AMD's GPU Performance API Library 3.5 Drops ROCm/HSA Support

[Naquatis]

Welcome to the club! After > 2 years of permanent frustration about amdgpu's abysmal stability, which got worse, not better with newer kernel versions over time, I came to that conclusion a few months ago. And I don't even need an OpenCL or high 3D performance... my only requirement is a kernel driver not crashing frequently, even on an idle desktop.

I need to be fair the RX 480 on Mesa was more stable for me than any other graphic card I ever had and I used custom Mesa patches on this thing. I would wish that AMD were able to see how big their Mesa advantages are and put some more work into this direction and at least finalize OpenCL 1.2.

[dwagner]

... no Radeon usage on Linux!
Now it will become a crappy Windows gaming card.

Waiting for Intel oneAPI and their graphics card.

Welcome to the club! After > 2 years of permanent frustration about amdgpu's abysmal stability, which got worse, not better with newer kernel versions over time, I came to that conclusion a few months ago. And I don't even need an OpenCL or high 3D performance... my only requirement is a kernel driver not crashing frequently, even on an idle desktop.

[Naquatis]

Well I am finally done with my Linux RX 5700 XT swap in and out.
No OpenCL with Mesa = no Radeon usage on Linux!
Now it will become a crappy Windows gaming card.

Waiting for Intel oneAPI and their graphics card.

And for everyone who like to watch a worthy speech:

https://www.youtube.com/watch?v=ZTq8wKnVUZ8

[Nille_kungen]

Let me guess ... whatever Oak Ridge sponsors them to develop for Frontier?

https://www.olcf.ornl.gov/wp-content..._specsheet.pdf
OpenCL is a dead end and the future seems to be CUDA over HIP so developers should start using HIP instead.
When a Swedish techsite had a 20yr anniversary Joe Macri was a speaker and i later read in the forum that one had asked him after the speech and the recording "doesn't AMD still invests in OpenCL?".
The answer was no, better to use CUDA over HIP.
I hope nothing got lost in translation.

Here's the speech if anyone is interested.

https://www.youtube.com/watch?v=D8m6CSlNmQA

[Naquatis]

According to David Airlie Vulkan is not able to run CL. Vulkan did not have pointers but it will get an extension for that. In Airlie's opinion Vulkan OpenCL is expected far more in the future than the other solutions.

I tested the RX 5700 XT again with Kernel 5.4.3 and amd-ucode but no way to get this Red Devil to work with OpenCL.
I tried the Blender release version and even build Blender from source with all bits and pieces but it still say:
"No compatible GPU found for path tracing"

So if I need to install the complete AMD closed source driver I would finally go with NVidia because without Mesa, RADV and Gallium Nine there is simply no benefit.

[duby229]

it's plainly obvious oss opencl on amd hardware is a disaster, so my question is, what's the possibility of opencl over vulkan? like a clvk or something?

[Naquatis]

When you say "if AMD OpenCL is already working with Mesa" are you talking about co-existence (AMDGPU-PRO OpenCL package working with Mesa GL/radv over an upstream kernel) or about the old clover implementation in Mesa ?

I'm asking because AFAIK the first one should work with a sufficiently new kernel (albeit the OpenCL support still being a bit rough) while we stopped working on the second years ago.

Separately you have probably noticed ongoing kernel work for the KFD/ROCm paths on Navi in preparation for getting the upper level components (OpenCL and HIP over ROCR).

Last time I tested the RX 5700 XT was with Kernel 5.4.1 with https://aur.archlinux.org/packages/opencl-amd/
The git link on this side do the same like the bash script you find as green comment below.

Maybe something missing for RX 5700 XT?
RX 480 works fine.

No idea if Kernel 5.4.3 will change anything on that situation?

[bridgman]

Yeah sorry, I can not see all the work you invest and from that point this are only harsh comments from a customer who bought a RX 5700 XT to swap it in and out from time to time to see if AMD OpenCL is already working with Mesa. My new plan is to throw this card into a Windows gaming rig and use the NVidia from there to run Cuda on Linux.

When you say "if AMD OpenCL is already working with Mesa" are you talking about co-existence (AMDGPU-PRO OpenCL package working with Mesa GL/radv over an upstream kernel) or about the old clover implementation in Mesa ?

I'm asking because AFAIK the first one should work with a sufficiently new kernel (albeit the OpenCL support still being a bit rough) while we stopped working on the second years ago.

Separately you have probably noticed ongoing kernel work for the KFD/ROCm paths on Navi in preparation for getting the upper level components (OpenCL and HIP over ROCR).

[bridgman]

really whats wrong with your AMD people? i have 4 vega64 on by desk right now and they do nothing because OpenCL on vanilla Fedora/Debian is super shit.

What problems are you seeing with Vega64 ? I didn't think you were running anything that used SYCL or SPIR/SPIR-V.

[Naquatis]

I think the intent of that comment was in response to a statement that we had dropped SPIR-V support, which AFAIK was not the case. Far be it from me to tell people what to complain about, but IMO the real issue here is that we dropped SPIR support before adding SPIR-V support.

Yeah sorry, I can not see all the work you invest and from that point this are only harsh comments from a customer who bought a RX 5700 XT to swap it in and out from time to time to see if AMD OpenCL is already working with Mesa. My new plan is to throw this card into a Windows gaming rig and use the NVidia from there to run Cuda on Linux.