Re: Octave 3.6.4 VS2010 and the C++ API

[Michael Goffioul]

On Wed, Sep 25, 2013 at 7:55 AM, Mike Puglia <address@hidden> wrote:

Hi Michael,

I think I've found the source of the complex matrix multiplication problem I'm having.  The xgemm function in CMatrix.h tests a couple cases before executing a LAPACK function to do the multiplication.  For most cases, ZGEMM is called and this works correctly in all builds.  However, if a 1xn matrix A is multiplied by a nx1 matrix B, as in c = A*B, XZDOTU is called instead and in the VS2010 build the result is not always (ever?) correct.  The octfile below should reproduce the problem.

It seems that this is a problem with the BLAS/LAPACK/ATLAS package being used in the VS2010 build.  Does this sound correct?  (I've never understood the difference between these three things, so maybe now is a good time to learn.)  There are some posts from 2007/08 that touch on this issue with XZDOTU.  Any information you might have about the packages used in the build (or how to find that information) would be greatly appreciated.

https://mailman.cae.wisc.edu/pipermail/bug-octave/2008-March/012009.html

https://mailman.cae.wisc.edu/pipermail/bug-octave/2007-October/010101.html

I didn't have time yet to investigate. But what surprizes me is that I would expect the same code path to be used for complex matrix multiplication in plain m-files.

That being said, it may be a problem of calling convention and returning a double complex object. What numerical library did you chose during installation? (OpenBLAS, ATLAS, Generic). Could you try if the problem exists with other flavours of the numerical libraries?

Michael.