Re: Octave 3.6.4 VS2010 and the C++ API

[Mike Puglia]

---

From: Michael Goffioul <address@hidden>
To: Mike Puglia <address@hidden>
Cc: oCTAVE <address@hidden>
Sent: Saturday, September 28, 2013 12:29 AM
Subject: Re: Octave 3.6.4 VS2010 and the C++ API

On Fri, Sep 27, 2013 at 9:40 AM, Mike Puglia <address@hidden> wrote:

---

From: Michael Goffioul <address@hidden>
To: Mike Puglia <address@hidden>
Cc: oCTAVE <address@hidden>

Sent: Friday, September 27, 2013 10:16 PM

Subject: Re: Octave 3.6.4 VS2010 and the C++ API

On Fri, Sep 27, 2013 at 9:09 AM, Mike Puglia <address@hidden> wrote:

---

From: Michael Goffioul <address@hidden>
To: Mike Puglia <address@hidden>

Cc: oCTAVE <address@hidden>
Sent: Friday, September 27, 2013 9:45 PM

Subject: Re: Octave 3.6.4 VS2010 and the C++ API

On Fri, Sep 27, 2013 at 8:00 AM, Mike Puglia <address@hidden> wrote:

The new OpenBLAS dll you sent fixes the problem.  Thanks again for all the help.

>I didn't check the performances of that new version. Did you noticed any degradation (or improvement)?

 

I did a quick check using the Octave2.m benchmarker.  The VS2010  OpenBLAS was about 20% slower (all in matrix calculations) than MINGW OpenBLAS, and about the same as VS2010 SSE3.

>How does it perform compared to the previous OpenBLAS DLL, the one that is buggy?

>For the record, I also compile OpenBLAS with MinGW, it's not possible to compile it with MSVC.

It looks like the new DLL performs about the same as the old on my machine.

Ok, thanks. Could you tell me what flavour of OpenBLAS/MinGW you're using? Looking at [1], there'sa bunch of them. Also could you perform comparison based on this piece of code:

n=2000; A=randn(n); B=randn(n);tic; C=A*B; t=toc, MFLOPS=2*n^3/t*1e-6

Not sure why the variance is so high, but here are the results on my machine.  The new DLL is only average 8% slower in this sample. 

Old DLL:

t =  18.847, MFLOPS =  848.94

t =  15.633, MFLOPS =  1023.5

t =  20.960, MFLOPS =  763.36

t =  24.496, MFLOPS =  653.17

New DLL:

t =  16.124, MFLOPS =  992.31

t =  18.596, MFLOPS =  860.40

t =  23.083, MFLOPS =  693.15

t =  15.733, MFLOPS =  1017.0

[1] http://sourceforge.net/projects/octave/files/Octave%20Windows%20binaries/Octave%203.6.4%20for%20Windows%20MinGW%20installer/

 

This is the closest post I could find, but it didn't really apply (I'm not using linux or any GUI) and it didn't mean much to me:

http://savannah.gnu.org/bugs/?36868

I don't think this is the same. The fact that csrss.exe is also getting crazy indicates another type of problem, which may be specific to your system.

Okay, not sure myself either.  I'm just running a very simple windows XP machine that was wiped clean a few months ago and doesn't even have much software on it. 

Michael.