Read ahead input lag feature and sound issues

[Tatsuya79]

The new read ahead feature in the Frame Throttle menu can work with some games and cause issues with some other ones in fbalpha.

Viewpoint or Armed Police Batrider will have messed up sound, mainly the bgm.

Is there anything possible to do about that? Cleaning up the save state or something?

Perhaps @Dwedit knows?

[barbudreadmon]

Hmm is that related to high cpu usage ? does the feature use the retro_serialize methods ?

[Tatsuya79]

It uses states yes, can't do without them. Not a CPU limit in my tests.

[barbudreadmon]

netplay/rewind working properly with those games ? did you run them at normal cpu clock ?

[furiadeoso]

Hi! I've just spent some time doing netplay in Armed Police Batrider and all ok. I hope that helps. Tell me if you need something more

[Tatsuya79]

The problem happens at normal clock. I don't do much netplay and never use rewind so I can't really say.

The problem could be that the save states need to be cleaned of any sound information? I think that's what Dwedit was doing in several cores starting with snes9x. That can be complicated in fba if every game has its particularities.

[barbudreadmon]

@Tatsuya79 well, 2 out of 3 features relying on retro_serialize methods work properly with sound information included, my guess is the read ahead feature is early beta and need some fix.

Removing sound data from the savestates in fba is not a realistic solution : it could affect other features and would require large amount of work on a per system basis.

[barbudreadmon]

I re-enabled USE_SPEEDHACKS in the neogeo driver, this code should be safe since https://github.com/libretro/fbalpha/commit/980171e5efff1730588c8a2d156f894b7655253b, is actually enabled upstream, and it seems to introduce something in the savestates related to sound. @Tatsuya79 let me know how it affects viewpoint and other neogeo games. @furiadeoso perhaps trying netplay after this change would be a nice idea too.

[furiadeoso]

Ok. I'll do ASAP, but certainly I'm very busy today. Thanks.

[furiadeoso]

I've just tested one minute metal slug in cross-play win64-android32 and seems to be ok.

[Tatsuya79]

Sadly it doesn't help for read ahead, same stuck notes in Viewpoint music.

[furiadeoso]

@Tatsuya79 Have you identified any FBA game or driver working well with read ahead? And what settings have you used to notice the effect? I would like to test it in netplay. Surely it will have undesirable effects, but I'm curious.

[Tatsuya79]

Dodonpachi works fine until the end, I remember playing some espgaluda too. I used "second instance" and 1 frame for both games.

[Tatsuya79]

Run Ahead isn't working at all since I updated the core to 0.2.97.43. The message "Core doesn't support save states, Run Ahead disabled." is always shown while save states are still working.

[barbudreadmon]

Well, if the core support save states and the feature says the core doesn't support save states, that doesn't seem like an issue in the core.

[furiadeoso]

@Tatsuya79 Have you been able to find out what happened?

[Dwedit]

Also how do I make MSVC builds? Would like a DLL with PDB file, don't care if it's my own build or not.

I'm seeing memory leaks.

[barbudreadmon]

@Dwedit https://github.com/libretro/libretro-super/tree/master/recipes/windows No idea how you use this since i'm a gcc user.

[Dwedit]

At least any idea how to build a non-stripped binary?

[barbudreadmon]

In gcc i build it with -O0 -g, toggleable by adding DEBUG=1 to make command.

[Tatsuya79]

@furiadeoso Yes, I think run ahead was defaulting to 1 instance mode. That would explain the sound issues. I opened this one in RA issues about the most recent problem (sorry if that's too many).

[barbudreadmon]

What's the status of this ?

[Tatsuya79]

Unchanged.

It works only if you start run ahead in 1 instance mode (which causes sound issues with many games) and slows the emulation down after a while (after several minutes of gameplay with hard sync 0 + some frame delay, I have to remove the frame delay, then the hard sync).

I tried hacking through runahead.cpp to force 2 instances:

When I'm in a game and activate it, I still hear the sound and gameplay of the 1st core but the picture is the one of the 2nd core booting the game from the beginning instead of loading the state.

[FrozenFish24]

As discussed in this issue, a fix for this problem is to make retro_serialize and retro_unserialize not dependent on a prior call to retro_serialize_size having happened, allowing secondary core mode to work. This could be done by removing the if statements in each function, or performing retro_serialize_size as part of each function. Opinions?

[barbudreadmon]

@FrozenFish24 Commited, let me know if it works.

[FrozenFish24]

Yep, working properly. Thanks @barbudreadmon!

[Tatsuya79]

Yes that works thanks! I have to see if that still slows down after a while now... but that would be a new issue. Closing this one.