[Mesa-dev] [PATCH 09/13] radeonsi: don't use struct si_descriptors for vertex buffer descriptors
Marek Olšák
maraeo at gmail.com
Sat Feb 17 19:43:24 UTC 2018
From: Marek Olšák <marek.olsak at amd.com>
VBO descriptor code will change a lot one day.
---
src/gallium/drivers/radeonsi/si_blit.c | 2 +-
src/gallium/drivers/radeonsi/si_cp_dma.c | 5 ++-
src/gallium/drivers/radeonsi/si_debug.c | 14 ++++++--
src/gallium/drivers/radeonsi/si_descriptors.c | 50 +++++++++++++--------------
src/gallium/drivers/radeonsi/si_hw_context.c | 2 +-
src/gallium/drivers/radeonsi/si_pipe.h | 6 +++-
6 files changed, 46 insertions(+), 33 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_blit.c b/src/gallium/drivers/radeonsi/si_blit.c
index 370ce04..f1c4f6d 100644
--- a/src/gallium/drivers/radeonsi/si_blit.c
+++ b/src/gallium/drivers/radeonsi/si_blit.c
@@ -79,21 +79,21 @@ void si_blitter_begin(struct pipe_context *ctx, enum si_blitter_op op)
void si_blitter_end(struct pipe_context *ctx)
{
struct si_context *sctx = (struct si_context *)ctx;
sctx->b.render_cond_force_off = false;
/* Restore shader pointers because the VS blit shader changed all
* non-global VS user SGPRs. */
sctx->shader_pointers_dirty |= SI_DESCS_SHADER_MASK(VERTEX);
- sctx->vertex_buffer_pointer_dirty = true;
+ sctx->vertex_buffer_pointer_dirty = sctx->vb_descriptors_buffer != NULL;
si_mark_atom_dirty(sctx, &sctx->shader_pointers.atom);
}
static unsigned u_max_sample(struct pipe_resource *r)
{
return r->nr_samples ? r->nr_samples - 1 : 0;
}
static unsigned
si_blit_dbcb_copy(struct si_context *sctx,
diff --git a/src/gallium/drivers/radeonsi/si_cp_dma.c b/src/gallium/drivers/radeonsi/si_cp_dma.c
index b665926..15bd305 100644
--- a/src/gallium/drivers/radeonsi/si_cp_dma.c
+++ b/src/gallium/drivers/radeonsi/si_cp_dma.c
@@ -509,23 +509,22 @@ static void cik_prefetch_shader_async(struct si_context *sctx,
assert(state->nbo == 1);
cik_prefetch_TC_L2_async(sctx, bo, 0, bo->width0);
}
static void cik_prefetch_VBO_descriptors(struct si_context *sctx)
{
if (!sctx->vertex_elements)
return;
- cik_prefetch_TC_L2_async(sctx, &sctx->vertex_buffers.buffer->b.b,
- sctx->vertex_buffers.gpu_address -
- sctx->vertex_buffers.buffer->gpu_address,
+ cik_prefetch_TC_L2_async(sctx, &sctx->vb_descriptors_buffer->b.b,
+ sctx->vb_descriptors_offset,
sctx->vertex_elements->desc_list_byte_size);
}
void cik_emit_prefetch_L2(struct si_context *sctx)
{
/* Prefetch shaders and VBO descriptors to TC L2. */
if (sctx->b.chip_class >= GFX9) {
/* Choose the right spot for the VBO prefetch. */
if (sctx->tes_shader.cso) {
if (sctx->prefetch_L2_mask & SI_PREFETCH_HS)
diff --git a/src/gallium/drivers/radeonsi/si_debug.c b/src/gallium/drivers/radeonsi/si_debug.c
index 385ce39..1f25f4e 100644
--- a/src/gallium/drivers/radeonsi/si_debug.c
+++ b/src/gallium/drivers/radeonsi/si_debug.c
@@ -733,24 +733,34 @@ static void si_dump_descriptors(struct si_context *sctx,
enabled_constbuf = sctx->const_and_shader_buffers[processor].enabled_mask >>
SI_NUM_SHADER_BUFFERS;
enabled_shaderbuf = sctx->const_and_shader_buffers[processor].enabled_mask &
u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS);
enabled_shaderbuf = util_bitreverse(enabled_shaderbuf) >>
(32 - SI_NUM_SHADER_BUFFERS);
enabled_samplers = sctx->samplers[processor].enabled_mask;
enabled_images = sctx->images[processor].enabled_mask;
}
- if (processor == PIPE_SHADER_VERTEX) {
