AMD Announces "Orochi" For HIP/CUDA Run-Time Handling

[phoronix]

Phoronix: AMD Announces "Orochi" For HIP/CUDA Run-Time Handling

AMD's GPUOpen team today announced "Orochi" as their latest open-source software offering in the HIP GPU compute space...

https://www.phoronix.com/scan.php?pa...&px=AMD-Orochi

[schmidtbag]

I would hope open-source programs utilizing CUDA take advantage of this.

Looking at you, Meshroom.

[kiffmet]

I fear that this isn't good enough. Asking to replace CUDA calls with calls to the Orochi API is already too much. It's the same reason why OpenCL, OpenMP, SYCL, C++ AMP and whatnot have had limited success in comparison to Nvidia's offering. When it comes to massively parallel, GPGPU accelerated application, the devs just want to use the de-facto standard CUDA and nothing else.

A reverse-engineered/open source no-frills CUDA reimplementation for AMD hardware would be a much better choice.

[coder]

When it comes to massively parallel, GPGPU accelerated application, the devs just want to use the de-facto standard CUDA and nothing else.

Some do. A lot of us would use OpenCL or SYCL, when given a choice.

It's the same reason why OpenCL, OpenMP, SYCL, C++ AMP and whatnot have had limited success in comparison to Nvidia's offering.

OpenCL got partly sabotaged by Nvidia & others dragging their feet on implementation, plus the slow pace at which new features were added. I'll never forgive Google for shutting it out of Android. Apple walked away from it, and Microsoft never had anything to do with it.

OpenMP is easy to use, but performance gains are limited, due to limited expressiveness and control.

SYCL was initially experimental and effectively limited to a niche of even the OpenCL market, for most of its history.

I think the reason C++ AMP didn't catch on is that it was a Microsoft proprietary thing. It had neither the portability of a true standard, nor the low-level control and advanced features of CUDA.

A reverse-engineered/open source no-frills CUDA reimplementation for AMD hardware would be a much better choice.

Uh, that's basically HiP.

[tehehe]

Uh, that's basically HiP.

Except it isn't. From documentation:

HIP allows developers to convert CUDA code to portable C++. The same source code can be compiled to run on NVIDIA or AMD GPUs....HIP is not intended to be a drop-in replacement for CUDA, and developers should expect to do some manual coding and performance tuning work to complete the port.

[coder]

Except it isn't. From documentation:

I've never used HiP, so I'm no authority on this, but I think it's basically as drop-in as reasonably possible.

You have to remember that it came out right after Google lost the API copyright lawsuit by Oracle, over Java. That would make any corporate lawyer slam the door shut on copying an API, verbatim. But AMD still wanted an easy way to claw back some marketshare from Nvidia, so the next best thing was a work-alike API + translation tools.

[Vermilion]

A lot of words to say it's a cross-platform abstraction layer.
If other vendors join in and more backends are developed, this could become a sane standard, and the porting examples make it look easy.

[billyswong]

I've never used HiP, so I'm no authority on this, but I think it's basically as drop-in as reasonably possible.

You have to remember that it came out right after Google lost the API copyright lawsuit by Oracle, over Java. That would make any corporate lawyer slam the door shut on copying an API, verbatim. But AMD still wanted an easy way to claw back some marketshare from Nvidia, so the next best thing was a work-alike API + translation tools.

So in this sense this new "Orochi" is a response to the latest ruling by the supreme court after Google won the lawsuit ultimately?