Lightning+Lightrec updates and other optimizations #862

ZachCook · 2022-09-25T02:28:53Z

Split out cpu_lightrec.cpp from cpu.cpp instead of being mashed together.

Added hugepage, codebuffer, mmap at 0x0, and SP GP hit RAM optimizations and fallbacks, see commits for more details.

Use lightrec's GTE registers directly when not using beetle interpreter (while maintaining backwards compatibility with savestates).

Removed the Cycle Timing Check dynarec option.

lgtm-com · 2022-09-25T02:43:46Z

This pull request introduces 1 alert when merging 8ae372a into bd6b9ef - view on LGTM.com

new alerts:

1 for Missing return statement

pcercuei · 2022-09-25T16:10:15Z

Makefile.common

@@ -375,7 +376,8 @@ ifeq ($(HAVE_LIGHTREC), 1)
                  $(DEPS_DIR)/lightrec/memmanager.c \
                  $(DEPS_DIR)/lightrec/optimizer.c \
                  $(DEPS_DIR)/lightrec/reaper.c \
-                  $(DEPS_DIR)/lightrec/regcache.c
+                  $(DEPS_DIR)/lightrec/regcache.c \
+                  $(DEPS_DIR)/lightrec/tlsf/tlsf.c


You shouldn't have to compile TLSF since you disabled the code buffer support (ENABLE_CODE_BUFFER=0)

ENABLE_CODE_BUFFER=1 was set in 3012b97 of this PR, there were many changes squashed together, and it could maybe use even more to make this more clear maybe?

Yes I saw it after my message. I guess this change should be squashed into 3012b97 then.

Ryunam · 2022-09-29T14:28:11Z

After testing this with a few games on my Windows 11 system, I have noticed a few Runahead-related issues:

As reported on Discord, enabling Runahead in RetroArch with either single instance or secondary instance mode and having the dynarec active seems to be triggering unpredictable crashes with this PR, at several spots. This is with the same values/settings used with current master. Also, the amount of dynarec cycles do not appear to affect the likelihood of the crashes.
Based on my testings so far, the crashes can occur either at startup, especially if one fastforwards right after launching the content, or when loading a save. The latter in particular is being triggered here consistently whenever I try to load my "Akumajou Dracula X: Gekka no Yasoukyoku" save from the memory card screen with Runahead active, which worked normally before this PR. Again, fastforwarding seems to greatly increase the chance of RetroArch and the core suddenly crashing with this setup.
If Dynarec SPU Samples is set to any value above 1 and Runahead is enabled with Single Instance mode, the audio will get noticeably choppy and distorted. Disabling Runahead or setting it to Second Instance Mode makes it sound normal again.

lgtm-com · 2022-10-01T02:00:05Z

This pull request introduces 1 alert when merging 1e71531 into bd6b9ef - view on LGTM.com

new alerts:

1 for Missing return statement

LibretroAdmin · 2022-10-01T03:05:12Z

@Ryunam Hi, are your issues addressed now?

Ryunam · 2022-10-01T11:47:23Z

Okay, I have performed a few more tests after the last commit:

The audio issue caused by the combination of SPU Samples set to a value higher than 1 and Runhead Single Instance mode seems to be fixed. I tried switching the values back and forth during gameplay and everything appears to sound normally.
However, the Dynarec SPU Samples by itself (with or without Runahead) seems not to play nicely with a few games. "Final Fantasy" (the PS1 remake of FF1) stood out during my testing, as any value higher than 1 triggers missing and sustained notes in the music of that game. A clear example can be seen by starting the game, then immediately entering the first city, Corneria. Set the samples to 16 and the music will start having missing parts or oddly prolonged notes. Setting the Samples value back to 1 makes the music play normally again. I assume this is partly expected behavior, as the feature in question might not work well with certain games depending on how the music files are streamed, but i figure it's worth pointing it out.
~~Interestingly, the crashes do not seem to happen anymore either! I have tried to trigger them multiple times and so far so good.~~

ZachCook · 2022-10-01T14:36:50Z

@Ryunam
I'm surprised the crashes disappeared, though I couldn't reproduce so just let me know if they come back.