+ if (processor == PIPE_SHADER_VERTEX &&
+ sctx->vb_descriptors_buffer &&
+ sctx->vb_descriptors_gpu_list &&
+ sctx->vertex_elements) {
assert(info); /* only CS may not have an info struct */
+ struct si_descriptors desc = {};
- si_dump_descriptor_list(sctx->screen, &sctx->vertex_buffers, name,
+ desc.buffer = sctx->vb_descriptors_buffer;
+ desc.list = sctx->vb_descriptors_gpu_list;
+ desc.gpu_list = sctx->vb_descriptors_gpu_list;
+ desc.element_dw_size = 4;
+ desc.num_active_slots = sctx->vertex_elements->desc_list_byte_size / 16;
+
+ si_dump_descriptor_list(sctx->screen, &desc, name,
" - Vertex buffer", 4, info->num_inputs,
si_identity, log);
}
si_dump_descriptor_list(sctx->screen,
&descs[SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS],
name, " - Constant buffer", 4,
util_last_bit(enabled_constbuf),
si_get_constbuf_slot, log);
si_dump_descriptor_list(sctx->screen,
diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index 76f2a3e..f6bc3cf 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -1040,84 +1040,82 @@ static void si_get_buffer_from_descriptors(struct si_buffer_resources *buffers,
assert(va >= res->gpu_address && va + *size <= res->gpu_address + res->bo_size);
*offset = va - res->gpu_address;
}
}
/* VERTEX BUFFERS */
static void si_vertex_buffers_begin_new_cs(struct si_context *sctx)
{
- struct si_descriptors *desc = &sctx->vertex_buffers;
int count = sctx->vertex_elements ? sctx->vertex_elements->count : 0;
int i;
for (i = 0; i < count; i++) {
int vb = sctx->vertex_elements->vertex_buffer_index[i];
if (vb >= ARRAY_SIZE(sctx->vertex_buffer))
continue;
if (!sctx->vertex_buffer[vb].buffer.resource)
continue;
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
(struct r600_resource*)sctx->vertex_buffer[vb].buffer.resource,
RADEON_USAGE_READ, RADEON_PRIO_VERTEX_BUFFER);
}
- if (!desc->buffer)
+ if (!sctx->vb_descriptors_buffer)
return;
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
- desc->buffer, RADEON_USAGE_READ,
- RADEON_PRIO_DESCRIPTORS);
+ sctx->vb_descriptors_buffer, RADEON_USAGE_READ,
+ RADEON_PRIO_DESCRIPTORS);
}
bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
{
struct si_vertex_elements *velems = sctx->vertex_elements;
- struct si_descriptors *desc = &sctx->vertex_buffers;
unsigned i, count;
unsigned desc_list_byte_size;
unsigned first_vb_use_mask;
uint32_t *ptr;
if (!sctx->vertex_buffers_dirty || !velems)
return true;
count = velems->count;
if (!count)
return true;
desc_list_byte_size = velems->desc_list_byte_size;
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.
*/
- unsigned buffer_offset = 0;
u_upload_alloc(sctx->b.b.const_uploader, 0,
desc_list_byte_size,
si_optimal_tcc_alignment(sctx, desc_list_byte_size),
- &buffer_offset,
- (struct pipe_resource**)&desc->buffer, (void**)&ptr);
- if (!desc->buffer) {
- desc->gpu_address = 0;
+ &sctx->vb_descriptors_offset,
+ (struct pipe_resource**)&sctx->vb_descriptors_buffer,
+ (void**)&ptr);
+ if (!sctx->vb_descriptors_buffer) {
+ sctx->vb_descriptors_offset = 0;
+ sctx->vb_descriptors_gpu_list = NULL;
return false;
}
- desc->gpu_address = desc->buffer->gpu_address + buffer_offset;
- desc->list = ptr;
+ sctx->vb_descriptors_gpu_list = ptr;
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
- desc->buffer, RADEON_USAGE_READ,
- RADEON_PRIO_DESCRIPTORS);
+ sctx->vb_descriptors_buffer, RADEON_USAGE_READ,
+ RADEON_PRIO_DESCRIPTORS);
