Renewed Interest In Background, Parallel Shader Compilation For Mesa

[phoronix]

Phoronix: Renewed Interest In Background, Parallel Shader Compilation For Mesa

Prolific Mesa contributor Ilia Mirkin has taken initial steps towards working on parallel shader compiles in Mesa...

http://www.phoronix.com/scan.php?page=news_item&px=Mesa-Parallel-Shader-Compiles

[cb88]

thread and memory safety? Write it in rust? seems like the obvious thing to suggest though it may or many not be right... Mesa already uses LLVM heavily... so I don't see why not.

[GabrielYYZ]

...he will pursue a design where only threading will be enabled when delivers set the ARB_parallel_shader_compile support and that initially will be off by default for users until all issues are worked out.

That is supposed to be drivers, yeah? couldn't come up with other words that made sense in that context.

[Michael]

That is supposed to be drivers, yeah? couldn't come up with other words that made sense in that context.

Yep, fixed. Thanks.

[zanny]

thread and memory safety? Write it in rust? seems like the obvious thing to suggest though it may or many not be right... Mesa already uses LLVM heavily... so I don't see why not.

All the current Mesa OpenGL function implementations are in C. Adding another language for just one extension (or even worse, to try to compartmentalize the C functions that compile shaders) adds a huge number of build dependences and build chain steps to make the Rust parts.

If you were starting an OpenGL implementation from scratch today, I do not think there is any other valid consideration besides Rust (cough, Vulkan). But when you already have a huge C codebase, mixing more languages into that can of course work, but its a lot of headaches.

Though honestly it might be less headaches than writing even one file of C in my experience. But that is why they wouldn't do it.

[cb88]

All the current Mesa OpenGL function implementations are in C. Adding another language for just one extension (or even worse, to try to compartmentalize the C functions that compile shaders) adds a huge number of build dependences and build chain steps to make the Rust parts.

If you were starting an OpenGL implementation from scratch today, I do not think there is any other valid consideration besides Rust (cough, Vulkan). But when you already have a huge C codebase, mixing more languages into that can of course work, but its a lot of headaches.

Though honestly it might be less headaches than writing even one file of C in my experience. But that is why they wouldn't do it.

Rust is designed to work with C and most of the dependencies are already required. https://doc.rust-lang.org/book/ffi.html

I agree in principle on the build dependencies issue though... it is nice to just install GCC and thats all you need.

[M@yeulC]

Thanks a lot for this, Ilia

[Adriannho]

Hmm ...just curious...what is the expected performance gain out of this (in percentage)?

[cb88]

Hmm ...just curious...what is the expected performance gain out of this (in percentage)?

I would imagine probably not much .... but also probably improved latency (less lag). And possibly faster startup/warmup of the rendering pipeline.

[pal666]

thread and memory safety? Write it in rust? seems like the obvious thing to suggest though it may or many not be right... Mesa already uses LLVM heavily... so I don't see why not.

you mean REwrite it in rust. noone is crazy enough to rewrite perfectly working software even if it is not as developer-starved as mesa. actually it is number one problem with all toy languages advertised as replacement for something