On 2022-01-12 00:04, Rebecca N. Palmer wrote:
That's probably a good idea for runtime Depends, but *not*
Build-Depends and test-Depends: those are likely to be run on GPU-less
servers, where only CPU ICDs (i.e. pocl-opencl-icd) will work.
(Roughly: pocl-opencl-icd = CPUs, mesa-opencl-icd = AMD/ATI GPUs,
intel-opencl-icd = Intel GPUs; the GPU ICDs work on integrated GPUs,
but server processors often don't even have those. I think of
pocl-opencl-icd as slow enough to be mostly for testing, but don't
know how generally true that is.)
It might be that I'm not understanding how pyopencl interacts with
opencl-icd, I'm no expert in opencl. What made me suggest building via
mesa-opencl-icd was that mesa-opencl-icd builds for all arches, even the
niche minor arches.
mesa-opencl-icd is running and passing its own tests on these arches.
What is pyopencl doing in build and tests that's different from what
mesa-opencl-icd itself does?
Drew