assert(count <= SI_MAX_ATTRIBS);
for (i = 0; i < count; i++) {
struct pipe_vertex_buffer *vb;
struct r600_resource *rbuffer;
unsigned vbo_index = velems->vertex_buffer_index[i];
uint32_t *desc = &ptr[i*4];
vb = &sctx->vertex_buffer[vbo_index];
@@ -1980,29 +1978,29 @@ void si_update_all_texture_descriptors(struct si_context *sctx)
/* SHADER USER DATA */
static void si_mark_shader_pointers_dirty(struct si_context *sctx,
unsigned shader)
{
sctx->shader_pointers_dirty |=
u_bit_consecutive(SI_DESCS_FIRST_SHADER + shader * SI_NUM_SHADER_DESCS,
SI_NUM_SHADER_DESCS);
if (shader == PIPE_SHADER_VERTEX)
- sctx->vertex_buffer_pointer_dirty = sctx->vertex_buffers.buffer != NULL;
+ sctx->vertex_buffer_pointer_dirty = sctx->vb_descriptors_buffer != NULL;
si_mark_atom_dirty(sctx, &sctx->shader_pointers.atom);
}
static void si_shader_pointers_begin_new_cs(struct si_context *sctx)
{
sctx->shader_pointers_dirty = u_bit_consecutive(0, SI_NUM_DESCS);
- sctx->vertex_buffer_pointer_dirty = sctx->vertex_buffers.buffer != NULL;
+ sctx->vertex_buffer_pointer_dirty = sctx->vb_descriptors_buffer != NULL;
si_mark_atom_dirty(sctx, &sctx->shader_pointers.atom);
sctx->graphics_bindless_pointer_dirty = sctx->bindless_descriptors.buffer != NULL;
sctx->compute_bindless_pointer_dirty = sctx->bindless_descriptors.buffer != NULL;
}
/* Set a base register address for user data constants in the given shader.
* This assigns a mapping from PIPE_SHADER_* to SPI_SHADER_USER_DATA_*.
*/
static void si_set_user_data_base(struct si_context *sctx,
unsigned shader, uint32_t new_base)
@@ -2182,22 +2180,28 @@ void si_emit_graphics_shader_pointers(struct si_context *sctx,
si_emit_disjoint_shader_pointers(sctx, SI_DESCS_SHADER_MASK(TESS_CTRL),
sh_base[PIPE_SHADER_TESS_CTRL]);
si_emit_disjoint_shader_pointers(sctx, SI_DESCS_SHADER_MASK(GEOMETRY),
sh_base[PIPE_SHADER_GEOMETRY]);
}
sctx->shader_pointers_dirty &=
~u_bit_consecutive(SI_DESCS_RW_BUFFERS, SI_DESCS_FIRST_COMPUTE);
if (sctx->vertex_buffer_pointer_dirty) {
- si_emit_shader_pointer(sctx, &sctx->vertex_buffers,
- sh_base[PIPE_SHADER_VERTEX]);
+ struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
+ unsigned sh_offset = sh_base[PIPE_SHADER_VERTEX] +
+ SI_SGPR_VERTEX_BUFFERS * 4;
+
+ si_emit_shader_pointer_head(cs, sh_offset, 1);
+ si_emit_shader_pointer_body(sctx->screen, cs,
+ sctx->vb_descriptors_buffer->gpu_address +
+ sctx->vb_descriptors_offset);
sctx->vertex_buffer_pointer_dirty = false;
}
if (sctx->graphics_bindless_pointer_dirty) {
si_emit_global_shader_pointers(sctx,
&sctx->bindless_descriptors);
sctx->graphics_bindless_pointer_dirty = false;
}
}
@@ -2719,25 +2723,20 @@ void si_init_all_descriptors(struct si_context *sctx)
si_init_buffer_resources(&sctx->rw_buffers,
&sctx->descriptors[SI_DESCS_RW_BUFFERS],
SI_NUM_RW_BUFFERS, SI_SGPR_RW_BUFFERS,
/* The second set of usage/priority is used by
* const buffers in RW buffer slots. */
RADEON_USAGE_READWRITE, RADEON_USAGE_READ,
RADEON_PRIO_SHADER_RINGS, RADEON_PRIO_CONST_BUFFER);
sctx->descriptors[SI_DESCS_RW_BUFFERS].num_active_slots = SI_NUM_RW_BUFFERS;
- si_init_descriptors(&sctx->vertex_buffers, SI_SGPR_VERTEX_BUFFERS,
- 4, SI_NUM_VERTEX_BUFFERS);
- FREE(sctx->vertex_buffers.list); /* not used */
- sctx->vertex_buffers.list = NULL;
-
/* Initialize an array of 1024 bindless descriptors, when the limit is
* reached, just make it larger and re-upload the whole array.
