SPU: Lock-line reservation optimizations + Savestates bugfix #12523

elad335 · 2022-08-18T07:47:01Z

"SPU GETLLAR polling detection" setting (which was yielding resources when it's set to true) has been transformed to "SPU Reservation Busy Waiting Percentage" and is off ("0") by default. (0 is yielding resources, 100 is always busy-waiting)
It's now responsible for SPU events waiting as well, so the yielding can be replaced with busy-waiting there. You can test if you get perf gains with it set to on if your CPU has many threads.
It's not allowed to be changed for TSX users because TSX is very sensitive to memory touching. It's always as if this setting is set to "0" for it.
Fix the detection of GETLLAR spin and optimize the actual yielding.
Add some more minor optimizations.
Fix PPU HLE injection.
Fix PPU function analysis for savestates.
Added an experimental SPU setting called "SPU Accurate Reservations" which may benefit CPUs, especially without TSX when set to "false" ("true" is the default and was the default before this pull request). Game config file only atm.
Detect PPU lwmutex updates, do not notify SPU threads when it's detected.
Implement specialized atomic waiting for SPU reservations - emphasizing on cheap thread sleep entering and exiting cost.
Disable notifications when no data was changed in PUTLLC.

This pull request is known to fix audio stutters and heavy framerate drops in Demon Souls for low threaded CPUs.
Fixes #12513

elad335 · 2022-08-18T09:17:51Z

Added experimental SPU setting called "SPU Inaccurate Reservations" which may benefit non-TSX users. Game config file only atm.

Nerboruto · 2022-08-18T10:54:44Z

generally it seems to keep 30 fps better and more fluidity on xeon e5450 quadcore 3 ghz, without changing the optimizations I have on at the moment, 3 spurs threads and 60fps patch with 30 vblank ...
the heaviest scene is the one in the picture only that the values I get are discordant and perhaps depend on windows scheduler, between 20 and 23fps we dance 3fps. some times it starts at 20 others at 23...
it would take a more empirical method to test however it seems more fluid...

elad335 · 2022-08-18T10:58:33Z

@Nerboruto Tested with inaccurate SPU reservations?

Nerboruto · 2022-08-18T11:00:45Z

@Nerboruto Tested with inaccurate SPU reservations?

I have not tried to change anything in config, now I'm going to try again ..

elad335 · 2022-08-18T11:41:43Z

Fixed a bug which made the new setting randomly not optimize in different runs.

Nerboruto · 2022-08-18T12:56:02Z

i tried with spu inaccurate reservation false and true...
unfortunately I still have the margin for error ... i also read 24fps.
it seems to be smoother but I have no way of having actual proof.
the game fluctuates between 20 and 30 but is more on 30, the slowdowns are there in some places.

elad335 · 2022-08-18T13:12:58Z

Added another optimization, please test and report performance regressions.

Nerboruto · 2022-08-18T17:42:55Z

well first i attach my config to see if i'm wrong something, i set it to true..
I have not seen any regression.
I put the master back to compare and in the opening scene there are about + 2fps of difference for the pr.
I tried to play a bit but the difference is obvious, pr has much more consistent and high minimum fps. in short, it is a nice performance improvement.
UPDATE: God of war gos crash if i set to true, false work good and fast...

config_BLES00932.zip

* Make this setting guard all reservation waitings. (renamed) * Revert atomic list usage: it's more expensive and not needed because it has a timeout and is not optimized for the rest of the waitables.

elad335 · 2022-08-20T18:27:36Z

@Hasster1 Thanks, seemingly fixed it and some performance miss-optimizations with it.

Hasster1 · 2022-08-20T19:53:29Z

@Hasster1 Thanks, seemingly fixed it and some performance miss-optimizations with it.

Wonderful job, nothing is freezing/crashing anymore with the inaccurate SPU reservations. Thank you.

Hasster1 · 2022-08-20T20:26:05Z

Looks like Inaccurate SPU Reservations breaks InFamous. Freezes right after the intro without any errors.