The SPU Samples option was expected to cause that, I'll add a note in the description about it, it's basically just running the sampling at a lower frequency and then running multiple times to catch up, it's a performance hack, I'm surprised it works as well as it does. Though now that you found a game I can test I might be able to improve it further by doing some run time detection or tweaking the timing code, but that may be a bigger project for the future.

Ryunam · 2022-10-01T16:08:28Z

Sorry @ZachCook, I stand corrected. The crashes appeared once again while I was testing other games, so they seem to be unrelated to the SPU changes.
They are also very unpredictable: sometimes the affected games will launch normally and then crash when loading a save, sometimes the crash will occur almost immediately at startup, some other times they won't happen at all.

One consistent situation where I can reproduce the crash right now is with the Japanese PS1 port of Street Fighter Zero 2:

Under Core Options -> Emulation Hacks, set Skip BIOS to ON.
Have Runahead enabled with Second Instance ON in a per-core override, to make it activate automatically upon launching the game.
The amount of Runahead frames does not seem to matter. I'm using the default value of 1 in this case.
This was tested with both the glcore and Vulkan video drivers in RA.
Launch Street Fighter Zero 2 and hold the Fastforward button right after starting the game. One thing that may also contribute to the crash being triggered is pressing/mashing the Circle button while it's fastforwarding, but this was not verified.
On my system the game and RA will crash either after the Capcom logo animation, or later on when selecting an option from the main menu (for example, "Versus Mode").

The same core as compiled from current master, before this PR, never triggers any crash with these same steps and/or settings.

LibretroAdmin · 2022-11-14T21:44:45Z

Hi there, what is the state of this PR as it stands? Are there any roadblocks?

ZachCook · 2022-11-14T23:21:49Z

Hi there, what is the state of this PR as it stands? Are there any roadblocks?

I'll update it again to latest lightrec+lightning, and see if I can reproduce the above mentioned crash, or at least narrow down what triggers it, I wouldn't really want to merge this until that is dealt with.

ZachCook · 2022-11-17T21:49:48Z

Updated to recently released Lightning 2.2.0 and Lightrec 0.6

Still working on crashes with runahead and loading, the hack I had there (restart entire dynarec and switch to interpreter for a while) looks awfully suspect and I'm working on a proper fix to at least rule that out as a cause if it doesn't just fix it.

lgtm-com · 2022-11-17T21:59:06Z

This pull request introduces 1 alert and fixes 1 when merging c41b8a7 into f8a904e - view on LGTM.com

new alerts:

1 for FIXME comment

fixed alerts:

1 for FIXME comment

Heads-up: LGTM.com's PR analysis will be disabled on the 5th of December, and LGTM.com will be shut down ⏻ completely on the 16th of December 2022. Please enable GitHub code scanning, which uses the same CodeQL engine ⚙️ that powers LGTM.com. For more information, please check out our post on the GitHub blog.

LibretroAdmin · 2022-11-19T14:08:42Z

@ZachCook Is this ready for merging now? Will this have any known regressions vs. the current state of things?

ZachCook · 2022-11-19T22:20:28Z

@ZachCook Is this ready for merging now? Will this have any known regressions vs. the current state of things?

@LibretroAdmin It is not ready yet, I am seeing lots of crashing now than can be triggered by any kind of run-ahead or loading a save state.

You could apply just 8814bcb for now, it is quite a useful improvement by itself.

LibretroAdmin · 2022-11-20T21:00:13Z

OK, I cherry-picked that commit for now and pushed it to master

ZachCook · 2024-09-16T00:23:32Z

Still have to re-fix the runahead/savestate issue right after bios since I ripped out the previous next_interpreter hack, but otherwise I have tried to update and clean up the code a bit.

