[Regression] MGS4 Performance

[Dormant-Hero]

Quick summary

Found a regression I can consistently reproduce from emulator build 32-16670 to 32-16676. There may be other regressions but from 32-16670 to the latest build there is a large difference also.

Details

Results below:

32-16670

32-16676

I had someone else confirm and they had ~20fps difference on their machine.

I think it relates to this: https://github.com/RPCS3/rpcs3/pull/15765

Result from latest build below:

Build with regression

v0.0.32-16676

Attach two log files

32-16670 RPCS3.log

32-16676 RPCS3.log

Attach capture files for visual issues

No response

System configuration

Other information

I did a test after my initial tests above with rpcs3 scheduler which seems to get me much more reliable results than my distros scheduler of choice.

[cipherxof]

+1

I lost about 20 FPS in this same scene with the mentioned PR.

[ItsLaguna]

Not losing as much, but still losing almost 5 fps on avg/max, while theres almost a 6 fps difference in 1%:

While in 16676 I get the following:

Bazzite 41 Ryzen 5800X

[cipherxof]

Seems to not be an issue with accurate spu reservations set to false, which is probably the intended behavior.

[cipherxof]

Disabling accurate spu reservations only restores the performance for me on Linux. The regression is still an issue in Windows.

EDIT: The prebuilt 16670 binary has better performance than my own compiled one for some reason. Disabling accurate spu res does seem to improve performance in Windows, but something else is causing the performance to regress a bit (maybe LLVM avx512 regression).

[cipherxof]

I did some more testing. https://github.com/RPCS3/rpcs3/pull/15933 restores the performance lost by the PR from the original post (as expected).

The primary regression happens with https://github.com/RPCS3/rpcs3/pull/15934/

Regression - 110fps

Previous build - 118fps

Master build - 114fps

[cipherxof]

Master build with https://github.com/RPCS3/rpcs3/pull/15934 reverted (https://github.com/cipherxof/rpcs3/commits/regression-test3/).

Note that I am only seeing the regression under Windows. These tests were done on a 8945HS.

[cipherxof]

The regression from https://github.com/RPCS3/rpcs3/pull/15934 seems to have been resolved by https://github.com/RPCS3/rpcs3/pull/16903, however there is another regression caused by https://github.com/RPCS3/rpcs3/pull/15759 (https://github.com/RPCS3/rpcs3/issues/16360). I suppose this issue can be closed when 16903 is merged.

[digant73]

The regression from #15934 seems to have been resolved by #16903, however there is another regression caused by #15759 (#16360). I suppose this issue can be closed when 16903 is merged.

#16903 has been merged since some time now. Is this specific issue fixed?

EDIT: Forget it. I wrongly saw that PR as merged while it is still open

[digant73]

those regressions have been probably fixed by recently merged PRs #17457 and #17472. Please retest on latest build 0.0.37-18149

[FarmerJoe7355]

Can this be retested?