Valve Has Been Developing A New Mesa Vulkan Shader Compiler For Radeon

[R41N3R]

Yeah I'd be lapping this up too if I actually had an amd card, but I'm a poorfag couldn't even afford the radeon vii, in fact I don't even have a desktop to put it into anymore, what is going on with my life?!

The interesting part for me is the technology itself, even if I can afford now to get some of the latest AMD hardware, it doesn't mean I'm using it to all extend, not really a gamer, but sure I like to test stuff like this. Wish you all the best.

[Peter Fodrek]

So for SI+ GPUs we now have:
1. AMD proprietary shader compiler used by proprietary OpenGL driver
2. AMD proprietary shader compiler used by proprietary Vulkan driver
3. LLVM-based compiler (modified AMDGPU backend?) used by AMDVLK
4. LLVM-based compiler (AMDGPU backend) used by radeonsi and RADV
5. AOC for RADV

Not bad but not perfect. As I see when we have 5 compilers and 3 Vulkan and 2 OpenGL drivers, then there may be 5 times 5 combinations or 5 combinations per driver to be allowed. Driver may do compilation with all 5 compilers and use result first finished compilation, or do simulation of performance of all result for all application shaders at application start or installation and use best performing shader compilation with compilation time mentioned.

Or do it even better. Load all Radeon drivers and switch them on the fly for best performance case.


Combination of both is too optimistic or too silly for me, but maybe baby.


Too ambitious is to use HBCC and asynchronous shaders to move most of the driver parts into compute shaders, as well.



But these ideas are only dreams, that will not become reality, i guess.

[woife]

I made a PPA for Ubuntu, please help me test it! https://launchpad.net/~ernstp/+archive/ubuntu/mesaaco

Thanks!

[smitty3268]

This is fascinating. I can't help but be a little pessimistic about whether or not Valve will be able to keep up with new hardware coming out in reasonable timeframes, but still this is very cool to see.

I think the main thing this is pointing out is that there are some fundamental problems with using LLVM to compile graphics code right now, and it's probably going to be really hard to fix them. LLVM was obviously designed around a different model entirely.

I really wonder what AMD thinks of this project. Are they watching it closely to see how it works out, or are they locked 100% into LLVM and ignoring everything outside that?

This makes a significant improvement for Guild Wars 2 with D9VK; with LLVM, the shader compile is horrible (full stops for up to 1 second). With ACO, I don't notice any stuttering.

Here is a quick Copr I threw together for Fedora 30: https://copr.fedorainfracloud.org/co...a-git-acotest/

[pal666]

3rd party does AMD's job.

high-risk experiments aren't amd's job

[andrei_me]

Remember Ubuntu's Drama? Steam have an official AUR ready to test ACO but no PPA at the time of writing

[puleglot]

could this eventually also be used for radeonSI?

Q: Is it radv-only or will RadeonSI work?

A: radv for now, but we intend to look at RadeonSI once things are farther along.

https://steamcommunity.com/app/221410/discussions/0/1640915206474070669/

[davidbepo]

could this eventually also be used for radeonSI?

[ernstp]

I made a PPA for Ubuntu, please help me test it! https://launchpad.net/~ernstp/+archive/ubuntu/mesaaco