anv: Support new local ID generation & cross-thread constants

The cross thread constant support appears on Haswell. It allows us to
upload a set of uniform data for all threads without duplicating it
per thread.

We also support per-thread data which allows us to store a per-thread
ID in one of the uniforms that can be used to calculate the
gl_LocalInvocationIndex and gl_LocalInvocationID variables.

v4:
 * Support the old local ID push constant layout as well (Jason)

Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

4 files changed, 42 insertions, 48 deletions

diff --git a/src/intel/vulkan/anv_cmd_buffer.c b/src/intel/vulkan/anv_cmd_buffer.c
index 63d096c6e8..edaaa3d2ef 100644
--- a/src/intel/vulkan/anv_cmd_buffer.c
+++ b/src/intel/vulkan/anv_cmd_buffer.c
@@ -1065,23 +1065,14 @@ anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
 
-   const unsigned local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
-   const unsigned push_constant_data_size =
-      (local_id_dwords + prog_data->nr_params) * 4;
-   const unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
-   const unsigned param_aligned_count =
-      reg_aligned_constant_size / sizeof(uint32_t);
-
    /* If we don't actually have any push constants, bail. */
-   if (reg_aligned_constant_size == 0)
+   if (cs_prog_data->push.total.size == 0)
       return (struct anv_state) { .offset = 0 };
 
-   const unsigned total_push_constants_size =
-      reg_aligned_constant_size * cs_prog_data->threads;
    const unsigned push_constant_alignment =
       cmd_buffer->device->info.gen < 8 ? 32 : 64;
    const unsigned aligned_total_push_constants_size =
-      ALIGN(total_push_constants_size, push_constant_alignment);
+      ALIGN(cs_prog_data->push.total.size, push_constant_alignment);
    struct anv_state state =
       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
                                          aligned_total_push_constants_size,
@@ -1090,21 +1081,36 @@ anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
    /* Walk through the param array and fill the buffer with data */
    uint32_t *u32_map = state.map;
 
-   brw_cs_fill_local_id_payload(cs_prog_data, u32_map, cs_prog_data->threads,
-                                reg_aligned_constant_size);
-
-   /* Setup uniform data for the first thread */
-   for (unsigned i = 0; i < prog_data->nr_params; i++) {
-      uint32_t offset = (uintptr_t)prog_data->param[i];
-      u32_map[local_id_dwords + i] = *(uint32_t *)((uint8_t *)data + offset);
+   if (cs_prog_data->push.cross_thread.size > 0) {
+      assert(cs_prog_data->thread_local_id_index < 0 ||
+             cs_prog_data->thread_local_id_index >=
+                cs_prog_data->push.cross_thread.dwords);
+      for (unsigned i = 0;
+           i < cs_prog_data->push.cross_thread.dwords;
+           i++) {
+         uint32_t offset = (uintptr_t)prog_data->param[i];
+         u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
+      }
    }
 
-   /* Copy uniform data from the first thread to every other thread */
-   const size_t uniform_data_size = prog_data->nr_params * sizeof(uint32_t);
-   for (unsigned t = 1; t < cs_prog_data->threads; t++) {
-      memcpy(&u32_map[t * param_aligned_count + local_id_dwords],
-             &u32_map[local_id_dwords],
-             uniform_data_size);
+   if (cs_prog_data->push.per_thread.size > 0) {
+      brw_cs_fill_local_id_payload(cs_prog_data, u32_map, cs_prog_data->threads,
+                                   cs_prog_data->push.per_thread.size);
+      for (unsigned t = 0; t < cs_prog_data->threads; t++) {
+         unsigned dst =
+            8 * (cs_prog_data->push.per_thread.regs * t +
+                 cs_prog_data->push.cross_thread.regs +
+                 cs_prog_data->local_invocation_id_regs);
+         unsigned src = cs_prog_data->push.cross_thread.dwords;
+         for ( ; src < prog_data->nr_params; src++, dst++) {
+            if (src != cs_prog_data->thread_local_id_index) {
+               uint32_t offset = (uintptr_t)prog_data->param[src];
+               u32_map[dst] = *(uint32_t *)((uint8_t *)data + offset);
+            } else {
+               u32_map[dst] = t * cs_prog_data->simd_size;
+            }
+         }
+      }
    }
 
    if (!cmd_buffer->device->info.has_llc)
diff --git a/src/intel/vulkan/gen7_cmd_buffer.c b/src/intel/vulkan/gen7_cmd_buffer.c
index 40ab008990..478122b9c0 100644
--- a/src/intel/vulkan/gen7_cmd_buffer.c
+++ b/src/intel/vulkan/gen7_cmd_buffer.c
@@ -234,12 +234,6 @@ flush_compute_descriptor_set(struct anv_cmd_buffer *cmd_buffer)
    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
 
