RetroArch shaders slower with vc4 than with bcrm

[vanfanel]

Hi, Eric

I am using VC4 as my videodriver for the Raspberry Pi 3, but I have noticed that the same shader is WAY slower on the VC4 driver than it is on the bcrm vendor driver (which I don't use anymore).
For example, the dedicated Raspberry Pi CTR shader here:
https://github.com/RetroPie/common-shaders/blob/vc4-fixes/shaders/crt-pi.glsl
This shader was fullspeed on any core on the bcrm driver, but it's SLOW on VC4.

Even this "simple" scanline shader
https://github.com/libretro/glsl-shaders/blob/master/misc/scanline.glsl
...isn't always fullspeed anymore. It's an 1-pass shader only.

Maybe there's some operation that's possible to optimize in order to get back the lost performance?
I am using recent upstream MESA, I believe all your VC4-related optimizations are there, aren't they?

[anholt]

I don't see anything in those shaders that should be unable to sample faster than refresh rate fullscreen. Without specific instructions on how to reproduce the problem, I (or anyone else helping) probably won't be able to help. For debugging on your own, I'd recommend taking a look at the docs at https://github.com/anholt/mesa/wiki/VC4-Performance-Debugging

[vanfanel]

I didn't explain the issue properly. The shaders themselves are fullspeed: but paired with CPU-demanding emulators (~70-80% CPU usage in one core, since emulation is not multithreaded), the Pi can't keep up with and fullspeed isn't possible. They ARE fullspeed with emulators or native games that don't use so much CPU time. The problem is that they are fullspeed always with the BCRM vendor driver, while they are not with your fantastic VC4 driver, which is the future Pi driver.

So, are shaders using CPU time, or maybe it could be due to memory bandwidth limitations?

[anholt]

That link is precisely the thing for figuring out what's going on with performance, instead of speculating about what might be going on.