmd_state->attachments[ds].pending_clear_aspects) {
+
+			VkClearAttachment clear_att = {
+				.aspectMask = cmd_state->attachments[ds].pending_clear_aspects,
+				.clearValue = cmd_state->attachments[ds].clear_value,
+			};
+
+			emit_clear(cmd_buffer, &clear_att, &clear_rect);
+			cmd_state->attachments[ds].pending_clear_aspects = 0;
+		}
+	}
+
+	radv_meta_restore(&saved_state, cmd_buffer);
+}
+
+static void
+radv_cmd_clear_image(struct radv_cmd_buffer *cmd_buffer,
+		     struct radv_image *image,
+		     VkImageLayout image_layout,
+		     const VkClearValue *clear_value,
+		     uint32_t range_count,
+		     const VkImageSubresourceRange *ranges)
+{
+	VkDevice device_h = radv_device_to_handle(cmd_buffer->device);
+	VkFormat format = image->vk_format;
+	VkClearValue internal_clear_value = *clear_value;
+
+	if (format == VK_FORMAT_E5B9G9R9_UFLOAT_PACK32) {
+		uint32_t value;
+		format = VK_FORMAT_R32_UINT;
+		value = float3_to_rgb9e5(clear_value->color.float32);
+		internal_clear_value.color.uint32[0] = value;
+	}
+
+	for (uint32_t r = 0; r < range_count; r++) {
+		const VkImageSubresourceRange *range = &ranges[r];
+		for (uint32_t l = 0; l < radv_get_levelCount(image, range); ++l) {
+			const uint32_t layer_count = image->type == VK_IMAGE_TYPE_3D ?
+				radv_minify(image->extent.depth, l) :
+				radv_get_layerCount(image, range);
+			for (uint32_t s = 0; s < layer_count; ++s) {
+				struct radv_image_view iview;
+				radv_image_view_init(&iview, cmd_buffer->device,
+						     &(VkImageViewCreateInfo) {
+							     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+								     .image = radv_image_to_handle(image),
+								     .viewType = radv_meta_get_view_type(image),
+								     .format = format,
+								     .subresourceRange = {
+								     .aspectMask = range->aspectMask,
+								     .baseMipLevel = range->baseMipLevel + l,
+								     .levelCount = 1,
+								     .baseArrayLayer = range->baseArrayLayer + s,
+								     .layerCount = 1
+							     },
+								     },
+						     cmd_buffer, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
+
+				VkFramebuffer fb;
+				radv_CreateFramebuffer(device_h,
+						       &(VkFramebufferCreateInfo) {
+							       .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
+								       .attachmentCount = 1,
+								       .pAttachments = (VkImageView[]) {
+								       radv_image_view_to_handle(&iview),
+							       },
+								       .width = iview.extent.width,
+										.height = iview.extent.height,
+										.layers = 1
+										},
+						       &cmd_buffer->pool->alloc,
+						       &fb);
+
+				VkAttachmentDescription att_desc = {
+					.format = iview.vk_format,
+					.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+					.storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+					.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
+					.stencilStoreOp = VK_ATTACHMENT_STORE_OP_STORE,
+					.initialLayout = image_layout,
+					.finalLayout = image_layout,
+				};
+
+				VkSubpassDescription subpass_desc = {
+					.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+					.inputAttachmentCount = 0,
+					.colorAttachmentCount = 0,
+					.pColorAttachments = NULL,
+					.pResolveAttachments = NULL,
+					.pDepthStencilAttachment = NULL,
+					.preserveAttachmentCount = 0,
+					.pPreserveAttachments = NULL,
+				};
+
+				const VkAttachmentReference att_ref = {
+					.attachment = 0,
+					.layout = image_layout,
+				};
+
+				if (range->aspectMask & VK_IMAGE_ASPECT_COLOR_BIT) {
+					subpass_desc.colorAttachmentCount = 1;
+					subpass_desc.pColorAttachments = &att_ref;
+				} else {
+					subpass_desc.pDepthStencilAttachment = &att_ref;
+				}
+
+				VkRenderPass pass;
+				radv_CreateRenderPass(device_h,
+						      &(VkRenderPassCreateInfo) {
+							      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+								      .attachmentCount = 1,
+								      .pAttachments = &att_desc,
+								      .subpassCount = 1,
+								      .pSubpasses = &subpass_desc,
+								      },
+						      &cmd_buffer->pool->alloc,
+						      &pass);
+
+				RADV_CALL(CmdBeginRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer),
+							      &(VkRenderPassBeginInfo) {
+								      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+									      .renderArea = {
+									      .offset = { 0, 0, },
+									      .extent = {
+										      .width = iview.extent.width,
+										      .height = iview.extent.height,
+									      },
+								      },
+									      .renderPass = pass,
+										       .framebuffer = fb,
+										       .clearValueCount = 0,
+										       .pClearValues = NULL,
+										       },
+							      VK_SUBPASS_CONTENTS_INLINE);
+
+				VkClearAttachment clear_att = {
+					.aspectMask = range->aspectMask,
+					.colorAttachment = 0,
+					.clearValue = internal_clear_value,
+				};
+
+				VkClearRect clear_rect = {
+					.rect = {
+						.offset = { 0, 0 },
+						.extent = { iview.extent.width, iview.extent.height },
+					},
+					.baseArrayLayer = range->baseArrayLayer,
+					.layerCount = 1, /* FINISHME: clear multi-layer framebuffer */
+				};
+
+				emit_clear(cmd_buffer, &clear_att, &clear_rect);
+
+				RADV_CALL(CmdEndRenderPass)(radv_cmd_buffer_to_handle(cmd_buffer));
+				RADV_CALL(DestroyRenderPass)(device_h, pass,
+							     &cmd_buffer->pool->alloc);
+				RADV_CALL(DestroyFramebuffer)(device_h, fb,
+							      &cmd_buffer->pool->alloc);
+			}
+		}
+	}
+}
+
+void radv_CmdClearColorImage(
+	VkCommandBuffer                             commandBuffer,
+	VkImage                                     image_h,
+	VkImageLayout                               imageLayout,
+	const VkClearColorValue*                    pColor,
+	uint32_t                                    rangeCount,
+	const VkImageSubresourceRange*              pRanges)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_image, image, image_h);
+	struct radv_meta_saved_state saved_state;
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	radv_cmd_clear_image(cmd_buffer, image, imageLayout,
+			     (const VkClearValue *) pColor,
+			     rangeCount, pRanges);
+
+	radv_meta_restore(&saved_state, cmd_buffer);
+}
+
+void radv_CmdClearDepthStencilImage(
+	VkCommandBuffer                             commandBuffer,
+	VkImage                                     image_h,
+	VkImageLayout                               imageLayout,
+	const VkClearDepthStencilValue*             pDepthStencil,
+	uint32_t                                    rangeCount,
+	const VkImageSubresourceRange*              pRanges)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_image, image, image_h);
+	struct radv_meta_saved_state saved_state;
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	radv_cmd_clear_image(cmd_buffer, image, imageLayout,
+			     (const VkClearValue *) pDepthStencil,
+			     rangeCount, pRanges);
+
+	radv_meta_restore(&saved_state, cmd_buffer);
+}
+
+void radv_CmdClearAttachments(
+	VkCommandBuffer                             commandBuffer,
+	uint32_t                                    attachmentCount,
+	const VkClearAttachment*                    pAttachments,
+	uint32_t                                    rectCount,
+	const VkClearRect*                          pRects)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	struct radv_meta_saved_state saved_state;
+
+	if (!cmd_buffer->state.subpass)
+		return;
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	/* FINISHME: We can do better than this dumb loop. It thrashes too much
+	 * state.
+	 */
+	for (uint32_t a = 0; a < attachmentCount; ++a) {
+		for (uint32_t r = 0; r < rectCount; ++r) {
+			emit_clear(cmd_buffer, &pAttachments[a], &pRects[r]);
+		}
+	}
+
+	radv_meta_restore(&saved_state, cmd_buffer);
+}
diff --git a/src/amd/vulkan/radv_meta_copy.c b/src/amd/vulkan/radv_meta_copy.c
new file mode 100644
index 0000000000..4c01eb7acb
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_copy.c
@@ -0,0 +1,399 @@
+/*
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "radv_meta.h"
+#include "vk_format.h"
+
+static VkExtent3D
+meta_image_block_size(const struct radv_image *image)
+{
+	const struct vk_format_description *desc = vk_format_description(image->vk_format);
+	return (VkExtent3D) { desc->block.width, desc->block.height, 1 };
+}
+
+/* Returns the user-provided VkBufferImageCopy::imageExtent in units of
+ * elements rather than texels. One element equals one texel or one block
+ * if Image is uncompressed or compressed, respectively.
+ */
+static struct VkExtent3D
+meta_region_extent_el(const struct radv_image *image,
+                      const struct VkExtent3D *extent)
+{
+	const VkExtent3D block = meta_image_block_size(image);
+	return radv_sanitize_image_extent(image->type, (VkExtent3D) {
+			.width  = DIV_ROUND_UP(extent->width , block.width),
+				.height = DIV_ROUND_UP(extent->height, block.height),
+				.depth  = DIV_ROUND_UP(extent->depth , block.depth),
+				});
+}
+
+/* Returns the user-provided VkBufferImageCopy::imageOffset in units of
+ * elements rather than texels. One element equals one texel or one block
+ * if Image is uncompressed or compressed, respectively.
+ */
+static struct VkOffset3D
+meta_region_offset_el(const struct radv_image *image,
+                      const struct VkOffset3D *offset)
+{
+	const VkExtent3D block = meta_image_block_size(image);
+	return radv_sanitize_image_offset(image->type, (VkOffset3D) {
+			.x = offset->x / block.width,
+				.y = offset->y / block.height,
+				.z = offset->z / block.depth,
+				});
+}
+
+static VkFormat
+vk_format_for_size(int bs)
+{
+	switch (bs) {
+	case 1: return VK_FORMAT_R8_UINT;
+	case 2: return VK_FORMAT_R8G8_UINT;
+	case 4: return VK_FORMAT_R8G8B8A8_UINT;
+	case 8: return VK_FORMAT_R16G16B16A16_UINT;
+	case 16: return VK_FORMAT_R32G32B32A32_UINT;
+	default:
+		unreachable("Invalid format block size");
+	}
+}
+
+static struct radv_meta_blit2d_surf
+blit_surf_for_image_level_layer(struct radv_image* image, VkImageAspectFlags aspectMask,
+				int level, int layer)
+{
+	VkFormat format = image->vk_format;
+	if (aspectMask & VK_IMAGE_ASPECT_DEPTH_BIT)
+		format = vk_format_depth_only(format);
+	else if (aspectMask & VK_IMAGE_ASPECT_STENCIL_BIT)
+		format = vk_format_stencil_only(format);
+
+	if (!image->surface.dcc_size)
+		format = vk_format_for_size(vk_format_get_blocksize(format));
+
+	return (struct radv_meta_blit2d_surf) {
+		.format = format,
+		.bs = vk_format_get_blocksize(format),
+		.level = level,
+		.layer = layer,
+		.image = image,
+		.aspect_mask = aspectMask,
+	};
+}
+
+static void
+meta_copy_buffer_to_image(struct radv_cmd_buffer *cmd_buffer,
+                          struct radv_buffer* buffer,
+                          struct radv_image* image,
+                          uint32_t regionCount,
+                          const VkBufferImageCopy* pRegions)
+{
+	struct radv_meta_saved_state saved_state;
+
+	/* The Vulkan 1.0 spec says "dstImage must have a sample count equal to
+	 * VK_SAMPLE_COUNT_1_BIT."
+	 */
+	assert(image->samples == 1);
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	for (unsigned r = 0; r < regionCount; r++) {
+
+		/**
+		 * From the Vulkan 1.0.6 spec: 18.3 Copying Data Between Images
+		 *    extent is the size in texels of the source image to copy in width,
+		 *    height and depth. 1D images use only x and width. 2D images use x, y,
+		 *    width and height. 3D images use x, y, z, width, height and depth.
+		 *
+		 *
+		 * Also, convert the offsets and extent from units of texels to units of
+		 * blocks - which is the highest resolution accessible in this command.
+		 */
+		const VkOffset3D img_offset_el =
+			meta_region_offset_el(image, &pRegions[r].imageOffset);
+		const VkExtent3D bufferExtent = {
+			.width  = pRegions[r].bufferRowLength ?
+			pRegions[r].bufferRowLength : pRegions[r].imageExtent.width,
+			.height = pRegions[r].bufferImageHeight ?
+			pRegions[r].bufferImageHeight : pRegions[r].imageExtent.height,
+		};
+		const VkExtent3D buf_extent_el =
+			meta_region_extent_el(image, &bufferExtent);
+
+		/* Start creating blit rect */
+		const VkExtent3D img_extent_el =
+			meta_region_extent_el(image, &pRegions[r].imageExtent);
+		struct radv_meta_blit2d_rect rect = {
+			.width = img_extent_el.width,
+			.height =  img_extent_el.height,
+		};
+
+		/* Create blit surfaces */
+		struct radv_meta_blit2d_surf img_bsurf =
+			blit_surf_for_image_level_layer(image,
+							pRegions[r].imageSubresource.aspectMask,
+							pRegions[r].imageSubresource.mipLevel,
+							pRegions[r].imageSubresource.baseArrayLayer);
+
+		struct radv_meta_blit2d_buffer buf_bsurf = {
+			.bs = img_bsurf.bs,
+			.format = img_bsurf.format,
+			.buffer = buffer,
+			.offset = pRegions[r].bufferOffset,
+			.pitch = buf_extent_el.width,
+		};
+
+		/* Loop through each 3D or array slice */
+		unsigned num_slices_3d = img_extent_el.depth;
+		unsigned num_slices_array = pRegions[r].imageSubresource.layerCount;
+		unsigned slice_3d = 0;
+		unsigned slice_array = 0;
+		while (slice_3d < num_slices_3d && slice_array < num_slices_array) {
+
+			rect.dst_x = img_offset_el.x;
+			rect.dst_y = img_offset_el.y;
+
+
+			/* Perform Blit */
+			radv_meta_blit2d(cmd_buffer, NULL, &buf_bsurf, &img_bsurf, 1, &rect);
+
+			/* Once we've done the blit, all of the actual information about
+			 * the image is embedded in the command buffer so we can just
+			 * increment the offset directly in the image effectively
+			 * re-binding it to different backing memory.
+			 */
+			buf_bsurf.offset += buf_extent_el.width *
+			                    buf_extent_el.height * buf_bsurf.bs;
+			img_bsurf.layer++;
+			if (image->type == VK_IMAGE_TYPE_3D)
+				slice_3d++;
+			else
+				slice_array++;
+		}
+	}
+	radv_meta_restore(&saved_state, cmd_buffer);
+}
+
+void radv_CmdCopyBufferToImage(
+	VkCommandBuffer                             commandBuffer,
+	VkBuffer                                    srcBuffer,
+	VkImage                                     destImage,
+	VkImageLayout                               destImageLayout,
+	uint32_t                                    regionCount,
+	const VkBufferImageCopy*                    pRegions)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_image, dest_image, destImage);
+	RADV_FROM_HANDLE(radv_buffer, src_buffer, srcBuffer);
+
+	meta_copy_buffer_to_image(cmd_buffer, src_buffer, dest_image,
+				  regionCount, pRegions);
+}
+
+static void
+meta_copy_image_to_buffer(struct radv_cmd_buffer *cmd_buffer,
+                          struct radv_buffer* buffer,
+                          struct radv_image* image,
+                          uint32_t regionCount,
+                          const VkBufferImageCopy* pRegions)
+{
+	struct radv_meta_saved_compute_state saved_state;
+
+	radv_meta_begin_bufimage(cmd_buffer, &saved_state);
+	for (unsigned r = 0; r < regionCount; r++) {
+
+		/**
+		 * From the Vulkan 1.0.6 spec: 18.3 Copying Data Between Images
+		 *    extent is the size in texels of the source image to copy in width,
+		 *    height and depth. 1D images use only x and width. 2D images use x, y,
+		 *    width and height. 3D images use x, y, z, width, height and depth.
+		 *
+		 *
+		 * Also, convert the offsets and extent from units of texels to units of
+		 * blocks - which is the highest resolution accessible in this command.
+		 */
+		const VkOffset3D img_offset_el =
+			meta_region_offset_el(image, &pRegions[r].imageOffset);
+		const VkExtent3D bufferExtent = {
+			.width  = pRegions[r].bufferRowLength ?
+			pRegions[r].bufferRowLength : pRegions[r].imageExtent.width,
+			.height = pRegions[r].bufferImageHeight ?
+			pRegions[r].bufferImageHeight : pRegions[r].imageExtent.height,
+		};
+		const VkExtent3D buf_extent_el =
+			meta_region_extent_el(image, &bufferExtent);
+
+		/* Start creating blit rect */
+		const VkExtent3D img_extent_el =
+			meta_region_extent_el(image, &pRegions[r].imageExtent);
+		struct radv_meta_blit2d_rect rect = {
+			.width = img_extent_el.width,
+			.height =  img_extent_el.height,
+		};
+
+		/* Create blit surfaces */
+		struct radv_meta_blit2d_surf img_info =
+			blit_surf_for_image_level_layer(image,
+							pRegions[r].imageSubresource.aspectMask,
+							pRegions[r].imageSubresource.mipLevel,
+							pRegions[r].imageSubresource.baseArrayLayer);
+		struct radv_meta_blit2d_buffer buf_info = {
+			.bs = img_info.bs,
+			.format = img_info.format,
+			.buffer = buffer,
+			.offset = pRegions[r].bufferOffset,
+			.pitch = buf_extent_el.width,
+		};
+
+		/* Loop through each 3D or array slice */
+		unsigned num_slices_3d = img_extent_el.depth;
+		unsigned num_slices_array = pRegions[r].imageSubresource.layerCount;
+		unsigned slice_3d = 0;
+		unsigned slice_array = 0;
+		while (slice_3d < num_slices_3d && slice_array < num_slices_array) {
+
+			rect.src_x = img_offset_el.x;
+			rect.src_y = img_offset_el.y;
+
+
+			/* Perform Blit */
+			radv_meta_image_to_buffer(cmd_buffer, &img_info, &buf_info, 1, &rect);
+
+			buf_info.offset += buf_extent_el.width *
+			                    buf_extent_el.height * buf_info.bs;
+			img_info.layer++;
+			if (image->type == VK_IMAGE_TYPE_3D)
+				slice_3d++;
+			else
+				slice_array++;
+		}
+	}
+	radv_meta_end_bufimage(cmd_buffer, &saved_state);
+}
+
+void radv_CmdCopyImageToBuffer(
+	VkCommandBuffer                             commandBuffer,
+	VkImage                                     srcImage,
+	VkImageLayout                               srcImageLayout,
+	VkBuffer                                    destBuffer,
+	uint32_t                                    regionCount,
+	const VkBufferImageCopy*                    pRegions)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_image, src_image, srcImage);
+	RADV_FROM_HANDLE(radv_buffer, dst_buffer, destBuffer);
+
+	meta_copy_image_to_buffer(cmd_buffer, dst_buffer, src_image,
+				  regionCount, pRegions);
+}
+
+void radv_CmdCopyImage(
+	VkCommandBuffer                             commandBuffer,
+	VkImage                                     srcImage,
+	VkImageLayout                               srcImageLayout,
+	VkImage                                     destImage,
+	VkImageLayout                               destImageLayout,
+	uint32_t                                    regionCount,
+	const VkImageCopy*                          pRegions)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_image, src_image, srcImage);
+	RADV_FROM_HANDLE(radv_image, dest_image, destImage);
+	struct radv_meta_saved_state saved_state;
+
+	/* From the Vulkan 1.0 spec:
+	 *
+	 *    vkCmdCopyImage can be used to copy image data between multisample
+	 *    images, but both images must have the same number of samples.
+	 */
+	assert(src_image->samples == dest_image->samples);
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	for (unsigned r = 0; r < regionCount; r++) {
+		assert(pRegions[r].srcSubresource.aspectMask ==
+		       pRegions[r].dstSubresource.aspectMask);
+
+		/* Create blit surfaces */
+		struct radv_meta_blit2d_surf b_src =
+			blit_surf_for_image_level_layer(src_image,
+							pRegions[r].srcSubresource.aspectMask,
+							pRegions[r].srcSubresource.mipLevel,
+							pRegions[r].srcSubresource.baseArrayLayer);
+		struct radv_meta_blit2d_surf b_dst =
+			blit_surf_for_image_level_layer(dest_image,
+							pRegions[r].dstSubresource.aspectMask,
+							pRegions[r].dstSubresource.mipLevel,
+							pRegions[r].dstSubresource.baseArrayLayer);
+
+		/* for DCC */
+		b_src.format = b_dst.format;
+
+		/**
+		 * From the Vulkan 1.0.6 spec: 18.4 Copying Data Between Buffers and Images
+		 *    imageExtent is the size in texels of the image to copy in width, height
+		 *    and depth. 1D images use only x and width. 2D images use x, y, width
+		 *    and height. 3D images use x, y, z, width, height and depth.
+		 *
+		 * Also, convert the offsets and extent from units of texels to units of
+		 * blocks - which is the highest resolution accessible in this command.
+		 */
+		const VkOffset3D dst_offset_el =
+			meta_region_offset_el(dest_image, &pRegions[r].dstOffset);
+		const VkOffset3D src_offset_el =
+			meta_region_offset_el(src_image, &pRegions[r].srcOffset);
+		const VkExtent3D img_extent_el =
+			meta_region_extent_el(src_image, &pRegions[r].extent);
+
+		/* Start creating blit rect */
+		struct radv_meta_blit2d_rect rect = {
+			.width = img_extent_el.width,
+			.height = img_extent_el.height,
+		};
+
+		/* Loop through each 3D or array slice */
+		unsigned num_slices_3d = img_extent_el.depth;
+		unsigned num_slices_array = pRegions[r].dstSubresource.layerCount;
+		unsigned slice_3d = 0;
+		unsigned slice_array = 0;
+		while (slice_3d < num_slices_3d && slice_array < num_slices_array) {
+
+			/* Finish creating blit rect */
+			rect.dst_x = dst_offset_el.x;
+			rect.dst_y = dst_offset_el.y;
+			rect.src_x = src_offset_el.x;
+			rect.src_y = src_offset_el.y;
+
+			/* Perform Blit */
+			radv_meta_blit2d(cmd_buffer, &b_src, NULL, &b_dst, 1, &rect);
+
+			b_src.layer++;
+			b_dst.layer++;
+			if (dest_image->type == VK_IMAGE_TYPE_3D)
+				slice_3d++;
+			else
+				slice_array++;
+		}
+	}
+
+	radv_meta_restore(&saved_state, cmd_buffer);
+}
diff --git a/src/amd/vulkan/radv_meta_decompress.c b/src/amd/vulkan/radv_meta_decompress.c
new file mode 100644
index 0000000000..498cc239bd
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_decompress.c
@@ -0,0 +1,463 @@
+/*
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <assert.h>
+#include <stdbool.h>
+
+#include "radv_meta.h"
+#include "radv_private.h"
+#include "nir/nir_builder.h"
+#include "sid.h"
+/**
+ * Vertex attributes used by all pipelines.
+ */
+struct vertex_attrs {
+	float position[2]; /**< 3DPRIM_RECTLIST */
+};
+
+/* passthrough vertex shader */
+static nir_shader *
+build_nir_vs(void)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+
+	nir_builder b;
+	nir_variable *a_position;
+	nir_variable *v_position;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, "meta_depth_decomp_vs");
+
+	a_position = nir_variable_create(b.shader, nir_var_shader_in, vec4,
+					 "a_position");
+	a_position->data.location = VERT_ATTRIB_GENERIC0;
+
+	v_position = nir_variable_create(b.shader, nir_var_shader_out, vec4,
+					 "gl_Position");
+	v_position->data.location = VARYING_SLOT_POS;
+
+	nir_copy_var(&b, v_position, a_position);
+
+	return b.shader;
+}
+
+/* simple passthrough shader */
+static nir_shader *
+build_nir_fs(void)
+{
+	nir_builder b;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+	b.shader->info.name = ralloc_asprintf(b.shader,
+					      "meta_depth_decomp_noop_fs");
+
+	return b.shader;
+}
+
+static VkResult
+create_pass(struct radv_device *device)
+{
+	VkResult result;
+	VkDevice device_h = radv_device_to_handle(device);
+	const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
+	VkAttachmentDescription attachment;
+
+	attachment.format = VK_FORMAT_UNDEFINED;
+	attachment.samples = 1;
+	attachment.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
+	attachment.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
+	attachment.initialLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
+	attachment.finalLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
+
+	result = radv_CreateRenderPass(device_h,
+				       &(VkRenderPassCreateInfo) {
+					       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+						       .attachmentCount = 1,
+						       .pAttachments = &attachment,
+						       .subpassCount = 1,
+							.pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 0,
+						       .pColorAttachments = NULL,
+						       .pResolveAttachments = NULL,
+						       .pDepthStencilAttachment = &(VkAttachmentReference) {
+							       .attachment = 0,
+							       .layout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
+						       },
+						       .preserveAttachmentCount = 0,
+						       .pPreserveAttachments = NULL,
+					       },
+								.dependencyCount = 0,
+								   },
+				       alloc,
+				       &device->meta_state.depth_decomp.pass);
+
+	return result;
+}
+
+static VkResult
+create_pipeline(struct radv_device *device,
+                VkShaderModule vs_module_h)
+{
+	VkResult result;
+	VkDevice device_h = radv_device_to_handle(device);
+
+	struct radv_shader_module fs_module = {
+		.nir = build_nir_fs(),
+	};
+
+	if (!fs_module.nir) {
+		/* XXX: Need more accurate error */
+		result = VK_ERROR_OUT_OF_HOST_MEMORY;
+		goto cleanup;
+	}
+
+	const VkGraphicsPipelineCreateInfo pipeline_create_info = {
+		.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+		.stageCount = 2,
+		.pStages = (VkPipelineShaderStageCreateInfo[]) {
+		       {
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+				.stage = VK_SHADER_STAGE_VERTEX_BIT,
+				.module = vs_module_h,
+				.pName = "main",
+			},
+			{
+				.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+				.stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+				.module = radv_shader_module_to_handle(&fs_module),
+				.pName = "main",
+			},
+		},
+		.pVertexInputState = &(VkPipelineVertexInputStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+			.vertexBindingDescriptionCount = 1,
+			.pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
+				{
+					.binding = 0,
+					.stride = sizeof(struct vertex_attrs),
+					.inputRate = VK_VERTEX_INPUT_RATE_VERTEX
+				},
+			},
+			.vertexAttributeDescriptionCount = 1,
+			.pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
+				{
+					/* Position */
+					.location = 0,
+					.binding = 0,
+					.format = VK_FORMAT_R32G32_SFLOAT,
+					.offset = offsetof(struct vertex_attrs, position),
+				},
+			},
+		},
+		.pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+			.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+			.primitiveRestartEnable = false,
+		},
+		.pViewportState = &(VkPipelineViewportStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+			.viewportCount = 0,
+			.scissorCount = 0,
+		},
+		.pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+			.depthClampEnable = false,
+			.rasterizerDiscardEnable = false,
+			.polygonMode = VK_POLYGON_MODE_FILL,
+			.cullMode = VK_CULL_MODE_NONE,
+			.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE,
+		},
+		.pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+			.rasterizationSamples = 1,
+			.sampleShadingEnable = false,
+			.pSampleMask = NULL,
+			.alphaToCoverageEnable = false,
+			.alphaToOneEnable = false,
+		},
+		.pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+			.logicOpEnable = false,
+			.attachmentCount = 0,
+			.pAttachments = NULL,
+		},
+		.pDepthStencilState = &(VkPipelineDepthStencilStateCreateInfo) {
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
+			.depthTestEnable = false,
+			.depthWriteEnable = false,
+			.depthBoundsTestEnable = false,
+			.stencilTestEnable = false,
+		},
+		.pDynamicState = NULL,
+		.renderPass = device->meta_state.depth_decomp.pass,
+		.subpass = 0,
+	};
+
+	result = radv_graphics_pipeline_create(device_h,
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &pipeline_create_info,
+					       &(struct radv_graphics_pipeline_create_info) {
+							.use_rectlist = true,
+							.db_flush_depth_inplace = true,
+							.db_flush_stencil_inplace = true,
+					       },
+					       &device->meta_state.alloc,
+					       &device->meta_state.depth_decomp.decompress_pipeline);
+	if (result != VK_SUCCESS)
+		goto cleanup;
+
+	result = radv_graphics_pipeline_create(device_h,
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &pipeline_create_info,
+					       &(struct radv_graphics_pipeline_create_info) {
+							.use_rectlist = true,
+							.db_flush_depth_inplace = true,
+							.db_flush_stencil_inplace = true,
+							.db_resummarize = true,
+					       },
+					       &device->meta_state.alloc,
+					       &device->meta_state.depth_decomp.resummarize_pipeline);
+	if (result != VK_SUCCESS)
+		goto cleanup;
+
+	goto cleanup;
+
+cleanup:
+	ralloc_free(fs_module.nir);
+	return result;
+}
+
+void
+radv_device_finish_meta_depth_decomp_state(struct radv_device *device)
+{
+	struct radv_meta_state *state = &device->meta_state;
+	VkDevice device_h = radv_device_to_handle(device);
+	VkRenderPass pass_h = device->meta_state.depth_decomp.pass;
+	const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
+
+	if (pass_h)
+		RADV_CALL(DestroyRenderPass)(device_h, pass_h,
+					     &device->meta_state.alloc);
+
+	VkPipeline pipeline_h = state->depth_decomp.decompress_pipeline;
+	if (pipeline_h) {
+		RADV_CALL(DestroyPipeline)(device_h, pipeline_h, alloc);
+	}
+	pipeline_h = state->depth_decomp.resummarize_pipeline;
+	if (pipeline_h) {
+		RADV_CALL(DestroyPipeline)(device_h, pipeline_h, alloc);
+	}
+}
+
+VkResult
+radv_device_init_meta_depth_decomp_state(struct radv_device *device)
+{
+	VkResult res = VK_SUCCESS;
+
+	zero(device->meta_state.depth_decomp);
+
+	struct radv_shader_module vs_module = { .nir = build_nir_vs() };
+	if (!vs_module.nir) {
+		/* XXX: Need more accurate error */
+		res = VK_ERROR_OUT_OF_HOST_MEMORY;
+		goto fail;
+	}
+
+	res = create_pass(device);
+	if (res != VK_SUCCESS)
+		goto fail;
+
+	VkShaderModule vs_module_h = radv_shader_module_to_handle(&vs_module);
+	res = create_pipeline(device, vs_module_h);
+	if (res != VK_SUCCESS)
+		goto fail;
+
+	goto cleanup;
+
+fail:
+	radv_device_finish_meta_depth_decomp_state(device);
+
+cleanup:
+	ralloc_free(vs_module.nir);
+
+	return res;
+}
+
+static void
+emit_depth_decomp(struct radv_cmd_buffer *cmd_buffer,
+		  const VkOffset2D *dest_offset,
+		  const VkExtent2D *depth_decomp_extent,
+		  VkPipeline pipeline_h)
+{
+	struct radv_device *device = cmd_buffer->device;
+	VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
+	uint32_t offset;
+	const struct vertex_attrs vertex_data[3] = {
+		{
+			.position = {
+				dest_offset->x,
+				dest_offset->y,
+			},
+		},
+		{
+			.position = {
+				dest_offset->x,
+				dest_offset->y + depth_decomp_extent->height,
+			},
+		},
+		{
+			.position = {
+				dest_offset->x + depth_decomp_extent->width,
+				dest_offset->y,
+			},
+		},
+	};
+
+	radv_cmd_buffer_upload_data(cmd_buffer, sizeof(vertex_data), 16, vertex_data, &offset);
+	struct radv_buffer vertex_buffer = {
+		.device = device,
+		.size = sizeof(vertex_data),
+		.bo = cmd_buffer->upload.upload_bo,
+		.offset = offset,
+	};
+
+	VkBuffer vertex_buffer_h = radv_buffer_to_handle(&vertex_buffer);
+
+	radv_CmdBindVertexBuffers(cmd_buffer_h,
+				  /*firstBinding*/ 0,
+				  /*bindingCount*/ 1,
+				  (VkBuffer[]) { vertex_buffer_h },
+				  (VkDeviceSize[]) { 0 });
+
+	RADV_FROM_HANDLE(radv_pipeline, pipeline, pipeline_h);
+
+	if (cmd_buffer->state.pipeline != pipeline) {
+		radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
+				     pipeline_h);
+	}
+
+	RADV_CALL(CmdDraw)(cmd_buffer_h, 3, 1, 0, 0);
+}
+
+
+static void radv_process_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
+					     struct radv_image *image,
+					     VkImageSubresourceRange *subresourceRange,
+					     VkPipeline pipeline_h)
+{
+	struct radv_meta_saved_state saved_state;
+	struct radv_meta_saved_pass_state saved_pass_state;
+	VkDevice device_h = radv_device_to_handle(cmd_buffer->device);
+	VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
+	uint32_t width = radv_minify(image->extent.width,
+				     subresourceRange->baseMipLevel);
+	uint32_t height = radv_minify(image->extent.height,
+				     subresourceRange->baseMipLevel);
+
+	if (!image->htile.size)
+		return;
+	radv_meta_save_pass(&saved_pass_state, cmd_buffer);
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	for (uint32_t layer = 0; layer < subresourceRange->layerCount; layer++) {
+		struct radv_image_view iview;
+
+		radv_image_view_init(&iview, cmd_buffer->device,
+				     &(VkImageViewCreateInfo) {
+					     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+					     .image = radv_image_to_handle(image),
+					     .format = image->vk_format,
+					     .subresourceRange = {
+						     .aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT,
+						     .baseMipLevel = subresourceRange->baseMipLevel,
+						     .levelCount = 1,
+						     .baseArrayLayer = subresourceRange->baseArrayLayer + layer,
+						     .layerCount = 1,
+					     },
+				     },
+				     cmd_buffer, VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT);
+
+
+		VkFramebuffer fb_h;
+		radv_CreateFramebuffer(device_h,
+				       &(VkFramebufferCreateInfo) {
+					       .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
+					       .attachmentCount = 1,
+						       .pAttachments = (VkImageView[]) {
+						       radv_image_view_to_handle(&iview)
+					       },
+					       .width = width,
+						.height = height,
+					       .layers = 1
+				       },
+				       &cmd_buffer->pool->alloc,
+				       &fb_h);
+
+		RADV_CALL(CmdBeginRenderPass)(cmd_buffer_h,
+					      &(VkRenderPassBeginInfo) {
+						      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+							      .renderPass = cmd_buffer->device->meta_state.depth_decomp.pass,
+							      .framebuffer = fb_h,
+							      .renderArea = {
+							      .offset = {
+								      0,
+								      0,
+							      },
+							      .extent = {
+								      width,
+								      height,
+							      }
+						       },
+						       .clearValueCount = 0,
+						       .pClearValues = NULL,
+					   },
+					   VK_SUBPASS_CONTENTS_INLINE);
+
+		emit_depth_decomp(cmd_buffer, &(VkOffset2D){0, 0 }, &(VkExtent2D){width, height}, pipeline_h);
+		RADV_CALL(CmdEndRenderPass)(cmd_buffer_h);
+
+		radv_DestroyFramebuffer(device_h, fb_h,
+					&cmd_buffer->pool->alloc);
+	}
+	radv_meta_restore(&saved_state, cmd_buffer);
+	radv_meta_restore_pass(&saved_pass_state, cmd_buffer);
+}
+
+void radv_decompress_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
+					 struct radv_image *image,
+					 VkImageSubresourceRange *subresourceRange)
+{
+	radv_process_depth_image_inplace(cmd_buffer, image, subresourceRange,
+					 cmd_buffer->device->meta_state.depth_decomp.decompress_pipeline);
+}
+
+void radv_resummarize_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
+					 struct radv_image *image,
+					 VkImageSubresourceRange *subresourceRange)
+{
+	radv_process_depth_image_inplace(cmd_buffer, image, subresourceRange,
+					 cmd_buffer->device->meta_state.depth_decomp.resummarize_pipeline);
+}
diff --git a/src/amd/vulkan/radv_meta_fast_clear.c b/src/amd/vulkan/radv_meta_fast_clear.c
new file mode 100644
index 0000000000..0fb10c7b3b
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_fast_clear.c
@@ -0,0 +1,536 @@
+/*
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <assert.h>
+#include <stdbool.h>
+
+#include "radv_meta.h"
+#include "radv_private.h"
+#include "nir/nir_builder.h"
+#include "sid.h"
+/**
+ * Vertex attributes used by all pipelines.
+ */
+struct vertex_attrs {
+	float position[2]; /**< 3DPRIM_RECTLIST */
+	float tex_position[2];
+};
+
+/* passthrough vertex shader */
+static nir_shader *
+build_nir_vs(void)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+
+	nir_builder b;
+	nir_variable *a_position;
+	nir_variable *v_position;
+	nir_variable *a_tex_position;
+	nir_variable *v_tex_position;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, "meta_fast_clear_vs");
+
+	a_position = nir_variable_create(b.shader, nir_var_shader_in, vec4,
+					 "a_position");
+	a_position->data.location = VERT_ATTRIB_GENERIC0;
+
+	v_position = nir_variable_create(b.shader, nir_var_shader_out, vec4,
+					 "gl_Position");
+	v_position->data.location = VARYING_SLOT_POS;
+
+	a_tex_position = nir_variable_create(b.shader, nir_var_shader_in, vec4,
+					     "a_tex_position");
+	a_tex_position->data.location = VERT_ATTRIB_GENERIC1;
+
+	v_tex_position = nir_variable_create(b.shader, nir_var_shader_out, vec4,
+					     "v_tex_position");
+	v_tex_position->data.location = VARYING_SLOT_VAR0;
+
+	nir_copy_var(&b, v_position, a_position);
+	nir_copy_var(&b, v_tex_position, a_tex_position);
+
+	return b.shader;
+}
+
+/* simple passthrough shader */
+static nir_shader *
+build_nir_fs(void)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	nir_builder b;
+	nir_variable *v_tex_position; /* vec4, varying texture coordinate */
+	nir_variable *f_color; /* vec4, fragment output color */
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+	b.shader->info.name = ralloc_asprintf(b.shader,
+					      "meta_fast_clear_fs");
+
+	v_tex_position = nir_variable_create(b.shader, nir_var_shader_in, vec4,
+					     "v_tex_position");
+	v_tex_position->data.location = VARYING_SLOT_VAR0;
+
+	f_color = nir_variable_create(b.shader, nir_var_shader_out, vec4,
+				      "f_color");
+	f_color->data.location = FRAG_RESULT_DATA0;
+
+	nir_copy_var(&b, f_color, v_tex_position);
+
+	return b.shader;
+}
+
+static VkResult
+create_pass(struct radv_device *device)
+{
+	VkResult result;
+	VkDevice device_h = radv_device_to_handle(device);
+	const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
+	VkAttachmentDescription attachment;
+
+	attachment.format = VK_FORMAT_UNDEFINED;
+	attachment.samples = 1;
+	attachment.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
+	attachment.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
+	attachment.initialLayout = VK_IMAGE_LAYOUT_GENERAL;
+	attachment.finalLayout = VK_IMAGE_LAYOUT_GENERAL;
+
+	result = radv_CreateRenderPass(device_h,
+				       &(VkRenderPassCreateInfo) {
+					       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+						       .attachmentCount = 1,
+						       .pAttachments = &attachment,
+						       .subpassCount = 1,
+						       .pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 1,
+						       .pColorAttachments = (VkAttachmentReference[]) {
+							       {
+								       .attachment = 0,
+								       .layout = VK_IMAGE_LAYOUT_GENERAL,
+							       },
+						       },
+						       .pResolveAttachments = NULL,
+						       .pDepthStencilAttachment = &(VkAttachmentReference) {
+							       .attachment = VK_ATTACHMENT_UNUSED,
+						       },
+						       .preserveAttachmentCount = 0,
+						       .pPreserveAttachments = NULL,
+					       },
+								.dependencyCount = 0,
+				       },
+				       alloc,
+				       &device->meta_state.fast_clear_flush.pass);
+
+	return result;
+}
+
+static VkResult
+create_pipeline(struct radv_device *device,
+                VkShaderModule vs_module_h)
+{
+	VkResult result;
+	VkDevice device_h = radv_device_to_handle(device);
+
+	struct radv_shader_module fs_module = {
+		.nir = build_nir_fs(),
+	};
+
+	if (!fs_module.nir) {
+		/* XXX: Need more accurate error */
+		result = VK_ERROR_OUT_OF_HOST_MEMORY;
+		goto cleanup;
+	}
+
+	const VkPipelineShaderStageCreateInfo stages[2] = {
+		{
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_VERTEX_BIT,
+			.module = vs_module_h,
+			.pName = "main",
+		},
+		{
+			.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+			.stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+			.module = radv_shader_module_to_handle(&fs_module),
+			.pName = "main",
+		},
+	};
+
+	const VkPipelineVertexInputStateCreateInfo vi_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+		.vertexBindingDescriptionCount = 1,
+		.pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
+			{
+				.binding = 0,
+				.stride = sizeof(struct vertex_attrs),
+				.inputRate = VK_VERTEX_INPUT_RATE_VERTEX
+			},
+		},
+		.vertexAttributeDescriptionCount = 2,
+		.pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
+			{
+				/* Position */
+				.location = 0,
+				.binding = 0,
+				.format = VK_FORMAT_R32G32_SFLOAT,
+				.offset = offsetof(struct vertex_attrs, position),
+			},
+			{
+				/* Texture Coordinate */
+				.location = 1,
+				.binding = 0,
+				.format = VK_FORMAT_R32G32_SFLOAT,
+				.offset = offsetof(struct vertex_attrs, tex_position),
+			},
+		}
+	};
+
+	const VkPipelineInputAssemblyStateCreateInfo ia_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+		.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+		.primitiveRestartEnable = false,
+	};
+
+	const VkPipelineColorBlendStateCreateInfo blend_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+		.logicOpEnable = false,
+		.attachmentCount = 1,
+		.pAttachments = (VkPipelineColorBlendAttachmentState []) {
+			{
+				.colorWriteMask = VK_COLOR_COMPONENT_R_BIT |
+				VK_COLOR_COMPONENT_G_BIT |
+				VK_COLOR_COMPONENT_B_BIT |
+				VK_COLOR_COMPONENT_A_BIT,
+			},
+		}
+	};
+	const VkPipelineRasterizationStateCreateInfo rs_state = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+		.depthClampEnable = false,
+		.rasterizerDiscardEnable = false,
+		.polygonMode = VK_POLYGON_MODE_FILL,
+		.cullMode = VK_CULL_MODE_NONE,
+		.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE,
+	};
+
+	result = radv_graphics_pipeline_create(device_h,
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &(VkGraphicsPipelineCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+						       .stageCount = 2,
+						       .pStages = stages,
+
+						       .pVertexInputState = &vi_state,
+						       .pInputAssemblyState = &ia_state,
+
+					       .pViewportState = &(VkPipelineViewportStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+						       .viewportCount = 0,
+						       .scissorCount = 0,
+					       },
+						       .pRasterizationState = &rs_state,
+					       .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+						       .rasterizationSamples = 1,
+						       .sampleShadingEnable = false,
+						       .pSampleMask = NULL,
+						       .alphaToCoverageEnable = false,
+						       .alphaToOneEnable = false,
+					       },
+						.pColorBlendState = &blend_state,
+						.pDynamicState = NULL,
+						.renderPass = device->meta_state.fast_clear_flush.pass,
+						.subpass = 0,
+					       },
+					       &(struct radv_graphics_pipeline_create_info) {
+						       .use_rectlist = true,
+						       .custom_blend_mode = V_028808_CB_ELIMINATE_FAST_CLEAR,
+					       },
+					       &device->meta_state.alloc,
+					       &device->meta_state.fast_clear_flush.cmask_eliminate_pipeline);
+	if (result != VK_SUCCESS)
+		goto cleanup;
+
+	result = radv_graphics_pipeline_create(device_h,
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &(VkGraphicsPipelineCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+						       .stageCount = 2,
+						       .pStages = stages,
+
+						       .pVertexInputState = &vi_state,
+						       .pInputAssemblyState = &ia_state,
+
+					       .pViewportState = &(VkPipelineViewportStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+						       .viewportCount = 0,
+						       .scissorCount = 0,
+					       },
+						       .pRasterizationState = &rs_state,
+					       .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+						       .rasterizationSamples = 1,
+						       .sampleShadingEnable = false,
+						       .pSampleMask = NULL,
+						       .alphaToCoverageEnable = false,
+						       .alphaToOneEnable = false,
+					       },
+						.pColorBlendState = &blend_state,
+						.pDynamicState = NULL,
+						.renderPass = device->meta_state.fast_clear_flush.pass,
+						.subpass = 0,
+					       },
+					       &(struct radv_graphics_pipeline_create_info) {
+						       .use_rectlist = true,
+						       .custom_blend_mode = V_028808_CB_FMASK_DECOMPRESS,
+					       },
+					       &device->meta_state.alloc,
+					       &device->meta_state.fast_clear_flush.fmask_decompress_pipeline);
+	if (result != VK_SUCCESS)
+		goto cleanup_cmask;
+
+	goto cleanup;
+cleanup_cmask:
+	RADV_CALL(DestroyPipeline)(device_h, device->meta_state.fast_clear_flush.cmask_eliminate_pipeline, &device->meta_state.alloc);
+cleanup:
+	ralloc_free(fs_module.nir);
+	return result;
+}
+
+void
+radv_device_finish_meta_fast_clear_flush_state(struct radv_device *device)
+{
+	struct radv_meta_state *state = &device->meta_state;
+	VkDevice device_h = radv_device_to_handle(device);
+	VkRenderPass pass_h = device->meta_state.fast_clear_flush.pass;
+	const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
+
+	if (pass_h)
+		RADV_CALL(DestroyRenderPass)(device_h, pass_h,
+					     &device->meta_state.alloc);
+
+	VkPipeline pipeline_h = state->fast_clear_flush.cmask_eliminate_pipeline;
+	if (pipeline_h) {
+		RADV_CALL(DestroyPipeline)(device_h, pipeline_h, alloc);
+	}
+
+	pipeline_h = state->fast_clear_flush.fmask_decompress_pipeline;
+	if (pipeline_h) {
+		RADV_CALL(DestroyPipeline)(device_h, pipeline_h, alloc);
+	}
+}
+
+VkResult
+radv_device_init_meta_fast_clear_flush_state(struct radv_device *device)
+{
+	VkResult res = VK_SUCCESS;
+
+	zero(device->meta_state.fast_clear_flush);
+
+	struct radv_shader_module vs_module = { .nir = build_nir_vs() };
+	if (!vs_module.nir) {
+		/* XXX: Need more accurate error */
+		res = VK_ERROR_OUT_OF_HOST_MEMORY;
+		goto fail;
+	}
+
+	res = create_pass(device);
+	if (res != VK_SUCCESS)
+		goto fail;
+
+	VkShaderModule vs_module_h = radv_shader_module_to_handle(&vs_module);
+	res = create_pipeline(device, vs_module_h);
+	if (res != VK_SUCCESS)
+		goto fail;
+
+	goto cleanup;
+
+fail:
+	radv_device_finish_meta_fast_clear_flush_state(device);
+
+cleanup:
+	ralloc_free(vs_module.nir);
+
+	return res;
+}
+
+static void
+emit_fast_clear_flush(struct radv_cmd_buffer *cmd_buffer,
+		      const VkExtent2D *resolve_extent,
+		      bool fmask_decompress)
+{
+	struct radv_device *device = cmd_buffer->device;
+	VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
+	uint32_t offset;
+	const struct vertex_attrs vertex_data[3] = {
+		{
+			.position = {
+				0,
+				0,
+			},
+			.tex_position = {
+				0,
+				0,
+			},
+		},
+		{
+			.position = {
+				0,
+				resolve_extent->height,
+			},
+			.tex_position = {
+				0,
+				resolve_extent->height,
+			},
+		},
+		{
+			.position = {
+				resolve_extent->width,
+				0,
+			},
+			.tex_position = {
+				resolve_extent->width,
+				0,
+			},
+		},
+	};
+
+	cmd_buffer->state.flush_bits |= (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+					 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META);
+	radv_cmd_buffer_upload_data(cmd_buffer, sizeof(vertex_data), 16, vertex_data, &offset);
+	struct radv_buffer vertex_buffer = {
+		.device = device,
+		.size = sizeof(vertex_data),
+		.bo = cmd_buffer->upload.upload_bo,
+		.offset = offset,
+	};
+
+	VkBuffer vertex_buffer_h = radv_buffer_to_handle(&vertex_buffer);
+
+	radv_CmdBindVertexBuffers(cmd_buffer_h,
+				  /*firstBinding*/ 0,
+				  /*bindingCount*/ 1,
+				  (VkBuffer[]) { vertex_buffer_h },
+				  (VkDeviceSize[]) { 0 });
+
+	VkPipeline pipeline_h;
+	if (fmask_decompress)
+		pipeline_h = device->meta_state.fast_clear_flush.fmask_decompress_pipeline;
+	else
+		pipeline_h = device->meta_state.fast_clear_flush.cmask_eliminate_pipeline;
+	RADV_FROM_HANDLE(radv_pipeline, pipeline, pipeline_h);
+
+	if (cmd_buffer->state.pipeline != pipeline) {
+		radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
+				     pipeline_h);
+	}
+
+	RADV_CALL(CmdDraw)(cmd_buffer_h, 3, 1, 0, 0);
+	cmd_buffer->state.flush_bits |= (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+					 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META);
+	si_emit_cache_flush(cmd_buffer);
+}
+
+/**
+ */
+void
+radv_fast_clear_flush_image_inplace(struct radv_cmd_buffer *cmd_buffer,
+				    struct radv_image *image)
+{
+	struct radv_meta_saved_state saved_state;
+	struct radv_meta_saved_pass_state saved_pass_state;
+	VkDevice device_h = radv_device_to_handle(cmd_buffer->device);
+	VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
+
+	if (!image->cmask.size)
+		return;
+
+	if (!cmd_buffer->device->allow_fast_clears)
+		return;
+
+	radv_meta_save_pass(&saved_pass_state, cmd_buffer);
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	struct radv_image_view iview;
+	radv_image_view_init(&iview, cmd_buffer->device,
+			     &(VkImageViewCreateInfo) {
+				     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+					     .image = radv_image_to_handle(image),
+					     .format = image->vk_format,
+					     .subresourceRange = {
+						     .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+						     .baseMipLevel = 0,
+						     .levelCount = 1,
+						     .baseArrayLayer = 0,
+						     .layerCount = 1,
+					     },
+				     },
+				     cmd_buffer, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
+
+	VkFramebuffer fb_h;
+	radv_CreateFramebuffer(device_h,
+			       &(VkFramebufferCreateInfo) {
+				       .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
+				       .attachmentCount = 1,
+				       .pAttachments = (VkImageView[]) {
+					       radv_image_view_to_handle(&iview)
+				       },
+				       .width = image->extent.width,
+				       .height = image->extent.height,
+				       .layers = 1
+			      },
+			      &cmd_buffer->pool->alloc,
+			      &fb_h);
+
+	RADV_CALL(CmdBeginRenderPass)(cmd_buffer_h,
+				      &(VkRenderPassBeginInfo) {
+					      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+						      .renderPass = cmd_buffer->device->meta_state.fast_clear_flush.pass,
+						      .framebuffer = fb_h,
+						      .renderArea = {
+						      .offset = {
+							      0,
+							      0,
+						      },
+						      .extent = {
+							      image->extent.width,
+							      image->extent.height,
+						      }
+					      },
+					      .clearValueCount = 0,
+					      .pClearValues = NULL,
+				     },
+				     VK_SUBPASS_CONTENTS_INLINE);
+
+	emit_fast_clear_flush(cmd_buffer,
+			      &(VkExtent2D) { image->extent.width, image->extent.height },
+			      image->fmask.size > 0);
+	RADV_CALL(CmdEndRenderPass)(cmd_buffer_h);
+
+	radv_DestroyFramebuffer(device_h, fb_h,
+				&cmd_buffer->pool->alloc);
+
+	radv_meta_restore(&saved_state, cmd_buffer);
+	radv_meta_restore_pass(&saved_pass_state, cmd_buffer);
+}
diff --git a/src/amd/vulkan/radv_meta_resolve.c b/src/amd/vulkan/radv_meta_resolve.c
new file mode 100644
index 0000000000..514aa8c7ef
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_resolve.c
@@ -0,0 +1,670 @@
+/*
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <assert.h>
+#include <stdbool.h>
+
+#include "radv_meta.h"
+#include "radv_private.h"
+#include "nir/nir_builder.h"
+#include "sid.h"
+/**
+ * Vertex attributes used by all pipelines.
+ */
+struct vertex_attrs {
+	float position[2]; /**< 3DPRIM_RECTLIST */
+	float tex_position[2];
+};
+
+/* passthrough vertex shader */
+static nir_shader *
+build_nir_vs(void)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+
+	nir_builder b;
+	nir_variable *a_position;
+	nir_variable *v_position;
+	nir_variable *a_tex_position;
+	nir_variable *v_tex_position;
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, "meta_resolve_vs");
+
+	a_position = nir_variable_create(b.shader, nir_var_shader_in, vec4,
+					 "a_position");
+	a_position->data.location = VERT_ATTRIB_GENERIC0;
+
+	v_position = nir_variable_create(b.shader, nir_var_shader_out, vec4,
+					 "gl_Position");
+	v_position->data.location = VARYING_SLOT_POS;
+
+	a_tex_position = nir_variable_create(b.shader, nir_var_shader_in, vec4,
+					     "a_tex_position");
+	a_tex_position->data.location = VERT_ATTRIB_GENERIC1;
+
+	v_tex_position = nir_variable_create(b.shader, nir_var_shader_out, vec4,
+					     "v_tex_position");
+	v_tex_position->data.location = VARYING_SLOT_VAR0;
+
+	nir_copy_var(&b, v_position, a_position);
+	nir_copy_var(&b, v_tex_position, a_tex_position);
+
+	return b.shader;
+}
+
+/* simple passthrough shader */
+static nir_shader *
+build_nir_fs(void)
+{
+	const struct glsl_type *vec4 = glsl_vec4_type();
+	nir_builder b;
+	nir_variable *v_tex_position; /* vec4, varying texture coordinate */
+	nir_variable *f_color; /* vec4, fragment output color */
+
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+	b.shader->info.name = ralloc_asprintf(b.shader,
+					      "meta_resolve_fs");
+
+	v_tex_position = nir_variable_create(b.shader, nir_var_shader_in, vec4,
+					     "v_tex_position");
+	v_tex_position->data.location = VARYING_SLOT_VAR0;
+
+	f_color = nir_variable_create(b.shader, nir_var_shader_out, vec4,
+				      "f_color");
+	f_color->data.location = FRAG_RESULT_DATA0;
+
+	nir_copy_var(&b, f_color, v_tex_position);
+
+	return b.shader;
+}
+
+static VkResult
+create_pass(struct radv_device *device)
+{
+	VkResult result;
+	VkDevice device_h = radv_device_to_handle(device);
+	const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
+	VkAttachmentDescription attachments[2];
+	int i;
+
+	for (i = 0; i < 2; i++) {
+		attachments[i].format = VK_FORMAT_UNDEFINED;
+		attachments[i].samples = 1;
+		attachments[i].loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
+		attachments[i].storeOp = VK_ATTACHMENT_STORE_OP_STORE;
+		attachments[i].initialLayout = VK_IMAGE_LAYOUT_GENERAL;
+		attachments[i].finalLayout = VK_IMAGE_LAYOUT_GENERAL;
+	}
+
+	result = radv_CreateRenderPass(device_h,
+				       &(VkRenderPassCreateInfo) {
+					       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+						       .attachmentCount = 2,
+						       .pAttachments = attachments,
+						       .subpassCount = 1,
+								.pSubpasses = &(VkSubpassDescription) {
+						       .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+						       .inputAttachmentCount = 0,
+						       .colorAttachmentCount = 2,
+						       .pColorAttachments = (VkAttachmentReference[]) {
+							       {
+								       .attachment = 0,
+								       .layout = VK_IMAGE_LAYOUT_GENERAL,
+							       },
+							       {
+								       .attachment = 1,
+								       .layout = VK_IMAGE_LAYOUT_GENERAL,
+							       },
+						       },
+						       .pResolveAttachments = NULL,
+						       .pDepthStencilAttachment = &(VkAttachmentReference) {
+							       .attachment = VK_ATTACHMENT_UNUSED,
+						       },
+						       .preserveAttachmentCount = 0,
+						       .pPreserveAttachments = NULL,
+					       },
+								.dependencyCount = 0,
+									 },
+				       alloc,
+				       &device->meta_state.resolve.pass);
+
+	return result;
+}
+
+static VkResult
+create_pipeline(struct radv_device *device,
+                VkShaderModule vs_module_h)
+{
+	VkResult result;
+	VkDevice device_h = radv_device_to_handle(device);
+
+	struct radv_shader_module fs_module = {
+		.nir = build_nir_fs(),
+	};
+
+	if (!fs_module.nir) {
+		/* XXX: Need more accurate error */
+		result = VK_ERROR_OUT_OF_HOST_MEMORY;
+		goto cleanup;
+	}
+
+	result = radv_graphics_pipeline_create(device_h,
+					       radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					       &(VkGraphicsPipelineCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+						       .stageCount = 2,
+						       .pStages = (VkPipelineShaderStageCreateInfo[]) {
+						       {
+							       .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+							       .stage = VK_SHADER_STAGE_VERTEX_BIT,
+							       .module = vs_module_h,
+							       .pName = "main",
+						       },
+						       {
+							       .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+							       .stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+							       .module = radv_shader_module_to_handle(&fs_module),
+							       .pName = "main",
+						       },
+					       },
+					       .pVertexInputState = &(VkPipelineVertexInputStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+						       .vertexBindingDescriptionCount = 1,
+						       .pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
+							       {
+								       .binding = 0,
+								       .stride = sizeof(struct vertex_attrs),
+								       .inputRate = VK_VERTEX_INPUT_RATE_VERTEX
+							       },
+						       },
+						       .vertexAttributeDescriptionCount = 2,
+						       .pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
+							       {
+								       /* Position */
+								       .location = 0,
+								       .binding = 0,
+								       .format = VK_FORMAT_R32G32_SFLOAT,
+								       .offset = offsetof(struct vertex_attrs, position),
+							       },
+							       {
+								       /* Texture Coordinate */
+								       .location = 1,
+								       .binding = 0,
+								       .format = VK_FORMAT_R32G32_SFLOAT,
+								       .offset = offsetof(struct vertex_attrs, tex_position),
+							       },
+						       },
+					       },
+					       .pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+						       .topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
+						       .primitiveRestartEnable = false,
+					       },
+					       .pViewportState = &(VkPipelineViewportStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+						       .viewportCount = 0,
+						       .scissorCount = 0,
+					       },
+					       .pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+						       .depthClampEnable = false,
+						       .rasterizerDiscardEnable = false,
+						       .polygonMode = VK_POLYGON_MODE_FILL,
+						       .cullMode = VK_CULL_MODE_NONE,
+						       .frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE,
+					       },
+					       .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+						       .rasterizationSamples = 1,
+						       .sampleShadingEnable = false,
+						       .pSampleMask = NULL,
+						       .alphaToCoverageEnable = false,
+						       .alphaToOneEnable = false,
+					       },
+					       .pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+						       .logicOpEnable = false,
+						       .attachmentCount = 2,
+						       .pAttachments = (VkPipelineColorBlendAttachmentState []) {
+							       {
+							       .colorWriteMask = VK_COLOR_COMPONENT_R_BIT |
+									       VK_COLOR_COMPONENT_G_BIT |
+									       VK_COLOR_COMPONENT_B_BIT |
+									       VK_COLOR_COMPONENT_A_BIT,
+							       },
+							       {
+							       .colorWriteMask = 0,
+
+							       }
+						       },
+						},
+						  .pDynamicState = NULL,
+																       .renderPass = device->meta_state.resolve.pass,
+																       .subpass = 0,
+																       },
+					       &(struct radv_graphics_pipeline_create_info) {
+						       .use_rectlist = true,
+						       .custom_blend_mode = V_028808_CB_RESOLVE,
+							       },
+					       &device->meta_state.alloc,
+					       &device->meta_state.resolve.pipeline);
+	if (result != VK_SUCCESS)
+		goto cleanup;
+
+	goto cleanup;
+
+cleanup:
+	ralloc_free(fs_module.nir);
+	return result;
+}
+
+void
+radv_device_finish_meta_resolve_state(struct radv_device *device)
+{
+	struct radv_meta_state *state = &device->meta_state;
+	VkDevice device_h = radv_device_to_handle(device);
+	VkRenderPass pass_h = device->meta_state.resolve.pass;
+	const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
+
+	if (pass_h)
+		RADV_CALL(DestroyRenderPass)(device_h, pass_h,
+					     &device->meta_state.alloc);
+
+	VkPipeline pipeline_h = state->resolve.pipeline;
+	if (pipeline_h) {
+		RADV_CALL(DestroyPipeline)(device_h, pipeline_h, alloc);
+	}
+}
+
+VkResult
+radv_device_init_meta_resolve_state(struct radv_device *device)
+{
+	VkResult res = VK_SUCCESS;
+
+	zero(device->meta_state.resolve);
+
+	struct radv_shader_module vs_module = { .nir = build_nir_vs() };
+	if (!vs_module.nir) {
+		/* XXX: Need more accurate error */
+		res = VK_ERROR_OUT_OF_HOST_MEMORY;
+		goto fail;
+	}
+
+	res = create_pass(device);
+	if (res != VK_SUCCESS)
+		goto fail;
+
+	VkShaderModule vs_module_h = radv_shader_module_to_handle(&vs_module);
+	res = create_pipeline(device, vs_module_h);
+	if (res != VK_SUCCESS)
+		goto fail;
+
+	goto cleanup;
+
+fail:
+	radv_device_finish_meta_resolve_state(device);
+
+cleanup:
+	ralloc_free(vs_module.nir);
+
+	return res;
+}
+
+static void
+emit_resolve(struct radv_cmd_buffer *cmd_buffer,
+             const VkOffset2D *src_offset,
+             const VkOffset2D *dest_offset,
+             const VkExtent2D *resolve_extent)
+{
+	struct radv_device *device = cmd_buffer->device;
+	VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
+	uint32_t offset;
+	const struct vertex_attrs vertex_data[3] = {
+		{
+			.position = {
+				dest_offset->x,
+				dest_offset->y,
+			},
+			.tex_position = {
+				src_offset->x,
+				src_offset->y,
+			},
+		},
+		{
+			.position = {
+				dest_offset->x,
+				dest_offset->y + resolve_extent->height,
+			},
+			.tex_position = {
+				src_offset->x,
+				src_offset->y + resolve_extent->height,
+			},
+		},
+		{
+			.position = {
+				dest_offset->x + resolve_extent->width,
+				dest_offset->y,
+			},
+			.tex_position = {
+				src_offset->x + resolve_extent->width,
+				src_offset->y,
+			},
+		},
+	};
+
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB;
+	radv_cmd_buffer_upload_data(cmd_buffer, sizeof(vertex_data), 16, vertex_data, &offset);
+	struct radv_buffer vertex_buffer = {
+		.device = device,
+		.size = sizeof(vertex_data),
+		.bo = cmd_buffer->upload.upload_bo,
+		.offset = offset,
+	};
+
+	VkBuffer vertex_buffer_h = radv_buffer_to_handle(&vertex_buffer);
+
+	radv_CmdBindVertexBuffers(cmd_buffer_h,
+				  /*firstBinding*/ 0,
+				  /*bindingCount*/ 1,
+				  (VkBuffer[]) { vertex_buffer_h },
+				  (VkDeviceSize[]) { 0 });
+
+	VkPipeline pipeline_h = device->meta_state.resolve.pipeline;
+	RADV_FROM_HANDLE(radv_pipeline, pipeline, pipeline_h);
+
+	if (cmd_buffer->state.pipeline != pipeline) {
+		radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
+				     pipeline_h);
+	}
+
+	RADV_CALL(CmdDraw)(cmd_buffer_h, 3, 1, 0, 0);
+	cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB;
+	si_emit_cache_flush(cmd_buffer);
+}
+
+void radv_CmdResolveImage(
+	VkCommandBuffer                             cmd_buffer_h,
+	VkImage                                     src_image_h,
+	VkImageLayout                               src_image_layout,
+	VkImage                                     dest_image_h,
+	VkImageLayout                               dest_image_layout,
+	uint32_t                                    region_count,
+	const VkImageResolve*                       regions)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, cmd_buffer_h);
+	RADV_FROM_HANDLE(radv_image, src_image, src_image_h);
+	RADV_FROM_HANDLE(radv_image, dest_image, dest_image_h);
+	struct radv_device *device = cmd_buffer->device;
+	struct radv_meta_saved_state saved_state;
+	VkDevice device_h = radv_device_to_handle(device);
+	bool use_compute_resolve = false;
+
+	/* we can use the hw resolve only for single full resolves */
+	if (region_count == 1) {
+		if (regions[0].srcOffset.x ||
+		    regions[0].srcOffset.y ||
+		    regions[0].srcOffset.z)
+			use_compute_resolve = true;
+		if (regions[0].dstOffset.x ||
+		    regions[0].dstOffset.y ||
+		    regions[0].dstOffset.z)
+			use_compute_resolve = true;
+
+		if (regions[0].extent.width != src_image->extent.width ||
+		    regions[0].extent.height != src_image->extent.height ||
+		    regions[0].extent.depth != src_image->extent.depth)
+			use_compute_resolve = true;
+	} else
+		use_compute_resolve = true;
+
+	if (use_compute_resolve) {
+		radv_meta_resolve_compute_image(cmd_buffer,
+						src_image,
+						src_image_layout,
+						dest_image,
+						dest_image_layout,
+						region_count, regions);
+		return;
+	}
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	assert(src_image->samples > 1);
+	assert(dest_image->samples == 1);
+
+	if (src_image->samples >= 16) {
+		/* See commit aa3f9aaf31e9056a255f9e0472ebdfdaa60abe54 for the
+		 * glBlitFramebuffer workaround for samples >= 16.
