radeonsi: move fence functions to si_fence.c

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

6 files changed, 312 insertions, 267 deletions

diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index 486d635527..f6e97e9e4e 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -36,23 +36,10 @@
 #include "amd/common/sid.h"
 #include <inttypes.h>
 #include <sys/utsname.h>
-#include <libsync.h>
 
 #include <llvm-c/TargetMachine.h>
 
 
-struct r600_multi_fence {
-	struct pipe_reference reference;
-	struct pipe_fence_handle *gfx;
-	struct pipe_fence_handle *sdma;
-
-	/* If the context wasn't flushed at fence creation, this is non-NULL. */
-	struct {
-		struct r600_common_context *ctx;
-		unsigned ib_index;
-	} gfx_unflushed;
-};
-
 /*
  * shader binary helpers.
  */
@@ -305,191 +292,6 @@ void si_postflush_resume_features(struct r600_common_context *ctx)
 		si_resume_queries(ctx);
 }
 
-static void r600_add_fence_dependency(struct r600_common_context *rctx,
-				      struct pipe_fence_handle *fence)
-{
-	struct radeon_winsys *ws = rctx->ws;
-
-	if (rctx->dma.cs)
-		ws->cs_add_fence_dependency(rctx->dma.cs, fence);
-	ws->cs_add_fence_dependency(rctx->gfx.cs, fence);
-}
-
-static void r600_fence_server_sync(struct pipe_context *ctx,
-				   struct pipe_fence_handle *fence)
-{
-	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
-	struct r600_multi_fence *rfence = (struct r600_multi_fence *)fence;
-
-	/* Only amdgpu needs to handle fence dependencies (for fence imports).
-	 * radeon synchronizes all rings by default and will not implement
-	 * fence imports.
-	 */
-	if (rctx->screen->info.drm_major == 2)
-		return;
-
-	/* Only imported fences need to be handled by fence_server_sync,
-	 * because the winsys handles synchronizations automatically for BOs
-	 * within the process.
-	 *
-	 * Simply skip unflushed fences here, and the winsys will drop no-op
-	 * dependencies (i.e. dependencies within the same ring).
-	 */
-	if (rfence->gfx_unflushed.ctx)
-		return;
-
-	/* All unflushed commands will not start execution before
-	 * this fence dependency is signalled.
-	 *
-	 * Should we flush the context to allow more GPU parallelism?
-	 */
-	if (rfence->sdma)
-		r600_add_fence_dependency(rctx, rfence->sdma);
-	if (rfence->gfx)
-		r600_add_fence_dependency(rctx, rfence->gfx);
-}
-
-static void r600_create_fence_fd(struct pipe_context *ctx,
-				 struct pipe_fence_handle **pfence, int fd)
-{
-	struct r600_common_screen *rscreen = (struct r600_common_screen*)ctx->screen;
-	struct radeon_winsys *ws = rscreen->ws;
-	struct r600_multi_fence *rfence;
-
-	*pfence = NULL;
-
-	if (!rscreen->info.has_sync_file)
-		return;
-
-	rfence = CALLOC_STRUCT(r600_multi_fence);
-	if (!rfence)
-		return;
-
-	pipe_reference_init(&rfence->reference, 1);
-	rfence->gfx = ws->fence_import_sync_file(ws, fd);
-	if (!rfence->gfx) {
-		FREE(rfence);
-		return;
-	}
-
-	*pfence = (struct pipe_fence_handle*)rfence;
-}
-
-static int r600_fence_get_fd(struct pipe_screen *screen,
-			     struct pipe_fence_handle *fence)
-{
-	struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
-	struct radeon_winsys *ws = rscreen->ws;
-	struct r600_multi_fence *rfence = (struct r600_multi_fence *)fence;
-	int gfx_fd = -1, sdma_fd = -1;
-
-	if (!rscreen->info.has_sync_file)
-		return -1;
-
-	/* Deferred fences aren't supported. */
-	assert(!rfence->gfx_unflushed.ctx);
-	if (rfence->gfx_unflushed.ctx)
-		return -1;
-
-	if (rfence->sdma) {
-		sdma_fd = ws->fence_export_sync_file(ws, rfence->sdma);
