We have small matrices we're attempting to multiply, and sometimes it crashes. Not every time, and when it works the results appear to be correct.
# SIGSEGV (0xb) at pc=0x00007f39486428ec, pid=15934, tid=0x00007f3934a26700
siginfo: si_signo: 11 (SIGSEGV), si_code: 2 (SEGV_ACCERR), si_addr: 0x00007f3bd40bc000
Problematic frame: # C [libmkl_mc3.so+0xcc08ec] mkl_dft_mc3_gather_d_d+0xcc
Arguments to the function:
A is 18x24
B is 18x28
C is 24x28
matdescra: "G F "
Are we calling the function incorrectly somehow? It appears like every time it crashes the process is attempting to read 32 bytes past the end of B's allocated space. Our theory is that it "works" when that memory is part of the process address space and crashes when it's not.