+		 */
+		radv_finishme("vkCmdResolveImage: need interpolation workaround when "
+			      "samples >= 16");
+	}
+
+	if (src_image->array_size > 1)
+		radv_finishme("vkCmdResolveImage: multisample array images");
+
+	for (uint32_t r = 0; r < region_count; ++r) {
+		const VkImageResolve *region = &regions[r];
+
+		/* From the Vulkan 1.0 spec:
+		 *
+		 *    - The aspectMask member of srcSubresource and dstSubresource must
+		 *      only contain VK_IMAGE_ASPECT_COLOR_BIT
+		 *
+		 *    - The layerCount member of srcSubresource and dstSubresource must
+		 *      match
+		 */
+		assert(region->srcSubresource.aspectMask == VK_IMAGE_ASPECT_COLOR_BIT);
+		assert(region->dstSubresource.aspectMask == VK_IMAGE_ASPECT_COLOR_BIT);
+		assert(region->srcSubresource.layerCount ==
+		       region->dstSubresource.layerCount);
+
+		const uint32_t src_base_layer =
+			radv_meta_get_iview_layer(src_image, &region->srcSubresource,
+						  &region->srcOffset);
+
+		const uint32_t dest_base_layer =
+			radv_meta_get_iview_layer(dest_image, &region->dstSubresource,
+						  &region->dstOffset);
+
+		/**
+		 * From Vulkan 1.0.6 spec: 18.6 Resolving Multisample Images
+		 *
+		 *    extent is the size in texels of the source image to resolve in width,
+		 *    height and depth. 1D images use only x and width. 2D images use x, y,
+		 *    width and height. 3D images use x, y, z, width, height and depth.
+		 *
+		 *    srcOffset and dstOffset select the initial x, y, and z offsets in
+		 *    texels of the sub-regions of the source and destination image data.
+		 *    extent is the size in texels of the source image to resolve in width,
+		 *    height and depth. 1D images use only x and width. 2D images use x, y,
+		 *    width and height. 3D images use x, y, z, width, height and depth.
+		 */
+		const struct VkExtent3D extent =
+			radv_sanitize_image_extent(src_image->type, region->extent);
+		const struct VkOffset3D srcOffset =
+			radv_sanitize_image_offset(src_image->type, region->srcOffset);
+		const struct VkOffset3D dstOffset =
+			radv_sanitize_image_offset(dest_image->type, region->dstOffset);
+
+
+		for (uint32_t layer = 0; layer < region->srcSubresource.layerCount;
+		     ++layer) {
+
+			struct radv_image_view src_iview;
+			radv_image_view_init(&src_iview, cmd_buffer->device,
+					     &(VkImageViewCreateInfo) {
+						     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+							     .image = src_image_h,
+							     .viewType = radv_meta_get_view_type(src_image),
+							     .format = src_image->vk_format,
+							     .subresourceRange = {
+							     .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+							     .baseMipLevel = region->srcSubresource.mipLevel,
+							     .levelCount = 1,
+							     .baseArrayLayer = src_base_layer + layer,
+							     .layerCount = 1,
+						     },
+							     },
+					     cmd_buffer, VK_IMAGE_USAGE_SAMPLED_BIT);
+
+			struct radv_image_view dest_iview;
+			radv_image_view_init(&dest_iview, cmd_buffer->device,
+					     &(VkImageViewCreateInfo) {
+						     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+							     .image = dest_image_h,
+							     .viewType = radv_meta_get_view_type(dest_image),
+							     .format = dest_image->vk_format,
+							     .subresourceRange = {
+							     .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+							     .baseMipLevel = region->dstSubresource.mipLevel,
+							     .levelCount = 1,
+							     .baseArrayLayer = dest_base_layer + layer,
+							     .layerCount = 1,
+						     },
+							     },
+					     cmd_buffer, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
+
+			VkFramebuffer fb_h;
+			radv_CreateFramebuffer(device_h,
+					       &(VkFramebufferCreateInfo) {
+						       .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
+							       .attachmentCount = 2,
+							       .pAttachments = (VkImageView[]) {
+							       radv_image_view_to_handle(&src_iview),
+							       radv_image_view_to_handle(&dest_iview),
+						       },
+						       .width = radv_minify(dest_image->extent.width,
+									    region->dstSubresource.mipLevel),
+						       .height = radv_minify(dest_image->extent.height,
+									      region->dstSubresource.mipLevel),
+						       .layers = 1
+					       },
+					       &cmd_buffer->pool->alloc,
+					       &fb_h);
+
+			RADV_CALL(CmdBeginRenderPass)(cmd_buffer_h,
+						      &(VkRenderPassBeginInfo) {
+							      .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+								      .renderPass = device->meta_state.resolve.pass,
+								      .framebuffer = fb_h,
+								      .renderArea = {
+								      .offset = {
+									      dstOffset.x,
+									      dstOffset.y,
+								      },
+								      .extent = {
+									      extent.width,
+									      extent.height,
+								      }
+							      },
+							      .clearValueCount = 0,
+							      .pClearValues = NULL,
+						      },
+						      VK_SUBPASS_CONTENTS_INLINE);
+
+			emit_resolve(cmd_buffer,
+				     &(VkOffset2D) {
+					     .x = srcOffset.x,
+					     .y = srcOffset.y,
+				     },
+				     &(VkOffset2D) {
+					     .x = dstOffset.x,
+					     .y = dstOffset.y,
+				     },
+				     &(VkExtent2D) {
+					     .width = extent.width,
+					     .height = extent.height,
+				     });
+
+			RADV_CALL(CmdEndRenderPass)(cmd_buffer_h);
+
+			radv_DestroyFramebuffer(device_h, fb_h,
+						&cmd_buffer->pool->alloc);
+		}
+	}
+
+	radv_meta_restore(&saved_state, cmd_buffer);
+}
+
+/**
+ * Emit any needed resolves for the current subpass.
+ */
+void
+radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer)
+{
+	struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
+	const struct radv_subpass *subpass = cmd_buffer->state.subpass;
+	struct radv_meta_saved_state saved_state;
+
+	/* FINISHME(perf): Skip clears for resolve attachments.
+	 *
+	 * From the Vulkan 1.0 spec:
+	 *
+	 *    If the first use of an attachment in a render pass is as a resolve
+	 *    attachment, then the loadOp is effectively ignored as the resolve is
+	 *    guaranteed to overwrite all pixels in the render area.
+	 */
+
+	if (!subpass->has_resolve)
+		return;
+
+	radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+
+	for (uint32_t i = 0; i < subpass->color_count; ++i) {
+		VkAttachmentReference src_att = subpass->color_attachments[i];
+		VkAttachmentReference dest_att = subpass->resolve_attachments[i];
+		struct radv_image *dst_img = cmd_buffer->state.framebuffer->attachments[dest_att.attachment].attachment->image;
+		if (dest_att.attachment == VK_ATTACHMENT_UNUSED)
+			continue;
+
+		if (dst_img->surface.dcc_size) {
+			radv_initialize_dcc(cmd_buffer, dst_img, 0xffffffff);
+			cmd_buffer->state.attachments[dest_att.attachment].current_layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
+		}
+
+		struct radv_subpass resolve_subpass = {
+			.color_count = 2,
+			.color_attachments = (VkAttachmentReference[]) { src_att, dest_att },
+			.depth_stencil_attachment = { .attachment = VK_ATTACHMENT_UNUSED },
+		};
+
+		radv_cmd_buffer_set_subpass(cmd_buffer, &resolve_subpass, false);
+
+		/* Subpass resolves must respect the render area. We can ignore the
+		 * render area here because vkCmdBeginRenderPass set the render area
+		 * with 3DSTATE_DRAWING_RECTANGLE.
+		 *
+		 * XXX(chadv): Does the hardware really respect
+		 * 3DSTATE_DRAWING_RECTANGLE when draing a 3DPRIM_RECTLIST?
+		 */
+		emit_resolve(cmd_buffer,
+			     &(VkOffset2D) { 0, 0 },
+			     &(VkOffset2D) { 0, 0 },
+			     &(VkExtent2D) { fb->width, fb->height });
+	}
+
+	cmd_buffer->state.subpass = subpass;
+	radv_meta_restore(&saved_state, cmd_buffer);
+}
diff --git a/src/amd/vulkan/radv_meta_resolve_cs.c b/src/amd/vulkan/radv_meta_resolve_cs.c
new file mode 100644
index 0000000000..c6525b6f36
--- /dev/null
+++ b/src/amd/vulkan/radv_meta_resolve_cs.c
@@ -0,0 +1,461 @@
+/*
+ * Copyright © 2016 Dave Airlie
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+
+#include <assert.h>
+#include <stdbool.h>
+
+#include "radv_meta.h"
+#include "radv_private.h"
+#include "nir/nir_builder.h"
+#include "sid.h"
+#include "vk_format.h"
+
+static nir_shader *
+build_resolve_compute_shader(struct radv_device *dev, bool is_integer, int samples)
+{
+	nir_builder b;
+	char name[64];
+	nir_if *outer_if = NULL;
+	const struct glsl_type *sampler_type = glsl_sampler_type(GLSL_SAMPLER_DIM_MS,
+								 false,
+								 false,
+								 GLSL_TYPE_FLOAT);
+	const struct glsl_type *img_type = glsl_sampler_type(GLSL_SAMPLER_DIM_2D,
+							     false,
+							     false,
+							     GLSL_TYPE_FLOAT);
+	snprintf(name, 64, "meta_resolve_cs-%d-%s", samples, is_integer ? "int" : "float");
+	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
+	b.shader->info.name = ralloc_strdup(b.shader, name);
+	b.shader->info.cs.local_size[0] = 16;
+	b.shader->info.cs.local_size[1] = 16;
+	b.shader->info.cs.local_size[2] = 1;
+
+	nir_variable *input_img = nir_variable_create(b.shader, nir_var_uniform,
+						      sampler_type, "s_tex");
+	input_img->data.descriptor_set = 0;
+	input_img->data.binding = 0;
+
+	nir_variable *output_img = nir_variable_create(b.shader, nir_var_uniform,
+						       img_type, "out_img");
+	output_img->data.descriptor_set = 0;
+	output_img->data.binding = 1;
+	nir_ssa_def *invoc_id = nir_load_system_value(&b, nir_intrinsic_load_local_invocation_id, 0);
+	nir_ssa_def *wg_id = nir_load_system_value(&b, nir_intrinsic_load_work_group_id, 0);
+	nir_ssa_def *block_size = nir_imm_ivec4(&b,
+						b.shader->info.cs.local_size[0],
+						b.shader->info.cs.local_size[1],
+						b.shader->info.cs.local_size[2], 0);
+
+	nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
+
+	nir_intrinsic_instr *src_offset = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
+	src_offset->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
+	src_offset->num_components = 2;
+	nir_ssa_dest_init(&src_offset->instr, &src_offset->dest, 2, 32, "src_offset");
+	nir_builder_instr_insert(&b, &src_offset->instr);
+
+	nir_intrinsic_instr *dst_offset = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
+	dst_offset->src[0] = nir_src_for_ssa(nir_imm_int(&b, 8));
+	dst_offset->num_components = 2;
+	nir_ssa_dest_init(&dst_offset->instr, &dst_offset->dest, 2, 32, "dst_offset");
+	nir_builder_instr_insert(&b, &dst_offset->instr);
+
+	nir_ssa_def *img_coord = nir_iadd(&b, global_id, &src_offset->dest.ssa);
+	/* do a txf_ms on each sample */
+	nir_ssa_def *tmp;
+
+	nir_tex_instr *tex = nir_tex_instr_create(b.shader, 2);
+	tex->sampler_dim = GLSL_SAMPLER_DIM_MS;
+	tex->op = nir_texop_txf_ms;
+	tex->src[0].src_type = nir_tex_src_coord;
+	tex->src[0].src = nir_src_for_ssa(img_coord);
+	tex->src[1].src_type = nir_tex_src_ms_index;
+	tex->src[1].src = nir_src_for_ssa(nir_imm_int(&b, 0));
+	tex->dest_type = nir_type_float;
+	tex->is_array = false;
+	tex->coord_components = 2;
+	tex->texture = nir_deref_var_create(tex, input_img);
+	tex->sampler = NULL;
+
+	nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
+	nir_builder_instr_insert(&b, &tex->instr);
+
+	tmp = &tex->dest.ssa;
+	nir_variable *color =
+		nir_local_variable_create(b.impl, glsl_vec4_type(), "color");
+
+	if (!is_integer && samples > 1) {
+		nir_tex_instr *tex_all_same = nir_tex_instr_create(b.shader, 1);
+		tex_all_same->sampler_dim = GLSL_SAMPLER_DIM_MS;
+		tex_all_same->op = nir_texop_samples_identical;
+		tex_all_same->src[0].src_type = nir_tex_src_coord;
+		tex_all_same->src[0].src = nir_src_for_ssa(img_coord);
+		tex_all_same->dest_type = nir_type_float;
+		tex_all_same->is_array = false;
+		tex_all_same->coord_components = 2;
+		tex_all_same->texture = nir_deref_var_create(tex_all_same, input_img);
+		tex_all_same->sampler = NULL;
+
+		nir_ssa_dest_init(&tex_all_same->instr, &tex_all_same->dest, 1, 32, "tex");
+		nir_builder_instr_insert(&b, &tex_all_same->instr);
+
+		nir_ssa_def *all_same = nir_ine(&b, &tex_all_same->dest.ssa, nir_imm_int(&b, 0));
+		nir_if *if_stmt = nir_if_create(b.shader);
+		if_stmt->condition = nir_src_for_ssa(all_same);
+		nir_cf_node_insert(b.cursor, &if_stmt->cf_node);
+
+		b.cursor = nir_after_cf_list(&if_stmt->then_list);
+		for (int i = 1; i < samples; i++) {
+			nir_tex_instr *tex_add = nir_tex_instr_create(b.shader, 2);
+			tex_add->sampler_dim = GLSL_SAMPLER_DIM_MS;
+			tex_add->op = nir_texop_txf_ms;
+			tex_add->src[0].src_type = nir_tex_src_coord;
+			tex_add->src[0].src = nir_src_for_ssa(img_coord);
+			tex_add->src[1].src_type = nir_tex_src_ms_index;
+			tex_add->src[1].src = nir_src_for_ssa(nir_imm_int(&b, i));
+			tex_add->dest_type = nir_type_float;
+			tex_add->is_array = false;
+			tex_add->coord_components = 2;
+			tex_add->texture = nir_deref_var_create(tex_add, input_img);
+			tex_add->sampler = NULL;
+
+			nir_ssa_dest_init(&tex_add->instr, &tex_add->dest, 4, 32, "tex");
+			nir_builder_instr_insert(&b, &tex_add->instr);
+
+			tmp = nir_fadd(&b, tmp, &tex_add->dest.ssa);
+		}
+
+		tmp = nir_fdiv(&b, tmp, nir_imm_float(&b, samples));
+		nir_store_var(&b, color, tmp, 0xf);
+		b.cursor = nir_after_cf_list(&if_stmt->else_list);
+		outer_if = if_stmt;
+	}
+	nir_store_var(&b, color, &tex->dest.ssa, 0xf);
+
+	if (outer_if)
+		b.cursor = nir_after_cf_node(&outer_if->cf_node);
+
+	nir_ssa_def *newv = nir_load_var(&b, color);
+	nir_ssa_def *coord = nir_iadd(&b, global_id, &dst_offset->dest.ssa);
+	nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_image_store);
+	store->src[0] = nir_src_for_ssa(coord);
+	store->src[1] = nir_src_for_ssa(nir_ssa_undef(&b, 1, 32));
+	store->src[2] = nir_src_for_ssa(newv);
+	store->variables[0] = nir_deref_var_create(store, output_img);
+	nir_builder_instr_insert(&b, &store->instr);
+	return b.shader;
+}
+
+
+static VkResult
+create_layout(struct radv_device *device)
+{
+	VkResult result;
+	/*
+	 * two descriptors one for the image being sampled
+	 * one for the buffer being written.
+	 */
+	VkDescriptorSetLayoutCreateInfo ds_create_info = {
+		.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+		.bindingCount = 2,
+		.pBindings = (VkDescriptorSetLayoutBinding[]) {
+			{
+				.binding = 0,
+				.descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
+				.descriptorCount = 1,
+				.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+				.pImmutableSamplers = NULL
+			},
+			{
+				.binding = 1,
+				.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+				.descriptorCount = 1,
+				.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+				.pImmutableSamplers = NULL
+			},
+		}
+	};
+
+	result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
+						&ds_create_info,
+						&device->meta_state.alloc,
+						&device->meta_state.resolve_compute.ds_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+
+	VkPipelineLayoutCreateInfo pl_create_info = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+		.setLayoutCount = 1,
+		.pSetLayouts = &device->meta_state.resolve_compute.ds_layout,
+		.pushConstantRangeCount = 1,
+		.pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 16},
+	};
+
+	result = radv_CreatePipelineLayout(radv_device_to_handle(device),
+					  &pl_create_info,
+					  &device->meta_state.alloc,
+					  &device->meta_state.resolve_compute.p_layout);
+	if (result != VK_SUCCESS)
+		goto fail;
+	return VK_SUCCESS;
+fail:
+	return result;
+}
+
+static VkResult
+create_resolve_pipeline(struct radv_device *device,
+			int samples,
+			bool is_integer,
+			VkPipeline *pipeline)
+{
+	VkResult result;
+	struct radv_shader_module cs = { .nir = NULL };
+
+	cs.nir = build_resolve_compute_shader(device, is_integer, samples);
+
+	/* compute shader */
+
+	VkPipelineShaderStageCreateInfo pipeline_shader_stage = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+		.stage = VK_SHADER_STAGE_COMPUTE_BIT,
+		.module = radv_shader_module_to_handle(&cs),
+		.pName = "main",
+		.pSpecializationInfo = NULL,
+	};
+
+	VkComputePipelineCreateInfo vk_pipeline_info = {
+		.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+		.stage = pipeline_shader_stage,
+		.flags = 0,
+		.layout = device->meta_state.resolve_compute.p_layout,
+	};
+
+	result = radv_CreateComputePipelines(radv_device_to_handle(device),
+					     radv_pipeline_cache_to_handle(&device->meta_state.cache),
+					     1, &vk_pipeline_info, NULL,
+					     pipeline);
+	if (result != VK_SUCCESS)
+		goto fail;
+
+	ralloc_free(cs.nir);
+	return VK_SUCCESS;
+fail:
+	ralloc_free(cs.nir);
+	return result;
+}
+
+VkResult
+radv_device_init_meta_resolve_compute_state(struct radv_device *device)
+{
+	struct radv_meta_state *state = &device->meta_state;
+	VkResult res;
+	memset(&device->meta_state.resolve_compute, 0, sizeof(device->meta_state.resolve_compute));
+
+	res = create_layout(device);
+	if (res != VK_SUCCESS)
+		return res;
+
+	for (uint32_t i = 0; i < MAX_SAMPLES_LOG2; ++i) {
+		uint32_t samples = 1 << i;
+
+		res = create_resolve_pipeline(device, samples, false,
+					      &state->resolve_compute.rc[i].pipeline);
+
+		res = create_resolve_pipeline(device, samples, true,
+					      &state->resolve_compute.rc[i].i_pipeline);
+
+	}
+
+	return res;
+}
+
+void
+radv_device_finish_meta_resolve_compute_state(struct radv_device *device)
+{
+	struct radv_meta_state *state = &device->meta_state;
+	for (uint32_t i = 0; i < MAX_SAMPLES_LOG2; ++i) {
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     state->resolve_compute.rc[i].pipeline,
+				     &state->alloc);
+
+		radv_DestroyPipeline(radv_device_to_handle(device),
+				     state->resolve_compute.rc[i].i_pipeline,
+				     &state->alloc);
+	}
+
+	radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
+					state->resolve_compute.ds_layout,
+					&state->alloc);
+	radv_DestroyPipelineLayout(radv_device_to_handle(device),
+				   state->resolve_compute.p_layout,
+				   &state->alloc);
+}
+
+void radv_meta_resolve_compute_image(struct radv_cmd_buffer *cmd_buffer,
+				     struct radv_image *src_image,
+				     VkImageLayout src_image_layout,
+				     struct radv_image *dest_image,
+				     VkImageLayout dest_image_layout,
+				     uint32_t region_count,
+				     const VkImageResolve *regions)
+{
+	struct radv_device *device = cmd_buffer->device;
+	struct radv_meta_saved_compute_state saved_state;
+	const uint32_t samples = src_image->samples;
+	const uint32_t samples_log2 = ffs(samples) - 1;
+	radv_meta_save_compute(&saved_state, cmd_buffer, 16);
+
+	for (uint32_t r = 0; r < region_count; ++r) {
+		const VkImageResolve *region = &regions[r];
+
+		assert(region->srcSubresource.aspectMask == VK_IMAGE_ASPECT_COLOR_BIT);
+		assert(region->dstSubresource.aspectMask == VK_IMAGE_ASPECT_COLOR_BIT);
+		assert(region->srcSubresource.layerCount == region->dstSubresource.layerCount);
+
+		const uint32_t src_base_layer =
+			radv_meta_get_iview_layer(src_image, &region->srcSubresource,
+						  &region->srcOffset);
+
+		const uint32_t dest_base_layer =
+			radv_meta_get_iview_layer(dest_image, &region->dstSubresource,
+						  &region->dstOffset);
+
+		const struct VkExtent3D extent =
+			radv_sanitize_image_extent(src_image->type, region->extent);
+		const struct VkOffset3D srcOffset =
+			radv_sanitize_image_offset(src_image->type, region->srcOffset);
+		const struct VkOffset3D dstOffset =
+			radv_sanitize_image_offset(dest_image->type, region->dstOffset);
+
+		for (uint32_t layer = 0; layer < region->srcSubresource.layerCount;
+		     ++layer) {
+
+			struct radv_image_view src_iview;
+			VkDescriptorSet set;
+			radv_image_view_init(&src_iview, cmd_buffer->device,
+					     &(VkImageViewCreateInfo) {
+						     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+							     .image = radv_image_to_handle(src_image),
+							     .viewType = radv_meta_get_view_type(src_image),
+							     .format = src_image->vk_format,
+							     .subresourceRange = {
+							     .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+							     .baseMipLevel = region->srcSubresource.mipLevel,
+							     .levelCount = 1,
+							     .baseArrayLayer = src_base_layer + layer,
+							     .layerCount = 1,
+						     },
+					     },
+					     cmd_buffer, VK_IMAGE_USAGE_SAMPLED_BIT);
+
+			struct radv_image_view dest_iview;
+			radv_image_view_init(&dest_iview, cmd_buffer->device,
+					     &(VkImageViewCreateInfo) {
+						     .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+							     .image = radv_image_to_handle(dest_image),
+							     .viewType = radv_meta_get_view_type(dest_image),
+							     .format = dest_image->vk_format,
+							     .subresourceRange = {
+							     .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+							     .baseMipLevel = region->dstSubresource.mipLevel,
+							     .levelCount = 1,
+							     .baseArrayLayer = dest_base_layer + layer,
+							     .layerCount = 1,
+						     },
+							     },
+					     cmd_buffer, VK_IMAGE_USAGE_STORAGE_BIT);
+
+
+			radv_temp_descriptor_set_create(device, cmd_buffer,
+							device->meta_state.resolve_compute.ds_layout,
+							&set);
+
+			radv_UpdateDescriptorSets(radv_device_to_handle(device),
+						  2, /* writeCount */
+						  (VkWriteDescriptorSet[]) {
+						  {
+						  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+						  .dstSet = set,
+						  .dstBinding = 0,
+						  .dstArrayElement = 0,
+						  .descriptorCount = 1,
+						  .descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
+						  .pImageInfo = (VkDescriptorImageInfo[]) {
+							  {
+								  .sampler = NULL,
+								  .imageView = radv_image_view_to_handle(&src_iview),
+								  .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
+							  },
+						  }
+					  },
+					  {
+						  .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+						  .dstSet = set,
+						  .dstBinding = 1,
+						  .dstArrayElement = 0,
+						  .descriptorCount = 1,
+						  .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+						  .pImageInfo = (VkDescriptorImageInfo[]) {
+							  {
+								  .sampler = NULL,
+								  .imageView = radv_image_view_to_handle(&dest_iview),
+								  .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
+							  },
+						  }
+					  }
+				  }, 0, NULL);
+
+			radv_CmdBindDescriptorSets(radv_cmd_buffer_to_handle(cmd_buffer),
+						   VK_PIPELINE_BIND_POINT_COMPUTE,
+						   device->meta_state.resolve_compute.p_layout, 0, 1,
+						   &set, 0, NULL);
+
+			VkPipeline pipeline;
+			if (vk_format_is_int(src_image->vk_format))
+				pipeline = device->meta_state.resolve_compute.rc[samples_log2].i_pipeline;
+			else
+				pipeline = device->meta_state.resolve_compute.rc[samples_log2].pipeline;
+			if (cmd_buffer->state.compute_pipeline != radv_pipeline_from_handle(pipeline)) {
+				radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+						     VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
+			}
+
+			unsigned push_constants[4] = {
+				srcOffset.x,
+				srcOffset.y,
+				dstOffset.x,
+				dstOffset.y,
+			};
+			radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
+					      device->meta_state.resolve_compute.p_layout,
+					      VK_SHADER_STAGE_COMPUTE_BIT, 0, 16,
+					      push_constants);
+			radv_unaligned_dispatch(cmd_buffer, extent.width, extent.height, 1);
+			radv_temp_descriptor_set_destroy(cmd_buffer->device, set);
+		}
+	}
+	radv_meta_restore_compute(&saved_state, cmd_buffer, 16);
+}
diff --git a/src/amd/vulkan/radv_pass.c b/src/amd/vulkan/radv_pass.c
new file mode 100644
index 0000000000..fa217977ca
--- /dev/null
+++ b/src/amd/vulkan/radv_pass.c
@@ -0,0 +1,183 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based in part on anv driver which is:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#include "radv_private.h"
+
+VkResult radv_CreateRenderPass(
+	VkDevice                                    _device,
+	const VkRenderPassCreateInfo*               pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkRenderPass*                               pRenderPass)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_render_pass *pass;
+	size_t size;
+	size_t attachments_offset;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO);
+
+	size = sizeof(*pass);
+	size += pCreateInfo->subpassCount * sizeof(pass->subpasses[0]);
+	attachments_offset = size;
+	size += pCreateInfo->attachmentCount * sizeof(pass->attachments[0]);
+
+	pass = radv_alloc2(&device->alloc, pAllocator, size, 8,
+			   VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (pass == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	memset(pass, 0, size);
+	pass->attachment_count = pCreateInfo->attachmentCount;
+	pass->subpass_count = pCreateInfo->subpassCount;
+	pass->attachments = (void *) pass + attachments_offset;
+
+	for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
+		struct radv_render_pass_attachment *att = &pass->attachments[i];
+
+		att->format = pCreateInfo->pAttachments[i].format;
+		att->samples = pCreateInfo->pAttachments[i].samples;
+		att->load_op = pCreateInfo->pAttachments[i].loadOp;
+		att->stencil_load_op = pCreateInfo->pAttachments[i].stencilLoadOp;
+		att->initial_layout =  pCreateInfo->pAttachments[i].initialLayout;
+		att->final_layout =  pCreateInfo->pAttachments[i].finalLayout;
+		// att->store_op = pCreateInfo->pAttachments[i].storeOp;
+		// att->stencil_store_op = pCreateInfo->pAttachments[i].stencilStoreOp;
+	}
+	uint32_t subpass_attachment_count = 0;
+	VkAttachmentReference *p;
+	for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
+		const VkSubpassDescription *desc = &pCreateInfo->pSubpasses[i];
+
+		subpass_attachment_count +=
+			desc->inputAttachmentCount +
+			desc->colorAttachmentCount +
+			/* Count colorAttachmentCount again for resolve_attachments */
+			desc->colorAttachmentCount;
+	}
+
+	if (subpass_attachment_count) {
+		pass->subpass_attachments =
+			radv_alloc2(&device->alloc, pAllocator,
+				    subpass_attachment_count * sizeof(VkAttachmentReference), 8,
+				    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+		if (pass->subpass_attachments == NULL) {
+			radv_free2(&device->alloc, pAllocator, pass);
+			return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		}
+	} else
+		pass->subpass_attachments = NULL;
+
+	p = pass->subpass_attachments;
+	for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
+		const VkSubpassDescription *desc = &pCreateInfo->pSubpasses[i];
+		struct radv_subpass *subpass = &pass->subpasses[i];
+
+		subpass->input_count = desc->inputAttachmentCount;
+		subpass->color_count = desc->colorAttachmentCount;
+
+		if (desc->inputAttachmentCount > 0) {
+			subpass->input_attachments = p;
+			p += desc->inputAttachmentCount;
+
+			for (uint32_t j = 0; j < desc->inputAttachmentCount; j++) {
+				subpass->input_attachments[j]
+					= desc->pInputAttachments[j];
+			}
+		}
+
+		if (desc->colorAttachmentCount > 0) {
+			subpass->color_attachments = p;
+			p += desc->colorAttachmentCount;
+
+			for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
+				subpass->color_attachments[j]
+					= desc->pColorAttachments[j];
+			}
+		}
+
+		subpass->has_resolve = false;
+		if (desc->pResolveAttachments) {
+			subpass->resolve_attachments = p;
+			p += desc->colorAttachmentCount;
+
+			for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
+				uint32_t a = desc->pResolveAttachments[j].attachment;
+				subpass->resolve_attachments[j]
+					= desc->pResolveAttachments[j];
+				if (a != VK_ATTACHMENT_UNUSED)
+					subpass->has_resolve = true;
+			}
+		}
+
+		if (desc->pDepthStencilAttachment) {
+			subpass->depth_stencil_attachment =
+				*desc->pDepthStencilAttachment;
+		} else {
+			subpass->depth_stencil_attachment.attachment = VK_ATTACHMENT_UNUSED;
+		}
+	}
+
+	for (unsigned i = 0; i < pCreateInfo->dependencyCount; ++i) {
+		uint32_t dst = pCreateInfo->pDependencies[i].dstSubpass;
+		if (dst == VK_SUBPASS_EXTERNAL) {
+			pass->end_barrier.src_stage_mask = pCreateInfo->pDependencies[i].srcStageMask;
+			pass->end_barrier.src_access_mask = pCreateInfo->pDependencies[i].srcAccessMask;
+			pass->end_barrier.dst_access_mask = pCreateInfo->pDependencies[i].dstAccessMask;
+		} else {
+			pass->subpasses[dst].start_barrier.src_stage_mask = pCreateInfo->pDependencies[i].srcStageMask;
+			pass->subpasses[dst].start_barrier.src_access_mask = pCreateInfo->pDependencies[i].srcAccessMask;
+			pass->subpasses[dst].start_barrier.dst_access_mask = pCreateInfo->pDependencies[i].dstAccessMask;
+		}
+	}
+
+	*pRenderPass = radv_render_pass_to_handle(pass);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroyRenderPass(
+	VkDevice                                    _device,
+	VkRenderPass                                _pass,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_render_pass, pass, _pass);
+
+	if (!_pass)
+		return;
+	radv_free2(&device->alloc, pAllocator, pass->subpass_attachments);
+	radv_free2(&device->alloc, pAllocator, pass);
+}
+
+void radv_GetRenderAreaGranularity(
+    VkDevice                                    device,
+    VkRenderPass                                renderPass,
+    VkExtent2D*                                 pGranularity)
+{
+	pGranularity->width = 1;
+	pGranularity->height = 1;
+}
+
diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
new file mode 100644
index 0000000000..89300e5992
--- /dev/null
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -0,0 +1,1408 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based in part on anv driver which is:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "util/mesa-sha1.h"
+#include "radv_private.h"
+#include "nir/nir.h"
+#include "nir/nir_builder.h"
+#include "spirv/nir_spirv.h"
+
+#include <llvm-c/Core.h>
+#include <llvm-c/TargetMachine.h>
+
+#include "sid.h"
+#include "r600d_common.h"
+#include "ac_binary.h"
+#include "ac_llvm_util.h"
+#include "ac_nir_to_llvm.h"
+#include "vk_format.h"
+#include "util/debug.h"
+void radv_shader_variant_destroy(struct radv_device *device,
+                                 struct radv_shader_variant *variant);
+
+static const struct nir_shader_compiler_options nir_options = {
+	.vertex_id_zero_based = true,
+	.lower_scmp = true,
+	.lower_flrp32 = true,
+	.lower_fsat = true,
+	.lower_pack_snorm_2x16 = true,
+	.lower_pack_snorm_4x8 = true,
+	.lower_pack_unorm_2x16 = true,
+	.lower_pack_unorm_4x8 = true,
+	.lower_unpack_snorm_2x16 = true,
+	.lower_unpack_snorm_4x8 = true,
+	.lower_unpack_unorm_2x16 = true,
+	.lower_unpack_unorm_4x8 = true,
+	.lower_extract_byte = true,
+	.lower_extract_word = true,
+};
+
+VkResult radv_CreateShaderModule(
+	VkDevice                                    _device,
+	const VkShaderModuleCreateInfo*             pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkShaderModule*                             pShaderModule)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_shader_module *module;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO);
+	assert(pCreateInfo->flags == 0);
+
+	module = radv_alloc2(&device->alloc, pAllocator,
+			     sizeof(*module) + pCreateInfo->codeSize, 8,
+			     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (module == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	module->nir = NULL;
+	module->size = pCreateInfo->codeSize;
+	memcpy(module->data, pCreateInfo->pCode, module->size);
+
+	_mesa_sha1_compute(module->data, module->size, module->sha1);
+
+	*pShaderModule = radv_shader_module_to_handle(module);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroyShaderModule(
+	VkDevice                                    _device,
+	VkShaderModule                              _module,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_shader_module, module, _module);
+
+	if (!module)
+		return;
+
+	radv_free2(&device->alloc, pAllocator, module);
+}
+
+void radv_DestroyPipeline(
+	VkDevice                                    _device,
+	VkPipeline                                  _pipeline,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
+
+	if (!_pipeline)
+		return;
+
+	for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i)
+		if (pipeline->shaders[i])
+			radv_shader_variant_destroy(device, pipeline->shaders[i]);
+
+	radv_free2(&device->alloc, pAllocator, pipeline);
+}
+
+
+static void
+radv_optimize_nir(struct nir_shader *shader)
+{
+        bool progress;
+
+        do {
+                progress = false;
+
+                NIR_PASS_V(shader, nir_lower_vars_to_ssa);
+                NIR_PASS_V(shader, nir_lower_alu_to_scalar);
+                NIR_PASS_V(shader, nir_lower_phis_to_scalar);
+
+                NIR_PASS(progress, shader, nir_copy_prop);
+                NIR_PASS(progress, shader, nir_opt_remove_phis);
+                NIR_PASS(progress, shader, nir_opt_dce);
+                NIR_PASS(progress, shader, nir_opt_dead_cf);
+                NIR_PASS(progress, shader, nir_opt_cse);
+                NIR_PASS(progress, shader, nir_opt_peephole_select, 8);
+                NIR_PASS(progress, shader, nir_opt_algebraic);
+                NIR_PASS(progress, shader, nir_opt_constant_folding);
+                NIR_PASS(progress, shader, nir_opt_undef);
+        } while (progress);
+}
+
+static nir_shader *
+radv_shader_compile_to_nir(struct radv_device *device,
+			   struct radv_shader_module *module,
+			   const char *entrypoint_name,
+			   gl_shader_stage stage,
+			   const VkSpecializationInfo *spec_info,
+			   bool dump)
+{
+	if (strcmp(entrypoint_name, "main") != 0) {
+		radv_finishme("Multiple shaders per module not really supported");
+	}
+
+	nir_shader *nir;
+	nir_function *entry_point;
+	if (module->nir) {
+		/* Some things such as our meta clear/blit code will give us a NIR
+		 * shader directly.  In that case, we just ignore the SPIR-V entirely
+		 * and just use the NIR shader */
+		nir = module->nir;
+		nir->options = &nir_options;
+		nir_validate_shader(nir);
+
+		assert(exec_list_length(&nir->functions) == 1);
+		struct exec_node *node = exec_list_get_head(&nir->functions);
+		entry_point = exec_node_data(nir_function, node, node);
+	} else {
+		uint32_t *spirv = (uint32_t *) module->data;
+		assert(module->size % 4 == 0);
+
+		uint32_t num_spec_entries = 0;
+		struct nir_spirv_specialization *spec_entries = NULL;
+		if (spec_info && spec_info->mapEntryCount > 0) {
+			num_spec_entries = spec_info->mapEntryCount;
+			spec_entries = malloc(num_spec_entries * sizeof(*spec_entries));
+			for (uint32_t i = 0; i < num_spec_entries; i++) {
+				VkSpecializationMapEntry entry = spec_info->pMapEntries[i];
+				const void *data = spec_info->pData + entry.offset;
+				assert(data + entry.size <= spec_info->pData + spec_info->dataSize);
+
+				spec_entries[i].id = spec_info->pMapEntries[i].constantID;
+				spec_entries[i].data = *(const uint32_t *)data;
+			}
+		}
+
+		entry_point = spirv_to_nir(spirv, module->size / 4,
+					   spec_entries, num_spec_entries,
+					   stage, entrypoint_name, &nir_options);
+		nir = entry_point->shader;
+		assert(nir->stage == stage);
+		nir_validate_shader(nir);
+
+		free(spec_entries);
+
+		nir_lower_returns(nir);
+		nir_validate_shader(nir);
+
+		nir_inline_functions(nir);
+		nir_validate_shader(nir);
+
+		/* Pick off the single entrypoint that we want */
+		foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
+			if (func != entry_point)
+				exec_node_remove(&func->node);
+		}
+		assert(exec_list_length(&nir->functions) == 1);
+		entry_point->name = ralloc_strdup(entry_point, "main");
+
+		nir_remove_dead_variables(nir, nir_var_shader_in);
+		nir_remove_dead_variables(nir, nir_var_shader_out);
+		nir_remove_dead_variables(nir, nir_var_system_value);
+		nir_validate_shader(nir);
+
+		nir_lower_system_values(nir);
+		nir_validate_shader(nir);
+	}
+
+	/* Vulkan uses the separate-shader linking model */
+	nir->info.separate_shader = true;
+
+	//   nir = brw_preprocess_nir(compiler, nir);
+
+	nir_shader_gather_info(nir, entry_point->impl);
+
+	nir_variable_mode indirect_mask = 0;
+	//   if (compiler->glsl_compiler_options[stage].EmitNoIndirectInput)
+	indirect_mask |= nir_var_shader_in;
+	//   if (compiler->glsl_compiler_options[stage].EmitNoIndirectTemp)
+	indirect_mask |= nir_var_local;
+
+	nir_lower_indirect_derefs(nir, indirect_mask);
+
+	static const nir_lower_tex_options tex_options = {
+	  .lower_txp = ~0,
+	};
+
+	nir_lower_tex(nir, &tex_options);
+
+	nir_lower_vars_to_ssa(nir);
+	nir_lower_var_copies(nir);
+	nir_lower_global_vars_to_local(nir);
+	nir_remove_dead_variables(nir, nir_var_local);
+	radv_optimize_nir(nir);
+
+	if (dump)
+		nir_print_shader(nir, stderr);
+
+	return nir;
+}
+
+void radv_shader_variant_destroy(struct radv_device *device,
+                                 struct radv_shader_variant *variant)
+{
+	if (__sync_fetch_and_sub(&variant->ref_count, 1) != 1)
+		return;
+
+	device->ws->buffer_destroy(variant->bo);
+	free(variant);
+}
+
+static
+struct radv_shader_variant *radv_shader_variant_create(struct radv_device *device,
+                                                       struct nir_shader *shader,
+                                                       struct radv_pipeline_layout *layout,
+                                                       const union ac_shader_variant_key *key,
+						       void** code_out,
+						       unsigned *code_size_out,
+						       bool dump)
+{
+	struct radv_shader_variant *variant = calloc(1, sizeof(struct radv_shader_variant));
+	enum radeon_family chip_family = device->instance->physicalDevice.rad_info.family;
+	LLVMTargetMachineRef tm;
+	if (!variant)
+		return NULL;
+
+	struct ac_nir_compiler_options options = {0};
+	options.layout = layout;
+	if (key)
+		options.key = *key;
+
+	struct ac_shader_binary binary;
+
+	options.unsafe_math = env_var_as_boolean("RADV_UNSAFE_MATH", false);
+	options.family = chip_family;
+	options.chip_class = device->instance->physicalDevice.rad_info.chip_class;
+	tm = ac_create_target_machine(chip_family);
+	ac_compile_nir_shader(tm, &binary, &variant->config,
+			      &variant->info, shader, &options, dump);
+	LLVMDisposeTargetMachine(tm);
+
+	bool scratch_enabled = variant->config.scratch_bytes_per_wave > 0;
+	unsigned vgpr_comp_cnt = 0;
+
+	if (scratch_enabled)
+		radv_finishme("shader scratch space");
+	switch (shader->stage) {
+	case MESA_SHADER_VERTEX:
+		variant->rsrc2 = S_00B12C_USER_SGPR(variant->info.num_user_sgprs) |
+			S_00B12C_SCRATCH_EN(scratch_enabled);
+		vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
+		break;
+	case MESA_SHADER_FRAGMENT:
+		variant->rsrc2 = S_00B12C_USER_SGPR(variant->info.num_user_sgprs) |
+			S_00B12C_SCRATCH_EN(scratch_enabled);
+		break;
+	case MESA_SHADER_COMPUTE:
+		variant->rsrc2 = S_00B84C_USER_SGPR(variant->info.num_user_sgprs) |
+			S_00B84C_SCRATCH_EN(scratch_enabled) |
+			S_00B84C_TGID_X_EN(1) | S_00B84C_TGID_Y_EN(1) |
+			S_00B84C_TGID_Z_EN(1) | S_00B84C_TIDIG_COMP_CNT(2) |
+			S_00B84C_TG_SIZE_EN(1) |
+			S_00B84C_LDS_SIZE(variant->config.lds_size);
+		break;
+	default:
+		unreachable("unsupported shader type");
+		break;
+	}
+
+	variant->rsrc1 =  S_00B848_VGPRS((variant->config.num_vgprs - 1) / 4) |
+		S_00B848_SGPRS((variant->config.num_sgprs - 1) / 8) |
+		S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt) |
+		S_00B848_DX10_CLAMP(1) |
+		S_00B848_FLOAT_MODE(variant->config.float_mode);
+
+	variant->bo = device->ws->buffer_create(device->ws, binary.code_size, 256,
+						RADEON_DOMAIN_GTT, RADEON_FLAG_CPU_ACCESS);
+
+	void *ptr = device->ws->buffer_map(variant->bo);
+	memcpy(ptr, binary.code, binary.code_size);
+	device->ws->buffer_unmap(variant->bo);
+
+	if (code_out) {
+		*code_out = binary.code;
+		*code_size_out = binary.code_size;
+	} else
+		free(binary.code);
+	free(binary.config);
+	free(binary.rodata);
+	free(binary.global_symbol_offsets);
+	free(binary.relocs);
+	free(binary.disasm_string);
+	variant->ref_count = 1;
+	return variant;
+}
+
+
+static struct radv_shader_variant *
+radv_pipeline_compile(struct radv_pipeline *pipeline,
+		      struct radv_pipeline_cache *cache,
+		      struct radv_shader_module *module,
+		      const char *entrypoint,
+		      gl_shader_stage stage,
+		      const VkSpecializationInfo *spec_info,
+		      struct radv_pipeline_layout *layout,
+		      const union ac_shader_variant_key *key,
+		      bool dump)
+{
+	unsigned char sha1[20];
+	struct radv_shader_variant *variant;
+	nir_shader *nir;
+	void *code = NULL;
+	unsigned code_size;
+
+	if (module->nir)
+		_mesa_sha1_compute(module->nir->info.name,
+				   strlen(module->nir->info.name),
+				   module->sha1);
+
+	radv_hash_shader(sha1, module, entrypoint, spec_info, layout, key);
+
+	if (cache) {
+		variant = radv_create_shader_variant_from_pipeline_cache(pipeline->device,
+									 cache,
+									 sha1);
+		if (variant)
+			return variant;
+	}
+
+	nir = radv_shader_compile_to_nir(pipeline->device,
+				         module, entrypoint, stage,
+					 spec_info, dump);
+	if (nir == NULL)
+		return NULL;
+
+	variant = radv_shader_variant_create(pipeline->device, nir, layout, key,
+					     &code, &code_size, dump);
+	if (!module->nir)
+			ralloc_free(nir);
+
+	if (variant && cache)
+		variant = radv_pipeline_cache_insert_shader(cache, sha1, variant,
+							    code, code_size);
+
+	if (code)
+		free(code);
+	return variant;
+}
+
+static uint32_t si_translate_blend_function(VkBlendOp op)
+{
+	switch (op) {
+	case VK_BLEND_OP_ADD:
+		return V_028780_COMB_DST_PLUS_SRC;
+	case VK_BLEND_OP_SUBTRACT:
+		return V_028780_COMB_SRC_MINUS_DST;
+	case VK_BLEND_OP_REVERSE_SUBTRACT:
+		return V_028780_COMB_DST_MINUS_SRC;
+	case VK_BLEND_OP_MIN:
+		return V_028780_COMB_MIN_DST_SRC;
+	case VK_BLEND_OP_MAX:
+		return V_028780_COMB_MAX_DST_SRC;
+	default:
+		return 0;
+	}
+}
+
+static uint32_t si_translate_blend_factor(VkBlendFactor factor)
+{
+	switch (factor) {
+	case VK_BLEND_FACTOR_ZERO:
+		return V_028780_BLEND_ZERO;
+	case VK_BLEND_FACTOR_ONE:
+		return V_028780_BLEND_ONE;
+	case VK_BLEND_FACTOR_SRC_COLOR:
+		return V_028780_BLEND_SRC_COLOR;
+	case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR:
+		return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
+	case VK_BLEND_FACTOR_DST_COLOR:
+		return V_028780_BLEND_DST_COLOR;
+	case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR:
+		return V_028780_BLEND_ONE_MINUS_DST_COLOR;
+	case VK_BLEND_FACTOR_SRC_ALPHA:
+		return V_028780_BLEND_SRC_ALPHA;
+	case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA:
+		return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
+	case VK_BLEND_FACTOR_DST_ALPHA:
+		return V_028780_BLEND_DST_ALPHA;
+	case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA:
+		return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
+	case VK_BLEND_FACTOR_CONSTANT_COLOR:
+		return V_028780_BLEND_CONSTANT_COLOR;
+	case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR:
+		return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR;
+	case VK_BLEND_FACTOR_CONSTANT_ALPHA:
+		return V_028780_BLEND_CONSTANT_ALPHA;
+	case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA:
+		return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA;
+	case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE:
+		return V_028780_BLEND_SRC_ALPHA_SATURATE;
+	case VK_BLEND_FACTOR_SRC1_COLOR:
+		return V_028780_BLEND_SRC1_COLOR;
+	case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
+		return V_028780_BLEND_INV_SRC1_COLOR;
+	case VK_BLEND_FACTOR_SRC1_ALPHA:
+		return V_028780_BLEND_SRC1_ALPHA;
+	case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
+		return V_028780_BLEND_INV_SRC1_ALPHA;
+	default:
+		return 0;
+	}
+}
+
+static bool is_dual_src(VkBlendFactor factor)
+{
+	switch (factor) {
+	case VK_BLEND_FACTOR_SRC1_COLOR:
+	case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
+	case VK_BLEND_FACTOR_SRC1_ALPHA:
+	case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
+		return true;
+	default:
+		return false;
+	}
+}
+
+static unsigned si_choose_spi_color_format(VkFormat vk_format,
+					    bool blend_enable,
+					    bool blend_need_alpha)
+{
+	const struct vk_format_description *desc = vk_format_description(vk_format);
+	unsigned format, ntype, swap;
+
+	/* Alpha is needed for alpha-to-coverage.
+	 * Blending may be with or without alpha.
+	 */
+	unsigned normal = 0; /* most optimal, may not support blending or export alpha */
+	unsigned alpha = 0; /* exports alpha, but may not support blending */
+	unsigned blend = 0; /* supports blending, but may not export alpha */
+	unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */
+
+	format = radv_translate_colorformat(vk_format);
+	ntype = radv_translate_color_numformat(vk_format, desc,
+					       vk_format_get_first_non_void_channel(vk_format));
+	swap = radv_translate_colorswap(vk_format, false);
+
+	/* Choose the SPI color formats. These are required values for Stoney/RB+.
+	 * Other chips have multiple choices, though they are not necessarily better.
