Re: Matlab, Mex, FFTW3

[Mike Miller]

On Tue, Dec 08, 2015 at 19:17:01 +0100, Hugo Raguet wrote:
> Hello,
> 
> I am switching from Octave to MATLAB, after having developed a series of
> tools for my research in the latter language. As is usual (or should be
> usual), I wrote the most time-consuming operations in C, interfaced with
> Mex.
> 
> I have two specific questions, which I will justify below.
> 1) is it possible that a mex-file properly written for Matlab (i.e.
> fully deterministic behavior, no segmentation violation, etc.) would not
> behave correctly when compiled and used within Octave ?
> 2) are there known incompatibilities between Mex (either Matlab or
> Octave) and the famous library FFTW3 ?
> 
> I am asking those questions because of the following observations:
> 1) Under Matlab, I encontered the following, very strange, bug:
> http://stackoverflow.com/questions/33921437/fftw-within-mex-crashes-with-very-odd-behavior
> . In brief, I compile C codes linked with FFTW3, and the resulting
> mex-file crashes in very specific and odd circumstances. The same bug is
> not reproducible within Octave, but it is expected, given the extreme
> specificity of this bug.
> 2) However, I have a huge routine in mex, which by the way also uses
> FFTW3, which seems to work perfectly fine : deterministic behavior, no
> crash, results coherent with the data.
> Now when I use it within Octave, I get *sometimes* what looks like the
> same results as in Matlab, sometimes different results (still
> 'resembling' in some sense the Matlab results, but different), and
> sometimes it crashes.
> 
> Any help, commentary, pointer, hint, idea concerning the above would be
> greatly appreciated.

You also asked here

  
http://www.mathworks.com/matlabcentral/answers/257492-fftw-within-mex-crashes-with-very-odd-behavior

which IMHO is the best place to ask. (Best to mention all
cross-postings).

Not having a lot of information, I can only guess at answers:

1) anything is possible,
2) not that I know of.

You say your simple code example always crashes in Matlab but never in
Octave. Just a guess, but does Matlab link against a different ABI
version of FFTW than the one you are trying to link your mex file to?

No way to even speculate on the "huge routine" without diving into the
code.

> PS: unrelated: where should I report very small errors, like typos in
> the French version of the GUI?

The bug tracker:

  https://savannah.gnu.org/bugs/?func=additem&group=octave

-- 
mike