NPJH50276 Adhoc Performance Regression (>=1.11.0)

Schtolteheim commented 3 years ago

What games and where in them does it happen?

Ys vs. Sora no Kiseki (NPJH50276) during 4-player adhoc multiplayer (4 instances on 1 machine, in my case)

What speed are you getting with PPSSPP speed display enabled?

Host's framerate slows to 24, while guests remain at 30. Tested on 1.11.0, 1.11.3, and 1.11.3-729-g69fa20744. Last observed build with full speed was 1.10.3-873-g6ec4e012e, where host and guests maintained stable 30. All settings remained the same between all builds, to my knowledge.

What hardware / device and operating system?

Windows 10 64-bit, Ryzen 9 5900X; also tested with i7 6700K

What graphics card (GPU) or mobile phone model?

RTX 2080Ti; also tested with GTX 970M

Graphics backend (Vulkan, OpenGL, etc.)?

Direct3D 11

Checklist

[X] Test in the latest git build in case it's already fixed.
[X] Search for other reports of the same issue.
[X] Try resetting settings or older versions and include if the issue is related.
[X] Try changing graphics settings to determine if one causes the slowdown.
[X] Include logs of screenshots of issue.

unknownbrackets commented 3 years ago

Given that it still says 100%, this means something is consuming more cycles than before. Presumably incorrectly, though it would be ideal to verify the framerate on a PSP with a CFW plugin.

-[Unknown]

anr2me commented 3 years ago

I couldn't find 1.10.3-873-g6ec4e012e at https://buildbot.orphis.net/ppsspp/index.php?m=fulllist Are you using test build?

Could you try a few version after 1.10.3-873 from https://buildbot.orphis.net/ppsspp/index.php?m=fulllist to find out the latest official dev build that worked full speed?

You can try those builds that have Merged ANR2ME repo

Schtolteheim commented 3 years ago

I was using the build from https://github.com/hrydgard/ppsspp/pull/13519

Will test those builds shortly.

Schtolteheim commented 3 years ago

Last adhoc-related build that ran at full speed was 1.10.3-965-gd738c9b78 (1.10.3-1026-g36a614072 also ran fine).

All builds from 1.10.3-1044-ga4a0b4ba0 onward seem to have this issue.

unknownbrackets commented 3 years ago

That points to #13579. Perhaps 3020efdbe was aggressive and could be reduced.

-[Unknown]

anr2me commented 3 years ago

Yeah, those hleEatMicro might be too much, since there is already an overhead on the actual socket it self.

anr2me commented 3 years ago

Please try the artifacts on this PR later https://github.com/hrydgard/ppsspp/pull/14513/checks?check_run_id=2760683525

hrydgard / ppsspp