Blitz++ Library

Sorry about by silly questions (but I'm new at nummeric programming):

I looked at your image http://www.opencascade.org/upload/87/image_1.html 
and now I have some questions:

* What does MFLOPS mean (MegaFlops, I suppose, but what is this?)
* What is better? More or less MFLOPS?
* Why is there a maximum at a matrix size of 100? What does this mean?
* Could you use the ifc|icc instead of f77|gcc they should be much faster?
* Why is MTL so much better(worse) than the rest and why are ATLAS and 
INTEL_BLAS so much worse(better) than the rest?



Thank you for enlighten me :-)

Malte




Laurent Plagne wrote:
> Hi all,
> 
> I am testing C++ libraries to be used in my simulation project. In order 
> to measure relative performance of different available libs, I built a 
> generic benchmark application :
> 
> Project page :  http://www.opencascade.org/upload/87/index.html
> 
> Rather good performances (compared to f77 or C) can be achieved using 
> Blitz++ arrays and C style programming :
> 
> static inline void matrix_matrix_product(const gene_matrix & A, const
> gene_matrix & B, gene_matrix & X, int N)
>    {
> 
>      double somme;
>      for (int i=0;i<N;i++){
>        for (int j=0;j<N;j++){
>     somme=0.0;
>     for (int k=0;k<N;k++){
>       somme+=A(i,k)*B(k,j);
>     }
>     X(i,j)=somme;
>        }
>      }
>    }
> 
> with
> 
> typedef blitz::Array<double, 2>  gene_matrix;
> 
> (I have tried blitz::Matrix<double> gene_matrix but the in cache perfs 
> where slightly less good (5-10%))
> 
> Is it possible to write an efficient dense matrix matrix product using 
> condensed blitz syntax ?
> 
> Laurent Plagne
> 
> 
> _______________________________________________
> Blitz-support mailing list
> Blitz-support@...
> http://www.oonumerics.org/mailman/listinfo.cgi/blitz-support
>

Hello,

Malte Thoma wrote:

> Sorry about by silly questions (but I'm new at nummeric programming):

Your questions ared welcomed since this gives me hints to improve my 
html page.

> I looked at your image 
> http://www.opencascade.org/upload/87/image_1.html and now I have some 
> questions:
>
> * What does MFLOPS mean (MegaFlops, I suppose, but what is this?) 

Million of Floating points operation per second (more or less the speed)

* What is better? More or less MFLOPS?

More, it is evaluated as MFLOPS=nb_of_operations/(calculation_time*1000000)

> * Why is there a maximum at a matrix size of 100? What does this mean? 

Well, it is actually 1000. It is an arbitrary size. it means 1000x1000 
elements which is a large enough
size to be an "out of the cache calculation" and fit in ram (no swapping).

> * Could you use the ifc|icc instead of f77|gcc they should be much 
> faster? 

I guess that ifc and icc are both native compiler on a given 
architecture. The comparison
between ifc and g77 would probably be not fair because most good native 
fortran compiler
can identify a BLAS like calculation written with loops and replace it 
with call to
vendor-blas implementation. Good performances obtained this way depends 
on commercial
compiler and will not be portable (on a linux box for example).
For gcc, I think it is a very good compiler which probably compares well 
with vendor compiler in
most cases (especially with C code).

> * Why is MTL so much better(worse) than the rest and why are ATLAS and 
> INTEL_BLAS so much worse(better) than the rest?

MTL is in MY CASE very bad because this library is (AFAIK) inefficient 
with gcc (it requires very
clever compiler such as KCC which is dead by now). I still have hope to 
improve my MTL installation.

Some of your questions (mflops, MTL) are adressed in the page (click on 
comments links) but
seems to be uneasy to find...

Laurent

Hi Laurent,

Yes, see the section on tensor notation in the Blitz++ manual.
You would write it as follows:

using namespace blitz::tensor; // bring index placeholders i,j,k into scope
X = sum(A(i,k) * B(k,j), k);


Regards, Julian C.


Dr. Julian C. Cummings
Staff Scientist, CACR/Caltech
(626) 395-2543
cummings@...
 

> Is it possible to write an efficient dense matrix matrix product using 
> condensed blitz syntax ?
> 
> Laurent Plagne

Laurent Plagne wrote:

>> * Why is there a maximum at a matrix size of 100? What does this mean? 
>
>
> Well, it is actually 1000. It is an arbitrary size. it means 1000x1000 
> elements which is a large enough
> size to be an "out of the cache calculation" and fit in ram (no 
> swapping).

Sorry, if I havn't expressed acuralty, but at a matrix size of ~100 
there is a maximum speed (mflops) for the most
libarys. My question points out to this point.

>> * Could you use the ifc|icc instead of f77|gcc they should be much 
>> faster? 
>
>
> I guess that ifc and icc are both native compiler on a given 
> architecture. The comparison
> between ifc and g77 would probably be not fair because most good 
> native fortran compiler
> can identify a BLAS like calculation written with loops and replace it 
> with call to
> vendor-blas implementation. Good performances obtained this way 
> depends on commercial
> compiler and will not be portable (on a linux box for example).
> For gcc, I think it is a very good compiler which probably compares 
> well with vendor compiler in
> most cases (especially with C code).

OK, ifc and icc are comercial compilers, but for university it is free. 
So if I write a nummerical code for me it is of interest
how fast it is on which architectur with which compiler. Because the gnu 
compiler runs nearly everywhere this allways
be should be given as a reference speed, but why shouldn't compare nativ 
compilers?


>> * Why is MTL so much better(worse) than the rest and why are ATLAS 
>> and INTEL_BLAS so much worse(better) than the rest? 
>
Ok, if ATLAS and INTEL_BLAS (whatever it is, I havn't a clue) ar so much 
faster, why don't we use them and forget about blitz,ublas and even stl?

Greetings,

Malte


>>

Hi Julian,

Thank you very much for your support.
I do test this condensed implentation but it is much slower :

http://www.opencascade.org/upload/87/blitz_matrix_matrix.html

Any idea to improve it ?


Julian C. Cummings wrote:

>Hi Laurent,
>
>Yes, see the section on tensor notation in the Blitz++ manual.
>You would write it as follows:
>
>using namespace blitz::tensor; // bring index placeholders i,j,k into scope
>X = sum(A(i,k) * B(k,j), k);
>
>
>Regards, Julian C.
>
>
>Dr. Julian C. Cummings
>Staff Scientist, CACR/Caltech
>(626) 395-2543
>cummings@...
> 
>
>  
>
>>Is it possible to write an efficient dense matrix matrix product using 
>>condensed blitz syntax ?
>>
>>Laurent Plagne
>>    
>>
>
>_______________________________________________
>Blitz-support mailing list
>Blitz-support@...
>http://www.oonumerics.org/mailman/listinfo.cgi/blitz-support
>
>
>  
>

Laurent Plagne wrote:
> Hi Julian,
> 
> Thank you very much for your support.
> I do test this condensed implentation but it is much slower :
> 
> http://www.opencascade.org/upload/87/blitz_matrix_matrix.html
> 
> Any idea to improve it ?
> 
> 
> Julian C. Cummings wrote:
> 
>

Hi Julian,

I get the same results as Laurent's ones with the sum() operator : very 
very slow (at least with gcc). In fact i avoid to use it.

Xavier Warin