about latest versions

[MHELHEFNY]

Hello Michael !!

As you mentioned that both  VC++/3.0.3/SSE2  and  MinGW/3.0.2/SSE2  on  your
computer
gave , the same performance of about 1350 MFLOPS , What astonishes  me is
that MinGw/3.0.2
on my computer gives a performance of  3600  MFLOPS  which is more than 2.5
times more than
 that you have mentioned  and which is also 5 times more than  that of
Matlab (ver 6.5) which on my machine gave 660 MFLOPS  i.e. 1/5 of
MinGw/3.0.2 .
It is also  astonishing that Tatsuro's computer gave even more performance
:-
mingw(gcc-4.3.2-dw2-TDM)  MFLOPS = 4830.2 .
May ,I dare to ask you , tatsuro and others what is more important in a
mathematical software than performance?.
Best regards.
mhelhefny



----- Original Message -----
From: "Michael Goffioul" <michael.goffioul at gmail.com>
To: "MHELHEFNY" <mhelhefny at gmail.com>
Cc: <bug-octave-request at octave.org>
Sent: Sunday, April 19, 2009 12:43 PM
Subject: Re: about latest versions


>To be honest, I don't think there will be any difference in the next
>VC++ compiled
>version. I have no idea why the VC++/ATLAS release gives such lower
>performance for you, compared to MinGW/ATLAS. I did the test on my PC:
>Pentium M 1.7GHz, 1GB RAM; I tried VC++/3.0.3/SSE2 and
>MinGW/3.0.2/SSE2, and both gave the same performance, about
>1350 MFLOPS. By comparison, VC++/3.0.5/Generic gives about 480 MFLOPS.

>ATLAS is part of VC++ release, I compiled it on the same laptop.
>It would be very unfortunate that the optimization of ATLAS for my
>CPU does lead to such dramatic performance decrease on your CPU.

On Sun, Apr 19, 2009 at 9:13 AM, MHELHEFNY <mhelhefny at gmail.com> wrote:
>
> ----- Original Message -----
> From: "Michael Goffioul" <michael.goffioul at gmail.com>
> To: "MHELHEFNY" <mhelhefny at gmail.com>
> Cc: <tmacchant at yahoo.co.jp>; <bug at octave.org>
> Sent: Saturday, April 18, 2009 9:16 PM
> Subject: Re: about latest versions
>
>
> On Sat, Apr 18, 2009 at 1:45 PM, MHELHEFNY <mhelhefny at gmail.com> wrote:
>>
>> ----- Original Message -----
>> From: "Tatsuro MATSUOKA" <tmacchant at yahoo.co.jp>
>> To: "Michael Goffioul" <michael.goffioul at gmail.com>; "MHELHEFNY"
>> <mhelhefny at gmail.com>
>> Cc: <bug at octave.org>
>> Sent: Saturday, April 18, 2009 4:16 AM
>> Subject: Re: about latest versions
>>
>>
>>> Hello
>>>
>>> Benjamin's octave-mingw will check cpu and sse2 and sse3 versions of the
>> ATLAS exist.
>>> In my computer in my my uinversity, the cpu of which is presscotte
>> (pentium4-HT 3.4GHz),
>>> the following test have given,
>>> n=2000; A=randn(n); B=randn(n);tic; C=A*B; t=toc, MFLOPS=2*n^3/t*1e-6
>>> Benjamin 3.0.2 (mingw) MFLOPS = 1973.0
>>> Michael 3.0,3 (MSVC) MFLOPS = 1845.9
>>>
>>> The results were almost the same.
>>>
>>> I do not think that Atlas of Benjamis's Octave is optimized for the
older
>> CPUs.
>>>
>>> As a refrence, in case of self build Atlas for my computer
(octave-3.0.4)
>> using
>>> mingw(gcc-4.3.2-dw2-TDM)
>>> MFLOPS = 4830.2
>>> If you want to get highest performance for matrix calculation, the
octave
>> will be reccomended to build
>>> by yourself with the optimized the ATLAS (or GotoBLAS*) for you computer
>> system.
>>> However it is hard way to the beginners.
>>>
>>> *GotoBLAS can be used for the self building case because its licesence
is
>> not compatible the GPL.
>>>
>>> Regards
>>>
>>> Tatsuro
>>>
>>
>> Hello
>>
>> as you said on your cpu (pentium4-ht 3.4GHZ) the results were almost the
>> same. but alas on my cpu (pentium4 2.8GHZ) the results were different
>>
>> Benjamin 3.0.2 (mingw) MFLOPS = 3580.4
>> Michael 3.0.3 (MSVC) MFLOPS =318.25
>>
>> which is 12 times better with mingw both versions was ready on the web.
on
>> my computer , ver. 3.0.2 is even better two folds than on yours. That is
>> very strange.
>
>>The version 3.0.3 you installed does not use ATLAS, but basic
>>unoptimized BLAS/LAPACK. Try to re-install and at early stage,
>>select the right CPU.
>
>>Michael.
>
> Hi Michael !!
> thanks for your message,
> I have re-installed the windows octave 3.0.3; and it gave the following
> results,for the test proposed by Tatsuro as follows:
> for sse2 850 MFLOPS,
> for sse 620 MFLOPS,
> that is an impovement compared my old results of 318 MFLOPS but is still
> much less than 3580 MFLOPS, for the (mingw) version.
> However, I hope that the next windows vesion supplied by octave-forge will
> use ATLAS,
> Since I am an engineer and I am not ready to do it myself.
>
> Best Regards
> mhelhefny.
>
>