Can I perform matrix-vector multiplication myself when using PARDISO ?

[afd_lml]

I want to use the PARDISO.

However, the matrix to be solved does not be explicitlygenerated, for example, the fast multiple method(FMM). In other words,I does not know the entries of the matrix,although I can carry out the matrix-vector multiplication.

So, the matrix itself can not be a input parameter for PARDISO.

Can I perform matrix-vector multiplication myself when using PARDISO ?

[ArturGuzik · Accepted solution]

Quoting - afd.lml

I want to use the PARDISO.

However, the matrix to be solved does not be explicitlygenerated, for example, the fast multiple method(FMM). In other words,I does not know the entries of the matrix,although I can carry out the matrix-vector multiplication.

So, the matrix itself can not be a input parameter for PARDISO.

Can I perform matrix-vector multiplication myself when using PARDISO ?

Usually FMM employs GMRES based algorithm for solution procedure. MKL offers FGMRES which allows you (actually requires, it is reverse communication type) to perform matrix-vector multiplication yourself.

I'm unaware of any way that PARDISO can be used in that fashion. Presumably somebody will provide some more definite answer to that.

A.

[Konstantin_A_Intel]

Hello,

Unfortunately, you cannot use PARDISO in a such way. It needs explicitly all matrix elements in order to make LLT or LU decompositions.

Artur gave yougood advice: use iterative solvers which support reverse communication interface.

Regards,
Konstantin