fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-22 07:00:12 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

gallium/u_blitter: remove UTIL_BLITTER_ATTRIB_COLOR, use a constant buffer

Browse source 
Pass the clear color via a constant buffer instead of a FS input.
This results in less driver code and it could be faster on some GPUs.

Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33837>
This commit is contained in:
Marek Olšák 2025-03-02 11:04:47 -05:00 • committed by Marge Bot
parent 99b5970eb2
commit ca09c173f6
16 changed files with 50 additions and 114 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

82
src/gallium/auxiliary/util/u_blitter.c
View file

@ -75,8 +75,7 @@ struct blitter_context_priv
/* Fragment shaders. */
void *fs_empty;
void *fs_write_one_cbuf;
void *fs_clear_all_cbufs;
void *fs_clear_color[2];
/* FS which outputs a color from a texture where
* the 1st index indicates the texture type / destination type,
@ -450,30 +449,18 @@ static void bind_fs_empty(struct blitter_context_priv *ctx)
ctx->bind_fs_state(pipe, ctx->fs_empty);
}
static void bind_fs_write_one_cbuf(struct blitter_context_priv *ctx)
static void bind_fs_clear_color(struct blitter_context_priv *ctx,
bool write_all_cbufs)
{
struct pipe_context *pipe = ctx->base.pipe;
if (!ctx->fs_write_one_cbuf) {
if (!ctx->fs_clear_color[write_all_cbufs]) {
assert(!ctx->cached_all_shaders);
ctx->fs_write_one_cbuf =
util_make_fragment_passthrough_shader(pipe, TGSI_SEMANTIC_GENERIC,
TGSI_INTERPOLATE_CONSTANT, false);
ctx->fs_clear_color[write_all_cbufs] =
util_make_fs_clear_color(pipe, write_all_cbufs);
}
ctx->bind_fs_state(pipe, ctx->fs_write_one_cbuf);
}
static void bind_fs_clear_all_cbufs(struct blitter_context_priv *ctx)
{
struct pipe_context *pipe = ctx->base.pipe;
if (!ctx->fs_clear_all_cbufs) {
assert(!ctx->cached_all_shaders);
ctx->fs_clear_all_cbufs = util_make_fs_clear_all_cbufs(pipe);
}
ctx->bind_fs_state(pipe, ctx->fs_clear_all_cbufs);
ctx->bind_fs_state(pipe, ctx->fs_clear_color[write_all_cbufs]);
}
void util_blitter_destroy(struct blitter_context *blitter)
@ -569,10 +556,11 @@ void util_blitter_destroy(struct blitter_context *blitter)
if (ctx->fs_empty)
ctx->delete_fs_state(pipe, ctx->fs_empty);
if (ctx->fs_write_one_cbuf)
ctx->delete_fs_state(pipe, ctx->fs_write_one_cbuf);
if (ctx->fs_clear_all_cbufs)
ctx->delete_fs_state(pipe, ctx->fs_clear_all_cbufs);
for (i = 0; i < ARRAY_SIZE(ctx->fs_clear_color); i++) {
if (ctx->fs_clear_color[i])
ctx->delete_fs_state(pipe, ctx->fs_clear_color[i]);
}
for (i = 0; i < ARRAY_SIZE(ctx->fs_stencil_blit_fallback); ++i)
if (ctx->fs_stencil_blit_fallback[i])
@ -863,20 +851,6 @@ static void blitter_set_rectangle(struct blitter_context_priv *ctx,
ctx->base.pipe->set_viewport_states(ctx->base.pipe, 0, 1, &viewport);
}
static void blitter_set_clear_color(struct blitter_context_priv *ctx,
const float color[4])
{
int i;
if (color) {
for (i = 0; i < 4; i++)
memcpy(&ctx->vertices[i][1][0], color, sizeof(uint32_t) * 4);
} else {
for (i = 0; i < 4; i++)
memset(&ctx->vertices[i][1][0], 0, sizeof(uint32_t) * 4);
}
}
static void get_texcoords(struct pipe_sampler_view *src,
unsigned src_width0, unsigned src_height0,
int x1, int y1, int x2, int y2,
@ -1354,11 +1328,8 @@ void util_blitter_cache_all_shaders(struct blitter_context *blitter)
ctx->fs_empty = util_make_empty_fragment_shader(pipe);
ctx->fs_write_one_cbuf =
