fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-20 16:00:08 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

nir/copy_prop_vars: Mask out no-op writes to variables.
Some checks are pending
macOS-CI / macOS-CI (dri) (push) Waiting to run
Details
macOS-CI / macOS-CI (xlib) (push) Waiting to run
Details

Browse source 
The pass previously supported removing complete no-op writes, but we can
do better by noticing if any channel being written is the current
channel's value and masking off those writes.  I noticed this happening in
Stray's fragment shader, where it looked like some translation layer had
turned var[x].zw = vec2(a, b) into var[x] = vec4(var[x].x, var[x].y, a,
b).  This in turn lets nir_shrink_vec_array_vars be more effective.

Totals:
MaxWaves: 22158876 -> 22156696 (-0.01%); split: +0.00%, -0.01%
Instrs: 401167243 -> 401007996 (-0.04%); split: -0.04%, +0.00%
CodeSize: 1004397302 -> 1004133728 (-0.03%); split: -0.03%, +0.00%
STPs: 369810 -> 234618 (-36.56%)
LDPs: 209430 -> 172011 (-17.87%)

Totals from 1884 (0.12% of 1560230) affected shaders:
MaxWaves: 12686 -> 10506 (-17.18%); split: +6.97%, -24.15%
Instrs: 2099486 -> 1940239 (-7.59%); split: -7.64%, +0.06%
CodeSize: 4570472 -> 4306898 (-5.77%); split: -5.81%, +0.05%
NOPs: 334399 -> 270881 (-18.99%); split: -20.58%, +1.58%
MOVs: 131003 -> 148034 (+13.00%); split: -11.59%, +24.59%
COVs: 14512 -> 16921 (+16.60%); split: -0.23%, +16.83%
Full: 58120 -> 72399 (+24.57%); split: -6.75%, +31.31%
(ss): 79215 -> 45331 (-42.77%); split: -48.46%, +5.68%
(sy): 33081 -> 11119 (-66.39%); split: -66.56%, +0.18%
(ss)-stall: 302152 -> 115528 (-61.76%); split: -64.34%, +2.57%
(sy)-stall: 2706110 -> 498998 (-81.56%); split: -81.68%, +0.12%
STPs: 212045 -> 76853 (-63.76%)
LDPs: 47337 -> 9918 (-79.05%)
Preamble Instrs: 413954 -> 413630 (-0.08%); split: -0.21%, +0.13%

Cat0: 370362 -> 306844 (-17.15%); split: -18.58%, +1.43%
Cat1: 145629 -> 165003 (+13.30%); split: -10.51%, +23.81%
Cat2: 687947 -> 683992 (-0.57%); split: -0.61%, +0.04%
Cat3: 362919 -> 360690 (-0.61%); split: -0.72%, +0.11%
Cat6: 461411 -> 352375 (-23.63%)
Cat7: 16857 -> 16974 (+0.69%); split: -0.35%, +1.04%
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37313>
This commit is contained in:
Emma Anholt 2025-09-10 11:21:38 -07:00 • committed by Marge Bot
parent 537cc4e0ff
commit 0781edc30f
1 changed files with 42 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

47
src/compiler/nir/nir_opt_copy_prop_vars.c
View file

@ -123,20 +123,22 @@ struct copy_prop_var_state {
bool progress;
};
static bool
static nir_component_mask_t
value_equals_store_src(struct value *value, nir_intrinsic_instr *intrin)
{
assert(intrin->intrinsic == nir_intrinsic_store_deref);
nir_component_mask_t write_mask = nir_intrinsic_write_mask(intrin);
nir_component_mask_t equals_mask = 0;
for (unsigned i = 0; i < intrin->num_components; i++) {
nir_scalar src = nir_scalar_resolved(intrin->src[1].ssa, i);
if ((write_mask & (1 << i)) &&
(value->ssa.def[i] != intrin->src[1].ssa ||
value->ssa.component[i] != i))
return false;
(value->ssa.def[i] == src.def &&
value->ssa.component[i] == src.comp))
equals_mask |= 1 << i;
}
return true;
return equals_mask;
}
static struct vars_written *
@ -1195,13 +1197,46 @@ copy_prop_vars_block(struct copy_prop_var_state *state,
struct copy_entry *entry =
lookup_entry_for_deref(state, copies, &dst, nir_derefs_equal_bit, NULL);
if (entry && value_equals_store_src(&entry->src, intrin)) {
nir_component_mask_t equals_mask = entry ? value_equals_store_src(&entry->src, intrin) : 0;
if (equals_mask == nir_intrinsic_write_mask(intrin)) {
/* If we are storing the value from a load of the same var the
* store is redundant so remove it.
*/
nir_instr_remove(instr);
state->progress = true;
} else {
if (!(b->shader->info.stage == MESA_SHADER_FRAGMENT &&
(nir_deref_mode_may_be(dst.instr, nir_var_shader_out)))) {
/* If any channels we wrote were already the dst's value, mask them
* off (which can lead to other dead code elimination). Apparently
* glslang does write masking with load-vec-store.
*
* Skip this for FS outputs, where multiple drivers don't like color
* writes getting channels writemasked out based on copying the
* fbfetched data through.
*/
nir_component_mask_t remove_mask = equals_mask & nir_intrinsic_write_mask(intrin);
if (remove_mask) {
nir_intrinsic_set_write_mask(intrin, nir_intrinsic_write_mask(intrin) & ~equals_mask);
/* For any channels we're trimming off the write mask, replace
* them with undefs. This lets them be dead-code eliminated,
* which no other pass would do on its own.
*/
b->cursor = nir_before_instr(instr);
nir_def *channels[NIR_MAX_VEC_COMPONENTS];
nir_def *undef = nir_undef(b, 1, intrin->src[1].ssa->bit_size);
for (int i = 0; i < intrin->num_components; i++) {
if (remove_mask & (1 << i)) {
channels[i] = undef;
} else {
channels[i] = nir_channel(b, intrin->src[1].ssa, i);
}
}
nir_src_rewrite(&intrin->src[1], nir_vec(b, channels, intrin->num_components));
state->progress = true;
}
}
struct value value = { 0 };
value_set_ssa_components(&value, intrin->src[1].ssa,
intrin->num_components);