libblas for windows（Re: Octave3.6.0_gcc4.6. 2 mingw binaries for testing）

[Tatsuro MATSUOKA]

Hello

As reported in the bug tracker:
http://savannah.gnu.org/bugs/?34822
matrix calculation is not stable for windows binary.

This might be due to the stability of the ATLAS.

Recently I have succeeded the ATLAS 3.8.4 using self complied gcc-4.5.3 on 
cygwin on the HT-Pentium.
This ATLAS seem to be more stable that have been used.

I will use this ATLAS for octave-3.6.1. (Octave 3.6.1 will appear in the very 
near future.)

However, this ATLAS uses the SSE3 so that it might be unstable for the CPU that 
does not have the SSE3.

OTOH, libblas.dll can be replaced.

You can use self-build libblas.dll.  One of the powerful candidate is OpenBLAS.
OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version.  

http://xianyi.github.com/OpenBLAS/

The performance of OpenBLAS seems to be higher than that of the ATLAS.

If I will have time,. I want to describe how to build ATLAS and OpenBLAS for 
MinGW version octave on the Octave Wiki.

Regards

Tatsuro



--- On Wed, 2012/2/1, Tatsuro MATSUOKA  wrote:

> Hello Nitzan
> 
> > Could you please describe how you built the binaries? and the patches needed
> > to build the octave-forge packages? I tried a few myself but got build
> > errors. I conclude that you have more insight, it would be nice if you could
> > share that.
> > I think I have asked this question before.....
> 
> In the above, who did you indicate by "you" ?
> 
> I do not have enough knowledge to make the octave-forge package so that "you" 
> do not indicate "me".
> 
> Anyway, the way of creating octave binary itself is written in
> Readme_octave-3.6.0.txt
> http://www.tatsuromatsuoka.com/octave/Eng/Win/
> 
> The patch is in http://savannah.gnu.org/bugs/?33798
> CTRL_CLOSE_EVENT.diff 
> 
> Other modifications are to avoid errors from gnulib faults(?) on MinGW.
>   #inculde <unistd.h>
>    |
>    V
>   #inculde <unistd.h>
>   #inculde <unistd.h>
> 
>   The above modification is required for liboctave/oct-env.cc, src/lex.cc, 
>   src/oct-hist.cc, src/sighandlers.cc, and src/sysdep.cc
> 
> That's all the fix the code for the build.
> Build commands are written in 'Readme_octave-3.6.0.txt'.
> 
> Regards
> 
> Tatsuro
> 
> --- On Tue, 2012/1/31, nitnit  wrote:
> 
> > Sorry about this site. Multiupload.com which I have used before, didn't
> > function very well during recent weekend.
> > 
> > Regarding the packages, most of the patches I have applied are due to minor
> > problems in the make/configure scripts (mostly missing LDFLAGS) which are
> > probably specific to windows mingw environment. Since I am not an expert
> > with such scripts I do not know how to make the patches properly in a way
> > that they will still be compatible with other systems. This is the reason I
> > didn't applied them to the svn.
> > 
> > It is still in my tasks list to improve the way these patches are documented
> > and to post them to the OF mailing lists so that the packages maintainers
> > will fix these issues properly.
> > 
> > 
> > BTW the upload site you used is a rather nasty one. One should click the
> > tiny "download" button at the bottom of the web page, not the big one next
> > to the archive name as that will lead one to a download manager (perhaps
> > riddled with malware, I obviously didn't try). 
> > To me this site looks extremely deceiving (on purpose I'd say).
> > In addition my desktop got littered with several pop-up windows that managed
> > to bypass by browser's pop-up blocker.
> > 
> > It may be a lot better if you could get permission to upload your 7z files
> > to octave-forge directly, or use dropbox or so.
> > 
> > Could you please describe how you built the binaries? and the patches needed
> > to build the octave-forge packages? I tried a few myself but got build
> > errors. I conclude that you have more insight, it would be nice if you could
> > share that.
> > I think I have asked this question before.....
> > 
> > 
> > 
> > --
> > View this message in context: 
> > http://octave.1599824.n4.nabble.com/Octave3-6-0-gcc4-6-2-mingw-binaries-for-testing-tp4338795p4343888.html
> > Sent from the Octave - Maintainers mailing list archive at Nabble.com.
> >
>