PoCL 3.1-RC1 Released With Improved SPIR-V Support For CPU & CUDA Drivers, Vulkan WIP

[phoronix]

Phoronix: PoCL 3.1-RC1 Released With Improved SPIR-V Support For CPU & CUDA Drivers, Vulkan WIP

PoCL 3.1 is nearing release as the "Portable Computing Language" that is most known for serving as a CPU-based OpenCL implementation but via its LLVM usage also allows supporting OpenCL execution atop NVIDIA CUDA and other targets...

https://www.phoronix.com/news/PoCL-3.1-RC1-Released

[Eirikr1848]

FYI for CUDA it states: Currently this backend has only been tested against CUDA 8.0.

[Developer12]

While I'm all for multiple implementations and pathways as much as the next guy, I fully expect rusticl will kill this thing.

[R41N3R]

Another solution of OpenCL on Vulkan :-) It will be interesting to compare PoCL to Rusticl on Zink at some point.

[Quackdoc]

Another solution of OpenCL on Vulkan :-) It will be interesting to compare PoCL to Rusticl on Zink at some point.

PoCL predates rusticl by a fair chunk, but rusticl has a couple large advantages, the most important being that it's built on the already mature and highly battle tested gallium stack, meaning a lot of work for rusticl was already done to get it working, (gallium is lovely for a reason).

EDIT: so you could say that while PoCL predates rusticl, gallium predates PoCL so a lot more work has been put into that.

[marlock]

Does using PoCL require OpenCL code refactoring in any way or can it be simply be selected to run unchanged OpenCL software that hasn't taken it into consideration?

Also isn't a direct path from OpenCL to Vulkan potentially better in the long run than the extra crazy roundtrip done via Zink and the less-close-to-metal features offered by Gallium? If that's the case I wouldn't scrape PoCL so quickly even if it currently lags behing...