+	 */
+	switch (format) {
+	case V_028C70_COLOR_5_6_5:
+	case V_028C70_COLOR_1_5_5_5:
+	case V_028C70_COLOR_5_5_5_1:
+	case V_028C70_COLOR_4_4_4_4:
+	case V_028C70_COLOR_10_11_11:
+	case V_028C70_COLOR_11_11_10:
+	case V_028C70_COLOR_8:
+	case V_028C70_COLOR_8_8:
+	case V_028C70_COLOR_8_8_8_8:
+	case V_028C70_COLOR_10_10_10_2:
+	case V_028C70_COLOR_2_10_10_10:
+		if (ntype == V_028C70_NUMBER_UINT)
+			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
+		else if (ntype == V_028C70_NUMBER_SINT)
+			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
+		else
+			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
+		break;
+
+	case V_028C70_COLOR_16:
+	case V_028C70_COLOR_16_16:
+	case V_028C70_COLOR_16_16_16_16:
+		if (ntype == V_028C70_NUMBER_UNORM ||
+		    ntype == V_028C70_NUMBER_SNORM) {
+			/* UNORM16 and SNORM16 don't support blending */
+			if (ntype == V_028C70_NUMBER_UNORM)
+				normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR;
+			else
+				normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR;
+
+			/* Use 32 bits per channel for blending. */
+			if (format == V_028C70_COLOR_16) {
+				if (swap == V_028C70_SWAP_STD) { /* R */
+					blend = V_028714_SPI_SHADER_32_R;
+					blend_alpha = V_028714_SPI_SHADER_32_AR;
+				} else if (swap == V_028C70_SWAP_ALT_REV) /* A */
+					blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
+				else
+					assert(0);
+			} else if (format == V_028C70_COLOR_16_16) {
+				if (swap == V_028C70_SWAP_STD) { /* RG */
+					blend = V_028714_SPI_SHADER_32_GR;
+					blend_alpha = V_028714_SPI_SHADER_32_ABGR;
+				} else if (swap == V_028C70_SWAP_ALT) /* RA */
+					blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
+				else
+					assert(0);
+			} else /* 16_16_16_16 */
+				blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
+		} else if (ntype == V_028C70_NUMBER_UINT)
+			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
+		else if (ntype == V_028C70_NUMBER_SINT)
+			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
+		else if (ntype == V_028C70_NUMBER_FLOAT)
+			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
+		else
+			assert(0);
+		break;
+
+	case V_028C70_COLOR_32:
+		if (swap == V_028C70_SWAP_STD) { /* R */
+			blend = normal = V_028714_SPI_SHADER_32_R;
+			alpha = blend_alpha = V_028714_SPI_SHADER_32_AR;
+		} else if (swap == V_028C70_SWAP_ALT_REV) /* A */
+			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
+		else
+			assert(0);
+		break;
+
+	case V_028C70_COLOR_32_32:
+		if (swap == V_028C70_SWAP_STD) { /* RG */
+			blend = normal = V_028714_SPI_SHADER_32_GR;
+			alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
+		} else if (swap == V_028C70_SWAP_ALT) /* RA */
+			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
+		else
+			assert(0);
+		break;
+
+	case V_028C70_COLOR_32_32_32_32:
+	case V_028C70_COLOR_8_24:
+	case V_028C70_COLOR_24_8:
+	case V_028C70_COLOR_X24_8_32_FLOAT:
+		alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
+		break;
+
+	default:
+		unreachable("unhandled blend format");
+	}
+
+	if (blend_enable && blend_need_alpha)
+		return blend_alpha;
+	else if(blend_need_alpha)
+		return alpha;
+	else if(blend_enable)
+		return blend;
+	else
+		return normal;
+}
+
+static unsigned si_get_cb_shader_mask(unsigned spi_shader_col_format)
+{
+	unsigned i, cb_shader_mask = 0;
+
+	for (i = 0; i < 8; i++) {
+		switch ((spi_shader_col_format >> (i * 4)) & 0xf) {
+		case V_028714_SPI_SHADER_ZERO:
+			break;
+		case V_028714_SPI_SHADER_32_R:
+			cb_shader_mask |= 0x1 << (i * 4);
+			break;
+		case V_028714_SPI_SHADER_32_GR:
+			cb_shader_mask |= 0x3 << (i * 4);
+			break;
+		case V_028714_SPI_SHADER_32_AR:
+			cb_shader_mask |= 0x9 << (i * 4);
+			break;
+		case V_028714_SPI_SHADER_FP16_ABGR:
+		case V_028714_SPI_SHADER_UNORM16_ABGR:
+		case V_028714_SPI_SHADER_SNORM16_ABGR:
+		case V_028714_SPI_SHADER_UINT16_ABGR:
+		case V_028714_SPI_SHADER_SINT16_ABGR:
+		case V_028714_SPI_SHADER_32_ABGR:
+			cb_shader_mask |= 0xf << (i * 4);
+			break;
+		default:
+			assert(0);
+		}
+	}
+	return cb_shader_mask;
+}
+
+static void
+radv_pipeline_compute_spi_color_formats(struct radv_pipeline *pipeline,
+					const VkGraphicsPipelineCreateInfo *pCreateInfo,
+					uint32_t blend_enable,
+					uint32_t blend_need_alpha,
+					bool single_cb_enable)
+{
+	RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
+	struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
+	struct radv_blend_state *blend = &pipeline->graphics.blend;
+	unsigned col_format = 0;
+
+	for (unsigned i = 0; i < (single_cb_enable ? 1 : subpass->color_count); ++i) {
+		struct radv_render_pass_attachment *attachment;
+		unsigned cf;
+
+		attachment = pass->attachments + subpass->color_attachments[i].attachment;
+
+		cf = si_choose_spi_color_format(attachment->format,
+						blend_enable & (1 << i),
+						blend_need_alpha & (1 << i));
+
+		col_format |= cf << (4 * i);
+	}
+
+	blend->cb_shader_mask = si_get_cb_shader_mask(col_format);
+
+	if (!col_format)
+		col_format |= V_028714_SPI_SHADER_32_R;
+	blend->spi_shader_col_format = col_format;
+}
+
+static bool
+format_is_int8(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+	int channel =  vk_format_get_first_non_void_channel(format);
+
+	return channel >= 0 && desc->channel[channel].pure_integer &&
+	       desc->channel[channel].size == 8;
+}
+
+unsigned radv_format_meta_fs_key(VkFormat format)
+{
+	unsigned col_format = si_choose_spi_color_format(format, false, false) - 1;
+	bool is_int8 = format_is_int8(format);
+
+	return col_format + (is_int8 ? 3 : 0);
+}
+
+static unsigned
+radv_pipeline_compute_is_int8(const VkGraphicsPipelineCreateInfo *pCreateInfo)
+{
+	RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
+	struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
+	unsigned is_int8 = 0;
+
+	for (unsigned i = 0; i < subpass->color_count; ++i) {
+		struct radv_render_pass_attachment *attachment;
+
+		attachment = pass->attachments + subpass->color_attachments[i].attachment;
+
+		if (format_is_int8(attachment->format))
+			is_int8 |= 1 << i;
+	}
+
+	return is_int8;
+}
+
+static void
+radv_pipeline_init_blend_state(struct radv_pipeline *pipeline,
+			       const VkGraphicsPipelineCreateInfo *pCreateInfo,
+			       const struct radv_graphics_pipeline_create_info *extra)
+{
+	const VkPipelineColorBlendStateCreateInfo *vkblend = pCreateInfo->pColorBlendState;
+	struct radv_blend_state *blend = &pipeline->graphics.blend;
+	unsigned mode = V_028808_CB_NORMAL;
+	uint32_t blend_enable = 0, blend_need_alpha = 0;
+	int i;
+	bool single_cb_enable = false;
+	if (extra && extra->custom_blend_mode) {
+		single_cb_enable = true;
+		mode = extra->custom_blend_mode;
+	}
+	blend->cb_color_control = 0;
+	if (vkblend->logicOpEnable)
+		blend->cb_color_control |= S_028808_ROP3(vkblend->logicOp | (vkblend->logicOp << 4));
+	else
+		blend->cb_color_control |= S_028808_ROP3(0xcc);
+
+	blend->db_alpha_to_mask = S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
+		S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
+		S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
+		S_028B70_ALPHA_TO_MASK_OFFSET3(2);
+
+	blend->cb_target_mask = 0;
+	for (i = 0; i < vkblend->attachmentCount; i++) {
+		const VkPipelineColorBlendAttachmentState *att = &vkblend->pAttachments[i];
+		unsigned blend_cntl = 0;
+		VkBlendOp eqRGB = att->colorBlendOp;
+		VkBlendFactor srcRGB = att->srcColorBlendFactor;
+		VkBlendFactor dstRGB = att->dstColorBlendFactor;
+		VkBlendOp eqA = att->alphaBlendOp;
+		VkBlendFactor srcA = att->srcAlphaBlendFactor;
+		VkBlendFactor dstA = att->dstAlphaBlendFactor;
+
+		blend->sx_mrt0_blend_opt[i] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
+
+		if (!att->colorWriteMask)
+			continue;
+
+		blend->cb_target_mask |= (unsigned)att->colorWriteMask << (4 * i);
+		if (!att->blendEnable) {
+			blend->cb_blend_control[i] = blend_cntl;
+			continue;
+		}
+
+		if (is_dual_src(srcRGB) || is_dual_src(dstRGB) || is_dual_src(srcA) || is_dual_src(dstA))
+			radv_finishme("dual source blending");
+		if (eqRGB == VK_BLEND_OP_MIN || eqRGB == VK_BLEND_OP_MAX) {
+			srcRGB = VK_BLEND_FACTOR_ONE;
+			dstRGB = VK_BLEND_FACTOR_ONE;
+		}
+		if (eqA == VK_BLEND_OP_MIN || eqA == VK_BLEND_OP_MAX) {
+			srcA = VK_BLEND_FACTOR_ONE;
+			dstA = VK_BLEND_FACTOR_ONE;
+		}
+
+		blend_cntl |= S_028780_ENABLE(1);
+
+		blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
+		blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB));
+		blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB));
+		if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
+			blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
+			blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
+			blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA));
+			blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA));
+		}
+		blend->cb_blend_control[i] = blend_cntl;
+
+		blend_enable |= 1 << i;
+
+		if (srcRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
+		    dstRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
+		    srcRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
+		    dstRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
+		    srcRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA ||
+		    dstRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA)
+			blend_need_alpha |= 1 << i;
+	}
+	for (i = vkblend->attachmentCount; i < 8; i++)
+		blend->cb_blend_control[i] = 0;
+
+	if (blend->cb_target_mask)
+		blend->cb_color_control |= S_028808_MODE(mode);
+	else
+		blend->cb_color_control |= S_028808_MODE(V_028808_CB_DISABLE);
+
+	radv_pipeline_compute_spi_color_formats(pipeline, pCreateInfo,
+						blend_enable, blend_need_alpha, single_cb_enable);
+}
+
+static uint32_t si_translate_stencil_op(enum VkStencilOp op)
+{
+	switch (op) {
+	case VK_STENCIL_OP_KEEP:
+		return V_02842C_STENCIL_KEEP;
+	case VK_STENCIL_OP_ZERO:
+		return V_02842C_STENCIL_ZERO;
+	case VK_STENCIL_OP_REPLACE:
+		return V_02842C_STENCIL_REPLACE_TEST;
+	case VK_STENCIL_OP_INCREMENT_AND_CLAMP:
+		return V_02842C_STENCIL_ADD_CLAMP;
+	case VK_STENCIL_OP_DECREMENT_AND_CLAMP:
+		return V_02842C_STENCIL_SUB_CLAMP;
+	case VK_STENCIL_OP_INVERT:
+		return V_02842C_STENCIL_INVERT;
+	case VK_STENCIL_OP_INCREMENT_AND_WRAP:
+		return V_02842C_STENCIL_ADD_WRAP;
+	case VK_STENCIL_OP_DECREMENT_AND_WRAP:
+		return V_02842C_STENCIL_SUB_WRAP;
+	default:
+		return 0;
+	}
+}
+static void
+radv_pipeline_init_depth_stencil_state(struct radv_pipeline *pipeline,
+				       const VkGraphicsPipelineCreateInfo *pCreateInfo,
+				       const struct radv_graphics_pipeline_create_info *extra)
+{
+	const VkPipelineDepthStencilStateCreateInfo *vkds = pCreateInfo->pDepthStencilState;
+	struct radv_depth_stencil_state *ds = &pipeline->graphics.ds;
+
+	memset(ds, 0, sizeof(*ds));
+	if (!vkds)
+		return;
+	ds->db_depth_control = S_028800_Z_ENABLE(vkds->depthTestEnable ? 1 : 0) |
+		S_028800_Z_WRITE_ENABLE(vkds->depthWriteEnable ? 1 : 0) |
+		S_028800_ZFUNC(vkds->depthCompareOp) |
+		S_028800_DEPTH_BOUNDS_ENABLE(vkds->depthBoundsTestEnable ? 1 : 0);
+
+	if (vkds->stencilTestEnable) {
+		ds->db_depth_control |= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
+		ds->db_depth_control |= S_028800_STENCILFUNC(vkds->front.compareOp);
+		ds->db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds->front.failOp));
+		ds->db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds->front.passOp));
+		ds->db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds->front.depthFailOp));
+
+		ds->db_depth_control |= S_028800_STENCILFUNC_BF(vkds->back.compareOp);
+		ds->db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds->back.failOp));
+		ds->db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds->back.passOp));
+		ds->db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds->back.depthFailOp));
+	}
+
+	if (extra) {
+
+		ds->db_render_control |= S_028000_DEPTH_CLEAR_ENABLE(extra->db_depth_clear);
+		ds->db_render_control |= S_028000_STENCIL_CLEAR_ENABLE(extra->db_stencil_clear);
+
+		ds->db_render_control |= S_028000_RESUMMARIZE_ENABLE(extra->db_resummarize);
+		ds->db_render_control |= S_028000_DEPTH_COMPRESS_DISABLE(extra->db_flush_depth_inplace);
+		ds->db_render_control |= S_028000_STENCIL_COMPRESS_DISABLE(extra->db_flush_stencil_inplace);
+		ds->db_render_override2 |= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra->db_depth_disable_expclear);
+		ds->db_render_override2 |= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra->db_stencil_disable_expclear);
+	}
+}
+
+static uint32_t si_translate_fill(VkPolygonMode func)
+{
+	switch(func) {
+	case VK_POLYGON_MODE_FILL:
+		return V_028814_X_DRAW_TRIANGLES;
+	case VK_POLYGON_MODE_LINE:
+		return V_028814_X_DRAW_LINES;
+	case VK_POLYGON_MODE_POINT:
+		return V_028814_X_DRAW_POINTS;
+	default:
+		assert(0);
+		return V_028814_X_DRAW_POINTS;
+	}
+}
+static void
+radv_pipeline_init_raster_state(struct radv_pipeline *pipeline,
+				const VkGraphicsPipelineCreateInfo *pCreateInfo)
+{
+	const VkPipelineRasterizationStateCreateInfo *vkraster = pCreateInfo->pRasterizationState;
+	struct radv_raster_state *raster = &pipeline->graphics.raster;
+
+	memset(raster, 0, sizeof(*raster));
+
+	raster->spi_interp_control =
+		S_0286D4_FLAT_SHADE_ENA(1) |
+		S_0286D4_PNT_SPRITE_ENA(1) |
+		S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
+		S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
+		S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
+		S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
+		S_0286D4_PNT_SPRITE_TOP_1(0); // vulkan is top to bottom - 1.0 at bottom
+
+	raster->pa_cl_vs_out_cntl = S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(1);
+	raster->pa_cl_clip_cntl = S_028810_PS_UCP_MODE(3) |
+		S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
+		S_028810_ZCLIP_NEAR_DISABLE(vkraster->depthClampEnable ? 1 : 0) |
+		S_028810_ZCLIP_FAR_DISABLE(vkraster->depthClampEnable ? 1 : 0) |
+		S_028810_DX_RASTERIZATION_KILL(vkraster->rasterizerDiscardEnable ? 1 : 0) |
+		S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
+
+	raster->pa_su_vtx_cntl =
+		S_028BE4_PIX_CENTER(1) | // TODO verify
+		S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN) |
+		S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH);
+
+	raster->pa_su_sc_mode_cntl =
+		S_028814_FACE(vkraster->frontFace) |
+		S_028814_CULL_FRONT(!!(vkraster->cullMode & VK_CULL_MODE_FRONT_BIT)) |
+		S_028814_CULL_BACK(!!(vkraster->cullMode & VK_CULL_MODE_BACK_BIT)) |
+		S_028814_POLY_MODE(vkraster->polygonMode != VK_POLYGON_MODE_FILL) |
+		S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster->polygonMode)) |
+		S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster->polygonMode)) |
+		S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
+		S_028814_POLY_OFFSET_BACK_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
+		S_028814_POLY_OFFSET_PARA_ENABLE(vkraster->depthBiasEnable ? 1 : 0);
+
+}
+
+static void
+radv_pipeline_init_multisample_state(struct radv_pipeline *pipeline,
+				     const VkGraphicsPipelineCreateInfo *pCreateInfo)
+{
+	const VkPipelineMultisampleStateCreateInfo *vkms = pCreateInfo->pMultisampleState;
+	struct radv_blend_state *blend = &pipeline->graphics.blend;
+	struct radv_multisample_state *ms = &pipeline->graphics.ms;
+	unsigned num_tile_pipes = pipeline->device->instance->physicalDevice.rad_info.num_tile_pipes;
+	int ps_iter_samples = 1;
+	uint32_t mask = 0xffff;
+
+	ms->num_samples = vkms->rasterizationSamples;
+	ms->pa_sc_line_cntl = S_028BDC_DX10_DIAMOND_TEST_ENA(1);
+	ms->pa_sc_aa_config = 0;
+	ms->db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
+		S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
+	ms->pa_sc_mode_cntl_1 =
+		S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
+		S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
+		/* always 1: */
+		S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
+		S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
+		S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
+		S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
+		EG_S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
+		EG_S_028A4C_FORCE_EOV_REZ_ENABLE(1);
+
+	if (vkms->rasterizationSamples > 1) {
+		unsigned log_samples = util_logbase2(vkms->rasterizationSamples);
+		unsigned log_ps_iter_samples = util_logbase2(util_next_power_of_two(ps_iter_samples));
+		ms->pa_sc_mode_cntl_0 = S_028A48_MSAA_ENABLE(1);
+		ms->pa_sc_line_cntl |= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */
+		ms->db_eqaa |= S_028804_MAX_ANCHOR_SAMPLES(log_samples) |
+			S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
+			S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
+			S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
+		ms->pa_sc_aa_config |= S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
+			S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples)) |
+			S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples); /* CM_R_028BE0_PA_SC_AA_CONFIG */
+		ms->pa_sc_mode_cntl_1 |= EG_S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1);
+	}
+
+	if (vkms->alphaToCoverageEnable)
+		blend->db_alpha_to_mask |= S_028B70_ALPHA_TO_MASK_ENABLE(1);
+
+	if (vkms->pSampleMask) {
+		mask = vkms->pSampleMask[0] & 0xffff;
+	}
+
+	ms->pa_sc_aa_mask[0] = mask | (mask << 16);
+	ms->pa_sc_aa_mask[1] = mask | (mask << 16);
+}
+
+static uint32_t
+si_translate_prim(enum VkPrimitiveTopology topology)
+{
+	switch (topology) {
+	case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
+		return V_008958_DI_PT_POINTLIST;
+	case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
+		return V_008958_DI_PT_LINELIST;
+	case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
+		return V_008958_DI_PT_LINESTRIP;
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
+		return V_008958_DI_PT_TRILIST;
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
+		return V_008958_DI_PT_TRISTRIP;
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
+		return V_008958_DI_PT_TRIFAN;
+	case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
+		return V_008958_DI_PT_LINELIST_ADJ;
+	case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
+		return V_008958_DI_PT_LINESTRIP_ADJ;
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
+		return V_008958_DI_PT_TRILIST_ADJ;
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
+		return V_008958_DI_PT_TRISTRIP_ADJ;
+	case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
+		return V_008958_DI_PT_PATCH;
+	default:
+		assert(0);
+		return 0;
+	}
+}
+
+static uint32_t
+si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology)
+{
+	switch (topology) {
+	case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
+	case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
+		return V_028A6C_OUTPRIM_TYPE_POINTLIST;
+	case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
+	case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
+	case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
+	case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
+		return V_028A6C_OUTPRIM_TYPE_LINESTRIP;
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
+	case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
+		return V_028A6C_OUTPRIM_TYPE_TRISTRIP;
+	default:
+		assert(0);
+		return 0;
+	}
+}
+
+static unsigned si_map_swizzle(unsigned swizzle)
+{
+	switch (swizzle) {
+	case VK_SWIZZLE_Y:
+		return V_008F0C_SQ_SEL_Y;
+	case VK_SWIZZLE_Z:
+		return V_008F0C_SQ_SEL_Z;
+	case VK_SWIZZLE_W:
+		return V_008F0C_SQ_SEL_W;
+	case VK_SWIZZLE_0:
+		return V_008F0C_SQ_SEL_0;
+	case VK_SWIZZLE_1:
+		return V_008F0C_SQ_SEL_1;
+	default: /* VK_SWIZZLE_X */
+		return V_008F0C_SQ_SEL_X;
+	}
+}
+
+static void
+radv_pipeline_init_dynamic_state(struct radv_pipeline *pipeline,
+				 const VkGraphicsPipelineCreateInfo *pCreateInfo)
+{
+	radv_cmd_dirty_mask_t states = RADV_CMD_DIRTY_DYNAMIC_ALL;
+	RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
+	struct radv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
+
+	pipeline->dynamic_state = default_dynamic_state;
+
+	if (pCreateInfo->pDynamicState) {
+		/* Remove all of the states that are marked as dynamic */
+		uint32_t count = pCreateInfo->pDynamicState->dynamicStateCount;
+		for (uint32_t s = 0; s < count; s++)
+			states &= ~(1 << pCreateInfo->pDynamicState->pDynamicStates[s]);
+	}
+
+	struct radv_dynamic_state *dynamic = &pipeline->dynamic_state;
+
+	dynamic->viewport.count = pCreateInfo->pViewportState->viewportCount;
+	if (states & (1 << VK_DYNAMIC_STATE_VIEWPORT)) {
+		typed_memcpy(dynamic->viewport.viewports,
+			     pCreateInfo->pViewportState->pViewports,
+			     pCreateInfo->pViewportState->viewportCount);
+	}
+
+	dynamic->scissor.count = pCreateInfo->pViewportState->scissorCount;
+	if (states & (1 << VK_DYNAMIC_STATE_SCISSOR)) {
+		typed_memcpy(dynamic->scissor.scissors,
+			     pCreateInfo->pViewportState->pScissors,
+			     pCreateInfo->pViewportState->scissorCount);
+	}
+
+	if (states & (1 << VK_DYNAMIC_STATE_LINE_WIDTH)) {
+		assert(pCreateInfo->pRasterizationState);
+		dynamic->line_width = pCreateInfo->pRasterizationState->lineWidth;
+	}
+
+	if (states & (1 << VK_DYNAMIC_STATE_DEPTH_BIAS)) {
+		assert(pCreateInfo->pRasterizationState);
+		dynamic->depth_bias.bias =
+			pCreateInfo->pRasterizationState->depthBiasConstantFactor;
+		dynamic->depth_bias.clamp =
+			pCreateInfo->pRasterizationState->depthBiasClamp;
+		dynamic->depth_bias.slope =
+			pCreateInfo->pRasterizationState->depthBiasSlopeFactor;
+	}
+
+	if (states & (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS)) {
+		assert(pCreateInfo->pColorBlendState);
+		typed_memcpy(dynamic->blend_constants,
+			     pCreateInfo->pColorBlendState->blendConstants, 4);
+	}
+
+	/* If there is no depthstencil attachment, then don't read
+	 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
+	 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
+	 * no need to override the depthstencil defaults in
+	 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
+	 *
+	 * From the Vulkan spec (20 Oct 2015, git-aa308cb):
+	 *
+	 *    pDepthStencilState [...] may only be NULL if renderPass and subpass
+	 *    specify a subpass that has no depth/stencil attachment.
+	 */
+	if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
+		if (states & (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS)) {
+			assert(pCreateInfo->pDepthStencilState);
+			dynamic->depth_bounds.min =
+				pCreateInfo->pDepthStencilState->minDepthBounds;
+			dynamic->depth_bounds.max =
+				pCreateInfo->pDepthStencilState->maxDepthBounds;
+		}
+
+		if (states & (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK)) {
+			assert(pCreateInfo->pDepthStencilState);
+			dynamic->stencil_compare_mask.front =
+				pCreateInfo->pDepthStencilState->front.compareMask;
+			dynamic->stencil_compare_mask.back =
+				pCreateInfo->pDepthStencilState->back.compareMask;
+		}
+
+		if (states & (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK)) {
+			assert(pCreateInfo->pDepthStencilState);
+			dynamic->stencil_write_mask.front =
+				pCreateInfo->pDepthStencilState->front.writeMask;
+			dynamic->stencil_write_mask.back =
+				pCreateInfo->pDepthStencilState->back.writeMask;
+		}
+
+		if (states & (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE)) {
+			assert(pCreateInfo->pDepthStencilState);
+			dynamic->stencil_reference.front =
+				pCreateInfo->pDepthStencilState->front.reference;
+			dynamic->stencil_reference.back =
+				pCreateInfo->pDepthStencilState->back.reference;
+		}
+	}
+
+	pipeline->dynamic_state_mask = states;
+}
+
+static union ac_shader_variant_key
+radv_compute_vs_key(const VkGraphicsPipelineCreateInfo *pCreateInfo)
+{
+	union ac_shader_variant_key key;
+	const VkPipelineVertexInputStateCreateInfo *input_state =
+	                                         pCreateInfo->pVertexInputState;
+
+	memset(&key, 0, sizeof(key));
+	key.vs.instance_rate_inputs = 0;
+
+	for (unsigned i = 0; i < input_state->vertexAttributeDescriptionCount; ++i) {
+		unsigned binding;
+		binding = input_state->pVertexAttributeDescriptions[i].binding;
+		if (input_state->pVertexBindingDescriptions[binding].inputRate)
+			key.vs.instance_rate_inputs |= 1u << input_state->pVertexAttributeDescriptions[i].location;
+	}
+	return key;
+}
+
+VkResult
+radv_pipeline_init(struct radv_pipeline *pipeline,
+		   struct radv_device *device,
+		   struct radv_pipeline_cache *cache,
+		   const VkGraphicsPipelineCreateInfo *pCreateInfo,
+		   const struct radv_graphics_pipeline_create_info *extra,
+		   const VkAllocationCallbacks *alloc)
+{
+	struct radv_shader_module fs_m = {0};
+
+	bool dump = getenv("RADV_DUMP_SHADERS");
+	if (alloc == NULL)
+		alloc = &device->alloc;
+
+	pipeline->device = device;
+	pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
+
+	radv_pipeline_init_dynamic_state(pipeline, pCreateInfo);
+	const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
+	struct radv_shader_module *modules[MESA_SHADER_STAGES] = { 0, };
+	for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
+		gl_shader_stage stage = ffs(pCreateInfo->pStages[i].stage) - 1;
+		pStages[stage] = &pCreateInfo->pStages[i];
+		modules[stage] = radv_shader_module_from_handle(pStages[stage]->module);
+	}
+
+	radv_pipeline_init_blend_state(pipeline, pCreateInfo, extra);
+
+	/* */
+	if (modules[MESA_SHADER_VERTEX]) {
+		union ac_shader_variant_key key = radv_compute_vs_key(pCreateInfo);
+
+		pipeline->shaders[MESA_SHADER_VERTEX] =
+			 radv_pipeline_compile(pipeline, cache, modules[MESA_SHADER_VERTEX],
+					       pStages[MESA_SHADER_VERTEX]->pName,
+					       MESA_SHADER_VERTEX,
+					       pStages[MESA_SHADER_VERTEX]->pSpecializationInfo,
+					       pipeline->layout, &key, dump);
+
+		pipeline->active_stages |= mesa_to_vk_shader_stage(MESA_SHADER_VERTEX);
+	}
+
+	if (!modules[MESA_SHADER_FRAGMENT]) {
+		nir_builder fs_b;
+		nir_builder_init_simple_shader(&fs_b, NULL, MESA_SHADER_FRAGMENT, NULL);
+		fs_b.shader->info.name = ralloc_strdup(fs_b.shader, "noop_fs");
+		fs_m.nir = fs_b.shader;
+		modules[MESA_SHADER_FRAGMENT] = &fs_m;
+	}
+
+	if (modules[MESA_SHADER_FRAGMENT]) {
+		union ac_shader_variant_key key;
+		key.fs.col_format = pipeline->graphics.blend.spi_shader_col_format;
+		key.fs.is_int8 = radv_pipeline_compute_is_int8(pCreateInfo);
+
+		const VkPipelineShaderStageCreateInfo *stage = pStages[MESA_SHADER_FRAGMENT];
+
+		pipeline->shaders[MESA_SHADER_FRAGMENT] =
+			 radv_pipeline_compile(pipeline, cache, modules[MESA_SHADER_FRAGMENT],
+					       stage ? stage->pName : "main",
+					       MESA_SHADER_FRAGMENT,
+					       stage ? stage->pSpecializationInfo : NULL,
+					       pipeline->layout, &key, dump);
+		pipeline->active_stages |= mesa_to_vk_shader_stage(MESA_SHADER_FRAGMENT);
+	}
+
+	if (fs_m.nir)
+		ralloc_free(fs_m.nir);
+
+	radv_pipeline_init_depth_stencil_state(pipeline, pCreateInfo, extra);
+	radv_pipeline_init_raster_state(pipeline, pCreateInfo);
+	radv_pipeline_init_multisample_state(pipeline, pCreateInfo);
+	pipeline->graphics.prim = si_translate_prim(pCreateInfo->pInputAssemblyState->topology);
+	pipeline->graphics.gs_out = si_conv_prim_to_gs_out(pCreateInfo->pInputAssemblyState->topology);
+	if (extra && extra->use_rectlist) {
+		pipeline->graphics.prim = V_008958_DI_PT_RECTLIST;
+		pipeline->graphics.gs_out = V_028A6C_OUTPRIM_TYPE_TRISTRIP;
+	}
+	pipeline->graphics.prim_restart_enable = !!pCreateInfo->pInputAssemblyState->primitiveRestartEnable;
+
+	const VkPipelineVertexInputStateCreateInfo *vi_info =
+		pCreateInfo->pVertexInputState;
+	for (uint32_t i = 0; i < vi_info->vertexAttributeDescriptionCount; i++) {
+		const VkVertexInputAttributeDescription *desc =
+			&vi_info->pVertexAttributeDescriptions[i];
+		unsigned loc = desc->location;
+		const struct vk_format_description *format_desc;
+		int first_non_void;
+		uint32_t num_format, data_format;
+		format_desc = vk_format_description(desc->format);
+		first_non_void = vk_format_get_first_non_void_channel(desc->format);
+
+		num_format = radv_translate_buffer_numformat(format_desc, first_non_void);
+		data_format = radv_translate_buffer_dataformat(format_desc, first_non_void);
+
+		pipeline->va_rsrc_word3[loc] = S_008F0C_DST_SEL_X(si_map_swizzle(format_desc->swizzle[0])) |
+			S_008F0C_DST_SEL_Y(si_map_swizzle(format_desc->swizzle[1])) |
+			S_008F0C_DST_SEL_Z(si_map_swizzle(format_desc->swizzle[2])) |
+			S_008F0C_DST_SEL_W(si_map_swizzle(format_desc->swizzle[3])) |
+			S_008F0C_NUM_FORMAT(num_format) |
+			S_008F0C_DATA_FORMAT(data_format);
+		pipeline->va_format_size[loc] = format_desc->block.bits / 8;
+		pipeline->va_offset[loc] = desc->offset;
+		pipeline->va_binding[loc] = desc->binding;
+		pipeline->num_vertex_attribs = MAX2(pipeline->num_vertex_attribs, loc + 1);
+	}
+
+	for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
+		const VkVertexInputBindingDescription *desc =
+			&vi_info->pVertexBindingDescriptions[i];
+
+		pipeline->binding_stride[desc->binding] = desc->stride;
+	}
+
+	return VK_SUCCESS;
+}
+
+VkResult
+radv_graphics_pipeline_create(
+	VkDevice _device,
+	VkPipelineCache _cache,
+	const VkGraphicsPipelineCreateInfo *pCreateInfo,
+	const struct radv_graphics_pipeline_create_info *extra,
+	const VkAllocationCallbacks *pAllocator,
+	VkPipeline *pPipeline)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
+	struct radv_pipeline *pipeline;
+	VkResult result;
+
+	pipeline = radv_alloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
+			       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (pipeline == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	memset(pipeline, 0, sizeof(*pipeline));
+	result = radv_pipeline_init(pipeline, device, cache,
+				    pCreateInfo, extra, pAllocator);
+	if (result != VK_SUCCESS) {
+		radv_free2(&device->alloc, pAllocator, pipeline);
+		return result;
+	}
+
+	*pPipeline = radv_pipeline_to_handle(pipeline);
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_CreateGraphicsPipelines(
+	VkDevice                                    _device,
+	VkPipelineCache                             pipelineCache,
+	uint32_t                                    count,
+	const VkGraphicsPipelineCreateInfo*         pCreateInfos,
+	const VkAllocationCallbacks*                pAllocator,
+	VkPipeline*                                 pPipelines)
+{
+	VkResult result = VK_SUCCESS;
+	unsigned i = 0;
+
+	for (; i < count; i++) {
+		result = radv_graphics_pipeline_create(_device,
+						       pipelineCache,
+						       &pCreateInfos[i],
+						       NULL, pAllocator, &pPipelines[i]);
+		if (result != VK_SUCCESS) {
+			for (unsigned j = 0; j < i; j++) {
+				radv_DestroyPipeline(_device, pPipelines[j], pAllocator);
+			}
+
+			return result;
+		}
+	}
+
+	return VK_SUCCESS;
+}
+
+static VkResult radv_compute_pipeline_create(
+	VkDevice                                    _device,
+	VkPipelineCache                             _cache,
+	const VkComputePipelineCreateInfo*          pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkPipeline*                                 pPipeline)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
+	RADV_FROM_HANDLE(radv_shader_module, module, pCreateInfo->stage.module);
+	struct radv_pipeline *pipeline;
+	bool dump = getenv("RADV_DUMP_SHADERS");
+
+	pipeline = radv_alloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
+			       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (pipeline == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	memset(pipeline, 0, sizeof(*pipeline));
+	pipeline->device = device;
+	pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
+
+	pipeline->shaders[MESA_SHADER_COMPUTE] =
+		 radv_pipeline_compile(pipeline, cache, module,
+				       pCreateInfo->stage.pName,
+				       MESA_SHADER_COMPUTE,
+				       pCreateInfo->stage.pSpecializationInfo,
+				       pipeline->layout, NULL, dump);
+
+	*pPipeline = radv_pipeline_to_handle(pipeline);
+	return VK_SUCCESS;
+}
+VkResult radv_CreateComputePipelines(
+	VkDevice                                    _device,
+	VkPipelineCache                             pipelineCache,
+	uint32_t                                    count,
+	const VkComputePipelineCreateInfo*          pCreateInfos,
+	const VkAllocationCallbacks*                pAllocator,
+	VkPipeline*                                 pPipelines)
+{
+	VkResult result = VK_SUCCESS;
+
+	unsigned i = 0;
+	for (; i < count; i++) {
+		result = radv_compute_pipeline_create(_device, pipelineCache,
+						      &pCreateInfos[i],
+						      pAllocator, &pPipelines[i]);
+		if (result != VK_SUCCESS) {
+			for (unsigned j = 0; j < i; j++) {
+				radv_DestroyPipeline(_device, pPipelines[j], pAllocator);
+			}
+
+			return result;
+		}
+	}
+
+	return VK_SUCCESS;
+}
diff --git a/src/amd/vulkan/radv_pipeline_cache.c b/src/amd/vulkan/radv_pipeline_cache.c
new file mode 100644
index 0000000000..032a7e4604
--- /dev/null
+++ b/src/amd/vulkan/radv_pipeline_cache.c
@@ -0,0 +1,475 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "util/mesa-sha1.h"
+#include "util/debug.h"
+#include "radv_private.h"
+
+#include "ac_nir_to_llvm.h"
+
+struct cache_entry {
+	unsigned char sha1[20];
+	uint32_t code_size;
+	struct ac_shader_variant_info variant_info;
+	struct ac_shader_config config;
+	uint32_t rsrc1, rsrc2;
+	struct radv_shader_variant *variant;
+	uint32_t code[0];
+};
+
+void
+radv_pipeline_cache_init(struct radv_pipeline_cache *cache,
+			 struct radv_device *device)
+{
+	cache->device = device;
+	pthread_mutex_init(&cache->mutex, NULL);
+
+	cache->modified = false;
+	cache->kernel_count = 0;
+	cache->total_size = 0;
+	cache->table_size = 1024;
+	const size_t byte_size = cache->table_size * sizeof(cache->hash_table[0]);
+	cache->hash_table = malloc(byte_size);
+
+	/* We don't consider allocation failure fatal, we just start with a 0-sized
+	 * cache. */
+	if (cache->hash_table == NULL ||
+	    !env_var_as_boolean("RADV_ENABLE_PIPELINE_CACHE", true))
+		cache->table_size = 0;
+	else
+		memset(cache->hash_table, 0, byte_size);
+}
+
+void
+radv_pipeline_cache_finish(struct radv_pipeline_cache *cache)
+{
+	for (unsigned i = 0; i < cache->table_size; ++i)
+		if (cache->hash_table[i]) {
+			if (cache->hash_table[i]->variant)
+				radv_shader_variant_destroy(cache->device,
+							    cache->hash_table[i]->variant);
+			radv_free(&cache->alloc, cache->hash_table[i]);
+		}
+	pthread_mutex_destroy(&cache->mutex);
+	free(cache->hash_table);
+}
+
+static uint32_t
+entry_size(struct cache_entry *entry)
+{
+	return sizeof(*entry) + entry->code_size;
+}
+
+void
+radv_hash_shader(unsigned char *hash, struct radv_shader_module *module,
+		 const char *entrypoint,
+		 const VkSpecializationInfo *spec_info,
+		 const struct radv_pipeline_layout *layout,
+		 const union ac_shader_variant_key *key)
+{
+	struct mesa_sha1 *ctx;
+
+	ctx = _mesa_sha1_init();
+	if (key)
+		_mesa_sha1_update(ctx, key, sizeof(*key));
+	_mesa_sha1_update(ctx, module->sha1, sizeof(module->sha1));
+	_mesa_sha1_update(ctx, entrypoint, strlen(entrypoint));
+	if (layout)
+		_mesa_sha1_update(ctx, layout->sha1, sizeof(layout->sha1));
+	if (spec_info) {
+		_mesa_sha1_update(ctx, spec_info->pMapEntries,
+				  spec_info->mapEntryCount * sizeof spec_info->pMapEntries[0]);
+		_mesa_sha1_update(ctx, spec_info->pData, spec_info->dataSize);
+	}
+	_mesa_sha1_final(ctx, hash);
+}
+
+
+static struct cache_entry *
+radv_pipeline_cache_search_unlocked(struct radv_pipeline_cache *cache,
+				    const unsigned char *sha1)
+{
+	const uint32_t mask = cache->table_size - 1;
+	const uint32_t start = (*(uint32_t *) sha1);
+
+	for (uint32_t i = 0; i < cache->table_size; i++) {
+		const uint32_t index = (start + i) & mask;
+		struct cache_entry *entry = cache->hash_table[index];
+
+		if (!entry)
+			return NULL;
+
+		if (memcmp(entry->sha1, sha1, sizeof(entry->sha1)) == 0) {
+			return entry;
+		}
+	}
+
+	unreachable("hash table should never be full");
+}
+
+static struct cache_entry *
+radv_pipeline_cache_search(struct radv_pipeline_cache *cache,
+			   const unsigned char *sha1)
+{
+	struct cache_entry *entry;
+
+	pthread_mutex_lock(&cache->mutex);
+
+	entry = radv_pipeline_cache_search_unlocked(cache, sha1);
+
+	pthread_mutex_unlock(&cache->mutex);
+
+	return entry;
+}
+
+struct radv_shader_variant *
+radv_create_shader_variant_from_pipeline_cache(struct radv_device *device,
+					       struct radv_pipeline_cache *cache,
+					       const unsigned char *sha1)
+{
+	struct cache_entry *entry = radv_pipeline_cache_search(cache, sha1);
+
+	if (!entry)
+		return NULL;
+
+	if (!entry->variant) {
+		struct radv_shader_variant *variant;
+
+		variant = calloc(1, sizeof(struct radv_shader_variant));
+		if (!variant)
+			return NULL;
+
+		variant->config = entry->config;
+		variant->info = entry->variant_info;
+		variant->rsrc1 = entry->rsrc1;
+		variant->rsrc2 = entry->rsrc2;
+		variant->ref_count = 1;
+
+		variant->bo = device->ws->buffer_create(device->ws, entry->code_size, 256,
+						RADEON_DOMAIN_GTT, RADEON_FLAG_CPU_ACCESS);
+
+		void *ptr = device->ws->buffer_map(variant->bo);
+		memcpy(ptr, entry->code, entry->code_size);
+		device->ws->buffer_unmap(variant->bo);
+
+		entry->variant = variant;
+	}
+
+	__sync_fetch_and_add(&entry->variant->ref_count, 1);
+	return entry->variant;
+}
+
+
+static void
+radv_pipeline_cache_set_entry(struct radv_pipeline_cache *cache,
+			      struct cache_entry *entry)
+{
+	const uint32_t mask = cache->table_size - 1;
+	const uint32_t start = (*(uint32_t *) entry->sha1);
+
+	/* We'll always be able to insert when we get here. */
+	assert(cache->kernel_count < cache->table_size / 2);
+
+	for (uint32_t i = 0; i < cache->table_size; i++) {
+		const uint32_t index = (start + i) & mask;
+		if (!cache->hash_table[index]) {
+			cache->hash_table[index] = entry;
+			break;
+		}
+	}
+
+	cache->total_size += entry_size(entry);
+	cache->kernel_count++;
+}
+
+
+static VkResult
+radv_pipeline_cache_grow(struct radv_pipeline_cache *cache)
+{
+	const uint32_t table_size = cache->table_size * 2;
+	const uint32_t old_table_size = cache->table_size;
+	const size_t byte_size = table_size * sizeof(cache->hash_table[0]);
+	struct cache_entry **table;
+	struct cache_entry **old_table = cache->hash_table;
+
+	table = malloc(byte_size);
+	if (table == NULL)
+		return VK_ERROR_OUT_OF_HOST_MEMORY;
+
+	cache->hash_table = table;
+	cache->table_size = table_size;
+	cache->kernel_count = 0;
+	cache->total_size = 0;
+
+	memset(cache->hash_table, 0, byte_size);
+	for (uint32_t i = 0; i < old_table_size; i++) {
+		struct cache_entry *entry = old_table[i];
+		if (!entry)
+			continue;
+
+		radv_pipeline_cache_set_entry(cache, entry);
+	}
+
+	free(old_table);
+
+	return VK_SUCCESS;
+}
+
+static void
+radv_pipeline_cache_add_entry(struct radv_pipeline_cache *cache,
+			      struct cache_entry *entry)
+{
+	if (cache->kernel_count == cache->table_size / 2)
+		radv_pipeline_cache_grow(cache);
+
+	/* Failing to grow that hash table isn't fatal, but may mean we don't
+	 * have enough space to add this new kernel. Only add it if there's room.
+	 */
+	if (cache->kernel_count < cache->table_size / 2)
+		radv_pipeline_cache_set_entry(cache, entry);
+}
+
+struct radv_shader_variant *
+radv_pipeline_cache_insert_shader(struct radv_pipeline_cache *cache,
+				  const unsigned char *sha1,
+				  struct radv_shader_variant *variant,
+				  const void *code, unsigned code_size)
+{
+	pthread_mutex_lock(&cache->mutex);
+	struct cache_entry *entry = radv_pipeline_cache_search_unlocked(cache, sha1);
+	if (entry) {
+		if (entry->variant) {
+			radv_shader_variant_destroy(cache->device, variant);
+			variant = entry->variant;
+		} else {
+			entry->variant = variant;
+		}
+		__sync_fetch_and_add(&variant->ref_count, 1);
+		pthread_mutex_unlock(&cache->mutex);
+		return variant;
+	}
+
+	entry = radv_alloc(&cache->alloc, sizeof(*entry) + code_size, 8,
+			   VK_SYSTEM_ALLOCATION_SCOPE_CACHE);
+	if (!entry) {
+		pthread_mutex_unlock(&cache->mutex);
+		return variant;
+	}
+
+	memcpy(entry->sha1, sha1, 20);
+	memcpy(entry->code, code, code_size);
+	entry->config = variant->config;
+	entry->variant_info = variant->info;
+	entry->rsrc1 = variant->rsrc1;
+	entry->rsrc2 = variant->rsrc2;
+	entry->code_size = code_size;
+	entry->variant = variant;
+	__sync_fetch_and_add(&variant->ref_count, 1);
+
+	radv_pipeline_cache_add_entry(cache, entry);
+
+	cache->modified = true;
+	pthread_mutex_unlock(&cache->mutex);
+	return variant;
+}
+
+struct cache_header {
+	uint32_t header_size;
+	uint32_t header_version;
+	uint32_t vendor_id;
+	uint32_t device_id;
+	uint8_t  uuid[VK_UUID_SIZE];
+};
+void
+radv_pipeline_cache_load(struct radv_pipeline_cache *cache,
+			 const void *data, size_t size)
+{
+	struct radv_device *device = cache->device;
+	struct cache_header header;
+	uint8_t uuid[VK_UUID_SIZE];
+
+	if (size < sizeof(header))
+		return;
+	memcpy(&header, data, sizeof(header));
+	if (header.header_size < sizeof(header))
+		return;
+	if (header.header_version != VK_PIPELINE_CACHE_HEADER_VERSION_ONE)
+		return;
+	if (header.vendor_id != 0x1002)
+		return;
+	if (header.device_id != device->instance->physicalDevice.rad_info.pci_id)
+		return;
+	radv_device_get_cache_uuid(uuid);
+	if (memcmp(header.uuid, uuid, VK_UUID_SIZE) != 0)
+		return;
+
+	char *end = (void *) data + size;
+	char *p = (void *) data + header.header_size;
+
+	while (end - p >= sizeof(struct cache_entry)) {
+		struct cache_entry *entry = (struct cache_entry*)p;
+		struct cache_entry *dest_entry;
+		if(end - p < sizeof(*entry) + entry->code_size)
+			break;
+
+		dest_entry = radv_alloc(&cache->alloc, sizeof(*entry) + entry->code_size,
+					8, VK_SYSTEM_ALLOCATION_SCOPE_CACHE);
+		if (dest_entry) {
+			memcpy(dest_entry, entry, sizeof(*entry) + entry->code_size);
+			dest_entry->variant = NULL;
+			radv_pipeline_cache_add_entry(cache, dest_entry);
+		}
+		p += sizeof (*entry) + entry->code_size;
+	}
+}
+
+VkResult radv_CreatePipelineCache(
+	VkDevice                                    _device,
+	const VkPipelineCacheCreateInfo*            pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkPipelineCache*                            pPipelineCache)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	struct radv_pipeline_cache *cache;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO);
+	assert(pCreateInfo->flags == 0);
+
+	cache = radv_alloc2(&device->alloc, pAllocator,
+			    sizeof(*cache), 8,
+			    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (cache == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	if (pAllocator)
+		cache->alloc = *pAllocator;
+	else
+		cache->alloc = device->alloc;
+
+	radv_pipeline_cache_init(cache, device);
+
+	if (pCreateInfo->initialDataSize > 0) {
+		radv_pipeline_cache_load(cache,
+					 pCreateInfo->pInitialData,
+					 pCreateInfo->initialDataSize);
+	}
+
+	*pPipelineCache = radv_pipeline_cache_to_handle(cache);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroyPipelineCache(
+	VkDevice                                    _device,
+	VkPipelineCache                             _cache,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
+
+	if (!cache)
+		return;
+	radv_pipeline_cache_finish(cache);
+
+	radv_free2(&device->alloc, pAllocator, cache);
+}
+
+VkResult radv_GetPipelineCacheData(
+	VkDevice                                    _device,
+	VkPipelineCache                             _cache,
+	size_t*                                     pDataSize,
+	void*                                       pData)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
+	struct cache_header *header;
+	VkResult result = VK_SUCCESS;
+	const size_t size = sizeof(*header) + cache->total_size;
+	if (pData == NULL) {
+		*pDataSize = size;
+		return VK_SUCCESS;
+	}
+	if (*pDataSize < sizeof(*header)) {
+		*pDataSize = 0;
+		return VK_INCOMPLETE;
+	}
+	void *p = pData, *end = pData + *pDataSize;
+	header = p;
+	header->header_size = sizeof(*header);
+	header->header_version = VK_PIPELINE_CACHE_HEADER_VERSION_ONE;
+	header->vendor_id = 0x1002;
+	header->device_id = device->instance->physicalDevice.rad_info.pci_id;
+	radv_device_get_cache_uuid(header->uuid);
+	p += header->header_size;
+
+	struct cache_entry *entry;
+	for (uint32_t i = 0; i < cache->table_size; i++) {
+		if (!cache->hash_table[i])
+			continue;
+		entry = cache->hash_table[i];
+		const uint32_t size = entry_size(entry);
+		if (end < p + size) {
+			result = VK_INCOMPLETE;
+			break;
+		}
+
+		memcpy(p, entry, size);
+		((struct cache_entry*)p)->variant = NULL;
+		p += size;
+	}
+	*pDataSize = p - pData;
+
+	return result;
+}
+
+static void
+radv_pipeline_cache_merge(struct radv_pipeline_cache *dst,
+			  struct radv_pipeline_cache *src)
+{
+	for (uint32_t i = 0; i < src->table_size; i++) {
+		struct cache_entry *entry = src->hash_table[i];
+		if (!entry || radv_pipeline_cache_search(dst, entry->sha1))
+			continue;
+
+		radv_pipeline_cache_add_entry(dst, entry);
+
+		src->hash_table[i] = NULL;
+	}
+}
+
+VkResult radv_MergePipelineCaches(
+	VkDevice                                    _device,
+	VkPipelineCache                             destCache,
+	uint32_t                                    srcCacheCount,
+	const VkPipelineCache*                      pSrcCaches)
+{
+	RADV_FROM_HANDLE(radv_pipeline_cache, dst, destCache);
+
+	for (uint32_t i = 0; i < srcCacheCount; i++) {
+		RADV_FROM_HANDLE(radv_pipeline_cache, src, pSrcCaches[i]);
+
+		radv_pipeline_cache_merge(dst, src);
+	}
+
+	return VK_SUCCESS;
+}
diff --git a/src/amd/vulkan/radv_private.h b/src/amd/vulkan/radv_private.h
new file mode 100644
index 0000000000..e738218fc1
--- /dev/null
+++ b/src/amd/vulkan/radv_private.h
@@ -0,0 +1,1402 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based in part on anv driver which is:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#pragma once
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <stdbool.h>
+#include <pthread.h>
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+#ifdef HAVE_VALGRIND
+#include <valgrind.h>
+#include <memcheck.h>
+#define VG(x) x
+#define __gen_validate_value(x) VALGRIND_CHECK_MEM_IS_DEFINED(&(x), sizeof(x))
+#else
+#define VG(x)
+#endif
+
+#include <amdgpu.h>
+#include "radv_device_info.h"
+#include "compiler/shader_enums.h"
+#include "util/macros.h"
+#include "util/list.h"
+#include "main/macros.h"
+#include "radv_radeon_winsys.h"
+#include "ac_binary.h"
+#include "ac_nir_to_llvm.h"
+#include "radv_descriptor_set.h"
+
+#include <llvm-c/TargetMachine.h>
+
+/* Pre-declarations needed for WSI entrypoints */
+struct wl_surface;
+struct wl_display;
+typedef struct xcb_connection_t xcb_connection_t;
+typedef uint32_t xcb_visualid_t;
+typedef uint32_t xcb_window_t;
+
+#include <vulkan/vulkan.h>
+#include <vulkan/vulkan_intel.h>
+#include <vulkan/vk_icd.h>
+
+#include "radv_entrypoints.h"
+
+
+#define MAX_VBS         32
+#define MAX_VERTEX_ATTRIBS 32
+#define MAX_RTS          8
+#define MAX_VIEWPORTS   16
+#define MAX_SCISSORS    16
+#define MAX_PUSH_CONSTANTS_SIZE 128
+#define MAX_DYNAMIC_BUFFERS 16
+#define MAX_IMAGES 8
+#define MAX_SAMPLES_LOG2 4 /* SKL supports 16 samples */
+#define NUM_META_FS_KEYS 11
+
+#define NUM_DEPTH_CLEAR_PIPELINES 3
+
+#define radv_noreturn __attribute__((__noreturn__))
+#define radv_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+static inline uint32_t
+align_u32(uint32_t v, uint32_t a)
+{
+	assert(a != 0 && a == (a & -a));
+	return (v + a - 1) & ~(a - 1);
+}
+
+static inline uint32_t
+align_u32_npot(uint32_t v, uint32_t a)
+{
+	return (v + a - 1) / a * a;
+}
+
+static inline uint64_t
+align_u64(uint64_t v, uint64_t a)
+{
+	assert(a != 0 && a == (a & -a));
+	return (v + a - 1) & ~(a - 1);
+}
+
+static inline int32_t
+align_i32(int32_t v, int32_t a)
+{
+	assert(a != 0 && a == (a & -a));
+	return (v + a - 1) & ~(a - 1);
+}
+
+/** Alignment must be a power of 2. */
+static inline bool
+radv_is_aligned(uintmax_t n, uintmax_t a)
+{
+	assert(a == (a & -a));
+	return (n & (a - 1)) == 0;
+}
+
+static inline uint32_t
+round_up_u32(uint32_t v, uint32_t a)
+{
+	return (v + a - 1) / a;
+}
+
+static inline uint64_t
+round_up_u64(uint64_t v, uint64_t a)
+{
+	return (v + a - 1) / a;
+}
+
+static inline uint32_t
+radv_minify(uint32_t n, uint32_t levels)
+{
+	if (unlikely(n == 0))
+		return 0;
+	else
+		return MAX(n >> levels, 1);
+}
+static inline float
+radv_clamp_f(float f, float min, float max)
+{
+	assert(min < max);
+
+	if (f > max)
+		return max;
+	else if (f < min)
+		return min;
+	else
+		return f;
+}
+
+static inline bool
+radv_clear_mask(uint32_t *inout_mask, uint32_t clear_mask)
+{
+	if (*inout_mask & clear_mask) {
+		*inout_mask &= ~clear_mask;
+		return true;
+	} else {
+		return false;
+	}
+}
+
+#define for_each_bit(b, dword)                          \
+	for (uint32_t __dword = (dword);		\
+	     (b) = __builtin_ffs(__dword) - 1, __dword;	\
+	     __dword &= ~(1 << (b)))
+
+#define typed_memcpy(dest, src, count) ({				\
+			static_assert(sizeof(*src) == sizeof(*dest), ""); \
+			memcpy((dest), (src), (count) * sizeof(*(src))); \
+		})
+
+#define zero(x) (memset(&(x), 0, sizeof(x)))
+
+/* Define no kernel as 1, since that's an illegal offset for a kernel */
+#define NO_KERNEL 1
+
+struct radv_common {
+	VkStructureType                             sType;
+	const void*                                 pNext;
+};
+
+/* Whenever we generate an error, pass it through this function. Useful for
+ * debugging, where we can break on it. Only call at error site, not when
+ * propagating errors. Might be useful to plug in a stack trace here.
+ */
+
+VkResult __vk_errorf(VkResult error, const char *file, int line, const char *format, ...);
+
+#ifdef DEBUG
+#define vk_error(error) __vk_errorf(error, __FILE__, __LINE__, NULL);
+#define vk_errorf(error, format, ...) __vk_errorf(error, __FILE__, __LINE__, format, ## __VA_ARGS__);
+#else
+#define vk_error(error) error
+#define vk_errorf(error, format, ...) error
+#endif
+
+void __radv_finishme(const char *file, int line, const char *format, ...)
+	radv_printflike(3, 4);
+void radv_loge(const char *format, ...) radv_printflike(1, 2);
+void radv_loge_v(const char *format, va_list va);
+
+/**
+ * Print a FINISHME message, including its source location.
+ */
+#define radv_finishme(format, ...)					\
+	__radv_finishme(__FILE__, __LINE__, format, ##__VA_ARGS__);
+
+/* A non-fatal assert.  Useful for debugging. */
+#ifdef DEBUG
+#define radv_assert(x) ({						\
+			if (unlikely(!(x)))				\
+				fprintf(stderr, "%s:%d ASSERT: %s\n", __FILE__, __LINE__, #x); \
+		})
+#else
+#define radv_assert(x)
+#endif
+
+/**
+ * If a block of code is annotated with radv_validate, then the block runs only
+ * in debug builds.
+ */
+#ifdef DEBUG
+#define radv_validate if (1)
+#else
+#define radv_validate if (0)
+#endif
+
+void radv_abortf(const char *format, ...) radv_noreturn radv_printflike(1, 2);
+void radv_abortfv(const char *format, va_list va) radv_noreturn;
+
+#define stub_return(v)					\
+	do {						\
+		radv_finishme("stub %s", __func__);	\
+		return (v);				\
+	} while (0)
+
+#define stub()						\
+	do {						\
+		radv_finishme("stub %s", __func__);	\
+		return;					\
+	} while (0)
+
+/**
+ * A dynamically growable, circular buffer.  Elements are added at head and
+ * removed from tail. head and tail are free-running uint32_t indices and we
+ * only compute the modulo with size when accessing the array.  This way,
+ * number of bytes in the queue is always head - tail, even in case of
+ * wraparound.
+ */
+
+struct radv_vector {
+   uint32_t head;
+   uint32_t tail;
+   uint32_t element_size;
+   uint32_t size;
+   void *data;
+};
+
+int radv_vector_init(struct radv_vector *queue, uint32_t element_size, uint32_t size);
+void *radv_vector_add(struct radv_vector *queue);
+void *radv_vector_remove(struct radv_vector *queue);
+
+static inline int
+radv_vector_length(struct radv_vector *queue)
+{
+   return (queue->head - queue->tail) / queue->element_size;
+}
+
+static inline void *
+radv_vector_head(struct radv_vector *vector)
+{
+   assert(vector->tail < vector->head);
+   return (void *)((char *)vector->data +
+                   ((vector->head - vector->element_size) &
+                    (vector->size - 1)));
+}
+
+static inline void *
+radv_vector_tail(struct radv_vector *vector)
+{
+   return (void *)((char *)vector->data + (vector->tail & (vector->size - 1)));
+}
+
+static inline void
+radv_vector_finish(struct radv_vector *queue)
+{
+   free(queue->data);
+}
+
+#define radv_vector_foreach(elem, queue)                                  \
+   static_assert(__builtin_types_compatible_p(__typeof__(queue), struct radv_vector *), ""); \
+   for (uint32_t __radv_vector_offset = (queue)->tail;                                \
+        elem = (queue)->data + (__radv_vector_offset & ((queue)->size - 1)), __radv_vector_offset < (queue)->head; \
+        __radv_vector_offset += (queue)->element_size)
+
+void *radv_resolve_entrypoint(uint32_t index);
+void *radv_lookup_entrypoint(const char *name);
+
+extern struct radv_dispatch_table dtable;
+
+#define RADV_CALL(func) ({						\
+			if (dtable.func == NULL) {			\
+				size_t idx = offsetof(struct radv_dispatch_table, func) / sizeof(void *); \
+				dtable.entrypoints[idx] = radv_resolve_entrypoint(idx); \
+			}						\
+			dtable.func;					\
+		})
+
+static inline void *
+radv_alloc(const VkAllocationCallbacks *alloc,
+	   size_t size, size_t align,
+	   VkSystemAllocationScope scope)
+{
+	return alloc->pfnAllocation(alloc->pUserData, size, align, scope);
+}
+
+static inline void *
+radv_realloc(const VkAllocationCallbacks *alloc,
+	     void *ptr, size_t size, size_t align,
+	     VkSystemAllocationScope scope)
+{
+	return alloc->pfnReallocation(alloc->pUserData, ptr, size, align, scope);
+}
+
+static inline void
+radv_free(const VkAllocationCallbacks *alloc, void *data)
+{
+	alloc->pfnFree(alloc->pUserData, data);
+}
+
+static inline void *
+radv_alloc2(const VkAllocationCallbacks *parent_alloc,
+	    const VkAllocationCallbacks *alloc,
+	    size_t size, size_t align,
+	    VkSystemAllocationScope scope)
+{
+	if (alloc)
+		return radv_alloc(alloc, size, align, scope);
+	else
+		return radv_alloc(parent_alloc, size, align, scope);
+}
+
+static inline void
+radv_free2(const VkAllocationCallbacks *parent_alloc,
+	   const VkAllocationCallbacks *alloc,
+	   void *data)
+{
+	if (alloc)
+		radv_free(alloc, data);
+	else
+		radv_free(parent_alloc, data);
+}
+
+struct radv_wsi_interaface;
+
+#define VK_ICD_WSI_PLATFORM_MAX 5
+
+struct radv_physical_device {
+	VK_LOADER_DATA                              _loader_data;
+
+	struct radv_instance *                       instance;
+
+	struct radeon_winsys *ws;
+	struct radeon_info rad_info;
+	uint32_t                                    chipset_id;
+	char                                        path[20];
+	const char *                                name;
+	uint64_t                                    aperture_size;
+	int                                         cmd_parser_version;
+	uint32_t                    pci_vendor_id;
+	uint32_t                    pci_device_id;
+
+	struct radv_wsi_interface *                  wsi[VK_ICD_WSI_PLATFORM_MAX];
+};
+
+struct radv_instance {
+	VK_LOADER_DATA                              _loader_data;
+
+	VkAllocationCallbacks                       alloc;
+
+	uint32_t                                    apiVersion;
+	int                                         physicalDeviceCount;
+	struct radv_physical_device                  physicalDevice;
+};
+
+VkResult radv_init_wsi(struct radv_physical_device *physical_device);
+void radv_finish_wsi(struct radv_physical_device *physical_device);
+
+struct cache_entry;
+
+struct radv_pipeline_cache {
+	struct radv_device *                          device;
+	pthread_mutex_t                              mutex;
+
+	uint32_t                                     total_size;
+	uint32_t                                     table_size;
+	uint32_t                                     kernel_count;
+	struct cache_entry **                        hash_table;
+	bool                                         modified;
+
+	VkAllocationCallbacks                        alloc;
+};
+
+void
+radv_pipeline_cache_init(struct radv_pipeline_cache *cache,
+			 struct radv_device *device);
+void
+radv_pipeline_cache_finish(struct radv_pipeline_cache *cache);
+void
+radv_pipeline_cache_load(struct radv_pipeline_cache *cache,
+			 const void *data, size_t size);
+
+struct radv_shader_variant *
+radv_create_shader_variant_from_pipeline_cache(struct radv_device *device,
+					       struct radv_pipeline_cache *cache,
+					       const unsigned char *sha1);
+
+struct radv_shader_variant *
+radv_pipeline_cache_insert_shader(struct radv_pipeline_cache *cache,
+				  const unsigned char *sha1,
+				  struct radv_shader_variant *variant,
+				  const void *code, unsigned code_size);
+
+void radv_shader_variant_destroy(struct radv_device *device,
+				 struct radv_shader_variant *variant);
+
+struct radv_meta_state {
+	VkAllocationCallbacks alloc;
+
+	struct radv_pipeline_cache cache;
+
+	/**
+	 * Use array element `i` for images with `2^i` samples.
+	 */
+	struct {
+		VkRenderPass render_pass[NUM_META_FS_KEYS];
+		struct radv_pipeline *color_pipelines[NUM_META_FS_KEYS];
+
+		VkRenderPass depth_only_rp[NUM_DEPTH_CLEAR_PIPELINES];
+		struct radv_pipeline *depth_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
+		VkRenderPass stencil_only_rp[NUM_DEPTH_CLEAR_PIPELINES];
+		struct radv_pipeline *stencil_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
+		VkRenderPass depthstencil_rp[NUM_DEPTH_CLEAR_PIPELINES];
+		struct radv_pipeline *depthstencil_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
+	} clear[1 + MAX_SAMPLES_LOG2];
+
+	struct {
+		VkRenderPass render_pass[NUM_META_FS_KEYS];
+
+		/** Pipeline that blits from a 1D image. */
+		VkPipeline pipeline_1d_src[NUM_META_FS_KEYS];
+
+		/** Pipeline that blits from a 2D image. */
+		VkPipeline pipeline_2d_src[NUM_META_FS_KEYS];
+
+		/** Pipeline that blits from a 3D image. */
+		VkPipeline pipeline_3d_src[NUM_META_FS_KEYS];
+
+		VkRenderPass depth_only_rp;
+		VkPipeline depth_only_1d_pipeline;
+		VkPipeline depth_only_2d_pipeline;
+		VkPipeline depth_only_3d_pipeline;
+
+		VkRenderPass stencil_only_rp;
+		VkPipeline stencil_only_1d_pipeline;
+		VkPipeline stencil_only_2d_pipeline;
+		VkPipeline stencil_only_3d_pipeline;
+		VkPipelineLayout                          pipeline_layout;
+		VkDescriptorSetLayout                     ds_layout;
+	} blit;
+
+	struct {
+		VkRenderPass render_passes[NUM_META_FS_KEYS];
+
+		VkPipelineLayout p_layouts[2];
+		VkDescriptorSetLayout ds_layouts[2];
+		VkPipeline pipelines[2][NUM_META_FS_KEYS];
+
+		VkRenderPass depth_only_rp;
+		VkPipeline depth_only_pipeline[2];
+
+		VkRenderPass stencil_only_rp;
+		VkPipeline stencil_only_pipeline[2];
+	} blit2d;
+
+	struct {
+		VkPipelineLayout                          img_p_layout;
+		VkDescriptorSetLayout                     img_ds_layout;
+		VkPipeline pipeline;
+	} itob;
+	struct {
+		VkRenderPass render_pass;
+		VkPipelineLayout                          img_p_layout;
+		VkDescriptorSetLayout                     img_ds_layout;
+		VkPipeline pipeline;
+	} btoi;
+
+	struct {
+		VkPipeline                                pipeline;
+		VkRenderPass                              pass;
+	} resolve;
+
+	struct {
+		VkDescriptorSetLayout                     ds_layout;
+		VkPipelineLayout                          p_layout;
+		struct {
+			VkPipeline                                pipeline;
+			VkPipeline                                i_pipeline;
+		} rc[MAX_SAMPLES_LOG2];
+	} resolve_compute;
+
+	struct {
+		VkPipeline                                decompress_pipeline;
+		VkPipeline                                resummarize_pipeline;
+		VkRenderPass                              pass;
+	} depth_decomp;
+
+	struct {
+		VkPipeline                                cmask_eliminate_pipeline;
+		VkPipeline                                fmask_decompress_pipeline;
+		VkRenderPass                              pass;
+	} fast_clear_flush;
+
+	struct {
+		VkPipelineLayout fill_p_layout;
+		VkPipelineLayout copy_p_layout;
+		VkDescriptorSetLayout fill_ds_layout;
+		VkDescriptorSetLayout copy_ds_layout;
+		VkPipeline fill_pipeline;
+		VkPipeline copy_pipeline;
+	} buffer;
+};
+
+struct radv_queue {
+	VK_LOADER_DATA                              _loader_data;
+
+	struct radv_device *                         device;
+
+	struct radv_state_pool *                     pool;
+};
+
+struct radv_device {
+	VK_LOADER_DATA                              _loader_data;
+
+	VkAllocationCallbacks                       alloc;
+
+	struct radv_instance *                       instance;
+	struct radeon_winsys *ws;
+	struct radeon_winsys_ctx *hw_ctx;
+
+	struct radv_meta_state                       meta_state;
+	struct radv_queue                            queue;
+	struct radeon_winsys_cs *empty_cs;
+
+	bool allow_fast_clears;
+	bool allow_dcc;
+
+	/* MSAA sample locations.
+	 * The first index is the sample index.
