AMD Clarifies ROCm Compute Support For GUI Applications

[bridgman]

What's ROCr? Is that different from ROCm? Anyway, I was under the impression that ROCm would be used only on fairly high end GPUs, while low-end GPUs and APUs would use PAL (https://www.phoronix.com/scan.php?pa...OCm-PAL-Future).

ROCR is the ROC Runtime, one of the layers in the ROCm stack. From top to bottom the core stack for ML on Linux is roughly:

ML frameworks (Tensorflow / Pytorch etc...)
MIOpen (CUDNN) and RCCL (NCCL)
Math libraries
HIP language runtime
VDI (virtual device interface, runs over CPU / Orca / PAL / ROCr back ends)
ROCr back end
ROCR
libhsakmt aka "thunk" aka ROCT
amdkfd
amdgpu

For OpenCL it's the same from amdgpu up to VDI, but with OpenCL language runtime replacing HIP language runtime. HPC is similar to ML except CUDNN and the ML frameworks are not normally used.

Do all RDNA GPUs work with ROCr, and will RDNA APUs do so?

Yes, right up to OpenCL so far, and yes.

But yes, open-sourcing PAL OpenCL would still leave a place for Clover on Polaris and older, except those that work with ROCm.
BTW, why can't PAL OpenCL be used on Polaris and earlier, given that AMDVLK works on these?

It could be - just another pile of work for relatively small benefit at a time when we are not exactly looking for more work.

[pete910]

Can someone please tell me what this clarification from AMD will mean for Folding@home? It needs openCL to work on their video cards.

To update this, 4.1 has fixed it for me it seems