3 files changed, 50 insertions, 1 deletions

diff --git a/src/gallium/drivers/radeonsi/si_cp_dma.c b/src/gallium/drivers/radeonsi/si_cp_dma.c
index 934fd8df2db..f06b8ddb79b 100644
--- a/src/gallium/drivers/radeonsi/si_cp_dma.c
+++ b/src/gallium/drivers/radeonsi/si_cp_dma.c
@@ -386,6 +386,18 @@ void si_copy_buffer(struct si_context *sctx,
 		sctx->b.num_cp_dma_calls++;
 }
 
+void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf,
+			      uint64_t offset, unsigned size)
+{
+	assert(sctx->b.chip_class >= CIK);
+
+	si_copy_buffer(sctx, buf, buf, offset, offset, size,
+		       SI_CPDMA_SKIP_CHECK_CS_SPACE |
+		       SI_CPDMA_SKIP_SYNC_AFTER |
+		       SI_CPDMA_SKIP_SYNC_BEFORE |
+		       SI_CPDMA_SKIP_GFX_SYNC);
+}
+
 void si_init_cp_dma_functions(struct si_context *sctx)
 {
 	sctx->b.clear_buffer = si_clear_buffer;
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index dc37c8d28f3..c0a4636cc63 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -381,6 +381,8 @@ void si_copy_buffer(struct si_context *sctx,
 		    struct pipe_resource *dst, struct pipe_resource *src,
 		    uint64_t dst_offset, uint64_t src_offset, unsigned size,
 		    unsigned user_flags);
+void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf,
+			      uint64_t offset, unsigned size);
 void si_init_cp_dma_functions(struct si_context *sctx);
 
 /* si_debug.c */
diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index b3f664eff2f..7b756025f81 100644
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -937,6 +937,17 @@ void si_ce_post_draw_synchronization(struct si_context *sctx)
 	}
 }
 
+static void cik_prefetch_shader_async(struct si_context *sctx,
+				      struct si_pm4_state *state)
+{
+	if (state) {
+		struct pipe_resource *bo = &state->bo[0]->b.b;
+		assert(state->nbo == 1);
+
+		cik_prefetch_TC_L2_async(sctx, bo, 0, bo->width0);
+	}
+}
+
 void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 {
 	struct si_context *sctx = (struct si_context *)ctx;
@@ -1114,10 +1125,34 @@ void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 	if (!si_upload_vertex_buffer_descriptors(sctx))
 		return;
 
-	/* Flushed caches prior to emitting states. */
+	/* Flushed caches prior to prefetching shaders. */
 	if (sctx->b.flags)
 		si_emit_cache_flush(sctx);
 
+	/* Prefetch shaders and VBO descriptors to TC L2. */
+	if (sctx->b.chip_class >= CIK) {
+		if (si_pm4_state_changed(sctx, ls))
+			cik_prefetch_shader_async(sctx, sctx->queued.named.ls);
+		if (si_pm4_state_changed(sctx, hs))
+			cik_prefetch_shader_async(sctx, sctx->queued.named.hs);
+		if (si_pm4_state_changed(sctx, es))
+			cik_prefetch_shader_async(sctx, sctx->queued.named.es);
+		if (si_pm4_state_changed(sctx, gs))
+			cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
+		if (si_pm4_state_changed(sctx, vs))
+			cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
+
+		/* Vertex buffer descriptors are uploaded uncached, so prefetch
+		 * them right after the VS binary. */
+		if (sctx->vertex_buffers.pointer_dirty) {
+			cik_prefetch_TC_L2_async(sctx, &sctx->vertex_buffers.buffer->b.b,
+						sctx->vertex_buffers.buffer_offset,
+						sctx->vertex_elements->count * 16);
+		}
+		if (si_pm4_state_changed(sctx, ps))
+			cik_prefetch_shader_async(sctx, sctx->queued.named.ps);
+	}
+
 	/* Emit states. */
 	mask = sctx->dirty_atoms;
 	while (mask) {