LLVM Merges Initial Support For OpenMP Kernel Language

[phoronix]

Phoronix: LLVM Merges Initial Support For OpenMP Kernel Language

Merged to LLVM 18 Git yesterday was the initial support for the OpenMP kernel language, an effort around having performance portable GPU codes as an alternative to the likes of the proprietary CUDA...

https://www.phoronix.com/news/LLVM-Lands-OpenMP-Kernel-Lang

[AlanTuring69]

Sad to see the open source community embracing EEE methodology

[V1tol]

Another one computing language. What industry needs - OpenCUDA which is 100% source compatible and can translate/compile to something else - OpenCL, Vulkan, doesn't matter. Ideally we need ability to translate CUDA binaries into anything else without source code. And fuck novideo if they try to bark something about this - they already did massive damage to the industry.

[ddriver]

I was under the impression people use cuda because of the libraries nvidia is spoon-feeding devs along with it.

Modern day gpu programmers appear to be terrified by the very prospect of writing their own kernels.

Translating binaries may also not be the silver bullet people are hoping for, as nvidia is likely heavy on architecture specific optimizations, meaning automatically translated code is likely to perform way below the optimal.

Take the M1 translated x86 binaries, which exhibit none of the high performance or efficiency of native binaries.

[OneTimeShot]

Pretty sure OpenMP predates CUDA by a decade or so…. It’s been in GCC for a while.

[geerge]

It's openmp, so that means kernels can be written in c/c++/fortran? With the exception of fortran it sounds pretty similar to working with opencl, except openmp deals hopefully more cleanly with threading and memory management etc? How is it noticeably better than opencl?

[Eirikr1848]

Does this mean with parallel OpenMP one could use the 12GB Quadro K6000 with an 8GB Radeon R9 390X?

The K6000 is on a lifeline with the 470 driver series and the R9 card is no longer supported in ROCM.

So: neither will receive a CUDA or ROCM/HIP update in the future and this could allow all that VRAM to continue to be used in a post-vendor-support situation.

If this works well: do we forecast vendors dropping driver support sooner? Or keeping it around just as long?

[toughy]

But isn't OpenCL meant to offer portable performance ? Which is what OpenMP here promises

[RealNC]

At school, they taught us CUDA. Everything else was viewed more or less like a joke. I'm not sure how easy it would be for this to change. CUDA is basically viewed as "what you use to actually get the job done."