Performance regression in recent version (April 2026)

[metchebe]

Hello all.

I noticed a big performance regression using build 1275bd7 (April 2026) on my Raspberry Pi 3 (Lakka-5.x aarch64).

The only recent significant change is PR #82, which I suspect is causing this. My question is, is a performance regression for this expected? Is there an option that I can change?

Going back to 8f837eb (Oct 2024) restores performance.

Thanks.

[hizzlekizzle]

It certainly doesn't affect performance on any non-potato, but I guess increasing the framebuffer size is too much for an RPi3.

We can probably make the FB width macro check against the core option. Are you able to compile and test if I give you a modified libretro.cpp?

[metchebe]

Thanks for looking into it. Yes, I can test.

[hizzlekizzle]

Okay, it seems we can keep the framebuffer small as long as we're okay with the 3D mode being a "requires restart" option. Otherwise, our screen gets all screwy because the stride and memory allocation, etc. aren't able to update accordingly until there's a full teardown and rebuild.

This should get you even better performance than before, since it makes the framebuffer a flat 384 with anaglyph mode and then doubles it for the other modes and adds 256 for side-by-side (to accommodate the gap setting): https://pastebin.com/8E86A4vF

[metchebe]

Thanks!

Builds OK, tested briefly for a couple of minutes with Mario Clash and VB Wario Land and indeed it has a slight performance bump over 2024 versions, around 2-5 FPS more when fast-forwarding.

With this the core is now usable again on the Pi 3.