util_make_fragment_passthrough_shader(pipe, TGSI_SEMANTIC_GENERIC,
TGSI_INTERPOLATE_CONSTANT, false);
ctx->fs_clear_all_cbufs = util_make_fs_clear_all_cbufs(pipe);
ctx->fs_clear_color[0] = util_make_fs_clear_color(pipe, false);
ctx->fs_clear_color[1] = util_make_fs_clear_color(pipe, true);
ctx->cached_all_shaders = true;
}
@ -1431,10 +1402,6 @@ void util_blitter_draw_rectangle(struct blitter_context *blitter,
unsigned i;
switch (type) {
case UTIL_BLITTER_ATTRIB_COLOR:
blitter_set_clear_color(ctx, attrib->color);
break;
case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
for (i = 0; i < 4; i++) {
ctx->vertices[i][1][2] = attrib->texcoord.z;
@ -1565,7 +1532,7 @@ static void util_blitter_clear_custom(struct blitter_context *blitter,
};
pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, blitter->cb_slot,
false, &cb);
bind_fs_clear_all_cbufs(ctx);
bind_fs_clear_color(ctx, true);
} else {
bind_fs_empty(ctx);
}
@ -2408,7 +2375,7 @@ void util_blitter_clear_render_target(struct blitter_context *blitter,
/* bind states */
pipe->bind_blend_state(pipe, ctx->blend[PIPE_MASK_RGBA][0]);
pipe->bind_depth_stencil_alpha_state(pipe, ctx->dsa_keep_depth_stencil);
bind_fs_write_one_cbuf(ctx);
bind_fs_clear_color(ctx, false);
/* set a framebuffer state */
fb_state.width = dstsurf->width;
@ -2426,8 +2393,12 @@ void util_blitter_clear_render_target(struct blitter_context *blitter,
blitter_set_dst_dimensions(ctx, dstsurf->width, dstsurf->height);
blitter_set_common_draw_rect_state(ctx, false, msaa);
union blitter_attrib attrib;
memcpy(attrib.color, color->ui, sizeof(color->ui));
struct pipe_constant_buffer cb = {
.user_buffer = color->f,
.buffer_size = 4 * sizeof(float),
};
pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, blitter->cb_slot,
false, &cb);
num_layers = dstsurf->u.tex.last_layer - dstsurf->u.tex.first_layer + 1;
@ -2440,10 +2411,11 @@ void util_blitter_clear_render_target(struct blitter_context *blitter,
blitter->draw_rectangle(blitter, ctx->velem_state, get_vs,
dstx, dsty, dstx+width, dsty+height, 0,
num_layers, UTIL_BLITTER_ATTRIB_COLOR, &attrib);
num_layers, UTIL_BLITTER_ATTRIB_NONE, NULL);
util_blitter_restore_vertex_states(blitter);
util_blitter_restore_fragment_states(blitter);
util_blitter_restore_constant_buffer_state(blitter);
util_blitter_restore_fb_state(blitter);
util_blitter_restore_render_cond(blitter);
util_blitter_unset_running_flag(blitter);
@ -2558,7 +2530,7 @@ void util_blitter_custom_depth_stencil(struct blitter_context *blitter,
ctx->blend[0][0]);
pipe->bind_depth_stencil_alpha_state(pipe, dsa_stage);
if (cbsurf)
bind_fs_write_one_cbuf(ctx);
bind_fs_clear_color(ctx, false);
else
bind_fs_empty(ctx);
@ -2685,7 +2657,7 @@ void util_blitter_custom_resolve_color(struct blitter_context *blitter,
/* bind states */
pipe->bind_blend_state(pipe, custom_blend);
pipe->bind_depth_stencil_alpha_state(pipe, ctx->dsa_keep_depth_stencil);
bind_fs_write_one_cbuf(ctx);
bind_fs_clear_color(ctx, false);
pipe->set_sample_mask(pipe, sample_mask);
if (pipe->set_min_samples)
pipe->set_min_samples(pipe, 1);
@ -2753,7 +2725,7 @@ void util_blitter_custom_color(struct blitter_context *blitter,
pipe->bind_blend_state(pipe, custom_blend ? custom_blend
: ctx->blend[PIPE_MASK_RGBA][0]);
pipe->bind_depth_stencil_alpha_state(pipe, ctx->dsa_keep_depth_stencil);
bind_fs_write_one_cbuf(ctx);
bind_fs_clear_color(ctx, false);
/* set a framebuffer state */
fb_state.width = dstsurf->width;

