Segmentation fault with openblas >0.3.10 and numpy

[Vringe]

I'm using numpy-1.16.6 with openblas-0.3.13 which causes a segmentation fault while importing.
Not sure if it is numpy or openblas related. But what I found out is that with openblas-0.3.13, much more virtual memory is required.

Here is how I tested it:

$ ulimit -v
2097152
$ python -c "import numpy as np"
Segmentation fault
$ ulimit -v 3670016
$ python -c "import numpy as np"
$

Downgrading to openblas-0.3.10 also fixes this issue.

Also mentioned here: numpy/numpy#18631

[martin-frbg]

OpenBLAS versions after 0.3.10 default to using a larger buffer for multihreaded GEMM to accomodate larger problem sizes. This can be adjusted at build time via the BUFFERSIZE parameter - recent numpy releases make use of this (setting BUFFERSIZE=20 to restore the pre-0.3.10 behaviour), which is probably why you were asked to upgrade to 1.20.x in the numpy ticket
(i.e. basically a duplicate of #2970 and numpy/numpy#18141)

[Vringe]

Thank you. Can confirm that setting BUFFERSIZE to 20 fixes the issue.