*/
si_init_bindless_descriptors(sctx, &sctx->bindless_descriptors,
SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
1024);
sctx->descriptors_dirty = u_bit_consecutive(0, SI_NUM_DESCS);
/* Set pipe_context functions. */
@@ -2825,22 +2824,23 @@ void si_release_all_descriptors(struct si_context *sctx)
si_release_image_views(&sctx->images[i]);
}
si_release_buffer_resources(&sctx->rw_buffers,
&sctx->descriptors[SI_DESCS_RW_BUFFERS]);
for (i = 0; i < SI_NUM_VERTEX_BUFFERS; i++)
pipe_vertex_buffer_unreference(&sctx->vertex_buffer[i]);
for (i = 0; i < SI_NUM_DESCS; ++i)
si_release_descriptors(&sctx->descriptors[i]);
- sctx->vertex_buffers.list = NULL; /* points into a mapped buffer */
- si_release_descriptors(&sctx->vertex_buffers);
+ r600_resource_reference(&sctx->vb_descriptors_buffer, NULL);
+ sctx->vb_descriptors_gpu_list = NULL; /* points into a mapped buffer */
+
si_release_bindless_descriptors(sctx);
}
void si_all_descriptors_begin_new_cs(struct si_context *sctx)
{
int i;
for (i = 0; i < SI_NUM_SHADERS; i++) {
si_buffer_resources_begin_new_cs(sctx, &sctx->const_and_shader_buffers[i]);
si_sampler_views_begin_new_cs(sctx, &sctx->samplers[i]);
diff --git a/src/gallium/drivers/radeonsi/si_hw_context.c b/src/gallium/drivers/radeonsi/si_hw_context.c
index 3823be0..61c8d70 100644
--- a/src/gallium/drivers/radeonsi/si_hw_context.c
+++ b/src/gallium/drivers/radeonsi/si_hw_context.c
@@ -209,21 +209,21 @@ void si_begin_new_cs(struct si_context *ctx)
if (ctx->queued.named.hs)
ctx->prefetch_L2_mask |= SI_PREFETCH_HS;
if (ctx->queued.named.es)
ctx->prefetch_L2_mask |= SI_PREFETCH_ES;
if (ctx->queued.named.gs)
ctx->prefetch_L2_mask |= SI_PREFETCH_GS;
if (ctx->queued.named.vs)
ctx->prefetch_L2_mask |= SI_PREFETCH_VS;
if (ctx->queued.named.ps)
ctx->prefetch_L2_mask |= SI_PREFETCH_PS;
- if (ctx->vertex_buffers.buffer && ctx->vertex_elements)
+ if (ctx->vb_descriptors_buffer && ctx->vertex_elements)
ctx->prefetch_L2_mask |= SI_PREFETCH_VBO_DESCRIPTORS;
/* CLEAR_STATE disables all colorbuffers, so only enable bound ones. */
bool has_clear_state = ctx->screen->has_clear_state;
if (has_clear_state) {
ctx->framebuffer.dirty_cbufs =
u_bit_consecutive(0, ctx->framebuffer.state.nr_cbufs);
/* CLEAR_STATE disables the zbuffer, so only enable it if it's bound. */
ctx->framebuffer.dirty_zsbuf = ctx->framebuffer.state.zsbuf != NULL;
} else {
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 896b640..2053dcb 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -490,22 +490,26 @@ struct si_context {
struct si_shader_ctx_state tcs_shader;
struct si_shader_ctx_state tes_shader;
struct si_cs_shader_state cs_shader_state;
/* shader information */
struct si_vertex_elements *vertex_elements;
unsigned sprite_coord_enable;
bool flatshade;
bool do_update_shaders;
+ /* vertex buffer descriptors */
+ uint32_t *vb_descriptors_gpu_list;
+ struct r600_resource *vb_descriptors_buffer;
+ unsigned vb_descriptors_offset;
+
/* shader descriptors */
- struct si_descriptors vertex_buffers;
struct si_descriptors descriptors[SI_NUM_DESCS];
unsigned descriptors_dirty;
unsigned shader_pointers_dirty;
unsigned shader_needs_decompress_mask;
struct si_buffer_resources rw_buffers;
struct si_buffer_resources const_and_shader_buffers[SI_NUM_SHADERS];
struct si_samplers samplers[SI_NUM_SHADERS];
struct si_images images[SI_NUM_SHADERS];
/* other shader resources */
--
2.7.4
More information about the mesa-dev
mailing list