Radeon ProRender SDK 3.1 Released - Finishes Transition From OpenCL To HIP

[phoronix]

Phoronix: Radeon ProRender SDK 3.1 Released - Finishes Transition From OpenCL To HIP

AMD today published Radeon ProRender SDK 3.1 as the newest version of this cross-platform and open-source physically-based rendering engine...

https://www.phoronix.com/news/Radeon-ProRender-SDK-3.1

[gsedej]

Hip works on Polaris??? In demo file or it actually work?

[FireBurn]

Just as OpenCL almost becomes usable with Mesa, things move away from it. Typical

[Developer12]

Just as OpenCL almost becomes usable with Mesa, things move away from it. Typical

The best we can hope is a catch-up game where next there's a mesa project to implement HIP, except this time it's a vendor API.

[Awesomeness]

The best we can hope is a catch-up game where next there's a mesa project to implement HIP, except this time it's a vendor API.

Why should Mesa implement HIP when HIP is already fully open source and supports multiple back-ends?

[gsedej]

HIP was designed and is being pushed for PRO and enterprise hardware like radeon W and MI

The consumer RX cards are badly supported like missing convolution kernels (written in assembly)

[boxerab]

Support for consumer cards is such a mess. And the writing is on the wall - OpenCL support will be neglected and eventually abandoned.
Meanwhile CUDA runs on every nvidia card back to Kepler, and on every majour OS.

[osw89]

Just as OpenCL almost becomes usable with Mesa, things move away from it. Typical

Does anything relevant apart from john/hashcat even use opencl anymore ?

[coder]

The best we can hope is a catch-up game where next there's a mesa project to implement HIP,

Screw HIP. It should die in a fire, along with CUDA.

OpenCL is an open standard. HIP is nothing more than a carbon copy of CUDA.

[coder]

Does anything relevant apart from john/hashcat even use opencl anymore ?

OpenCL has a lot more users than HIP does. Intel's oneAPI, their OpenVINO deep learning framework, and their Embree renderer are all built on OpenCL.

OpenCL support will be neglected and eventually abandoned.

AMD has a lot more incentive to continue supporting OpenCL than Nvidia does. But Nvidia has maintained their OpenCL, even if it still lacks significant features.

Right now, AMD is sitting at like 1% marketshare of GPU compute. The main incentive anyone has not to use CUDA is to avoid vendor lock-in. And, in spite of what AMD says, relying on continued HIP support and performance across all hardware is a risky bet. If AMD wants to grow its GPU Compute userbase, they'd better maintain a decent OpenCL backend or else a lot of people will walk right by them and go straight to Intel.