AMD ROCm 3.9 Released With AOMP OpenMP Offloading Integrated

[phoronix]

Phoronix: AMD ROCm 3.9 Released With AOMP OpenMP Offloading Integrated

A new version of the AMD Radeon Open eCosystem (ROCm) has been released on the same day as the company announcing the Radeon RX 6800/6900 series. Meet ROCm 3.9...

http://www.phoronix.com/scan.php?page=news_item&px=AMD-ROCm-3.9-Released

[perpetually high]

FYI- not sure why they did this, but the rocm-smi tool is now rocm_smi_deprecated.py

Had to run the following to get it back:

$ sudo ln -s /opt/rocm-3.9.0/bin/rocm_smi_deprecated.py /usr/bin/rocm-smi

Would also be nice if they just used /opt/rocm so the symlinks stick on each new version.

[saltyhashy]

I was hoping that a new ROCm release on the day of the RX 6000 series announcement meant that they're rolling out support for RDNA2 and optionally RDNA1, or at least a tiny hint that they're working on it in the release notes. Sadly not the case, but here's hoping for the next release. If not, I'll probably put in an order for an Ampere card and see what happens first - the Ampere card being in stock and getting shipped or RDNA2 support in ROCm.

[Imroy]

Sigh, still buggy. Something broke after ROCm 3.3. And I'm getting lots of "Evicting/Restoring PASID 0x8009 queues" messages in the kernel log. Back to the AMDGPU PRO OpenCL driver again.

[oleid]

Sigh, still buggy. Something broke after ROCm 3.3. And I'm getting lots of "Evicting/Restoring PASID 0x8009 queues" messages in the kernel log. Back to the AMDGPU PRO OpenCL driver again.

Which kernel driver are you using? Upstream or bundled? The former (ArchLinux's 5.8 for example) works fine for me when using rocm.

[oleid]

ROCm 3.9 has no mention of supporting these GPUs starting to ship in November.

From what I heard OpenCL should (TM) work and MIOpen is WIP, but having a look at the git repo, they are actively working on it.

[Setif]

Navi and Big Navi Cards are for gaming.

[bridgman]

Sigh, still buggy. Something broke after ROCm 3.3. And I'm getting lots of "Evicting/Restoring PASID 0x8009 queues" messages in the kernel log. Back to the AMDGPU PRO OpenCL driver again.

The evicting/restoring messages are not errors, just informative. Most common cause is the OS wanting to move a pinned memory page, so we have to temporarily evict so that the page(s) can be unpinned. Are you having functional problems or just concerned about the messages ?

Regarding Luxmark, are you seeing the same problem as in the ticket, where Luxmark 3.1 has issues with -cl-fast-relaxed-math enabled in the CL option menu but works OK with it disabled ?

[Danny3]

Navi and Big Navi Cards are for gaming.

Then they should should sell it to gamers only!
I'm not gaming much, so good luck selling me an expensive GPU without compute support.

[wizard69]

I was hoping that a new ROCm release on the day of the RX 6000 series announcement meant that they're rolling out support for RDNA2 and optionally RDNA1, or at least a tiny hint that they're working on it in the release notes. Sadly not the case, but here's hoping for the next release. If not, I'll probably put in an order for an Ampere card and see what happens first - the Ampere card being in stock and getting shipped or RDNA2 support in ROCm.

This has become a bit pathetic. I didn't expect immediate support in Navi 1 but hell Navi 2 will be shipping in a couple of weeks.