-		if (sdma_fd == -1)
-			return -1;
-	}
-	if (rfence->gfx) {
-		gfx_fd = ws->fence_export_sync_file(ws, rfence->gfx);
-		if (gfx_fd == -1) {
-			if (sdma_fd != -1)
-				close(sdma_fd);
-			return -1;
-		}
-	}
-
-	/* If we don't have FDs at this point, it means we don't have fences
-	 * either. */
-	if (sdma_fd == -1)
-		return gfx_fd;
-	if (gfx_fd == -1)
-		return sdma_fd;
-
-	/* Get a fence that will be a combination of both fences. */
-	sync_accumulate("radeonsi", &gfx_fd, sdma_fd);
-	close(sdma_fd);
-	return gfx_fd;
-}
-
-static void r600_flush_from_st(struct pipe_context *ctx,
-			       struct pipe_fence_handle **fence,
-			       unsigned flags)
-{
-	struct pipe_screen *screen = ctx->screen;
-	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
-	struct radeon_winsys *ws = rctx->ws;
-	struct pipe_fence_handle *gfx_fence = NULL;
-	struct pipe_fence_handle *sdma_fence = NULL;
-	bool deferred_fence = false;
-	unsigned rflags = RADEON_FLUSH_ASYNC;
-
-	if (flags & PIPE_FLUSH_END_OF_FRAME)
-		rflags |= RADEON_FLUSH_END_OF_FRAME;
-
-	/* DMA IBs are preambles to gfx IBs, therefore must be flushed first. */
-	if (rctx->dma.cs)
-		rctx->dma.flush(rctx, rflags, fence ? &sdma_fence : NULL);
-
-	if (!radeon_emitted(rctx->gfx.cs, rctx->initial_gfx_cs_size)) {
-		if (fence)
-			ws->fence_reference(&gfx_fence, rctx->last_gfx_fence);
-		if (!(flags & PIPE_FLUSH_DEFERRED))
-			ws->cs_sync_flush(rctx->gfx.cs);
-	} else {
-		/* Instead of flushing, create a deferred fence. Constraints:
-		 * - The state tracker must allow a deferred flush.
-		 * - The state tracker must request a fence.
-		 * - fence_get_fd is not allowed.
-		 * Thread safety in fence_finish must be ensured by the state tracker.
-		 */
-		if (flags & PIPE_FLUSH_DEFERRED &&
-		    !(flags & PIPE_FLUSH_FENCE_FD) &&
-		    fence) {
-			gfx_fence = rctx->ws->cs_get_next_fence(rctx->gfx.cs);
-			deferred_fence = true;
-		} else {
-			rctx->gfx.flush(rctx, rflags, fence ? &gfx_fence : NULL);
-		}
-	}
-
-	/* Both engines can signal out of order, so we need to keep both fences. */
-	if (fence) {
-		struct r600_multi_fence *multi_fence =
-			CALLOC_STRUCT(r600_multi_fence);
-		if (!multi_fence) {
-			ws->fence_reference(&sdma_fence, NULL);
-			ws->fence_reference(&gfx_fence, NULL);
-			goto finish;
-		}
-
-		multi_fence->reference.count = 1;
-		/* If both fences are NULL, fence_finish will always return true. */
-		multi_fence->gfx = gfx_fence;
-		multi_fence->sdma = sdma_fence;
-
-		if (deferred_fence) {
-			multi_fence->gfx_unflushed.ctx = rctx;
-			multi_fence->gfx_unflushed.ib_index = rctx->num_gfx_cs_flushes;
-		}
-
-		screen->fence_reference(screen, fence, NULL);
-		*fence = (struct pipe_fence_handle*)multi_fence;
-	}
-finish:
-	if (!(flags & PIPE_FLUSH_DEFERRED)) {
-		if (rctx->dma.cs)
-			ws->cs_sync_flush(rctx->dma.cs);
-		ws->cs_sync_flush(rctx->gfx.cs);
-	}
-}
-
 static void r600_flush_dma_ring(void *ctx, unsigned flags,
 				struct pipe_fence_handle **fence)
 {
@@ -681,9 +483,6 @@ bool si_common_context_init(struct r600_common_context *rctx,
 	rctx->b.transfer_unmap = u_transfer_unmap_vtbl;
 	rctx->b.texture_subdata = u_default_texture_subdata;
 	rctx->b.memory_barrier = r600_memory_barrier;
-	rctx->b.flush = r600_flush_from_st;
-	rctx->b.create_fence_fd = r600_create_fence_fd;
-	rctx->b.fence_server_sync = r600_fence_server_sync;
 	rctx->dma_clear_buffer = r600_dma_clear_buffer_fallback;
 	rctx->b.buffer_subdata = si_buffer_subdata;
 
