panfrost: Remove MRT indirection in blend shaders

Since we have a separate blend shader for each render target, let's
simplify this structure and reduce the options memory footprint by 88%
or something goofy like that.

Should also enable separate blending per render target.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This commit is contained in:
Alyssa Rosenzweig 2019-12-30 12:13:45 -05:00
parent 67fe2afa51
commit 8f4b15636b
4 changed files with 39 additions and 39 deletions

View File

@ -175,7 +175,7 @@ nir_blend(
for (unsigned c = 0; c < 4; ++c) {
/* Decide properties based on channel */
nir_lower_blend_channel chan =
(c < 3) ? options.rt[0].rgb : options.rt[0].alpha;
(c < 3) ? options.rgb : options.alpha;
nir_ssa_def *psrc = nir_channel(b, src, c);
nir_ssa_def *pdst = nir_channel(b, dst, c);
@ -197,7 +197,7 @@ nir_blend(
/* Then just recombine with an applied colormask */
nir_ssa_def *blended = nir_vec(b, channels, 4);
return nir_color_mask(b, options.rt[0].colormask, blended, dst);
return nir_color_mask(b, options.colormask, blended, dst);
}
static bool
@ -214,8 +214,8 @@ static bool
nir_is_blend_replace(nir_lower_blend_options options)
{
return
nir_is_blend_channel_replace(options.rt[0].rgb) &&
nir_is_blend_channel_replace(options.rt[0].alpha);
nir_is_blend_channel_replace(options.rgb) &&
nir_is_blend_channel_replace(options.alpha);
}
void

View File

@ -43,13 +43,11 @@ typedef struct {
} nir_lower_blend_channel;
typedef struct {
struct {
nir_lower_blend_channel rgb;
nir_lower_blend_channel alpha;
nir_lower_blend_channel rgb;
nir_lower_blend_channel alpha;
/* 4-bit colormask. 0x0 for none, 0xF for RGBA, 0x1 for R */
unsigned colormask;
} rt[8];
/* 4-bit colormask. 0x0 for none, 0xF for RGBA, 0x1 for R */
unsigned colormask;
} nir_lower_blend_options;
void nir_lower_blend(nir_shader *shader, nir_lower_blend_options options);

View File

@ -85,43 +85,41 @@
*/
static nir_lower_blend_options
nir_make_options(const struct pipe_blend_state *blend, unsigned nr_cbufs)
nir_make_options(const struct pipe_blend_state *blend, unsigned i)
{
nir_lower_blend_options options;
for (unsigned i = 0; i < nr_cbufs; ++i) {
/* If blend is disabled, we just use replace mode */
/* If blend is disabled, we just use replace mode */
nir_lower_blend_channel rgb = {
.func = BLEND_FUNC_ADD,
.src_factor = BLEND_FACTOR_ZERO,
.invert_src_factor = true,
.dst_factor = BLEND_FACTOR_ZERO,
.invert_dst_factor = false
};
nir_lower_blend_channel rgb = {
.func = BLEND_FUNC_ADD,
.src_factor = BLEND_FACTOR_ZERO,
.invert_src_factor = true,
.dst_factor = BLEND_FACTOR_ZERO,
.invert_dst_factor = false
};
nir_lower_blend_channel alpha = rgb;
nir_lower_blend_channel alpha = rgb;
if (blend->rt[i].blend_enable) {
rgb.func = util_blend_func_to_shader(blend->rt[i].rgb_func);
rgb.src_factor = util_blend_factor_to_shader(blend->rt[i].rgb_src_factor);
rgb.dst_factor = util_blend_factor_to_shader(blend->rt[i].rgb_dst_factor);
rgb.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_src_factor);
rgb.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_dst_factor);
if (blend->rt[i].blend_enable) {
rgb.func = util_blend_func_to_shader(blend->rt[i].rgb_func);
rgb.src_factor = util_blend_factor_to_shader(blend->rt[i].rgb_src_factor);
rgb.dst_factor = util_blend_factor_to_shader(blend->rt[i].rgb_dst_factor);
rgb.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_src_factor);
rgb.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_dst_factor);
alpha.func = util_blend_func_to_shader(blend->rt[i].alpha_func);
alpha.src_factor = util_blend_factor_to_shader(blend->rt[i].alpha_src_factor);
alpha.dst_factor = util_blend_factor_to_shader(blend->rt[i].alpha_dst_factor);
alpha.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_src_factor);
alpha.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_dst_factor);
}
options.rt[i].rgb = rgb;
options.rt[i].alpha = alpha;
options.rt[i].colormask = blend->rt[i].colormask;
alpha.func = util_blend_func_to_shader(blend->rt[i].alpha_func);
alpha.src_factor = util_blend_factor_to_shader(blend->rt[i].alpha_src_factor);
alpha.dst_factor = util_blend_factor_to_shader(blend->rt[i].alpha_dst_factor);
alpha.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_src_factor);
alpha.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_dst_factor);
}
options.rgb = rgb;
options.alpha = alpha;
options.colormask = blend->rt[i].colormask;
return options;
}
@ -166,7 +164,7 @@ panfrost_compile_blend_shader(
nir_store_var(b, c_out, s_src, 0xFF);
nir_lower_blend_options options =
nir_make_options(cso, 1);
nir_make_options(cso, rt);
NIR_PASS_V(shader, nir_lower_blend, options);
NIR_PASS_V(shader, nir_lower_framebuffer, format, screen->gpu_id);

View File

@ -1454,6 +1454,7 @@ disassemble_midgard(uint8_t *code, size_t size, unsigned gpu_id, gl_shader_stage
uint32_t *words = (uint32_t *) code;
unsigned num_words = size / 4;
int tabs = 0;
num_words = MIN2(num_words, 100);
bool prefetch_flag = false;
@ -1470,6 +1471,7 @@ disassemble_midgard(uint8_t *code, size_t size, unsigned gpu_id, gl_shader_stage
while (i < num_words) {
unsigned tag = words[i] & 0xF;
unsigned next_tag = (words[i] >> 4) & 0xF;
printf("\t%X -> %X\n", tag, next_tag);
unsigned num_quad_words = midgard_word_size[tag];
if (midg_tags[i] && midg_tags[i] != tag) {
@ -1544,12 +1546,14 @@ disassemble_midgard(uint8_t *code, size_t size, unsigned gpu_id, gl_shader_stage
/* Break based on instruction prefetch flag */
#if 0
if (i < num_words && next == 1) {
prefetch_flag = true;
if (midgard_word_types[words[i] & 0xF] != midgard_word_type_alu)
break;
}
#endif
i += 4 * num_quad_words;
}