ifort and lapack for dummies

[guillaume]

Hi

i'm currently trying to improve my fortran code using LAPACK routines. My compiler is ifort. But among the compiler options i've found to invoque lapack (-llapack...) not one worked.
Do you know any website where to understand clearly how to do that or can anyone explain me ? I'd like to found something like : "ifort -o tr.o -llapack tr.f90"...

Guillaume

[Julien Langou]

Bonjour Guillaume,
several options:

• either you make your own liblapack.a, for this you
  
  • get lapack.tgz from netlib http://www.netlib.org/lapack/lapack.tgz
    
  • apply the patch from: http://www.netlib.org/lapack-dev/lapack--3.0--patch--10042002.tgz
    
  • edit the make.inc file so that it uses ifort
    
  • type make install lib
    
  • this will create a lapack${PLAT}.a library
    
  • you certainly want an optimized BLAS library under LAPACK, have a look at: Where can I find optimized BLAS libraries?
• You might want to use MKL if you have on your system (it's the Intel LAPACK optimized and precompiled library). Or any precompiler LAPACK library.

I see that your main is in Fortran 90, be careful when you call LAPACK or BLAS routines. If you call for example: DGESV( N, NRHS, A, LDA, ... ) then there is lots of chance that your compiler makes a copy of A before calling DGESV. (I do not know specifically about ifort but let's be cautious.)

There is no much way to avoid this if you want to be compatible with the Fortran standard, however you can make some assumptions on the compiler and use a little trick to avoid the copy.

To avoid the copy, just call: DGESV( N, NRHS, A(1,1), LDA, ... ) . The compiler is not going to make any copy in that case.

The problem is that the latter approach ( DGESV( N, NRHS, A(1,1), LDA, ... ) ) assumes that A is stored in a contiguous piece of memory. And this is not necessarily guarantee by allocate. However this has always worked for me in practice.

Julien

[guillaume]

Thanks

now i think lapack is correctly installed. To begin with using lapack, i use such a command:
$ ifort /usr/lib/liblapack.so.2.0 truc.f

is there anything wrong with it (it seems to word with a few examples) ?

Guillaume

[Julien Langou]

If your compile line works with some examples, you should be good.

Where do you get your liblapack from?
In general you want to make sure you know which BLAS you are using behind LAPACK. In your compile line I do not see which BLAS you are using so I guess it is included in liblapack.so in which case I wonder how much optimized it is.

Also if you want to fully test your liblapack.so you can use the TESTING in the LAPACK distribution.

Julien

[guillaume]

I have run BLAS routines with others examples before, but without any special command line. I juste had to include : external [...] in the fortran code (for example : external dgeem, and nothing special for the compiler : ifort truc.f )
Is that strange ?
I will try to understand where my lapack comes from and to use this TESTING thing, and i reply again.
Thanks

[guillaume]

In fact my /usr/lib contains all this things (find /usr/lib -name "*lapack*" -print ):

./atlas/liblapack.so.2.3
./atlas/liblapack.so.3.0
./atlas/liblapack.so.2
./atlas/liblapack.so.3
./liblapack_atlas.so.2.3
./liblapack_atlas.so.2
./liblapack.so.2.0
./liblapack.so.2
./liblapack_atlas.so.3.0
./liblapack_atlas.so.3

Some seem to have the same function according to the examples i have run. I have installed some by myself but perharps some were already there...