radeonsi: don't emit partial flushes for internal CS flushes only

Tested-by: Benedikt Schemmer <ben@besd.de>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

10 files changed, 32 insertions, 20 deletions

diff --git a/src/gallium/drivers/radeonsi/si_buffer.c b/src/gallium/drivers/radeonsi/si_buffer.c
index 1420702d8d4..d17b2c6a831 100644
--- a/src/gallium/drivers/radeonsi/si_buffer.c
+++ b/src/gallium/drivers/radeonsi/si_buffer.c
@@ -64,10 +64,10 @@ void *si_buffer_map_sync_with_rings(struct si_context *sctx,
 	    sctx->ws->cs_is_buffer_referenced(sctx->gfx_cs,
 						resource->buf, rusage)) {
 		if (usage & PIPE_TRANSFER_DONTBLOCK) {
-			si_flush_gfx_cs(sctx, PIPE_FLUSH_ASYNC, NULL);
+			si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 			return NULL;
 		} else {
-			si_flush_gfx_cs(sctx, 0, NULL);
+			si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 			busy = true;
 		}
 	}
@@ -725,7 +725,7 @@ static bool si_resource_commit(struct pipe_context *pctx,
 	if (radeon_emitted(ctx->gfx_cs, ctx->initial_gfx_cs_size) &&
 	    ctx->ws->cs_is_buffer_referenced(ctx->gfx_cs,
 					       res->buf, RADEON_USAGE_READWRITE)) {
-		si_flush_gfx_cs(ctx, PIPE_FLUSH_ASYNC, NULL);
+		si_flush_gfx_cs(ctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 	}
 	if (radeon_emitted(ctx->dma_cs, 0) &&
 	    ctx->ws->cs_is_buffer_referenced(ctx->dma_cs,
diff --git a/src/gallium/drivers/radeonsi/si_dma_cs.c b/src/gallium/drivers/radeonsi/si_dma_cs.c
index 7af7c5623b7..1eefaeb6ad5 100644
--- a/src/gallium/drivers/radeonsi/si_dma_cs.c
+++ b/src/gallium/drivers/radeonsi/si_dma_cs.c
@@ -58,7 +58,7 @@ void si_need_dma_space(struct si_context *ctx, unsigned num_dw,
 	     (src &&
 	      ctx->ws->cs_is_buffer_referenced(ctx->gfx_cs, src->buf,
 						 RADEON_USAGE_WRITE))))
-		si_flush_gfx_cs(ctx, PIPE_FLUSH_ASYNC, NULL);
+		si_flush_gfx_cs(ctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 
 	/* Flush if there's not enough space, or if the memory usage per IB
 	 * is too large.
diff --git a/src/gallium/drivers/radeonsi/si_fence.c b/src/gallium/drivers/radeonsi/si_fence.c
index 26d6c43b34d..19fcb96041f 100644
--- a/src/gallium/drivers/radeonsi/si_fence.c
+++ b/src/gallium/drivers/radeonsi/si_fence.c
@@ -374,7 +374,10 @@ static boolean si_fence_finish(struct pipe_screen *screen,
 			 * not going to wait.
 			 */
 			threaded_context_unwrap_sync(ctx);
-			si_flush_gfx_cs(sctx, timeout ? 0 : PIPE_FLUSH_ASYNC, NULL);
+			si_flush_gfx_cs(sctx,
+					(timeout ? 0 : PIPE_FLUSH_ASYNC) |
+					 RADEON_FLUSH_START_NEXT_GFX_IB_NOW,
+					NULL);
 			rfence->gfx_unflushed.ctx = NULL;
 
