AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

aco: disable wqm for tex loads when not needed

Browse Source 
By only executing VMEM loads for lanes where the result is used, we can save
bandwidth.

The NIR pass only handles tex for now, but those are most common anyway.
We can extend it handle image/ssbo/ubo/global loads in the future.

Foz-DB GFX1201:
Totals from 32633 (40.66% of 80251) affected shaders:
Instrs: 22635910 -> 23193509 (+2.46%); split: -0.00%, +2.46%
CodeSize: 122880044 -> 125093428 (+1.80%); split: -0.00%, +1.81%
VGPRs: 1481868 -> 1481712 (-0.01%)
SpillSGPRs: 3877 -> 4301 (+10.94%); split: -0.52%, +11.45%
Latency: 171480552 -> 171685219 (+0.12%); split: -0.18%, +0.30%
InvThroughput: 24364743 -> 24373441 (+0.04%); split: -0.08%, +0.12%
VClause: 388318 -> 388557 (+0.06%); split: -0.06%, +0.13%
SClause: 774781 -> 776492 (+0.22%); split: -0.29%, +0.51%
Copies: 1416586 -> 1541199 (+8.80%); split: -0.16%, +8.96%
Branches: 419591 -> 419673 (+0.02%); split: -0.02%, +0.04%
PreSGPRs: 1330303 -> 1416540 (+6.48%)
PreVGPRs: 964864 -> 964863 (-0.00%)
VALU: 12919601 -> 12920254 (+0.01%); split: -0.01%, +0.01%
SALU: 2685402 -> 3224147 (+20.06%); split: -0.00%, +20.07%

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35970>
This commit is contained in:
Georg Lehmann
2025-07-06 23:12:58 +02:00
committed by Marge Bot
parent 7159fd21f8
commit 883b1ca364
3 changed files with 13 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/compiler/instruction_selection/aco_isel_setup.cpp
+8 -5
View File
@@ -367,6 +367,12 @@ init_context(isel_context* ctx, nir_shader* shader)
apply_nuw_to_offsets(ctx, impl);
ac_nir_flag_smem_for_loads(shader, ctx->program->gfx_level, false, true);
if (shader->info.stage == MESA_SHADER_FRAGMENT) {
nir_opt_load_skip_helpers_options skip_helper_options = {};
skip_helper_options.no_add_divergence = true;
nir_opt_load_skip_helpers(shader, &skip_helper_options);
}
/* sanitize control flow */
sanitize_cf_list(impl, &impl->body);
nir_progress(true, impl, nir_metadata_none);
@@ -621,11 +627,8 @@ init_context(isel_context* ctx, nir_shader* shader)
}
case nir_instr_type_tex: {
nir_tex_instr* tex = nir_instr_as_tex(instr);
RegType type = tex->def.divergent ? RegType::vgpr : RegType::sgpr;
if (tex->op == nir_texop_texture_samples) {
assert(!tex->def.divergent);
}
RegType type =
tex->def.divergent || tex->skip_helpers ? RegType::vgpr : RegType::sgpr;
RegClass rc = get_reg_class(ctx, type, tex->def.num_components, tex->def.bit_size);
regclasses[tex->def.index] = rc;