Diagonal complex matrix

[martas]

Hi, I am trying to use the BLAS/LAPACK libraries in C to perform some memory intensive calculations. One way to speed this up would be for me to produce a sparse (actually, diagonal) complex matrix, of size at least 1000 * 1000. I noticed that BLAS has methods to generate sparse matrices by listing the non zero entries and their indices, but I haven't been able to find a method to generate complex sparse/diagonal matrices, and I don't really want to populate an entire array with predominantly zeroes.

Does anyone have any ideas?

Thanks very much

Marta

[Danesh_D]

Hi Marta,

I think you have a simpler solution rather than using BLAS. You can simply save all elemenets of you matrix (I assume that it is diagonal) in a vector of 1000 elements. Then you can use ScaLAPACK functions which is tailored for diagonal matrices or simply write routines which for example adds a digonal matrix with another matrix. If you needed source code of these functions let me know.

Danesh

[martas]

Hi Danesh,

Thanks for your reply.

I am running 32 bit Windows, unfortunately, and it seems to me that in order to get the best performance out of ScaLapack, I need a 64 bit machine, or else the Linux OS.

Is that correct?

Sorry for my ignorance - I only learnt about ACML, BLAS and Lapack in the last couple of days. I had previously written my code in scilab, which allowed me to do element-wise matrix operations, so the things I am asking about were not an issue. But the calculations were too slow, which is why I am trying to switch to ACML.

Thanks

Marta

[Julien Langou]

Hello Marta,

you are right ACML has a kind of sparse BLAS implemented in it. I have never used
it personnally. It is not the exactly what we call the sparse BLAS though (I think).
From what I have understood they have an implementation of some routines
proposed in: David S. Dodson, Roger G. Grimes, John G. Lewis, Sparse Extensions to the
FORTRAN Basic Linear Algebra Subprograms, ACM Trans. Math. Soft., 17 (1991), pp.
253–263. If this is what you need you might go and try and use ACML.

Now for the BLAS in itself, I am not sure what kind of operation you want to do, but it
looks to me that you would be better off simply rewriting your Scilab code in C or Fortran.
The executable should be faster than the Scilab equivallent.

Your operation looks very specific that might be the best way to go.

Julien.

[martas]

Thanks to both of you for your help.

Marta