+	 * The second index is the coordinate: X, Y. */
+	float sample_locations_1x[1][2];
+	float sample_locations_2x[2][2];
+	float sample_locations_4x[4][2];
+	float sample_locations_8x[8][2];
+	float sample_locations_16x[16][2];
+};
+
+void radv_device_get_cache_uuid(void *uuid);
+
+struct radv_device_memory {
+	struct radeon_winsys_bo                      *bo;
+	uint32_t                                     type_index;
+	VkDeviceSize                                 map_size;
+	void *                                       map;
+};
+
+
+struct radv_descriptor_range {
+	uint64_t va;
+	uint32_t size;
+};
+
+struct radv_descriptor_set {
+	const struct radv_descriptor_set_layout *layout;
+	struct list_head descriptor_pool;
+	uint32_t size;
+
+	struct radv_buffer_view *buffer_views;
+	struct radeon_winsys_bo *bo;
+	uint64_t va;
+	uint32_t *mapped_ptr;
+	struct radv_descriptor_range *dynamic_descriptors;
+	struct radeon_winsys_bo *descriptors[0];
+};
+
+struct radv_descriptor_pool_free_node {
+	int next;
+	uint32_t offset;
+	uint32_t size;
+};
+
+struct radv_descriptor_pool {
+	struct list_head descriptor_sets;
+
+	struct radeon_winsys_bo *bo;
+	uint8_t *mapped_ptr;
+	uint64_t current_offset;
+	uint64_t size;
+
+	int free_list;
+	int full_list;
+	uint32_t max_sets;
+	struct radv_descriptor_pool_free_node free_nodes[];
+};
+
+struct radv_buffer {
+	struct radv_device *                          device;
+	VkDeviceSize                                 size;
+
+	VkBufferUsageFlags                           usage;
+
+	/* Set when bound */
+	struct radeon_winsys_bo *                      bo;
+	VkDeviceSize                                 offset;
+};
+
+
+enum radv_cmd_dirty_bits {
+	RADV_CMD_DIRTY_DYNAMIC_VIEWPORT                  = 1 << 0, /* VK_DYNAMIC_STATE_VIEWPORT */
+	RADV_CMD_DIRTY_DYNAMIC_SCISSOR                   = 1 << 1, /* VK_DYNAMIC_STATE_SCISSOR */
+	RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH                = 1 << 2, /* VK_DYNAMIC_STATE_LINE_WIDTH */
+	RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS                = 1 << 3, /* VK_DYNAMIC_STATE_DEPTH_BIAS */
+	RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS           = 1 << 4, /* VK_DYNAMIC_STATE_BLEND_CONSTANTS */
+	RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS              = 1 << 5, /* VK_DYNAMIC_STATE_DEPTH_BOUNDS */
+	RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK      = 1 << 6, /* VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK */
+	RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK        = 1 << 7, /* VK_DYNAMIC_STATE_STENCIL_WRITE_MASK */
+	RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE         = 1 << 8, /* VK_DYNAMIC_STATE_STENCIL_REFERENCE */
+	RADV_CMD_DIRTY_DYNAMIC_ALL                       = (1 << 9) - 1,
+	RADV_CMD_DIRTY_PIPELINE                          = 1 << 9,
+	RADV_CMD_DIRTY_INDEX_BUFFER                      = 1 << 10,
+	RADV_CMD_DIRTY_RENDER_TARGETS                    = 1 << 11,
+};
+typedef uint32_t radv_cmd_dirty_mask_t;
+
+enum radv_cmd_flush_bits {
+	RADV_CMD_FLAG_INV_ICACHE = 1 << 0,
+	/* SMEM L1, other names: KCACHE, constant cache, DCACHE, data cache */
+	RADV_CMD_FLAG_INV_SMEM_L1 = 1 << 1,
+	/* VMEM L1 can optionally be bypassed (GLC=1). Other names: TC L1 */
+	RADV_CMD_FLAG_INV_VMEM_L1 = 1 << 2,
+	/* Used by everything except CB/DB, can be bypassed (SLC=1). Other names: TC L2 */
+	RADV_CMD_FLAG_INV_GLOBAL_L2 = 1 << 3,
+	/* Framebuffer caches */
+	RADV_CMD_FLAG_FLUSH_AND_INV_CB_META = 1 << 4,
+	RADV_CMD_FLAG_FLUSH_AND_INV_DB_META = 1 << 5,
+	RADV_CMD_FLAG_FLUSH_AND_INV_DB = 1 << 6,
+	RADV_CMD_FLAG_FLUSH_AND_INV_CB = 1 << 7,
+	/* Engine synchronization. */
+	RADV_CMD_FLAG_VS_PARTIAL_FLUSH = 1 << 8,
+	RADV_CMD_FLAG_PS_PARTIAL_FLUSH = 1 << 9,
+	RADV_CMD_FLAG_CS_PARTIAL_FLUSH = 1 << 10,
+	RADV_CMD_FLAG_VGT_FLUSH        = 1 << 11,
+
+	RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER = (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+					      RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
+					      RADV_CMD_FLAG_FLUSH_AND_INV_DB |
+					      RADV_CMD_FLAG_FLUSH_AND_INV_DB_META)
+};
+
+struct radv_vertex_binding {
+	struct radv_buffer *                          buffer;
+	VkDeviceSize                                 offset;
+};
+
+struct radv_dynamic_state {
+	struct {
+		uint32_t                                  count;
+		VkViewport                                viewports[MAX_VIEWPORTS];
+	} viewport;
+
+	struct {
+		uint32_t                                  count;
+		VkRect2D                                  scissors[MAX_SCISSORS];
+	} scissor;
+
+	float                                        line_width;
+
+	struct {
+		float                                     bias;
+		float                                     clamp;
+		float                                     slope;
+	} depth_bias;
+
+	float                                        blend_constants[4];
+
+	struct {
+		float                                     min;
+		float                                     max;
+	} depth_bounds;
+
+	struct {
+		uint32_t                                  front;
+		uint32_t                                  back;
+	} stencil_compare_mask;
+
+	struct {
+		uint32_t                                  front;
+		uint32_t                                  back;
+	} stencil_write_mask;
+
+	struct {
+		uint32_t                                  front;
+		uint32_t                                  back;
+	} stencil_reference;
+};
+
+extern const struct radv_dynamic_state default_dynamic_state;
+
+void radv_dynamic_state_copy(struct radv_dynamic_state *dest,
+			     const struct radv_dynamic_state *src,
+			     uint32_t copy_mask);
+/**
+ * Attachment state when recording a renderpass instance.
+ *
+ * The clear value is valid only if there exists a pending clear.
+ */
+struct radv_attachment_state {
+	VkImageAspectFlags                           pending_clear_aspects;
+	VkClearValue                                 clear_value;
+	VkImageLayout                                current_layout;
+};
+
+struct radv_cmd_state {
+	uint32_t                                      vb_dirty;
+	bool                                          vertex_descriptors_dirty;
+	radv_cmd_dirty_mask_t                         dirty;
+
+	struct radv_pipeline *                        pipeline;
+	struct radv_pipeline *                        emitted_pipeline;
+	struct radv_pipeline *                        compute_pipeline;
+	struct radv_pipeline *                        emitted_compute_pipeline;
+	struct radv_framebuffer *                     framebuffer;
+	struct radv_render_pass *                     pass;
+	const struct radv_subpass *                         subpass;
+	struct radv_dynamic_state                     dynamic;
+	struct radv_vertex_binding                    vertex_bindings[MAX_VBS];
+	struct radv_descriptor_set *                  descriptors[MAX_SETS];
+	struct radv_attachment_state *                attachments;
+	VkRect2D                                     render_area;
+	struct radv_buffer *                         index_buffer;
+	uint32_t                                     index_type;
+	uint32_t                                     index_offset;
+	uint32_t                                     last_primitive_reset_index;
+	enum radv_cmd_flush_bits                     flush_bits;
+	unsigned                                     active_occlusion_queries;
+	float					     offset_scale;
+};
+struct radv_cmd_pool {
+	VkAllocationCallbacks                        alloc;
+	struct list_head                             cmd_buffers;
+};
+
+struct radv_cmd_buffer_upload {
+	uint8_t *map;
+	unsigned offset;
+	uint64_t size;
+	struct radeon_winsys_bo *upload_bo;
+	struct list_head list;
+};
+
+struct radv_cmd_buffer {
+	VK_LOADER_DATA                               _loader_data;
+
+	struct radv_device *                          device;
+
+	struct radv_cmd_pool *                        pool;
+	struct list_head                             pool_link;
+
+	VkCommandBufferUsageFlags                    usage_flags;
+	VkCommandBufferLevel                         level;
+	struct radeon_winsys_cs *cs;
+	struct radv_cmd_state state;
+
+	uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
+	uint32_t dynamic_buffers[16 * MAX_DYNAMIC_BUFFERS];
+	VkShaderStageFlags push_constant_stages;
+
+	struct radv_cmd_buffer_upload upload;
+
+	bool record_fail;
+};
+
+struct radv_image;
+
+void si_init_config(struct radv_physical_device *physical_device,
+		    struct radv_cmd_buffer *cmd_buffer);
+void si_write_viewport(struct radeon_winsys_cs *cs, int first_vp,
+		       int count, const VkViewport *viewports);
+void si_write_scissors(struct radeon_winsys_cs *cs, int first,
+		       int count, const VkRect2D *scissors);
+uint32_t si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer);
+void si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer);
+void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
+			   uint64_t src_va, uint64_t dest_va,
+			   uint64_t size);
+void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
+			    uint64_t size, unsigned value);
+void radv_set_db_count_control(struct radv_cmd_buffer *cmd_buffer);
+void radv_bind_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
+			      struct radv_descriptor_set *set,
+			      unsigned idx);
+bool
+radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
+			     unsigned size,
+			     unsigned alignment,
+			     unsigned *out_offset,
+			     void **ptr);
+void
+radv_cmd_buffer_set_subpass(struct radv_cmd_buffer *cmd_buffer,
+			    const struct radv_subpass *subpass,
+			    bool transitions);
+bool
+radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
+			    unsigned size, unsigned alignmnet,
+			    const void *data, unsigned *out_offset);
+void
+radv_emit_framebuffer_state(struct radv_cmd_buffer *cmd_buffer);
+void radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer *cmd_buffer);
+void radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer);
+void radv_cayman_emit_msaa_sample_locs(struct radeon_winsys_cs *cs, int nr_samples);
+unsigned radv_cayman_get_maxdist(int log_samples);
+void radv_device_init_msaa(struct radv_device *device);
+void radv_set_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
+			       struct radv_image *image,
+			       VkClearDepthStencilValue ds_clear_value,
+			       VkImageAspectFlags aspects);
+void radv_set_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
+			       struct radv_image *image,
+			       int idx,
+			       uint32_t color_values[2]);
+void radv_fill_buffer(struct radv_cmd_buffer *cmd_buffer,
+		      struct radeon_winsys_bo *bo,
+		      uint64_t offset, uint64_t size, uint32_t value);
+
+/*
+ * Takes x,y,z as exact numbers of invocations, instead of blocks.
+ *
+ * Limitations: Can't call normal dispatch functions without binding or rebinding
+ *              the compute pipeline.
+ */
+void radv_unaligned_dispatch(
+	struct radv_cmd_buffer                      *cmd_buffer,
+	uint32_t                                    x,
+	uint32_t                                    y,
+	uint32_t                                    z);
+
+struct radv_event {
+	struct radeon_winsys_bo *bo;
+	uint64_t *map;
+};
+
+struct nir_shader;
+
+struct radv_shader_module {
+	struct nir_shader *                          nir;
+	unsigned char                                sha1[20];
+	uint32_t                                     size;
+	char                                         data[0];
+};
+
+union ac_shader_variant_key;
+
+void
+radv_hash_shader(unsigned char *hash, struct radv_shader_module *module,
+		 const char *entrypoint,
+		 const VkSpecializationInfo *spec_info,
+		 const struct radv_pipeline_layout *layout,
+		 const union ac_shader_variant_key *key);
+
+static inline gl_shader_stage
+vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
+{
+	assert(__builtin_popcount(vk_stage) == 1);
+	return ffs(vk_stage) - 1;
+}
+
+static inline VkShaderStageFlagBits
+mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
+{
+	return (1 << mesa_stage);
+}
+
+#define RADV_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
+
+#define radv_foreach_stage(stage, stage_bits)				\
+	for (gl_shader_stage stage,					\
+		     __tmp = (gl_shader_stage)((stage_bits) & RADV_STAGE_MASK);	\
+	     stage = __builtin_ffs(__tmp) - 1, __tmp;			\
+	     __tmp &= ~(1 << (stage)))
+
+struct radv_shader_variant {
+	uint32_t ref_count;
+
+	struct radeon_winsys_bo *bo;
+	struct ac_shader_config config;
+	struct ac_shader_variant_info info;
+	unsigned rsrc1;
+	unsigned rsrc2;
+};
+
+struct radv_depth_stencil_state {
+	uint32_t db_depth_control;
+	uint32_t db_stencil_control;
+	uint32_t db_render_control;
+	uint32_t db_render_override2;
+};
+
+struct radv_blend_state {
+	uint32_t cb_color_control;
+	uint32_t cb_target_mask;
+	uint32_t sx_mrt0_blend_opt[8];
+	uint32_t cb_blend_control[8];
+
+	uint32_t spi_shader_col_format;
+	uint32_t cb_shader_mask;
+	uint32_t db_alpha_to_mask;
+};
+
+unsigned radv_format_meta_fs_key(VkFormat format);
+
+struct radv_raster_state {
+	uint32_t pa_cl_clip_cntl;
+	uint32_t pa_cl_vs_out_cntl;
+	uint32_t spi_interp_control;
+	uint32_t pa_su_point_size;
+	uint32_t pa_su_point_minmax;
+	uint32_t pa_su_line_cntl;
+	uint32_t pa_su_vtx_cntl;
+	uint32_t pa_su_sc_mode_cntl;
+};
+
+struct radv_multisample_state {
+	uint32_t db_eqaa;
+	uint32_t pa_sc_line_cntl;
+	uint32_t pa_sc_mode_cntl_0;
+	uint32_t pa_sc_mode_cntl_1;
+	uint32_t pa_sc_aa_config;
+	uint32_t pa_sc_aa_mask[2];
+	unsigned num_samples;
+};
+
+struct radv_pipeline {
+	struct radv_device *                          device;
+	uint32_t                                     dynamic_state_mask;
+	struct radv_dynamic_state                     dynamic_state;
+
+	struct radv_pipeline_layout *                 layout;
+
+	bool                                         needs_data_cache;
+
+	struct radv_shader_variant *                 shaders[MESA_SHADER_STAGES];
+	VkShaderStageFlags                           active_stages;
+
+	uint32_t va_rsrc_word3[MAX_VERTEX_ATTRIBS];
+	uint32_t va_format_size[MAX_VERTEX_ATTRIBS];
+	uint32_t va_binding[MAX_VERTEX_ATTRIBS];
+	uint32_t va_offset[MAX_VERTEX_ATTRIBS];
+	uint32_t num_vertex_attribs;
+	uint32_t                                     binding_stride[MAX_VBS];
+
+	union {
+		struct {
+			struct radv_blend_state blend;
+			struct radv_depth_stencil_state ds;
+			struct radv_raster_state raster;
+			struct radv_multisample_state ms;
+			unsigned prim;
+			unsigned gs_out;
+			bool prim_restart_enable;
+		} graphics;
+	};
+};
+
+struct radv_graphics_pipeline_create_info {
+	bool use_rectlist;
+	bool db_depth_clear;
+	bool db_stencil_clear;
+	bool db_depth_disable_expclear;
+	bool db_stencil_disable_expclear;
+	bool db_flush_depth_inplace;
+	bool db_flush_stencil_inplace;
+	bool db_resummarize;
+	uint32_t custom_blend_mode;
+};
+
+VkResult
+radv_pipeline_init(struct radv_pipeline *pipeline, struct radv_device *device,
+		   struct radv_pipeline_cache *cache,
+		   const VkGraphicsPipelineCreateInfo *pCreateInfo,
+		   const struct radv_graphics_pipeline_create_info *extra,
+		   const VkAllocationCallbacks *alloc);
+
+VkResult
+radv_graphics_pipeline_create(VkDevice device,
+			      VkPipelineCache cache,
+			      const VkGraphicsPipelineCreateInfo *pCreateInfo,
+			      const struct radv_graphics_pipeline_create_info *extra,
+			      const VkAllocationCallbacks *alloc,
+			      VkPipeline *pPipeline);
+
+struct vk_format_description;
+uint32_t radv_translate_buffer_dataformat(const struct vk_format_description *desc,
+					  int first_non_void);
+uint32_t radv_translate_buffer_numformat(const struct vk_format_description *desc,
+					 int first_non_void);
+uint32_t radv_translate_colorformat(VkFormat format);
+uint32_t radv_translate_color_numformat(VkFormat format,
+					const struct vk_format_description *desc,
+					int first_non_void);
+uint32_t radv_colorformat_endian_swap(uint32_t colorformat);
+unsigned radv_translate_colorswap(VkFormat format, bool do_endian_swap);
+uint32_t radv_translate_dbformat(VkFormat format);
+uint32_t radv_translate_tex_dataformat(VkFormat format,
+				       const struct vk_format_description *desc,
+				       int first_non_void);
+uint32_t radv_translate_tex_numformat(VkFormat format,
+				      const struct vk_format_description *desc,
+				      int first_non_void);
+bool radv_format_pack_clear_color(VkFormat format,
+				  uint32_t clear_vals[2],
+				  VkClearColorValue *value);
+bool radv_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
+
+struct radv_fmask_info {
+	uint64_t offset;
+	uint64_t size;
+	unsigned alignment;
+	unsigned pitch_in_pixels;
+	unsigned bank_height;
+	unsigned slice_tile_max;
+	unsigned tile_mode_index;
+};
+
+struct radv_cmask_info {
+	uint64_t offset;
+	uint64_t size;
+	unsigned alignment;
+	unsigned pitch;
+	unsigned height;
+	unsigned xalign;
+	unsigned yalign;
+	unsigned slice_tile_max;
+	unsigned base_address_reg;
+};
+
+struct r600_htile_info {
+	uint64_t offset;
+	uint64_t size;
+	unsigned pitch;
+	unsigned height;
+	unsigned xalign;
+	unsigned yalign;
+};
+
+struct radv_image {
+	VkImageType type;
+	/* The original VkFormat provided by the client.  This may not match any
+	 * of the actual surface formats.
+	 */
+	VkFormat vk_format;
+	VkImageAspectFlags aspects;
+	VkExtent3D extent;
+	uint32_t levels;
+	uint32_t array_size;
+	uint32_t samples; /**< VkImageCreateInfo::samples */
+	VkImageUsageFlags usage; /**< Superset of VkImageCreateInfo::usage. */
+	VkImageTiling tiling; /** VkImageCreateInfo::tiling */
+
+	VkDeviceSize size;
+	uint32_t alignment;
+
+	/* Set when bound */
+	struct radeon_winsys_bo *bo;
+	VkDeviceSize offset;
+	uint32_t dcc_offset;
+	struct radeon_surf surface;
+
+	struct radv_fmask_info fmask;
+	struct radv_cmask_info cmask;
+	uint32_t clear_value_offset;
+
+	/* Depth buffer compression and fast clear. */
+	struct r600_htile_info htile;
+};
+
+bool radv_layout_has_htile(const struct radv_image *image,
+                           VkImageLayout layout);
+bool radv_layout_is_htile_compressed(const struct radv_image *image,
+                                     VkImageLayout layout);
+bool radv_layout_can_expclear(const struct radv_image *image,
+                              VkImageLayout layout);
+bool radv_layout_has_cmask(const struct radv_image *image,
+			   VkImageLayout layout);
+static inline uint32_t
+radv_get_layerCount(const struct radv_image *image,
+		    const VkImageSubresourceRange *range)
+{
+	return range->layerCount == VK_REMAINING_ARRAY_LAYERS ?
+		image->array_size - range->baseArrayLayer : range->layerCount;
+}
+
+static inline uint32_t
+radv_get_levelCount(const struct radv_image *image,
+		    const VkImageSubresourceRange *range)
+{
+	return range->levelCount == VK_REMAINING_MIP_LEVELS ?
+		image->levels - range->baseMipLevel : range->levelCount;
+}
+
+struct radeon_bo_metadata;
+void
+radv_init_metadata(struct radv_device *device,
+		   struct radv_image *image,
+		   struct radeon_bo_metadata *metadata);
+
+struct radv_image_view {
+	struct radv_image *image; /**< VkImageViewCreateInfo::image */
+	struct radeon_winsys_bo *bo;
+
+	VkImageViewType type;
+	VkImageAspectFlags aspect_mask;
+	VkFormat vk_format;
+	uint32_t base_layer;
+	uint32_t layer_count;
+	uint32_t base_mip;
+	VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
+
+	uint32_t descriptor[8];
+	uint32_t fmask_descriptor[8];
+};
+
+struct radv_image_create_info {
+	const VkImageCreateInfo *vk_info;
+	uint32_t stride;
+	bool scanout;
+};
+
+VkResult radv_image_create(VkDevice _device,
+			   const struct radv_image_create_info *info,
+			   const VkAllocationCallbacks* alloc,
+			   VkImage *pImage);
+
+void radv_image_view_init(struct radv_image_view *view,
+			  struct radv_device *device,
+			  const VkImageViewCreateInfo* pCreateInfo,
+			  struct radv_cmd_buffer *cmd_buffer,
+			  VkImageUsageFlags usage_mask);
+void radv_image_set_optimal_micro_tile_mode(struct radv_device *device,
+					    struct radv_image *image, uint32_t micro_tile_mode);
+struct radv_buffer_view {
+	struct radeon_winsys_bo *bo;
+	VkFormat vk_format;
+	uint64_t range; /**< VkBufferViewCreateInfo::range */
+	uint32_t state[4];
+};
+void radv_buffer_view_init(struct radv_buffer_view *view,
+			   struct radv_device *device,
+			   const VkBufferViewCreateInfo* pCreateInfo,
+			   struct radv_cmd_buffer *cmd_buffer);
+
+static inline struct VkExtent3D
+radv_sanitize_image_extent(const VkImageType imageType,
+			   const struct VkExtent3D imageExtent)
+{
+	switch (imageType) {
+	case VK_IMAGE_TYPE_1D:
+		return (VkExtent3D) { imageExtent.width, 1, 1 };
+	case VK_IMAGE_TYPE_2D:
+		return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
+	case VK_IMAGE_TYPE_3D:
+		return imageExtent;
+	default:
+		unreachable("invalid image type");
+	}
+}
+
+static inline struct VkOffset3D
+radv_sanitize_image_offset(const VkImageType imageType,
+			   const struct VkOffset3D imageOffset)
+{
+	switch (imageType) {
+	case VK_IMAGE_TYPE_1D:
+		return (VkOffset3D) { imageOffset.x, 0, 0 };
+	case VK_IMAGE_TYPE_2D:
+		return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
+	case VK_IMAGE_TYPE_3D:
+		return imageOffset;
+	default:
+		unreachable("invalid image type");
+	}
+}
+
+struct radv_sampler {
+	uint32_t state[4];
+};
+
+struct radv_color_buffer_info {
+	uint32_t cb_color_base;
+	uint32_t cb_color_pitch;
+	uint32_t cb_color_slice;
+	uint32_t cb_color_view;
+	uint32_t cb_color_info;
+	uint32_t cb_color_attrib;
+	uint32_t cb_dcc_control;
+	uint32_t cb_color_cmask;
+	uint32_t cb_color_cmask_slice;
+	uint32_t cb_color_fmask;
+	uint32_t cb_color_fmask_slice;
+	uint32_t cb_clear_value0;
+	uint32_t cb_clear_value1;
+	uint32_t cb_dcc_base;
+	uint32_t micro_tile_mode;
+};
+
+struct radv_ds_buffer_info {
+	uint32_t db_depth_info;
+	uint32_t db_z_info;
+	uint32_t db_stencil_info;
+	uint32_t db_z_read_base;
+	uint32_t db_stencil_read_base;
+	uint32_t db_z_write_base;
+	uint32_t db_stencil_write_base;
+	uint32_t db_depth_view;
+	uint32_t db_depth_size;
+	uint32_t db_depth_slice;
+	uint32_t db_htile_surface;
+	uint32_t db_htile_data_base;
+	uint32_t pa_su_poly_offset_db_fmt_cntl;
+	float offset_scale;
+};
+
+struct radv_attachment_info {
+	union {
+		struct radv_color_buffer_info cb;
+		struct radv_ds_buffer_info ds;
+	};
+	struct radv_image_view *attachment;
+};
+
+struct radv_framebuffer {
+	uint32_t                                     width;
+	uint32_t                                     height;
+	uint32_t                                     layers;
+
+	uint32_t                                     attachment_count;
+	struct radv_attachment_info                  attachments[0];
+};
+
+struct radv_subpass_barrier {
+	VkPipelineStageFlags src_stage_mask;
+	VkAccessFlags        src_access_mask;
+	VkAccessFlags        dst_access_mask;
+};
+
+struct radv_subpass {
+	uint32_t                                     input_count;
+	VkAttachmentReference *                      input_attachments;
+	uint32_t                                     color_count;
+	VkAttachmentReference *                      color_attachments;
+	VkAttachmentReference *                      resolve_attachments;
+	VkAttachmentReference                        depth_stencil_attachment;
+
+	/** Subpass has at least one resolve attachment */
+	bool                                         has_resolve;
+
+	struct radv_subpass_barrier                  start_barrier;
+};
+
+struct radv_render_pass_attachment {
+	VkFormat                                     format;
+	uint32_t                                     samples;
+	VkAttachmentLoadOp                           load_op;
+	VkAttachmentLoadOp                           stencil_load_op;
+	VkImageLayout                                initial_layout;
+	VkImageLayout                                final_layout;
+};
+
+struct radv_render_pass {
+	uint32_t                                     attachment_count;
+	uint32_t                                     subpass_count;
+	VkAttachmentReference *                      subpass_attachments;
+	struct radv_render_pass_attachment *         attachments;
+	struct radv_subpass_barrier                  end_barrier;
+	struct radv_subpass                          subpasses[0];
+};
+
+VkResult radv_device_init_meta(struct radv_device *device);
+void radv_device_finish_meta(struct radv_device *device);
+
+struct radv_query_pool {
+	struct radeon_winsys_bo *bo;
+	uint32_t stride;
+	uint32_t availability_offset;
+	char *ptr;
+	VkQueryType type;
+};
+
+VkResult
+radv_temp_descriptor_set_create(struct radv_device *device,
+				struct radv_cmd_buffer *cmd_buffer,
+				VkDescriptorSetLayout _layout,
+				VkDescriptorSet *_set);
+
+void
+radv_temp_descriptor_set_destroy(struct radv_device *device,
+				 VkDescriptorSet _set);
+void radv_initialise_cmask(struct radv_cmd_buffer *cmd_buffer,
+			   struct radv_image *image, uint32_t value);
+void radv_initialize_dcc(struct radv_cmd_buffer *cmd_buffer,
+			 struct radv_image *image, uint32_t value);
+#define RADV_DEFINE_HANDLE_CASTS(__radv_type, __VkType)		\
+								\
+	static inline struct __radv_type *			\
+	__radv_type ## _from_handle(__VkType _handle)		\
+	{							\
+		return (struct __radv_type *) _handle;		\
+	}							\
+								\
+	static inline __VkType					\
+	__radv_type ## _to_handle(struct __radv_type *_obj)	\
+	{							\
+		return (__VkType) _obj;				\
+	}
+
+#define RADV_DEFINE_NONDISP_HANDLE_CASTS(__radv_type, __VkType)		\
+									\
+	static inline struct __radv_type *				\
+	__radv_type ## _from_handle(__VkType _handle)			\
+	{								\
+		return (struct __radv_type *)(uintptr_t) _handle;	\
+	}								\
+									\
+	static inline __VkType						\
+	__radv_type ## _to_handle(struct __radv_type *_obj)		\
+	{								\
+		return (__VkType)(uintptr_t) _obj;			\
+	}
+
+#define RADV_FROM_HANDLE(__radv_type, __name, __handle)			\
+	struct __radv_type *__name = __radv_type ## _from_handle(__handle)
+
+RADV_DEFINE_HANDLE_CASTS(radv_cmd_buffer, VkCommandBuffer)
+RADV_DEFINE_HANDLE_CASTS(radv_device, VkDevice)
+RADV_DEFINE_HANDLE_CASTS(radv_instance, VkInstance)
+RADV_DEFINE_HANDLE_CASTS(radv_physical_device, VkPhysicalDevice)
+RADV_DEFINE_HANDLE_CASTS(radv_queue, VkQueue)
+
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_cmd_pool, VkCommandPool)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer, VkBuffer)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer_view, VkBufferView)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_pool, VkDescriptorPool)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set, VkDescriptorSet)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set_layout, VkDescriptorSetLayout)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_device_memory, VkDeviceMemory)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_fence, VkFence)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_event, VkEvent)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_framebuffer, VkFramebuffer)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image, VkImage)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image_view, VkImageView);
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_cache, VkPipelineCache)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline, VkPipeline)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_layout, VkPipelineLayout)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_query_pool, VkQueryPool)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_render_pass, VkRenderPass)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_sampler, VkSampler)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_shader_module, VkShaderModule)
+
+#define RADV_DEFINE_STRUCT_CASTS(__radv_type, __VkType)			\
+									\
+	static inline const __VkType *					\
+	__radv_type ## _to_ ## __VkType(const struct __radv_type *__radv_obj) \
+	{								\
+		return (const __VkType *) __radv_obj;			\
+	}
+
+#define RADV_COMMON_TO_STRUCT(__VkType, __vk_name, __common_name)	\
+	const __VkType *__vk_name = radv_common_to_ ## __VkType(__common_name)
+
+RADV_DEFINE_STRUCT_CASTS(radv_common, VkMemoryBarrier)
+RADV_DEFINE_STRUCT_CASTS(radv_common, VkBufferMemoryBarrier)
+RADV_DEFINE_STRUCT_CASTS(radv_common, VkImageMemoryBarrier)
+
+
diff --git a/src/amd/vulkan/radv_query.c b/src/amd/vulkan/radv_query.c
new file mode 100644
index 0000000000..f60c10a3a8
--- /dev/null
+++ b/src/amd/vulkan/radv_query.c
@@ -0,0 +1,415 @@
+/*
+ * Copyrigh 2016 Red Hat Inc.
+ * Based on anv:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <assert.h>
+#include <stdbool.h>
+#include <string.h>
+#include <unistd.h>
+#include <fcntl.h>
+
+#include "radv_private.h"
+#include "radv_cs.h"
+#include "sid.h"
+
+static unsigned get_max_db(struct radv_device *device)
+{
+	unsigned num_db = device->instance->physicalDevice.rad_info.num_render_backends;
+	unsigned rb_mask = device->instance->physicalDevice.rad_info.enabled_rb_mask;
+
+	if (device->instance->physicalDevice.rad_info.chip_class == SI)
+		num_db = 8;
+	else
+		num_db = MAX2(8, num_db);
+
+	/* Otherwise we need to change the query reset procedure */
+	assert(rb_mask == ((1ull << num_db) - 1));
+
+	return num_db;
+}
+
+VkResult radv_CreateQueryPool(
+	VkDevice                                    _device,
+	const VkQueryPoolCreateInfo*                pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkQueryPool*                                pQueryPool)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	uint64_t size;
+	struct radv_query_pool *pool = radv_alloc2(&device->alloc, pAllocator,
+					       sizeof(*pool), 8,
+					       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+	if (!pool)
+		return VK_ERROR_OUT_OF_HOST_MEMORY;
+
+
+	switch(pCreateInfo->queryType) {
+	case VK_QUERY_TYPE_OCCLUSION:
+		/* 16 bytes tmp. buffer as the compute packet writes 64 bits, but
+		 * the app. may have 32 bits of space. */
+		pool->stride = 16 * get_max_db(device) + 16;
+		break;
+	case VK_QUERY_TYPE_PIPELINE_STATISTICS:
+		pool->stride = 16 * 11;
+		break;
+	case VK_QUERY_TYPE_TIMESTAMP:
+		pool->stride = 8;
+		break;
+	default:
+		unreachable("creating unhandled query type");
+	}
+
+	pool->type = pCreateInfo->queryType;
+	pool->availability_offset = pool->stride * pCreateInfo->queryCount;
+	size = pool->availability_offset + 4 * pCreateInfo->queryCount;
+
+	pool->bo = device->ws->buffer_create(device->ws, size,
+					     64, RADEON_DOMAIN_GTT, 0);
+
+	if (!pool->bo) {
+		radv_free2(&device->alloc, pAllocator, pool);
+		return VK_ERROR_OUT_OF_DEVICE_MEMORY;
+	}
+
+	pool->ptr = device->ws->buffer_map(pool->bo);
+
+	if (!pool->ptr) {
+		device->ws->buffer_destroy(pool->bo);
+		radv_free2(&device->alloc, pAllocator, pool);
+		return VK_ERROR_OUT_OF_DEVICE_MEMORY;
+	}
+	memset(pool->ptr, 0, size);
+
+	*pQueryPool = radv_query_pool_to_handle(pool);
+	return VK_SUCCESS;
+}
+
+void radv_DestroyQueryPool(
+	VkDevice                                    _device,
+	VkQueryPool                                 _pool,
+	const VkAllocationCallbacks*                pAllocator)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(radv_query_pool, pool, _pool);
+
+	if (!pool)
+		return;
+
+	device->ws->buffer_destroy(pool->bo);
+	radv_free2(&device->alloc, pAllocator, pool);
+}
+
+VkResult radv_GetQueryPoolResults(
+	VkDevice                                    _device,
+	VkQueryPool                                 queryPool,
+	uint32_t                                    firstQuery,
+	uint32_t                                    queryCount,
+	size_t                                      dataSize,
+	void*                                       pData,
+	VkDeviceSize                                stride,
+	VkQueryResultFlags                          flags)
+{
+	RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
+	char *data = pData;
+	VkResult result = VK_SUCCESS;
+
+	for(unsigned i = 0; i < queryCount; ++i, data += stride) {
+		char *dest = data;
+		unsigned query = firstQuery + i;
+		char *src = pool->ptr + query * pool->stride;
+		uint32_t available;
+
+		if (flags & VK_QUERY_RESULT_WAIT_BIT) {
+			while(!*(volatile uint32_t*)(pool->ptr + pool->availability_offset + 4 * query))
+				;
+		}
+
+		if (!*(uint32_t*)(pool->ptr + pool->availability_offset + 4 * query) &&
+		    !(flags & VK_QUERY_RESULT_PARTIAL_BIT)) {
+			if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
+				*(uint32_t*)dest = 0;
+			result = VK_NOT_READY;
+			continue;
+
+		}
+
+		available = *(uint32_t*)(pool->ptr + pool->availability_offset + 4 * query);
+		switch (pool->type) {
+		case VK_QUERY_TYPE_TIMESTAMP:
+			if (flags & VK_QUERY_RESULT_64_BIT) {
+				*(uint64_t*)dest = *(uint64_t*)src;
+				dest += 8;
+			} else {
+				*(uint32_t*)dest = *(uint32_t*)src;
+				dest += 4;
+			}
+			break;
+		case VK_QUERY_TYPE_OCCLUSION: {
+			uint64_t result = *(uint64_t*)(src + pool->stride - 16);
+
+			if (flags & VK_QUERY_RESULT_64_BIT) {
+				*(uint64_t*)dest = result;
+				dest += 8;
+			} else {
+				*(uint32_t*)dest = result;
+				dest += 4;
+			}
+			break;
+		default:
+			unreachable("trying to get results of unhandled query type");
+		}
+		}
+
+		if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
+			*(uint32_t*)dest = available;
+			dest += 4;
+		}
+	}
+
+	return result;
+}
+
+void radv_CmdCopyQueryPoolResults(
+    VkCommandBuffer                             commandBuffer,
+    VkQueryPool                                 queryPool,
+    uint32_t                                    firstQuery,
+    uint32_t                                    queryCount,
+    VkBuffer                                    dstBuffer,
+    VkDeviceSize                                dstOffset,
+    VkDeviceSize                                stride,
+    VkQueryResultFlags                          flags)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
+	RADV_FROM_HANDLE(radv_buffer, dst_buffer, dstBuffer);
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
+	uint64_t dest_va = cmd_buffer->device->ws->buffer_get_va(dst_buffer->bo);
+	dest_va += dst_buffer->offset + dstOffset;
+
+	cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, pool->bo, 8);
+	cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, dst_buffer->bo, 8);
+
+	for(unsigned i = 0; i < queryCount; ++i, dest_va += stride) {
+		unsigned query = firstQuery + i;
+		uint64_t local_src_va = va  + query * pool->stride;
+		unsigned elem_size = (flags & VK_QUERY_RESULT_64_BIT) ? 8 : 4;
+
+		unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 26);
+
+		if (flags & VK_QUERY_RESULT_WAIT_BIT) {
+			/* TODO, not sure if there is any case where we won't always be ready yet */
+			uint64_t avail_va = va + pool->availability_offset + 4 * query;
+
+
+			/* This waits on the ME. All copies below are done on the ME */
+			radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, 0));
+			radeon_emit(cs, WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_MEM_SPACE(1));
+			radeon_emit(cs, avail_va);
+			radeon_emit(cs, avail_va >> 32);
+			radeon_emit(cs, 1); /* reference value */
+			radeon_emit(cs, 0xffffffff); /* mask */
+			radeon_emit(cs, 4); /* poll interval */
+		}
+
+		switch (pool->type) {
+		case VK_QUERY_TYPE_OCCLUSION:
+			local_src_va += pool->stride - 16;
+
+		case VK_QUERY_TYPE_TIMESTAMP:
+			radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
+			radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
+					COPY_DATA_DST_SEL(COPY_DATA_MEM) |
+					((flags & VK_QUERY_RESULT_64_BIT) ? COPY_DATA_COUNT_SEL : 0));
+			radeon_emit(cs, local_src_va);
+			radeon_emit(cs, local_src_va >> 32);
+			radeon_emit(cs, dest_va);
+			radeon_emit(cs, dest_va >> 32);
+			break;
+		default:
+			unreachable("trying to get results of unhandled query type");
+		}
+
+		/* The flag could be still changed while the data copy is busy and we
+		 * then might have invalid data, but a ready flag. However, the availability
+		 * writes happen on the ME too, so they should be synchronized. Might need to
+		 * revisit this with multiple queues.
+		 */
+		if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
+			uint64_t avail_va = va + pool->availability_offset + 4 * query;
+			uint64_t avail_dest_va = dest_va;
+			if (pool->type != VK_QUERY_TYPE_PIPELINE_STATISTICS)
+				avail_dest_va += elem_size;
+			else
+				abort();
+
+			radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
+			radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
+					COPY_DATA_DST_SEL(COPY_DATA_MEM));
+			radeon_emit(cs, avail_va);
+			radeon_emit(cs, avail_va >> 32);
+			radeon_emit(cs, avail_dest_va);
+			radeon_emit(cs, avail_dest_va >> 32);
+		}
+
+		assert(cs->cdw <= cdw_max);
+	}
+
+}
+
+void radv_CmdResetQueryPool(
+	VkCommandBuffer                             commandBuffer,
+	VkQueryPool                                 queryPool,
+	uint32_t                                    firstQuery,
+	uint32_t                                    queryCount)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
+
+	cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, pool->bo, 8);
+
+	si_cp_dma_clear_buffer(cmd_buffer, va + firstQuery * pool->stride,
+			       queryCount * pool->stride, 0);
+	si_cp_dma_clear_buffer(cmd_buffer, va + pool->availability_offset + firstQuery * 4,
+			       queryCount * 4, 0);
+}
+
+void radv_CmdBeginQuery(
+    VkCommandBuffer                             commandBuffer,
+    VkQueryPool                                 queryPool,
+    uint32_t                                    query,
+    VkQueryControlFlags                         flags)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
+	va += pool->stride * query;
+
+	cmd_buffer->device->ws->cs_add_buffer(cs, pool->bo, 8);
+
+	switch (pool->type) {
+	case VK_QUERY_TYPE_OCCLUSION:
+		radeon_check_space(cmd_buffer->device->ws, cs, 7);
+
+		++cmd_buffer->state.active_occlusion_queries;
+		if (cmd_buffer->state.active_occlusion_queries == 1)
+			radv_set_db_count_control(cmd_buffer);
+
+		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+		radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+		break;
+	default:
+		unreachable("beginning unhandled query type");
+	}
+}
+
+
+void radv_CmdEndQuery(
+    VkCommandBuffer                             commandBuffer,
+    VkQueryPool                                 queryPool,
+    uint32_t                                    query)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
+	uint64_t avail_va = va + pool->availability_offset + 4 * query;
+	va += pool->stride * query;
+
+	cmd_buffer->device->ws->cs_add_buffer(cs, pool->bo, 8);
+
+	switch (pool->type) {
+	case VK_QUERY_TYPE_OCCLUSION:
+		radeon_check_space(cmd_buffer->device->ws, cs, 14);
+
+		cmd_buffer->state.active_occlusion_queries--;
+		if (cmd_buffer->state.active_occlusion_queries == 0)
+			radv_set_db_count_control(cmd_buffer);
+
+		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+		radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
+		radeon_emit(cs, va + 8);
+		radeon_emit(cs, (va + 8) >> 32);
+
+		radeon_emit(cs, PKT3(PKT3_OCCLUSION_QUERY, 3, 0));
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+		radeon_emit(cs, va + pool->stride - 16);
+		radeon_emit(cs, (va + pool->stride - 16) >> 32);
+
+		break;
+	default:
+		unreachable("ending unhandled query type");
+	}
+
+	radeon_check_space(cmd_buffer->device->ws, cs, 5);
+
+	radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
+	radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
+		    S_370_WR_CONFIRM(1) |
+		    S_370_ENGINE_SEL(V_370_ME));
+	radeon_emit(cs, avail_va);
+	radeon_emit(cs, avail_va >> 32);
+	radeon_emit(cs, 1);
+}
+
+void radv_CmdWriteTimestamp(
+    VkCommandBuffer                             commandBuffer,
+    VkPipelineStageFlagBits                     pipelineStage,
+    VkQueryPool                                 queryPool,
+    uint32_t                                    query)
+{
+	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
+	RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+	uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
+	uint64_t avail_va = va + pool->availability_offset + 4 * query;
+	uint64_t query_va = va + pool->stride * query;
+
+	cmd_buffer->device->ws->cs_add_buffer(cs, pool->bo, 5);
+
+	unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 11);
+
+	radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+	radeon_emit(cs, EVENT_TYPE(V_028A90_BOTTOM_OF_PIPE_TS) | EVENT_INDEX(5));
+	radeon_emit(cs, query_va);
+	radeon_emit(cs, (3 << 29) | ((query_va >> 32) & 0xFFFF));
+	radeon_emit(cs, 0);
+	radeon_emit(cs, 0);
+
+	radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
+	radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
+		    S_370_WR_CONFIRM(1) |
+		    S_370_ENGINE_SEL(V_370_ME));
+	radeon_emit(cs, avail_va);
+	radeon_emit(cs, avail_va >> 32);
+	radeon_emit(cs, 1);
+
+	assert(cmd_buffer->cs->cdw <= cdw_max);
+}
diff --git a/src/amd/vulkan/radv_radeon_winsys.h b/src/amd/vulkan/radv_radeon_winsys.h
new file mode 100644
index 0000000000..29a4ee3e2d
--- /dev/null
+++ b/src/amd/vulkan/radv_radeon_winsys.h
@@ -0,0 +1,336 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * Based on radeon_winsys.h which is:
+ * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
+ * Copyright 2010 Marek Olšák <maraeo@gmail.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#pragma once
+
+#include <stdint.h>
+#include <stdbool.h>
+#include <stdlib.h>
+#include "main/macros.h"
+#include "amd_family.h"
+
+#define FREE(x) free(x)
+
+enum radeon_bo_domain { /* bitfield */
+	RADEON_DOMAIN_GTT  = 2,
+	RADEON_DOMAIN_VRAM = 4,
+	RADEON_DOMAIN_VRAM_GTT = RADEON_DOMAIN_VRAM | RADEON_DOMAIN_GTT
+};
+
+enum radeon_bo_flag { /* bitfield */
+	RADEON_FLAG_GTT_WC =        (1 << 0),
+	RADEON_FLAG_CPU_ACCESS =    (1 << 1),
+	RADEON_FLAG_NO_CPU_ACCESS = (1 << 2),
+};
+
+enum radeon_bo_usage { /* bitfield */
+	RADEON_USAGE_READ = 2,
+	RADEON_USAGE_WRITE = 4,
+	RADEON_USAGE_READWRITE = RADEON_USAGE_READ | RADEON_USAGE_WRITE
+};
+
+enum ring_type {
+	RING_GFX = 0,
+	RING_COMPUTE,
+	RING_DMA,
+	RING_UVD,
+	RING_VCE,
+	RING_LAST,
+};
+
+struct radeon_winsys_cs {
+	unsigned cdw;  /* Number of used dwords. */
+	unsigned max_dw; /* Maximum number of dwords. */
+	uint32_t *buf; /* The base pointer of the chunk. */
+};
+
+struct radeon_info {
+	/* PCI info: domain:bus:dev:func */
+	uint32_t                    pci_domain;
+	uint32_t                    pci_bus;
+	uint32_t                    pci_dev;
+	uint32_t                    pci_func;
+
+	/* Device info. */
+	uint32_t                    pci_id;
+	enum radeon_family          family;
+	const char                  *name;
+	enum chip_class             chip_class;
+	uint32_t                    gart_page_size;
+	uint64_t                    gart_size;
+	uint64_t                    vram_size;
+	bool                        has_dedicated_vram;
+	bool                     has_virtual_memory;
+	bool                        gfx_ib_pad_with_type2;
+	bool                     has_sdma;
+	bool                     has_uvd;
+	uint32_t                    vce_fw_version;
+	uint32_t                    vce_harvest_config;
+	uint32_t                    clock_crystal_freq;
+
+	/* Kernel info. */
+	uint32_t                    drm_major; /* version */
+	uint32_t                    drm_minor;
+	uint32_t                    drm_patchlevel;
+	bool                     has_userptr;
+
+	/* Shader cores. */
+	uint32_t                    r600_max_quad_pipes; /* wave size / 16 */
+	uint32_t                    max_shader_clock;
+	uint32_t                    num_good_compute_units;
+	uint32_t                    max_se; /* shader engines */
+	uint32_t                    max_sh_per_se; /* shader arrays per shader engine */
+
+	/* Render backends (color + depth blocks). */
+	uint32_t                    r300_num_gb_pipes;
+	uint32_t                    r300_num_z_pipes;
+	uint32_t                    r600_gb_backend_map; /* R600 harvest config */
+	bool                     r600_gb_backend_map_valid;
+	uint32_t                    r600_num_banks;
+	uint32_t                    num_render_backends;
+	uint32_t                    num_tile_pipes; /* pipe count from PIPE_CONFIG */
+	uint32_t                    pipe_interleave_bytes;
+	uint32_t                    enabled_rb_mask; /* GCN harvest config */
+
+	/* Tile modes. */
+	uint32_t                    si_tile_mode_array[32];
+	uint32_t                    cik_macrotile_mode_array[16];
+};
+
+#define RADEON_SURF_MAX_LEVEL                   32
+
+#define RADEON_SURF_TYPE_MASK                   0xFF
+#define RADEON_SURF_TYPE_SHIFT                  0
+#define     RADEON_SURF_TYPE_1D                     0
+#define     RADEON_SURF_TYPE_2D                     1
+#define     RADEON_SURF_TYPE_3D                     2
+#define     RADEON_SURF_TYPE_CUBEMAP                3
+#define     RADEON_SURF_TYPE_1D_ARRAY               4
+#define     RADEON_SURF_TYPE_2D_ARRAY               5
+#define RADEON_SURF_MODE_MASK                   0xFF
+#define RADEON_SURF_MODE_SHIFT                  8
+#define     RADEON_SURF_MODE_LINEAR_ALIGNED         1
+#define     RADEON_SURF_MODE_1D                     2
+#define     RADEON_SURF_MODE_2D                     3
+#define RADEON_SURF_SCANOUT                     (1 << 16)
+#define RADEON_SURF_ZBUFFER                     (1 << 17)
+#define RADEON_SURF_SBUFFER                     (1 << 18)
+#define RADEON_SURF_Z_OR_SBUFFER                (RADEON_SURF_ZBUFFER | RADEON_SURF_SBUFFER)
+#define RADEON_SURF_HAS_SBUFFER_MIPTREE         (1 << 19)
+#define RADEON_SURF_HAS_TILE_MODE_INDEX         (1 << 20)
+#define RADEON_SURF_FMASK                       (1 << 21)
+#define RADEON_SURF_DISABLE_DCC                 (1 << 22)
+
+#define RADEON_SURF_GET(v, field)   (((v) >> RADEON_SURF_ ## field ## _SHIFT) & RADEON_SURF_ ## field ## _MASK)
+#define RADEON_SURF_SET(v, field)   (((v) & RADEON_SURF_ ## field ## _MASK) << RADEON_SURF_ ## field ## _SHIFT)
+#define RADEON_SURF_CLR(v, field)   ((v) & ~(RADEON_SURF_ ## field ## _MASK << RADEON_SURF_ ## field ## _SHIFT))
+
+struct radeon_surf_level {
+	uint64_t                    offset;
+	uint64_t                    slice_size;
+	uint32_t                    npix_x;
+	uint32_t                    npix_y;
+	uint32_t                    npix_z;
+	uint32_t                    nblk_x;
+	uint32_t                    nblk_y;
+	uint32_t                    nblk_z;
+	uint32_t                    pitch_bytes;
+	uint32_t                    mode;
+	uint64_t                    dcc_offset;
+	uint64_t                    dcc_fast_clear_size;
+	bool                        dcc_enabled;
+};
+
+
+/* surface defintions from the winsys */
+struct radeon_surf {
+	/* These are inputs to the calculator. */
+	uint32_t                    npix_x;
+	uint32_t                    npix_y;
+	uint32_t                    npix_z;
+	uint32_t                    blk_w;
+	uint32_t                    blk_h;
+	uint32_t                    blk_d;
+	uint32_t                    array_size;
+	uint32_t                    last_level;
+	uint32_t                    bpe;
+	uint32_t                    nsamples;
+	uint32_t                    flags;
+
+	/* These are return values. Some of them can be set by the caller, but
+	 * they will be treated as hints (e.g. bankw, bankh) and might be
+	 * changed by the calculator.
+	 */
+	uint64_t                    bo_size;
+	uint64_t                    bo_alignment;
+	/* This applies to EG and later. */
+	uint32_t                    bankw;
+	uint32_t                    bankh;
+	uint32_t                    mtilea;
+	uint32_t                    tile_split;
+	uint32_t                    stencil_tile_split;
+	uint64_t                    stencil_offset;
+	struct radeon_surf_level    level[RADEON_SURF_MAX_LEVEL];
+	struct radeon_surf_level    stencil_level[RADEON_SURF_MAX_LEVEL];
+	uint32_t                    tiling_index[RADEON_SURF_MAX_LEVEL];
+	uint32_t                    stencil_tiling_index[RADEON_SURF_MAX_LEVEL];
+	uint32_t                    pipe_config;
+	uint32_t                    num_banks;
+	uint32_t                    macro_tile_index;
+	uint32_t                    micro_tile_mode; /* displayable, thin, depth, rotated */
+
+	/* Whether the depth miptree or stencil miptree as used by the DB are
+	 * adjusted from their TC compatible form to ensure depth/stencil
+	 * compatibility. If either is true, the corresponding plane cannot be
+	 * sampled from.
+	 */
+	bool                        depth_adjusted;
+	bool                        stencil_adjusted;
+
+	uint64_t                    dcc_size;
+	uint64_t                    dcc_alignment;
+};
+
+enum radeon_bo_layout {
+	RADEON_LAYOUT_LINEAR = 0,
+	RADEON_LAYOUT_TILED,
+	RADEON_LAYOUT_SQUARETILED,
+
+	RADEON_LAYOUT_UNKNOWN
+};
+
+/* Tiling info for display code, DRI sharing, and other data. */
+struct radeon_bo_metadata {
+	/* Tiling flags describing the texture layout for display code
+	 * and DRI sharing.
+	 */
+	enum radeon_bo_layout   microtile;
+	enum radeon_bo_layout   macrotile;
+	unsigned                pipe_config;
+	unsigned                bankw;
+	unsigned                bankh;
+	unsigned                tile_split;
+	unsigned                mtilea;
+	unsigned                num_banks;
+	unsigned                stride;
+	bool                    scanout;
+
+	/* Additional metadata associated with the buffer, in bytes.
+	 * The maximum size is 64 * 4. This is opaque for the winsys & kernel.
+	 * Supported by amdgpu only.
+	 */
+	uint32_t                size_metadata;
+	uint32_t                metadata[64];
+};
+
+struct radeon_winsys_bo;
+struct radeon_winsys_fence;
+
+struct radeon_winsys {
+	void (*destroy)(struct radeon_winsys *ws);
+
+	void (*query_info)(struct radeon_winsys *ws,
+			   struct radeon_info *info);
+
+	struct radeon_winsys_bo *(*buffer_create)(struct radeon_winsys *ws,
+						  uint64_t size,
+						  unsigned alignment,
+						  enum radeon_bo_domain domain,
+						  enum radeon_bo_flag flags);
+
+	void (*buffer_destroy)(struct radeon_winsys_bo *bo);
+	void *(*buffer_map)(struct radeon_winsys_bo *bo);
+
+	struct radeon_winsys_bo *(*buffer_from_fd)(struct radeon_winsys *ws,
+						   int fd,
+						   unsigned *stride, unsigned *offset);
+
+	bool (*buffer_get_fd)(struct radeon_winsys *ws,
+			      struct radeon_winsys_bo *bo,
+			      int *fd);
+
+	void (*buffer_unmap)(struct radeon_winsys_bo *bo);
+
+	uint64_t (*buffer_get_va)(struct radeon_winsys_bo *bo);
+
+	void (*buffer_set_metadata)(struct radeon_winsys_bo *bo,
+				    struct radeon_bo_metadata *md);
+	struct radeon_winsys_ctx *(*ctx_create)(struct radeon_winsys *ws);
+	void (*ctx_destroy)(struct radeon_winsys_ctx *ctx);
+
+	bool (*ctx_wait_idle)(struct radeon_winsys_ctx *ctx);
+
+	struct radeon_winsys_cs *(*cs_create)(struct radeon_winsys *ws,
+					      enum ring_type ring_type);
+
+	void (*cs_destroy)(struct radeon_winsys_cs *cs);
+
+	void (*cs_reset)(struct radeon_winsys_cs *cs);
+
+	bool (*cs_finalize)(struct radeon_winsys_cs *cs);
+
+	void (*cs_grow)(struct radeon_winsys_cs * cs, size_t min_size);
+
+	int (*cs_submit)(struct radeon_winsys_ctx *ctx,
+			 struct radeon_winsys_cs **cs_array,
+			 unsigned cs_count,
+			 bool can_patch,
+			 struct radeon_winsys_fence *fence);
+
+	void (*cs_add_buffer)(struct radeon_winsys_cs *cs,
+			      struct radeon_winsys_bo *bo,
+			      uint8_t priority);
+
+	void (*cs_execute_secondary)(struct radeon_winsys_cs *parent,
+				    struct radeon_winsys_cs *child);
+
+	int (*surface_init)(struct radeon_winsys *ws,
+			    struct radeon_surf *surf);
+
+	int (*surface_best)(struct radeon_winsys *ws,
+			    struct radeon_surf *surf);
+
+	struct radeon_winsys_fence *(*create_fence)();
+	void (*destroy_fence)(struct radeon_winsys_fence *fence);
+	bool (*fence_wait)(struct radeon_winsys *ws,
+			   struct radeon_winsys_fence *fence,
+			   bool absolute,
+			   uint64_t timeout);
+};
+
+static inline void radeon_emit(struct radeon_winsys_cs *cs, uint32_t value)
+{
+	cs->buf[cs->cdw++] = value;
+}
+
+static inline void radeon_emit_array(struct radeon_winsys_cs *cs,
+				     const uint32_t *values, unsigned count)
+{
+	memcpy(cs->buf + cs->cdw, values, count * 4);
+	cs->cdw += count;
+}
+
diff --git a/src/amd/vulkan/radv_util.c b/src/amd/vulkan/radv_util.c
new file mode 100644
index 0000000000..bf7abd49d1
--- /dev/null
+++ b/src/amd/vulkan/radv_util.c
@@ -0,0 +1,204 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <assert.h>
+
+#include "radv_private.h"
+
+#include "util/u_math.h"
+
+/** Log an error message.  */
+void radv_printflike(1, 2)
+	radv_loge(const char *format, ...)
+{
+	va_list va;
+
+	va_start(va, format);
+	radv_loge_v(format, va);
+	va_end(va);
+}
+
+/** \see radv_loge() */
+void
+radv_loge_v(const char *format, va_list va)
+{
+	fprintf(stderr, "vk: error: ");
+	vfprintf(stderr, format, va);
+	fprintf(stderr, "\n");
+}
+
+void radv_printflike(3, 4)
+	__radv_finishme(const char *file, int line, const char *format, ...)
+{
+	va_list ap;
+	char buffer[256];
+
+	va_start(ap, format);
+	vsnprintf(buffer, sizeof(buffer), format, ap);
+	va_end(ap);
+
+	fprintf(stderr, "%s:%d: FINISHME: %s\n", file, line, buffer);
+}
+
+void radv_noreturn radv_printflike(1, 2)
+	radv_abortf(const char *format, ...)
+{
+	va_list va;
+
+	va_start(va, format);
+	radv_abortfv(format, va);
+	va_end(va);
+}
+
+void radv_noreturn
+radv_abortfv(const char *format, va_list va)
+{
+	fprintf(stderr, "vk: error: ");
+	vfprintf(stderr, format, va);
+	fprintf(stderr, "\n");
+	abort();
+}
+
+VkResult
+__vk_errorf(VkResult error, const char *file, int line, const char *format, ...)
+{
+	va_list ap;
+	char buffer[256];
+
+#define ERROR_CASE(error) case error: error_str = #error; break;
+
+	const char *error_str;
+	switch ((int32_t)error) {
+
+		/* Core errors */
+		ERROR_CASE(VK_ERROR_OUT_OF_HOST_MEMORY)
+			ERROR_CASE(VK_ERROR_OUT_OF_DEVICE_MEMORY)
+			ERROR_CASE(VK_ERROR_INITIALIZATION_FAILED)
+			ERROR_CASE(VK_ERROR_DEVICE_LOST)
+			ERROR_CASE(VK_ERROR_MEMORY_MAP_FAILED)
+			ERROR_CASE(VK_ERROR_LAYER_NOT_PRESENT)
+			ERROR_CASE(VK_ERROR_EXTENSION_NOT_PRESENT)
+			ERROR_CASE(VK_ERROR_INCOMPATIBLE_DRIVER)
+
+			/* Extension errors */
+			ERROR_CASE(VK_ERROR_OUT_OF_DATE_KHR)
+
+	default:
+		assert(!"Unknown error");
+		error_str = "unknown error";
+	}
+
+#undef ERROR_CASE
+
+	if (format) {
+		va_start(ap, format);
+		vsnprintf(buffer, sizeof(buffer), format, ap);
+		va_end(ap);
+
+		fprintf(stderr, "%s:%d: %s (%s)\n", file, line, buffer, error_str);
+	} else {
+		fprintf(stderr, "%s:%d: %s\n", file, line, error_str);
+	}
+
+	return error;
+}
+
+int
+radv_vector_init(struct radv_vector *vector, uint32_t element_size, uint32_t size)
+{
+	assert(util_is_power_of_two(size));
+	assert(element_size < size && util_is_power_of_two(element_size));
+
+	vector->head = 0;
+	vector->tail = 0;
+	vector->element_size = element_size;
+	vector->size = size;
+	vector->data = malloc(size);
+
+	return vector->data != NULL;
+}
+
+void *
+radv_vector_add(struct radv_vector *vector)
+{
+	uint32_t offset, size, split, src_tail, dst_tail;
+	void *data;
+
+	if (vector->head - vector->tail == vector->size) {
+		size = vector->size * 2;
+		data = malloc(size);
+		if (data == NULL)
+			return NULL;
+		src_tail = vector->tail & (vector->size - 1);
+		dst_tail = vector->tail & (size - 1);
+		if (src_tail == 0) {
+			/* Since we know that the vector is full, this means that it's
+			 * linear from start to end so we can do one copy.
+			 */
+			memcpy(data + dst_tail, vector->data, vector->size);
+		} else {
+			/* In this case, the vector is split into two pieces and we have
+			 * to do two copies.  We have to be careful to make sure each
+			 * piece goes to the right locations.  Thanks to the change in
+			 * size, it may or may not still wrap around.