RPCS3.log

rpcs3/Emu/system_config.h

Nekotekina · 2022-08-21T10:39:22Z

rpcs3/Emu/Cell/PPUThread.cpp

+
+				if (far_jump)
+				{
+					// Replace the function with ppu_far_jump


I don't think it works as intended. Module-local jumps don't use the global mapping.

Yes but it updates its PPU hash and forces recompilation. That's how it was handled before.

That's good but how forcing recompilation would handle local branches? Especially with PPUTranslator having far jump part removed.

It uses the manually installed value because the function is removed from compilation.

The code below is responsible for its compilation in llvm and that's skipped.

Okay, I see now.

Implement cellSpursRequestIdleSpu

Hasster1 · 2022-08-21T11:28:48Z

Excuse me, what?

Also i can't find that the inaccurate reservations are enabled in the logs.

Hasster1 · 2022-08-21T11:31:16Z

Oh, i see that it's been reversed to avoid confusion. Sorry.

elad335 changed the title ~~SPU Enable thread yield on GETLLAR spinning by default~~ SPU/Savestates: Enable thread yield on GETLLAR spinning by default Aug 18, 2022

elad335 force-pushed the savestates branch 6 times, most recently from 8f59609 to 53fa746 Compare August 18, 2022 09:16

elad335 force-pushed the savestates branch from 53fa746 to ca47a0d Compare August 18, 2022 09:22

This comment was marked as off-topic.

Sign in to view

elad335 force-pushed the savestates branch 2 times, most recently from 64445ea to 47110ac Compare August 18, 2022 11:31

This comment was marked as off-topic.

Sign in to view

elad335 force-pushed the savestates branch 3 times, most recently from 71ae3c7 to 8ccd92d Compare August 18, 2022 15:08

elad335 changed the title ~~SPU/Savestates: Enable thread yield on GETLLAR spinning by default~~ SPU: Enable thread yield on GETLLAR spinning by default + Savestates bugfix Aug 18, 2022

elad335 force-pushed the savestates branch 6 times, most recently from 0a79106 to c25b5f5 Compare August 18, 2022 23:35

elad335 added 4 commits August 20, 2022 21:26

SPU: Optimize and enable SPU GETLLAR Polling detection by default

3adf1c8

* Make this setting guard all reservation waitings. (renamed) * Revert atomic list usage: it's more expensive and not needed because it has a timeout and is not optimized for the rest of the waitables.

SPU: Optimize spu_thread::get_events()

847f198

PPU LLVM: Fix HLE function injection

7c6ca7e

Savestates: Fix PPU analysis when patches are applied

7409db5

Nekotekina reviewed Aug 21, 2022

View reviewed changes

rpcs3/Emu/system_config.h Outdated Show resolved Hide resolved

Nekotekina reviewed Aug 21, 2022

View reviewed changes

elad335 force-pushed the savestates branch from 89faffd to ce0a711 Compare August 21, 2022 10:48

elad335 added 6 commits August 21, 2022 13:53

SPU: add the concept of inaccurate reservations

4e5b8b2

Implement cellSpursRequestIdleSpu

PPU: Detect lwmutex reservation pattern and optimize it

beb9307

SPU: Implement custom reservation condition in atomic wait

eba27b5

SPU: Disable notification if no changes were made in PUTLLC

fcd1259

vm: Attempt to not remove cpu_flag::wait flag during access violation

ca4f97f

SPU Add ability to specify percentage of busy waiting

8c8c37b

elad335 force-pushed the savestates branch from ce0a711 to 8c8c37b Compare August 21, 2022 10:53

elad335 requested a review from Nekotekina August 21, 2022 10:54

Nekotekina merged commit bf63a18 into RPCS3:master Aug 21, 2022

solarmystic mentioned this pull request Aug 21, 2022

Regression: Severe performance regression from 0.0.23-14079 onwards (#12523) #12541

Closed

elad335 mentioned this pull request Aug 21, 2022

Fix GETLLAR spin detection and restore default configuration as before #12523 #12544