3
src/gallium/auxiliary/util/u_blitter.h
View file

@ -40,14 +40,11 @@ struct pipe_context;
enum blitter_attrib_type {
UTIL_BLITTER_ATTRIB_NONE,
UTIL_BLITTER_ATTRIB_COLOR,
UTIL_BLITTER_ATTRIB_TEXCOORD_XY,
UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW,
};
union blitter_attrib {
float color[4];
struct {
float x1, y1, x2, y2, z, w;
} texcoord;

9
src/gallium/auxiliary/util/u_simple_shaders.c
View file

@ -1313,16 +1313,19 @@ util_make_fs_stencil_blit(struct pipe_context *pipe, bool msaa_src, bool has_txq
}
void *
util_make_fs_clear_all_cbufs(struct pipe_context *pipe)
util_make_fs_clear_color(struct pipe_context *pipe, bool write_all_cbufs)
{
static const char text[] =
static const char text_templ[] =
"FRAG\n"
"PROPERTY FS_COLOR0_WRITES_ALL_CBUFS 1\n"
"PROPERTY FS_COLOR0_WRITES_ALL_CBUFS %u\n"
"DCL OUT[0], COLOR[0]\n"
"DCL CONST[0][0]\n"
"MOV OUT[0], CONST[0][0]\n"
"END\n";
char text[1000];
snprintf(text, ARRAY_SIZE(text), text_templ, write_all_cbufs);
struct tgsi_token tokens[1000];
struct pipe_shader_state state = { 0 };

2
src/gallium/auxiliary/util/u_simple_shaders.h
View file

@ -161,7 +161,7 @@ void *
util_make_fs_stencil_blit(struct pipe_context *pipe, bool msaa_src, bool has_txq);
void *
util_make_fs_clear_all_cbufs(struct pipe_context *pipe);
util_make_fs_clear_color(struct pipe_context *pipe, bool write_all_cbufs);
#ifdef __cplusplus
}

2
src/gallium/drivers/panfrost/pan_resource.c
View file

@ -85,7 +85,7 @@ panfrost_clear_render_target(struct pipe_context *pipe,
return;
panfrost_blitter_save(
ctx, render_condition_enabled ? PAN_RENDER_COND : PAN_RENDER_BASE);
ctx, (render_condition_enabled ? PAN_RENDER_COND : PAN_RENDER_BASE) | PAN_SAVE_FRAGMENT_CONSTANT);
util_blitter_clear_render_target(ctx->blitter, dst, color, dstx, dsty, width,
height);
}

