radeonsi: upload constants into VRAM instead of GTT

This lowers lgkm wait cycles by 30% on VI and normal conditions.
The might be a measurable improvement when CE is disabled (radeon)
or under L2 thrashing.

Reviewed-by: Nicolai Hähnle <[email protected]>

4 files changed, 18 insertions, 10 deletions

diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index d573b39d7c0..1781584f5ff 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -607,7 +607,11 @@ bool r600_common_context_init(struct r600_common_context *rctx,
 						  0, PIPE_USAGE_STREAM);
 	if (!rctx->b.stream_uploader)
 		return false;
-	rctx->b.const_uploader = rctx->b.stream_uploader;
+
+	rctx->b.const_uploader = u_upload_create(&rctx->b, 128 * 1024,
+						 0, PIPE_USAGE_DEFAULT);
+	if (!rctx->b.const_uploader)
+		return false;
 
 	rctx->ctx = rctx->ws->ctx_create(rctx->ws);
 	if (!rctx->ctx)
@@ -649,9 +653,10 @@ void r600_common_context_cleanup(struct r600_common_context *rctx)
 	if (rctx->ctx)
 		rctx->ws->ctx_destroy(rctx->ctx);
 
-	if (rctx->b.stream_uploader) {
+	if (rctx->b.stream_uploader)
 		u_upload_destroy(rctx->b.stream_uploader);
-	}
+	if (rctx->b.const_uploader)
+		u_upload_destroy(rctx->b.const_uploader);
 
 	slab_destroy_child(&rctx->pool_transfers);
 
diff --git a/src/gallium/drivers/radeonsi/si_compute.c b/src/gallium/drivers/radeonsi/si_compute.c
index 381837c8a57..88d72c1ea2a 100644
--- a/src/gallium/drivers/radeonsi/si_compute.c
+++ b/src/gallium/drivers/radeonsi/si_compute.c
@@ -503,7 +503,7 @@ static void si_setup_user_sgprs_co_v2(struct si_context *sctx,
 
 		dispatch.kernarg_address = kernel_args_va;
 
-		u_upload_data(sctx->b.b.stream_uploader, 0, sizeof(dispatch),
+		u_upload_data(sctx->b.b.const_uploader, 0, sizeof(dispatch),
                               256, &dispatch, &dispatch_offset,
                               (struct pipe_resource**)&dispatch_buf);
 
@@ -565,7 +565,7 @@ static void si_upload_compute_input(struct si_context *sctx,
 	/* The extra num_work_size_bytes are for work group / work item size information */
 	kernel_args_size = program->input_size + num_work_size_bytes;
 
-	u_upload_alloc(sctx->b.b.stream_uploader, 0, kernel_args_size,
+	u_upload_alloc(sctx->b.b.const_uploader, 0, kernel_args_size,
 		       sctx->screen->b.info.tcc_cache_line_size,
 		       &kernel_args_offset,
 		       (struct pipe_resource**)&input_buffer, &kernel_args_ptr);
diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index b4f1fbfb213..a41b243eec8 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -235,7 +235,7 @@ static bool si_upload_descriptors(struct si_context *sctx,
 	} else {
 		void *ptr;
 
-		u_upload_alloc(sctx->b.b.stream_uploader, 0, list_size,
+		u_upload_alloc(sctx->b.b.const_uploader, 0, list_size,
 			       sctx->screen->b.info.tcc_cache_line_size,
 			       &desc->buffer_offset,
 			       (struct pipe_resource**)&desc->buffer, &ptr);
@@ -963,7 +963,7 @@ bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
 	 * directly through a staging buffer and don't go through
 	 * the fine-grained upload path.
 	 */
-	u_upload_alloc(sctx->b.b.stream_uploader, 0,
+	u_upload_alloc(sctx->b.b.const_uploader, 0,
 		       desc_list_byte_size,
 		       si_optimal_tcc_alignment(sctx, desc_list_byte_size),
 		       &desc->buffer_offset,
@@ -1051,7 +1051,7 @@ void si_upload_const_buffer(struct si_context *sctx, struct r600_resource **rbuf
 {
 	void *tmp;
 
-	u_upload_alloc(sctx->b.b.stream_uploader, 0, size,
+	u_upload_alloc(sctx->b.b.const_uploader, 0, size,
 		       si_optimal_tcc_alignment(sctx, size),
 		       const_offset,
 		       (struct pipe_resource**)rbuffer, &tmp);
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index f53f8dd8ee9..81592a7e858 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -3526,8 +3526,11 @@ static void si_set_vertex_buffers(struct pipe_context *ctx,
 				assert(src->stride == 0);
 
 				/* Assume the attrib has 4 dwords like the vbo
-				 * module. This is also a good upper bound. */
-				u_upload_data(sctx->b.b.stream_uploader, 0, 16, 16,
+				 * module. This is also a good upper bound.
+				 *
+				 * Use const_uploader to upload into VRAM directly.
+				 */
+				u_upload_data(sctx->b.b.const_uploader, 0, 16, 16,
 					      src->user_buffer,
 					      &dsti->buffer_offset,
 					      &dsti->buffer);