I think some of the older commits are a bit messy/WIP that are fixed later in the series since there was such a long period of time since I last was working on this, so it might be easier to view the changes as a whole for some parts in cpu/cpu_lightrec.

Fix bios freeze with lightrec, AssertIRQ must change lightrec's CAUSE too cleanup unused cop0 functions

Codebuffer is much faster due to lightrec's 32-bit LUT optimization Moved lightning.h to lightning-actual.h and made lightning.h include lightning-actual.h and define HAVE_MMAP for lightning, instead of defining in the Makefile which applied it to other parts of beetle, causing issues on Windows in libretro-common/vfs

…uilds

This allows less code to be emitted by lightrec, but will only be used if root user enables it with sysctl -w vm.mmap_min_addr=0

…htrec api This was only useful for debugging, users don't benefit from it, the config will fallback to using the beetle interpreter if it was selected The lightrec execute_one function this option called was removed, for equivalent debugging pass current timestamp instead of next_event_ts to lightrec_execute Use lightrec cop2_notify for PGXP, update interpreter to new api

subrepo: subdir: "deps/lightning" merged: "edc57df2" upstream: origin: "https://git.savannah.gnu.org/git/lightning.git" branch: "master" commit: "bfd695a9" git-subrepo: version: "0.4.9" origin: "https://github.com/ingydotnet/git-subrepo" commit: "ea10886"

subrepo: subdir: "deps/lightrec" merged: "ea20362c" upstream: origin: "https://github.com/pcercuei/lightrec.git" branch: "master" commit: "ea20362c" git-subrepo: version: "0.4.9" origin: "https://github.com/ingydotnet/git-subrepo" commit: "ea10886"

tweak lightrec startup and syncing stuff remove next_interpreter hack tweak setoptions, should be safer

Disabled by default for max compatibility Enabling should generate faster code for most normal games that do not have sp and gp registers pointing outside of RAM

This was broken by cpu.cpp and cpu_lightrec.cpp being split apart

Add breakpoint handling to match interpreter Fix setting of options to lightrec

Replaces supported_io_bases array with min/inc/max_io_base Makes it easier to avoid ranges of address for different platforms Fix unmap of lightrec codebuffer

pcercuei reviewed Sep 25, 2022

View reviewed changes

LibretroAdmin marked this pull request as ready for review September 25, 2022 18:29

ZachCook force-pushed the libretro-dynarec-backup branch from 1e71531 to c41b8a7 Compare November 17, 2022 21:41

ZachCook force-pushed the libretro-dynarec-backup branch from c41b8a7 to 1f36680 Compare September 16, 2024 00:14

ZachCook marked this pull request as draft September 16, 2024 00:16

ZachCook added 10 commits September 16, 2024 22:46

lightrec: update include files and folder

e328824

lightrec: Adjust to lightrec API changes and HAVE_MMAP for lightning

05f51c2

Fix bios freeze with lightrec, AssertIRQ must change lightrec's CAUSE too cleanup unused cop0 functions

Change GTE to allow using lightrec's registers

39f5b8a

lightrec: attempt memfd and hugepages for ram mmap

3f7d110

add cpu_timestamp hack to so lightrec tracks more cycles

ba5d04a

improve lightrec debugging by allowing threaded recompiler in debug b…

778c1b6

…uilds

clean and optimize loop around lightrec's execution functions

d34dea9

support MainRAM at 0x0, use INVALID_PTR instead of NULL

8a86773

This allows less code to be emitted by lightrec, but will only be used if root user enables it with sysctl -w vm.mmap_min_addr=0

ZachCook added 13 commits September 16, 2024 22:46

Fix Windows and cleanup error handling in init_mmap

9a38b02

Update lightning and lightrec include files

3541cf9

Makefile.common: Add missing files for lightning lightrec update

fd88631

Extraction of lightrec specific code to cpu_lightrec.cpp

59a2e9e

tweak lightrec startup and syncing stuff remove next_interpreter hack tweak setoptions, should be safer

