Temporarily allow MKL/OPENBLAS threading?

I noticed this Issue #317 and its associated fix: https://github.com/pyscf/pyscf/blob/e064645ee08728c64412ed9bc8e7ffe45fc8e1f4/pyscf/__init__.py#L42

Suppresses MKL and OPENBLAS threading for the entire pyscf calculation. With @hongzhouye’s RSDF (#1034) speedup over GDF, calls to scipy solve_triangular become a bottleneck in the process, e.g. https://github.com/pyscf/pyscf/blob/8a4d973aeab9f7f3adb08cf1d15ba73a05fc7c17/pyscf/pbc/df/df.py#L362 which appears to be single threaded due to the aforementioned setting.

I wonder if there is a way to temporarily allow BLAS/LAPACK multithreading for these calls to solve_triangular? They only show in df modules.

[jamesETsmith]

@btglau I’m 99% sure scipy.linalg.solve_triangular should not be parallelized across threads. That function calls the LAPACK function trtrs, see the SciPy source here: https://github.com/scipy/scipy/blob/47bb6febaa10658c72962b9615d5d5aa2513fa3a/scipy/linalg/basic.py#L349.

The MKL docs don’t list trtrs as a function that’s parallelized with threads. I’ve tried out using scipy.linalg.solve_triangular on it’s own (with MKL and several OpenBLAS versions as the SciPy backend) and none show threading dependence while other methods like scipy.linalg.solve do (see the mini-benchmark here).

[fishjojo]

You may try the threadpoolctl package, which allows to set BLAS threading dynamically.