OpenGL Shader Cache Support For RadeonSI Is Making Progress

[xxmitsu]

tarceri ,

Are the shaders cached after all the optimizations took place ?

If yes, would this allow something like previously cached shaders optimized with NIR for intel backend, to be consumed on a subsequent run by the amdgpu backend .. just to simulate if some optimizations would bring benefit for radeon ?

Or this wouldn't be possible because NIR is just an intermediate representation and the actual optimizations happen in the backend if backend chooses them with NIR_PASS_V or whatever the recommended method is..

Thank you in advance for the response!

[ResponseWriter]

Is this coming to AMDGPU as well? Playing Rocket League with an RX480 can get pretty annoying when the menus hang the entire game for 3,5, even 10+ seconds at a time under the Unreal Engine.

The issue in Rocket League is due to a different issue (not shader compilation) that may now be fixed. See https://bugs.freedesktop.org/show_bug.cgi?id=97879

Let's hope that update doesn't take too long.

tarceri ,

Are the shaders cached after all the optimizations took place ?

If yes, would this allow something like previously cached shaders optimized with NIR for intel backend, to be consumed on a subsequent run by the amdgpu backend .. just to simulate if some optimizations would bring benefit for radeon ?

Or this wouldn't be possible because NIR is just an intermediate representation and the actual optimizations happen in the backend if backend chooses them with NIR_PASS_V or whatever the recommended method is..

Thank you in advance for the response!

Radeonsi will not be able to make use of shaders compiled for Intel, and vice versa, plus people are more likely to change Mesa version and recompile the whole set of shaders more than they'll change GPU vendor on the same machine.

Edit: I made a mistake in that it's cached at a higher level than I thought but still fairly specific to the hardware. Intel and AMD both use different IR so it would need to be compiled and cached independently for both.