AMD Releases Orochi 3.0 For HIP & CUDA API Switching At Run-Time

[phoronix]

Phoronix: AMD Releases Orochi 3.0 For HIP & CUDA API Switching At Run-Time

Back in April 2022 was the announcement by AMD's GPUOpen team of Orochi as a library for HIP/CUDA API run-time switching. Making use of Orochi allows for dynamically targeting either AMD HIP or NVIDIA CUDA at run-time to ease the distribution/usage of apps wanting to target both AMD and NVIDIA GPUs from the same software build. Today Orochi 3.0 was released...

https://www.phoronix.com/news/AMD-Orochi-3.0

[nazar-pc]

I wish they supported static linking like Nvidia allows

[ET3D]

I wish they supported static linking like Nvidia allows

Orochi's point is dynamically loading CUDA or HIP, so I'm not sure what you're asking for.

[nazar-pc]

If HIP runtime was possible to link statically then application would be able to support ROCm and CUDA at the same time in statically built application (either with ability to select in runtime or to use both concurrently). Right now supporting ROCm means depending on HIP runtime libraries even for those users that only need CUDA.

I'm coming from a background of Rust development where most things outside of libc are statically linked and application can make decisions in runtime on its own.

[ET3D]

If HIP runtime was possible to link statically then application would be able to support ROCm and CUDA at the same time in statically built application (either with ability to select in runtime or to use both concurrently). Right now supporting ROCm means depending on HIP runtime libraries even for those users that only need CUDA.

But this is precisely what Orochi is meant to solve. It's meant to enable a single application to run on CUDA if there's a CUDA runtime and on HIP if there's a HIP runtime.

Disclaimer: I haven't used Orochi, so I don't know how well it achieves this. When I tried it, it had missing functionality (which I needed) compared to HIP. Still, as I understand it, what you're asking for is exactly what it's meant to offer.