HIPCL Lets CUDA Run On OpenCL+SPIR-V

[phoronix]

Phoronix: HIPCL Lets CUDA Run On OpenCL+SPIR-V

Based off AMD's GPUOpen HIP as part of their ROCm stack, researchers at Tampere University in Finland have created HIPCL as leveraging HIP as well as POCL for routing CUDA codes to run on any hardware supporting OpenCL+SPIR-V...

http://www.phoronix.com/scan.php?page=news_item&px=HIPCL-CUDA-Over-OpenCL-SPIR-V

[tildearrow]

Congrats researching team! As an AMD user this effort will be much appreciated.

*insert Torvalds showing middle finger to NVIDIA here*

[Ipkh]

Now now, even Linus doesn't do that anymore....
I'm not sure how this helps, surely efforts are better spent recompiling or something similar. Certainly the implicit performance reduction due to translation will make AMD and Intel gpus too slow vs Nvidia cards. Its pretty much an OpenCL and CUDA workd with Nvidia ruling the roost by default bu supporting both.

[Meteorhead]

So there is work based off of HIP which is incompatible with ROCm. What a time to be an AMD owner.

[Mathias]

Correct me if I'm wrong, but this is not a drop in replacement that helps running CUDA stuff on other GPUs. It is a porting framework to convert CUDA code into platform agnostic code. So you need the source and invest time porting that code from Cuda to HIPCL.

I have no real experience with HIP, but since its inception, none of my favourite Cuda code has been ported to HIP. So I'm unsure about the impact this will have.

[tildearrow]

Correct me if I'm wrong, but this is not a drop in replacement that helps running CUDA stuff on other GPUs. It is a porting framework to convert CUDA code into platform agnostic code. So you need the source and invest time porting that code from Cuda to HIPCL.

Oh... I thought it was something like MoltenVK or Wine...

[schmidtbag]

What great timing. I was looking for something like this earlier this week.

[nuetzel]

What great timing. I was looking for something like this earlier this week.

And do you have numbers (with AMD OpenCL)?

[schmidtbag]

And do you have numbers (with AMD OpenCL)?

Numbers of/for what? All I want to do is run a program that has CUDA as a dependency. I don't especially care if it runs super efficiently.

[Mathias]

Numbers of/for what? All I want to do is run a program that has CUDA as a dependency. I don't especially care if it runs super efficiently.

This will only help you, when you actively want to port the CUDA code to non AMD or NVIDIA Platforms. HIP for AMD ROCm enabled GPUs is available for a long time (yet nobody upstream uses it).