Fortran - Using Lapack with Atlas BLAS from C

Author

Using Lapack with Atlas BLAS from C

Harald Grossauer

2004-11-16, 3:57 pm

Hi,

I want to use Lapack routines making use of the optimized Atlas BLAS
routines and call them from a C application.

Since I have had troubles getting CLAPACK to run correctly I now thought
of the following:

I download and compile Lapack (Fortran version). In the Makefile I
specify the correct BLASLIB setting (i.e., my Atlas version, which is a
precompiled version from netlib).
This should then give me a full lapack library employing the Atlas BLAS
routines (as compared to the partial Lapack support in Atlas).
When writing a C program I only need a clapack.h header file which
contains the correct prototypes for the functions. Then I link my
executable against the Lapack library that I compiled before.

1. Does it work the way I think it does?
2. Is the performance "optimal" (i.e., as fast a Atlas can be)?

Thanks in advance,
Harald

Janne Blomqvist

2004-11-16, 3:57 pm

In article <419a33b4$1@sia.uibk.ac.at>, Harald Grossauer wrote:
> Hi,
>
> I want to use Lapack routines making use of the optimized Atlas BLAS
> routines and call them from a C application.
>
> Since I have had troubles getting CLAPACK to run correctly I now thought
> of the following:
>
> I download and compile Lapack (Fortran version). In the Makefile I
> specify the correct BLASLIB setting (i.e., my Atlas version, which is a
> precompiled version from netlib).
> This should then give me a full lapack library employing the Atlas BLAS
> routines (as compared to the partial Lapack support in Atlas).
> When writing a C program I only need a clapack.h header file which
> contains the correct prototypes for the functions. Then I link my
> executable against the Lapack library that I compiled before.
>
> 1. Does it work the way I think it does?

I guess you need to test it yourself.

> 2. Is the performance "optimal" (i.e., as fast a Atlas can be)?

Depends. The point of the atlas library is that during the build process it
runs an extensive set of tests which are designed to find optimal
settings for your particular computer (block size and such). Thus, by
using a precompiled version these settings are probably not optimal.

If you decide to compile atlas yourself, note that it takes a _very_
long time to compile (IIRC it was something like 18 h when I did it
some years ago on a 800 MHz PC).

Also, take a look at

http://math-atlas.sourceforge.net/e...html#completelp

--
Janne Blomqvist

Roman Werpachowski

2004-11-17, 3:59 pm

On the Tue, 16 Nov 2004 20:20:06 +0200 (EET), Janne Blomqvist wrote:

> If you decide to compile atlas yourself, note that it takes a _very_
> long time to compile (IIRC it was something like 18 h when I did it
> some years ago on a 800 MHz PC).

1-2 hours on an Athlon 1333MHz, about an hour on an Athlon 64 3000+ with 1GB
RAM. (memory is probably an issue here as well).

--
Roman Werpachowski
/--------==============--------\
| http://www.cft.edu.pl/~roman |
\--------==============--------/

Sponsored Links