+			 */
+			split = align_u32(vector->tail, vector->size);
+			assert(vector->tail <= split && split < vector->head);
+			memcpy(data + dst_tail, vector->data + src_tail,
+			       split - vector->tail);
+			memcpy(data + (split & (size - 1)), vector->data,
+			       vector->head - split);
+		}
+		free(vector->data);
+		vector->data = data;
+		vector->size = size;
+	}
+
+	assert(vector->head - vector->tail < vector->size);
+
+	offset = vector->head & (vector->size - 1);
+	vector->head += vector->element_size;
+
+	return vector->data + offset;
+}
+
+void *
+radv_vector_remove(struct radv_vector *vector)
+{
+	uint32_t offset;
+
+	if (vector->head == vector->tail)
+		return NULL;
+
+	assert(vector->head - vector->tail <= vector->size);
+
+	offset = vector->tail & (vector->size - 1);
+	vector->tail += vector->element_size;
+
+	return vector->data + offset;
+}
diff --git a/src/amd/vulkan/radv_util.h b/src/amd/vulkan/radv_util.h
new file mode 100644
index 0000000000..57a4c95adb
--- /dev/null
+++ b/src/amd/vulkan/radv_util.h
@@ -0,0 +1,9 @@
+#pragma once
+
+#ifdef HAVE___BUILTIN_POPCOUNT
+#define util_bitcount(i) __builtin_popcount(i)
+#else
+extern unsigned int
+util_bitcount(unsigned int n);
+#endif
+
diff --git a/src/amd/vulkan/radv_wsi.c b/src/amd/vulkan/radv_wsi.c
new file mode 100644
index 0000000000..5922d6ab8b
--- /dev/null
+++ b/src/amd/vulkan/radv_wsi.c
@@ -0,0 +1,246 @@
+/*
+ * Copyright © 2016 Red Hat
+ * based on intel anv code:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "radv_wsi.h"
+
+
+VkResult
+radv_init_wsi(struct radv_physical_device *physical_device)
+{
+	VkResult result;
+
+	memset(physical_device->wsi, 0, sizeof(physical_device->wsi));
+
+#ifdef VK_USE_PLATFORM_XCB_KHR
+	result = radv_x11_init_wsi(physical_device);
+	if (result != VK_SUCCESS)
+		return result;
+#endif
+
+#ifdef VK_USE_PLATFORM_WAYLAND_KHR
+	result = radv_wl_init_wsi(physical_device);
+	if (result != VK_SUCCESS) {
+#ifdef VK_USE_PLATFORM_XCB_KHR
+		radv_x11_finish_wsi(physical_device);
+#endif
+		return result;
+	}
+#endif
+
+	return VK_SUCCESS;
+}
+
+void
+radv_finish_wsi(struct radv_physical_device *physical_device)
+{
+#ifdef VK_USE_PLATFORM_WAYLAND_KHR
+	radv_wl_finish_wsi(physical_device);
+#endif
+#ifdef VK_USE_PLATFORM_XCB_KHR
+	radv_x11_finish_wsi(physical_device);
+#endif
+}
+
+void radv_DestroySurfaceKHR(
+	VkInstance                                   _instance,
+	VkSurfaceKHR                                 _surface,
+	const VkAllocationCallbacks*                 pAllocator)
+{
+	RADV_FROM_HANDLE(radv_instance, instance, _instance);
+	RADV_FROM_HANDLE(_VkIcdSurfaceBase, surface, _surface);
+
+	radv_free2(&instance->alloc, pAllocator, surface);
+}
+
+VkResult radv_GetPhysicalDeviceSurfaceSupportKHR(
+	VkPhysicalDevice                            physicalDevice,
+	uint32_t                                    queueFamilyIndex,
+	VkSurfaceKHR                                _surface,
+	VkBool32*                                   pSupported)
+{
+	RADV_FROM_HANDLE(radv_physical_device, device, physicalDevice);
+	RADV_FROM_HANDLE(_VkIcdSurfaceBase, surface, _surface);
+	struct radv_wsi_interface *iface = device->wsi[surface->platform];
+
+	return iface->get_support(surface, device, queueFamilyIndex, pSupported);
+}
+
+VkResult radv_GetPhysicalDeviceSurfaceCapabilitiesKHR(
+	VkPhysicalDevice                            physicalDevice,
+	VkSurfaceKHR                                _surface,
+	VkSurfaceCapabilitiesKHR*                   pSurfaceCapabilities)
+{
+	RADV_FROM_HANDLE(radv_physical_device, device, physicalDevice);
+	RADV_FROM_HANDLE(_VkIcdSurfaceBase, surface, _surface);
+	struct radv_wsi_interface *iface = device->wsi[surface->platform];
+
+	return iface->get_capabilities(surface, device, pSurfaceCapabilities);
+}
+
+VkResult radv_GetPhysicalDeviceSurfaceFormatsKHR(
+	VkPhysicalDevice                            physicalDevice,
+	VkSurfaceKHR                                _surface,
+	uint32_t*                                   pSurfaceFormatCount,
+	VkSurfaceFormatKHR*                         pSurfaceFormats)
+{
+	RADV_FROM_HANDLE(radv_physical_device, device, physicalDevice);
+	RADV_FROM_HANDLE(_VkIcdSurfaceBase, surface, _surface);
+	struct radv_wsi_interface *iface = device->wsi[surface->platform];
+
+	return iface->get_formats(surface, device, pSurfaceFormatCount,
+				  pSurfaceFormats);
+}
+
+VkResult radv_GetPhysicalDeviceSurfacePresentModesKHR(
+	VkPhysicalDevice                            physicalDevice,
+	VkSurfaceKHR                                _surface,
+	uint32_t*                                   pPresentModeCount,
+	VkPresentModeKHR*                           pPresentModes)
+{
+	RADV_FROM_HANDLE(radv_physical_device, device, physicalDevice);
+	RADV_FROM_HANDLE(_VkIcdSurfaceBase, surface, _surface);
+	struct radv_wsi_interface *iface = device->wsi[surface->platform];
+
+	return iface->get_present_modes(surface, device, pPresentModeCount,
+					pPresentModes);
+}
+
+VkResult radv_CreateSwapchainKHR(
+	VkDevice                                     _device,
+	const VkSwapchainCreateInfoKHR*              pCreateInfo,
+	const VkAllocationCallbacks*                 pAllocator,
+	VkSwapchainKHR*                              pSwapchain)
+{
+	RADV_FROM_HANDLE(radv_device, device, _device);
+	RADV_FROM_HANDLE(_VkIcdSurfaceBase, surface, pCreateInfo->surface);
+	struct radv_wsi_interface *iface =
+		device->instance->physicalDevice.wsi[surface->platform];
+	struct radv_swapchain *swapchain;
+
+	VkResult result = iface->create_swapchain(surface, device, pCreateInfo,
+						  pAllocator, &swapchain);
+	if (result != VK_SUCCESS)
+		return result;
+
+	if (pAllocator)
+		swapchain->alloc = *pAllocator;
+	else
+		swapchain->alloc = device->alloc;
+
+	for (unsigned i = 0; i < ARRAY_SIZE(swapchain->fences); i++)
+		swapchain->fences[i] = VK_NULL_HANDLE;
+
+	*pSwapchain = radv_swapchain_to_handle(swapchain);
+
+	return VK_SUCCESS;
+}
+
+void radv_DestroySwapchainKHR(
+	VkDevice                                     device,
+	VkSwapchainKHR                               _swapchain,
+	const VkAllocationCallbacks*                 pAllocator)
+{
+	RADV_FROM_HANDLE(radv_swapchain, swapchain, _swapchain);
+
+	for (unsigned i = 0; i < ARRAY_SIZE(swapchain->fences); i++) {
+		if (swapchain->fences[i] != VK_NULL_HANDLE)
+			radv_DestroyFence(device, swapchain->fences[i], pAllocator);
+	}
+
+	swapchain->destroy(swapchain, pAllocator);
+}
+
+VkResult radv_GetSwapchainImagesKHR(
+	VkDevice                                     device,
+	VkSwapchainKHR                               _swapchain,
+	uint32_t*                                    pSwapchainImageCount,
+	VkImage*                                     pSwapchainImages)
+{
+	RADV_FROM_HANDLE(radv_swapchain, swapchain, _swapchain);
+
+	return swapchain->get_images(swapchain, pSwapchainImageCount,
+				     pSwapchainImages);
+}
+
+VkResult radv_AcquireNextImageKHR(
+	VkDevice                                     device,
+	VkSwapchainKHR                               _swapchain,
+	uint64_t                                     timeout,
+	VkSemaphore                                  semaphore,
+	VkFence                                      fence,
+	uint32_t*                                    pImageIndex)
+{
+	RADV_FROM_HANDLE(radv_swapchain, swapchain, _swapchain);
+
+	return swapchain->acquire_next_image(swapchain, timeout, semaphore,
+					     pImageIndex);
+}
+
+VkResult radv_QueuePresentKHR(
+	VkQueue                                  _queue,
+	const VkPresentInfoKHR*                  pPresentInfo)
+{
+	RADV_FROM_HANDLE(radv_queue, queue, _queue);
+	VkResult result = VK_SUCCESS;
+
+	for (uint32_t i = 0; i < pPresentInfo->swapchainCount; i++) {
+		RADV_FROM_HANDLE(radv_swapchain, swapchain, pPresentInfo->pSwapchains[i]);
+
+		assert(swapchain->device == queue->device);
+		if (swapchain->fences[0] == VK_NULL_HANDLE) {
+			result = radv_CreateFence(radv_device_to_handle(queue->device),
+						  &(VkFenceCreateInfo) {
+							  .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
+								  .flags = 0,
+								  }, &swapchain->alloc, &swapchain->fences[0]);
+			if (result != VK_SUCCESS)
+				return result;
+		} else {
+			radv_ResetFences(radv_device_to_handle(queue->device),
+					 1, &swapchain->fences[0]);
+		}
+
+		radv_QueueSubmit(_queue, 0, NULL, swapchain->fences[0]);
+
+		result = swapchain->queue_present(swapchain, queue,
+						  pPresentInfo->pImageIndices[i]);
+		/* TODO: What if one of them returns OUT_OF_DATE? */
+		if (result != VK_SUCCESS)
+			return result;
+
+		VkFence last = swapchain->fences[2];
+		swapchain->fences[2] = swapchain->fences[1];
+		swapchain->fences[1] = swapchain->fences[0];
+		swapchain->fences[0] = last;
+
+		if (last != VK_NULL_HANDLE) {
+			radv_WaitForFences(radv_device_to_handle(queue->device),
+					   1, &last, true, 1);
+		}
+
+	}
+
+	return VK_SUCCESS;
+}
diff --git a/src/amd/vulkan/radv_wsi.h b/src/amd/vulkan/radv_wsi.h
new file mode 100644
index 0000000000..3e453fbe77
--- /dev/null
+++ b/src/amd/vulkan/radv_wsi.h
@@ -0,0 +1,79 @@
+/*
+ * Copyright © 2016 Red Hat
+ * based on intel anv code:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#pragma once
+
+#include "radv_private.h"
+
+struct radv_swapchain;
+struct radv_wsi_interface {
+	VkResult (*get_support)(VkIcdSurfaceBase *surface,
+				struct radv_physical_device *device,
+				uint32_t queueFamilyIndex,
+				VkBool32* pSupported);
+	VkResult (*get_capabilities)(VkIcdSurfaceBase *surface,
+				     struct radv_physical_device *device,
+				     VkSurfaceCapabilitiesKHR* pSurfaceCapabilities);
+	VkResult (*get_formats)(VkIcdSurfaceBase *surface,
+				struct radv_physical_device *device,
+				uint32_t* pSurfaceFormatCount,
+				VkSurfaceFormatKHR* pSurfaceFormats);
+	VkResult (*get_present_modes)(VkIcdSurfaceBase *surface,
+				      struct radv_physical_device *device,
+				      uint32_t* pPresentModeCount,
+				      VkPresentModeKHR* pPresentModes);
+	VkResult (*create_swapchain)(VkIcdSurfaceBase *surface,
+				     struct radv_device *device,
+				     const VkSwapchainCreateInfoKHR* pCreateInfo,
+				     const VkAllocationCallbacks* pAllocator,
+				     struct radv_swapchain **swapchain);
+};
+
+struct radv_swapchain {
+	struct radv_device *device;
+
+	VkAllocationCallbacks alloc;
+
+	VkFence fences[3];
+
+	VkResult (*destroy)(struct radv_swapchain *swapchain,
+			    const VkAllocationCallbacks *pAllocator);
+	VkResult (*get_images)(struct radv_swapchain *swapchain,
+			       uint32_t *pCount, VkImage *pSwapchainImages);
+	VkResult (*acquire_next_image)(struct radv_swapchain *swap_chain,
+				       uint64_t timeout, VkSemaphore semaphore,
+				       uint32_t *image_index);
+	VkResult (*queue_present)(struct radv_swapchain *swap_chain,
+				  struct radv_queue *queue,
+				  uint32_t image_index);
+};
+
+RADV_DEFINE_NONDISP_HANDLE_CASTS(_VkIcdSurfaceBase, VkSurfaceKHR)
+RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_swapchain, VkSwapchainKHR)
+
+VkResult radv_x11_init_wsi(struct radv_physical_device *physical_device);
+void radv_x11_finish_wsi(struct radv_physical_device *physical_device);
+VkResult radv_wl_init_wsi(struct radv_physical_device *physical_device);
+void radv_wl_finish_wsi(struct radv_physical_device *physical_device);
diff --git a/src/amd/vulkan/radv_wsi_wayland.c b/src/amd/vulkan/radv_wsi_wayland.c
new file mode 100644
index 0000000000..2b4a3d32a4
--- /dev/null
+++ b/src/amd/vulkan/radv_wsi_wayland.c
@@ -0,0 +1,880 @@
+/*
+ * Copyright © 2016 Red Hat
+ * based on intel anv code:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <wayland-client.h>
+#include <wayland-drm-client-protocol.h>
+
+#include "radv_wsi.h"
+
+#include "vk_format.h"
+#include <util/hash_table.h>
+
+#define MIN_NUM_IMAGES 2
+
+struct wsi_wl_display {
+	struct radv_physical_device                  *physical_device;
+	struct wl_display *                          display;
+	struct wl_drm *                              drm;
+
+	/* Vector of VkFormats supported */
+	struct radv_vector                            formats;
+
+	uint32_t                                     capabilities;
+};
+
+struct wsi_wayland {
+	struct radv_wsi_interface                     base;
+
+	struct radv_physical_device *                 physical_device;
+
+	pthread_mutex_t                              mutex;
+	/* Hash table of wl_display -> wsi_wl_display mappings */
+	struct hash_table *                          displays;
+};
+
+static void
+wsi_wl_display_add_vk_format(struct wsi_wl_display *display, VkFormat format)
+{
+	/* Don't add a format that's already in the list */
+	VkFormat *f;
+	radv_vector_foreach(f, &display->formats)
+		if (*f == format)
+			return;
+
+	/* Don't add formats that aren't renderable. */
+	VkFormatProperties props;
+	radv_GetPhysicalDeviceFormatProperties(
+		radv_physical_device_to_handle(display->physical_device), format, &props);
+	if (!(props.optimalTilingFeatures & VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT))
+		return;
+
+	f = radv_vector_add(&display->formats);
+	if (f)
+		*f = format;
+}
+
+static void
+drm_handle_device(void *data, struct wl_drm *drm, const char *name)
+{
+	fprintf(stderr, "wl_drm.device(%s)\n", name);
+}
+
+static uint32_t
+wl_drm_format_for_vk_format(VkFormat vk_format, bool alpha)
+{
+	switch (vk_format) {
+		/* TODO: Figure out what all the formats mean and make this table
+		 * correct.
+		 */
+#if 0
+	case VK_FORMAT_R4G4B4A4_UNORM:
+		return alpha ? WL_DRM_FORMAT_ABGR4444 : WL_DRM_FORMAT_XBGR4444;
+	case VK_FORMAT_R5G6B5_UNORM:
+		return WL_DRM_FORMAT_BGR565;
+	case VK_FORMAT_R5G5B5A1_UNORM:
+		return alpha ? WL_DRM_FORMAT_ABGR1555 : WL_DRM_FORMAT_XBGR1555;
+	case VK_FORMAT_R8G8B8_UNORM:
+		return WL_DRM_FORMAT_XBGR8888;
+	case VK_FORMAT_R8G8B8A8_UNORM:
+		return alpha ? WL_DRM_FORMAT_ABGR8888 : WL_DRM_FORMAT_XBGR8888;
+	case VK_FORMAT_R10G10B10A2_UNORM:
+		return alpha ? WL_DRM_FORMAT_ABGR2101010 : WL_DRM_FORMAT_XBGR2101010;
+	case VK_FORMAT_B4G4R4A4_UNORM:
+		return alpha ? WL_DRM_FORMAT_ARGB4444 : WL_DRM_FORMAT_XRGB4444;
+	case VK_FORMAT_B5G6R5_UNORM:
+		return WL_DRM_FORMAT_RGB565;
+	case VK_FORMAT_B5G5R5A1_UNORM:
+		return alpha ? WL_DRM_FORMAT_XRGB1555 : WL_DRM_FORMAT_XRGB1555;
+#endif
+	case VK_FORMAT_B8G8R8_SRGB:
+		return WL_DRM_FORMAT_BGRX8888;
+	case VK_FORMAT_B8G8R8A8_SRGB:
+		return alpha ? WL_DRM_FORMAT_ARGB8888 : WL_DRM_FORMAT_XRGB8888;
+#if 0
+	case VK_FORMAT_B10G10R10A2_UNORM:
+		return alpha ? WL_DRM_FORMAT_ARGB2101010 : WL_DRM_FORMAT_XRGB2101010;
+#endif
+
+	default:
+		assert(!"Unsupported Vulkan format");
+		return 0;
+	}
+}
+
+static void
+drm_handle_format(void *data, struct wl_drm *drm, uint32_t wl_format)
+{
+	struct wsi_wl_display *display = data;
+
+	switch (wl_format) {
+#if 0
+	case WL_DRM_FORMAT_ABGR4444:
+	case WL_DRM_FORMAT_XBGR4444:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_R4G4B4A4_UNORM);
+		break;
+	case WL_DRM_FORMAT_BGR565:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_R5G6B5_UNORM);
+		break;
+	case WL_DRM_FORMAT_ABGR1555:
+	case WL_DRM_FORMAT_XBGR1555:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_R5G5B5A1_UNORM);
+		break;
+	case WL_DRM_FORMAT_XBGR8888:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_R8G8B8_UNORM);
+		/* fallthrough */
+	case WL_DRM_FORMAT_ABGR8888:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_R8G8B8A8_UNORM);
+		break;
+	case WL_DRM_FORMAT_ABGR2101010:
+	case WL_DRM_FORMAT_XBGR2101010:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_R10G10B10A2_UNORM);
+		break;
+	case WL_DRM_FORMAT_ARGB4444:
+	case WL_DRM_FORMAT_XRGB4444:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_B4G4R4A4_UNORM);
+		break;
+	case WL_DRM_FORMAT_RGB565:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_B5G6R5_UNORM);
+		break;
+	case WL_DRM_FORMAT_ARGB1555:
+	case WL_DRM_FORMAT_XRGB1555:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_B5G5R5A1_UNORM);
+		break;
+#endif
+	case WL_DRM_FORMAT_XRGB8888:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_B8G8R8_SRGB);
+		/* fallthrough */
+	case WL_DRM_FORMAT_ARGB8888:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_B8G8R8A8_SRGB);
+		break;
+#if 0
+	case WL_DRM_FORMAT_ARGB2101010:
+	case WL_DRM_FORMAT_XRGB2101010:
+		wsi_wl_display_add_vk_format(display, VK_FORMAT_B10G10R10A2_UNORM);
+		break;
+#endif
+	}
+}
+
+static void
+drm_handle_authenticated(void *data, struct wl_drm *drm)
+{
+}
+
+static void
+drm_handle_capabilities(void *data, struct wl_drm *drm, uint32_t capabilities)
+{
+	struct wsi_wl_display *display = data;
+
+	display->capabilities = capabilities;
+}
+
+static const struct wl_drm_listener drm_listener = {
+	drm_handle_device,
+	drm_handle_format,
+	drm_handle_authenticated,
+	drm_handle_capabilities,
+};
+
+static void
+registry_handle_global(void *data, struct wl_registry *registry,
+                       uint32_t name, const char *interface, uint32_t version)
+{
+	struct wsi_wl_display *display = data;
+
+	if (strcmp(interface, "wl_drm") == 0) {
+		assert(display->drm == NULL);
+
+		assert(version >= 2);
+		display->drm = wl_registry_bind(registry, name, &wl_drm_interface, 2);
+
+		if (display->drm)
+			wl_drm_add_listener(display->drm, &drm_listener, display);
+	}
+}
+
+static void
+registry_handle_global_remove(void *data, struct wl_registry *registry,
+                              uint32_t name)
+{ /* No-op */ }
+
+static const struct wl_registry_listener registry_listener = {
+	registry_handle_global,
+	registry_handle_global_remove
+};
+
+static void
+wsi_wl_display_destroy(struct wsi_wayland *wsi, struct wsi_wl_display *display)
+{
+	radv_vector_finish(&display->formats);
+	if (display->drm)
+		wl_drm_destroy(display->drm);
+	radv_free(&wsi->physical_device->instance->alloc, display);
+}
+
+static struct wsi_wl_display *
+wsi_wl_display_create(struct wsi_wayland *wsi, struct wl_display *wl_display)
+{
+	struct wsi_wl_display *display =
+		radv_alloc(&wsi->physical_device->instance->alloc, sizeof(*display), 8,
+			   VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
+	if (!display)
+		return NULL;
+
+	memset(display, 0, sizeof(*display));
+
+	display->display = wl_display;
+	display->physical_device = wsi->physical_device;
+
+	if (!radv_vector_init(&display->formats, sizeof(VkFormat), 8))
+		goto fail;
+
+	struct wl_registry *registry = wl_display_get_registry(wl_display);
+	if (!registry)
+		return NULL;
+
+	wl_registry_add_listener(registry, &registry_listener, display);
+
+	/* Round-rip to get the wl_drm global */
+	wl_display_roundtrip(wl_display);
+
+	if (!display->drm)
+		goto fail;
+
+	/* Round-rip to get wl_drm formats and capabilities */
+	wl_display_roundtrip(wl_display);
+
+	/* We need prime support */
+	if (!(display->capabilities & WL_DRM_CAPABILITY_PRIME))
+		goto fail;
+
+	/* We don't need this anymore */
+	wl_registry_destroy(registry);
+
+	return display;
+
+fail:
+	if (registry)
+		wl_registry_destroy(registry);
+
+	wsi_wl_display_destroy(wsi, display);
+	return NULL;
+}
+
+static struct wsi_wl_display *
+wsi_wl_get_display(struct radv_physical_device *device,
+                   struct wl_display *wl_display)
+{
+	struct wsi_wayland *wsi =
+		(struct wsi_wayland *)device->wsi[VK_ICD_WSI_PLATFORM_WAYLAND];
+
+	pthread_mutex_lock(&wsi->mutex);
+
+	struct hash_entry *entry = _mesa_hash_table_search(wsi->displays,
+							   wl_display);
+	if (!entry) {
+		/* We're about to make a bunch of blocking calls.  Let's drop the
+		 * mutex for now so we don't block up too badly.
+		 */
+		pthread_mutex_unlock(&wsi->mutex);
+
+		struct wsi_wl_display *display = wsi_wl_display_create(wsi, wl_display);
+
+		pthread_mutex_lock(&wsi->mutex);
+
+		entry = _mesa_hash_table_search(wsi->displays, wl_display);
+		if (entry) {
+			/* Oops, someone raced us to it */
+			wsi_wl_display_destroy(wsi, display);
+		} else {
+			entry = _mesa_hash_table_insert(wsi->displays, wl_display, display);
+		}
+	}
+
+	pthread_mutex_unlock(&wsi->mutex);
+
+	return entry->data;
+}
+
+VkBool32 radv_GetPhysicalDeviceWaylandPresentationSupportKHR(
+	VkPhysicalDevice                            physicalDevice,
+	uint32_t                                    queueFamilyIndex,
+	struct wl_display*                          display)
+{
+	RADV_FROM_HANDLE(radv_physical_device, physical_device, physicalDevice);
+
+	return wsi_wl_get_display(physical_device, display) != NULL;
+}
+
+static VkResult
+wsi_wl_surface_get_support(VkIcdSurfaceBase *surface,
+                           struct radv_physical_device *device,
+                           uint32_t queueFamilyIndex,
+                           VkBool32* pSupported)
+{
+	*pSupported = true;
+
+	return VK_SUCCESS;
+}
+
+static const VkPresentModeKHR present_modes[] = {
+	VK_PRESENT_MODE_MAILBOX_KHR,
+	VK_PRESENT_MODE_FIFO_KHR,
+};
+
+static VkResult
+wsi_wl_surface_get_capabilities(VkIcdSurfaceBase *surface,
+                                struct radv_physical_device *device,
+                                VkSurfaceCapabilitiesKHR* caps)
+{
+	caps->minImageCount = MIN_NUM_IMAGES;
+	caps->maxImageCount = 4;
+	caps->currentExtent = (VkExtent2D) { -1, -1 };
+	caps->minImageExtent = (VkExtent2D) { 1, 1 };
+	caps->maxImageExtent = (VkExtent2D) { INT16_MAX, INT16_MAX };
+	caps->supportedTransforms = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
+	caps->currentTransform = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
+	caps->maxImageArrayLayers = 1;
+
+	caps->supportedCompositeAlpha =
+		VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR |
+		VK_COMPOSITE_ALPHA_PRE_MULTIPLIED_BIT_KHR;
+
+	caps->supportedUsageFlags =
+		VK_IMAGE_USAGE_TRANSFER_SRC_BIT |
+		VK_IMAGE_USAGE_SAMPLED_BIT |
+		VK_IMAGE_USAGE_TRANSFER_DST_BIT |
+		VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+wsi_wl_surface_get_formats(VkIcdSurfaceBase *icd_surface,
+                           struct radv_physical_device *device,
+                           uint32_t* pSurfaceFormatCount,
+                           VkSurfaceFormatKHR* pSurfaceFormats)
+{
+	VkIcdSurfaceWayland *surface = (VkIcdSurfaceWayland *)icd_surface;
+	struct wsi_wl_display *display =
+		wsi_wl_get_display(device, surface->display);
+
+	uint32_t count = radv_vector_length(&display->formats);
+
+	if (pSurfaceFormats == NULL) {
+		*pSurfaceFormatCount = count;
+		return VK_SUCCESS;
+	}
+
+	assert(*pSurfaceFormatCount >= count);
+	*pSurfaceFormatCount = count;
+
+	VkFormat *f;
+	radv_vector_foreach(f, &display->formats) {
+		*(pSurfaceFormats++) = (VkSurfaceFormatKHR) {
+			.format = *f,
+			/* TODO: We should get this from the compositor somehow */
+			.colorSpace = VK_COLORSPACE_SRGB_NONLINEAR_KHR,
+		};
+	}
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+wsi_wl_surface_get_present_modes(VkIcdSurfaceBase *surface,
+                                 struct radv_physical_device *device,
+                                 uint32_t* pPresentModeCount,
+                                 VkPresentModeKHR* pPresentModes)
+{
+	if (pPresentModes == NULL) {
+		*pPresentModeCount = ARRAY_SIZE(present_modes);
+		return VK_SUCCESS;
+	}
+
+	assert(*pPresentModeCount >= ARRAY_SIZE(present_modes));
+	typed_memcpy(pPresentModes, present_modes, *pPresentModeCount);
+	*pPresentModeCount = ARRAY_SIZE(present_modes);
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+wsi_wl_surface_create_swapchain(VkIcdSurfaceBase *surface,
+                                struct radv_device *device,
+                                const VkSwapchainCreateInfoKHR* pCreateInfo,
+                                const VkAllocationCallbacks* pAllocator,
+                                struct radv_swapchain **swapchain);
+
+VkResult radv_CreateWaylandSurfaceKHR(
+	VkInstance                                  _instance,
+	const VkWaylandSurfaceCreateInfoKHR*        pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkSurfaceKHR*                               pSurface)
+{
+	RADV_FROM_HANDLE(radv_instance, instance, _instance);
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_WAYLAND_SURFACE_CREATE_INFO_KHR);
+
+	VkIcdSurfaceWayland *surface;
+
+	surface = radv_alloc2(&instance->alloc, pAllocator, sizeof *surface, 8,
+			      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (surface == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	surface->base.platform = VK_ICD_WSI_PLATFORM_WAYLAND;
+	surface->display = pCreateInfo->display;
+	surface->surface = pCreateInfo->surface;
+
+	*pSurface = _VkIcdSurfaceBase_to_handle(&surface->base);
+
+	return VK_SUCCESS;
+}
+
+struct wsi_wl_image {
+	struct radv_image *                           image;
+	struct radv_device_memory *                   memory;
+	struct wl_buffer *                           buffer;
+	bool                                         busy;
+};
+
+struct wsi_wl_swapchain {
+	struct radv_swapchain                        base;
+
+	struct wsi_wl_display *                      display;
+	struct wl_event_queue *                      queue;
+	struct wl_surface *                          surface;
+
+	VkExtent2D                                   extent;
+	VkFormat                                     vk_format;
+	uint32_t                                     drm_format;
+
+	VkPresentModeKHR                             present_mode;
+	bool                                         fifo_ready;
+
+	uint32_t                                     image_count;
+	struct wsi_wl_image                          images[0];
+};
+
+static VkResult
+wsi_wl_swapchain_get_images(struct radv_swapchain *radv_chain,
+                            uint32_t *pCount, VkImage *pSwapchainImages)
+{
+	struct wsi_wl_swapchain *chain = (struct wsi_wl_swapchain *)radv_chain;
+
+	if (pSwapchainImages == NULL) {
+		*pCount = chain->image_count;
+		return VK_SUCCESS;
+	}
+
+	assert(chain->image_count <= *pCount);
+	for (uint32_t i = 0; i < chain->image_count; i++)
+		pSwapchainImages[i] = radv_image_to_handle(chain->images[i].image);
+
+	*pCount = chain->image_count;
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+wsi_wl_swapchain_acquire_next_image(struct radv_swapchain *radv_chain,
+                                    uint64_t timeout,
+                                    VkSemaphore semaphore,
+                                    uint32_t *image_index)
+{
+	struct wsi_wl_swapchain *chain = (struct wsi_wl_swapchain *)radv_chain;
+
+	int ret = wl_display_dispatch_queue_pending(chain->display->display,
+						    chain->queue);
+	/* XXX: I'm not sure if out-of-date is the right error here.  If
+	 * wl_display_dispatch_queue_pending fails it most likely means we got
+	 * kicked by the server so this seems more-or-less correct.
+	 */
+	if (ret < 0)
+		return vk_error(VK_ERROR_OUT_OF_DATE_KHR);
+
+	while (1) {
+		for (uint32_t i = 0; i < chain->image_count; i++) {
+			if (!chain->images[i].busy) {
+				/* We found a non-busy image */
+				*image_index = i;
+				return VK_SUCCESS;
+			}
+		}
+
+		/* This time we do a blocking dispatch because we can't go
+		 * anywhere until we get an event.
+		 */
+		int ret = wl_display_roundtrip_queue(chain->display->display,
+						     chain->queue);
+		if (ret < 0)
+			return vk_error(VK_ERROR_OUT_OF_DATE_KHR);
+	}
+}
+
+static void
+frame_handle_done(void *data, struct wl_callback *callback, uint32_t serial)
+{
+	struct wsi_wl_swapchain *chain = data;
+
+	chain->fifo_ready = true;
+
+	wl_callback_destroy(callback);
+}
+
+static const struct wl_callback_listener frame_listener = {
+	frame_handle_done,
+};
+
+static VkResult
+wsi_wl_swapchain_queue_present(struct radv_swapchain *radv_chain,
+                               struct radv_queue *queue,
+                               uint32_t image_index)
+{
+	struct wsi_wl_swapchain *chain = (struct wsi_wl_swapchain *)radv_chain;
+
+	if (chain->present_mode == VK_PRESENT_MODE_FIFO_KHR) {
+		while (!chain->fifo_ready) {
+			int ret = wl_display_dispatch_queue(chain->display->display,
+							    chain->queue);
+			if (ret < 0)
+				return vk_error(VK_ERROR_OUT_OF_DATE_KHR);
+		}
+	}
+
+	assert(image_index < chain->image_count);
+	wl_surface_attach(chain->surface, chain->images[image_index].buffer, 0, 0);
+	wl_surface_damage(chain->surface, 0, 0, INT32_MAX, INT32_MAX);
+
+	if (chain->present_mode == VK_PRESENT_MODE_FIFO_KHR) {
+		struct wl_callback *frame = wl_surface_frame(chain->surface);
+		wl_proxy_set_queue((struct wl_proxy *)frame, chain->queue);
+		wl_callback_add_listener(frame, &frame_listener, chain);
+		chain->fifo_ready = false;
+	}
+
+	chain->images[image_index].busy = true;
+	wl_surface_commit(chain->surface);
+	wl_display_flush(chain->display->display);
+
+	return VK_SUCCESS;
+}
+
+static void
+wsi_wl_image_finish(struct wsi_wl_swapchain *chain, struct wsi_wl_image *image,
+                    const VkAllocationCallbacks* pAllocator)
+{
+	VkDevice vk_device = radv_device_to_handle(chain->base.device);
+	radv_FreeMemory(vk_device, radv_device_memory_to_handle(image->memory),
+			pAllocator);
+	radv_DestroyImage(vk_device, radv_image_to_handle(image->image),
+			  pAllocator);
+}
+
+static void
+buffer_handle_release(void *data, struct wl_buffer *buffer)
+{
+	struct wsi_wl_image *image = data;
+
+	assert(image->buffer == buffer);
+
+	image->busy = false;
+}
+
+static const struct wl_buffer_listener buffer_listener = {
+	buffer_handle_release,
+};
+
+static VkResult
+wsi_wl_image_init(struct wsi_wl_swapchain *chain,
+                  struct wsi_wl_image *image,
+                  const VkSwapchainCreateInfoKHR *pCreateInfo,
+                  const VkAllocationCallbacks* pAllocator)
+{
+	VkDevice vk_device = radv_device_to_handle(chain->base.device);
+	VkResult result;
+	bool bret;
+	VkImage vk_image;
+	struct radeon_surf *surface;
+	int fd;
+	result = radv_image_create(vk_device,
+				   &(struct radv_image_create_info) {
+					   .vk_info =
+						   &(VkImageCreateInfo) {
+						   .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
+						   .imageType = VK_IMAGE_TYPE_2D,
+						   .format = chain->vk_format,
+						   .extent = {
+							   .width = chain->extent.width,
+							   .height = chain->extent.height,
+							   .depth = 1
+						   },
+						   .mipLevels = 1,
+						   .arrayLayers = 1,
+						   .samples = 1,
+						   /* FIXME: Need a way to use X tiling to allow scanout */
+						   .tiling = VK_IMAGE_TILING_OPTIMAL,
+						   .usage = (VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT |
+							     pCreateInfo->imageUsage),
+						   .flags = 0,
+					   },
+						   .scanout = true},
+				   pAllocator,
+				   &vk_image);
+
+	if (result != VK_SUCCESS)
+		return result;
+
+	image->image = radv_image_from_handle(vk_image);
+	assert(vk_format_is_color(image->image->vk_format));
+
+	VkDeviceMemory vk_memory;
+	result = radv_AllocateMemory(vk_device,
+				     &(VkMemoryAllocateInfo) {
+					     .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
+						     .allocationSize = image->image->size,
+						     .memoryTypeIndex = 0,
+						     },
+				     pAllocator,
+				     &vk_memory);
+
+	if (result != VK_SUCCESS)
+		goto fail_image;
+
+	image->memory = radv_device_memory_from_handle(vk_memory);
+
+	result = radv_BindImageMemory(vk_device, vk_image, vk_memory, 0);
+
+	if (result != VK_SUCCESS)
+		goto fail_mem;
+
+	bret = chain->base.device->ws->buffer_get_fd(chain->base.device->ws,
+						     image->memory->bo, &fd);
+	if (bret == false)
+		goto fail_mem;
+
+	{
+		struct radeon_bo_metadata metadata;
+		radv_init_metadata(chain->base.device, image->image, &metadata);
+		chain->base.device->ws->buffer_set_metadata(image->memory->bo, &metadata);
+	}
+	surface = &image->image->surface;
+
+	image->buffer = wl_drm_create_prime_buffer(chain->display->drm,
+						   fd, /* name */
+						   chain->extent.width,
+						   chain->extent.height,
+						   chain->drm_format,
+						   surface->level[0].offset,
+						   surface->level[0].pitch_bytes,
+						   0, 0, 0, 0 /* unused */);
+	wl_display_roundtrip(chain->display->display);
+	close(fd);
+
+	wl_proxy_set_queue((struct wl_proxy *)image->buffer, chain->queue);
+	wl_buffer_add_listener(image->buffer, &buffer_listener, image);
+
+	return VK_SUCCESS;
+
+fail_mem:
+	radv_FreeMemory(vk_device, vk_memory, pAllocator);
+fail_image:
+	radv_DestroyImage(vk_device, vk_image, pAllocator);
+
+	return result;
+}
+
+static VkResult
+wsi_wl_swapchain_destroy(struct radv_swapchain *radv_chain,
+                         const VkAllocationCallbacks *pAllocator)
+{
+	struct wsi_wl_swapchain *chain = (struct wsi_wl_swapchain *)radv_chain;
+
+	for (uint32_t i = 0; i < chain->image_count; i++) {
+		if (chain->images[i].buffer)
+			wsi_wl_image_finish(chain, &chain->images[i], pAllocator);
+	}
+
+	radv_free2(&chain->base.device->alloc, pAllocator, chain);
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+wsi_wl_surface_create_swapchain(VkIcdSurfaceBase *icd_surface,
+                                struct radv_device *device,
+                                const VkSwapchainCreateInfoKHR* pCreateInfo,
+                                const VkAllocationCallbacks* pAllocator,
+                                struct radv_swapchain **swapchain_out)
+{
+	VkIcdSurfaceWayland *surface = (VkIcdSurfaceWayland *)icd_surface;
+	struct wsi_wl_swapchain *chain;
+	VkResult result;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SWAPCHAIN_CREATE_INFO_KHR);
+
+	int num_images = pCreateInfo->minImageCount;
+
+	assert(num_images >= MIN_NUM_IMAGES);
+
+	/* For true mailbox mode, we need at least 4 images:
+	 *  1) One to scan out from
+	 *  2) One to have queued for scan-out
+	 *  3) One to be currently held by the Wayland compositor
+	 *  4) One to render to
+	 */
+	if (pCreateInfo->presentMode == VK_PRESENT_MODE_MAILBOX_KHR)
+		num_images = MAX2(num_images, 4);
+
+	size_t size = sizeof(*chain) + num_images * sizeof(chain->images[0]);
+	chain = radv_alloc2(&device->alloc, pAllocator, size, 8,
+			    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (chain == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	chain->base.device = device;
+	chain->base.destroy = wsi_wl_swapchain_destroy;
+	chain->base.get_images = wsi_wl_swapchain_get_images;
+	chain->base.acquire_next_image = wsi_wl_swapchain_acquire_next_image;
+	chain->base.queue_present = wsi_wl_swapchain_queue_present;
+
+	chain->surface = surface->surface;
+	chain->extent = pCreateInfo->imageExtent;
+	chain->vk_format = pCreateInfo->imageFormat;
+	chain->drm_format = wl_drm_format_for_vk_format(chain->vk_format, false);
+
+	chain->present_mode = pCreateInfo->presentMode;
+	chain->fifo_ready = true;
+
+	chain->image_count = num_images;
+
+	/* Mark a bunch of stuff as NULL.  This way we can just call
+	 * destroy_swapchain for cleanup.
+	 */
+	for (uint32_t i = 0; i < chain->image_count; i++)
+		chain->images[i].buffer = NULL;
+	chain->queue = NULL;
+
+	chain->display = wsi_wl_get_display(&device->instance->physicalDevice,
+					    surface->display);
+	if (!chain->display) {
+		result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
+		goto fail;
+	}
+
+	chain->queue = wl_display_create_queue(chain->display->display);
+	if (!chain->queue) {
+		result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
+		goto fail;
+	}
+
+	for (uint32_t i = 0; i < chain->image_count; i++) {
+		result = wsi_wl_image_init(chain, &chain->images[i],
+					   pCreateInfo, pAllocator);
+		if (result != VK_SUCCESS)
+			goto fail;
+		chain->images[i].busy = false;
+	}
+
+	*swapchain_out = &chain->base;
+
+	return VK_SUCCESS;
+
+fail:
+	wsi_wl_swapchain_destroy(&chain->base, pAllocator);
+
+	return result;
+}
+
+VkResult
+radv_wl_init_wsi(struct radv_physical_device *device)
+{
+	struct wsi_wayland *wsi;
+	VkResult result;
+
+	wsi = radv_alloc(&device->instance->alloc, sizeof(*wsi), 8,
+			 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
+	if (!wsi) {
+		result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		goto fail;
+	}
+
+	wsi->physical_device = device;
+
+	int ret = pthread_mutex_init(&wsi->mutex, NULL);
+	if (ret != 0) {
+		if (ret == ENOMEM) {
+			result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		} else {
+			/* FINISHME: Choose a better error. */
+			result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		}
+
+		goto fail_alloc;
+	}
+
+	wsi->displays = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+						_mesa_key_pointer_equal);
+	if (!wsi->displays) {
+		result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		goto fail_mutex;
+	}
+
+	wsi->base.get_support = wsi_wl_surface_get_support;
+	wsi->base.get_capabilities = wsi_wl_surface_get_capabilities;
+	wsi->base.get_formats = wsi_wl_surface_get_formats;
+	wsi->base.get_present_modes = wsi_wl_surface_get_present_modes;
+	wsi->base.create_swapchain = wsi_wl_surface_create_swapchain;
+
+	device->wsi[VK_ICD_WSI_PLATFORM_WAYLAND] = &wsi->base;
+
+	return VK_SUCCESS;
+
+fail_mutex:
+	pthread_mutex_destroy(&wsi->mutex);
+
+fail_alloc:
+	radv_free(&device->instance->alloc, wsi);
+fail:
+	device->wsi[VK_ICD_WSI_PLATFORM_WAYLAND] = NULL;
+
+	return result;
+}
+
+void
+radv_wl_finish_wsi(struct radv_physical_device *device)
+{
+	struct wsi_wayland *wsi =
+		(struct wsi_wayland *)device->wsi[VK_ICD_WSI_PLATFORM_WAYLAND];
+
+	if (wsi) {
+		_mesa_hash_table_destroy(wsi->displays, NULL);
+
+		pthread_mutex_destroy(&wsi->mutex);
+
+		radv_free(&device->instance->alloc, wsi);
+	}
+}
diff --git a/src/amd/vulkan/radv_wsi_x11.c b/src/amd/vulkan/radv_wsi_x11.c
new file mode 100644
index 0000000000..0aae2a3530
--- /dev/null
+++ b/src/amd/vulkan/radv_wsi_x11.c
@@ -0,0 +1,963 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based mostly on anv driver which is:
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <X11/Xlib-xcb.h>
+#include <X11/xshmfence.h>
+#include <xcb/xcb.h>
+#include <xcb/dri3.h>
+#include <xcb/present.h>
+#include <errno.h>
+
+#include <unistd.h>
+#include "radv_wsi.h"
+#include "vk_format.h"
+#include "util/hash_table.h"
+
+struct wsi_x11_connection {
+	bool has_dri3;
+	bool has_present;
+};
+
+struct wsi_x11 {
+	struct radv_wsi_interface base;
+
+	pthread_mutex_t                              mutex;
+	/* Hash table of xcb_connection -> wsi_x11_connection mappings */
+	struct hash_table *connections;
+};
+
+static struct wsi_x11_connection *
+wsi_x11_connection_create(struct radv_physical_device *device,
+                          xcb_connection_t *conn)
+{
+	xcb_query_extension_cookie_t dri3_cookie, pres_cookie;
+	xcb_query_extension_reply_t *dri3_reply, *pres_reply;
+
+	struct wsi_x11_connection *wsi_conn =
+		radv_alloc(&device->instance->alloc, sizeof(*wsi_conn), 8,
+			   VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
+	if (!wsi_conn)
+		return NULL;
+
+	dri3_cookie = xcb_query_extension(conn, 4, "DRI3");
+	pres_cookie = xcb_query_extension(conn, 7, "PRESENT");
+
+	dri3_reply = xcb_query_extension_reply(conn, dri3_cookie, NULL);
+	pres_reply = xcb_query_extension_reply(conn, pres_cookie, NULL);
+	if (dri3_reply == NULL || pres_reply == NULL) {
+		free(dri3_reply);
+		free(pres_reply);
+		radv_free(&device->instance->alloc, wsi_conn);
+		return NULL;
+	}
+
+	wsi_conn->has_dri3 = dri3_reply->present != 0;
+	wsi_conn->has_present = pres_reply->present != 0;
+
+	free(dri3_reply);
+	free(pres_reply);
+
+	return wsi_conn;
+}
+
+static void
+wsi_x11_connection_destroy(struct radv_physical_device *device,
+                           struct wsi_x11_connection *conn)
+{
+	radv_free(&device->instance->alloc, conn);
+}
+
+static struct wsi_x11_connection *
+wsi_x11_get_connection(struct radv_physical_device *device,
+                       xcb_connection_t *conn)
+{
+	struct wsi_x11 *wsi =
+		(struct wsi_x11 *)device->wsi[VK_ICD_WSI_PLATFORM_XCB];
+
+	pthread_mutex_lock(&wsi->mutex);
+
+	struct hash_entry *entry = _mesa_hash_table_search(wsi->connections, conn);
+	if (!entry) {
+		/* We're about to make a bunch of blocking calls.  Let's drop the
+		 * mutex for now so we don't block up too badly.
+		 */
+		pthread_mutex_unlock(&wsi->mutex);
+
+		struct wsi_x11_connection *wsi_conn =
+			wsi_x11_connection_create(device, conn);
+
+		pthread_mutex_lock(&wsi->mutex);
+
+		entry = _mesa_hash_table_search(wsi->connections, conn);
+		if (entry) {
+			/* Oops, someone raced us to it */
+			wsi_x11_connection_destroy(device, wsi_conn);
+		} else {
+			entry = _mesa_hash_table_insert(wsi->connections, conn, wsi_conn);
+		}
+	}
+
+	pthread_mutex_unlock(&wsi->mutex);
+
+	return entry->data;
+}
+
+static const VkSurfaceFormatKHR formats[] = {
+	{ .format = VK_FORMAT_B8G8R8A8_UNORM, },
+	{ .format = VK_FORMAT_B8G8R8A8_SRGB, },
+};
+
+static const VkPresentModeKHR present_modes[] = {
+	VK_PRESENT_MODE_MAILBOX_KHR,
+};
+
+static xcb_screen_t *
+get_screen_for_root(xcb_connection_t *conn, xcb_window_t root)
+{
+	xcb_screen_iterator_t screen_iter =
+		xcb_setup_roots_iterator(xcb_get_setup(conn));
+
+	for (; screen_iter.rem; xcb_screen_next (&screen_iter)) {
+		if (screen_iter.data->root == root)
+			return screen_iter.data;
+	}
+
+	return NULL;
+}
+
+static xcb_visualtype_t *
+screen_get_visualtype(xcb_screen_t *screen, xcb_visualid_t visual_id,
+                      unsigned *depth)
+{
+	xcb_depth_iterator_t depth_iter =
+		xcb_screen_allowed_depths_iterator(screen);
+
+	for (; depth_iter.rem; xcb_depth_next (&depth_iter)) {
+		xcb_visualtype_iterator_t visual_iter =
+			xcb_depth_visuals_iterator (depth_iter.data);
+
+		for (; visual_iter.rem; xcb_visualtype_next (&visual_iter)) {
+			if (visual_iter.data->visual_id == visual_id) {
+				if (depth)
+					*depth = depth_iter.data->depth;
+				return visual_iter.data;
+			}
+		}
+	}
+
+	return NULL;
+}
+
+static xcb_visualtype_t *
+connection_get_visualtype(xcb_connection_t *conn, xcb_visualid_t visual_id,
+                          unsigned *depth)
+{
+	xcb_screen_iterator_t screen_iter =
+		xcb_setup_roots_iterator(xcb_get_setup(conn));
+
+	/* For this we have to iterate over all of the screens which is rather
+	 * annoying.  Fortunately, there is probably only 1.
+	 */
+	for (; screen_iter.rem; xcb_screen_next (&screen_iter)) {
+		xcb_visualtype_t *visual = screen_get_visualtype(screen_iter.data,
+								 visual_id, depth);
+		if (visual)
+			return visual;
+	}
+
+	return NULL;
+}
+
+static xcb_visualtype_t *
+get_visualtype_for_window(xcb_connection_t *conn, xcb_window_t window,
+                          unsigned *depth)
+{
+	xcb_query_tree_cookie_t tree_cookie;
+	xcb_get_window_attributes_cookie_t attrib_cookie;
+	xcb_query_tree_reply_t *tree;
+	xcb_get_window_attributes_reply_t *attrib;
+
+	tree_cookie = xcb_query_tree(conn, window);
+	attrib_cookie = xcb_get_window_attributes(conn, window);
+
+	tree = xcb_query_tree_reply(conn, tree_cookie, NULL);
+	attrib = xcb_get_window_attributes_reply(conn, attrib_cookie, NULL);
+	if (attrib == NULL || tree == NULL) {
+		free(attrib);
+		free(tree);
+		return NULL;
+	}
+
+	xcb_window_t root = tree->root;
+	xcb_visualid_t visual_id = attrib->visual;
+	free(attrib);
+	free(tree);
+
+	xcb_screen_t *screen = get_screen_for_root(conn, root);
+	if (screen == NULL)
+		return NULL;
+
+	return screen_get_visualtype(screen, visual_id, depth);
+}
+
+static bool
+visual_has_alpha(xcb_visualtype_t *visual, unsigned depth)
+{
+	uint32_t rgb_mask = visual->red_mask |
+		visual->green_mask |
+		visual->blue_mask;
+
+	uint32_t all_mask = 0xffffffff >> (32 - depth);
+
+	/* Do we have bits left over after RGB? */
+	return (all_mask & ~rgb_mask) != 0;
+}
+
+VkBool32 radv_GetPhysicalDeviceXcbPresentationSupportKHR(
+	VkPhysicalDevice                            physicalDevice,
+	uint32_t                                    queueFamilyIndex,
+	xcb_connection_t*                           connection,
+	xcb_visualid_t                              visual_id)
+{
+	RADV_FROM_HANDLE(radv_physical_device, device, physicalDevice);
+
+	struct wsi_x11_connection *wsi_conn =
+		wsi_x11_get_connection(device, connection);
+
+	if (!wsi_conn->has_dri3) {
+		fprintf(stderr, "vulkan: No DRI3 support\n");
+		return false;
+	}
+
+	unsigned visual_depth;
+	if (!connection_get_visualtype(connection, visual_id, &visual_depth))
+		return false;
+
+	if (visual_depth != 24 && visual_depth != 32)
+		return false;
+
+	return true;
+}
+
+VkBool32 radv_GetPhysicalDeviceXlibPresentationSupportKHR(
+	VkPhysicalDevice                            physicalDevice,
+	uint32_t                                    queueFamilyIndex,
+	Display*                                    dpy,
+	VisualID                                    visualID)
+{
+	return radv_GetPhysicalDeviceXcbPresentationSupportKHR(physicalDevice,
+							       queueFamilyIndex,
+							       XGetXCBConnection(dpy),
+							       visualID);
+}
+
+static xcb_connection_t*
+x11_surface_get_connection(VkIcdSurfaceBase *icd_surface)
+{
+	if (icd_surface->platform == VK_ICD_WSI_PLATFORM_XLIB)
+		return XGetXCBConnection(((VkIcdSurfaceXlib *)icd_surface)->dpy);
+	else
+		return ((VkIcdSurfaceXcb *)icd_surface)->connection;
+}
+
+static xcb_window_t
+x11_surface_get_window(VkIcdSurfaceBase *icd_surface)
+{
+	if (icd_surface->platform == VK_ICD_WSI_PLATFORM_XLIB)
+		return ((VkIcdSurfaceXlib *)icd_surface)->window;
+	else
+		return ((VkIcdSurfaceXcb *)icd_surface)->window;
+}
+
+static VkResult
+x11_surface_get_support(VkIcdSurfaceBase *icd_surface,
+                        struct radv_physical_device *device,
+                        uint32_t queueFamilyIndex,
+                        VkBool32* pSupported)
+{
+	xcb_connection_t *conn = x11_surface_get_connection(icd_surface);
+	xcb_window_t window = x11_surface_get_window(icd_surface);
+
+	struct wsi_x11_connection *wsi_conn =
+		wsi_x11_get_connection(device, conn);
+	if (!wsi_conn)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	if (!wsi_conn->has_dri3) {
+		fprintf(stderr, "vulkan: No DRI3 support\n");
+		*pSupported = false;
+		return VK_SUCCESS;
+	}
+
+	unsigned visual_depth;
+	if (!get_visualtype_for_window(conn, window, &visual_depth)) {
+		*pSupported = false;
+		return VK_SUCCESS;
+	}
+
+	if (visual_depth != 24 && visual_depth != 32) {
+		*pSupported = false;
+		return VK_SUCCESS;
+	}
+
+	*pSupported = true;
+	return VK_SUCCESS;
+}
+
+static VkResult
+x11_surface_get_capabilities(VkIcdSurfaceBase *icd_surface,
+                             struct radv_physical_device *device,
+                             VkSurfaceCapabilitiesKHR *caps)
+{
+	xcb_connection_t *conn = x11_surface_get_connection(icd_surface);
+	xcb_window_t window = x11_surface_get_window(icd_surface);
+	xcb_get_geometry_cookie_t geom_cookie;
+	xcb_generic_error_t *err;
+	xcb_get_geometry_reply_t *geom;
+	unsigned visual_depth;
+
+	geom_cookie = xcb_get_geometry(conn, window);
+
+	/* This does a round-trip.  This is why we do get_geometry first and
+	 * wait to read the reply until after we have a visual.
+	 */
+	xcb_visualtype_t *visual =
+		get_visualtype_for_window(conn, window, &visual_depth);
+
+	geom = xcb_get_geometry_reply(conn, geom_cookie, &err);
+	if (geom) {
+		VkExtent2D extent = { geom->width, geom->height };
+		caps->currentExtent = extent;
+		caps->minImageExtent = extent;
+		caps->maxImageExtent = extent;
+	} else {
+		/* This can happen if the client didn't wait for the configure event
+		 * to come back from the compositor.  In that case, we don't know the
+		 * size of the window so we just return valid "I don't know" stuff.
