Skip to content

aco/gfx11: use v_cmp_class_f16 with opsel for bitnz/bitz

Georg Lehmann requested to merge DadSchoorse/mesa:aco-gfx11-bitnz-opsel into main

Allows to check 22 out of 32 bits with a single VALU instruction instead of just 12.

Foz-DB GFX11:
Totals from 1607 (1.21% of 132657) affected shaders:
MaxWaves: 36511 -> 36513 (+0.01%)
Instrs: 2412601 -> 2408770 (-0.16%); split: -0.16%, +0.00%
CodeSize: 12434716 -> 12414536 (-0.16%); split: -0.17%, +0.00%
VGPRs: 108872 -> 108860 (-0.01%)
Latency: 27555570 -> 27541897 (-0.05%); split: -0.05%, +0.00%
InvThroughput: 3144382 -> 3140929 (-0.11%); split: -0.11%, +0.00%
SClause: 65762 -> 65770 (+0.01%)
Copies: 127148 -> 127170 (+0.02%); split: -0.06%, +0.08%
PreVGPRs: 96737 -> 96447 (-0.30%)

Merge request reports