Skip to content

aco,nir: optimize partially undef vec/pack

Georg Lehmann requested to merge DadSchoorse/mesa:aco-isel-partial-vec-undef into main

RDNA3+ attribute ring stores store vecs that are partially undef, for better cache utilization. But we can avoid moves by not handling undef elements like 0.

Foz-DB Navi31:
Totals from 27464 (34.59% of 79395) affected shaders:
Instrs: 9595601 -> 9535260 (-0.63%); split: -0.63%, +0.00%
CodeSize: 47900112 -> 47658648 (-0.50%); split: -0.50%, +0.00%
Latency: 43928471 -> 43918448 (-0.02%); split: -0.05%, +0.02%
InvThroughput: 4940105 -> 4903447 (-0.74%); split: -0.75%, +0.01%
Copies: 667294 -> 604603 (-9.39%); split: -9.39%, +0.00%
VALU: 5282264 -> 5219604 (-1.19%); split: -1.19%, +0.00%
VOPD: 342 -> 311 (-9.06%)

Merge request reports