Calling Iterative Sparse Solver from C++, Preconditioning Problems, Multiple Right Hand Side,

[recage]

Problem: when solving a MRHS problem with the iterative sparse solver (called from C++), it does not seem to put the current residual in the right place, so that it can be utilized by the preconditioner (i.e. when RCI_Request==3).

For a single RHS problems the function 'dcg' puts the current residual into the array tmp(:,3), which can be used by the preconditioner. That's fine.
However, for multiple RHS problems the function 'dcgmrhs' is supposed to put the residual into tmp(:,3+ipar(3)), but this doesn't seem to be true. Hence, the preconditioner can not proceed with its job. How can I instruct 'dcgmrhs' to store the residuals in tmp(:,3+ipar(3))?
Or is there a misunderstanding on my side?

I have observed this behavior when calling the functions from C++.

Best regards, Recage

[Alexander_K_Intel2]

Quoting - recage

Problem: when solving a MRHS problem with the iterative sparse solver (called from C++), it does not seem to put the current residual in the right place, so that it can be utilized by the preconditioner (i.e. when RCI_Request==3).

For a single RHS problems the function 'dcg' puts the current residual into the array tmp(:,3), which can be used by the preconditioner. That's fine.
However, for multiple RHS problems the function 'dcgmrhs' is supposed to put the residual into tmp(:,3+ipar(3)), but this doesn't seem to be true. Hence, the preconditioner can not proceed with its job. How can I instruct 'dcgmrhs' to store the residuals in tmp(:,3+ipar(3))?
Or is there a misunderstanding on my side?

I have observed this behavior when calling the functions from C++.

Best regards, Recage

Hi Recage,

To more deep understanding situation could you send example of your code? The cause of my misunderstanding is that you wrote about C++ program but use Fortran definition.

With best regards,

Alexander Kalinkin

[recage]

Dear Alexander, thanks for replying again.

Here ist the code segment for the preconditioning. The full source of the function is in the attachments.

[cpp]/*---------------------------------------------------------------------------*/
	/* If rci_request=3, then compute apply the preconditioner matrix C_inverse  */
	/* on vector tmp(:,3+ipar(3)) and put the result in vector tmp(:,3)          */
	/* (Fortran notation)                                                        */
	/*---------------------------------------------------------------------------*/
	if (rci_request==3)
	{
		mkl_dcsrsv(&matdes[2], &n, &one, matdes, a, ja, ia, &ia[1], &tmp[(3+ipar[2]-1)*n], &tmp[2*n]);
		goto rci;
	}[/cpp]

Best regards,
Recage

[Alexander_K_Intel2 · Accepted solution]

Quoting - recage

Dear Alexander, thanks for replying again.

Here ist the code segment for the preconditioning. The full source of the function is in the attachments.

[cpp]/*---------------------------------------------------------------------------*/
	/* If rci_request=3, then compute apply the preconditioner matrix C_inverse  */
	/* on vector tmp(:,3+ipar(3)) and put the result in vector tmp(:,3)          */
	/* (Fortran notation)                                                        */
	/*---------------------------------------------------------------------------*/
	if (rci_request==3)
	{
		mkl_dcsrsv(&matdes[2], &n, &one, matdes, a, ja, ia, &ia[1], &tmp[(3+ipar[2]-1)*n], &tmp[2*n]);
		goto rci;
	}[/cpp]

Best regards,
Recage

Recage,

First of all, what is not good in such precondition... CG and CGMHRS work with symmetric positive define system, thats mean that precondition must be symmetric positive define! If it's not nobody can guarantee the right solution of your system by CGMHRS. If you interested how construct precondition based on inversion L and U parts of initial matrix, you can look incg_ssor_precon_c.c in MKL examples. In that example precondition named SSOR construct.

With best regards,

Alexander Kalinkin

[recage]

Quoting - Alexander Kalinkin (Intel)

Recage,

First of all, what is not good in such precondition... CG and CGMHRS work with symmetric positive define system, thats mean that precondition must be symmetric positive define! If it's not nobody can guarantee the right solution of your system by CGMHRS. If you interested how construct precondition based on inversion L and U parts of initial matrix, you can look incg_ssor_precon_c.c in MKL examples. In that example precondition named SSOR construct.

With best regards,

Alexander Kalinkin

Thanks Alexander. Based on your hints I have successfully implemented a preconditioner for CGMRHS.

Best regards, Recage

[Alexander_K_Intel2]

Quoting - recage

Thanks Alexander. Based on your hints I have successfully implemented a preconditioner for CGMRHS.

Best regards, Recage

Recage,
You are welcome, fill free to ask more :)
With best regards,
Alexander Kalinkin