[Scipy-tickets] [SciPy] #1523: another Arpack crash on OS X python 2.7, 64-bit

SciPy Trac scipy-tickets@scipy....
Mon Sep 26 01:11:27 CDT 2011

#1523: another Arpack crash on OS X python 2.7, 64-bit
 Reporter:  rgommers             |       Owner:  wnbell
     Type:  defect               |      Status:  new   
 Priority:  highest              |   Milestone:  0.10.0
Component:  scipy.sparse.linalg  |     Version:  0.9.0 
 Keywords:  arpack, segfault     |  

Comment(by rgommers):

 Yes, must be OS X specific.  Moreover, must be also 64-bit specific.
 Based on searching for issues with that combination, the most common
 problem is with sdot, which is used in arpack.  From
 The issue with SDOT on 64-bit Intel is a legitimate bug with the BLAS not
 obeying the g77/f2c calling conventions. Specifically, the g77/f2c
 convention is for
 single precision results to be returned as double precision, but the OS X
 BLAS is
 incorrectly returning the result in single precision. When that single
 result is interpreted as a double precision value by the calling routine,
 the result
 is a very tiny (denormal) double-precision value, which rounds to exactly
 zero when
 converted to single.
 Something like that could explain an error like
 malloc: *** error for object 0x105df7af8: incorrect checksum for freed
 object -
 object was probably modified after being freed.

 There's also this thread:
 I believe I have found a bug in ARPACK for the case of an unsymmetric
 matrix where the values need reordering while trying to write a binding
 to ARPACK for octave. The issue is with xTRSEN running a value of M that
 exceeds NCONV, M being an output value of xTRSEN, that is returned in
 NCONV effectively replacing it. This results in a segmentation fault in
 the follow xCOPY commands. This can only be guarenteed to be avoided
 with the current code if the size of the vectors for the eigenvalues and
 vectors are ncv (rather than k+1) in size. Please find attached three
 files that I believe demonstrate the issue.

Ticket URL: <http://projects.scipy.org/scipy/ticket/1523#comment:4>
SciPy <http://www.scipy.org>
SciPy is open-source software for mathematics, science, and engineering.

More information about the Scipy-tickets mailing list