Skip to content

aco: Add pass to vectorize spill instructions

Friedrich Vock requested to merge pixelcluster/mesa:aco-spill-vectorize into main

Stats on main:

Totals from 14 (0.02% of 79395) affected shaders:
Instrs: 136242 -> 135883 (-0.26%); split: -0.27%, +0.01%
CodeSize: 777604 -> 774564 (-0.39%)
Latency: 2058554 -> 2053374 (-0.25%); split: -0.25%, +0.00%
InvThroughput: 1016809 -> 1014214 (-0.26%); split: -0.26%, +0.00%
Copies: 31283 -> 31404 (+0.39%); split: -0.02%, +0.40%
Branches: 4564 -> 4562 (-0.04%)
VALU: 97748 -> 97868 (+0.12%); split: -0.01%, +0.13%
SALU: 20206 -> 20207 (+0.00%)
VMEM: 4577 -> 4222 (-7.76%)

More shaders are affected on my RT function call branch:

Totals from 107 (0.13% of 79665) affected shaders:
Instrs: 476104 -> 475618 (-0.10%); split: -0.21%, +0.11%
CodeSize: 2673112 -> 2662652 (-0.39%); split: -0.46%, +0.07%
Latency: 5817091 -> 5814079 (-0.05%); split: -0.12%, +0.07%
InvThroughput: 1923728 -> 1923692 (-0.00%); split: -0.09%, +0.09%
VClause: 12667 -> 12774 (+0.84%); split: -0.03%, +0.88%
Copies: 77000 -> 79463 (+3.20%); split: -0.04%, +3.24%
Branches: 16080 -> 16069 (-0.07%)
VALU: 314117 -> 316582 (+0.78%); split: -0.01%, +0.79%
SALU: 69649 -> 69647 (-0.00%)
VMEM: 29839 -> 26858 (-9.99%)

Merge request reports