+		 */
+		caps->currentExtent = (VkExtent2D) { -1, -1 };
+		caps->minImageExtent = (VkExtent2D) { 1, 1 };
+		caps->maxImageExtent = (VkExtent2D) { INT16_MAX, INT16_MAX };
+	}
+	free(err);
+	free(geom);
+
+	if (visual_has_alpha(visual, visual_depth)) {
+		caps->supportedCompositeAlpha = VK_COMPOSITE_ALPHA_INHERIT_BIT_KHR |
+			VK_COMPOSITE_ALPHA_PRE_MULTIPLIED_BIT_KHR;
+	} else {
+		caps->supportedCompositeAlpha = VK_COMPOSITE_ALPHA_INHERIT_BIT_KHR |
+			VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR;
+	}
+
+	caps->minImageCount = 2;
+	caps->maxImageCount = 4;
+	caps->supportedTransforms = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
+	caps->currentTransform = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
+	caps->maxImageArrayLayers = 1;
+	caps->supportedUsageFlags =
+		VK_IMAGE_USAGE_TRANSFER_SRC_BIT |
+		VK_IMAGE_USAGE_SAMPLED_BIT |
+		VK_IMAGE_USAGE_TRANSFER_DST_BIT |
+		VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+x11_surface_get_formats(VkIcdSurfaceBase *surface,
+                        struct radv_physical_device *device,
+                        uint32_t *pSurfaceFormatCount,
+                        VkSurfaceFormatKHR *pSurfaceFormats)
+{
+	if (pSurfaceFormats == NULL) {
+		*pSurfaceFormatCount = ARRAY_SIZE(formats);
+		return VK_SUCCESS;
+	}
+
+	assert(*pSurfaceFormatCount >= ARRAY_SIZE(formats));
+	typed_memcpy(pSurfaceFormats, formats, *pSurfaceFormatCount);
+	*pSurfaceFormatCount = ARRAY_SIZE(formats);
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+x11_surface_get_present_modes(VkIcdSurfaceBase *surface,
+                              struct radv_physical_device *device,
+                              uint32_t *pPresentModeCount,
+                              VkPresentModeKHR *pPresentModes)
+{
+	if (pPresentModes == NULL) {
+		*pPresentModeCount = ARRAY_SIZE(present_modes);
+		return VK_SUCCESS;
+	}
+
+	assert(*pPresentModeCount >= ARRAY_SIZE(present_modes));
+	typed_memcpy(pPresentModes, present_modes, *pPresentModeCount);
+	*pPresentModeCount = ARRAY_SIZE(present_modes);
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+x11_surface_create_swapchain(VkIcdSurfaceBase *surface,
+                             struct radv_device *device,
+                             const VkSwapchainCreateInfoKHR* pCreateInfo,
+                             const VkAllocationCallbacks* pAllocator,
+                             struct radv_swapchain **swapchain);
+
+VkResult radv_CreateXcbSurfaceKHR(
+	VkInstance                                  _instance,
+	const VkXcbSurfaceCreateInfoKHR*            pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkSurfaceKHR*                               pSurface)
+{
+	RADV_FROM_HANDLE(radv_instance, instance, _instance);
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_XCB_SURFACE_CREATE_INFO_KHR);
+
+	VkIcdSurfaceXcb *surface;
+
+	surface = radv_alloc2(&instance->alloc, pAllocator, sizeof *surface, 8,
+			      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (surface == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	surface->base.platform = VK_ICD_WSI_PLATFORM_XCB;
+	surface->connection = pCreateInfo->connection;
+	surface->window = pCreateInfo->window;
+
+	*pSurface = _VkIcdSurfaceBase_to_handle(&surface->base);
+
+	return VK_SUCCESS;
+}
+
+VkResult radv_CreateXlibSurfaceKHR(
+	VkInstance                                  _instance,
+	const VkXlibSurfaceCreateInfoKHR*           pCreateInfo,
+	const VkAllocationCallbacks*                pAllocator,
+	VkSurfaceKHR*                               pSurface)
+{
+	RADV_FROM_HANDLE(radv_instance, instance, _instance);
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_XLIB_SURFACE_CREATE_INFO_KHR);
+
+	VkIcdSurfaceXlib *surface;
+
+	surface = radv_alloc2(&instance->alloc, pAllocator, sizeof *surface, 8,
+			      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (surface == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	surface->base.platform = VK_ICD_WSI_PLATFORM_XLIB;
+	surface->dpy = pCreateInfo->dpy;
+	surface->window = pCreateInfo->window;
+
+	*pSurface = _VkIcdSurfaceBase_to_handle(&surface->base);
+
+	return VK_SUCCESS;
+}
+
+struct x11_image {
+	struct radv_image *                        image;
+	struct radv_device_memory *                memory;
+	xcb_pixmap_t                              pixmap;
+	bool                                      busy;
+	struct xshmfence *                        shm_fence;
+	uint32_t                                  sync_fence;
+};
+
+struct x11_swapchain {
+	struct radv_swapchain                        base;
+
+	xcb_connection_t *                           conn;
+	xcb_window_t                                 window;
+	xcb_gc_t                                     gc;
+	VkExtent2D                                   extent;
+	uint32_t                                     image_count;
+
+	xcb_present_event_t                          event_id;
+	xcb_special_event_t *                        special_event;
+	uint64_t                                     send_sbc;
+	uint32_t                                     stamp;
+
+	struct x11_image                             images[0];
+};
+
+static VkResult
+x11_get_images(struct radv_swapchain *radv_chain,
+               uint32_t* pCount, VkImage *pSwapchainImages)
+{
+	struct x11_swapchain *chain = (struct x11_swapchain *)radv_chain;
+
+	if (pSwapchainImages == NULL) {
+		*pCount = chain->image_count;
+		return VK_SUCCESS;
+	}
+
+	assert(chain->image_count <= *pCount);
+	for (uint32_t i = 0; i < chain->image_count; i++)
+		pSwapchainImages[i] = radv_image_to_handle(chain->images[i].image);
+
+	*pCount = chain->image_count;
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+x11_handle_dri3_present_event(struct x11_swapchain *chain,
+                              xcb_present_generic_event_t *event)
+{
+	switch (event->evtype) {
+	case XCB_PRESENT_CONFIGURE_NOTIFY: {
+		xcb_present_configure_notify_event_t *config = (void *) event;
+
+		if (config->width != chain->extent.width ||
+		    config->height != chain->extent.height)
+			return vk_error(VK_ERROR_OUT_OF_DATE_KHR);
+
+		break;
+	}
+
+	case XCB_PRESENT_EVENT_IDLE_NOTIFY: {
+		xcb_present_idle_notify_event_t *idle = (void *) event;
+
+		for (unsigned i = 0; i < chain->image_count; i++) {
+			if (chain->images[i].pixmap == idle->pixmap) {
+				chain->images[i].busy = false;
+				break;
+			}
+		}
+
+		break;
+	}
+
+	case XCB_PRESENT_COMPLETE_NOTIFY:
+	default:
+		break;
+	}
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+x11_acquire_next_image(struct radv_swapchain *radv_chain,
+                       uint64_t timeout,
+                       VkSemaphore semaphore,
+                       uint32_t *image_index)
+{
+	struct x11_swapchain *chain = (struct x11_swapchain *)radv_chain;
+
+	while (1) {
+		for (uint32_t i = 0; i < chain->image_count; i++) {
+			if (!chain->images[i].busy) {
+				/* We found a non-busy image */
+				xshmfence_await(chain->images[i].shm_fence);
+				*image_index = i;
+				return VK_SUCCESS;
+			}
+		}
+
+		xcb_flush(chain->conn);
+		xcb_generic_event_t *event =
+			xcb_wait_for_special_event(chain->conn, chain->special_event);
+		if (!event)
+			return vk_error(VK_ERROR_OUT_OF_DATE_KHR);
+
+		VkResult result = x11_handle_dri3_present_event(chain, (void *)event);
+		free(event);
+		if (result != VK_SUCCESS)
+			return result;
+	}
+}
+
+static VkResult
+x11_queue_present(struct radv_swapchain *radv_chain,
+                  struct radv_queue *queue,
+                  uint32_t image_index)
+{
+	struct x11_swapchain *chain = (struct x11_swapchain *)radv_chain;
+	struct x11_image *image = &chain->images[image_index];
+
+	assert(image_index < chain->image_count);
+
+	uint32_t options = XCB_PRESENT_OPTION_NONE;
+
+	int64_t target_msc = 0;
+	int64_t divisor = 0;
+	int64_t remainder = 0;
+
+	options |= XCB_PRESENT_OPTION_ASYNC;
+
+	xshmfence_reset(image->shm_fence);
+
+	++chain->send_sbc;
+	xcb_void_cookie_t cookie =
+		xcb_present_pixmap(chain->conn,
+				   chain->window,
+				   image->pixmap,
+				   (uint32_t) chain->send_sbc,
+				   0,                                    /* valid */
+				   0,                                    /* update */
+				   0,                                    /* x_off */
+				   0,                                    /* y_off */
+				   XCB_NONE,                             /* target_crtc */
+				   XCB_NONE,
+				   image->sync_fence,
+				   options,
+				   target_msc,
+				   divisor,
+				   remainder, 0, NULL);
+	xcb_discard_reply(chain->conn, cookie.sequence);
+	image->busy = true;
+
+	xcb_flush(chain->conn);
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+x11_image_init(struct radv_device *device, struct x11_swapchain *chain,
+               const VkSwapchainCreateInfoKHR *pCreateInfo,
+               const VkAllocationCallbacks* pAllocator,
+               struct x11_image *image)
+{
+	xcb_void_cookie_t cookie;
+	VkResult result = VK_SUCCESS;
+	int fd;
+	VkImage image_h;
+	bool bret;
+	struct radeon_surf *surface;
+	result = radv_image_create(radv_device_to_handle(device),
+				   &(struct radv_image_create_info) {
+					   .vk_info =
+						   &(VkImageCreateInfo) {
+						   .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
+						   .imageType = VK_IMAGE_TYPE_2D,
+						   .format = pCreateInfo->imageFormat,
+						   .extent = {
+							   .width = pCreateInfo->imageExtent.width,
+							   .height = pCreateInfo->imageExtent.height,
+							   .depth = 1
+						   },
+						   .mipLevels = 1,
+						   .arrayLayers = 1,
+						   .samples = 1,
+						   /* FIXME: Need a way to use X tiling to allow scanout */
+						   .tiling = VK_IMAGE_TILING_OPTIMAL,
+						   .usage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT,
+						   .flags = 0,
+					   },
+						   .scanout = true},
+				   NULL,
+				   &image_h);
+	if (result != VK_SUCCESS)
+		return result;
+
+	image->image = radv_image_from_handle(image_h);
+
+	VkDeviceMemory memory_h;
+	result = radv_AllocateMemory(radv_device_to_handle(device),
+				     &(VkMemoryAllocateInfo) {
+					     .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
+						     .allocationSize = image->image->size,
+						     .memoryTypeIndex = 0,
+						     },
+				     NULL /* XXX: pAllocator */,
+				     &memory_h);
+	if (result != VK_SUCCESS)
+		goto fail_create_image;
+
+	image->memory = radv_device_memory_from_handle(memory_h);
+	//   image->memory->bo.is_winsys_bo = true;
+
+	radv_BindImageMemory(VK_NULL_HANDLE, image_h, memory_h, 0);
+
+	bret = device->ws->buffer_get_fd(device->ws,
+					 image->memory->bo, &fd);
+	if (bret == false)
+		goto fail_alloc_memory;
+
+	{
+		struct radeon_bo_metadata metadata;
+		radv_init_metadata(device, image->image, &metadata);
+		device->ws->buffer_set_metadata(image->memory->bo, &metadata);
+	}
+	surface = &image->image->surface;
+	uint32_t bpp = 32;
+	uint32_t depth = 24;
+	image->pixmap = xcb_generate_id(chain->conn);
+
+	cookie =
+		xcb_dri3_pixmap_from_buffer_checked(chain->conn,
+						    image->pixmap,
+						    chain->window,
+						    image->image->size,
+						    pCreateInfo->imageExtent.width,
+						    pCreateInfo->imageExtent.height,
+						    surface->level[0].pitch_bytes,
+						    depth, bpp, fd);
+	xcb_discard_reply(chain->conn, cookie.sequence);
+
+	int fence_fd = xshmfence_alloc_shm();
+	if (fence_fd < 0)
+		goto fail_pixmap;
+
+	image->shm_fence = xshmfence_map_shm(fence_fd);
+	if (image->shm_fence == NULL)
+		goto fail_shmfence_alloc;
+
+	image->sync_fence = xcb_generate_id(chain->conn);
+	xcb_dri3_fence_from_fd(chain->conn,
+			       image->pixmap,
+			       image->sync_fence,
+			       false,
+			       fence_fd);
+
+	image->busy = false;
+	xshmfence_trigger(image->shm_fence);
+
+	return VK_SUCCESS;
+
+fail_shmfence_alloc:
+	close(fence_fd);
+
+fail_pixmap:
+	cookie = xcb_free_pixmap(chain->conn, image->pixmap);
+	xcb_discard_reply(chain->conn, cookie.sequence);
+
+fail_alloc_memory:
+	radv_FreeMemory(radv_device_to_handle(chain->base.device),
+			radv_device_memory_to_handle(image->memory), pAllocator);
+
+fail_create_image:
+	radv_DestroyImage(radv_device_to_handle(chain->base.device),
+			  radv_image_to_handle(image->image), pAllocator);
+
+	return result;
+}
+
+static void
+x11_image_finish(struct x11_swapchain *chain,
+                 const VkAllocationCallbacks* pAllocator,
+                 struct x11_image *image)
+{
+	xcb_void_cookie_t cookie;
+
+	cookie = xcb_sync_destroy_fence(chain->conn, image->sync_fence);
+	xcb_discard_reply(chain->conn, cookie.sequence);
+	xshmfence_unmap_shm(image->shm_fence);
+
+	cookie = xcb_free_pixmap(chain->conn, image->pixmap);
+	xcb_discard_reply(chain->conn, cookie.sequence);
+
+	radv_DestroyImage(radv_device_to_handle(chain->base.device),
+			  radv_image_to_handle(image->image), pAllocator);
+
+	radv_FreeMemory(radv_device_to_handle(chain->base.device),
+			radv_device_memory_to_handle(image->memory), pAllocator);
+
+}
+
+static VkResult
+x11_swapchain_destroy(struct radv_swapchain *radv_chain,
+                      const VkAllocationCallbacks *pAllocator)
+{
+	struct x11_swapchain *chain = (struct x11_swapchain *)radv_chain;
+
+	for (uint32_t i = 0; i < chain->image_count; i++)
+		x11_image_finish(chain, pAllocator, &chain->images[i]);
+
+	xcb_unregister_for_special_event(chain->conn, chain->special_event);
+
+	radv_free2(&chain->base.device->alloc, pAllocator, chain);
+
+	return VK_SUCCESS;
+}
+
+static VkResult
+x11_surface_create_swapchain(VkIcdSurfaceBase *icd_surface,
+                             struct radv_device *device,
+                             const VkSwapchainCreateInfoKHR *pCreateInfo,
+                             const VkAllocationCallbacks* pAllocator,
+                             struct radv_swapchain **swapchain_out)
+{
+	struct x11_swapchain *chain;
+	xcb_void_cookie_t cookie;
+	VkResult result;
+
+	assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SWAPCHAIN_CREATE_INFO_KHR);
+
+	int num_images = pCreateInfo->minImageCount;
+
+	/* For true mailbox mode, we need at least 4 images:
+	 *  1) One to scan out from
+	 *  2) One to have queued for scan-out
+	 *  3) One to be currently held by the Wayland compositor
+	 *  4) One to render to
+	 */
+	if (pCreateInfo->presentMode == VK_PRESENT_MODE_MAILBOX_KHR)
+		num_images = MAX(num_images, 4);
+
+	size_t size = sizeof(*chain) + num_images * sizeof(chain->images[0]);
+	chain = radv_alloc2(&device->alloc, pAllocator, size, 8,
+			    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+	if (chain == NULL)
+		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+	chain->base.device = device;
+	chain->base.destroy = x11_swapchain_destroy;
+	chain->base.get_images = x11_get_images;
+	chain->base.acquire_next_image = x11_acquire_next_image;
+	chain->base.queue_present = x11_queue_present;
+
+	chain->conn = x11_surface_get_connection(icd_surface);
+	chain->window = x11_surface_get_window(icd_surface);
+	chain->extent = pCreateInfo->imageExtent;
+	chain->image_count = num_images;
+
+	chain->send_sbc = 0;
+	chain->event_id = xcb_generate_id(chain->conn);
+	xcb_present_select_input(chain->conn, chain->event_id, chain->window,
+				 XCB_PRESENT_EVENT_MASK_CONFIGURE_NOTIFY |
+				 XCB_PRESENT_EVENT_MASK_COMPLETE_NOTIFY |
+				 XCB_PRESENT_EVENT_MASK_IDLE_NOTIFY);
+
+	/* Create an XCB event queue to hold present events outside of the usual
+	 * application event queue
+	 */
+	chain->special_event =
+		xcb_register_for_special_xge(chain->conn, &xcb_present_id,
+					     chain->event_id, NULL);
+
+	chain->gc = xcb_generate_id(chain->conn);
+	if (!chain->gc) {
+		/* FINISHME: Choose a better error. */
+		result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		goto fail_register;
+	}
+
+	cookie = xcb_create_gc(chain->conn,
+			       chain->gc,
+			       chain->window,
+			       XCB_GC_GRAPHICS_EXPOSURES,
+			       (uint32_t []) { 0 });
+	xcb_discard_reply(chain->conn, cookie.sequence);
+
+	uint32_t image = 0;
+	for (; image < chain->image_count; image++) {
+		result = x11_image_init(device, chain, pCreateInfo, pAllocator,
+					&chain->images[image]);
+		if (result != VK_SUCCESS)
+			goto fail_init_images;
+	}
+
+	*swapchain_out = &chain->base;
+
+	return VK_SUCCESS;
+
+fail_init_images:
+	for (uint32_t j = 0; j < image; j++)
+		x11_image_finish(chain, pAllocator, &chain->images[j]);
+
+fail_register:
+	xcb_unregister_for_special_event(chain->conn, chain->special_event);
+
+	radv_free2(&device->alloc, pAllocator, chain);
+
+	return result;
+}
+
+VkResult
+radv_x11_init_wsi(struct radv_physical_device *device)
+{
+	struct wsi_x11 *wsi;
+	VkResult result;
+
+	wsi = radv_alloc(&device->instance->alloc, sizeof(*wsi), 8,
+			 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
+	if (!wsi) {
+		result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		goto fail;
+	}
+
+	int ret = pthread_mutex_init(&wsi->mutex, NULL);
+	if (ret != 0) {
+		if (ret == ENOMEM) {
+			result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		} else {
+			/* FINISHME: Choose a better error. */
+			result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		}
+
+		goto fail_alloc;
+	}
+
+	wsi->connections = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+						   _mesa_key_pointer_equal);
+	if (!wsi->connections) {
+		result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+		goto fail_mutex;
+	}
+
+	wsi->base.get_support = x11_surface_get_support;
+	wsi->base.get_capabilities = x11_surface_get_capabilities;
+	wsi->base.get_formats = x11_surface_get_formats;
+	wsi->base.get_present_modes = x11_surface_get_present_modes;
+	wsi->base.create_swapchain = x11_surface_create_swapchain;
+
+	device->wsi[VK_ICD_WSI_PLATFORM_XCB] = &wsi->base;
+	device->wsi[VK_ICD_WSI_PLATFORM_XLIB] = &wsi->base;
+
+	return VK_SUCCESS;
+
+fail_mutex:
+	pthread_mutex_destroy(&wsi->mutex);
+fail_alloc:
+	radv_free(&device->instance->alloc, wsi);
+fail:
+	device->wsi[VK_ICD_WSI_PLATFORM_XCB] = NULL;
+	device->wsi[VK_ICD_WSI_PLATFORM_XLIB] = NULL;
+
+	return result;
+}
+
+void
+radv_x11_finish_wsi(struct radv_physical_device *device)
+{
+	struct wsi_x11 *wsi =
+		(struct wsi_x11 *)device->wsi[VK_ICD_WSI_PLATFORM_XCB];
+
+	if (wsi) {
+		_mesa_hash_table_destroy(wsi->connections, NULL);
+
+		pthread_mutex_destroy(&wsi->mutex);
+
+		radv_free(&device->instance->alloc, wsi);
+	}
+}
diff --git a/src/amd/vulkan/si_cmd_buffer.c b/src/amd/vulkan/si_cmd_buffer.c
new file mode 100644
index 0000000000..a61a950de6
--- /dev/null
+++ b/src/amd/vulkan/si_cmd_buffer.c
@@ -0,0 +1,1119 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based on si_state.c
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+/* command buffer handling for SI */
+
+#include "radv_private.h"
+#include "radv_cs.h"
+#include "sid.h"
+#include "radv_util.h"
+#include "main/macros.h"
+
+#define SI_GS_PER_ES 128
+
+static void
+si_write_harvested_raster_configs(struct radv_physical_device *physical_device,
+                                  struct radeon_winsys_cs *cs,
+				  unsigned raster_config,
+				  unsigned raster_config_1)
+{
+	unsigned sh_per_se = MAX2(physical_device->rad_info.max_sh_per_se, 1);
+	unsigned num_se = MAX2(physical_device->rad_info.max_se, 1);
+	unsigned rb_mask = physical_device->rad_info.enabled_rb_mask;
+	unsigned num_rb = MIN2(physical_device->rad_info.num_render_backends, 16);
+	unsigned rb_per_pkr = MIN2(num_rb / num_se / sh_per_se, 2);
+	unsigned rb_per_se = num_rb / num_se;
+	unsigned se_mask[4];
+	unsigned se;
+
+	se_mask[0] = ((1 << rb_per_se) - 1) & rb_mask;
+	se_mask[1] = (se_mask[0] << rb_per_se) & rb_mask;
+	se_mask[2] = (se_mask[1] << rb_per_se) & rb_mask;
+	se_mask[3] = (se_mask[2] << rb_per_se) & rb_mask;
+
+	assert(num_se == 1 || num_se == 2 || num_se == 4);
+	assert(sh_per_se == 1 || sh_per_se == 2);
+	assert(rb_per_pkr == 1 || rb_per_pkr == 2);
+
+	/* XXX: I can't figure out what the *_XSEL and *_YSEL
+	 * fields are for, so I'm leaving them as their default
+	 * values. */
+
+	if ((num_se > 2) && ((!se_mask[0] && !se_mask[1]) ||
+			     (!se_mask[2] && !se_mask[3]))) {
+		raster_config_1 &= C_028354_SE_PAIR_MAP;
+
+		if (!se_mask[0] && !se_mask[1]) {
+			raster_config_1 |=
+				S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3);
+		} else {
+			raster_config_1 |=
+				S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0);
+		}
+	}
+
+	for (se = 0; se < num_se; se++) {
+		unsigned raster_config_se = raster_config;
+		unsigned pkr0_mask = ((1 << rb_per_pkr) - 1) << (se * rb_per_se);
+		unsigned pkr1_mask = pkr0_mask << rb_per_pkr;
+		int idx = (se / 2) * 2;
+
+		if ((num_se > 1) && (!se_mask[idx] || !se_mask[idx + 1])) {
+			raster_config_se &= C_028350_SE_MAP;
+
+			if (!se_mask[idx]) {
+				raster_config_se |=
+					S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3);
+			} else {
+				raster_config_se |=
+					S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0);
+			}
+		}
+
+		pkr0_mask &= rb_mask;
+		pkr1_mask &= rb_mask;
+		if (rb_per_se > 2 && (!pkr0_mask || !pkr1_mask)) {
+			raster_config_se &= C_028350_PKR_MAP;
+
+			if (!pkr0_mask) {
+				raster_config_se |=
+					S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3);
+			} else {
+				raster_config_se |=
+					S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0);
+			}
+		}
+
+		if (rb_per_se >= 2) {
+			unsigned rb0_mask = 1 << (se * rb_per_se);
+			unsigned rb1_mask = rb0_mask << 1;
+
+			rb0_mask &= rb_mask;
+			rb1_mask &= rb_mask;
+			if (!rb0_mask || !rb1_mask) {
+				raster_config_se &= C_028350_RB_MAP_PKR0;
+
+				if (!rb0_mask) {
+					raster_config_se |=
+						S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3);
+				} else {
+					raster_config_se |=
+						S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0);
+				}
+			}
+
+			if (rb_per_se > 2) {
+				rb0_mask = 1 << (se * rb_per_se + rb_per_pkr);
+				rb1_mask = rb0_mask << 1;
+				rb0_mask &= rb_mask;
+				rb1_mask &= rb_mask;
+				if (!rb0_mask || !rb1_mask) {
+					raster_config_se &= C_028350_RB_MAP_PKR1;
+
+					if (!rb0_mask) {
+						raster_config_se |=
+							S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3);
+					} else {
+						raster_config_se |=
+							S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0);
+					}
+				}
+			}
+		}
+
+		/* GRBM_GFX_INDEX has a different offset on SI and CI+ */
+		if (physical_device->rad_info.chip_class < CIK)
+			radeon_set_config_reg(cs, GRBM_GFX_INDEX,
+					      SE_INDEX(se) | SH_BROADCAST_WRITES |
+					      INSTANCE_BROADCAST_WRITES);
+		else
+			radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
+					       S_030800_SE_INDEX(se) | S_030800_SH_BROADCAST_WRITES(1) |
+					       S_030800_INSTANCE_BROADCAST_WRITES(1));
+		radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG, raster_config_se);
+		if (physical_device->rad_info.chip_class >= CIK)
+			radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
+	}
+
+	/* GRBM_GFX_INDEX has a different offset on SI and CI+ */
+	if (physical_device->rad_info.chip_class < CIK)
+		radeon_set_config_reg(cs, GRBM_GFX_INDEX,
+				      SE_BROADCAST_WRITES | SH_BROADCAST_WRITES |
+				      INSTANCE_BROADCAST_WRITES);
+	else
+		radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
+				       S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
+				       S_030800_INSTANCE_BROADCAST_WRITES(1));
+}
+
+static void
+si_init_compute(struct radv_physical_device *physical_device,
+                struct radeon_winsys_cs *cs)
+{
+	radeon_set_sh_reg_seq(cs, R_00B810_COMPUTE_START_X, 3);
+	radeon_emit(cs, 0);
+	radeon_emit(cs, 0);
+	radeon_emit(cs, 0);
+
+	radeon_set_sh_reg_seq(cs, R_00B854_COMPUTE_RESOURCE_LIMITS, 3);
+	radeon_emit(cs, 0);
+	/* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
+	radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
+	radeon_emit(cs, S_00B85C_SH0_CU_EN(0xffff) | S_00B85C_SH1_CU_EN(0xffff));
+
+	if (physical_device->rad_info.chip_class >= CIK) {
+		/* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
+		radeon_set_sh_reg_seq(cs,
+				      R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2, 2);
+		radeon_emit(cs, S_00B864_SH0_CU_EN(0xffff) |
+			    S_00B864_SH1_CU_EN(0xffff));
+		radeon_emit(cs, S_00B868_SH0_CU_EN(0xffff) |
+			    S_00B868_SH1_CU_EN(0xffff));
+	}
+
+	/* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
+	 * and is now per pipe, so it should be handled in the
+	 * kernel if we want to use something other than the default value,
+	 * which is now 0x22f.
+	 */
+	if (physical_device->rad_info.chip_class <= SI) {
+		/* XXX: This should be:
+		 * (number of compute units) * 4 * (waves per simd) - 1 */
+
+		radeon_set_sh_reg(cs, R_00B82C_COMPUTE_MAX_WAVE_ID,
+		                  0x190 /* Default value */);
+	}
+}
+
+
+void si_init_config(struct radv_physical_device *physical_device,
+		    struct radv_cmd_buffer *cmd_buffer)
+{
+	unsigned num_rb = MIN2(physical_device->rad_info.num_render_backends, 16);
+	unsigned rb_mask = physical_device->rad_info.enabled_rb_mask;
+	unsigned raster_config, raster_config_1;
+	int i;
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+	radeon_emit(cs, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
+	radeon_emit(cs, CONTEXT_CONTROL_LOAD_ENABLE(1));
+	radeon_emit(cs, CONTEXT_CONTROL_SHADOW_ENABLE(1));
+
+	radeon_set_context_reg(cs, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
+	radeon_set_context_reg(cs, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
+
+	/* FIXME calculate these values somehow ??? */
+	radeon_set_context_reg(cs, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
+	radeon_set_context_reg(cs, R_028A58_VGT_ES_PER_GS, 0x40);
+	radeon_set_context_reg(cs, R_028A5C_VGT_GS_PER_VS, 0x2);
+
+	radeon_set_context_reg(cs, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
+	radeon_set_context_reg(cs, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
+
+	radeon_set_context_reg(cs, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
+	radeon_set_context_reg(cs, R_028AB8_VGT_VTX_CNT_EN, 0x0);
+	if (physical_device->rad_info.chip_class < CIK)
+		radeon_set_config_reg(cs, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
+				      S_008A14_CLIP_VTX_REORDER_ENA(1));
+
+	radeon_set_context_reg(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 0x76543210);
+	radeon_set_context_reg(cs, R_028BD8_PA_SC_CENTROID_PRIORITY_1, 0xfedcba98);
+
+	radeon_set_context_reg(cs, R_02882C_PA_SU_PRIM_FILTER_CNTL, 0);
+
+	for (i = 0; i < 16; i++) {
+		radeon_set_context_reg(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 + i*8, 0);
+		radeon_set_context_reg(cs, R_0282D4_PA_SC_VPORT_ZMAX_0 + i*8, fui(1.0));
+	}
+
+	switch (physical_device->rad_info.family) {
+	case CHIP_TAHITI:
+	case CHIP_PITCAIRN:
+		raster_config = 0x2a00126a;
+		raster_config_1 = 0x00000000;
+		break;
+	case CHIP_VERDE:
+		raster_config = 0x0000124a;
+		raster_config_1 = 0x00000000;
+		break;
+	case CHIP_OLAND:
+		raster_config = 0x00000082;
+		raster_config_1 = 0x00000000;
+		break;
+	case CHIP_HAINAN:
+		raster_config = 0x00000000;
+		raster_config_1 = 0x00000000;
+		break;
+	case CHIP_BONAIRE:
+		raster_config = 0x16000012;
+		raster_config_1 = 0x00000000;
+		break;
+	case CHIP_HAWAII:
+		raster_config = 0x3a00161a;
+		raster_config_1 = 0x0000002e;
+		break;
+	case CHIP_FIJI:
+		if (physical_device->rad_info.cik_macrotile_mode_array[0] == 0x000000e8) {
+			/* old kernels with old tiling config */
+			raster_config = 0x16000012;
+			raster_config_1 = 0x0000002a;
+		} else {
+			raster_config = 0x3a00161a;
+			raster_config_1 = 0x0000002e;
+		}
+		break;
+	case CHIP_POLARIS10:
+		raster_config = 0x16000012;
+		raster_config_1 = 0x0000002a;
+		break;
+	case CHIP_POLARIS11:
+		raster_config = 0x16000012;
+		raster_config_1 = 0x00000000;
+		break;
+	case CHIP_TONGA:
+		raster_config = 0x16000012;
+		raster_config_1 = 0x0000002a;
+		break;
+	case CHIP_ICELAND:
+		if (num_rb == 1)
+			raster_config = 0x00000000;
+		else
+			raster_config = 0x00000002;
+		raster_config_1 = 0x00000000;
+		break;
+	case CHIP_CARRIZO:
+		raster_config = 0x00000002;
+		raster_config_1 = 0x00000000;
+		break;
+	case CHIP_KAVERI:
+		/* KV should be 0x00000002, but that causes problems with radeon */
+		raster_config = 0x00000000; /* 0x00000002 */
+		raster_config_1 = 0x00000000;
+		break;
+	case CHIP_KABINI:
+	case CHIP_MULLINS:
+	case CHIP_STONEY:
+		raster_config = 0x00000000;
+		raster_config_1 = 0x00000000;
+		break;
+	default:
+		fprintf(stderr,
+			"radeonsi: Unknown GPU, using 0 for raster_config\n");
+		raster_config = 0x00000000;
+		raster_config_1 = 0x00000000;
+		break;
+	}
+
+	/* Always use the default config when all backends are enabled
+	 * (or when we failed to determine the enabled backends).
+	 */
+	if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
+		radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG,
+				       raster_config);
+		if (physical_device->rad_info.chip_class >= CIK)
+			radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1,
+					       raster_config_1);
+	} else {
+		si_write_harvested_raster_configs(physical_device, cs, raster_config, raster_config_1);
+	}
+
+	radeon_set_context_reg(cs, R_028204_PA_SC_WINDOW_SCISSOR_TL, S_028204_WINDOW_OFFSET_DISABLE(1));
+	radeon_set_context_reg(cs, R_028240_PA_SC_GENERIC_SCISSOR_TL, S_028240_WINDOW_OFFSET_DISABLE(1));
+	radeon_set_context_reg(cs, R_028244_PA_SC_GENERIC_SCISSOR_BR,
+			       S_028244_BR_X(16384) | S_028244_BR_Y(16384));
+	radeon_set_context_reg(cs, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
+	radeon_set_context_reg(cs, R_028034_PA_SC_SCREEN_SCISSOR_BR,
+			       S_028034_BR_X(16384) | S_028034_BR_Y(16384));
+
+	radeon_set_context_reg(cs, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
+	radeon_set_context_reg(cs, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
+	/* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
+	radeon_set_context_reg(cs, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET, 0);
+	radeon_set_context_reg(cs, R_028820_PA_CL_NANINF_CNTL, 0);
+
+	radeon_set_context_reg(cs, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ, fui(1.0));
+	radeon_set_context_reg(cs, R_028BEC_PA_CL_GB_VERT_DISC_ADJ, fui(1.0));
+	radeon_set_context_reg(cs, R_028BF0_PA_CL_GB_HORZ_CLIP_ADJ, fui(1.0));
+	radeon_set_context_reg(cs, R_028BF4_PA_CL_GB_HORZ_DISC_ADJ, fui(1.0));
+
+	radeon_set_context_reg(cs, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
+	radeon_set_context_reg(cs, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
+	radeon_set_context_reg(cs, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
+	radeon_set_context_reg(cs, R_02800C_DB_RENDER_OVERRIDE,
+			       S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
+			       S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE));
+
+	radeon_set_context_reg(cs, R_028400_VGT_MAX_VTX_INDX, ~0);
+	radeon_set_context_reg(cs, R_028404_VGT_MIN_VTX_INDX, 0);
+	radeon_set_context_reg(cs, R_028408_VGT_INDX_OFFSET, 0);
+
+	if (physical_device->rad_info.chip_class >= CIK) {
+		radeon_set_sh_reg(cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS, 0);
+		radeon_set_sh_reg(cs, R_00B31C_SPI_SHADER_PGM_RSRC3_ES, S_00B31C_CU_EN(0xffff));
+		radeon_set_sh_reg(cs, R_00B21C_SPI_SHADER_PGM_RSRC3_GS, S_00B21C_CU_EN(0xffff));
+
+		if (physical_device->rad_info.num_good_compute_units /
+		    (physical_device->rad_info.max_se * physical_device->rad_info.max_sh_per_se) <= 4) {
+			/* Too few available compute units per SH. Disallowing
+			 * VS to run on CU0 could hurt us more than late VS
+			 * allocation would help.
+			 *
+			 * LATE_ALLOC_VS = 2 is the highest safe number.
+			 */
+			radeon_set_sh_reg(cs, R_00B51C_SPI_SHADER_PGM_RSRC3_LS, S_00B51C_CU_EN(0xffff));
+			radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS, S_00B118_CU_EN(0xffff));
+			radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(2));
+		} else {
+			/* Set LATE_ALLOC_VS == 31. It should be less than
+			 * the number of scratch waves. Limitations:
+			 * - VS can't execute on CU0.
+			 * - If HS writes outputs to LDS, LS can't execute on CU0.
+			 */
+			radeon_set_sh_reg(cs, R_00B51C_SPI_SHADER_PGM_RSRC3_LS, S_00B51C_CU_EN(0xfffe));
+			radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS, S_00B118_CU_EN(0xfffe));
+			radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(31));
+		}
+
+		radeon_set_sh_reg(cs, R_00B01C_SPI_SHADER_PGM_RSRC3_PS, S_00B01C_CU_EN(0xffff));
+	}
+
+	if (physical_device->rad_info.chip_class >= VI) {
+		radeon_set_context_reg(cs, R_028424_CB_DCC_CONTROL,
+				       S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
+				       S_028424_OVERWRITE_COMBINER_WATERMARK(4));
+		radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 30);
+		radeon_set_context_reg(cs, R_028C5C_VGT_OUT_DEALLOC_CNTL, 32);
+		radeon_set_context_reg(cs, R_028B50_VGT_TESS_DISTRIBUTION,
+				       S_028B50_ACCUM_ISOLINE(32) |
+				       S_028B50_ACCUM_TRI(11) |
+				       S_028B50_ACCUM_QUAD(11) |
+				       S_028B50_DONUT_SPLIT(16));
+	} else {
+		radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
+		radeon_set_context_reg(cs, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
+	}
+
+	if (physical_device->rad_info.family == CHIP_STONEY)
+		radeon_set_context_reg(cs, R_028C40_PA_SC_SHADER_CONTROL, 0);
+
+	si_init_compute(physical_device, cs);
+}
+
+static void
+get_viewport_xform(const VkViewport *viewport,
+                   float scale[3], float translate[3])
+{
+	float x = viewport->x;
+	float y = viewport->y;
+	float half_width = 0.5f * viewport->width;
+	float half_height = 0.5f * viewport->height;
+	double n = viewport->minDepth;
+	double f = viewport->maxDepth;
+
+	scale[0] = half_width;
+	translate[0] = half_width + x;
+	scale[1] = half_height;
+	translate[1] = half_height + y;
+
+	scale[2] = (f - n);
+	translate[2] = n;
+}
+
+void
+si_write_viewport(struct radeon_winsys_cs *cs, int first_vp,
+                  int count, const VkViewport *viewports)
+{
+	int i;
+
+	if (count == 0) {
+		radeon_set_context_reg_seq(cs, R_02843C_PA_CL_VPORT_XSCALE, 6);
+		radeon_emit(cs, fui(1.0));
+		radeon_emit(cs, fui(0.0));
+		radeon_emit(cs, fui(1.0));
+		radeon_emit(cs, fui(0.0));
+		radeon_emit(cs, fui(1.0));
+		radeon_emit(cs, fui(0.0));
+
+		radeon_set_context_reg_seq(cs, R_0282D0_PA_SC_VPORT_ZMIN_0, 2);
+		radeon_emit(cs, fui(0.0));
+		radeon_emit(cs, fui(1.0));
+
+		return;
+	}
+	radeon_set_context_reg_seq(cs, R_02843C_PA_CL_VPORT_XSCALE +
+				   first_vp * 4 * 6, count * 6);
+
+	for (i = 0; i < count; i++) {
+		float scale[3], translate[3];
+
+
+		get_viewport_xform(&viewports[i], scale, translate);
+		radeon_emit(cs, fui(scale[0]));
+		radeon_emit(cs, fui(translate[0]));
+		radeon_emit(cs, fui(scale[1]));
+		radeon_emit(cs, fui(translate[1]));
+		radeon_emit(cs, fui(scale[2]));
+		radeon_emit(cs, fui(translate[2]));
+	}
+
+	for (i = 0; i < count; i++) {
+		float zmin = MIN2(viewports[i].minDepth, viewports[i].maxDepth);
+		float zmax = MAX2(viewports[i].minDepth, viewports[i].maxDepth);
+		radeon_set_context_reg_seq(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 +
+					   first_vp * 4 * 2, count * 2);
+		radeon_emit(cs, fui(zmin));
+		radeon_emit(cs, fui(zmax));
+	}
+}
+
+void
+si_write_scissors(struct radeon_winsys_cs *cs, int first,
+                  int count, const VkRect2D *scissors)
+{
+	int i;
+	if (count == 0)
+		return;
+
+	radeon_set_context_reg_seq(cs, R_028250_PA_SC_VPORT_SCISSOR_0_TL + first * 4 * 2, count * 2);
+	for (i = 0; i < count; i++) {
+		radeon_emit(cs, S_028250_TL_X(scissors[i].offset.x) |
+			    S_028250_TL_Y(scissors[i].offset.y) |
+			    S_028250_WINDOW_OFFSET_DISABLE(1));
+		radeon_emit(cs, S_028254_BR_X(scissors[i].offset.x + scissors[i].extent.width) |
+			    S_028254_BR_Y(scissors[i].offset.y + scissors[i].extent.height));
+	}
+}
+
+uint32_t
+si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer)
+{
+	enum chip_class chip_class = cmd_buffer->device->instance->physicalDevice.rad_info.chip_class;
+	struct radeon_info *info = &cmd_buffer->device->instance->physicalDevice.rad_info;
+	unsigned prim = cmd_buffer->state.pipeline->graphics.prim;
+	unsigned primgroup_size = 128; /* recommended without a GS */
+	unsigned max_primgroup_in_wave = 2;
+	/* SWITCH_ON_EOP(0) is always preferable. */
+	bool wd_switch_on_eop = false;
+	bool ia_switch_on_eop = false;
+	bool ia_switch_on_eoi = false;
+	bool partial_vs_wave = false;
+	bool partial_es_wave = false;
+
+	/* TODO GS */
+
+	/* TODO TES */
+
+	/* TODO linestipple */
+
+	if (chip_class >= CIK) {
+		/* WD_SWITCH_ON_EOP has no effect on GPUs with less than
+		 * 4 shader engines. Set 1 to pass the assertion below.
+		 * The other cases are hardware requirements. */
+		if (info->max_se < 4 ||
+		    prim == V_008958_DI_PT_POLYGON ||
+		    prim == V_008958_DI_PT_LINELOOP ||
+		    prim == V_008958_DI_PT_TRIFAN ||
+		    prim == V_008958_DI_PT_TRISTRIP_ADJ)
+			//	    info->primitive_restart ||
+			//	    info->count_from_stream_output)
+			wd_switch_on_eop = true;
+
+		/* TODO HAWAII */
+
+		/* Required on CIK and later. */
+		if (info->max_se > 2 && !wd_switch_on_eop)
+			ia_switch_on_eoi = true;
+
+		/* Required by Hawaii and, for some special cases, by VI. */
+#if 0
+		if (ia_switch_on_eoi &&
+		    (sctx->b.family == CHIP_HAWAII ||
+		     (sctx->b.chip_class == VI &&
+		      (sctx->gs_shader.cso || max_primgroup_in_wave != 2))))
+			partial_vs_wave = true;
+#endif
+
+#if 0
+		/* Instancing bug on Bonaire. */
+		if (sctx->b.family == CHIP_BONAIRE && ia_switch_on_eoi &&
+		    (info->indirect || info->instance_count > 1))
+			partial_vs_wave = true;
+#endif
+		/* If the WD switch is false, the IA switch must be false too. */
+		assert(wd_switch_on_eop || !ia_switch_on_eop);
+	}
+	/* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
+	if (ia_switch_on_eoi)
+		partial_es_wave = true;
+
+	/* GS requirement. */
+#if 0
+	if (SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
+		partial_es_wave = true;
+#endif
+
+	/* Hw bug with single-primitive instances and SWITCH_ON_EOI
+	 * on multi-SE chips. */
+#if 0
+	if (sctx->b.screen->info.max_se >= 2 && ia_switch_on_eoi &&
+	    (info->indirect ||
+	     (info->instance_count > 1 &&
+	      si_num_prims_for_vertices(info) <= 1)))
+		sctx->b.flags |= SI_CONTEXT_VGT_FLUSH;
+#endif
+	return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) |
+		S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
+		S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
+		S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
+		S_028AA8_PRIMGROUP_SIZE(primgroup_size - 1) |
+		S_028AA8_WD_SWITCH_ON_EOP(chip_class >= CIK ? wd_switch_on_eop : 0) |
+		S_028AA8_MAX_PRIMGRP_IN_WAVE(chip_class >= VI ?
+					     max_primgroup_in_wave : 0);
+
+}
+
+void
+si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer)
+{
+	enum chip_class chip_class = cmd_buffer->device->instance->physicalDevice.rad_info.chip_class;
+	unsigned cp_coher_cntl = 0;
+
+	radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 128);
+
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_INV_ICACHE)
+		cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_INV_SMEM_L1)
+		cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_INV_VMEM_L1)
+		cp_coher_cntl |= S_0085F0_TCL1_ACTION_ENA(1);
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_INV_GLOBAL_L2) {
+		cp_coher_cntl |= S_0085F0_TC_ACTION_ENA(1);
+		if (chip_class >= VI)
+			cp_coher_cntl |= S_0301F0_TC_WB_ACTION_ENA(1);
+	}
+
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
+		cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
+			S_0085F0_CB0_DEST_BASE_ENA(1) |
+			S_0085F0_CB1_DEST_BASE_ENA(1) |
+			S_0085F0_CB2_DEST_BASE_ENA(1) |
+			S_0085F0_CB3_DEST_BASE_ENA(1) |
+			S_0085F0_CB4_DEST_BASE_ENA(1) |
+			S_0085F0_CB5_DEST_BASE_ENA(1) |
+			S_0085F0_CB6_DEST_BASE_ENA(1) |
+			S_0085F0_CB7_DEST_BASE_ENA(1);
+
+		/* Necessary for DCC */
+		if (cmd_buffer->device->instance->physicalDevice.rad_info.chip_class >= VI) {
+			radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+			radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_DATA_TS) |
+			                            EVENT_INDEX(5));
+			radeon_emit(cmd_buffer->cs, 0);
+			radeon_emit(cmd_buffer->cs, 0);
+			radeon_emit(cmd_buffer->cs, 0);
+			radeon_emit(cmd_buffer->cs, 0);
+		}
+	}
+
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
+		cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
+			S_0085F0_DB_DEST_BASE_ENA(1);
+	}
+
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB_META) {
+		radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+		radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
+	}
+
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB_META) {
+		radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+		radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
+	}
+
+	if (!(cmd_buffer->state.flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+					      RADV_CMD_FLAG_FLUSH_AND_INV_DB))) {
+		if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_PS_PARTIAL_FLUSH) {
+			radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+			radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
+		} else if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_VS_PARTIAL_FLUSH) {
+			radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+			radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
+		}
+	}
+
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_CS_PARTIAL_FLUSH) {
+		radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+		radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
+	}
+
+	/* VGT state sync */
+	if (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_VGT_FLUSH) {
+		radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+		radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
+	}
+
+	/* Make sure ME is idle (it executes most packets) before continuing.
+	 * This prevents read-after-write hazards between PFP and ME.
+	 */
+	if (cp_coher_cntl || (cmd_buffer->state.flush_bits & RADV_CMD_FLAG_CS_PARTIAL_FLUSH)) {
+		radeon_emit(cmd_buffer->cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
+		radeon_emit(cmd_buffer->cs, 0);
+	}
+
+	/* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
+	 * Therefore, it should be last. Done in PFP.
+	 */
+	if (cp_coher_cntl) {
+		/* ACQUIRE_MEM is only required on a compute ring. */
+		radeon_emit(cmd_buffer->cs, PKT3(PKT3_SURFACE_SYNC, 3, 0));
+		radeon_emit(cmd_buffer->cs, cp_coher_cntl);   /* CP_COHER_CNTL */
+		radeon_emit(cmd_buffer->cs, 0xffffffff);      /* CP_COHER_SIZE */
+		radeon_emit(cmd_buffer->cs, 0);               /* CP_COHER_BASE */
+		radeon_emit(cmd_buffer->cs, 0x0000000A);      /* POLL_INTERVAL */
+	}
+
+	cmd_buffer->state.flush_bits = 0;
+}
+
+
+/* Set this if you want the 3D engine to wait until CP DMA is done.
+ * It should be set on the last CP DMA packet. */
+#define R600_CP_DMA_SYNC	(1 << 0) /* R600+ */
+
+/* Set this if the source data was used as a destination in a previous CP DMA
+ * packet. It's for preventing a read-after-write (RAW) hazard between two
+ * CP DMA packets. */
+#define SI_CP_DMA_RAW_WAIT	(1 << 1) /* SI+ */
+#define CIK_CP_DMA_USE_L2	(1 << 2)
+
+/* Alignment for optimal performance. */
+#define CP_DMA_ALIGNMENT	32
+/* The max number of bytes to copy per packet. */
+#define CP_DMA_MAX_BYTE_COUNT	((1 << 21) - CP_DMA_ALIGNMENT)
+
+static void si_emit_cp_dma_copy_buffer(struct radv_cmd_buffer *cmd_buffer,
+				       uint64_t dst_va, uint64_t src_va,
+				       unsigned size, unsigned flags)
+{
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+	uint32_t sync_flag = flags & R600_CP_DMA_SYNC ? S_411_CP_SYNC(1) : 0;
+	uint32_t wr_confirm = !(flags & R600_CP_DMA_SYNC) ? S_414_DISABLE_WR_CONFIRM(1) : 0;
+	uint32_t raw_wait = flags & SI_CP_DMA_RAW_WAIT ? S_414_RAW_WAIT(1) : 0;
+	uint32_t sel = flags & CIK_CP_DMA_USE_L2 ?
+			   S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2) |
+			   S_411_DSL_SEL(V_411_DST_ADDR_TC_L2) : 0;
+
+	assert(size);
+	assert((size & ((1<<21)-1)) == size);
+
+	radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 9);
+
+	if (cmd_buffer->device->instance->physicalDevice.rad_info.chip_class >= CIK) {
+		radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0));
+		radeon_emit(cs, sync_flag | sel);	/* CP_SYNC [31] */
+		radeon_emit(cs, src_va);		/* SRC_ADDR_LO [31:0] */
+		radeon_emit(cs, src_va >> 32);		/* SRC_ADDR_HI [31:0] */
+		radeon_emit(cs, dst_va);		/* DST_ADDR_LO [31:0] */
+		radeon_emit(cs, dst_va >> 32);		/* DST_ADDR_HI [31:0] */
+		radeon_emit(cs, size | wr_confirm | raw_wait);	/* COMMAND [29:22] | BYTE_COUNT [20:0] */
+	} else {
+		radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
+		radeon_emit(cs, src_va);			/* SRC_ADDR_LO [31:0] */
+		radeon_emit(cs, sync_flag | ((src_va >> 32) & 0xffff)); /* CP_SYNC [31] | SRC_ADDR_HI [15:0] */
+		radeon_emit(cs, dst_va);			/* DST_ADDR_LO [31:0] */
+		radeon_emit(cs, (dst_va >> 32) & 0xffff);	/* DST_ADDR_HI [15:0] */
+		radeon_emit(cs, size | wr_confirm | raw_wait);	/* COMMAND [29:22] | BYTE_COUNT [20:0] */
+	}
+
+	/* CP DMA is executed in ME, but index buffers are read by PFP.
+	 * This ensures that ME (CP DMA) is idle before PFP starts fetching
+	 * indices. If we wanted to execute CP DMA in PFP, this packet
+	 * should precede it.
+	 */
+	if (sync_flag) {
+		radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
+		radeon_emit(cs, 0);
+	}
+}
+
+/* Emit a CP DMA packet to clear a buffer. The size must fit in bits [20:0]. */
+static void si_emit_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer,
+					uint64_t dst_va, unsigned size,
+					uint32_t clear_value, unsigned flags)
+{
+	struct radeon_winsys_cs *cs = cmd_buffer->cs;
+	uint32_t sync_flag = flags & R600_CP_DMA_SYNC ? S_411_CP_SYNC(1) : 0;
+	uint32_t wr_confirm = !(flags & R600_CP_DMA_SYNC) ? S_414_DISABLE_WR_CONFIRM(1) : 0;
+	uint32_t raw_wait = flags & SI_CP_DMA_RAW_WAIT ? S_414_RAW_WAIT(1) : 0;
+	uint32_t dst_sel = flags & CIK_CP_DMA_USE_L2 ? S_411_DSL_SEL(V_411_DST_ADDR_TC_L2) : 0;
+
+	assert(size);
+	assert((size & ((1<<21)-1)) == size);
+
+	radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 9);
+
+	if (cmd_buffer->device->instance->physicalDevice.rad_info.chip_class >= CIK) {
+		radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0));
+		radeon_emit(cs, sync_flag | dst_sel | S_411_SRC_SEL(V_411_DATA)); /* CP_SYNC [31] | SRC_SEL[30:29] */
+		radeon_emit(cs, clear_value);		/* DATA [31:0] */
+		radeon_emit(cs, 0);
+		radeon_emit(cs, dst_va);		/* DST_ADDR_LO [31:0] */
+		radeon_emit(cs, dst_va >> 32);		/* DST_ADDR_HI [15:0] */
+		radeon_emit(cs, size | wr_confirm | raw_wait);	/* COMMAND [29:22] | BYTE_COUNT [20:0] */
+	} else {
+		radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
+		radeon_emit(cs, clear_value);		/* DATA [31:0] */
+		radeon_emit(cs, sync_flag | S_411_SRC_SEL(V_411_DATA)); /* CP_SYNC [31] | SRC_SEL[30:29] */
+		radeon_emit(cs, dst_va);			/* DST_ADDR_LO [31:0] */
+		radeon_emit(cs, (dst_va >> 32) & 0xffff);	/* DST_ADDR_HI [15:0] */
+		radeon_emit(cs, size | wr_confirm | raw_wait);	/* COMMAND [29:22] | BYTE_COUNT [20:0] */
+	}
+
+	/* See "copy_buffer" for explanation. */
+	if (sync_flag) {
+		radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
+		radeon_emit(cs, 0);
+	}
+}
+
+static void si_cp_dma_prepare(struct radv_cmd_buffer *cmd_buffer, uint64_t byte_count,
+			      uint64_t remaining_size, unsigned *flags)
+{
+
+	/* Flush the caches for the first copy only.
+	 * Also wait for the previous CP DMA operations.
+	 */
+	if (cmd_buffer->state.flush_bits) {
+		si_emit_cache_flush(cmd_buffer);
+		*flags |= SI_CP_DMA_RAW_WAIT;
+	}
+
+	/* Do the synchronization after the last dma, so that all data
+	 * is written to memory.
+	 */
+	if (byte_count == remaining_size)
+		*flags |= R600_CP_DMA_SYNC;
+}
+
+static void si_cp_dma_realign_engine(struct radv_cmd_buffer *cmd_buffer, unsigned size)
+{
+	uint64_t va;
+	uint32_t offset;
+	unsigned dma_flags = 0;
+	unsigned buf_size = CP_DMA_ALIGNMENT * 2;
+	void *ptr;
+
+	assert(size < CP_DMA_ALIGNMENT);
+
+	radv_cmd_buffer_upload_alloc(cmd_buffer, buf_size, CP_DMA_ALIGNMENT,  &offset, &ptr);
+
+	va = cmd_buffer->device->ws->buffer_get_va(cmd_buffer->upload.upload_bo);
+	va += offset;
+
+	si_cp_dma_prepare(cmd_buffer, size, size, &dma_flags);
+
+	si_emit_cp_dma_copy_buffer(cmd_buffer, va, va + CP_DMA_ALIGNMENT, size,
+				   dma_flags);
+}
+
+void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
+			   uint64_t src_va, uint64_t dest_va,
+			   uint64_t size)
+{
+	uint64_t main_src_va, main_dest_va;
+	uint64_t skipped_size = 0, realign_size = 0;
+
+
+	if (cmd_buffer->device->instance->physicalDevice.rad_info.family <= CHIP_CARRIZO ||
+	    cmd_buffer->device->instance->physicalDevice.rad_info.family == CHIP_STONEY) {
+		/* If the size is not aligned, we must add a dummy copy at the end
+		 * just to align the internal counter. Otherwise, the DMA engine
+		 * would slow down by an order of magnitude for following copies.
+		 */
+		if (size % CP_DMA_ALIGNMENT)
+			realign_size = CP_DMA_ALIGNMENT - (size % CP_DMA_ALIGNMENT);
+
+		/* If the copy begins unaligned, we must start copying from the next
+		 * aligned block and the skipped part should be copied after everything
+		 * else has been copied. Only the src alignment matters, not dst.
+		 */
+		if (src_va % CP_DMA_ALIGNMENT) {
+			skipped_size = CP_DMA_ALIGNMENT - (src_va % CP_DMA_ALIGNMENT);
+			/* The main part will be skipped if the size is too small. */
+			skipped_size = MIN2(skipped_size, size);
+			size -= skipped_size;
+		}
+	}
+	main_src_va = src_va + skipped_size;
+	main_dest_va = dest_va + skipped_size;
+
+	while (size) {
+		unsigned dma_flags = 0;
+		unsigned byte_count = MIN2(size, CP_DMA_MAX_BYTE_COUNT);
+
+		si_cp_dma_prepare(cmd_buffer, byte_count,
+				  size + skipped_size + realign_size,
+				  &dma_flags);
+
+		si_emit_cp_dma_copy_buffer(cmd_buffer, main_dest_va, main_src_va,
+					   byte_count, dma_flags);
+
+		size -= byte_count;
+		main_src_va += byte_count;
+		main_dest_va += byte_count;
+	}
+
+	if (skipped_size) {
+		unsigned dma_flags = 0;
+
+		si_cp_dma_prepare(cmd_buffer, skipped_size,
+				  size + skipped_size + realign_size,
+				  &dma_flags);
+
+		si_emit_cp_dma_copy_buffer(cmd_buffer, dest_va, src_va,
+					   skipped_size, dma_flags);
+	}
+	if (realign_size)
+		si_cp_dma_realign_engine(cmd_buffer, realign_size);
+}
+
+void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
+			    uint64_t size, unsigned value)
+{
+
+	if (!size)
+		return;
+
+	assert(va % 4 == 0 && size % 4 == 0);
+
+	while (size) {
+		unsigned byte_count = MIN2(size, CP_DMA_MAX_BYTE_COUNT);
+		unsigned dma_flags = 0;
+
+		si_cp_dma_prepare(cmd_buffer, byte_count, size, &dma_flags);
+
+		/* Emit the clear packet. */
+		si_emit_cp_dma_clear_buffer(cmd_buffer, va, byte_count, value,
+					    dma_flags);
+
+		size -= byte_count;
+		va += byte_count;
+	}
+}
+
+/* For MSAA sample positions. */
+#define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y)  \
+	(((s0x) & 0xf) | (((unsigned)(s0y) & 0xf) << 4) |		   \
+	(((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) |	   \
+	(((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) |	   \
+	 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
+
+
+/* 2xMSAA
+ * There are two locations (4, 4), (-4, -4). */
+const uint32_t eg_sample_locs_2x[4] = {
+	FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
+	FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
+	FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
+	FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
+};
+const unsigned eg_max_dist_2x = 4;
+/* 4xMSAA
+ * There are 4 locations: (-2, 6), (6, -2), (-6, 2), (2, 6). */
+const uint32_t eg_sample_locs_4x[4] = {
+	FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
+	FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
+	FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
+	FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
+};
+const unsigned eg_max_dist_4x = 6;
+
+/* Cayman 8xMSAA */
+static const uint32_t cm_sample_locs_8x[] = {
+	FILL_SREG( 1, -3, -1,  3, 5,  1, -3, -5),
+	FILL_SREG( 1, -3, -1,  3, 5,  1, -3, -5),
+	FILL_SREG( 1, -3, -1,  3, 5,  1, -3, -5),
+	FILL_SREG( 1, -3, -1,  3, 5,  1, -3, -5),
+	FILL_SREG(-5,  5, -7, -1, 3,  7,  7, -7),
+	FILL_SREG(-5,  5, -7, -1, 3,  7,  7, -7),
+	FILL_SREG(-5,  5, -7, -1, 3,  7,  7, -7),
+	FILL_SREG(-5,  5, -7, -1, 3,  7,  7, -7),
+};
+static const unsigned cm_max_dist_8x = 8;
+/* Cayman 16xMSAA */
+static const uint32_t cm_sample_locs_16x[] = {
+	FILL_SREG( 1,  1, -1, -3, -3,  2,  4, -1),
+	FILL_SREG( 1,  1, -1, -3, -3,  2,  4, -1),
+	FILL_SREG( 1,  1, -1, -3, -3,  2,  4, -1),
+	FILL_SREG( 1,  1, -1, -3, -3,  2,  4, -1),
+	FILL_SREG(-5, -2,  2,  5,  5,  3,  3, -5),
+	FILL_SREG(-5, -2,  2,  5,  5,  3,  3, -5),
+	FILL_SREG(-5, -2,  2,  5,  5,  3,  3, -5),
+	FILL_SREG(-5, -2,  2,  5,  5,  3,  3, -5),
+	FILL_SREG(-2,  6,  0, -7, -4, -6, -6,  4),
+	FILL_SREG(-2,  6,  0, -7, -4, -6, -6,  4),
+	FILL_SREG(-2,  6,  0, -7, -4, -6, -6,  4),
+	FILL_SREG(-2,  6,  0, -7, -4, -6, -6,  4),
+	FILL_SREG(-8,  0,  7, -4,  6,  7, -7, -8),
+	FILL_SREG(-8,  0,  7, -4,  6,  7, -7, -8),
+	FILL_SREG(-8,  0,  7, -4,  6,  7, -7, -8),
+	FILL_SREG(-8,  0,  7, -4,  6,  7, -7, -8),
+};
+static const unsigned cm_max_dist_16x = 8;
+
+unsigned radv_cayman_get_maxdist(int log_samples)
+{
+	unsigned max_dist[] = {
+		0,
+		eg_max_dist_2x,
+		eg_max_dist_4x,
+		cm_max_dist_8x,
+		cm_max_dist_16x
+	};
+	return max_dist[log_samples];
+}
+
+void radv_cayman_emit_msaa_sample_locs(struct radeon_winsys_cs *cs, int nr_samples)
+{
+	switch (nr_samples) {
+	default:
+	case 1:
+		radeon_set_context_reg(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 0);
+		radeon_set_context_reg(cs, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, 0);
+		radeon_set_context_reg(cs, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, 0);
+		radeon_set_context_reg(cs, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, 0);
+		break;
+	case 2:
+		radeon_set_context_reg(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_2x[0]);
+		radeon_set_context_reg(cs, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_2x[1]);
+		radeon_set_context_reg(cs, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_2x[2]);
+		radeon_set_context_reg(cs, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_2x[3]);
+		break;
+	case 4:
+		radeon_set_context_reg(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_4x[0]);
+		radeon_set_context_reg(cs, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_4x[1]);
+		radeon_set_context_reg(cs, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_4x[2]);
+		radeon_set_context_reg(cs, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_4x[3]);
+		break;
+	case 8:
+		radeon_set_context_reg_seq(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 14);
+		radeon_emit(cs, cm_sample_locs_8x[0]);
+		radeon_emit(cs, cm_sample_locs_8x[4]);
+		radeon_emit(cs, 0);
+		radeon_emit(cs, 0);
+		radeon_emit(cs, cm_sample_locs_8x[1]);
+		radeon_emit(cs, cm_sample_locs_8x[5]);
+		radeon_emit(cs, 0);
+		radeon_emit(cs, 0);
+		radeon_emit(cs, cm_sample_locs_8x[2]);
+		radeon_emit(cs, cm_sample_locs_8x[6]);
+		radeon_emit(cs, 0);
+		radeon_emit(cs, 0);
+		radeon_emit(cs, cm_sample_locs_8x[3]);
+		radeon_emit(cs, cm_sample_locs_8x[7]);
+		break;
+	case 16:
+		radeon_set_context_reg_seq(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 16);
+		radeon_emit(cs, cm_sample_locs_16x[0]);
+		radeon_emit(cs, cm_sample_locs_16x[4]);
+		radeon_emit(cs, cm_sample_locs_16x[8]);
+		radeon_emit(cs, cm_sample_locs_16x[12]);
+		radeon_emit(cs, cm_sample_locs_16x[1]);
+		radeon_emit(cs, cm_sample_locs_16x[5]);
+		radeon_emit(cs, cm_sample_locs_16x[9]);
+		radeon_emit(cs, cm_sample_locs_16x[13]);
+		radeon_emit(cs, cm_sample_locs_16x[2]);
+		radeon_emit(cs, cm_sample_locs_16x[6]);
+		radeon_emit(cs, cm_sample_locs_16x[10]);
+		radeon_emit(cs, cm_sample_locs_16x[14]);
+		radeon_emit(cs, cm_sample_locs_16x[3]);
+		radeon_emit(cs, cm_sample_locs_16x[7]);
+		radeon_emit(cs, cm_sample_locs_16x[11]);
+		radeon_emit(cs, cm_sample_locs_16x[15]);
+		break;
+	}
+}
+
+static void radv_cayman_get_sample_position(struct radv_device *device,
+					    unsigned sample_count,
+					    unsigned sample_index, float *out_value)
+{
+	int offset, index;
+	struct {
+		int idx:4;
+	} val;
+	switch (sample_count) {
+	case 1:
+	default:
+		out_value[0] = out_value[1] = 0.5;
+		break;
+	case 2:
+		offset = 4 * (sample_index * 2);
+		val.idx = (eg_sample_locs_2x[0] >> offset) & 0xf;
+		out_value[0] = (float)(val.idx + 8) / 16.0f;
+		val.idx = (eg_sample_locs_2x[0] >> (offset + 4)) & 0xf;
+		out_value[1] = (float)(val.idx + 8) / 16.0f;
+		break;
+	case 4:
+		offset = 4 * (sample_index * 2);
+		val.idx = (eg_sample_locs_4x[0] >> offset) & 0xf;
+		out_value[0] = (float)(val.idx + 8) / 16.0f;
+		val.idx = (eg_sample_locs_4x[0] >> (offset + 4)) & 0xf;
+		out_value[1] = (float)(val.idx + 8) / 16.0f;
+		break;
+	case 8:
+		offset = 4 * (sample_index % 4 * 2);
+		index = (sample_index / 4) * 4;
+		val.idx = (cm_sample_locs_8x[index] >> offset) & 0xf;
+		out_value[0] = (float)(val.idx + 8) / 16.0f;
+		val.idx = (cm_sample_locs_8x[index] >> (offset + 4)) & 0xf;
+		out_value[1] = (float)(val.idx + 8) / 16.0f;
+		break;
+	case 16:
+		offset = 4 * (sample_index % 4 * 2);
+		index = (sample_index / 4) * 4;
+		val.idx = (cm_sample_locs_16x[index] >> offset) & 0xf;
+		out_value[0] = (float)(val.idx + 8) / 16.0f;
+		val.idx = (cm_sample_locs_16x[index] >> (offset + 4)) & 0xf;
+		out_value[1] = (float)(val.idx + 8) / 16.0f;
+		break;
+	}
+}
+
+void radv_device_init_msaa(struct radv_device *device)
+{
+	int i;
+	radv_cayman_get_sample_position(device, 1, 0, device->sample_locations_1x[0]);
+
+	for (i = 0; i < 2; i++)
+		radv_cayman_get_sample_position(device, 2, i, device->sample_locations_2x[i]);
+	for (i = 0; i < 4; i++)
+		radv_cayman_get_sample_position(device, 4, i, device->sample_locations_4x[i]);
+	for (i = 0; i < 8; i++)
+		radv_cayman_get_sample_position(device, 8, i, device->sample_locations_8x[i]);
+	for (i = 0; i < 16; i++)
+		radv_cayman_get_sample_position(device, 16, i, device->sample_locations_16x[i]);
+}
diff --git a/src/amd/vulkan/vk_format.h b/src/amd/vulkan/vk_format.h
new file mode 100644
index 0000000000..e0087f1dea
--- /dev/null
+++ b/src/amd/vulkan/vk_format.h
@@ -0,0 +1,449 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * Based on u_format.h which is:
+ * Copyright 2009-2010 Vmware, Inc.
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#pragma once
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <assert.h>
+#include <vulkan/vulkan.h>
+enum vk_format_layout {
+	/**
+	 * Formats with vk_format_block::width == vk_format_block::height == 1
+	 * that can be described as an ordinary data structure.
+	 */
+	VK_FORMAT_LAYOUT_PLAIN = 0,
+
+	/**
+	 * Formats with sub-sampled channels.
+	 *
+	 * This is for formats like YVYU where there is less than one sample per
+	 * pixel.
+	 */
+	VK_FORMAT_LAYOUT_SUBSAMPLED = 3,
+
+	/**
+	 * S3 Texture Compression formats.
+	 */
+	VK_FORMAT_LAYOUT_S3TC = 4,
+
+	/**
+	 * Red-Green Texture Compression formats.
+	 */
+	VK_FORMAT_LAYOUT_RGTC = 5,
+
+	/**
+	 * Ericsson Texture Compression
+	 */
+	VK_FORMAT_LAYOUT_ETC = 6,
+
+	/**
+	 * BC6/7 Texture Compression
+	 */
+	VK_FORMAT_LAYOUT_BPTC = 7,
+
+	/**
+	 * ASTC
+	 */
+	VK_FORMAT_LAYOUT_ASTC = 8,
+
+	/**
+	 * Everything else that doesn't fit in any of the above layouts.
+	 */
+	VK_FORMAT_LAYOUT_OTHER = 9
+};
+
+struct vk_format_block
+{
+	/** Block width in pixels */
+	unsigned width;
+
+	/** Block height in pixels */
+	unsigned height;
+
+	/** Block size in bits */
+	unsigned bits;
+};
+
+enum vk_format_type {
+	VK_FORMAT_TYPE_VOID = 0,
+	VK_FORMAT_TYPE_UNSIGNED = 1,
+	VK_FORMAT_TYPE_SIGNED = 2,
+	VK_FORMAT_TYPE_FIXED = 3,
+	VK_FORMAT_TYPE_FLOAT = 4
+};
+
+
+enum vk_format_colorspace {
+	VK_FORMAT_COLORSPACE_RGB = 0,
+	VK_FORMAT_COLORSPACE_SRGB = 1,
+	VK_FORMAT_COLORSPACE_YUV = 2,
+	VK_FORMAT_COLORSPACE_ZS = 3
+};
+
+struct vk_format_channel_description {
+	unsigned type:5;
+	unsigned normalized:1;
+	unsigned pure_integer:1;
+	unsigned scaled:1;
+	unsigned size:8;
+	unsigned shift:16;
+};
+
+struct vk_format_description
+{
+	VkFormat format;
+	const char *name;
+	const char *short_name;
+
+	struct vk_format_block block;
+	enum vk_format_layout layout;
+
+	unsigned nr_channels:3;
+	unsigned is_array:1;
+	unsigned is_bitmask:1;
+	unsigned is_mixed:1;
+
+	struct vk_format_channel_description channel[4];
+
+	unsigned char swizzle[4];
+
+	enum vk_format_colorspace colorspace;
+};
+
+extern const struct vk_format_description vk_format_description_table[];
+
+const struct vk_format_description *vk_format_description(VkFormat format);
+
+/**
+ * Return total bits needed for the pixel format per block.
+ */
+static inline uint
+vk_format_get_blocksizebits(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+
+	assert(desc);
+	if (!desc) {
+		return 0;
+	}
+
+	return desc->block.bits;
+}
+
+/**
+ * Return bytes per block (not pixel) for the given format.
