Defer GX pipeline UID cache writes in libretro

[rtomasa]

Summary

Several games experience frequent brief stalls, sometimes severe enough to make them unplayable. Each newly discovered GX pipeline UID is currently written to disk from the draw path, making rendering sensitive to storage delays.

In the libretro core, this change queues new UIDs in memory and writes them together when the cache closes on reload or shutdown.

Following the PR rules, I want to disclose that this issue was identified during an AI review of the draw path that I requested. I reviewed the code myself, and other users and I tested the change.

Testing

• In Resident Evil Zero on RePlay with Vulkan, the first launch logged Read 0 pipeline UIDs. After a clean unload, the next launch logged Read 6 pipeline UIDs from GBZP08.uidcache, confirming the entries were saved and reloaded.
• Several testers reported substantially smoother gameplay in Resident Evil Zero and other games.

[cscd98]

Thanks for the PR! Is this something that can be upstream to dolphin-emu?

[rtomasa]

I guarded the code with #ifdef LIBRETRO because of your previous recommendation of avoiding any change on non-libretro code, but definitely this is something that could be added upstream.
Standalone emulator uses buffered writes, so probably it is not so heavily affected because in libretro this operation is passed to the VFS implementation that normally perform immediate operations in the same thread.
That said, this will help removing any stall also in the standalone if any. The only tradeoff of this change is that you are not writing buffered UID that are in RAM in case that the emulator/core crashes, but this should be totally fine.

[cscd98]

Quite possibly this is due to me implementing VFS for everything and not specific to ISO loading 👍

Could you restore the comments? I try to keep this repo as close to upstream as possible.

[rtomasa]

Using VFS is definitely the right thing to do, but the original code was already writing to disk from the draw path. Cached or not, I think deferring those writes is a much better approach. The improvement is pretty noticeable.

BTW, I’ve already restored the original comments.