Fix build with lightrec update

972b562

Add Dynarec SP GP hit RAM optimization option

75d2a4b

Disabled by default for max compatibility Enabling should generate faster code for most normal games that do not have sp and gp registers pointing outside of RAM

Fix GTE bugs due to uninitialized lightrec_regs

7b88a39

This was broken by cpu.cpp and cpu_lightrec.cpp being split apart

Make cpu_lightrec friend class to cpu

472493e

Add breakpoint handling to match interpreter Fix setting of options to lightrec

Add Parrel Port Mirror support for PS2/PS3 BIOS

98d42a5

Fix PGXP for LWU/SWU meta-opcodes in lightrec

b1e476e

Remove kunseg for pgxp rw, since lightrec_rw does that now

2bf0f5e

ZachCook force-pushed the libretro-dynarec-backup branch 2 times, most recently from ea881ad to c3a4eaa Compare September 18, 2024 01:12

ZachCook added 2 commits September 27, 2024 04:19

lightrec_init_mmap: Clean and rewrite io_base mmap code

272fb20

Replaces supported_io_bases array with min/inc/max_io_base Makes it easier to avoid ranges of address for different platforms Fix unmap of lightrec codebuffer

Fix libretro.cpp with HAVE_LIGHTREC=0

a38d0ae

ZachCook force-pushed the libretro-dynarec-backup branch from cc3e61b to a38d0ae Compare September 27, 2024 08:36

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Lightning+Lightrec updates and other optimizations #862

Lightning+Lightrec updates and other optimizations #862

ZachCook commented Sep 25, 2022 •

edited

Loading

lgtm-com bot commented Sep 25, 2022

pcercuei Sep 25, 2022

ZachCook Sep 25, 2022

pcercuei Sep 25, 2022

Ryunam commented Sep 29, 2022

lgtm-com bot commented Oct 1, 2022

LibretroAdmin commented Oct 1, 2022

Ryunam commented Oct 1, 2022 •

edited

Loading

ZachCook commented Oct 1, 2022

Ryunam commented Oct 1, 2022 •

edited

Loading

LibretroAdmin commented Nov 14, 2022

ZachCook commented Nov 14, 2022

ZachCook commented Nov 17, 2022

lgtm-com bot commented Nov 17, 2022

LibretroAdmin commented Nov 19, 2022

ZachCook commented Nov 19, 2022

LibretroAdmin commented Nov 20, 2022

ZachCook commented Sep 16, 2024 •

edited

Loading

Lightning+Lightrec updates and other optimizations #862

Are you sure you want to change the base?

Lightning+Lightrec updates and other optimizations #862

Conversation

ZachCook commented Sep 25, 2022 • edited Loading

lgtm-com bot commented Sep 25, 2022

pcercuei Sep 25, 2022

Choose a reason for hiding this comment

ZachCook Sep 25, 2022

Choose a reason for hiding this comment

pcercuei Sep 25, 2022

Choose a reason for hiding this comment

Ryunam commented Sep 29, 2022

lgtm-com bot commented Oct 1, 2022

LibretroAdmin commented Oct 1, 2022

Ryunam commented Oct 1, 2022 • edited Loading

ZachCook commented Oct 1, 2022

Ryunam commented Oct 1, 2022 • edited Loading

LibretroAdmin commented Nov 14, 2022

ZachCook commented Nov 14, 2022

ZachCook commented Nov 17, 2022

lgtm-com bot commented Nov 17, 2022

LibretroAdmin commented Nov 19, 2022

ZachCook commented Nov 19, 2022

LibretroAdmin commented Nov 20, 2022

ZachCook commented Sep 16, 2024 • edited Loading

ZachCook commented Sep 25, 2022 •

edited

Loading

Ryunam commented Oct 1, 2022 •

edited

Loading

Ryunam commented Oct 1, 2022 •

edited

Loading

ZachCook commented Sep 16, 2024 •

edited

Loading