Merged

solarmystic mentioned this pull request Aug 22, 2022

Regression: Ratchet & Clank: Tools of Destruction [BCES00052] freezes after loading save files since 0.0.23-14079 (#12523) #12548

Closed

elad335 added a commit to elad335/rpcs3 that referenced this pull request Aug 22, 2022

Bugfix after RPCS3#12523

cd1ae29

Nekotekina pushed a commit that referenced this pull request Aug 22, 2022

Bugfix after #12523

9d8f4d5

JimScript mentioned this pull request Sep 5, 2022

[Regression] Windows 10 does not boot with PPU Recompiler after #12523 #12611

Closed

solarmystic mentioned this pull request Sep 11, 2022

[Regression] Driver: San Francisco [BLES00891] freezes after a short period of time driving from 0.0.23-14079 onwards #12640

Open

AniLeo mentioned this pull request Sep 25, 2022

Regression: Sky Fighter [NPUB30144] freezes when starting a level (#12544, #12523) #12642

Closed

jeremybris mentioned this pull request Oct 9, 2023

[Regression] Performance Regressions in Hot Shots Golf: Out of Bounds (#11904, #12523) #14724

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

SPU: Lock-line reservation optimizations + Savestates bugfix #12523

SPU: Lock-line reservation optimizations + Savestates bugfix #12523

elad335 commented Aug 18, 2022 •

edited

elad335 commented Aug 18, 2022

Nerboruto commented Aug 18, 2022 •

edited

elad335 commented Aug 18, 2022

Nerboruto commented Aug 18, 2022

This comment was marked as off-topic.

elad335 commented Aug 18, 2022 •

edited

Nerboruto commented Aug 18, 2022

elad335 commented Aug 18, 2022

This comment was marked as off-topic.

Nerboruto commented Aug 18, 2022 •

edited

elad335 commented Aug 20, 2022

Hasster1 commented Aug 20, 2022

Hasster1 commented Aug 20, 2022

Nekotekina Aug 21, 2022

elad335 Aug 21, 2022

Nekotekina Aug 21, 2022

elad335 Aug 21, 2022 •

edited

elad335 Aug 21, 2022 •

edited

Nekotekina Aug 21, 2022

Hasster1 commented Aug 21, 2022

Hasster1 commented Aug 21, 2022

SPU: Lock-line reservation optimizations + Savestates bugfix #12523

SPU: Lock-line reservation optimizations + Savestates bugfix #12523

Conversation

elad335 commented Aug 18, 2022 • edited

elad335 commented Aug 18, 2022

Nerboruto commented Aug 18, 2022 • edited

elad335 commented Aug 18, 2022

Nerboruto commented Aug 18, 2022

This comment was marked as off-topic.

elad335 commented Aug 18, 2022 • edited

Nerboruto commented Aug 18, 2022

elad335 commented Aug 18, 2022

This comment was marked as off-topic.

Nerboruto commented Aug 18, 2022 • edited

elad335 commented Aug 20, 2022

Hasster1 commented Aug 20, 2022

Hasster1 commented Aug 20, 2022

Nekotekina Aug 21, 2022

Choose a reason for hiding this comment

elad335 Aug 21, 2022

Choose a reason for hiding this comment

Nekotekina Aug 21, 2022

Choose a reason for hiding this comment

elad335 Aug 21, 2022 • edited

Choose a reason for hiding this comment

elad335 Aug 21, 2022 • edited

Choose a reason for hiding this comment

Nekotekina Aug 21, 2022

Choose a reason for hiding this comment

Hasster1 commented Aug 21, 2022

Hasster1 commented Aug 21, 2022

elad335 commented Aug 18, 2022 •

edited

Nerboruto commented Aug 18, 2022 •

edited

elad335 commented Aug 18, 2022 •

edited

Nerboruto commented Aug 18, 2022 •

edited

elad335 Aug 21, 2022 •

edited

elad335 Aug 21, 2022 •

edited