[AMD] ICD enumerates fine, but no devices found

[Richardk2n]

pyopencl finds a platform

>>> cl.get_platforms()
[<pyopencl.Platform 'AMD Accelerated Parallel Processing' at 0x7f1feeb07f10>]

but no devices

>>> cl.get_platforms()[0].get_devices()
[]

Other opencl software works just fine. For example clinfo:

$ clinfo -l
Platform #0: AMD Accelerated Parallel Processing
 +-- Device #0: gfx906:sramecc+:xnack-
 +-- Device #1: gfx906:sramecc+:xnack-
 +-- Device #2: gfx906:sramecc+:xnack-
 +-- Device #3: gfx906:sramecc+:xnack-
 +-- Device #4: gfx906:sramecc+:xnack-
 +-- Device #5: gfx906:sramecc+:xnack-
 +-- Device #6: gfx90a:sramecc+:xnack-
 `-- Device #7: gfx906:sramecc+:xnack-

pyopencl has been installed via pip. I have no rights on the server.

icds are present:

$ ls /etc/OpenCL/vendors
amdocl64_60103_sles154_122.icd  amdocl-orca64.icd

Any Ideas/Debugging steps?

[inducer]

That's super weird, I'm not sure I've seen that failure mode. Some things to try:

• Use ldd on clinfo and the PyOpenCL shared object to make sure they are both talking to the same libOpenCL.so.
• Use strace on clinfo and device enumeration via PyOpenCL to see how the two paths diverge.

[Richardk2n]

ldd revealed, that they use vastly different libOpenCL.so
clinfo uses /opt/amdgpu-pro/lib64/libOpenCL.so.1
_cl.cpython-310-x86_64-linux-gnu.so uses .myVenv/lib/python3.10/site-packages/pyopencl/./.libs/libOpenCL-b54a1ea0.so.1.0.0, which is a lot larger.
Replacing it with the lib from /opt/ produced an additional error libstdc++.so.6: version 'GLIBCXX_3.4.30' not found, because it somehow grabs the wrong libstadc++.so.6. It seems like python prepends its own lib directory to the LD_LIBRARY_PATH).

My understanding of the strace command is practically 0. The output is very long and different from the beginning. Is there anything specific to look for?

[inducer]

My understanding of the strace command is practically 0.

You won't need it. You've already found the root cause.

Replacing it with the lib from /opt/ produced an additional error libstdc++.so.6: version 'GLIBCXX_3.4.30' not found, because it somehow grabs the wrong libstadc++.so.6.

It should be safe to move that libstdc++.so.6 aside (or delete it) if you're feeling brave.

[Richardk2n]

Thank you, I "replaced" libstdc++.so.6 using LD_PRELOADbecause I have no rights on the server to actually modify the file and this worked.

[kif]

Could it be a problems of rights to access some devices in /dev?
The strace path Andreas suggest should be able to point out which device is the problem.