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

ac/nir/lower_ps: move exports after packing alu

Browse Source 
If ACO's wqm section ends just before the first export, this mixing alu and
exports means the alu in question can't be reordered as much by the ILP
scheduler.

Foz-DB Navi31:
Totals from 8959 (11.31% of 79188) affected shaders:
Instrs: 5977212 -> 5978494 (+0.02%); split: -0.02%, +0.04%
CodeSize: 32982732 -> 32987876 (+0.02%); split: -0.01%, +0.03%
Latency: 35218073 -> 35216277 (-0.01%); split: -0.02%, +0.02%
InvThroughput: 5149751 -> 5149696 (-0.00%); split: -0.00%, +0.00%
SClause: 220552 -> 220551 (-0.00%); split: -0.01%, +0.01%
PreVGPRs: 313203 -> 313069 (-0.04%); split: -0.06%, +0.01%

Foz-DB Navi21:
Totals from 8895 (11.21% of 79377) affected shaders:
MaxWaves: 219280 -> 219272 (-0.00%); split: +0.00%, -0.01%
Instrs: 5393330 -> 5393366 (+0.00%); split: -0.00%, +0.00%
CodeSize: 29921900 -> 29922024 (+0.00%); split: -0.00%, +0.00%
VGPRs: 406664 -> 406688 (+0.01%); split: -0.00%, +0.01%
Latency: 35653975 -> 35652220 (-0.00%); split: -0.02%, +0.02%
InvThroughput: 7992134 -> 7992032 (-0.00%); split: -0.00%, +0.00%
SClause: 223784 -> 223786 (+0.00%)
Copies: 370984 -> 370983 (-0.00%)
PreVGPRs: 314323 -> 314330 (+0.00%); split: -0.01%, +0.01%
VALU: 3800023 -> 3800022 (-0.00%)

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33417>
This commit is contained in:
Georg Lehmann
2025-02-05 19:44:44 +01:00
committed by Marge Bot
parent 0bc1bffe9a
commit fd77cc7c32
1 changed files with 4 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/amd/common/nir/ac_nir_lower_ps_late.c
View File

@@ -650,6 +650,10 @@ export_ps_outputs(nir_builder *b, lower_ps_state *s)
}
if (s->exp_num) {
/* Move exports to the end to avoid mixing alu and exports. */
for (unsigned i = 0; i < s->exp_num; i++)
nir_instr_move(nir_after_impl(b->impl), &s->exp[i]->instr);
if (s->options->dual_src_blend_swizzle) {
emit_ps_dual_src_blend_swizzle(b, s, first_color_export);
/* Skip last export flag setting because they have been replaced by