-   unsigned local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
-   unsigned push_constant_data_size =
-      (prog_data->nr_params + local_id_dwords) * 4;
-   unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
-   unsigned push_constant_regs = reg_aligned_constant_size / 32;
-
    if (push_state.alloc_size) {
       anv_batch_emit(&cmd_buffer->batch, GENX(MEDIA_CURBE_LOAD), curbe) {
          curbe.CURBETotalDataLength    = push_state.alloc_size;
@@ -264,8 +258,11 @@ flush_compute_descriptor_set(struct anv_cmd_buffer *cmd_buffer)
                           .BindingTablePointer = surfaces.offset,
                           .SamplerStatePointer = samplers.offset,
                           .ConstantURBEntryReadLength =
-                             push_constant_regs,
-#if !GEN_IS_HASWELL
+                             cs_prog_data->push.per_thread.regs,
+#if GEN_IS_HASWELL
+                          .CrossThreadConstantDataReadLength =
+                             cs_prog_data->push.cross_thread.regs,
+#else
                           .ConstantURBEntryReadOffset = 0,
 #endif
                           .BarrierEnable = cs_prog_data->uses_barrier,
diff --git a/src/intel/vulkan/gen8_cmd_buffer.c b/src/intel/vulkan/gen8_cmd_buffer.c
index e139e8af96..1fdc7d29ea 100644
--- a/src/intel/vulkan/gen8_cmd_buffer.c
+++ b/src/intel/vulkan/gen8_cmd_buffer.c
@@ -319,12 +319,6 @@ flush_compute_descriptor_set(struct anv_cmd_buffer *cmd_buffer)
    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
 
-   unsigned local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
-   unsigned push_constant_data_size =
-      (prog_data->nr_params + local_id_dwords) * 4;
-   unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
-   unsigned push_constant_regs = reg_aligned_constant_size / 32;
-
    if (push_state.alloc_size) {
       anv_batch_emit(&cmd_buffer->batch, GENX(MEDIA_CURBE_LOAD), curbe) {
          curbe.CURBETotalDataLength    = push_state.alloc_size;
@@ -351,12 +345,15 @@ flush_compute_descriptor_set(struct anv_cmd_buffer *cmd_buffer)
                           .BindingTableEntryCount = 0,
                           .SamplerStatePointer = samplers.offset,
                           .SamplerCount = 0,
-                          .ConstantIndirectURBEntryReadLength = push_constant_regs,
+                          .ConstantIndirectURBEntryReadLength =
+                             cs_prog_data->push.per_thread.regs,
                           .ConstantURBEntryReadOffset = 0,
                           .BarrierEnable = cs_prog_data->uses_barrier,
                           .SharedLocalMemorySize = slm_size,
                           .NumberofThreadsinGPGPUThreadGroup =
-                             cs_prog_data->threads);
+                             cs_prog_data->threads,
+                          .CrossThreadConstantDataReadLength =
+                             cs_prog_data->push.cross_thread.regs);
 
    uint32_t size = GENX(INTERFACE_DESCRIPTOR_DATA_length) * sizeof(uint32_t);
    anv_batch_emit(&cmd_buffer->batch,
diff --git a/src/intel/vulkan/genX_pipeline.c b/src/intel/vulkan/genX_pipeline.c
index 1776577d0a..458e80c82b 100644
--- a/src/intel/vulkan/genX_pipeline.c
+++ b/src/intel/vulkan/genX_pipeline.c
@@ -87,13 +87,6 @@ genX(compute_pipeline_create)(
    anv_setup_pipeline_l3_config(pipeline);
 
    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
-   const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
-
-   unsigned local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
-   unsigned push_constant_data_size =
-      (prog_data->nr_params + local_id_dwords) * 4;
-   unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
-   unsigned push_constant_regs = reg_aligned_constant_size / 32;
 
    uint32_t group_size = cs_prog_data->local_size[0] *
       cs_prog_data->local_size[1] * cs_prog_data->local_size[2];
@@ -105,7 +98,8 @@ genX(compute_pipeline_create)(
       pipeline->cs_right_mask = ~0u >> (32 - cs_prog_data->simd_size);
 
    const uint32_t vfe_curbe_allocation =
-      push_constant_regs * cs_prog_data->threads;
+      ALIGN(cs_prog_data->push.per_thread.regs * cs_prog_data->threads +
+            cs_prog_data->push.cross_thread.regs, 2);
 
    anv_batch_emit(&pipeline->batch, GENX(MEDIA_VFE_STATE), vfe) {
       vfe.ScratchSpaceBasePointer = pipeline->scratch_start[MESA_SHADER_COMPUTE];