 			if (!timeout)
diff --git a/src/gallium/drivers/radeonsi/si_gfx_cs.c b/src/gallium/drivers/radeonsi/si_gfx_cs.c
index 147433b69b6..ddfdb497364 100644
--- a/src/gallium/drivers/radeonsi/si_gfx_cs.c
+++ b/src/gallium/drivers/radeonsi/si_gfx_cs.c
@@ -47,7 +47,7 @@ void si_need_gfx_cs_space(struct si_context *ctx)
 						   ctx->vram, ctx->gtt))) {
 		ctx->gtt = 0;
 		ctx->vram = 0;
-		si_flush_gfx_cs(ctx, PIPE_FLUSH_ASYNC, NULL);
+		si_flush_gfx_cs(ctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 		return;
 	}
 	ctx->gtt = 0;
@@ -61,7 +61,7 @@ void si_need_gfx_cs_space(struct si_context *ctx)
 	 */
 	unsigned need_dwords = 2048 + ctx->num_cs_dw_queries_suspend;
 	if (!ctx->ws->cs_check_space(cs, need_dwords))
-		si_flush_gfx_cs(ctx, PIPE_FLUSH_ASYNC, NULL);
+		si_flush_gfx_cs(ctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 }
 
 void si_flush_gfx_cs(struct si_context *ctx, unsigned flags,
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 125b3a72bfb..351c9f4cd38 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -1340,7 +1340,7 @@ radeon_add_to_gfx_buffer_list_check_mem(struct si_context *sctx,
 	    !radeon_cs_memory_below_limit(sctx->screen, sctx->gfx_cs,
 					  sctx->vram + rbo->vram_usage,
 					  sctx->gtt + rbo->gart_usage))
-		si_flush_gfx_cs(sctx, PIPE_FLUSH_ASYNC, NULL);
+		si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 
 	radeon_add_to_buffer_list(sctx, sctx->gfx_cs, rbo, usage, priority);
 }
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 7e1660415f5..67ab75bbd2d 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -2774,7 +2774,7 @@ static bool si_update_gs_ring_buffers(struct si_context *sctx)
 
 	/* Flush the context to re-emit both init_config states. */
 	sctx->initial_gfx_cs_size = 0; /* force flush */
-	si_flush_gfx_cs(sctx, PIPE_FLUSH_ASYNC, NULL);
+	si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 
 	/* Set ring bindings. */
 	if (sctx->esgs_ring) {
@@ -3051,7 +3051,7 @@ static void si_init_tess_factor_ring(struct si_context *sctx)
 	 */
 	si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
 	sctx->initial_gfx_cs_size = 0; /* force flush */
-	si_flush_gfx_cs(sctx, PIPE_FLUSH_ASYNC, NULL);
+	si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 }
 
 /**
diff --git a/src/gallium/drivers/radeonsi/si_texture.c b/src/gallium/drivers/radeonsi/si_texture.c
index 17957f18a5f..b41a0d1b925 100644
--- a/src/gallium/drivers/radeonsi/si_texture.c
+++ b/src/gallium/drivers/radeonsi/si_texture.c
@@ -1869,7 +1869,7 @@ static void si_texture_transfer_unmap(struct pipe_context *ctx,
 	 * The result is that the kernel memory manager is never a bottleneck.
 	 */
 	if (sctx->num_alloc_tex_transfer_bytes > sctx->screen->info.gart_size / 4) {
-		si_flush_gfx_cs(sctx, PIPE_FLUSH_ASYNC, NULL);
+		si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 		sctx->num_alloc_tex_transfer_bytes = 0;
 	}
 
diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
index 22b5a73143d..9b6d6e83032 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
@@ -239,7 +239,8 @@ static void *amdgpu_bo_map(struct pb_buffer *buf,
              * Only check whether the buffer is being used for write. */
             if (cs && amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo,
                                                                RADEON_USAGE_WRITE)) {
-               cs->flush_cs(cs->flush_data, PIPE_FLUSH_ASYNC, NULL);
+               cs->flush_cs(cs->flush_data,
+			    RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
                return NULL;
             }
 
