Arceri Is Still Working To Improve GLSL Compile Times For Mesa

[phoronix]

Phoronix: Arceri Is Still Working To Improve GLSL Compile Times For Mesa

For those wondering what Valve Linux developer Timothy Arceri has been working on with KHR_no_error settling down and the like, he still is striving towards faster GLSL compilation times to in turn speed up Linux game loading performance...

http://www.phoronix.com/scan.php?page=news_item&px=Arceri-Faster-GLSL-Compile

[duby229]

Cool work. I don't personally consider load times to be that big of an issue now days, but I seriously expect these compiler enhancements to improve runtime performance too. Lighter always ends up being faster. Good job.

[shmerl]

Shader compilation can benefit a lot from being multithreaded.

[microcode]

Cool work. I don't personally consider load times to be that big of an issue now days, but I seriously expect these compiler enhancements to improve runtime performance too. Lighter always ends up being faster. Good job.

It's not so much about load times, shader compiles often happen in the middle of play/operation, and even if managed well (by the game) they can delay a frame substantially.

Shader compilation can benefit a lot from being multithreaded.

That's not always true, even assuming you mean that it would run asynchronously until a draw. If you mean multithreading within a given shader compile, you're probably not going to see any benefit from that, at least not without a lot of work.

[shmerl]

That's not always true, even assuming you mean that it would run asynchronously until a draw. If you mean multithreading within a given shader compile, you're probably not going to see any benefit from that, at least not without a lot of work.

Take something like Trine 3, which takes some ridiculous time on startup compiling shaders with Mesa, while Nvidia blob does it way way faster. I suppose it's because all that it heavily parallelized there.

[Masush5]

It's not so much about load times, shader compiles often happen in the middle of play/operation, and even if managed well (by the game) they can delay a frame substantially.


That's not always true, even assuming you mean that it would run asynchronously until a draw. If you mean multithreading within a given shader compile, you're probably not going to see any benefit from that, at least not without a lot of work.

Well radeonsi has a shader cache now, so shader compilation speed should be pretty much irrelevant once the cache is warmed up.
Still, doing the same thing faster is always appreciated.

[tomtomme]

Take something like Trine 3, which takes some ridiculous time on startup compiling shaders with Mesa, while Nvidia blob does it way way faster. I suppose it's because all that it heavily parallelized there.

mesas shader compiler is threaded since 1 or 2 years already...

[shmerl]

mesas shader compiler is threaded since 1 or 2 years already...

Not on every step as far as I know (there is frontend, backend and etc.).

[geearf]

Well radeonsi has a shader cache now, so shader compilation speed should be pretty much irrelevant once the cache is warmed up.
Still, doing the same thing faster is always appreciated.

It already had one before, just in memory not on disk.

[microcode]

Well radeonsi has a shader cache now, so shader compilation speed should be pretty much irrelevant once the cache is warmed up.
Still, doing the same thing faster is always appreciated.

Games sometimes generate new shaders at runtime now, who knows. I think the baseline shader compile speed is still very important, even if we're talking about a Vulkan driver. It has to finish compiling the shaders at some point, sooner seems better than later.