Radeon ROCm 3.5.1 Open-Source Compute Stack Released

[phoronix]

Phoronix: Radeon ROCm 3.5.1 Open-Source Compute Stack Released

Two weeks after ROCm 3.5, the AMD Radeon team has now issued a patch update to this Radeon Open Compute stack...

http://www.phoronix.com/scan.php?page=news_item&px=Radeon-ROCm-3.5.1

[bridgman]

I'm not 100% sure, but AFAIK we have been using the ROCm back end for OpenCL on Navi+Linux from day one so even the packaged drivers contain fully functional lower level ROCm components including amdkfd, libhsakmt and the ROC runtime.

What was missing was HIP/HCC and associated library support, but that would have been a bit of a throw-away because of the timing of the HIP/HCC to HIP/Clang transition.

We don't really have anything to announce in the release notes re: Navi support until all of the components are ready.

[halo9en]

Has anyone tried ROCm on the new APUs? I was wondering if it can be used on the Ideapad 5 (4700U) that I ordered, or if unofficial builds are still required.

[finalzone]

Has anyone tried ROCm on the new APUs? I was wondering if it can be used on the Ideapad 5 (4700U) that I ordered, or if unofficial builds are still required.

ROCm for Ryzen APU is still unspported. Installing the dkms-rocm is a bad idea as it will trash the initrd. For OpenCL, amdpgu-pro is still a better option despite detecting the GPU part of Ryzen APU as unknown.
The attached link for the unofficial builds is specific for Ubuntu. For Fedora users, no luck so far.

[bridgman]

I believe Renoir was the first shipping APU where we switched over to dGPU code paths, using GPUVM page tables rather than ATC/IOMMUv2. What I don't remember is whether we are using the ROCm back end for OpenCL in the packaged drivers.

Don't remember if Michael tested the packaged drivers on his Renoir laptops (IIRC running clinfo would tell whether PAL or ROCm back end was being used) but will check.

[angrypie]

Does it really need PCIe atomics? This requirement leaves many old (and performant) machines out of ROCm.

[bridgman]

Does it really need PCIe atomics? This requirement leaves many old (and performant) machines out of ROCm.

Vega does not, Polaris and earlier do.

Not sure about Navi but assume it needs them as well for now (ie unless you hear otherwise).

[finalzone]

I believe Renoir was the first shipping APU where we switched over to dGPU code paths, using GPUVM page tables rather than ATC/IOMMUv2. What I don't remember is whether we are using the ROCm back end for OpenCL in the packaged drivers.

Will these changes get backport for the integrated GPU part of Raven Ridge like Ryzen 2500u?

[marty1885]

I'm not 100% sure, but AFAIK we have been using the ROCm back end for OpenCL on Navi+Linux from day one so even the packaged drivers contain fully functional lower level ROCm components including amdkfd, libhsakmt and the ROC runtime.

What was missing was HIP/HCC and associated library support, but that would have been a bit of a throw-away because of the timing of the HIP/HCC to HIP/Clang transition.

We don't really have anything to announce in the release notes re: Navi support until all of the components are ready.

Could AMD just say OpenCL is working but HIP isn't? ROCm not supporting Navi is becoming a meme.

[vlad]

we have been using the ROCm back end for OpenCL on Navi+Linux from day one so even the packaged drivers contain fully functional lower level ROCm components including amdkfd, libhsakmt and the ROC runtime.

Is it possible that the ROCm back end for OpenCL works correctly for gfx7 despite the HIP backend does not?