I’m trying to link the precompiled libtorch_cpu v1.9.0 to an existing application which is linked to its own MKL dependency.
it is built with no problem but when I try to run forward from a script inside the app, I receive a SIGILL with following trace:
#0 0x00007f6fcb0d9550 mkl_blas_def_sgemm_kernel_0_bdz (libtorch_cpu.so)
Two puzzling leads to the problem:
- I ran this app under two different machines with/without AVX512 support. And it runs ok with AVX512 support. So it appears to be related to the AVX512.
- I tested an independent example solely with libtorch on non-AVX512 machine and it works fine.
So it seems something in my original application confuses libtorch to choose a wrong AVX… but I have no clue what could cause this kind of confusion.
I wonder if anyone has an idea on this
Thanks