@@ -1193,69 +992,6 @@ static uint64_t r600_get_timestamp(struct pipe_screen *screen)
 			rscreen->info.clock_crystal_freq;
 }
 
-static void r600_fence_reference(struct pipe_screen *screen,
-				 struct pipe_fence_handle **dst,
-				 struct pipe_fence_handle *src)
-{
-	struct radeon_winsys *ws = ((struct r600_common_screen*)screen)->ws;
-	struct r600_multi_fence **rdst = (struct r600_multi_fence **)dst;
-	struct r600_multi_fence *rsrc = (struct r600_multi_fence *)src;
-
-	if (pipe_reference(&(*rdst)->reference, &rsrc->reference)) {
-		ws->fence_reference(&(*rdst)->gfx, NULL);
-		ws->fence_reference(&(*rdst)->sdma, NULL);
-		FREE(*rdst);
-	}
-        *rdst = rsrc;
-}
-
-static boolean r600_fence_finish(struct pipe_screen *screen,
-				 struct pipe_context *ctx,
-				 struct pipe_fence_handle *fence,
-				 uint64_t timeout)
-{
-	struct radeon_winsys *rws = ((struct r600_common_screen*)screen)->ws;
-	struct r600_multi_fence *rfence = (struct r600_multi_fence *)fence;
-	struct r600_common_context *rctx;
-	int64_t abs_timeout = os_time_get_absolute_timeout(timeout);
-
-	ctx = threaded_context_unwrap_sync(ctx);
-	rctx = ctx ? (struct r600_common_context*)ctx : NULL;
-
-	if (rfence->sdma) {
-		if (!rws->fence_wait(rws, rfence->sdma, timeout))
-			return false;
-
-		/* Recompute the timeout after waiting. */
-		if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
-			int64_t time = os_time_get_nano();
-			timeout = abs_timeout > time ? abs_timeout - time : 0;
-		}
-	}
-
-	if (!rfence->gfx)
-		return true;
-
-	/* Flush the gfx IB if it hasn't been flushed yet. */
-	if (rctx &&
-	    rfence->gfx_unflushed.ctx == rctx &&
-	    rfence->gfx_unflushed.ib_index == rctx->num_gfx_cs_flushes) {
-		rctx->gfx.flush(rctx, timeout ? 0 : RADEON_FLUSH_ASYNC, NULL);
-		rfence->gfx_unflushed.ctx = NULL;
-
-		if (!timeout)
-			return false;
-
-		/* Recompute the timeout after all that. */
-		if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
-			int64_t time = os_time_get_nano();
-			timeout = abs_timeout > time ? abs_timeout - time : 0;
-		}
-	}
-
-	return rws->fence_wait(rws, rfence->gfx, timeout);
-}
-
 static void r600_query_memory_info(struct pipe_screen *screen,
 				   struct pipe_memory_info *info)
 {
@@ -1346,12 +1082,9 @@ bool si_common_screen_init(struct r600_common_screen *rscreen,
 	rscreen->b.get_compute_param = r600_get_compute_param;
 	rscreen->b.get_paramf = r600_get_paramf;
 	rscreen->b.get_timestamp = r600_get_timestamp;
-	rscreen->b.fence_finish = r600_fence_finish;
-	rscreen->b.fence_reference = r600_fence_reference;
 	rscreen->b.resource_destroy = u_resource_destroy_vtbl;
 	rscreen->b.resource_from_user_memory = si_buffer_from_user_memory;
 	rscreen->b.query_memory_info = r600_query_memory_info;
-	rscreen->b.fence_get_fd = r600_fence_get_fd;
 
 	if (rscreen->info.has_hw_decode) {
 		rscreen->b.get_video_param = si_vid_get_video_param;
diff --git a/src/gallium/drivers/radeonsi/Makefile.sources b/src/gallium/drivers/radeonsi/Makefile.sources
index 2d746ab4ba..8dd33f60a8 100644
--- a/src/gallium/drivers/radeonsi/Makefile.sources
+++ b/src/gallium/drivers/radeonsi/Makefile.sources
@@ -12,6 +12,7 @@ C_SOURCES := \
 	si_debug.c \
 	si_descriptors.c \
 	si_dma.c \
+	si_fence.c \
 	si_hw_context.c \
 	si_pipe.c \
 	si_pipe.h \
diff --git a/src/gallium/drivers/radeonsi/meson.build b/src/gallium/drivers/radeonsi/meson.build
index b0089c0d56..4392184dbb 100644
--- a/src/gallium/drivers/radeonsi/meson.build
+++ b/src/gallium/drivers/radeonsi/meson.build
@@ -28,6 +28,7 @@ files_libradeonsi = files(
   'si_debug.c',
   'si_descriptors.c',
   'si_dma.c',
+  'si_fence.c',
   'si_hw_context.c',
   'si_pipe.c',
   'si_pipe.h',
diff --git a/src/gallium/drivers/radeonsi/si_fence.c b/src/gallium/drivers/radeonsi/si_fence.c
new file mode 100644
index 0000000000..b416c47aa3
--- /dev/null
+++ b/src/gallium/drivers/radeonsi/si_fence.c
@@ -0,0 +1,304 @@
+/*
+ * Copyright 2013-2017 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 <libsync.h>
+
+#include "util/os_time.h"
+#include "util/u_memory.h"
+
+#include "si_pipe.h"
+
+struct si_multi_fence {
+	struct pipe_reference reference;
+	struct pipe_fence_handle *gfx;
+	struct pipe_fence_handle *sdma;
+
+	/* If the context wasn't flushed at fence creation, this is non-NULL. */
+	struct {
+		struct r600_common_context *ctx;
+		unsigned ib_index;
+	} gfx_unflushed;
+};
+
+static void si_add_fence_dependency(struct r600_common_context *rctx,
+				    struct pipe_fence_handle *fence)
+{
+	struct radeon_winsys *ws = rctx->ws;
+
+	if (rctx->dma.cs)
+		ws->cs_add_fence_dependency(rctx->dma.cs, fence);
+	ws->cs_add_fence_dependency(rctx->gfx.cs, fence);
+}
+
+static void si_fence_reference(struct pipe_screen *screen,
+			       struct pipe_fence_handle **dst,
+			       struct pipe_fence_handle *src)
+{
+	struct radeon_winsys *ws = ((struct r600_common_screen*)screen)->ws;
+	struct si_multi_fence **rdst = (struct si_multi_fence **)dst;
+	struct si_multi_fence *rsrc = (struct si_multi_fence *)src;
+
+	if (pipe_reference(&(*rdst)->reference, &rsrc->reference)) {
+		ws->fence_reference(&(*rdst)->gfx, NULL);
+		ws->fence_reference(&(*rdst)->sdma, NULL);
+		FREE(*rdst);
+	}
+        *rdst = rsrc;
+}
+
+static void si_fence_server_sync(struct pipe_context *ctx,
+				 struct pipe_fence_handle *fence)
+{
+	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
+	struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+
+	/* Only amdgpu needs to handle fence dependencies (for fence imports).
+	 * radeon synchronizes all rings by default and will not implement
+	 * fence imports.
+	 */
+	if (rctx->screen->info.drm_major == 2)
+		return;
+
+	/* Only imported fences need to be handled by fence_server_sync,
+	 * because the winsys handles synchronizations automatically for BOs
+	 * within the process.
+	 *
+	 * Simply skip unflushed fences here, and the winsys will drop no-op
+	 * dependencies (i.e. dependencies within the same ring).
+	 */
+	if (rfence->gfx_unflushed.ctx)
+		return;
+
+	/* All unflushed commands will not start execution before
+	 * this fence dependency is signalled.
+	 *
+	 * Should we flush the context to allow more GPU parallelism?
+	 */
+	if (rfence->sdma)
+		si_add_fence_dependency(rctx, rfence->sdma);
+	if (rfence->gfx)
+		si_add_fence_dependency(rctx, rfence->gfx);
+}
+
+static boolean si_fence_finish(struct pipe_screen *screen,
