OpenBLAS 0.3.8 issue on AMD Threadripper cpu

[shyrma]

OpenBLAS 0.3.8 crashes on AMD Threadripper 3970X cpu, with SIGILL error code.
We always got following crash while testing TFGraphTestAllHelper.log_determinant.rank3:

• A fatal error has been detected by the Java Runtime Environment:
• SIGILL (0x4) at pc=0x00007f86f33181ff, pid=130947, tid=0x00007f8839009700
• C [libopenblas_nolapack.so.0+0x10a61ff] sgemm_kernel_direct+0x126f

This issue is a blocker for us, since we basically unable to use devbox for java tests

[saudet]

I'm sure OpenBLAS 0.3.7 does the same. Work around that with OPENBLAS_CORETYPE, the same as was done for issue #4287.

[raver119]

re: 0.3.7. I don't think so, we've been using this box for 3 months now, and got into this problem only after update to 0.3.8 i think. It's pretty new issue there.

re: OPENBLAS_CORETYPE - nope, doesn't help.

[raver119]

@saudet you're right, 0.3.7 reproduces the same issue. So probably this crash is caused by the graph being added to resources.

I'll close this issue, we'll sort it out in some other way i think now.

[saudet]

Setting the OPENBLAS_CORETYPE environment variable to "Athlon" or "Core2" doesn't help? That's strange. That should disable pretty much all optimizations. Anyway, one more thing to try is to use MKL instead. The presets for OpenBLAS can load it with the system properties given here:
https://github.com/bytedeco/javacpp-presets/tree/master/openblas#documentation

[raver119]

Athlon doesn't help. We'll try Core2 now

[shyrma]

I've just checked, same situation - no success unfortunately

[saudet]

This is probably the cause: OpenMathLib/OpenBLAS#2526

[raver119]

Who's building OpenBLAS for JavaCPP?

[saudet]

The changes from commit bytedeco/javacpp-presets@cb00315 have been deployed.
Please give it a try with 0.3.9-1.5.4-SNAPSHOT: http://bytedeco.org/builds/

[AlexDBlack]

ND4J tests (including those that were consistently crashing on this system) are confirmed passing on 3990x with Linux (ubuntu 20.04) with the following version modifications:
https://github.com/KonduitAI/deeplearning4j/compare/master...KonduitAI:ab_openblas_test?expand=1

[raver119]

Everything is fine on 3950 and 3970 as well 👍