Work Is Underway To Upstream LLVM Clang's CUDA Toolchain For AMDGPU/HIP

[phoronix]

Phoronix: Work Is Underway To Upstream LLVM Clang's CUDA Toolchain For AMDGPU/HIP

A long available tool has been AMD's ROCm HIP that allows converting CUDA code to portable C++ code that in turn can be executed on Radeon GPUs. There is now work on getting the upstream LLVM Clang compiler's CUDA toolchain support to also support HIP...

http://www.phoronix.com/scan.php?page=news_item&px=LLVM-CUDA-To-AMD-HIP

[boxie]

This is rather cool!

[Meteorhead]

It's still something. You have to start somewhere if you want to break the CUDA haegemonia. SYCL is much more promising in this regard, and we advocate it wherever possible, but it's progressing far too slowly. I feel that we'll sooner have ISO C++ support for GPGPU sooner via std::executor other than any of these efforts. These projects are neccessary preliminary steps both for legacy code and for cooking the back-ends, but are not the ulitmate solution.

[boxie]

It's not. You still don't have a unified API set. It's just one more framework on the market.

sure, but you can take cuda code and make it run in places it could not before. I am sure you could even get some SPIR-V output later on too

[gsedej]

Is there any possibility that existing binnary cuda apps will in future use radeon gpus via hip?

[bachchain]

Is there any possibility that existing binnary cuda apps will in future use radeon gpus via hip?

No. That's completely out of the project's scope.

[Jumbotron]

"There is now work on getting the upstream LLVM Clang compiler's CUDA toolchain support to also support HIP."

HIP, HIP, Hooray !! (Sorry...had to do it.)

[bridgman]

It's not. You still don't have a unified API set. It's just one more framework on the market.

I don't understand. You can take HIP output and run it through either NVidia or AMD compilers - the whole idea is single framework based on C++17.

[dwagner]

Call me a pessimist, but this sounds rather like a proof-of-concept hacking project than like something that could actually see any relevant use in the future. I would sure like to see some non-proprietary GPU programming standard to gain traction, but a CUDA-to-something translator does not seem to help that much, IMHO.

[bridgman]

I remember they are still trying to port Tensorflow to HIP, and the worked version is 1.3 while the CUDA version is 1.7 now(You can find HIP port in Rocm application repository hosted on GitHub). So I dare say that HIP does not work in the way as you described, at least not that perfect, since even with human intervention, it's still not catching up.

Actually no - the CUDA to HIP/HCC part went pretty smoothly. What took the time (about 7 months between Tensorflow 1.3 and 1.7) was MIOpen, an open source replacement for cuDNN.

That's a one-ish-time thing (one time plus ongoing support & enhancement/optimization).