@@ -249,7 +250,8 @@ static void *amdgpu_bo_map(struct pb_buffer *buf,
             }
          } else {
             if (cs && amdgpu_bo_is_referenced_by_cs(cs, bo)) {
-               cs->flush_cs(cs->flush_data, PIPE_FLUSH_ASYNC, NULL);
+               cs->flush_cs(cs->flush_data,
+			    RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
                return NULL;
             }
 
@@ -272,7 +274,8 @@ static void *amdgpu_bo_map(struct pb_buffer *buf,
             if (cs) {
                if (amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo,
                                                             RADEON_USAGE_WRITE)) {
-                  cs->flush_cs(cs->flush_data, 0, NULL);
+                  cs->flush_cs(cs->flush_data,
+			       RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL);
                } else {
                   /* Try to avoid busy-waiting in amdgpu_bo_wait. */
                   if (p_atomic_read(&bo->num_active_ioctls))
@@ -286,7 +289,8 @@ static void *amdgpu_bo_map(struct pb_buffer *buf,
             /* Mapping for write. */
             if (cs) {
                if (amdgpu_bo_is_referenced_by_cs(cs, bo)) {
-                  cs->flush_cs(cs->flush_data, 0, NULL);
+                  cs->flush_cs(cs->flush_data,
+			       RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL);
                } else {
                   /* Try to avoid busy-waiting in amdgpu_bo_wait. */
                   if (p_atomic_read(&bo->num_active_ioctls))
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
index 1617a2fe32e..6652977e586 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
@@ -516,7 +516,8 @@ static void *radeon_bo_map(struct pb_buffer *buf,
                  *
                  * Only check whether the buffer is being used for write. */
                 if (cs && radeon_bo_is_referenced_by_cs_for_write(cs, bo)) {
-                    cs->flush_cs(cs->flush_data, PIPE_FLUSH_ASYNC, NULL);
+                    cs->flush_cs(cs->flush_data,
+				 RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
                     return NULL;
                 }
 
@@ -526,7 +527,8 @@ static void *radeon_bo_map(struct pb_buffer *buf,
                 }
             } else {
                 if (cs && radeon_bo_is_referenced_by_cs(cs, bo)) {
-                    cs->flush_cs(cs->flush_data, PIPE_FLUSH_ASYNC, NULL);
+                    cs->flush_cs(cs->flush_data,
+				 RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
                     return NULL;
                 }
 
@@ -547,7 +549,8 @@ static void *radeon_bo_map(struct pb_buffer *buf,
                  *
                  * Only check whether the buffer is being used for write. */
                 if (cs && radeon_bo_is_referenced_by_cs_for_write(cs, bo)) {
-                    cs->flush_cs(cs->flush_data, 0, NULL);
+                    cs->flush_cs(cs->flush_data,
+				 RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL);
                 }
                 radeon_bo_wait((struct pb_buffer*)bo, PIPE_TIMEOUT_INFINITE,
                                RADEON_USAGE_WRITE);
@@ -555,7 +558,8 @@ static void *radeon_bo_map(struct pb_buffer *buf,
                 /* Mapping for write. */
                 if (cs) {
                     if (radeon_bo_is_referenced_by_cs(cs, bo)) {
-                        cs->flush_cs(cs->flush_data, 0, NULL);
+                        cs->flush_cs(cs->flush_data,
+				     RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL);
                     } else {
                         /* Try to avoid busy-waiting in radeon_bo_wait. */
                         if (p_atomic_read(&bo->num_active_ioctls))
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_cs.c b/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
index a1975dff8df..9070464bec8 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
@@ -407,7 +407,8 @@ static bool radeon_drm_cs_validate(struct radeon_winsys_cs *rcs)
 
         /* Flush if there are any relocs. Clean up otherwise. */
         if (cs->csc->num_relocs) {
-            cs->flush_cs(cs->flush_data, PIPE_FLUSH_ASYNC, NULL);
+            cs->flush_cs(cs->flush_data,
+			 RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
         } else {
             radeon_cs_context_cleanup(cs->csc);
             cs->base.used_vram = 0;