[Mesa-dev] [PATCH 11/11] radeonsi: determine in advance which VBOs should be added to the buffer list

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

---
 src/gallium/drivers/radeonsi/si_descriptors.c | 8 ++++----
 src/gallium/drivers/radeonsi/si_state.c       | 6 ++++++
 src/gallium/drivers/radeonsi/si_state.h       | 1 +
 3 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index 837f393..057f374 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -932,29 +932,29 @@ static void si_vertex_buffers_begin_new_cs(struct si_context *sctx)
 		return;
 	radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
 			      desc->buffer, RADEON_USAGE_READ,
 			      RADEON_PRIO_DESCRIPTORS);
 }
 
 bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
 {
 	struct si_vertex_element *velems = sctx->vertex_elements;
 	struct si_descriptors *desc = &sctx->vertex_buffers;
-	bool bound[SI_NUM_VERTEX_BUFFERS] = {};
 	unsigned i, count = velems->count;
 	uint64_t va;
 	uint32_t *ptr;
 
 	if (!sctx->vertex_buffers_dirty || !count || !velems)
 		return true;
 
 	unsigned fix_size3 = velems->fix_size3;
+	unsigned first_vb_use_mask = velems->first_vb_use_mask;
 
 	/* Vertex buffer descriptors are the only ones which are uploaded
 	 * directly through a staging buffer and don't go through
 	 * the fine-grained upload path.
 	 */
 	u_upload_alloc(sctx->b.uploader, 0, count * 16, 256, &desc->buffer_offset,
 		       (struct pipe_resource**)&desc->buffer, (void**)&ptr);
 	if (!desc->buffer)
 		return false;
 
@@ -962,23 +962,24 @@ bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
 			      desc->buffer, RADEON_USAGE_READ,
 			      RADEON_PRIO_DESCRIPTORS);
 
 	assert(count <= SI_NUM_VERTEX_BUFFERS);
 
 	for (i = 0; i < count; i++) {
 		struct pipe_vertex_element *ve = &velems->elements[i];
 		struct pipe_vertex_buffer *vb;
 		struct r600_resource *rbuffer;
 		unsigned offset;
+		unsigned vbo_index = ve->vertex_buffer_index;
 		uint32_t *desc = &ptr[i*4];
 
-		vb = &sctx->vertex_buffer[ve->vertex_buffer_index];
+		vb = &sctx->vertex_buffer[vbo_index];
 		rbuffer = (struct r600_resource*)vb->buffer;
 		if (!rbuffer) {
 			memset(desc, 0, 16);
 			continue;
 		}
 
 		offset = vb->buffer_offset + ve->src_offset;
 		va = rbuffer->gpu_address + offset;
 
 		/* Fill in T# buffer resource description */
@@ -1011,25 +1012,24 @@ bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
 			size3 = (fix_size3 >> (2 * i)) & 3;
 			if (vb->stride && size3) {
 				assert(offset % 4 == 0 && vb->stride % 4 == 0);
 				assert(size3 <= 2);
 				desc[2] = align(desc[2], size3 * 2);
 			}
 		}
 
 		desc[3] = velems->rsrc_word3[i];
 
-		if (!bound[ve->vertex_buffer_index]) {
+		if (first_vb_use_mask & (1 << vbo_index)) {
 			radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
 					      (struct r600_resource*)vb->buffer,
 					      RADEON_USAGE_READ, RADEON_PRIO_VERTEX_BUFFER);
-			bound[ve->vertex_buffer_index] = true;
 		}
 	}
 
 	/* Don't flush the const cache. It would have a very negative effect
 	 * on performance (confirmed by testing). New descriptors are always
 	 * uploaded to a fresh new buffer, so I don't think flushing the const
 	 * cache is needed. */
 	si_mark_atom_dirty(sctx, &sctx->shader_userdata.atom);
 	sctx->vertex_buffers_dirty = false;
 	sctx->vertex_buffer_pointer_dirty = true;
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 3022260..568e169 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -3316,39 +3316,45 @@ static void si_delete_sampler_state(struct pipe_context *ctx, void *state)
 
 /*
  * Vertex elements & buffers
  */
 
 static void *si_create_vertex_elements(struct pipe_context *ctx,
 				       unsigned count,
 				       const struct pipe_vertex_element *elements)
 {
 	struct si_vertex_element *v = CALLOC_STRUCT(si_vertex_element);
+	bool used[SI_NUM_VERTEX_BUFFERS] = {};
 	int i;
 
 	assert(count <= SI_MAX_ATTRIBS);
 	if (!v)
 		return NULL;
 
 	v->count = count;
 	for (i = 0; i < count; ++i) {
 		const struct util_format_description *desc;
 		const struct util_format_channel_description *channel;
 		unsigned data_format, num_format;
 		int first_non_void;
 		unsigned vbo_index = elements[i].vertex_buffer_index;
 
 		if (vbo_index >= SI_NUM_VERTEX_BUFFERS) {
 			FREE(v);
 			return NULL;
 		}
 
+		if (!used[vbo_index]) {
+			v->first_vb_use_mask |= 1 << vbo_index;
+			used[vbo_index] = true;
+		}
+
 		desc = util_format_description(elements[i].src_format);
 		first_non_void = util_format_get_first_non_void_channel(elements[i].src_format);
 		data_format = si_translate_buffer_dataformat(ctx->screen, desc, first_non_void);
 		num_format = si_translate_buffer_numformat(ctx->screen, desc, first_non_void);
 		channel = &desc->channel[first_non_void];
 
 		v->rsrc_word3[i] = S_008F0C_DST_SEL_X(si_map_swizzle(desc->swizzle[0])) |
 				   S_008F0C_DST_SEL_Y(si_map_swizzle(desc->swizzle[1])) |
 				   S_008F0C_DST_SEL_Z(si_map_swizzle(desc->swizzle[2])) |
 				   S_008F0C_DST_SEL_W(si_map_swizzle(desc->swizzle[3])) |
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 34a0f57..03e5011 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -92,20 +92,21 @@ struct si_state_dsa {
 
 struct si_stencil_ref {
 	struct r600_atom		atom;
 	struct pipe_stencil_ref		state;
 	struct si_dsa_stencil_ref_part	dsa_part;
 };
 
 struct si_vertex_element
 {
 	unsigned			count;
+	unsigned			first_vb_use_mask;
 
 	/* Two bits per attribute indicating the size of each vector component
 	 * in bytes if the size 3-workaround must be applied.
 	 */
 	uint32_t			fix_size3;
 	uint64_t			fix_fetch;
 
 	uint32_t			rsrc_word3[SI_MAX_ATTRIBS];
 	uint32_t			format_size[SI_MAX_ATTRIBS];
 	struct pipe_vertex_element	elements[SI_MAX_ATTRIBS];
-- 
2.7.4