+			       struct pipe_context *ctx,
+			       struct pipe_fence_handle *fence,
+			       uint64_t timeout)
+{
+	struct radeon_winsys *rws = ((struct r600_common_screen*)screen)->ws;
+	struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+	struct r600_common_context *rctx;
+	int64_t abs_timeout = os_time_get_absolute_timeout(timeout);
+
+	ctx = threaded_context_unwrap_sync(ctx);
+	rctx = ctx ? (struct r600_common_context*)ctx : NULL;
+
+	if (rfence->sdma) {
+		if (!rws->fence_wait(rws, rfence->sdma, timeout))
+			return false;
+
+		/* Recompute the timeout after waiting. */
+		if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
+			int64_t time = os_time_get_nano();
+			timeout = abs_timeout > time ? abs_timeout - time : 0;
+		}
+	}
+
+	if (!rfence->gfx)
+		return true;
+
+	/* Flush the gfx IB if it hasn't been flushed yet. */
+	if (rctx &&
+	    rfence->gfx_unflushed.ctx == rctx &&
+	    rfence->gfx_unflushed.ib_index == rctx->num_gfx_cs_flushes) {
+		rctx->gfx.flush(rctx, timeout ? 0 : RADEON_FLUSH_ASYNC, NULL);
+		rfence->gfx_unflushed.ctx = NULL;
+
+		if (!timeout)
+			return false;
+
+		/* Recompute the timeout after all that. */
+		if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
+			int64_t time = os_time_get_nano();
+			timeout = abs_timeout > time ? abs_timeout - time : 0;
+		}
+	}
+
+	return rws->fence_wait(rws, rfence->gfx, timeout);
+}
+
+static void si_create_fence_fd(struct pipe_context *ctx,
+			       struct pipe_fence_handle **pfence, int fd)
+{
+	struct r600_common_screen *rscreen = (struct r600_common_screen*)ctx->screen;
+	struct radeon_winsys *ws = rscreen->ws;
+	struct si_multi_fence *rfence;
+
+	*pfence = NULL;
+
+	if (!rscreen->info.has_sync_file)
+		return;
+
+	rfence = CALLOC_STRUCT(si_multi_fence);
+	if (!rfence)
+		return;
+
+	pipe_reference_init(&rfence->reference, 1);
+	rfence->gfx = ws->fence_import_sync_file(ws, fd);
+	if (!rfence->gfx) {
+		FREE(rfence);
+		return;
+	}
+
+	*pfence = (struct pipe_fence_handle*)rfence;
+}
+
+static int si_fence_get_fd(struct pipe_screen *screen,
+			   struct pipe_fence_handle *fence)
+{
+	struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
+	struct radeon_winsys *ws = rscreen->ws;
+	struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+	int gfx_fd = -1, sdma_fd = -1;
+
+	if (!rscreen->info.has_sync_file)
+		return -1;
+
+	/* Deferred fences aren't supported. */
+	assert(!rfence->gfx_unflushed.ctx);
+	if (rfence->gfx_unflushed.ctx)
+		return -1;
+
+	if (rfence->sdma) {
+		sdma_fd = ws->fence_export_sync_file(ws, rfence->sdma);
+		if (sdma_fd == -1)
+			return -1;
+	}
+	if (rfence->gfx) {
+		gfx_fd = ws->fence_export_sync_file(ws, rfence->gfx);
+		if (gfx_fd == -1) {
+			if (sdma_fd != -1)
+				close(sdma_fd);
+			return -1;
+		}
+	}
+
+	/* If we don't have FDs at this point, it means we don't have fences
+	 * either. */
+	if (sdma_fd == -1)
+		return gfx_fd;
+	if (gfx_fd == -1)
+		return sdma_fd;
+
+	/* Get a fence that will be a combination of both fences. */
+	sync_accumulate("radeonsi", &gfx_fd, sdma_fd);
+	close(sdma_fd);
+	return gfx_fd;
+}
+
+static void si_flush_from_st(struct pipe_context *ctx,
+			     struct pipe_fence_handle **fence,
+			     unsigned flags)
+{
+	struct pipe_screen *screen = ctx->screen;
+	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
+	struct radeon_winsys *ws = rctx->ws;
+	struct pipe_fence_handle *gfx_fence = NULL;
+	struct pipe_fence_handle *sdma_fence = NULL;
+	bool deferred_fence = false;
+	unsigned rflags = RADEON_FLUSH_ASYNC;
+
+	if (flags & PIPE_FLUSH_END_OF_FRAME)
