Coriander Project: Compile CUDA Codes To OpenCL, Run Everywhere

[coder]

Anyone know how tough it would be to add OpenCL 2.2 support to Coriander?

I don't understand this. As described, the project is for running CUDA code on OpenCL devices. So, if OpenCL 1.2 is good enough for that, why would you want to ratchet-up the requirements?

Imagine being able to use IGPs for PhysX.

Can it work like that? I assume it requires a re-compile.

[coder]

Great news, although I worry this enables CUDA to become the default tool for GPU programming, if you can truly run it everywhere. I know it's already dominant, but from what I've seen it has enough baggage and OpenCL seems like a cleaner/leaner CUDA (having had the benefit of coming later).

There's also work on implementing portions of the CUDA API in OpenCL.

It's more interesting to me if we could run OpenCL 2.2 on all recent CUDA devices, including the Tegra GPUs that support no version of OpenCL.

[nanonyme]

I'm sure if NVIDIA could stop this, they would. Their precious CUDA's can't be let to roam free!

Haven't they usually done that by hiring the developers?

[theriddick]

I'm sure if NVIDIA could stop this, they would. Their precious CUDA's can't be let to roam free!

[schmidtbag]

This is really cool but I'd like to have seen it years ago. I'd be interested to see results on Nvidia GPUs, to see the potential performance losses.

Imagine being able to use IGPs for PhysX.

[Marc Driftmeyer]

But OpenCL is dying. Khronos is merging it with Vulkan.

That's not dying, it's providing a streamlined bridge to make V/OCL work as efficiently as possible.

[grndzro]

Anyone know how tough it would be to add OpenCL 2.2 support to Coriander?

Cool! I wonder if this can be used to run Physx work on GPU using OpenCL? I believe Physx uses CUDA, and it is hardware accelerated on NVIDIA GPUs.

[grndzro]

OpenCL is faster than Cuda. With more optimizations it might make Vega a very attractive card indeed.

[Serafean]

But OpenCL is dying. Khronos is merging it with Vulkan.

Haven't heard that, and even if : so what? with llvm, changing the destination target should be a moderate amount of work.