9
src/gallium/drivers/r300/r300_render.c
View file

@ -1161,11 +1161,9 @@ void r300_blitter_draw_rectangle(struct blitter_context *blitter,
unsigned last_is_point = r300->is_point;
unsigned width = x2 - x1;
unsigned height = y2 - y1;
unsigned vertex_size =
type == UTIL_BLITTER_ATTRIB_COLOR || !r300->draw ? 8 : 4;
unsigned vertex_size = !r300->draw ? 8 : 4;
unsigned dwords = 13 + vertex_size +
(type == UTIL_BLITTER_ATTRIB_TEXCOORD_XY ? 7 : 0);
static const union blitter_attrib zeros;
CS_LOCALS(r300);
/* XXX workaround for a lockup in MSAA resolve on SWTCL chipsets, this
@ -1234,9 +1232,8 @@ void r300_blitter_draw_rectangle(struct blitter_context *blitter,
OUT_CS_32F(1);
if (vertex_size == 8) {
if (!attrib)
attrib = &zeros;
OUT_CS_TABLE(attrib->color, 4);
static const float zeros[4];
OUT_CS_TABLE(zeros, 4);
}
END_CS;

2
src/gallium/drivers/r600/r600_blit.c
View file

@ -20,7 +20,7 @@ enum r600_blitter_op /* bitmask */
R600_CLEAR = R600_SAVE_FRAGMENT_STATE | R600_SAVE_CONST_BUF0,
R600_CLEAR_SURFACE = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER,
R600_CLEAR_SURFACE = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER | R600_SAVE_CONST_BUF0,
R600_COPY_BUFFER = R600_DISABLE_RENDER_COND,

5
src/gallium/drivers/r600/r600_pipe_common.c
View file

@ -163,11 +163,6 @@ void r600_draw_rectangle(struct blitter_context *blitter,
vb[19] = 1;
switch (type) {
case UTIL_BLITTER_ATTRIB_COLOR:
memcpy(vb+4, attrib->color, sizeof(float)*4);
memcpy(vb+12, attrib->color, sizeof(float)*4);
memcpy(vb+20, attrib->color, sizeof(float)*4);
break;
case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
case UTIL_BLITTER_ATTRIB_TEXCOORD_XY:
vb[6] = vb[14] = vb[22] = attrib->texcoord.z;

2
src/gallium/drivers/radeonsi/si_clear.c
View file

@ -13,7 +13,7 @@
enum {
SI_CLEAR = SI_SAVE_FRAGMENT_STATE | SI_SAVE_FRAGMENT_CONSTANT,
SI_CLEAR_SURFACE = SI_SAVE_FRAMEBUFFER | SI_SAVE_FRAGMENT_STATE,
SI_CLEAR_SURFACE = SI_SAVE_FRAMEBUFFER | SI_SAVE_FRAGMENT_STATE | SI_SAVE_FRAGMENT_CONSTANT,
};
void si_init_buffer_clear(struct si_clear_info *info,

23
src/gallium/drivers/radeonsi/si_nir_lower_vs_inputs.c
View file

@ -132,22 +132,17 @@ load_vs_input_from_blit_sgpr(nir_builder *b, unsigned input_index,
assert(input_index == 1);
unsigned vs_blit_property = b->shader->info.vs.blit_sgprs_amd;
if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR + has_attribute_ring_address) {
for (int i = 0; i < 4; i++)
out[i] = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 3 + i);
} else {
assert(vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD + has_attribute_ring_address);
assert(vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD + has_attribute_ring_address);
nir_def *x1 = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 3);
nir_def *y1 = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 4);
nir_def *x2 = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 5);
nir_def *y2 = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 6);
nir_def *x1 = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 3);
nir_def *y1 = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 4);
nir_def *x2 = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 5);
nir_def *y2 = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 6);
out[0] = nir_bcsel(b, sel_x1, x1, x2);
out[1] = nir_bcsel(b, sel_y1, y1, y2);
out[2] = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 7);
out[3] = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 8);
}
out[0] = nir_bcsel(b, sel_x1, x1, x2);
out[1] = nir_bcsel(b, sel_y1, y1, y2);
out[2] = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 7);
out[3] = ac_nir_load_arg_at_offset(b, &s->args->ac, s->args->vs_blit_inputs, 8);
}
}

