AMD Rebases Their OpenMP For Radeon GPUs Against LLVM 11

[phoronix]

Phoronix: AMD Rebases Their OpenMP For Radeon GPUs Against LLVM 11

At the end of last year with ROCm 3.0 AMD introduced the AOMP compiler for OpenMP support targeting Radeon GPUs. AOMP is another downstream of LLVM Clang and on Tuesday marked the latest update...

http://www.phoronix.com/scan.php?page=news_item&px=AOMP-OpenMP-LLVM-11

[boxie]

I am idly wondering how much work would be required to make ROCm a oneAPI level 0 compliant?

[bridgman]

I am idly wondering how much work would be required to make ROCm a oneAPI level 0 compliant?

Or make OneAPI ROCm compliant

[jrch2k8]

Honestly either way doesn't matter for the FOSS crowd because ROCm requires 3 Ph'd including chemistry and physics just to get a clue on how to build it just to realize it just don't work properly cuz you missed a kernel patch 125 steps ago and you were using the wrong Clang tree the whole time and oneAPI i'm not really sure how much use could get with suck weak GPU's from Intel assuming XE ever launches

For RH closed source customers i guess it is functional with the driver install but haven't even checked to be sure.

[Linuxhippy]

Too bad nothing of this / the whole AMD open-source compute stack finds its way into downstream distributions.
AMD has a really powerful open-source ecosystem, however as-is it seems to be targeted simply at compute focused data centeres where installation-overhead simply does not matter - the usual desktop user won't benefit at al.

[Shevchen]

Honestly either way doesn't matter for the FOSS crowd because ROCm requires 3 Ph'd including chemistry and physics just to get a clue on how to build it just to realize it just don't work properly cuz you missed a kernel patch 125 steps ago and you were using the wrong Clang tree the whole time and oneAPI

Sadly - this. Being able to execute CUDA code on AMD GPUs would bring a lot of people from the fence to finally get an AMD GPU. Hire a software developer or two and get this thing out for the open source community in a way that normal people get it up and running. (And honestly, I don't think AMD has cash problems right now. While there are software devs needed on every front, running CUDA on AMD GPUs is a critical part)

I'm pretty sure that the community will come up with some optimizations to increase the speed of the translation layer as it did with other projects. I see no reason why it isn't open source - everyone would win (aside Nvidia). Get some peeps, if needed even under NDA and get this thing up an running. After that we can explore how to optimize it and kill the roadblockers as a community.

[darkbasic]

I'll never understand why AMD doesn't care about ROCm being easy to compile/distribute. Almost no one uses it because of that, which is a true pity.

[ms178]

What about merging the AOMP with the AOCC efforts? I get it why they experiment with new functionality in a downstream fork but I don't get it why they keep these efforts separate from each other. From a user perspective, I'd like to use all advanced features with a single compiler (the optimum would be to have these upstream).

[FireBurn]

I'll never understand why AMD doesn't care about ROCm being easy to compile/distribute. Almost no one uses it because of that, which is a true pity.

I actually got it compiled once, submitted a pull request to get it working with split llvm/clang libraries (which was accepted quickly) only to discover it wouldn't work right on Raven, and isn't even supported on Tonga

All I was planning on running was clinfo

[darkbasic]

I actually got it compiled once, submitted a pull request to get it working with split llvm/clang libraries (which was accepted quickly) only to discover it wouldn't work right on Raven, and isn't even supported on Tonga

All I was planning on running was clinfo

I've never been so glad that someone coded radv, otherwise we would have been in the very same spot for Vulkan as well.