aco/gfx11: fix FS input loads in quad-divergent control flow
authorRhys Perry <pendingchaos02@gmail.com>
Wed, 26 Oct 2022 20:11:31 +0000 (21:11 +0100)
committerMarge Bot <emma+marge@anholt.net>
Tue, 1 Nov 2022 12:42:43 +0000 (12:42 +0000)
commit6113ee650a272dc737a200e276de474e083a2fdf
tree1e2a309793c6662db91a431f2a659c7ad47a7039
parent16d2c7ad557b46104f91365ab3405f0a3ed7e36d
aco/gfx11: fix FS input loads in quad-divergent control flow

This is not ideal and it would be great to somehow make it better some
day.

fossil-db (gfx1100):
Totals from 5208 (3.86% of 135032) affected shaders:
MaxWaves: 127058 -> 126962 (-0.08%); split: +0.01%, -0.09%
Instrs: 3983440 -> 4072736 (+2.24%); split: -0.00%, +2.24%
CodeSize: 21872468 -> 22230852 (+1.64%); split: -0.00%, +1.64%
VGPRs: 206688 -> 206984 (+0.14%); split: -0.05%, +0.20%
Latency: 37447383 -> 37491197 (+0.12%); split: -0.05%, +0.17%
InvThroughput: 6421955 -> 6422348 (+0.01%); split: -0.03%, +0.03%
VClause: 71579 -> 71545 (-0.05%); split: -0.09%, +0.04%
SClause: 148289 -> 147146 (-0.77%); split: -0.84%, +0.07%
Copies: 259011 -> 258084 (-0.36%); split: -0.61%, +0.25%
Branches: 101366 -> 101314 (-0.05%); split: -0.10%, +0.05%
PreSGPRs: 223482 -> 223460 (-0.01%); split: -0.21%, +0.20%
PreVGPRs: 184448 -> 184744 (+0.16%)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19370>
src/amd/compiler/aco_builder_h.py
src/amd/compiler/aco_instruction_selection.cpp
src/amd/compiler/aco_instruction_selection.h
src/amd/compiler/aco_lower_to_hw_instr.cpp
src/amd/compiler/aco_opcodes.py
src/amd/compiler/aco_optimizer.cpp
src/amd/compiler/aco_reduce_assign.cpp
src/amd/compiler/aco_register_allocation.cpp
src/amd/compiler/aco_validate.cpp