You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
MKL doesn't load correctly into a VEC yet. All our current examples have had to run with OpenBLAS because of this. When we try to load MKL we get a dynamic linker error claiming that omp_num_threads can't be found, but it's not clear why there would be any problems with that.
The text was updated successfully, but these errors were encountered:
Note: anaconda has some special compatibility code for loading MKL. They ship it in a Python-level mkl package that their MKL-based numpy builds import first to get MKL loaded correctly. That's one likely point of failure that could cause an issue like this.
The first place to look is in the dlopen wrapper for code inside VECs. It may be overriding or changing handling of the GLOBAL and LOCAL flags since those were broken in the unpatched glibc. Now that we use the patched glibc, it should work, but it's very possible that the wrapper still overrides the behavior since I may not have properly updated the wrapper and it might be overriding the GLOBAL/LOCAL flag in some way.
Weirdly enough, this works now. If I install the latest version MKL via conda-forge it no longer uses the mkl-service package that had all that business with RTLD_GLOBAL. So MKL works now, but numba is broken by a similar issue. I'm closing this one. I'll file a separate issue to discuss.
MKL doesn't load correctly into a VEC yet. All our current examples have had to run with OpenBLAS because of this. When we try to load MKL we get a dynamic linker error claiming that omp_num_threads can't be found, but it's not clear why there would be any problems with that.
The text was updated successfully, but these errors were encountered: