LAPACK has segfault with large matrix SVD.

[cyrusshaoul]

Dear LAPACK team,

I am sorry that I can't send in a report that uses the LAPACK routines directly, but I am not sure how to do this.
I can report two equivalent reproducible examples, one in numpy and one in R. Both are using LAPACK, to do the SVD within the pseudoinverse calculation:

See the attached code for the reproducible examples.

I am running all of this on a machine with 128Gb or RAM. These matrices require around 40Gb to process. I have tried this with OPENMP num threads of 1 or higher, and
the result is the same.

Do you have any large-memory machines you can use to try and reproduce this? Is there any reason why it does not like large matrices?

Thanks so much,

Cyrus

[cyrusshaoul]

Also, Here is the screenshot from the crash report that I got when numpy/python finished dumping core (many hours after it crashed.)
I am not sure what DLASD2_ is. Does it help trace the problem?

-Cyrus