4
src/gallium/drivers/radeonsi/si_pipe.c
View file

@ -251,10 +251,6 @@ static void si_destroy_context(struct pipe_context *context)
sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_pos);
if (sctx->vs_blit_pos_layered)
sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_pos_layered);
if (sctx->vs_blit_color)
sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_color);
if (sctx->vs_blit_color_layered)
sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_color_layered);
if (sctx->vs_blit_texcoord)
sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_texcoord);
if (sctx->cs_clear_buffer_rmw)

2
src/gallium/drivers/radeonsi/si_pipe.h
View file

@ -973,8 +973,6 @@ struct si_context {
void *custom_blend_dcc_decompress;
void *vs_blit_pos;
void *vs_blit_pos_layered;
void *vs_blit_color;
void *vs_blit_color_layered;
void *vs_blit_texcoord;
void *cs_clear_buffer_rmw;
void *cs_ubyte_to_ushort;

10
src/gallium/drivers/radeonsi/si_shader.c
View file

@ -294,15 +294,7 @@ static void declare_vs_blit_inputs(struct si_shader *shader, struct si_shader_ar
ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* depth */
if (info->vs.blit_sgprs_amd ==
SI_VS_BLIT_SGPRS_POS_COLOR + has_attribute_ring_address) {
ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color0 */
ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color1 */
ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color2 */
ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color3 */
if (has_attribute_ring_address)
ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* attribute ring address */
} else if (info->vs.blit_sgprs_amd ==
SI_VS_BLIT_SGPRS_POS_TEXCOORD + has_attribute_ring_address) {
SI_VS_BLIT_SGPRS_POS_TEXCOORD + has_attribute_ring_address) {
ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x1 */
ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.y1 */
ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x2 */

1
src/gallium/drivers/radeonsi/si_shader.h
View file

@ -326,7 +326,6 @@ enum
{
/* These represent the number of SGPRs the shader uses. */
SI_VS_BLIT_SGPRS_POS = 3,
SI_VS_BLIT_SGPRS_POS_COLOR = 7,
SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
MAX_SI_VS_BLIT_SGPRS = 10, /* +1 for the attribute ring address */

4
src/gallium/drivers/radeonsi/si_shaderlib_nir.c
View file

@ -300,10 +300,6 @@ void *si_get_blitter_vs(struct si_context *sctx, enum blitter_attrib_type type,
vs = num_layers > 1 ? &sctx->vs_blit_pos_layered : &sctx->vs_blit_pos;
vs_blit_property = SI_VS_BLIT_SGPRS_POS;
break;
case UTIL_BLITTER_ATTRIB_COLOR:
vs = num_layers > 1 ? &sctx->vs_blit_color_layered : &sctx->vs_blit_color;
vs_blit_property = SI_VS_BLIT_SGPRS_POS_COLOR;
break;
case UTIL_BLITTER_ATTRIB_TEXCOORD_XY:
case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
assert(num_layers == 1);

4
src/gallium/drivers/radeonsi/si_state_draw.cpp
View file

@ -2498,10 +2498,6 @@ static void si_draw_rectangle(struct blitter_context *blitter, void *vertex_elem
sctx->vs_blit_sh_data[2] = fui(depth);
switch (type) {
case UTIL_BLITTER_ATTRIB_COLOR:
memcpy(&sctx->vs_blit_sh_data[3], attrib->color, sizeof(float) * 4);
sctx->vs_blit_sh_data[7] = attribute_ring_address_lo;
break;
case UTIL_BLITTER_ATTRIB_TEXCOORD_XY:
case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
memcpy(&sctx->vs_blit_sh_data[3], &attrib->texcoord, sizeof(attrib->texcoord));