+		rflags |= RADEON_FLUSH_END_OF_FRAME;
+
+	/* DMA IBs are preambles to gfx IBs, therefore must be flushed first. */
+	if (rctx->dma.cs)
+		rctx->dma.flush(rctx, rflags, fence ? &sdma_fence : NULL);
+
+	if (!radeon_emitted(rctx->gfx.cs, rctx->initial_gfx_cs_size)) {
+		if (fence)
+			ws->fence_reference(&gfx_fence, rctx->last_gfx_fence);
+		if (!(flags & PIPE_FLUSH_DEFERRED))
+			ws->cs_sync_flush(rctx->gfx.cs);
+	} else {
+		/* Instead of flushing, create a deferred fence. Constraints:
+		 * - The state tracker must allow a deferred flush.
+		 * - The state tracker must request a fence.
+		 * - fence_get_fd is not allowed.
+		 * Thread safety in fence_finish must be ensured by the state tracker.
+		 */
+		if (flags & PIPE_FLUSH_DEFERRED &&
+		    !(flags & PIPE_FLUSH_FENCE_FD) &&
+		    fence) {
+			gfx_fence = rctx->ws->cs_get_next_fence(rctx->gfx.cs);
+			deferred_fence = true;
+		} else {
+			rctx->gfx.flush(rctx, rflags, fence ? &gfx_fence : NULL);
+		}
+	}
+
+	/* Both engines can signal out of order, so we need to keep both fences. */
+	if (fence) {
+		struct si_multi_fence *multi_fence =
+				CALLOC_STRUCT(si_multi_fence);
+		if (!multi_fence) {
+			ws->fence_reference(&sdma_fence, NULL);
+			ws->fence_reference(&gfx_fence, NULL);
+			goto finish;
+		}
+
+		multi_fence->reference.count = 1;
+		/* If both fences are NULL, fence_finish will always return true. */
+		multi_fence->gfx = gfx_fence;
+		multi_fence->sdma = sdma_fence;
+
+		if (deferred_fence) {
+			multi_fence->gfx_unflushed.ctx = rctx;
+			multi_fence->gfx_unflushed.ib_index = rctx->num_gfx_cs_flushes;
+		}
+
+		screen->fence_reference(screen, fence, NULL);
+		*fence = (struct pipe_fence_handle*)multi_fence;
+	}
+finish:
+	if (!(flags & PIPE_FLUSH_DEFERRED)) {
+		if (rctx->dma.cs)
+			ws->cs_sync_flush(rctx->dma.cs);
+		ws->cs_sync_flush(rctx->gfx.cs);
+	}
+}
+
+void si_init_fence_functions(struct si_context *ctx)
+{
+	ctx->b.b.flush = si_flush_from_st;
+	ctx->b.b.create_fence_fd = si_create_fence_fd;
+	ctx->b.b.fence_server_sync = si_fence_server_sync;
+}
+
+void si_init_screen_fence_functions(struct si_screen *screen)
+{
+	screen->b.b.fence_finish = si_fence_finish;
+	screen->b.b.fence_reference = si_fence_reference;
+	screen->b.b.fence_get_fd = si_fence_get_fd;
+}
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 4b7c002837..8d7fb52350 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -256,6 +256,7 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen,
 		goto fail;
 
 	si_init_all_descriptors(sctx);
+	si_init_fence_functions(sctx);
 	si_init_state_functions(sctx);
 	si_init_shader_functions(sctx);
 	si_init_viewport_functions(sctx);
@@ -1006,6 +1007,7 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
 	sscreen->b.b.get_driver_uuid = radeonsi_get_driver_uuid;
 	sscreen->b.b.resource_create = si_resource_create_common;
 
+	si_init_screen_fence_functions(sscreen);
 	si_init_screen_state_functions(sscreen);
 
 	/* Set these flags in debug_flags early, so that the shader cache takes
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 8e10e89a5c..5253dbc43e 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -600,6 +600,10 @@ bool si_replace_shader(unsigned num, struct ac_shader_binary *binary);
 /* si_dma.c */
 void si_init_dma_functions(struct si_context *sctx);
 
+/* si_fence.c */
+void si_init_fence_functions(struct si_context *ctx);
+void si_init_screen_fence_functions(struct si_screen *screen);
+
 /* si_hw_context.c */
 void si_destroy_saved_cs(struct si_saved_cs *scs);
 void si_context_gfx_flush(void *context, unsigned flags,