[Mesa-dev] [PATCH 3/6] radeonsi: align vertex buffer descriptor list size for optimal prefetch
Marek Olšák
maraeo at gmail.com
Thu Feb 9 11:21:46 UTC 2017
From: Marek Olšák <marek.olsak at amd.com>
---
src/gallium/drivers/radeonsi/si_cp_dma.c | 2 +-
src/gallium/drivers/radeonsi/si_descriptors.c | 3 ++-
src/gallium/drivers/radeonsi/si_state.c | 2 ++
src/gallium/drivers/radeonsi/si_state.h | 2 ++
4 files changed, 7 insertions(+), 2 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_cp_dma.c b/src/gallium/drivers/radeonsi/si_cp_dma.c
index 9fa3ccb..ea999d9 100644
--- a/src/gallium/drivers/radeonsi/si_cp_dma.c
+++ b/src/gallium/drivers/radeonsi/si_cp_dma.c
@@ -427,21 +427,21 @@ static void cik_emit_prefetch_L2(struct si_context *sctx, struct r600_atom *atom
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_buffer_pointer_dirty) {
cik_prefetch_TC_L2_async(sctx, &sctx->vertex_buffers.buffer->b.b,
sctx->vertex_buffers.buffer_offset,
- sctx->vertex_elements->count * 16);
+ sctx->vertex_elements->desc_list_byte_size);
}
if (si_pm4_state_changed(sctx, ps))
cik_prefetch_shader_async(sctx, sctx->queued.named.ps);
}
void si_init_cp_dma_functions(struct si_context *sctx)
{
sctx->b.clear_buffer = si_clear_buffer;
si_init_atom(sctx, &sctx->prefetch_L2, &sctx->atoms.s.prefetch_L2,
diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index 9acc423..b0faf42 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -954,21 +954,22 @@ bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
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,
+ u_upload_alloc(sctx->b.uploader, 0, velems->desc_list_byte_size, 256,
+ &desc->buffer_offset,
(struct pipe_resource**)&desc->buffer, (void**)&ptr);
if (!desc->buffer)
return false;
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
desc->buffer, RADEON_USAGE_READ,
RADEON_PRIO_DESCRIPTORS);
assert(count <= SI_NUM_VERTEX_BUFFERS);
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 5a163b1..1e0729c 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -3344,20 +3344,22 @@ static void *si_create_vertex_elements(struct pipe_context *ctx,
{
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;
+ v->desc_list_byte_size = align(count * 16, SI_CPDMA_ALIGNMENT);
+
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;
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 3f08f54..07b7d58 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -93,20 +93,22 @@ 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;
+ /* Vertex buffer descriptor list size aligned for optimal prefetch. */
+ unsigned desc_list_byte_size;
/* 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
More information about the mesa-dev
mailing list