segmentation fault when use cblas_sgemm. Occasionally, this core appears, and most of the time it works normally.

[johnson-xu01]

Program terminated with signal 11, Segmentation fault.
#0 0x0000000002e4297d in sgemm_incopy ()
#1 0x0000000002e3bc66 in sgemm_tn ()
#2 0x0000000002e3afaa in cblas_sgemm ()

[martin-frbg]

This could be related to the thread safety issue just reported as #3787 - but hard to tell without more information. What is your hardware, and which version of OpenBLAS please ?

[johnson-xu01]

This could be related to the thread safety issue just reported as #3787 - but hard to tell without more information. What is your hardware, and which version of OpenBLAS please ?

thank @martin-frbg . I just want to run normally and ignore the occasional error. add try catch for cblas_sgemm, does it work?

[martin-frbg]

I really cannot answer this with the little information you gave. aybe you are using an outdated version of OpenBLAS with an error that has long since been fixed; or maybe you can get around the problem by compiling OpenBLAS with USE_SIMPLE_THREADED_LEVEL3=1

[johnson-xu01]

I really cannot answer this with the little information you gave. aybe you are using an outdated version of OpenBLAS with an error that has long since been fixed; or maybe you can get around the problem by compiling OpenBLAS with USE_SIMPLE_THREADED_LEVEL3=1

@martin-frbg OpenBlas version 0.3.20

[johnson-xu01]

I really cannot answer this with the little information you gave. aybe you are using an outdated version of OpenBLAS with an error that has long since been fixed; or maybe you can get around the problem by compiling OpenBLAS with USE_SIMPLE_THREADED_LEVEL3=1

really thank you @martin-frbg . I use version 0.3.20, does 0.3.21 have fix this problem?

[brada4]

First to try is threads vs no-threads setting OPENBLAS_NUM_THREADS=1 before running sample - if that fails sporadicaly too then it is certain problem is threading.
Could be related to vectorizer dilemma too in case threads are cleared.

[martin-frbg]

Does not look like anything fixed between 0.3.20 and 0.3.21 unfortunately. What is your hardware, x86_64 or something like arm64 ?

[johnson-xu01]

Does not look like anything fixed between 0.3.20 and 0.3.21 unfortunately. What is your hardware, x86_64 or something like arm64 ?

maybe AMD EPYC ROME

[martin-frbg]

So a big system - how big is your matrix size in the SGEMM ? (Could be you sometimes exceed the size of the internal buffer, a static array used to communicate subsets of the data between threads, see BUFFERSIZE setting in Makefile.rule)

[martin-frbg]

closing as there was not much to go on