topic Re: Poisson Solver: issue with corner values in Intel® oneAPI Math Kernel Library

Poisson Solver: issue with corner values

ivanp — Sat, 08 Jun 2024 12:26:54 GMT

I'm attempting to use the Fast Poisson Solver routines from the oneMKL library. The driver, in Fortran, is provided in the attachment.

The solver works, however the corner values at locations (1,1) and (1,0) seem to be wrong, when compared with the output of a different solver.

What could be the problem?

Re: Poisson Solver: issue with corner values

Gennady_F_Intel — Mon, 10 Jun 2024 11:59:26 GMT

Ivan, I am not quite sure which results you are expecting to see here.

running with the current version of oneMKL, I see

1.00000000000000 0.000000000000000E+000 0.000000000000000E+000

.....
1.00000000000000 1.00000000000000 0.000000000000000E+000

Which results do you see with another solvers?

Re: Poisson Solver: issue with corner values

ivanp — Mon, 10 Jun 2024 12:10:45 GMT

I realized it was an error on my side. The boundary arrays are of length nx+1 and ny+1, but I was wrongly only setting the first nx, ny values. Hence, there was by default a 0.0 in the last value.

The boundary routine is supposed to read

subroutine fbd(nx,ny,x,y,bd_ax,bd_bx,bd_ay,bd_by) integer, intent(in) :: nx, ny real(dp), intent(in) :: x(nx+1), y(ny+1) real(dp), intent(out) :: bd_ax(ny+1), bd_bx(ny+1) real(dp), intent(out) :: bd_ay(nx+1), bd_by(nx+1) ! boundaries are given by ! u(x,y) = cos( pi x ) - sin( 2 pi y ) bd_ax = cos(pi*0.0_dp) - sin(2*pi*y) ! x = 0 bd_bx = cos(pi*1.0_dp) - sin(2*pi*y) ! x = 1 bd_ay = cos(pi*x) - sin(2*pi*0.0_dp) ! y = 0 bd_by = cos(pi*x) - sin(2*pi*1.0_dp) ! y = 1 end subroutine

Now the solution matches the multigrid solver in the link in my original post.

Poisson solution (correct)

Re:Poisson Solver: issue with corner values

Gennady_F_Intel — Mon, 10 Jun 2024 12:36:19 GMT

ok, then the thread is closing.