[amdgpu] split vgpr regalloc pipeline caused perf degradation about 10% in buildbot Blender test

[yxsamliu]

#93526 caused perf degradation in buildbot Blender test.

for example, the sky demo scene has about 10% kernel execution time increase.

2024-09-30 12:32:38,735a5f67,0.019419
2024-09-30 12:40:26,8e6bba23,0.019443
2024-09-30 12:48:22,18df9d23,0.019464
2024-09-30 12:57:20,5df7d88c,0.019439
2024-09-30 13:45:33,bfa8519b,0.019486
2024-09-30 13:57:00,be6a5dc9,0.01932
2024-09-30 14:05:10,06176293,0.019433
2024-09-30 14:38:16,ac0f64f0,0.02194
2024-09-30 14:46:29,9f6cd56d,0.021921
2024-09-30 15:05:01,93af9d6c,0.021789
2024-09-30 15:13:05,f9fbfc58,0.021948
2024-09-30 15:22:02,4ae0c505,0.021871
2024-09-30 15:30:14,01135de4,0.021882
2024-09-30 15:38:44,f2562229,0.021994

will post the dumped files for the good/bad cases.

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Yaxun (Sam) Liu (yxsamliu)

https://github.com//pull/93526 caused perf degradation in buildbot Blender test.

for example, the sky demo scene has about 10% kernel execution time increase.

2024-09-30 12:32:38,735a5f67,0.019419
2024-09-30 12:40:26,8e6bba23,0.019443
2024-09-30 12:48:22,18df9d23,0.019464
2024-09-30 12:57:20,5df7d88c,0.019439
2024-09-30 13:45:33,bfa8519b,0.019486
2024-09-30 13:57:00,be6a5dc9,0.01932
2024-09-30 14:05:10,06176293,0.019433
2024-09-30 14:38:16,ac0f64f0,0.02194
2024-09-30 14:46:29,9f6cd56d,0.021921
2024-09-30 15:05:01,93af9d6c,0.021789
2024-09-30 15:13:05,f9fbfc58,0.021948
2024-09-30 15:22:02,4ae0c505,0.021871
2024-09-30 15:30:14,01135de4,0.021882
2024-09-30 15:38:44,f2562229,0.021994

will post the dumped files for the good/bad cases.

[yxsamliu]

blender-perf-regression.zip
the dumped files for the good/bad cases are attached.

[arsenm]

Haven't looked at the outputs, but I was wondering if this could be due to the loss of the optimization to compact multiple SGPR spills into one VGPR