+ */
+static inline uint
+vk_format_get_blocksize(VkFormat format)
+{
+	uint bits = vk_format_get_blocksizebits(format);
+	uint bytes = bits / 8;
+
+	assert(bits % 8 == 0);
+	assert(bytes > 0);
+	if (bytes == 0) {
+		bytes = 1;
+	}
+
+	return bytes;
+}
+
+static inline uint
+vk_format_get_blockwidth(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+
+	assert(desc);
+	if (!desc) {
+		return 1;
+	}
+
+	return desc->block.width;
+}
+
+static inline uint
+vk_format_get_blockheight(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+
+	assert(desc);
+	if (!desc) {
+		return 1;
+	}
+
+	return desc->block.height;
+}
+
+/**
+ * Return the index of the first non-void channel
+ * -1 if no non-void channels
+ */
+static inline int
+vk_format_get_first_non_void_channel(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+	int i;
+
+	for (i = 0; i < 4; i++)
+		if (desc->channel[i].type != VK_FORMAT_TYPE_VOID)
+			break;
+
+	if (i == 4)
+		return -1;
+
+	return i;
+}
+
+enum vk_swizzle {
+	VK_SWIZZLE_X,
+	VK_SWIZZLE_Y,
+	VK_SWIZZLE_Z,
+	VK_SWIZZLE_W,
+	VK_SWIZZLE_0,
+	VK_SWIZZLE_1,
+	VK_SWIZZLE_NONE,
+	VK_SWIZZLE_MAX, /**< Number of enums counter (must be last) */
+};
+
+static inline VkImageAspectFlags
+vk_format_aspects(VkFormat format)
+{
+	switch (format) {
+	case VK_FORMAT_UNDEFINED:
+		return 0;
+
+	case VK_FORMAT_S8_UINT:
+		return VK_IMAGE_ASPECT_STENCIL_BIT;
+
+	case VK_FORMAT_D16_UNORM_S8_UINT:
+	case VK_FORMAT_D24_UNORM_S8_UINT:
+	case VK_FORMAT_D32_SFLOAT_S8_UINT:
+		return VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT;
+
+	case VK_FORMAT_D16_UNORM:
+	case VK_FORMAT_X8_D24_UNORM_PACK32:
+	case VK_FORMAT_D32_SFLOAT:
+		return VK_IMAGE_ASPECT_DEPTH_BIT;
+
+	default:
+		return VK_IMAGE_ASPECT_COLOR_BIT;
+	}
+}
+
+static inline enum vk_swizzle
+radv_swizzle_conv(int idx, const unsigned char chan[4], VkComponentSwizzle vk_swiz)
+{
+	int x;
+	switch (vk_swiz) {
+	case VK_COMPONENT_SWIZZLE_IDENTITY:
+		return chan[idx];
+	case VK_COMPONENT_SWIZZLE_ZERO:
+		return VK_SWIZZLE_0;
+	case VK_COMPONENT_SWIZZLE_ONE:
+		return VK_SWIZZLE_1;
+	case VK_COMPONENT_SWIZZLE_R:
+		for (x = 0; x < 4; x++)
+			if (chan[x] == 0)
+				return x;
+		return VK_SWIZZLE_0;
+	case VK_COMPONENT_SWIZZLE_G:
+		for (x = 0; x < 4; x++)
+			if (chan[x] == 1)
+				return x;
+		return VK_SWIZZLE_0;
+	case VK_COMPONENT_SWIZZLE_B:
+		for (x = 0; x < 4; x++)
+			if (chan[x] == 2)
+				return x;
+		return VK_SWIZZLE_0;
+	case VK_COMPONENT_SWIZZLE_A:
+		for (x = 0; x < 4; x++)
+			if (chan[x] == 3)
+				return x;
+		return VK_SWIZZLE_1;
+	default:
+		return chan[idx];
+	}
+}
+
+static inline void vk_format_compose_swizzles(const VkComponentMapping *mapping,
+					      const unsigned char swz[4],
+					      enum vk_swizzle dst[4])
+{
+	dst[0] = radv_swizzle_conv(0, swz, mapping->r);
+	dst[1] = radv_swizzle_conv(1, swz, mapping->g);
+	dst[2] = radv_swizzle_conv(2, swz, mapping->b);
+	dst[3] = radv_swizzle_conv(3, swz, mapping->a);
+}
+
+static inline bool
+vk_format_is_compressed(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+
+	assert(desc);
+	if (!desc) {
+		return false;
+	}
+
+	switch (desc->layout) {
+	case VK_FORMAT_LAYOUT_S3TC:
+	case VK_FORMAT_LAYOUT_RGTC:
+	case VK_FORMAT_LAYOUT_ETC:
+	case VK_FORMAT_LAYOUT_BPTC:
+	case VK_FORMAT_LAYOUT_ASTC:
+		/* XXX add other formats in the future */
+		return true;
+	default:
+		return false;
+	}
+}
+
+static inline bool
+vk_format_has_depth(const struct vk_format_description *desc)
+{
+	return desc->colorspace == VK_FORMAT_COLORSPACE_ZS &&
+		desc->swizzle[0] != VK_SWIZZLE_NONE;
+}
+
+static inline bool
+vk_format_has_stencil(const struct vk_format_description *desc)
+{
+	return desc->colorspace == VK_FORMAT_COLORSPACE_ZS &&
+		desc->swizzle[1] != VK_SWIZZLE_NONE;
+}
+
+static inline bool
+vk_format_is_depth_or_stencil(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+
+	assert(desc);
+	if (!desc) {
+		return false;
+	}
+
+	return vk_format_has_depth(desc) ||
+		vk_format_has_stencil(desc);
+}
+
+static inline bool
+vk_format_is_depth(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+
+	assert(desc);
+	if (!desc) {
+		return false;
+	}
+
+	return vk_format_has_depth(desc);
+}
+
+static inline bool
+vk_format_is_color(VkFormat format)
+{
+	return !vk_format_is_depth_or_stencil(format);
+}
+
+static inline VkFormat
+vk_format_depth_only(VkFormat format)
+{
+	switch (format) {
+	case VK_FORMAT_D16_UNORM_S8_UINT:
+		return VK_FORMAT_D16_UNORM;
+	case VK_FORMAT_D24_UNORM_S8_UINT:
+		return VK_FORMAT_X8_D24_UNORM_PACK32;
+	case VK_FORMAT_D32_SFLOAT_S8_UINT:
+		return VK_FORMAT_D32_SFLOAT;
+	default:
+		return format;
+	}
+}
+
+static inline bool
+vk_format_is_int(VkFormat format)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+	int channel =  vk_format_get_first_non_void_channel(format);
+
+	return channel >= 0 && desc->channel[channel].pure_integer;
+}
+
+static inline VkFormat
+vk_format_stencil_only(VkFormat format)
+{
+	return VK_FORMAT_S8_UINT;
+}
+
+static inline uint
+vk_format_get_component_bits(VkFormat format,
+			     enum vk_format_colorspace colorspace,
+			     uint component)
+{
+	const struct vk_format_description *desc = vk_format_description(format);
+	enum vk_format_colorspace desc_colorspace;
+
+	assert(format);
+	if (!format) {
+		return 0;
+	}
+
+	assert(component < 4);
+
+	/* Treat RGB and SRGB as equivalent. */
+	if (colorspace == VK_FORMAT_COLORSPACE_SRGB) {
+		colorspace = VK_FORMAT_COLORSPACE_RGB;
+	}
+	if (desc->colorspace == VK_FORMAT_COLORSPACE_SRGB) {
+		desc_colorspace = VK_FORMAT_COLORSPACE_RGB;
+	} else {
+		desc_colorspace = desc->colorspace;
+	}
+
+	if (desc_colorspace != colorspace) {
+		return 0;
+	}
+
+	switch (desc->swizzle[component]) {
+	case VK_SWIZZLE_X:
+		return desc->channel[0].size;
+	case VK_SWIZZLE_Y:
+		return desc->channel[1].size;
+	case VK_SWIZZLE_Z:
+		return desc->channel[2].size;
+	case VK_SWIZZLE_W:
+		return desc->channel[3].size;
+	default:
+		return 0;
+	}
+}
+#ifdef __cplusplus
+} // extern "C" {
+#endif
diff --git a/src/amd/vulkan/vk_format_layout.csv b/src/amd/vulkan/vk_format_layout.csv
new file mode 100644
index 0000000000..ae9ceda08e
--- /dev/null
+++ b/src/amd/vulkan/vk_format_layout.csv
@@ -0,0 +1,188 @@
+/* this is pretty much taken from the gallium one. */
+
+
+VK_FORMAT_UNDEFINED                  , plain, 1, 1, u8  ,     ,     ,     , x001, rgb
+VK_FORMAT_R4G4_UNORM_PACK8           , plain, 1, 1, un4 , un4 ,     ,     , xy01, rgb
+VK_FORMAT_R4G4B4A4_UNORM_PACK16      , plain, 1, 1, un4 , un4 , un4 , un4 , wzyx, rgb
+VK_FORMAT_B4G4R4A4_UNORM_PACK16      , plain, 1, 1, un4 , un4 , un4 , un4 , wxyz, rgb
+VK_FORMAT_R5G6B5_UNORM_PACK16        , plain, 1, 1, un5 , un6 , un5 ,     , zyx1, rgb
+VK_FORMAT_B5G6R5_UNORM_PACK16        , plain, 1, 1, un5 , un6 , un5 ,     , xyz1, rgb
+VK_FORMAT_R5G5B5A1_UNORM_PACK16      , plain, 1, 1, un1 , un5 , un5 , un5 , wzyx, rgb
+VK_FORMAT_B5G5R5A1_UNORM_PACK16      , plain, 1, 1, un1 , un5 , un5 , un5 , wxyz, rgb
+VK_FORMAT_A1R5G5B5_UNORM_PACK16      , plain, 1, 1, un5 , un5 , un5 , un1 , zyxw, rgb
+VK_FORMAT_R8_UNORM                   , plain, 1, 1, un8 ,     ,     ,     , x001, rgb
+VK_FORMAT_R8_SNORM                   , plain, 1, 1, sn8 ,     ,     ,     , x001, rgb
+VK_FORMAT_R8_USCALED                 , plain, 1, 1, us8 ,     ,     ,     , x001, rgb
+VK_FORMAT_R8_SSCALED                 , plain, 1, 1, ss8 ,     ,     ,     , x001, rgb
+VK_FORMAT_R8_UINT                    , plain, 1, 1, up8 ,     ,     ,     , x001, rgb
+VK_FORMAT_R8_SINT                    , plain, 1, 1, sp8 ,     ,     ,     , x001, rgb
+VK_FORMAT_R8_SRGB                    , plain, 1, 1, un8 ,     ,     ,     , x001, srgb
+VK_FORMAT_R8G8_UNORM                 , plain, 1, 1, un8 , un8 ,     ,     , xy01, rgb
+VK_FORMAT_R8G8_SNORM                 , plain, 1, 1, sn8 , sn8 ,     ,     , xy01, rgb
+VK_FORMAT_R8G8_USCALED               , plain, 1, 1, us8 , us8 ,     ,     , xy01, rgb
+VK_FORMAT_R8G8_SSCALED               , plain, 1, 1, ss8 , ss8 ,     ,     , xy01, rgb
+VK_FORMAT_R8G8_UINT                  , plain, 1, 1, up8 , up8 ,     ,     , xy01, rgb
+VK_FORMAT_R8G8_SINT                  , plain, 1, 1, sp8 , sp8 ,     ,     , xy01, rgb
+VK_FORMAT_R8G8_SRGB                  , plain, 1, 1, un8 , un8 ,     ,     , xy01, srgb
+VK_FORMAT_R8G8B8_UNORM               , plain, 1, 1, un8 , un8 , un8 ,     , xyz1, rgb
+VK_FORMAT_R8G8B8_SNORM               , plain, 1, 1, sn8 , sn8 , sn8 ,     , xyz1, rgb
+VK_FORMAT_R8G8B8_USCALED             , plain, 1, 1, us8 , us8 , us8 ,     , xyz1, rgb
+VK_FORMAT_R8G8B8_SSCALED             , plain, 1, 1, ss8 , ss8 , ss8 ,     , xyz1, rgb
+VK_FORMAT_R8G8B8_UINT                , plain, 1, 1, up8 , up8 , up8 ,     , xyz1, rgb
+VK_FORMAT_R8G8B8_SINT                , plain, 1, 1, sp8 , sp8 , sp8 ,     , xyz1, rgb
+VK_FORMAT_R8G8B8_SRGB                , plain, 1, 1, un8 , un8 , un8 ,     , xyz1, srgb
+VK_FORMAT_B8G8R8_UNORM               , plain, 1, 1, un8 , un8 , un8 ,     , zyx1, rgb
+VK_FORMAT_B8G8R8_SNORM               , plain, 1, 1, sn8 , sn8 , sn8 ,     , zyx1, rgb
+VK_FORMAT_B8G8R8_USCALED             , plain, 1, 1, us8 , us8 , us8 ,     , zyx1, rgb
+VK_FORMAT_B8G8R8_SSCALED             , plain, 1, 1, ss8 , ss8 , ss8 ,     , zyx1, rgb
+VK_FORMAT_B8G8R8_UINT                , plain, 1, 1, up8 , up8 , up8 ,     , zyx1, rgb
+VK_FORMAT_B8G8R8_SINT                , plain, 1, 1, sp8 , sp8 , sp8 ,     , zyx1, rgb
+VK_FORMAT_B8G8R8_SRGB                , plain, 1, 1, un8 , un8 , un8 ,     , zyx1, srgb
+VK_FORMAT_R8G8B8A8_UNORM             , plain, 1, 1, un8 , un8 , un8 , un8 , xyzw, rgb
+VK_FORMAT_R8G8B8A8_SNORM             , plain, 1, 1, sn8 , sn8 , sn8 , sn8 , xyzw, rgb
+VK_FORMAT_R8G8B8A8_USCALED           , plain, 1, 1, us8 , us8 , us8 , us8 , xyzw, rgb
+VK_FORMAT_R8G8B8A8_SSCALED           , plain, 1, 1, ss8 , ss8 , ss8 , ss8 , xyzw, rgb
+VK_FORMAT_R8G8B8A8_UINT              , plain, 1, 1, up8 , up8 , up8 , up8 , xyzw, rgb
+VK_FORMAT_R8G8B8A8_SINT              , plain, 1, 1, sp8 , sp8 , sp8 , sp8 , xyzw, rgb
+VK_FORMAT_R8G8B8A8_SRGB              , plain, 1, 1, un8 , un8 , un8 , un8 , xyzw, srgb
+VK_FORMAT_B8G8R8A8_UNORM             , plain, 1, 1, un8 , un8 , un8 , un8 , zyxw, rgb
+VK_FORMAT_B8G8R8A8_SNORM             , plain, 1, 1, sn8 , sn8 , sn8 , sn8 , zyxw, rgb
+VK_FORMAT_B8G8R8A8_USCALED           , plain, 1, 1, us8 , us8 , us8 , us8 , zyxw, rgb
+VK_FORMAT_B8G8R8A8_SSCALED           , plain, 1, 1, ss8 , ss8 , ss8 , ss8 , zyxw, rgb
+VK_FORMAT_B8G8R8A8_UINT              , plain, 1, 1, up8 , up8 , up8 , up8 , zyxw, rgb
+VK_FORMAT_B8G8R8A8_SINT              , plain, 1, 1, sp8 , sp8 , sp8 , sp8 , zyxw, rgb
+VK_FORMAT_B8G8R8A8_SRGB              , plain, 1, 1, un8 , un8 , un8 , un8 , zyxw, srgb
+VK_FORMAT_A8B8G8R8_UNORM_PACK32      , plain, 1, 1, un8 , un8 , un8 , un8 , xyzw, rgb
+VK_FORMAT_A8B8G8R8_SNORM_PACK32      , plain, 1, 1, sn8 , sn8 , sn8 , sn8 , xyzw, rgb
+VK_FORMAT_A8B8G8R8_USCALED_PACK32    , plain, 1, 1, us8 , us8 , us8 , us8 , xyzw, rgb
+VK_FORMAT_A8B8G8R8_SSCALED_PACK32    , plain, 1, 1, ss8 , ss8 , ss8 , ss8 , xyzw, rgb
+VK_FORMAT_A8B8G8R8_UINT_PACK32       , plain, 1, 1, up8 , up8 , up8 , up8 , xyzw, rgb
+VK_FORMAT_A8B8G8R8_SINT_PACK32       , plain, 1, 1, sp8 , sp8 , sp8 , sp8 , xyzw, rgb
+VK_FORMAT_A8B8G8R8_SRGB_PACK32       , plain, 1, 1, un8 , un8 , un8 , un8 , xyzw, srgb
+VK_FORMAT_A2R10G10B10_UNORM_PACK32   , plain, 1, 1, un10, un10, un10, un2 , zyxw, rgb
+VK_FORMAT_A2R10G10B10_SNORM_PACK32   , plain, 1, 1, sn10, sn10, sn10, sn2 , zyxw, rgb
+VK_FORMAT_A2R10G10B10_USCALED_PACK32 , plain, 1, 1, us10, us10, us10, us2 , zyxw, rgb
+VK_FORMAT_A2R10G10B10_SSCALED_PACK32 , plain, 1, 1, ss10, ss10, ss10, ss2 , zyxw, rgb
+VK_FORMAT_A2R10G10B10_UINT_PACK32    , plain, 1, 1, up10, up10, up10, up2 , zyxw, rgb
+VK_FORMAT_A2R10G10B10_SINT_PACK32    , plain, 1, 1, sp10, sp10, sp10, sp2 , zyxw, rgb
+VK_FORMAT_A2B10G10R10_UNORM_PACK32   , plain, 1, 1, un10, un10, un10, un2 , xyzw, rgb
+VK_FORMAT_A2B10G10R10_SNORM_PACK32   , plain, 1, 1, sn10, sn10, sn10, sn2 , xyzw, rgb
+VK_FORMAT_A2B10G10R10_USCALED_PACK32 , plain, 1, 1, us10, us10, us10, us2 , xyzw, rgb
+VK_FORMAT_A2B10G10R10_SSCALED_PACK32 , plain, 1, 1, ss10, ss10, ss10, ss2 , xyzw, rgb
+VK_FORMAT_A2B10G10R10_UINT_PACK32    , plain, 1, 1, up10, up10, up10, up2 , xyzw, rgb
+VK_FORMAT_A2B10G10R10_SINT_PACK32    , plain, 1, 1, sp10, sp10, sp10, sp2 , xyzw, rgb
+VK_FORMAT_R16_UNORM                  , plain, 1, 1, un16,     ,     ,     , x001, rgb
+VK_FORMAT_R16_SNORM                  , plain, 1, 1, sn16,     ,     ,     , x001, rgb
+VK_FORMAT_R16_USCALED                , plain, 1, 1, us16,     ,     ,     , x001, rgb
+VK_FORMAT_R16_SSCALED                , plain, 1, 1, ss16,     ,     ,     , x001, rgb
+VK_FORMAT_R16_UINT                   , plain, 1, 1, up16,     ,     ,     , x001, rgb
+VK_FORMAT_R16_SINT                   , plain, 1, 1, sp16,     ,     ,     , x001, rgb
+VK_FORMAT_R16_SFLOAT                 , plain, 1, 1, f16 ,     ,     ,     , x001, rgb
+VK_FORMAT_R16G16_UNORM               , plain, 1, 1, un16, un16,     ,     , xy01, rgb
+VK_FORMAT_R16G16_SNORM               , plain, 1, 1, sn16, sn16,     ,     , xy01, rgb
+VK_FORMAT_R16G16_USCALED             , plain, 1, 1, us16, us16,     ,     , xy01, rgb
+VK_FORMAT_R16G16_SSCALED             , plain, 1, 1, ss16, ss16,     ,     , xy01, rgb
+VK_FORMAT_R16G16_UINT                , plain, 1, 1, up16, up16,     ,     , xy01, rgb
+VK_FORMAT_R16G16_SINT                , plain, 1, 1, sp16, sp16,     ,     , xy01, rgb
+VK_FORMAT_R16G16_SFLOAT              , plain, 1, 1, f16 , f16 ,     ,     , xy01, rgb
+VK_FORMAT_R16G16B16_UNORM            , plain, 1, 1, un16, un16, un16,     , xyz1, rgb
+VK_FORMAT_R16G16B16_SNORM            , plain, 1, 1, sn16, sn16, sn16,     , xyz1, rgb
+VK_FORMAT_R16G16B16_USCALED          , plain, 1, 1, us16, us16, us16,     , xyz1, rgb
+VK_FORMAT_R16G16B16_SSCALED          , plain, 1, 1, ss16, ss16, ss16,     , xyz1, rgb
+VK_FORMAT_R16G16B16_UINT             , plain, 1, 1, up16, up16, up16,     , xyz1, rgb
+VK_FORMAT_R16G16B16_SINT             , plain, 1, 1, sp16, sp16, sp16,     , xyz1, rgb
+VK_FORMAT_R16G16B16_SFLOAT           , plain, 1, 1, f16 , f16 , f16 ,     , xyz1, rgb
+VK_FORMAT_R16G16B16A16_UNORM         , plain, 1, 1, un16, un16, un16, un16, xyzw, rgb
+VK_FORMAT_R16G16B16A16_SNORM         , plain, 1, 1, sn16, sn16, sn16, sn16, xyzw, rgb
+VK_FORMAT_R16G16B16A16_USCALED       , plain, 1, 1, us16, us16, us16, us16, xyzw, rgb
+VK_FORMAT_R16G16B16A16_SSCALED       , plain, 1, 1, ss16, ss16, ss16, ss16, xyzw, rgb
+VK_FORMAT_R16G16B16A16_UINT          , plain, 1, 1, up16, up16, up16, up16, xyzw, rgb
+VK_FORMAT_R16G16B16A16_SINT          , plain, 1, 1, sp16, sp16, sp16, sp16, xyzw, rgb
+VK_FORMAT_R16G16B16A16_SFLOAT        , plain, 1, 1, f16 , f16 , f16 , f16 , xyzw, rgb
+VK_FORMAT_R32_UINT                   , plain, 1, 1, up32,     ,     ,     , x001, rgb
+VK_FORMAT_R32_SINT                   , plain, 1, 1, sp32,     ,     ,     , x001, rgb
+VK_FORMAT_R32_SFLOAT                 , plain, 1, 1, f32 ,     ,     ,     , x001, rgb
+VK_FORMAT_R32G32_UINT                , plain, 1, 1, up32, up32,     ,     , xy01, rgb
+VK_FORMAT_R32G32_SINT                , plain, 1, 1, sp32, sp32,     ,     , xy01, rgb
+VK_FORMAT_R32G32_SFLOAT              , plain, 1, 1, f32 , f32 ,     ,     , xy01, rgb
+VK_FORMAT_R32G32B32_UINT             , plain, 1, 1, up32, up32, up32,     , xyz1, rgb
+VK_FORMAT_R32G32B32_SINT             , plain, 1, 1, sp32, sp32, sp32,     , xyz1, rgb
+VK_FORMAT_R32G32B32_SFLOAT           , plain, 1, 1, f32 , f32 , f32 ,     , xyz1, rgb
+VK_FORMAT_R32G32B32A32_UINT          , plain, 1, 1, up32, up32, up32, up32, xyzw, rgb
+VK_FORMAT_R32G32B32A32_SINT          , plain, 1, 1, sp32, sp32, sp32, sp32, xyzw, rgb
+VK_FORMAT_R32G32B32A32_SFLOAT        , plain, 1, 1, f32 , f32 , f32 , f32 , xyzw, rgb
+VK_FORMAT_R64_UINT                   , plain, 1, 1, up64,     ,     ,     , x001, rgb
+VK_FORMAT_R64_SINT                   , plain, 1, 1, sp64,     ,     ,     , x001, rgb
+VK_FORMAT_R64_SFLOAT                 , plain, 1, 1, f64 ,     ,     ,     , x001, rgb
+VK_FORMAT_R64G64_UINT                , plain, 1, 1, up64, up64,     ,     , xy01, rgb
+VK_FORMAT_R64G64_SINT                , plain, 1, 1, sp64, sp64,     ,     , xy01, rgb
+VK_FORMAT_R64G64_SFLOAT              , plain, 1, 1, f64 , f64 ,     ,     , xy01, rgb
+VK_FORMAT_R64G64B64_UINT             , plain, 1, 1, up64, up64, up64,     , xyz1, rgb
+VK_FORMAT_R64G64B64_SINT             , plain, 1, 1, sp64, sp64, sp64,     , xyz1, rgb
+VK_FORMAT_R64G64B64_SFLOAT           , plain, 1, 1, f64 , f64 , f64 ,     , xyz1, rgb
+VK_FORMAT_R64G64B64A64_UINT          , plain, 1, 1, up64, up64, up64, up64, xyzw, rgb
+VK_FORMAT_R64G64B64A64_SINT          , plain, 1, 1, sp64, sp64, sp64, sp64, xyzw, rgb
+VK_FORMAT_R64G64B64A64_SFLOAT        , plain, 1, 1, f64 , f64 , f64 , f64 , xyzw, rgb
+VK_FORMAT_B10G11R11_UFLOAT_PACK32    , other, 1, 1, x32 ,     ,     ,     , xyz1, rgb
+VK_FORMAT_E5B9G9R9_UFLOAT_PACK32     , other, 1, 1, x32 ,     ,     ,     , xyz1, rgb
+VK_FORMAT_D16_UNORM                  , plain, 1, 1, un16,     ,     ,     , x___, zs
+VK_FORMAT_X8_D24_UNORM_PACK32        , plain, 1, 1, un24, x8  ,     ,     , x___, zs
+VK_FORMAT_D32_SFLOAT                 , plain, 1, 1, f32 ,     ,     ,     , x___, zs
+VK_FORMAT_S8_UINT                    , plain, 1, 1, up8 ,     ,     ,     , _x__, zs
+VK_FORMAT_D16_UNORM_S8_UINT          , plain, 1, 1, un16, up8 ,     ,     , xy__, zs
+VK_FORMAT_D24_UNORM_S8_UINT          , plain, 1, 1, un24, up8 ,     ,     , xy__, zs
+VK_FORMAT_D32_SFLOAT_S8_UINT         , plain, 1, 1, f32 , up8 ,     ,     , xy__, zs
+VK_FORMAT_BC1_RGB_UNORM_BLOCK        , s3tc,  4, 4, x64 ,     ,     ,     , xyz1, rgb
+VK_FORMAT_BC1_RGB_SRGB_BLOCK         , s3tc,  4, 4, x64 ,     ,     ,     , xyz1, srgb
+VK_FORMAT_BC1_RGBA_UNORM_BLOCK       , s3tc,  4, 4, x64 ,     ,     ,     , xyzw, rgb
+VK_FORMAT_BC1_RGBA_SRGB_BLOCK        , s3tc,  4, 4, x64 ,     ,     ,     , xyzw, srgb
+VK_FORMAT_BC2_UNORM_BLOCK            , s3tc,  4, 4, x128,     ,     ,     , xyzw, rgb
+VK_FORMAT_BC2_SRGB_BLOCK             , s3tc,  4, 4, x128,     ,     ,     , xyzw, srgb
+VK_FORMAT_BC3_UNORM_BLOCK            , s3tc,  4, 4, x128,     ,     ,     , xyzw, rgb
+VK_FORMAT_BC3_SRGB_BLOCK             , s3tc,  4, 4, x128,     ,     ,     , xyzw, srgb
+VK_FORMAT_BC4_UNORM_BLOCK            , rgtc,  4, 4, x64,      ,     ,     , x001, rgb
+VK_FORMAT_BC4_SNORM_BLOCK            , rgtc,  4, 4, x64,      ,     ,     , x001, rgb
+VK_FORMAT_BC5_UNORM_BLOCK            , rgtc,  4, 4, x128,     ,     ,     , xy01, rgb
+VK_FORMAT_BC5_SNORM_BLOCK            , rgtc,  4, 4, x128,     ,     ,     , xy01, rgb
+VK_FORMAT_BC6H_UFLOAT_BLOCK          , bptc,  4, 4, x128,     ,     ,     , xyz1, rgb
+VK_FORMAT_BC6H_SFLOAT_BLOCK          , bptc,  4, 4, x128,     ,     ,     , xyz1, rgb
+VK_FORMAT_BC7_UNORM_BLOCK            , bptc,  4, 4, x128,     ,     ,     , xyzw, rgb
+VK_FORMAT_BC7_SRGB_BLOCK             , bptc,  4, 4, x128,     ,     ,     , xyzw, srgb
+VK_FORMAT_ETC2_R8G8B8_UNORM_BLOCK,
+VK_FORMAT_ETC2_R8G8B8_SRGB_BLOCK,
+VK_FORMAT_ETC2_R8G8B8A1_UNORM_BLOCK,
+VK_FORMAT_ETC2_R8G8B8A1_SRGB_BLOCK,
+VK_FORMAT_ETC2_R8G8B8A8_UNORM_BLOCK,
+VK_FORMAT_ETC2_R8G8B8A8_SRGB_BLOCK,
+VK_FORMAT_EAC_R11_UNORM_BLOCK,
+VK_FORMAT_EAC_R11_SNORM_BLOCK,
+VK_FORMAT_EAC_R11G11_UNORM_BLOCK,
+VK_FORMAT_EAC_R11G11_SNORM_BLOCK,
+VK_FORMAT_ASTC_4x4_UNORM_BLOCK,
+VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
+VK_FORMAT_ASTC_5x4_UNORM_BLOCK,
+VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
+VK_FORMAT_ASTC_5x5_UNORM_BLOCK,
+VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
+VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
+VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
+VK_FORMAT_ASTC_6x6_UNORM_BLOCK,
+VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
+VK_FORMAT_ASTC_8x5_UNORM_BLOCK,
+VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
+VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
+VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
+VK_FORMAT_ASTC_8x8_UNORM_BLOCK,
+VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
+VK_FORMAT_ASTC_10x5_UNORM_BLOCK,
+VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
+VK_FORMAT_ASTC_10x6_UNORM_BLOCK,
+VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
+VK_FORMAT_ASTC_10x8_UNORM_BLOCK,
+VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
+VK_FORMAT_ASTC_10x10_UNORM_BLOCK,
+VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
+VK_FORMAT_ASTC_12x10_UNORM_BLOCK,
+VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
+VK_FORMAT_ASTC_12x12_UNORM_BLOCK,
+VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
diff --git a/src/amd/vulkan/vk_format_parse.py b/src/amd/vulkan/vk_format_parse.py
new file mode 100755
index 0000000000..b743fc2bdb
--- /dev/null
+++ b/src/amd/vulkan/vk_format_parse.py
@@ -0,0 +1,384 @@
+#!/usr/bin/env python
+
+'''
+/**************************************************************************
+ *
+ * Copyright 2009 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+'''
+
+
+VOID, UNSIGNED, SIGNED, FIXED, FLOAT = range(5)
+
+SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W, SWIZZLE_0, SWIZZLE_1, SWIZZLE_NONE, = range(7)
+
+PLAIN = 'plain'
+SCALED = 'scaled'
+
+RGB = 'rgb'
+SRGB = 'srgb'
+YUV = 'yuv'
+ZS = 'zs'
+
+
+def is_pot(x):
+   return (x & (x - 1)) == 0
+
+
+VERY_LARGE = 99999999999999999999999
+
+
+class Channel:
+    '''Describe the channel of a color channel.'''
+    
+    def __init__(self, type, norm, pure, scaled, size, name = ''):
+        self.type = type
+        self.norm = norm
+        self.pure = pure
+        self.size = size
+        self.scaled = scaled
+        self.sign = type in (SIGNED, FIXED, FLOAT)
+        self.name = name
+
+    def __str__(self):
+        s = str(self.type)
+        if self.norm:
+            s += 'n'
+        if self.pure:
+            s += 'p'
+        if self.scaled:
+            s += 's'
+        s += str(self.size)
+        return s
+
+    def __eq__(self, other):
+        return self.type == other.type and self.norm == other.norm and self.pure == other.pure and self.size == other.size and self.scaled == other.scaled
+
+    def max(self):
+        '''Maximum representable number.'''
+        if self.type == FLOAT:
+            return VERY_LARGE
+        if self.type == FIXED:
+            return (1 << (self.size/2)) - 1
+        if self.norm:
+            return 1
+        if self.type == UNSIGNED:
+            return (1 << self.size) - 1
+        if self.type == SIGNED:
+            return (1 << (self.size - 1)) - 1
+        assert False
+    
+    def min(self):
+        '''Minimum representable number.'''
+        if self.type == FLOAT:
+            return -VERY_LARGE
+        if self.type == FIXED:
+            return -(1 << (self.size/2))
+        if self.type == UNSIGNED:
+            return 0
+        if self.norm:
+            return -1
+        if self.type == SIGNED:
+            return -(1 << (self.size - 1))
+        assert False
+
+
+class Format:
+    '''Describe a pixel format.'''
+
+    def __init__(self, name, layout, block_width, block_height, le_channels, le_swizzles, be_channels, be_swizzles, colorspace):
+        self.name = name
+        self.layout = layout
+        self.block_width = block_width
+        self.block_height = block_height
+        self.le_channels = le_channels
+        self.le_swizzles = le_swizzles
+        self.be_channels = be_channels
+        self.be_swizzles = be_swizzles
+        self.name = name
+        self.colorspace = colorspace
+
+    def __str__(self):
+        return self.name
+
+    def short_name(self):
+        '''Make up a short norm for a format, suitable to be used as suffix in
+        function names.'''
+
+        name = self.name
+        if name.startswith('VK_FORMAT_'):
+            name = name[len('VK_FORMAT_'):]
+        name = name.lower()
+        return name
+
+    def block_size(self):
+        size = 0
+        for channel in self.le_channels:
+            size += channel.size
+        return size
+
+    def nr_channels(self):
+        nr_channels = 0
+        for channel in self.le_channels:
+            if channel.size:
+                nr_channels += 1
+        return nr_channels
+
+    def array_element(self):
+        if self.layout != PLAIN:
+            return None
+        ref_channel = self.le_channels[0]
+        if ref_channel.type == VOID:
+           ref_channel = self.le_channels[1]
+        for channel in self.le_channels:
+            if channel.size and (channel.size != ref_channel.size or channel.size % 8):
+                return None
+            if channel.type != VOID:
+                if channel.type != ref_channel.type:
+                    return None
+                if channel.norm != ref_channel.norm:
+                    return None
+                if channel.pure != ref_channel.pure:
+                    return None
+                if channel.scaled != ref_channel.scaled:
+                    return None
+        return ref_channel
+
+    def is_array(self):
+        return self.array_element() != None
+
+    def is_mixed(self):
+        if self.layout != PLAIN:
+            return False
+        ref_channel = self.le_channels[0]
+        if ref_channel.type == VOID:
+           ref_channel = self.le_channels[1]
+        for channel in self.le_channels[1:]:
+            if channel.type != VOID:
+                if channel.type != ref_channel.type:
+                    return True
+                if channel.norm != ref_channel.norm:
+                    return True
+                if channel.pure != ref_channel.pure:
+                    return True
+                if channel.scaled != ref_channel.scaled:
+                    return True
+        return False
+
+    def is_pot(self):
+        return is_pot(self.block_size())
+
+    def is_int(self):
+        if self.layout != PLAIN:
+            return False
+        for channel in self.le_channels:
+            if channel.type not in (VOID, UNSIGNED, SIGNED):
+                return False
+        return True
+
+    def is_float(self):
+        if self.layout != PLAIN:
+            return False
+        for channel in self.le_channels:
+            if channel.type not in (VOID, FLOAT):
+                return False
+        return True
+
+    def is_bitmask(self):
+        if self.layout != PLAIN:
+            return False
+        if self.block_size() not in (8, 16, 32):
+            return False
+        for channel in self.le_channels:
+            if channel.type not in (VOID, UNSIGNED, SIGNED):
+                return False
+        return True
+
+    def is_pure_color(self):
+        if self.layout != PLAIN or self.colorspace == ZS:
+            return False
+        pures = [channel.pure
+                 for channel in self.le_channels
+                 if channel.type != VOID]
+        for x in pures:
+           assert x == pures[0]
+        return pures[0]
+
+    def channel_type(self):
+        types = [channel.type
+                 for channel in self.le_channels
+                 if channel.type != VOID]
+        for x in types:
+           assert x == types[0]
+        return types[0]
+
+    def is_pure_signed(self):
+        return self.is_pure_color() and self.channel_type() == SIGNED
+
+    def is_pure_unsigned(self):
+        return self.is_pure_color() and self.channel_type() == UNSIGNED
+
+    def has_channel(self, id):
+        return self.le_swizzles[id] != SWIZZLE_NONE
+
+    def has_depth(self):
+        return self.colorspace == ZS and self.has_channel(0)
+
+    def has_stencil(self):
+        return self.colorspace == ZS and self.has_channel(1)
+
+    def stride(self):
+        return self.block_size()/8
+
+
+_type_parse_map = {
+    '':  VOID,
+    'x': VOID,
+    'u': UNSIGNED,
+    's': SIGNED,
+    'h': FIXED,
+    'f': FLOAT,
+}
+
+_swizzle_parse_map = {
+    'x': SWIZZLE_X,
+    'y': SWIZZLE_Y,
+    'z': SWIZZLE_Z,
+    'w': SWIZZLE_W,
+    '0': SWIZZLE_0,
+    '1': SWIZZLE_1,
+    '_': SWIZZLE_NONE,
+}
+
+def _parse_channels(fields, layout, colorspace, swizzles):
+    if layout == PLAIN:
+        names = ['']*4
+        if colorspace in (RGB, SRGB):
+            for i in range(4):
+                swizzle = swizzles[i]
+                if swizzle < 4:
+                    names[swizzle] += 'rgba'[i]
+        elif colorspace == ZS:
+            for i in range(4):
+                swizzle = swizzles[i]
+                if swizzle < 4:
+                    names[swizzle] += 'zs'[i]
+        else:
+            assert False
+        for i in range(4):
+            if names[i] == '':
+                names[i] = 'x'
+    else:
+        names = ['x', 'y', 'z', 'w']
+
+    channels = []
+    for i in range(0, 4):
+        field = fields[i]
+        if field:
+            type = _type_parse_map[field[0]]
+            if field[1] == 'n':
+                norm = True
+                pure = False
+                scaled = False
+                size = int(field[2:])
+            elif field[1] == 'p':
+                pure = True
+                norm = False
+                scaled = False
+                size = int(field[2:])
+            elif field[1] == 's':
+                pure = False
+                norm = False
+                scaled = True
+                size = int(field[2:])
+            else:
+                norm = False
+                pure = False
+                scaled = False
+                size = int(field[1:])
+        else:
+            type = VOID
+            norm = False
+            pure = False
+            scaled = False
+            size = 0
+        channel = Channel(type, norm, pure, scaled, size, names[i])
+        channels.append(channel)
+
+    return channels
+
+def parse(filename):
+    '''Parse the format description in CSV format in terms of the
+    Channel and Format classes above.'''
+
+    stream = open(filename)
+    formats = []
+    for line in stream:
+        try:
+            comment = line.index('#')
+        except ValueError:
+            pass
+        else:
+            line = line[:comment]
+        line = line.strip()
+        if not line:
+            continue
+
+        fields = [field.strip() for field in line.split(',')]
+        if len (fields) < 10:
+           continue
+        if len (fields) == 10:
+           fields += fields[4:9]
+        assert len (fields) == 15
+        
+        name = fields[0]
+        layout = fields[1]
+        block_width, block_height = map(int, fields[2:4])
+        colorspace = fields[9]
+
+        le_swizzles = [_swizzle_parse_map[swizzle] for swizzle in fields[8]]
+        le_channels = _parse_channels(fields[4:8], layout, colorspace, le_swizzles)
+
+        be_swizzles = [_swizzle_parse_map[swizzle] for swizzle in fields[14]]
+        be_channels = _parse_channels(fields[10:14], layout, colorspace, be_swizzles)
+
+        le_shift = 0
+        for channel in le_channels:
+            channel.shift = le_shift
+            le_shift += channel.size
+
+        be_shift = 0
+        for channel in be_channels[3::-1]:
+            channel.shift = be_shift
+            be_shift += channel.size
+
+        assert le_shift == be_shift
+        for i in range(4):
+            assert (le_swizzles[i] != SWIZZLE_NONE) == (be_swizzles[i] != SWIZZLE_NONE)
+
+        format = Format(name, layout, block_width, block_height, le_channels, le_swizzles, be_channels, be_swizzles, colorspace)
+        formats.append(format)
+    return formats
+
diff --git a/src/amd/vulkan/vk_format_table.py b/src/amd/vulkan/vk_format_table.py
new file mode 100755
index 0000000000..06b98e568b
--- /dev/null
+++ b/src/amd/vulkan/vk_format_table.py
@@ -0,0 +1,173 @@
+#!/usr/bin/env python
+
+CopyRight = '''
+/**************************************************************************
+ *
+ * Copyright 2010 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+'''
+
+
+import sys
+
+from vk_format_parse import *
+
+def layout_map(layout):
+    return 'VK_FORMAT_LAYOUT_' + str(layout).upper()
+
+
+def colorspace_map(colorspace):
+    return 'VK_FORMAT_COLORSPACE_' + str(colorspace).upper()
+
+
+colorspace_channels_map = {
+    'rgb': ['r', 'g', 'b', 'a'],
+    'srgb': ['sr', 'sg', 'sb', 'a'],
+    'zs': ['z', 's'],
+    'yuv': ['y', 'u', 'v'],
+}
+
+
+type_map = {
+    VOID:     "VK_FORMAT_TYPE_VOID",
+    UNSIGNED: "VK_FORMAT_TYPE_UNSIGNED",
+    SIGNED:   "VK_FORMAT_TYPE_SIGNED",
+    FIXED:    "VK_FORMAT_TYPE_FIXED",
+    FLOAT:    "VK_FORMAT_TYPE_FLOAT",
+}
+
+
+def bool_map(value):
+    if value:
+        return "true"
+    else:
+        return "false"
+
+
+swizzle_map = {
+    SWIZZLE_X:    "VK_SWIZZLE_X",
+    SWIZZLE_Y:    "VK_SWIZZLE_Y",
+    SWIZZLE_Z:    "VK_SWIZZLE_Z",
+    SWIZZLE_W:    "VK_SWIZZLE_W",
+    SWIZZLE_0:    "VK_SWIZZLE_0",
+    SWIZZLE_1:    "VK_SWIZZLE_1",
+    SWIZZLE_NONE: "VK_SWIZZLE_NONE",
+}
+
+def print_channels(format, func):
+    if format.nr_channels() <= 1:
+        func(format.le_channels, format.le_swizzles)
+    else:
+        print '#ifdef PIPE_ARCH_BIG_ENDIAN'
+        func(format.be_channels, format.be_swizzles)
+        print '#else'
+        func(format.le_channels, format.le_swizzles)
+        print '#endif'
+
+def write_format_table(formats):
+    print '/* This file is autogenerated by u_format_table.py from u_format.csv. Do not edit directly. */'
+    print
+    # This will print the copyright message on the top of this file
+    print CopyRight.strip()
+    print
+    print '#include "stdbool.h"'
+    print '#include "vk_format.h"'
+    print
+    
+    def do_channel_array(channels, swizzles):
+        print "   {"
+        for i in range(4):
+            channel = channels[i]
+            if i < 3:
+                sep = ","
+            else:
+                sep = ""
+            if channel.size:
+                print "      {%s, %s, %s, %s, %u, %u}%s\t/* %s = %s */" % (type_map[channel.type], bool_map(channel.norm), bool_map(channel.pure), bool_map(channel.scaled), channel.size, channel.shift, sep, "xyzw"[i], channel.name)
+            else:
+                print "      {0, 0, 0, 0, 0}%s" % (sep,)
+        print "   },"
+
+    def do_swizzle_array(channels, swizzles):
+        print "   {"
+        for i in range(4):
+            swizzle = swizzles[i]
+            if i < 3:
+                sep = ","
+            else:
+                sep = ""
+            try:
+                comment = colorspace_channels_map[format.colorspace][i]
+            except (KeyError, IndexError):
+                comment = 'ignored'
+            print "      %s%s\t/* %s */" % (swizzle_map[swizzle], sep, comment)
+        print "   },"
+
+    for format in formats:
+        print 'const struct vk_format_description'
+        print 'vk_format_%s_description = {' % (format.short_name(),)
+        print "   %s," % (format.name,)
+        print "   \"%s\"," % (format.name,)
+        print "   \"%s\"," % (format.short_name(),)
+        print "   {%u, %u, %u},\t/* block */" % (format.block_width, format.block_height, format.block_size())
+        print "   %s," % (layout_map(format.layout),)
+        print "   %u,\t/* nr_channels */" % (format.nr_channels(),)
+        print "   %s,\t/* is_array */" % (bool_map(format.is_array()),)
+        print "   %s,\t/* is_bitmask */" % (bool_map(format.is_bitmask()),)
+        print "   %s,\t/* is_mixed */" % (bool_map(format.is_mixed()),)
+        print_channels(format, do_channel_array)
+        print_channels(format, do_swizzle_array)
+        print "   %s," % (colorspace_map(format.colorspace),)
+        print "};"
+        print
+        
+    print "const struct vk_format_description *"
+    print "vk_format_description(VkFormat format)"
+    print "{"
+    print "   if (format > VK_FORMAT_END_RANGE) {"
+    print "      return NULL;"
+    print "   }"
+    print
+    print "   switch (format) {"
+    for format in formats:
+        print "   case %s:" % format.name
+        print "      return &vk_format_%s_description;" % (format.short_name(),)
+    print "   default:"
+    print "      return NULL;"
+    print "   }"
+    print "}"
+    print
+
+
+def main():
+
+    formats = []
+    for arg in sys.argv[1:]:
+        formats.extend(parse(arg))
+    write_format_table(formats)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
new file mode 100644
index 0000000000..7319a98887
--- /dev/null
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
@@ -0,0 +1,297 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based on amdgpu winsys.
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <stdio.h>
+
+#include "radv_amdgpu_bo.h"
+
+#include <amdgpu.h>
+#include <amdgpu_drm.h>
+#include <inttypes.h>
+
+static void radv_amdgpu_winsys_bo_destroy(struct radeon_winsys_bo *_bo)
+{
+	struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
+
+	if (bo->ws->debug_all_bos) {
+		pthread_mutex_lock(&bo->ws->global_bo_list_lock);
+		LIST_DEL(&bo->global_list_item);
+		bo->ws->num_buffers--;
+		pthread_mutex_unlock(&bo->ws->global_bo_list_lock);
+	}
+	amdgpu_bo_va_op(bo->bo, 0, bo->size, bo->va, 0, AMDGPU_VA_OP_UNMAP);
+	amdgpu_va_range_free(bo->va_handle);
+	amdgpu_bo_free(bo->bo);
+	FREE(bo);
+}
+
+static void radv_amdgpu_add_buffer_to_global_list(struct radv_amdgpu_winsys_bo *bo)
+{
+	struct radv_amdgpu_winsys *ws = bo->ws;
+
+	if (bo->ws->debug_all_bos) {
+		pthread_mutex_lock(&ws->global_bo_list_lock);
+		LIST_ADDTAIL(&bo->global_list_item, &ws->global_bo_list);
+		ws->num_buffers++;
+		pthread_mutex_unlock(&ws->global_bo_list_lock);
+	}
+}
+
+static struct radeon_winsys_bo *
+radv_amdgpu_winsys_bo_create(struct radeon_winsys *_ws,
+			     uint64_t size,
+			     unsigned alignment,
+			     enum radeon_bo_domain initial_domain,
+			     unsigned flags)
+{
+	struct radv_amdgpu_winsys *ws = radv_amdgpu_winsys(_ws);
+	struct radv_amdgpu_winsys_bo *bo;
+	struct amdgpu_bo_alloc_request request = {0};
+	amdgpu_bo_handle buf_handle;
+	uint64_t va = 0;
+	amdgpu_va_handle va_handle;
+	int r;
+	bo = CALLOC_STRUCT(radv_amdgpu_winsys_bo);
+	if (!bo) {
+		return NULL;
+	}
+
+	request.alloc_size = size;
+	request.phys_alignment = alignment;
+
+	if (initial_domain & RADEON_DOMAIN_VRAM)
+		request.preferred_heap |= AMDGPU_GEM_DOMAIN_VRAM;
+	if (initial_domain & RADEON_DOMAIN_GTT)
+		request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT;
+
+	if (flags & RADEON_FLAG_CPU_ACCESS)
+		request.flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
+	if (flags & RADEON_FLAG_NO_CPU_ACCESS)
+		request.flags |= AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
+	if (flags & RADEON_FLAG_GTT_WC)
+		request.flags |= AMDGPU_GEM_CREATE_CPU_GTT_USWC;
+
+	r = amdgpu_bo_alloc(ws->dev, &request, &buf_handle);
+	if (r) {
+		fprintf(stderr, "amdgpu: Failed to allocate a buffer:\n");
+		fprintf(stderr, "amdgpu:    size      : %"PRIu64" bytes\n", size);
+		fprintf(stderr, "amdgpu:    alignment : %u bytes\n", alignment);
+		fprintf(stderr, "amdgpu:    domains   : %u\n", initial_domain);
+		goto error_bo_alloc;
+	}
+
+	r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
+				  size, alignment, 0, &va, &va_handle, 0);
+	if (r)
+		goto error_va_alloc;
+
+	r = amdgpu_bo_va_op(buf_handle, 0, size, va, 0, AMDGPU_VA_OP_MAP);
+	if (r)
+		goto error_va_map;
+
+	bo->bo = buf_handle;
+	bo->va = va;
+	bo->va_handle = va_handle;
+	bo->initial_domain = initial_domain;
+	bo->size = size;
+	bo->is_shared = false;
+	bo->ws = ws;
+	radv_amdgpu_add_buffer_to_global_list(bo);
+	return (struct radeon_winsys_bo *)bo;
+error_va_map:
+	amdgpu_va_range_free(va_handle);
+
+error_va_alloc:
+	amdgpu_bo_free(buf_handle);
+
+error_bo_alloc:
+	FREE(bo);
+	return NULL;
+}
+
+static uint64_t radv_amdgpu_winsys_bo_get_va(struct radeon_winsys_bo *_bo)
+{
+	struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
+	return bo->va;
+}
+
+static void *
+radv_amdgpu_winsys_bo_map(struct radeon_winsys_bo *_bo)
+{
+	struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
+	int ret;
+	void *data;
+	ret = amdgpu_bo_cpu_map(bo->bo, &data);
+	if (ret)
+		return NULL;
+	return data;
+}
+
+static void
+radv_amdgpu_winsys_bo_unmap(struct radeon_winsys_bo *_bo)
+{
+	struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
+	amdgpu_bo_cpu_unmap(bo->bo);
+}
+
+static struct radeon_winsys_bo *
+radv_amdgpu_winsys_bo_from_fd(struct radeon_winsys *_ws,
+			      int fd, unsigned *stride,
+			      unsigned *offset)
+{
+	struct radv_amdgpu_winsys *ws = radv_amdgpu_winsys(_ws);
+	struct radv_amdgpu_winsys_bo *bo;
+	uint64_t va;
+	amdgpu_va_handle va_handle;
+	enum amdgpu_bo_handle_type type = amdgpu_bo_handle_type_dma_buf_fd;
+	struct amdgpu_bo_import_result result = {0};
+	struct amdgpu_bo_info info = {0};
+	enum radeon_bo_domain initial = 0;
+	int r;
+	bo = CALLOC_STRUCT(radv_amdgpu_winsys_bo);
+	if (!bo)
+		return NULL;
+
+	r = amdgpu_bo_import(ws->dev, type, fd, &result);
+	if (r)
+		goto error;
+
+	r = amdgpu_bo_query_info(result.buf_handle, &info);
+	if (r)
+		goto error_query;
+
+	r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
+				  result.alloc_size, 1 << 20, 0, &va, &va_handle, 0);
+	if (r)
+		goto error_query;
+
+	r = amdgpu_bo_va_op(result.buf_handle, 0, result.alloc_size, va, 0, AMDGPU_VA_OP_MAP);
+	if (r)
+		goto error_va_map;
+
+	if (info.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM)
+		initial |= RADEON_DOMAIN_VRAM;
+	if (info.preferred_heap & AMDGPU_GEM_DOMAIN_GTT)
+		initial |= RADEON_DOMAIN_GTT;
+
+	bo->bo = result.buf_handle;
+	bo->va = va;
+	bo->va_handle = va_handle;
+	bo->initial_domain = initial;
+	bo->size = result.alloc_size;
+	bo->is_shared = true;
+	return (struct radeon_winsys_bo *)bo;
+error_va_map:
+	amdgpu_va_range_free(va_handle);
+
+error_query:
+	amdgpu_bo_free(result.buf_handle);
+
+error:
+	FREE(bo);
+	return NULL;
+}
+
+static bool
+radv_amdgpu_winsys_get_fd(struct radeon_winsys *_ws,
+			  struct radeon_winsys_bo *_bo,
+			  int *fd)
+{
+	struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
+	enum amdgpu_bo_handle_type type = amdgpu_bo_handle_type_dma_buf_fd;
+	int r;
+	unsigned handle;
+	r = amdgpu_bo_export(bo->bo, type, &handle);
+	if (r)
+		return false;
+
+	*fd = (int)handle;
+	bo->is_shared = true;
+	return true;
+}
+
+static unsigned radv_eg_tile_split_rev(unsigned eg_tile_split)
+{
+	switch (eg_tile_split) {
+	case 64:    return 0;
+	case 128:   return 1;
+	case 256:   return 2;
+	case 512:   return 3;
+	default:
+	case 1024:  return 4;
+	case 2048:  return 5;
+	case 4096:  return 6;
+	}
+}
+
+static void
+radv_amdgpu_winsys_bo_set_metadata(struct radeon_winsys_bo *_bo,
+				   struct radeon_bo_metadata *md)
+{
+	struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
+	struct amdgpu_bo_metadata metadata = {0};
+	uint32_t tiling_flags = 0;
+
+	if (md->macrotile == RADEON_LAYOUT_TILED)
+		tiling_flags |= AMDGPU_TILING_SET(ARRAY_MODE, 4); /* 2D_TILED_THIN1 */
+	else if (md->microtile == RADEON_LAYOUT_TILED)
+		tiling_flags |= AMDGPU_TILING_SET(ARRAY_MODE, 2); /* 1D_TILED_THIN1 */
+	else
+		tiling_flags |= AMDGPU_TILING_SET(ARRAY_MODE, 1); /* LINEAR_ALIGNED */
+
+	tiling_flags |= AMDGPU_TILING_SET(PIPE_CONFIG, md->pipe_config);
+	tiling_flags |= AMDGPU_TILING_SET(BANK_WIDTH, util_logbase2(md->bankw));
+	tiling_flags |= AMDGPU_TILING_SET(BANK_HEIGHT, util_logbase2(md->bankh));
+	if (md->tile_split)
+		tiling_flags |= AMDGPU_TILING_SET(TILE_SPLIT, radv_eg_tile_split_rev(md->tile_split));
+	tiling_flags |= AMDGPU_TILING_SET(MACRO_TILE_ASPECT, util_logbase2(md->mtilea));
+	tiling_flags |= AMDGPU_TILING_SET(NUM_BANKS, util_logbase2(md->num_banks)-1);
+
+	if (md->scanout)
+		tiling_flags |= AMDGPU_TILING_SET(MICRO_TILE_MODE, 0); /* DISPLAY_MICRO_TILING */
+	else
+		tiling_flags |= AMDGPU_TILING_SET(MICRO_TILE_MODE, 1); /* THIN_MICRO_TILING */
+
+	metadata.tiling_info = tiling_flags;
+	metadata.size_metadata = md->size_metadata;
+	memcpy(metadata.umd_metadata, md->metadata, sizeof(md->metadata));
+
+	amdgpu_bo_set_metadata(bo->bo, &metadata);
+}
+
+void radv_amdgpu_bo_init_functions(struct radv_amdgpu_winsys *ws)
+{
+	ws->base.buffer_create = radv_amdgpu_winsys_bo_create;
+	ws->base.buffer_destroy = radv_amdgpu_winsys_bo_destroy;
+	ws->base.buffer_get_va = radv_amdgpu_winsys_bo_get_va;
+	ws->base.buffer_map = radv_amdgpu_winsys_bo_map;
+	ws->base.buffer_unmap = radv_amdgpu_winsys_bo_unmap;
+	ws->base.buffer_from_fd = radv_amdgpu_winsys_bo_from_fd;
+	ws->base.buffer_get_fd = radv_amdgpu_winsys_get_fd;
+	ws->base.buffer_set_metadata = radv_amdgpu_winsys_bo_set_metadata;
+}
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h
new file mode 100644
index 0000000000..59a1bb7650
--- /dev/null
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h
@@ -0,0 +1,50 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based on amdgpu winsys.
