fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-27 07:40:31 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

nir/opt_uniform_subgroup: optimize uniform ddx/ddy
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 
We can't just use 0.0 as the replacement because of NaN/Inf.
But turning the intrinsic into a simple fsub should still be better
or at least equal.

Foz-DB Navi48:
Totals from 128 (0.10% of 125402) affected shaders:
MaxWaves: 3684 -> 3708 (+0.65%)
Instrs: 111150 -> 111055 (-0.09%); split: -0.20%, +0.11%
CodeSize: 587176 -> 590800 (+0.62%); split: -0.01%, +0.63%
VGPRs: 6540 -> 6480 (-0.92%)
Latency: 382775 -> 383332 (+0.15%); split: -0.15%, +0.29%
InvThroughput: 80909 -> 80530 (-0.47%); split: -0.51%, +0.04%
VClause: 1433 -> 1430 (-0.21%)
SClause: 1834 -> 1841 (+0.38%); split: -0.11%, +0.49%
Copies: 6130 -> 6096 (-0.55%); split: -1.29%, +0.73%
PreSGPRs: 7352 -> 7356 (+0.05%)
PreVGPRs: 4797 -> 4721 (-1.58%)
VALU: 71892 -> 71435 (-0.64%); split: -0.64%, +0.01%
SALU: 12665 -> 13056 (+3.09%); split: -0.06%, +3.14%

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39112>
This commit is contained in:
Georg Lehmann 2025-12-30 21:56:24 +01:00 • committed by Marge Bot
parent 7ed6679361
commit 369a3b22b4
1 changed files with 15 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

15
src/compiler/nir/nir_opt_uniform_subgroup.c
View file

@ -221,6 +221,21 @@ opt_uniform_subgroup_instr(nir_builder *b, nir_intrinsic_instr *intrin, void *_s
}
break;
case nir_intrinsic_ddx:
case nir_intrinsic_ddx_coarse:
case nir_intrinsic_ddx_fine:
case nir_intrinsic_ddy:
case nir_intrinsic_ddy_coarse:
case nir_intrinsic_ddy_fine:
if (nir_src_is_divergent(&intrin->src[0]))
return false;
nir_def *x = intrin->src[0].ssa;
b->fp_math_ctrl = nir_fp_no_fast_math;
replacement = nir_fsub(b, x, x);
b->fp_math_ctrl = nir_fp_fast_math;
break;
case nir_intrinsic_reduce:
case nir_intrinsic_exclusive_scan:
case nir_intrinsic_inclusive_scan: {