illegal instruction error in scipy.linalg.decomp.svd
On a Pentium 4 Linux box running python 2.5.1, scipy 0.6.0, numpy 1.0.4, and lapack 3.0, I recently noticed that scipy.linalg.decomp.svd() fails (and causes python to crash) with an "illegal instruction" error. A bit of debugging revealed that the error occurs in the line lwork = calc_lwork.gesdd(gesdd.prefix,m,n,compute_uv)[1] in scipy/linalg/decomp.py. Interestingly, scipy 0.5.2.1 on the same box (with all of the other software packages unchanged) does not exhibit this problem. Moreover, when I install scipy 0.6.0 along with all of the other packages on a Linux machine containing a Pentium D CPU, I do not observe the problem either. Being that I am running Mandriva Linux, closed scipy bug 540 caught my eye. I'm not sure how it could be related to the above problem, though (and I also do not know what the lapack patch mentioned in the ticket could have been - even though I have been maintaining Mandriva's lapack lately :-). Thoughts? L.G.
Received from Lev Givon on Mon, Dec 17, 2007 at 07:38:18PM EST:
On a Pentium 4 Linux box running python 2.5.1, scipy 0.6.0, numpy 1.0.4, and lapack 3.0, I recently noticed that scipy.linalg.decomp.svd() fails (and causes python to crash) with an "illegal instruction" error. A bit of debugging revealed that the error occurs in the line
lwork = calc_lwork.gesdd(gesdd.prefix,m,n,compute_uv)[1]
in scipy/linalg/decomp.py. Interestingly, scipy 0.5.2.1 on the same box (with all of the other software packages unchanged) does not exhibit this problem. Moreover, when I install scipy 0.6.0 along with all of the other packages on a Linux machine containing a Pentium D CPU, I do not observe the problem either.
Being that I am running Mandriva Linux, closed scipy bug 540 caught my eye. I'm not sure how it could be related to the above problem, though (and I also do not know what the lapack patch mentioned in the ticket could have been - even though I have been maintaining Mandriva's lapack lately :-).
Thoughts?
This problem is also present in the latest svn release of scipy (3781). L.G.
Received from Lev Givon on Fri, Jan 04, 2008 at 11:44:11AM EST:
Received from Lev Givon on Mon, Dec 17, 2007 at 07:38:18PM EST:
On a Pentium 4 Linux box running python 2.5.1, scipy 0.6.0, numpy 1.0.4, and lapack 3.0, I recently noticed that scipy.linalg.decomp.svd() fails (and causes python to crash) with an "illegal instruction" error. A bit of debugging revealed that the error occurs in the line
lwork = calc_lwork.gesdd(gesdd.prefix,m,n,compute_uv)[1]
in scipy/linalg/decomp.py. Interestingly, scipy 0.5.2.1 on the same box (with all of the other software packages unchanged) does not exhibit this problem. Moreover, when I install scipy 0.6.0 along with all of the other packages on a Linux machine containing a Pentium D CPU, I do not observe the problem either.
Being that I am running Mandriva Linux, closed scipy bug 540 caught my eye. I'm not sure how it could be related to the above problem, though (and I also do not know what the lapack patch mentioned in the ticket could have been - even though I have been maintaining Mandriva's lapack lately :-).
Thoughts?
This problem is also present in the latest svn release of scipy (3781).
L.G.
This problem seems to have been resolved; numpy revision 4863 / scipy revision 4010 (built against the same python and lapack libraries described earlier) do not exhibit the above behavior. Some other illegal instruction events I had observed on a Pentium 4 system involving routines in cephes also appear to have been resolved. L.G.
participants (1)
-
Lev Givon