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#pragma once
+#include "radv_amdgpu_winsys.h"
+struct radv_amdgpu_winsys_bo {
+	amdgpu_bo_handle bo;
+	amdgpu_va_handle va_handle;
+
+	uint64_t va;
+	enum radeon_bo_domain initial_domain;
+	uint64_t size;
+	bool is_shared;
+
+	struct radv_amdgpu_winsys *ws;
+	struct list_head global_list_item;
+};
+
+static inline
+struct radv_amdgpu_winsys_bo *radv_amdgpu_winsys_bo(struct radeon_winsys_bo *bo)
+{
+	return (struct radv_amdgpu_winsys_bo *)bo;
+}
+
+void radv_amdgpu_bo_init_functions(struct radv_amdgpu_winsys *ws);
+
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
new file mode 100644
index 0000000000..dedc778f1c
--- /dev/null
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
@@ -0,0 +1,778 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <stdlib.h>
+#include <amdgpu.h>
+#include <amdgpu_drm.h>
+#include <assert.h>
+
+#include "amdgpu_id.h"
+#include "radv_radeon_winsys.h"
+#include "radv_amdgpu_cs.h"
+#include "radv_amdgpu_bo.h"
+#include "sid.h"
+
+struct radv_amdgpu_cs {
+	struct radeon_winsys_cs base;
+	struct radv_amdgpu_winsys *ws;
+
+	struct amdgpu_cs_ib_info    ib;
+
+	struct radeon_winsys_bo     *ib_buffer;
+	uint8_t                 *ib_mapped;
+	unsigned                    max_num_buffers;
+	unsigned                    num_buffers;
+	amdgpu_bo_handle            *handles;
+	uint8_t                     *priorities;
+
+	struct radeon_winsys_bo     **old_ib_buffers;
+	unsigned                    num_old_ib_buffers;
+	unsigned                    max_num_old_ib_buffers;
+	unsigned                    *ib_size_ptr;
+	bool                        failed;
+	bool                        is_chained;
+
+	int                         buffer_hash_table[1024];
+};
+
+static inline struct radv_amdgpu_cs *
+radv_amdgpu_cs(struct radeon_winsys_cs *base)
+{
+	return (struct radv_amdgpu_cs*)base;
+}
+
+
+static struct radeon_winsys_fence *radv_amdgpu_create_fence()
+{
+	struct radv_amdgpu_cs_fence *fence = calloc(1, sizeof(struct amdgpu_cs_fence));
+	return (struct radeon_winsys_fence*)fence;
+}
+
+static void radv_amdgpu_destroy_fence(struct radeon_winsys_fence *_fence)
+{
+	struct amdgpu_cs_fence *fence = (struct amdgpu_cs_fence *)_fence;
+	free(fence);
+}
+
+static bool radv_amdgpu_fence_wait(struct radeon_winsys *_ws,
+			      struct radeon_winsys_fence *_fence,
+			      bool absolute,
+			      uint64_t timeout)
+{
+	struct amdgpu_cs_fence *fence = (struct amdgpu_cs_fence *)_fence;
+	unsigned flags = absolute ? AMDGPU_QUERY_FENCE_TIMEOUT_IS_ABSOLUTE : 0;
+	int r;
+	uint32_t expired = 0;
+	/* Now use the libdrm query. */
+	r = amdgpu_cs_query_fence_status(fence,
+					 timeout,
+					 flags,
+					 &expired);
+
+	if (r) {
+		fprintf(stderr, "amdgpu: radv_amdgpu_cs_query_fence_status failed.\n");
+		return false;
+	}
+
+	if (expired) {
+		return true;
+	}
+	return false;
+
+}
+
+static void radv_amdgpu_cs_destroy(struct radeon_winsys_cs *rcs)
+{
+	struct radv_amdgpu_cs *cs = radv_amdgpu_cs(rcs);
+	if (cs->ib_buffer)
+		cs->ws->base.buffer_destroy(cs->ib_buffer);
+	else
+		free(cs->base.buf);
+
+	for (unsigned i = 0; i < cs->num_old_ib_buffers; ++i)
+		cs->ws->base.buffer_destroy(cs->old_ib_buffers[i]);
+	free(cs->old_ib_buffers);
+	free(cs->handles);
+	free(cs->priorities);
+	free(cs);
+}
+
+static boolean radv_amdgpu_init_cs(struct radv_amdgpu_cs *cs,
+				   enum ring_type ring_type)
+{
+	for (int i = 0; i < ARRAY_SIZE(cs->buffer_hash_table); ++i) {
+		cs->buffer_hash_table[i] = -1;
+	}
+	return true;
+}
+
+static struct radeon_winsys_cs *
+radv_amdgpu_cs_create(struct radeon_winsys *ws,
+		      enum ring_type ring_type)
+{
+	struct radv_amdgpu_cs *cs;
+	uint32_t ib_size = 20 * 1024 * 4;
+	cs = calloc(1, sizeof(struct radv_amdgpu_cs));
+	if (!cs)
+		return NULL;
+
+	cs->ws = radv_amdgpu_winsys(ws);
+	radv_amdgpu_init_cs(cs, RING_GFX);
+
+	if (cs->ws->use_ib_bos) {
+		cs->ib_buffer = ws->buffer_create(ws, ib_size, 0,
+						RADEON_DOMAIN_GTT,
+						RADEON_FLAG_CPU_ACCESS);
+		if (!cs->ib_buffer) {
+			free(cs);
+			return NULL;
+		}
+
+		cs->ib_mapped = ws->buffer_map(cs->ib_buffer);
+		if (!cs->ib_mapped) {
+			ws->buffer_destroy(cs->ib_buffer);
+			free(cs);
+			return NULL;
+		}
+
+		cs->ib.ib_mc_address = radv_amdgpu_winsys_bo(cs->ib_buffer)->va;
+		cs->base.buf = (uint32_t *)cs->ib_mapped;
+		cs->base.max_dw = ib_size / 4 - 4;
+		cs->ib_size_ptr = &cs->ib.size;
+		cs->ib.size = 0;
+
+		ws->cs_add_buffer(&cs->base, cs->ib_buffer, 8);
+	} else {
+		cs->base.buf = malloc(16384);
+		cs->base.max_dw = 4096;
+		if (!cs->base.buf) {
+			free(cs);
+			return NULL;
+		}
+	}
+
+	return &cs->base;
+}
+
+static void radv_amdgpu_cs_grow(struct radeon_winsys_cs *_cs, size_t min_size)
+{
+	struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
+	uint64_t ib_size = MAX2(min_size * 4 + 16, cs->base.max_dw * 4 * 2);
+
+	/* max that fits in the chain size field. */
+	ib_size = MIN2(ib_size, 0xfffff);
+
+	if (cs->failed) {
+		cs->base.cdw = 0;
+		return;
+	}
+
+	if (!cs->ws->use_ib_bos) {
+		uint32_t *new_buf = realloc(cs->base.buf, ib_size);
+		if (new_buf) {
+			cs->base.buf = new_buf;
+			cs->base.max_dw = ib_size / 4;
+		} else {
+			cs->failed = true;
+			cs->base.cdw = 0;
+		}
+		return;
+	}
+
+	while (!cs->base.cdw || (cs->base.cdw & 7) != 4)
+		cs->base.buf[cs->base.cdw++] = 0xffff1000;
+
+	*cs->ib_size_ptr |= cs->base.cdw + 4;
+
+	if (cs->num_old_ib_buffers == cs->max_num_old_ib_buffers) {
+		cs->max_num_old_ib_buffers = MAX2(1, cs->max_num_old_ib_buffers * 2);
+		cs->old_ib_buffers = realloc(cs->old_ib_buffers,
+					     cs->max_num_old_ib_buffers * sizeof(void*));
+	}
+
+	cs->old_ib_buffers[cs->num_old_ib_buffers++] = cs->ib_buffer;
+
+	cs->ib_buffer = cs->ws->base.buffer_create(&cs->ws->base, ib_size, 0,
+						   RADEON_DOMAIN_GTT,
+						   RADEON_FLAG_CPU_ACCESS);
+
+	if (!cs->ib_buffer) {
+		cs->base.cdw = 0;
+		cs->failed = true;
+		cs->ib_buffer = cs->old_ib_buffers[--cs->num_old_ib_buffers];
+	}
+
+	cs->ib_mapped = cs->ws->base.buffer_map(cs->ib_buffer);
+	if (!cs->ib_mapped) {
+		cs->ws->base.buffer_destroy(cs->ib_buffer);
+		cs->base.cdw = 0;
+		cs->failed = true;
+		cs->ib_buffer = cs->old_ib_buffers[--cs->num_old_ib_buffers];
+	}
+
+	cs->ws->base.cs_add_buffer(&cs->base, cs->ib_buffer, 8);
+
+	cs->base.buf[cs->base.cdw++] = PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0);
+	cs->base.buf[cs->base.cdw++] = radv_amdgpu_winsys_bo(cs->ib_buffer)->va;
+	cs->base.buf[cs->base.cdw++] = radv_amdgpu_winsys_bo(cs->ib_buffer)->va >> 32;
+	cs->ib_size_ptr = cs->base.buf + cs->base.cdw;
+	cs->base.buf[cs->base.cdw++] = S_3F2_CHAIN(1) | S_3F2_VALID(1);
+
+	cs->base.buf = (uint32_t *)cs->ib_mapped;
+	cs->base.cdw = 0;
+	cs->base.max_dw = ib_size / 4 - 4;
+
+}
+
+static bool radv_amdgpu_cs_finalize(struct radeon_winsys_cs *_cs)
+{
+	struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
+
+	if (cs->ws->use_ib_bos) {
+		while (!cs->base.cdw || (cs->base.cdw & 7) != 0)
+			cs->base.buf[cs->base.cdw++] = 0xffff1000;
+
+		*cs->ib_size_ptr |= cs->base.cdw;
+
+		cs->is_chained = false;
+	}
+
+	return !cs->failed;
+}
+
+static void radv_amdgpu_cs_reset(struct radeon_winsys_cs *_cs)
+{
+	struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
+	cs->base.cdw = 0;
+	cs->failed = false;
+
+	for (unsigned i = 0; i < cs->num_buffers; ++i) {
+		unsigned hash = ((uintptr_t)cs->handles[i] >> 6) &
+		                 (ARRAY_SIZE(cs->buffer_hash_table) - 1);
+		cs->buffer_hash_table[hash] = -1;
+	}
+
+	cs->num_buffers = 0;
+
+	if (cs->ws->use_ib_bos) {
+		cs->ws->base.cs_add_buffer(&cs->base, cs->ib_buffer, 8);
+
+		for (unsigned i = 0; i < cs->num_old_ib_buffers; ++i)
+			cs->ws->base.buffer_destroy(cs->old_ib_buffers[i]);
+
+		cs->num_old_ib_buffers = 0;
+		cs->ib.ib_mc_address = radv_amdgpu_winsys_bo(cs->ib_buffer)->va;
+		cs->ib_size_ptr = &cs->ib.size;
+		cs->ib.size = 0;
+	}
+}
+
+static int radv_amdgpu_cs_find_buffer(struct radv_amdgpu_cs *cs,
+				      amdgpu_bo_handle bo)
+{
+	unsigned hash = ((uintptr_t)bo >> 6) & (ARRAY_SIZE(cs->buffer_hash_table) - 1);
+	int index = cs->buffer_hash_table[hash];
+
+	if (index == -1)
+		return -1;
+
+	if(cs->handles[index] == bo)
+		return index;
+
+	for (unsigned i = 0; i < cs->num_buffers; ++i) {
+		if (cs->handles[i] == bo) {
+			cs->buffer_hash_table[hash] = i;
+			return i;
+		}
+	}
+	return -1;
+}
+
+static void radv_amdgpu_cs_add_buffer_internal(struct radv_amdgpu_cs *cs,
+					       amdgpu_bo_handle bo,
+					       uint8_t priority)
+{
+	unsigned hash;
+	int index = radv_amdgpu_cs_find_buffer(cs, bo);
+
+	if (index != -1) {
+		cs->priorities[index] = MAX2(cs->priorities[index], priority);
+		return;
+	}
+
+	if (cs->num_buffers == cs->max_num_buffers) {
+		unsigned new_count = MAX2(1, cs->max_num_buffers * 2);
+		cs->handles = realloc(cs->handles, new_count * sizeof(amdgpu_bo_handle));
+		cs->priorities = realloc(cs->priorities, new_count * sizeof(uint8_t));
+		cs->max_num_buffers = new_count;
+	}
+
+	cs->handles[cs->num_buffers] = bo;
+	cs->priorities[cs->num_buffers] = priority;
+
+	hash = ((uintptr_t)bo >> 6) & (ARRAY_SIZE(cs->buffer_hash_table) - 1);
+	cs->buffer_hash_table[hash] = cs->num_buffers;
+
+	++cs->num_buffers;
+}
+
+static void radv_amdgpu_cs_add_buffer(struct radeon_winsys_cs *_cs,
+				 struct radeon_winsys_bo *_bo,
+				 uint8_t priority)
+{
+	struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
+	struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
+
+	radv_amdgpu_cs_add_buffer_internal(cs, bo->bo, priority);
+}
+
+static void radv_amdgpu_cs_execute_secondary(struct radeon_winsys_cs *_parent,
+					     struct radeon_winsys_cs *_child)
+{
+	struct radv_amdgpu_cs *parent = radv_amdgpu_cs(_parent);
+	struct radv_amdgpu_cs *child = radv_amdgpu_cs(_child);
+
+	for (unsigned i = 0; i < child->num_buffers; ++i) {
+		radv_amdgpu_cs_add_buffer_internal(parent, child->handles[i],
+						   child->priorities[i]);
+	}
+
+	if (parent->ws->use_ib_bos) {
+		if (parent->base.cdw + 4 > parent->base.max_dw)
+			radv_amdgpu_cs_grow(&parent->base, 4);
+
+		parent->base.buf[parent->base.cdw++] = PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0);
+		parent->base.buf[parent->base.cdw++] = child->ib.ib_mc_address;
+		parent->base.buf[parent->base.cdw++] = child->ib.ib_mc_address >> 32;
+		parent->base.buf[parent->base.cdw++] = child->ib.size;
+	} else {
+		if (parent->base.cdw + child->base.cdw > parent->base.max_dw)
+			radv_amdgpu_cs_grow(&parent->base, child->base.cdw);
+
+		memcpy(parent->base.buf + parent->base.cdw, child->base.buf, 4 * child->base.cdw);
+		parent->base.cdw += child->base.cdw;
+	}
+}
+
+static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
+				      struct radeon_winsys_cs **cs_array,
+				      unsigned count,
+				      struct radv_amdgpu_winsys_bo *extra_bo,
+				      amdgpu_bo_list_handle *bo_list)
+{
+	int r;
+	if (ws->debug_all_bos) {
+		struct radv_amdgpu_winsys_bo *bo;
+		amdgpu_bo_handle *handles;
+		unsigned num = 0;
+
+		pthread_mutex_lock(&ws->global_bo_list_lock);
+
+		handles = malloc(sizeof(handles[0]) * ws->num_buffers);
+		if (!handles) {
+			pthread_mutex_unlock(&ws->global_bo_list_lock);
+			return -ENOMEM;
+		}
+
+		LIST_FOR_EACH_ENTRY(bo, &ws->global_bo_list, global_list_item) {
+			assert(num < ws->num_buffers);
+			handles[num++] = bo->bo;
+		}
+
+		r = amdgpu_bo_list_create(ws->dev, ws->num_buffers,
+					  handles, NULL,
+					  bo_list);
+		free(handles);
+		pthread_mutex_unlock(&ws->global_bo_list_lock);
+	} else if (count == 1 && !extra_bo) {
+		struct radv_amdgpu_cs *cs = (struct radv_amdgpu_cs*)cs_array[0];
+		r = amdgpu_bo_list_create(ws->dev, cs->num_buffers, cs->handles,
+					  cs->priorities, bo_list);
+	} else {
+		unsigned total_buffer_count = !!extra_bo;
+		unsigned unique_bo_count = !!extra_bo;
+		for (unsigned i = 0; i < count; ++i) {
+			struct radv_amdgpu_cs *cs = (struct radv_amdgpu_cs*)cs_array[i];
+			total_buffer_count += cs->num_buffers;
+		}
+
+		amdgpu_bo_handle *handles = malloc(sizeof(amdgpu_bo_handle) * total_buffer_count);
+		uint8_t *priorities = malloc(sizeof(uint8_t) * total_buffer_count);
+		if (!handles || !priorities) {
+			free(handles);
+			free(priorities);
+			return -ENOMEM;
+		}
+
+		if (extra_bo) {
+			handles[0] = extra_bo->bo;
+			priorities[0] = 8;
+		}
+
+		for (unsigned i = 0; i < count; ++i) {
+			struct radv_amdgpu_cs *cs = (struct radv_amdgpu_cs*)cs_array[i];
+			for (unsigned j = 0; j < cs->num_buffers; ++j) {
+				bool found = false;
+				for (unsigned k = 0; k < unique_bo_count; ++k) {
+					if (handles[k] == cs->handles[j]) {
+						found = true;
+						priorities[k] = MAX2(priorities[k],
+								     cs->priorities[j]);
+						break;
+					}
+				}
+				if (!found) {
+					handles[unique_bo_count] = cs->handles[j];
+					priorities[unique_bo_count] = cs->priorities[j];
+					++unique_bo_count;
+				}
+			}
+		}
+		r = amdgpu_bo_list_create(ws->dev, unique_bo_count, handles,
+					  priorities, bo_list);
+
+		free(handles);
+		free(priorities);
+	}
+	return r;
+}
+
+static int radv_amdgpu_winsys_cs_submit_chained(struct radeon_winsys_ctx *_ctx,
+						struct radeon_winsys_cs **cs_array,
+						unsigned cs_count,
+						struct radeon_winsys_fence *_fence)
+{
+	int r;
+	struct radv_amdgpu_ctx *ctx = radv_amdgpu_ctx(_ctx);
+	struct amdgpu_cs_fence *fence = (struct amdgpu_cs_fence *)_fence;
+	struct radv_amdgpu_cs *cs0 = radv_amdgpu_cs(cs_array[0]);
+	amdgpu_bo_list_handle bo_list;
+	struct amdgpu_cs_request request = {0};
+
+	for (unsigned i = cs_count; i--;) {
+		struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[i]);
+
+		if (cs->is_chained) {
+			*cs->ib_size_ptr -= 4;
+			cs->is_chained = false;
+		}
+
+		if (i + 1 < cs_count) {
+			struct radv_amdgpu_cs *next = radv_amdgpu_cs(cs_array[i + 1]);
+			assert(cs->base.cdw + 4 <= cs->base.max_dw);
+
+			cs->is_chained = true;
+			*cs->ib_size_ptr += 4;
+
+			cs->base.buf[cs->base.cdw + 0] = PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0);
+			cs->base.buf[cs->base.cdw + 1] = next->ib.ib_mc_address;
+			cs->base.buf[cs->base.cdw + 2] = next->ib.ib_mc_address >> 32;
+			cs->base.buf[cs->base.cdw + 3] = S_3F2_CHAIN(1) | S_3F2_VALID(1) | next->ib.size;
+		}
+	}
+
+	r = radv_amdgpu_create_bo_list(cs0->ws, cs_array, cs_count, NULL, &bo_list);
+	if (r) {
+		fprintf(stderr, "amdgpu: Failed to created the BO list for submission\n");
+		return r;
+	}
+
+	request.ip_type = AMDGPU_HW_IP_GFX;
+	request.number_of_ibs = 1;
+	request.ibs = &cs0->ib;
+	request.resources = bo_list;
+
+	r = amdgpu_cs_submit(ctx->ctx, 0, &request, 1);
+	if (r) {
+		if (r == -ENOMEM)
+			fprintf(stderr, "amdgpu: Not enough memory for command submission.\n");
+		else
+			fprintf(stderr, "amdgpu: The CS has been rejected, "
+					"see dmesg for more information.\n");
+	}
+
+	amdgpu_bo_list_destroy(bo_list);
+
+	if (fence) {
+		fence->context = ctx->ctx;
+		fence->ip_type = request.ip_type;
+		fence->ip_instance = request.ip_instance;
+		fence->ring = request.ring;
+		fence->fence = request.seq_no;
+	}
+	ctx->last_seq_no = request.seq_no;
+
+	return r;
+}
+
+static int radv_amdgpu_winsys_cs_submit_fallback(struct radeon_winsys_ctx *_ctx,
+						 struct radeon_winsys_cs **cs_array,
+						 unsigned cs_count,
+						 struct radeon_winsys_fence *_fence)
+{
+	int r;
+	struct radv_amdgpu_ctx *ctx = radv_amdgpu_ctx(_ctx);
+	struct amdgpu_cs_fence *fence = (struct amdgpu_cs_fence *)_fence;
+	amdgpu_bo_list_handle bo_list;
+	struct amdgpu_cs_request request;
+
+	assert(cs_count);
+
+	for (unsigned i = 0; i < cs_count;) {
+		struct radv_amdgpu_cs *cs0 = radv_amdgpu_cs(cs_array[i]);
+		struct amdgpu_cs_ib_info ibs[AMDGPU_CS_MAX_IBS_PER_SUBMIT];
+		unsigned cnt = MIN2(AMDGPU_CS_MAX_IBS_PER_SUBMIT, cs_count - i);
+
+		memset(&request, 0, sizeof(request));
+
+		r = radv_amdgpu_create_bo_list(cs0->ws, &cs_array[i], cnt, NULL, &bo_list);
+		if (r) {
+			fprintf(stderr, "amdgpu: Failed to created the BO list for submission\n");
+			return r;
+		}
+
+		request.ip_type = AMDGPU_HW_IP_GFX;
+		request.resources = bo_list;
+		request.number_of_ibs = cnt;
+		request.ibs = ibs;
+
+		for (unsigned j = 0; j < cnt; ++j) {
+			struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[i + j]);
+			ibs[j] = cs->ib;
+
+			if (cs->is_chained) {
+				*cs->ib_size_ptr -= 4;
+				cs->is_chained = false;
+			}
+		}
+
+		r = amdgpu_cs_submit(ctx->ctx, 0, &request, 1);
+		if (r) {
+			if (r == -ENOMEM)
+				fprintf(stderr, "amdgpu: Not enough memory for command submission.\n");
+			else
+				fprintf(stderr, "amdgpu: The CS has been rejected, "
+						"see dmesg for more information.\n");
+		}
+
+		amdgpu_bo_list_destroy(bo_list);
+
+		if (r)
+			return r;
+
+		i += cnt;
+	}
+	if (fence) {
+		fence->context = ctx->ctx;
+		fence->ip_type = request.ip_type;
+		fence->ip_instance = request.ip_instance;
+		fence->ring = request.ring;
+		fence->fence = request.seq_no;
+	}
+	ctx->last_seq_no = request.seq_no;
+
+	return 0;
+}
+
+static int radv_amdgpu_winsys_cs_submit_sysmem(struct radeon_winsys_ctx *_ctx,
+					       struct radeon_winsys_cs **cs_array,
+					       unsigned cs_count,
+					       struct radeon_winsys_fence *_fence)
+{
+	int r;
+	struct radv_amdgpu_ctx *ctx = radv_amdgpu_ctx(_ctx);
+	struct amdgpu_cs_fence *fence = (struct amdgpu_cs_fence *)_fence;
+	struct radv_amdgpu_cs *cs0 = radv_amdgpu_cs(cs_array[0]);
+	struct radeon_winsys *ws = (struct radeon_winsys*)cs0->ws;
+	amdgpu_bo_list_handle bo_list;
+	struct amdgpu_cs_request request;
+	uint32_t pad_word = 0xffff1000U;
+
+	if (radv_amdgpu_winsys(ws)->family == FAMILY_SI)
+		pad_word = 0x80000000;
+
+	assert(cs_count);
+
+	for (unsigned i = 0; i < cs_count;) {
+		struct amdgpu_cs_ib_info ib = {0};
+		struct radeon_winsys_bo *bo = NULL;
+		uint32_t *ptr;
+		unsigned cnt = 0;
+		unsigned size = 0;
+
+		while (i + cnt < cs_count && 0xffff8 - size >= radv_amdgpu_cs(cs_array[i + cnt])->base.cdw) {
+			size += radv_amdgpu_cs(cs_array[i + cnt])->base.cdw;
+			++cnt;
+		}
+
+		assert(cnt);
+
+		bo = ws->buffer_create(ws, 4 * size, 4096, RADEON_DOMAIN_GTT, RADEON_FLAG_CPU_ACCESS);
+		ptr = ws->buffer_map(bo);
+
+		for (unsigned j = 0; j < cnt; ++j) {
+			struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[i + j]);
+			memcpy(ptr, cs->base.buf, 4 * cs->base.cdw);
+			ptr += cs->base.cdw;
+
+		}
+
+		while(!size || (size & 7)) {
+			*ptr++ = pad_word;
+			++size;
+		}
+
+		memset(&request, 0, sizeof(request));
+
+
+		r = radv_amdgpu_create_bo_list(cs0->ws, &cs_array[i], cnt,
+		                               (struct radv_amdgpu_winsys_bo*)bo, &bo_list);
+		if (r) {
+			fprintf(stderr, "amdgpu: Failed to created the BO list for submission\n");
+			return r;
+		}
+
+		ib.size = size;
+		ib.ib_mc_address = ws->buffer_get_va(bo);
+
+		request.ip_type = AMDGPU_HW_IP_GFX;
+		request.resources = bo_list;
+		request.number_of_ibs = 1;
+		request.ibs = &ib;
+
+		r = amdgpu_cs_submit(ctx->ctx, 0, &request, 1);
+		if (r) {
+			if (r == -ENOMEM)
+				fprintf(stderr, "amdgpu: Not enough memory for command submission.\n");
+			else
+				fprintf(stderr, "amdgpu: The CS has been rejected, "
+						"see dmesg for more information.\n");
+		}
+
+		amdgpu_bo_list_destroy(bo_list);
+
+		ws->buffer_destroy(bo);
+		if (r)
+			return r;
+
+		i += cnt;
+	}
+	if (fence) {
+		fence->context = ctx->ctx;
+		fence->ip_type = request.ip_type;
+		fence->ip_instance = request.ip_instance;
+		fence->ring = request.ring;
+		fence->fence = request.seq_no;
+	}
+	ctx->last_seq_no = request.seq_no;
+
+	return 0;
+}
+
+static int radv_amdgpu_winsys_cs_submit(struct radeon_winsys_ctx *_ctx,
+					struct radeon_winsys_cs **cs_array,
+					unsigned cs_count,
+					bool can_patch,
+					struct radeon_winsys_fence *_fence)
+{
+	struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[0]);
+	if (!cs->ws->use_ib_bos) {
+		return radv_amdgpu_winsys_cs_submit_sysmem(_ctx, cs_array,
+							   cs_count, _fence);
+	} else if (can_patch && cs_count > AMDGPU_CS_MAX_IBS_PER_SUBMIT && false) {
+		return radv_amdgpu_winsys_cs_submit_chained(_ctx, cs_array,
+							    cs_count, _fence);
+	} else {
+		return radv_amdgpu_winsys_cs_submit_fallback(_ctx, cs_array,
+							     cs_count, _fence);
+	}
+}
+
+static struct radeon_winsys_ctx *radv_amdgpu_ctx_create(struct radeon_winsys *_ws)
+{
+	struct radv_amdgpu_winsys *ws = radv_amdgpu_winsys(_ws);
+	struct radv_amdgpu_ctx *ctx = CALLOC_STRUCT(radv_amdgpu_ctx);
+	int r;
+
+	if (!ctx)
+		return NULL;
+	r = amdgpu_cs_ctx_create(ws->dev, &ctx->ctx);
+	if (r) {
+		fprintf(stderr, "amdgpu: radv_amdgpu_cs_ctx_create failed. (%i)\n", r);
+		goto error_create;
+	}
+	ctx->ws = ws;
+	return (struct radeon_winsys_ctx *)ctx;
+error_create:
+	return NULL;
+}
+
+static void radv_amdgpu_ctx_destroy(struct radeon_winsys_ctx *rwctx)
+{
+	struct radv_amdgpu_ctx *ctx = (struct radv_amdgpu_ctx *)rwctx;
+	amdgpu_cs_ctx_free(ctx->ctx);
+	FREE(ctx);
+}
+
+static bool radv_amdgpu_ctx_wait_idle(struct radeon_winsys_ctx *rwctx)
+{
+	struct radv_amdgpu_ctx *ctx = (struct radv_amdgpu_ctx *)rwctx;
+
+	if (ctx->last_seq_no) {
+		uint32_t expired;
+		struct amdgpu_cs_fence fence;
+
+		fence.context = ctx->ctx;
+		fence.ip_type = RING_GFX;
+		fence.ip_instance = 0;
+		fence.ring = 0;
+		fence.fence = ctx->last_seq_no;
+
+		int ret = amdgpu_cs_query_fence_status(&fence, 1000000000ull, 0,
+						       &expired);
+
+		if (ret || !expired)
+			return false;
+	}
+
+	return true;
+}
+
+void radv_amdgpu_cs_init_functions(struct radv_amdgpu_winsys *ws)
+{
+	ws->base.ctx_create = radv_amdgpu_ctx_create;
+	ws->base.ctx_destroy = radv_amdgpu_ctx_destroy;
+	ws->base.ctx_wait_idle = radv_amdgpu_ctx_wait_idle;
+	ws->base.cs_create = radv_amdgpu_cs_create;
+	ws->base.cs_destroy = radv_amdgpu_cs_destroy;
+	ws->base.cs_grow = radv_amdgpu_cs_grow;
+	ws->base.cs_finalize = radv_amdgpu_cs_finalize;
+	ws->base.cs_reset = radv_amdgpu_cs_reset;
+	ws->base.cs_add_buffer = radv_amdgpu_cs_add_buffer;
+	ws->base.cs_execute_secondary = radv_amdgpu_cs_execute_secondary;
+	ws->base.cs_submit = radv_amdgpu_winsys_cs_submit;
+	ws->base.create_fence = radv_amdgpu_create_fence;
+	ws->base.destroy_fence = radv_amdgpu_destroy_fence;
+	ws->base.fence_wait = radv_amdgpu_fence_wait;
+}
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.h b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.h
new file mode 100644
index 0000000000..230639a258
--- /dev/null
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.h
@@ -0,0 +1,51 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based on amdgpu winsys.
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#pragma once
+
+#include <string.h>
+#include <stdint.h>
+#include <assert.h>
+#include "r600d_common.h"
+#include <amdgpu.h>
+
+#include "radv_radeon_winsys.h"
+
+#include "radv_amdgpu_winsys.h"
+struct radv_amdgpu_ctx {
+	struct radv_amdgpu_winsys *ws;
+	amdgpu_context_handle ctx;
+	uint64_t last_seq_no;
+};
+
+static inline struct radv_amdgpu_ctx *
+radv_amdgpu_ctx(struct radeon_winsys_ctx *base)
+{
+	return (struct radv_amdgpu_ctx *)base;
+}
+
+void radv_amdgpu_cs_init_functions(struct radv_amdgpu_winsys *ws);
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_surface.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_surface.c
new file mode 100644
index 0000000000..a3c24115a1
--- /dev/null
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_surface.c
@@ -0,0 +1,523 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based on amdgpu winsys.
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <errno.h>
+#include "radv_private.h"
+#include "addrlib/addrinterface.h"
+#include "util/bitset.h"
+#include "radv_amdgpu_winsys.h"
+#include "radv_amdgpu_surface.h"
+#include "sid.h"
+#ifndef NO_ENTRIES
+#define NO_ENTRIES 32
+#endif
+
+#ifndef NO_MACRO_ENTRIES
+#define NO_MACRO_ENTRIES 16
+#endif
+
+#ifndef CIASICIDGFXENGINE_SOUTHERNISLAND
+#define CIASICIDGFXENGINE_SOUTHERNISLAND 0x0000000A
+#endif
+
+static int radv_amdgpu_surface_sanity(const struct radeon_surf *surf)
+{
+	unsigned type = RADEON_SURF_GET(surf->flags, TYPE);
+
+	if (!(surf->flags & RADEON_SURF_HAS_TILE_MODE_INDEX))
+		return -EINVAL;
+
+	/* all dimension must be at least 1 ! */
+	if (!surf->npix_x || !surf->npix_y || !surf->npix_z ||
+	    !surf->array_size)
+		return -EINVAL;
+
+	if (!surf->blk_w || !surf->blk_h || !surf->blk_d)
+		return -EINVAL;
+
+	switch (surf->nsamples) {
+	case 1:
+	case 2:
+	case 4:
+	case 8:
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	switch (type) {
+	case RADEON_SURF_TYPE_1D:
+		if (surf->npix_y > 1)
+			return -EINVAL;
+		/* fall through */
+	case RADEON_SURF_TYPE_2D:
+	case RADEON_SURF_TYPE_CUBEMAP:
+		if (surf->npix_z > 1 || surf->array_size > 1)
+			return -EINVAL;
+		break;
+	case RADEON_SURF_TYPE_3D:
+		if (surf->array_size > 1)
+			return -EINVAL;
+		break;
+	case RADEON_SURF_TYPE_1D_ARRAY:
+		if (surf->npix_y > 1)
+			return -EINVAL;
+		/* fall through */
+	case RADEON_SURF_TYPE_2D_ARRAY:
+		if (surf->npix_z > 1)
+			return -EINVAL;
+		break;
+	default:
+		return -EINVAL;
+	}
+	return 0;
+}
+
+static void *ADDR_API radv_allocSysMem(const ADDR_ALLOCSYSMEM_INPUT * pInput)
+{
+	return malloc(pInput->sizeInBytes);
+}
+
+static ADDR_E_RETURNCODE ADDR_API radv_freeSysMem(const ADDR_FREESYSMEM_INPUT * pInput)
+{
+	free(pInput->pVirtAddr);
+	return ADDR_OK;
+}
+
+ADDR_HANDLE radv_amdgpu_addr_create(struct amdgpu_gpu_info *amdinfo, int family, int rev_id,
+				    enum chip_class chip_class)
+{
+	ADDR_CREATE_INPUT addrCreateInput = {0};
+	ADDR_CREATE_OUTPUT addrCreateOutput = {0};
+	ADDR_REGISTER_VALUE regValue = {0};
+	ADDR_CREATE_FLAGS createFlags = {{0}};
+	ADDR_E_RETURNCODE addrRet;
+
+	addrCreateInput.size = sizeof(ADDR_CREATE_INPUT);
+	addrCreateOutput.size = sizeof(ADDR_CREATE_OUTPUT);
+
+	regValue.noOfBanks = amdinfo->mc_arb_ramcfg & 0x3;
+	regValue.gbAddrConfig = amdinfo->gb_addr_cfg;
+	regValue.noOfRanks = (amdinfo->mc_arb_ramcfg & 0x4) >> 2;
+
+	regValue.backendDisables = amdinfo->backend_disable[0];
+	regValue.pTileConfig = amdinfo->gb_tile_mode;
+	regValue.noOfEntries = ARRAY_SIZE(amdinfo->gb_tile_mode);
+	if (chip_class == SI) {
+		regValue.pMacroTileConfig = NULL;
+		regValue.noOfMacroEntries = 0;
+	} else {
+		regValue.pMacroTileConfig = amdinfo->gb_macro_tile_mode;
+		regValue.noOfMacroEntries = ARRAY_SIZE(amdinfo->gb_macro_tile_mode);
+	}
+
+	createFlags.value = 0;
+	createFlags.useTileIndex = 1;
+	createFlags.degradeBaseLevel = 1;
+
+	addrCreateInput.chipEngine = CIASICIDGFXENGINE_SOUTHERNISLAND;
+	addrCreateInput.chipFamily = family;
+	addrCreateInput.chipRevision = rev_id;
+	addrCreateInput.createFlags = createFlags;
+	addrCreateInput.callbacks.allocSysMem = radv_allocSysMem;
+	addrCreateInput.callbacks.freeSysMem = radv_freeSysMem;
+	addrCreateInput.callbacks.debugPrint = 0;
+	addrCreateInput.regValue = regValue;
+
+	addrRet = AddrCreate(&addrCreateInput, &addrCreateOutput);
+	if (addrRet != ADDR_OK)
+		return NULL;
+
+	return addrCreateOutput.hLib;
+}
+
+static int radv_compute_level(ADDR_HANDLE addrlib,
+                              struct radeon_surf *surf, bool is_stencil,
+                              unsigned level, unsigned type, bool compressed,
+                              ADDR_COMPUTE_SURFACE_INFO_INPUT *AddrSurfInfoIn,
+                              ADDR_COMPUTE_SURFACE_INFO_OUTPUT *AddrSurfInfoOut,
+                              ADDR_COMPUTE_DCCINFO_INPUT *AddrDccIn,
+                              ADDR_COMPUTE_DCCINFO_OUTPUT *AddrDccOut)
+{
+	struct radeon_surf_level *surf_level;
+	ADDR_E_RETURNCODE ret;
+
+	AddrSurfInfoIn->mipLevel = level;
+	AddrSurfInfoIn->width = u_minify(surf->npix_x, level);
+	AddrSurfInfoIn->height = u_minify(surf->npix_y, level);
+
+	if (type == RADEON_SURF_TYPE_3D)
+		AddrSurfInfoIn->numSlices = u_minify(surf->npix_z, level);
+	else if (type == RADEON_SURF_TYPE_CUBEMAP)
+		AddrSurfInfoIn->numSlices = 6;
+	else
+		AddrSurfInfoIn->numSlices = surf->array_size;
+
+	if (level > 0) {
+		/* Set the base level pitch. This is needed for calculation
+		 * of non-zero levels. */
+		if (is_stencil)
+			AddrSurfInfoIn->basePitch = surf->stencil_level[0].nblk_x;
+		else
+			AddrSurfInfoIn->basePitch = surf->level[0].nblk_x;
+
+		/* Convert blocks to pixels for compressed formats. */
+		if (compressed)
+			AddrSurfInfoIn->basePitch *= surf->blk_w;
+	}
+
+	ret = AddrComputeSurfaceInfo(addrlib,
+				     AddrSurfInfoIn,
+				     AddrSurfInfoOut);
+	if (ret != ADDR_OK) {
+		return ret;
+	}
+
+	surf_level = is_stencil ? &surf->stencil_level[level] : &surf->level[level];
+	surf_level->offset = align64(surf->bo_size, AddrSurfInfoOut->baseAlign);
+	surf_level->slice_size = AddrSurfInfoOut->sliceSize;
+	surf_level->pitch_bytes = AddrSurfInfoOut->pitch * (is_stencil ? 1 : surf->bpe);
+	surf_level->npix_x = u_minify(surf->npix_x, level);
+	surf_level->npix_y = u_minify(surf->npix_y, level);
+	surf_level->npix_z = u_minify(surf->npix_z, level);
+	surf_level->nblk_x = AddrSurfInfoOut->pitch;
+	surf_level->nblk_y = AddrSurfInfoOut->height;
+	if (type == RADEON_SURF_TYPE_3D)
+		surf_level->nblk_z = AddrSurfInfoOut->depth;
+	else
+		surf_level->nblk_z = 1;
+
+	switch (AddrSurfInfoOut->tileMode) {
+	case ADDR_TM_LINEAR_ALIGNED:
+		surf_level->mode = RADEON_SURF_MODE_LINEAR_ALIGNED;
+		break;
+	case ADDR_TM_1D_TILED_THIN1:
+		surf_level->mode = RADEON_SURF_MODE_1D;
+		break;
+	case ADDR_TM_2D_TILED_THIN1:
+		surf_level->mode = RADEON_SURF_MODE_2D;
+		break;
+	default:
+		assert(0);
+	}
+
+	if (is_stencil)
+		surf->stencil_tiling_index[level] = AddrSurfInfoOut->tileIndex;
+	else
+		surf->tiling_index[level] = AddrSurfInfoOut->tileIndex;
+
+	surf->bo_size = surf_level->offset + AddrSurfInfoOut->surfSize;
+
+	/* Clear DCC fields at the beginning. */
+	surf_level->dcc_offset = 0;
+	surf_level->dcc_enabled = false;
+
+	/* The previous level's flag tells us if we can use DCC for this level. */
+	if (AddrSurfInfoIn->flags.dccCompatible &&
+	    (level == 0 || AddrDccOut->subLvlCompressible)) {
+		AddrDccIn->colorSurfSize = AddrSurfInfoOut->surfSize;
+		AddrDccIn->tileMode = AddrSurfInfoOut->tileMode;
+		AddrDccIn->tileInfo = *AddrSurfInfoOut->pTileInfo;
+		AddrDccIn->tileIndex = AddrSurfInfoOut->tileIndex;
+		AddrDccIn->macroModeIndex = AddrSurfInfoOut->macroModeIndex;
+
+		ret = AddrComputeDccInfo(addrlib,
+					 AddrDccIn,
+					 AddrDccOut);
+
+		if (ret == ADDR_OK) {
+			surf_level->dcc_offset = surf->dcc_size;
+			surf_level->dcc_fast_clear_size = AddrDccOut->dccFastClearSize;
+			surf_level->dcc_enabled = true;
+			surf->dcc_size = surf_level->dcc_offset + AddrDccOut->dccRamSize;
+			surf->dcc_alignment = MAX(surf->dcc_alignment, AddrDccOut->dccRamBaseAlign);
+		}
+	}
+
+	return 0;
+}
+
+static void radv_set_micro_tile_mode(struct radeon_surf *surf,
+                                     struct radeon_info *info)
+{
+	uint32_t tile_mode = info->si_tile_mode_array[surf->tiling_index[0]];
+
+	if (info->chip_class >= CIK)
+		surf->micro_tile_mode = G_009910_MICRO_TILE_MODE_NEW(tile_mode);
+	else
+		surf->micro_tile_mode = G_009910_MICRO_TILE_MODE(tile_mode);
+}
+
+
+static int radv_amdgpu_winsys_surface_init(struct radeon_winsys *_ws,
+					   struct radeon_surf *surf)
+{
+	struct radv_amdgpu_winsys *ws = radv_amdgpu_winsys(_ws);
+	unsigned level, mode, type;
+	bool compressed;
+	ADDR_COMPUTE_SURFACE_INFO_INPUT AddrSurfInfoIn = {0};
+	ADDR_COMPUTE_SURFACE_INFO_OUTPUT AddrSurfInfoOut = {0};
+	ADDR_COMPUTE_DCCINFO_INPUT AddrDccIn = {0};
+	ADDR_COMPUTE_DCCINFO_OUTPUT AddrDccOut = {0};
+	ADDR_TILEINFO AddrTileInfoIn = {0};
+	ADDR_TILEINFO AddrTileInfoOut = {0};
+	int r;
+
+	r = radv_amdgpu_surface_sanity(surf);
+	if (r)
+		return r;
+
+	AddrSurfInfoIn.size = sizeof(ADDR_COMPUTE_SURFACE_INFO_INPUT);
+	AddrSurfInfoOut.size = sizeof(ADDR_COMPUTE_SURFACE_INFO_OUTPUT);
+	AddrDccIn.size = sizeof(ADDR_COMPUTE_DCCINFO_INPUT);
+	AddrDccOut.size = sizeof(ADDR_COMPUTE_DCCINFO_OUTPUT);
+	AddrSurfInfoOut.pTileInfo = &AddrTileInfoOut;
+
+	type = RADEON_SURF_GET(surf->flags, TYPE);
+	mode = RADEON_SURF_GET(surf->flags, MODE);
+	compressed = surf->blk_w == 4 && surf->blk_h == 4;
+
+	/* MSAA and FMASK require 2D tiling. */
+	if (surf->nsamples > 1 ||
+	    (surf->flags & RADEON_SURF_FMASK))
+		mode = RADEON_SURF_MODE_2D;
+
+	/* DB doesn't support linear layouts. */
+	if (surf->flags & (RADEON_SURF_Z_OR_SBUFFER) &&
+	    mode < RADEON_SURF_MODE_1D)
+		mode = RADEON_SURF_MODE_1D;
+
+	/* Set the requested tiling mode. */
+	switch (mode) {
+	case RADEON_SURF_MODE_LINEAR_ALIGNED:
+		AddrSurfInfoIn.tileMode = ADDR_TM_LINEAR_ALIGNED;
+		break;
+	case RADEON_SURF_MODE_1D:
+		AddrSurfInfoIn.tileMode = ADDR_TM_1D_TILED_THIN1;
+		break;
+	case RADEON_SURF_MODE_2D:
+		AddrSurfInfoIn.tileMode = ADDR_TM_2D_TILED_THIN1;
+		break;
+	default:
+		assert(0);
+	}
+
+	/* The format must be set correctly for the allocation of compressed
+	 * textures to work. In other cases, setting the bpp is sufficient. */
+	if (compressed) {
+		switch (surf->bpe) {
+		case 8:
+			AddrSurfInfoIn.format = ADDR_FMT_BC1;
+			break;
+		case 16:
+			AddrSurfInfoIn.format = ADDR_FMT_BC3;
+			break;
+		default:
+			assert(0);
+		}
+	}
+	else {
+		AddrDccIn.bpp = AddrSurfInfoIn.bpp = surf->bpe * 8;
+	}
+
+	AddrDccIn.numSamples = AddrSurfInfoIn.numSamples = surf->nsamples;
+	AddrSurfInfoIn.tileIndex = -1;
+
+	/* Set the micro tile type. */
+	if (surf->flags & RADEON_SURF_SCANOUT)
+		AddrSurfInfoIn.tileType = ADDR_DISPLAYABLE;
+	else if (surf->flags & RADEON_SURF_Z_OR_SBUFFER)
+		AddrSurfInfoIn.tileType = ADDR_DEPTH_SAMPLE_ORDER;
+	else
+		AddrSurfInfoIn.tileType = ADDR_NON_DISPLAYABLE;
+
+	AddrSurfInfoIn.flags.color = !(surf->flags & RADEON_SURF_Z_OR_SBUFFER);
+	AddrSurfInfoIn.flags.depth = (surf->flags & RADEON_SURF_ZBUFFER) != 0;
+	AddrSurfInfoIn.flags.cube = type == RADEON_SURF_TYPE_CUBEMAP;
+	AddrSurfInfoIn.flags.display = (surf->flags & RADEON_SURF_SCANOUT) != 0;
+	AddrSurfInfoIn.flags.pow2Pad = surf->last_level > 0;
+	AddrSurfInfoIn.flags.degrade4Space = 1;
+
+	/* DCC notes:
+	 * - If we add MSAA support, keep in mind that CB can't decompress 8bpp
+	 *   with samples >= 4.
+	 * - Mipmapped array textures have low performance (discovered by a closed
+	 *   driver team).
+	 */
+	AddrSurfInfoIn.flags.dccCompatible = !(surf->flags & RADEON_SURF_Z_OR_SBUFFER) &&
+		!(surf->flags & RADEON_SURF_DISABLE_DCC) &&
+		!compressed && AddrDccIn.numSamples <= 1 &&
+		((surf->array_size == 1 && surf->npix_z == 1) ||
+		 surf->last_level == 0);
+
+	AddrSurfInfoIn.flags.noStencil = (surf->flags & RADEON_SURF_SBUFFER) == 0;
+	AddrSurfInfoIn.flags.compressZ = AddrSurfInfoIn.flags.depth;
+
+	/* noStencil = 0 can result in a depth part that is incompatible with
+	 * mipmapped texturing. So set noStencil = 1 when mipmaps are requested (in
+	 * this case, we may end up setting stencil_adjusted).
+	 *
+	 * TODO: update addrlib to a newer version, remove this, and
+	 * use flags.matchStencilTileCfg = 1 as an alternative fix.
+	 */
+	if (surf->last_level > 0)
+		AddrSurfInfoIn.flags.noStencil = 1;
+
+	/* Set preferred macrotile parameters. This is usually required
+	 * for shared resources. This is for 2D tiling only. */
+	if (AddrSurfInfoIn.tileMode >= ADDR_TM_2D_TILED_THIN1 &&
+	    surf->bankw && surf->bankh && surf->mtilea && surf->tile_split) {
+		/* If any of these parameters are incorrect, the calculation
+		 * will fail. */
+		AddrTileInfoIn.banks = surf->num_banks;
+		AddrTileInfoIn.bankWidth = surf->bankw;
+		AddrTileInfoIn.bankHeight = surf->bankh;
+		AddrTileInfoIn.macroAspectRatio = surf->mtilea;
+		AddrTileInfoIn.tileSplitBytes = surf->tile_split;
+		AddrTileInfoIn.pipeConfig = surf->pipe_config + 1; /* +1 compared to GB_TILE_MODE */
+		AddrSurfInfoIn.flags.degrade4Space = 0;
+		AddrSurfInfoIn.pTileInfo = &AddrTileInfoIn;
+
+		/* If AddrSurfInfoIn.pTileInfo is set, Addrlib doesn't set
+		 * the tile index, because we are expected to know it if
+		 * we know the other parameters.
+		 *
+		 * This is something that can easily be fixed in Addrlib.
+		 * For now, just figure it out here.
+		 * Note that only 2D_TILE_THIN1 is handled here.
+		 */
+		assert(!(surf->flags & RADEON_SURF_Z_OR_SBUFFER));
+		assert(AddrSurfInfoIn.tileMode == ADDR_TM_2D_TILED_THIN1);
+
+		if (ws->info.chip_class == SI) {
+			if (AddrSurfInfoIn.tileType == ADDR_DISPLAYABLE) {
+				if (surf->bpe == 2)
+					AddrSurfInfoIn.tileIndex = 11; /* 16bpp */
+				else
+					AddrSurfInfoIn.tileIndex = 12; /* 32bpp */
+			} else {
+				if (surf->bpe == 1)
+					AddrSurfInfoIn.tileIndex = 14; /* 8bpp */
+				else if (surf->bpe == 2)
+					AddrSurfInfoIn.tileIndex = 15; /* 16bpp */
+				else if (surf->bpe == 4)
+					AddrSurfInfoIn.tileIndex = 16; /* 32bpp */
+				else
+					AddrSurfInfoIn.tileIndex = 17; /* 64bpp (and 128bpp) */
+			}
+		} else {
+			if (AddrSurfInfoIn.tileType == ADDR_DISPLAYABLE)
+				AddrSurfInfoIn.tileIndex = 10; /* 2D displayable */
+			else
+				AddrSurfInfoIn.tileIndex = 14; /* 2D non-displayable */
+		}
+	}
+
+	surf->bo_size = 0;
+	surf->dcc_size = 0;
+	surf->dcc_alignment = 1;
+
+	/* Calculate texture layout information. */
+	for (level = 0; level <= surf->last_level; level++) {
+		r = radv_compute_level(ws->addrlib, surf, false, level, type, compressed,
+				       &AddrSurfInfoIn, &AddrSurfInfoOut, &AddrDccIn, &AddrDccOut);
+		if (r)
+			return r;
+
+		if (level == 0) {
+			surf->bo_alignment = AddrSurfInfoOut.baseAlign;
+			surf->pipe_config = AddrSurfInfoOut.pTileInfo->pipeConfig - 1;
+			radv_set_micro_tile_mode(surf, &ws->info);
+
+			/* For 2D modes only. */
+			if (AddrSurfInfoOut.tileMode >= ADDR_TM_2D_TILED_THIN1) {
+				surf->bankw = AddrSurfInfoOut.pTileInfo->bankWidth;
+				surf->bankh = AddrSurfInfoOut.pTileInfo->bankHeight;
+				surf->mtilea = AddrSurfInfoOut.pTileInfo->macroAspectRatio;
+				surf->tile_split = AddrSurfInfoOut.pTileInfo->tileSplitBytes;
+				surf->num_banks = AddrSurfInfoOut.pTileInfo->banks;
+				surf->macro_tile_index = AddrSurfInfoOut.macroModeIndex;
+			} else {
+				surf->macro_tile_index = 0;
+			}
+		}
+	}
+
+	/* Calculate texture layout information for stencil. */
+	if (surf->flags & RADEON_SURF_SBUFFER) {
+		AddrSurfInfoIn.bpp = 8;
+		AddrSurfInfoIn.flags.depth = 0;
+		AddrSurfInfoIn.flags.stencil = 1;
+		/* This will be ignored if AddrSurfInfoIn.pTileInfo is NULL. */
+		AddrTileInfoIn.tileSplitBytes = surf->stencil_tile_split;
+
+		for (level = 0; level <= surf->last_level; level++) {
+			r = radv_compute_level(ws->addrlib, surf, true, level, type, compressed,
+					       &AddrSurfInfoIn, &AddrSurfInfoOut, &AddrDccIn, &AddrDccOut);
+			if (r)
+				return r;
+
+			/* DB uses the depth pitch for both stencil and depth. */
+			if (surf->stencil_level[level].nblk_x != surf->level[level].nblk_x)
+				surf->stencil_adjusted = true;
+
+			if (level == 0) {
+				/* For 2D modes only. */
+				if (AddrSurfInfoOut.tileMode >= ADDR_TM_2D_TILED_THIN1) {
+					surf->stencil_tile_split =
+						AddrSurfInfoOut.pTileInfo->tileSplitBytes;
+				}
+			}
+		}
+	}
+
+	/* Recalculate the whole DCC miptree size including disabled levels.
+	 * This is what addrlib does, but calling addrlib would be a lot more
+	 * complicated.
+	 */
+#if 0
+	if (surf->dcc_size && surf->last_level > 0) {
+		surf->dcc_size = align64(surf->bo_size >> 8,
+					 ws->info.pipe_interleave_bytes *
+					 ws->info.num_tile_pipes);
+	}
+#endif
+	return 0;
+}
+
+static int radv_amdgpu_winsys_surface_best(struct radeon_winsys *rws,
+					   struct radeon_surf *surf)
+{
+	return 0;
+}
+
+void radv_amdgpu_surface_init_functions(struct radv_amdgpu_winsys *ws)
+{
+	ws->base.surface_init = radv_amdgpu_winsys_surface_init;
+	ws->base.surface_best = radv_amdgpu_winsys_surface_best;
+}
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_surface.h b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_surface.h
new file mode 100644
index 0000000000..acc12af3d0
--- /dev/null
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_surface.h
@@ -0,0 +1,29 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#pragma once
+
+#include <amdgpu.h>
+
+void radv_amdgpu_surface_init_functions(struct radv_amdgpu_winsys *ws);
+ADDR_HANDLE radv_amdgpu_addr_create(struct amdgpu_gpu_info *amdinfo, int family, int rev_id, enum chip_class chip_class);
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c
new file mode 100644
index 0000000000..94505367e2
--- /dev/null
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c
@@ -0,0 +1,359 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ * based on amdgpu winsys.
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#include "radv_amdgpu_winsys.h"
+#include "radv_amdgpu_winsys_public.h"
+#include "radv_amdgpu_surface.h"
+#include "amdgpu_id.h"
+#include "xf86drm.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <amdgpu_drm.h>
+#include <assert.h>
+#include "radv_amdgpu_cs.h"
+#include "radv_amdgpu_bo.h"
+#include "radv_amdgpu_surface.h"
+#define CIK_TILE_MODE_COLOR_2D			14
+
+#define CIK__GB_TILE_MODE__PIPE_CONFIG(x)        (((x) >> 6) & 0x1f)
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P2               0
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16          4
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16         5
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32         6
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32         7
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16    8
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16    9
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16    10
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16   11
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16   12
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32   13
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32   14
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16   16
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16  17
+
+static unsigned radv_cik_get_num_tile_pipes(struct amdgpu_gpu_info *info)
+{
+	unsigned mode2d = info->gb_tile_mode[CIK_TILE_MODE_COLOR_2D];
+
+	switch (CIK__GB_TILE_MODE__PIPE_CONFIG(mode2d)) {
+	case CIK__PIPE_CONFIG__ADDR_SURF_P2:
+		return 2;
+	case CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32:
+		return 4;
+	case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32:
+		return 8;
+	case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16:
+	case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16:
+		return 16;
+	default:
+		fprintf(stderr, "Invalid CIK pipe configuration, assuming P2\n");
+		assert(!"this should never occur");
+		return 2;
+	}
+}
+
+static const char *
+get_chip_name(enum radeon_family family)
+{
+	switch (family) {
+	case CHIP_TAHITI: return "AMD RADV TAHITI";
+	case CHIP_PITCAIRN: return "AMD RADV PITCAIRN";
+	case CHIP_VERDE: return "AMD RADV CAPE VERDE";
+	case CHIP_OLAND: return "AMD RADV OLAND";
+	case CHIP_HAINAN: return "AMD RADV HAINAN";
+	case CHIP_BONAIRE: return "AMD RADV BONAIRE";
+	case CHIP_KAVERI: return "AMD RADV KAVERI";
+	case CHIP_KABINI: return "AMD RADV KABINI";
+	case CHIP_HAWAII: return "AMD RADV HAWAII";
+	case CHIP_MULLINS: return "AMD RADV MULLINS";
+	case CHIP_TONGA: return "AMD RADV TONGA";
+	case CHIP_ICELAND: return "AMD RADV ICELAND";
+	case CHIP_CARRIZO: return "AMD RADV CARRIZO";
+	case CHIP_FIJI: return "AMD RADV FIJI";
+	case CHIP_POLARIS10: return "AMD RADV POLARIS10";
+	case CHIP_POLARIS11: return "AMD RADV POLARIS11";
+	case CHIP_STONEY: return "AMD RADV STONEY";
+	default: return "AMD RADV unknown";
+	}
+}
+
+
+static bool
+do_winsys_init(struct radv_amdgpu_winsys *ws, int fd)
+{
+	struct amdgpu_buffer_size_alignments alignment_info = {};
+	struct amdgpu_heap_info vram, gtt;
+	struct drm_amdgpu_info_hw_ip dma = {};
+	drmDevicePtr devinfo;
+	int r;
+	int i, j;
+	/* Get PCI info. */
+	r = drmGetDevice(fd, &devinfo);
+	if (r) {
+		fprintf(stderr, "amdgpu: drmGetDevice failed.\n");
+		goto fail;
+	}
+	ws->info.pci_domain = devinfo->businfo.pci->domain;
+	ws->info.pci_bus = devinfo->businfo.pci->bus;
+	ws->info.pci_dev = devinfo->businfo.pci->dev;
+	ws->info.pci_func = devinfo->businfo.pci->func;
+	drmFreeDevice(&devinfo);
+
+	/* Query hardware and driver information. */
+	r = amdgpu_query_gpu_info(ws->dev, &ws->amdinfo);
+	if (r) {
+		fprintf(stderr, "amdgpu: amdgpu_query_gpu_info failed.\n");
+		goto fail;
+	}
+
+	r = amdgpu_query_buffer_size_alignment(ws->dev, &alignment_info);
+	if (r) {
+		fprintf(stderr, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
+		goto fail;
+	}
+
+	r = amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_VRAM, 0, &vram);
+	if (r) {
+		fprintf(stderr, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
+		goto fail;
+	}
+
+	r = amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_GTT, 0, &gtt);
+	if (r) {
+		fprintf(stderr, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
+		goto fail;
+	}
+
+	r = amdgpu_query_hw_ip_info(ws->dev, AMDGPU_HW_IP_DMA, 0, &dma);
+	if (r) {
+		fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(dma) failed.\n");
+		goto fail;
+	}
+	ws->info.pci_id = ws->amdinfo.asic_id; /* TODO: is this correct? */
+	ws->info.vce_harvest_config = ws->amdinfo.vce_harvest_config;
+
+	switch (ws->info.pci_id) {
+#define CHIPSET(pci_id, name, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; break;
+#include "pci_ids/radeonsi_pci_ids.h"
+#undef CHIPSET
+	default:
+		fprintf(stderr, "amdgpu: Invalid PCI ID.\n");
+		goto fail;
+	}
+
+	if (ws->info.family >= CHIP_TONGA)
+		ws->info.chip_class = VI;
+	else if (ws->info.family >= CHIP_BONAIRE)
+		ws->info.chip_class = CIK;
+	else if (ws->info.family >= CHIP_TAHITI)
+		ws->info.chip_class = SI;
+	else {
+		fprintf(stderr, "amdgpu: Unknown family.\n");
+		goto fail;
+	}
+
+	/* family and rev_id are for addrlib */
+	switch (ws->info.family) {
+	case CHIP_TAHITI:
+		ws->family = FAMILY_SI;
+		ws->rev_id = SI_TAHITI_P_A0;
+		break;
+	case CHIP_PITCAIRN:
+		ws->family = FAMILY_SI;
+		ws->rev_id = SI_PITCAIRN_PM_A0;
+	  break;
+	case CHIP_VERDE:
+		ws->family = FAMILY_SI;
+		ws->rev_id = SI_CAPEVERDE_M_A0;
+		break;
+	case CHIP_OLAND:
+		ws->family = FAMILY_SI;
+		ws->rev_id = SI_OLAND_M_A0;
+		break;
+	case CHIP_HAINAN:
+		ws->family = FAMILY_SI;
+		ws->rev_id = SI_HAINAN_V_A0;
+		break;
+	case CHIP_BONAIRE:
+		ws->family = FAMILY_CI;
+		ws->rev_id = CI_BONAIRE_M_A0;
+		break;
+	case CHIP_KAVERI:
+		ws->family = FAMILY_KV;
+		ws->rev_id = KV_SPECTRE_A0;
+		break;
+	case CHIP_KABINI:
+		ws->family = FAMILY_KV;
+		ws->rev_id = KB_KALINDI_A0;
+		break;
+	case CHIP_HAWAII:
+		ws->family = FAMILY_CI;
+		ws->rev_id = CI_HAWAII_P_A0;
+		break;
+	case CHIP_MULLINS:
+		ws->family = FAMILY_KV;
+		ws->rev_id = ML_GODAVARI_A0;
+		break;
+	case CHIP_TONGA:
+		ws->family = FAMILY_VI;
+		ws->rev_id = VI_TONGA_P_A0;
+		break;
+	case CHIP_ICELAND:
+		ws->family = FAMILY_VI;
+		ws->rev_id = VI_ICELAND_M_A0;
+		break;
+	case CHIP_CARRIZO:
+		ws->family = FAMILY_CZ;
+		ws->rev_id = CARRIZO_A0;
+		break;
+	case CHIP_STONEY:
+		ws->family = FAMILY_CZ;
+		ws->rev_id = STONEY_A0;
+		break;
+	case CHIP_FIJI:
+		ws->family = FAMILY_VI;
+		ws->rev_id = VI_FIJI_P_A0;
+		break;
+	case CHIP_POLARIS10:
+		ws->family = FAMILY_VI;
+		ws->rev_id = VI_POLARIS10_P_A0;
+		break;
+	case CHIP_POLARIS11:
+		ws->family = FAMILY_VI;
+		ws->rev_id = VI_POLARIS11_M_A0;
+		break;
+	default:
+		fprintf(stderr, "amdgpu: Unknown family.\n");
+		goto fail;
+	}
+
+	ws->addrlib = radv_amdgpu_addr_create(&ws->amdinfo, ws->family, ws->rev_id, ws->info.chip_class);
+	if (!ws->addrlib) {
+		fprintf(stderr, "amdgpu: Cannot create addrlib.\n");
+		goto fail;
+	}
+	/* Set hardware information. */
+	ws->info.name = get_chip_name(ws->info.family);
+	ws->info.gart_size = gtt.heap_size;
+	ws->info.vram_size = vram.heap_size;
+	/* convert the shader clock from KHz to MHz */
+	ws->info.max_shader_clock = ws->amdinfo.max_engine_clk / 1000;
+	ws->info.max_se = ws->amdinfo.num_shader_engines;
+	ws->info.max_sh_per_se = ws->amdinfo.num_shader_arrays_per_engine;
+	ws->info.has_uvd = 0;
+	ws->info.vce_fw_version = 0;
+	ws->info.has_userptr = TRUE;
+	ws->info.num_render_backends = ws->amdinfo.rb_pipes;
+	ws->info.clock_crystal_freq = ws->amdinfo.gpu_counter_freq;
+	ws->info.num_tile_pipes = radv_cik_get_num_tile_pipes(&ws->amdinfo);
+	ws->info.pipe_interleave_bytes = 256 << ((ws->amdinfo.gb_addr_cfg >> 4) & 0x7);
+	ws->info.has_virtual_memory = TRUE;
+	ws->info.has_sdma = dma.available_rings != 0;
+
+	/* Get the number of good compute units. */
+	ws->info.num_good_compute_units = 0;
+	for (i = 0; i < ws->info.max_se; i++)
+		for (j = 0; j < ws->info.max_sh_per_se; j++)
+			ws->info.num_good_compute_units +=
+				util_bitcount(ws->amdinfo.cu_bitmap[i][j]);
+
+	memcpy(ws->info.si_tile_mode_array, ws->amdinfo.gb_tile_mode,
+	       sizeof(ws->amdinfo.gb_tile_mode));
+	ws->info.enabled_rb_mask = ws->amdinfo.enabled_rb_pipes_mask;
+
+	memcpy(ws->info.cik_macrotile_mode_array, ws->amdinfo.gb_macro_tile_mode,
+	       sizeof(ws->amdinfo.gb_macro_tile_mode));
+
+	ws->info.gart_page_size = alignment_info.size_remote;
+
+	if (ws->info.chip_class == SI)
+		ws->info.gfx_ib_pad_with_type2 = TRUE;
+
+	ws->use_ib_bos = ws->family >= FAMILY_CI;
+	return true;
+fail:
+	return false;
+}
+
+static void radv_amdgpu_winsys_query_info(struct radeon_winsys *rws,
+                                     struct radeon_info *info)
+{
+	*info = ((struct radv_amdgpu_winsys *)rws)->info;
+}
+
+static void radv_amdgpu_winsys_destroy(struct radeon_winsys *rws)
+{
+	struct radv_amdgpu_winsys *ws = (struct radv_amdgpu_winsys*)rws;
+
+	AddrDestroy(ws->addrlib);
+	amdgpu_device_deinitialize(ws->dev);
+	FREE(rws);
+}
+
+struct radeon_winsys *
+radv_amdgpu_winsys_create(int fd)
+{
+	uint32_t drm_major, drm_minor, r;
+	amdgpu_device_handle dev;
+	struct radv_amdgpu_winsys *ws;
+
+	r = amdgpu_device_initialize(fd, &drm_major, &drm_minor, &dev);
+	if (r)
+		return NULL;
+
+	ws = calloc(1, sizeof(struct radv_amdgpu_winsys));
+	if (!ws)
+		return NULL;
+
+
+	ws->dev = dev;
+	ws->info.drm_major = drm_major;
+	ws->info.drm_minor = drm_minor;
+	if (!do_winsys_init(ws, fd))
+		goto fail;
+
+	ws->debug_all_bos = getenv("RADV_DEBUG_ALL_BOS") ? true : false;
+	LIST_INITHEAD(&ws->global_bo_list);
+	pthread_mutex_init(&ws->global_bo_list_lock, NULL);
+	ws->base.query_info = radv_amdgpu_winsys_query_info;
+	ws->base.destroy = radv_amdgpu_winsys_destroy;
+	radv_amdgpu_bo_init_functions(ws);
+	radv_amdgpu_cs_init_functions(ws);
+	radv_amdgpu_surface_init_functions(ws);
+	return &ws->base;
+fail:
+	return NULL;
+}
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.h b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.h
new file mode 100644
index 0000000000..b79495d995
--- /dev/null
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.h
@@ -0,0 +1,57 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ * based on amdgpu winsys.
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#pragma once
+
+#include "radv_radeon_winsys.h"
+#include "addrlib/addrinterface.h"
+#include <amdgpu.h>
+#include "util/list.h"
+
+struct radv_amdgpu_winsys {
+	struct radeon_winsys base;
+	amdgpu_device_handle dev;
+
+	struct radeon_info info;
+	struct amdgpu_gpu_info amdinfo;
+	ADDR_HANDLE addrlib;
+
+	uint32_t rev_id;
+	unsigned family;
+
+	bool debug_all_bos;
+	pthread_mutex_t global_bo_list_lock;
+	struct list_head global_bo_list;
+	unsigned num_buffers;
+
+	bool use_ib_bos;
+};
+
+static inline struct radv_amdgpu_winsys *
+radv_amdgpu_winsys(struct radeon_winsys *base)
+{
+	return (struct radv_amdgpu_winsys*)base;
+}
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys_public.h b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys_public.h
new file mode 100644
index 0000000000..cf066011c2
--- /dev/null
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys_public.h
@@ -0,0 +1,30 @@
+/*
+ * Copyright © 2016 Red Hat.
+ * Copyright © 2016 Bas Nieuwenhuizen
+ *
+ * based on amdgpu winsys.
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+#pragma once
+
+struct radeon_winsys *radv_amdgpu_winsys_create(int fd);