Numpy 1.7.1 Crashing with MKL and AVX instructions

[Paul Blelloch]

I've tried posting this on the numpy list, but it keeps on getting bounced, so I'll try here:

 

I found that when I went from numpy 1.7.0 to 1.7.1 I get a crash whenever I try an eigenvalue calculation (or any other linalg calculation) on matrices bigger than about 200x200.  This happens with both the latest Anaconda and WinPython 64-bit Windows distributions (both of which use numpy 1.7.1) and occurs on all the HP workstations at my company.  The folks at Continuum helped me debug this and determined that the failure was most likely in use of AVX instructions in MKL, but we haven't found a workaround yet short of sticking to numpy 1.7.0.  The problem is apparently specific to some combination of processor and OS.  Here's what I have:

 

OS Name Microsoft Windows 7 Professional
Version 6.1.7601 Service Pack 1 Build 7601
OS Manufacturer Microsoft Corporation
System Name Z420-6
System Manufacturer Hewlett-Packard
System Model HP Z420 Workstation
System Type x64-based PC
Processor Intel(R) Xeon(R) CPU E5-1650 0 @ 3.20GHz, 3201 Mhz, 6 Core(s), 12 Logical Processor(s)
BIOS Version/Date Hewlett-Packard J61 v01.14, 7/17/2012
SMBIOS Version 2.7
Windows Directory C:\Windows
System Directory C:\Windows\system32
Boot Device \Device\HarddiskVolume1
Locale United States
Hardware Abstraction Layer Version = "6.1.7601.17514"
Installed Physical Memory (RAM) 32.0 GB
Total Physical Memory 31.9 GB
Available Physical Memory 28.4 GB
Total Virtual Memory 95.8 GB
Available Virtual Memory 92.1 GB
Page File Space 63.9 GB
Page File C:\pagefile.sys

 

I'm surprised that I'm the only person out there running into this.  Is there anyone else who's running into this problem with numpy 1.7.1 with MKL on 64-bit Windows?

 

-Paul Blelloch

[Paul Blelloch]

This is just an extra piece of information that I also downloaded numpy 1.7.1 directly from http://www.lfd.uci.edu/~gohlke/pythonlibs/#numpy, so this seems to be fairly universal for the MKL version of numpy 1.7.1.

[Paul Blelloch]

And another piece of information is that I get the same behavior with the scipy 0.12.0 linalg routines such as sqrtm.

On Friday, May 24, 2013 8:24:37 AM UTC-7, Paul Blelloch wrote:

[Paul Blelloch]

Christoph Gohlke figured this out and has updated the MKL 64-bit builds of both numpy 1.7.1 and scipy 0.12.0 on his site. 

On Friday, May 24, 2013 8:24:37 AM UTC-7, Paul Blelloch wrote: