Isuru Fernando
Isuru Fernando
I cannot reproduce this bug with pocl 4.0
I tried only https://github.com/pocl/pocl/issues/1288#issuecomment-1696570750
I cannot reproduce with, ``` PYTHONPATH=. PYOPENCL_CTX=port:nvidia ./run_tests.py pyFAI.opencl.test.test_openCL.TestDoubleWord ```
> PYOPENCL_CTX=port:nvidia selects Portable Computing Language:cpu-broadwell-Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz Ah, you are overriding pyopencl's parsing of `PYOPENCL_CTX` env var and has different behaviour. I can reproduce now.
https://github.com/pocl/pocl/pull/1227 is the culprit.
For now you can add `#pragma clang fp contract(on)` to the top of the `doubleword.cl` file. It's strange that NVIDIA OpenCL don't run into this issue.
Did you try the main branch or the release?
It does have the unused code, but I believe https://github.com/pocl/pocl/pull/1252 fixes the issue that you are running into.
If you are okay with using binaries, you can try the instructions at https://github.com/pocl/pocl#pocl-with-cuda-driver
I've tried it on an aarch64 server, but not a Jetson in particular. If it does not work there, I'll be happy to fix it.