!> @file poisfft_mod.f90 !------------------------------------------------------------------------------! ! This file is part of the PALM model system. ! ! PALM is free software: you can redistribute it and/or modify it under the ! terms of the GNU General Public License as published by the Free Software ! Foundation, either version 3 of the License, or (at your option) any later ! version. ! ! PALM is distributed in the hope that it will be useful, but WITHOUT ANY ! WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR ! A PARTICULAR PURPOSE. See the GNU General Public License for more details. ! ! You should have received a copy of the GNU General Public License along with ! PALM. If not, see . ! ! Copyright 1997-2020 Leibniz Universitaet Hannover !------------------------------------------------------------------------------! ! ! Current revisions: ! ----------------- ! ! ! Former revisions: ! ----------------- ! $Id: poisfft_mod.f90 4366 2020-01-09 08:12:43Z monakurppa $ ! modification concerning NEC vectorizatio ! ! 4360 2020-01-07 11:25:50Z suehring ! Corrected "Former revisions" section ! ! 3690 2019-01-22 22:56:42Z knoop ! OpenACC port for SPEC ! ! Revision 1.1 1997/07/24 11:24:14 raasch ! Initial revision ! ! ! Description: ! ------------ !> Solves the Poisson equation with a 2D spectral method !> d^2 p / dx^2 + d^2 p / dy^2 + d^2 p / dz^2 = s !> !> Input: !> real ar contains (nnz,nny,nnx) elements of the velocity divergence, !> starting from (1,nys,nxl) !> !> Output: !> real ar contains the solution for perturbation pressure p !------------------------------------------------------------------------------! MODULE poisfft_mod USE fft_xy, & ONLY: fft_init, fft_y, fft_y_1d, fft_y_m, fft_x, fft_x_1d, fft_x_m, & temperton_fft_vec USE indices, & ONLY: nnx, nny, nx, nxl, nxr, ny, nys, nyn, nz USE transpose_indices, & ONLY: nxl_y, nxl_z, nxr_y, nxr_z, nys_x, nys_z, nyn_x, nyn_z, nzb_x, & nzb_y, nzt_x, nzt_y USE tridia_solver, & ONLY: tridia_1dd, tridia_init, tridia_substi, tridia_substi_overlap IMPLICIT NONE LOGICAL, SAVE :: poisfft_initialized = .FALSE. PRIVATE PUBLIC poisfft, poisfft_init INTERFACE poisfft MODULE PROCEDURE poisfft END INTERFACE poisfft INTERFACE poisfft_init MODULE PROCEDURE poisfft_init END INTERFACE poisfft_init CONTAINS !------------------------------------------------------------------------------! ! Description: ! ------------ !> Setup coefficients for FFT and the tridiagonal solver !------------------------------------------------------------------------------! SUBROUTINE poisfft_init IMPLICIT NONE CALL fft_init CALL tridia_init poisfft_initialized = .TRUE. END SUBROUTINE poisfft_init !------------------------------------------------------------------------------! ! Description: ! ------------ !> Two-dimensional Fourier Transformation in x- and y-direction. !------------------------------------------------------------------------------! SUBROUTINE poisfft( ar ) USE control_parameters, & ONLY: transpose_compute_overlap USE cpulog, & ONLY: cpu_log, cpu_log_nowait, log_point_s USE kinds USE pegrid IMPLICIT NONE INTEGER(iwp) :: ii !< INTEGER(iwp) :: iind !< INTEGER(iwp) :: inew !< INTEGER(iwp) :: jj !< INTEGER(iwp) :: jind !< INTEGER(iwp) :: jnew !< INTEGER(iwp) :: ki !< INTEGER(iwp) :: kk !< INTEGER(iwp) :: knew !< INTEGER(iwp) :: n !< INTEGER(iwp) :: nblk !< INTEGER(iwp) :: nnx_y !< INTEGER(iwp) :: nny_z !< INTEGER(iwp) :: nnz_x !< INTEGER(iwp) :: nxl_y_bound !< INTEGER(iwp) :: nxr_y_bound !< INTEGER(iwp), DIMENSION(4) :: isave !< REAL(wp), DIMENSION(1:nz,nys:nyn,nxl:nxr) :: ar !< REAL(wp), DIMENSION(nys:nyn,nxl:nxr,1:nz) :: ar_inv !< #define __acc_fft_device ( defined( _OPENACC ) && ( defined ( __cuda_fft ) ) ) #if __acc_fft_device !$ACC DECLARE CREATE(ar_inv) #endif REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: ar1 !< REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: f_in !< REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: f_inv !< REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: f_out_y !< REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: f_out_z !< CALL cpu_log( log_point_s(3), 'poisfft', 'start' ) IF ( .NOT. poisfft_initialized ) CALL poisfft_init #if !__acc_fft_device !$ACC UPDATE HOST(ar) #endif #ifndef _OPENACC ! !-- Two-dimensional Fourier Transformation in x- and y-direction. IF ( pdims(2) == 1 .AND. pdims(1) > 1 ) THEN ! !-- 1d-domain-decomposition along x: !-- FFT along y and transposition y --> x CALL ffty_tr_yx( ar, ar ) ! !-- FFT along x, solving the tridiagonal system and backward FFT CALL fftx_tri_fftx( ar ) ! !-- Transposition x --> y and backward FFT along y CALL tr_xy_ffty( ar, ar ) ELSEIF ( pdims(1) == 1 .AND. pdims(2) > 1 ) THEN ! !-- 1d-domain-decomposition along y: !-- FFT along x and transposition x --> y CALL fftx_tr_xy( ar, ar ) ! !-- FFT along y, solving the tridiagonal system and backward FFT CALL ffty_tri_ffty( ar ) ! !-- Transposition y --> x and backward FFT along x CALL tr_yx_fftx( ar, ar ) ELSEIF ( .NOT. transpose_compute_overlap ) THEN #endif ! !-- 2d-domain-decomposition or no decomposition (1 PE run) !-- Transposition z --> x CALL cpu_log( log_point_s(5), 'transpo forward', 'start' ) CALL resort_for_zx( ar, ar_inv ) CALL transpose_zx( ar_inv, ar ) CALL cpu_log( log_point_s(5), 'transpo forward', 'pause' ) CALL cpu_log( log_point_s(4), 'fft_x', 'start' ) IF ( temperton_fft_vec ) THEN ! !-- Vector version outputs a transformed array ar_inv that does not require resorting !-- (which is done for ar further below) CALL fft_x( ar, 'forward', ar_inv=ar_inv) ELSE CALL fft_x( ar, 'forward') ENDIF CALL cpu_log( log_point_s(4), 'fft_x', 'pause' ) ! !-- Transposition x --> y CALL cpu_log( log_point_s(5), 'transpo forward', 'continue' ) IF( .NOT. temperton_fft_vec ) CALL resort_for_xy( ar, ar_inv ) CALL transpose_xy( ar_inv, ar ) CALL cpu_log( log_point_s(5), 'transpo forward', 'pause' ) CALL cpu_log( log_point_s(7), 'fft_y', 'start' ) IF ( temperton_fft_vec ) THEN ! !-- Input array ar_inv from fft_x can be directly used here. !-- The output (also in array ar_inv) does not require resorting below. CALL fft_y( ar, 'forward', ar_inv = ar_inv, nxl_y_bound = nxl_y, nxr_y_bound = nxr_y, & nxl_y_l = nxl_y, nxr_y_l = nxr_y ) ELSE CALL fft_y( ar, 'forward', ar_tr = ar, nxl_y_bound = nxl_y, nxr_y_bound = nxr_y, & nxl_y_l = nxl_y, nxr_y_l = nxr_y ) ENDIF CALL cpu_log( log_point_s(7), 'fft_y', 'pause' ) ! !-- Transposition y --> z CALL cpu_log( log_point_s(5), 'transpo forward', 'continue' ) IF ( .NOT. temperton_fft_vec ) CALL resort_for_yz( ar, ar_inv ) CALL transpose_yz( ar_inv, ar ) CALL cpu_log( log_point_s(5), 'transpo forward', 'stop' ) ! !-- Solve the tridiagonal equation system along z CALL cpu_log( log_point_s(6), 'tridia', 'start' ) CALL tridia_substi( ar ) CALL cpu_log( log_point_s(6), 'tridia', 'stop' ) ! !-- Inverse Fourier Transformation !-- Transposition z --> y CALL cpu_log( log_point_s(8), 'transpo invers', 'start' ) CALL transpose_zy( ar, ar_inv ) ! !-- The fft_y below (vector branch) can directly process ar_inv (i.e. does not require a !-- resorting) IF ( .NOT. temperton_fft_vec ) CALL resort_for_zy( ar_inv, ar ) CALL cpu_log( log_point_s(8), 'transpo invers', 'pause' ) CALL cpu_log( log_point_s(7), 'fft_y', 'continue' ) IF ( temperton_fft_vec ) THEN ! !-- Output array ar_inv can be used as input to the below fft_x routine without resorting CALL fft_y( ar, 'backward', ar_inv = ar_inv, nxl_y_bound = nxl_y, nxr_y_bound = nxr_y,& nxl_y_l = nxl_y, nxr_y_l = nxr_y ) ELSE CALL fft_y( ar, 'backward', ar_tr = ar, nxl_y_bound = nxl_y, nxr_y_bound = nxr_y, & nxl_y_l = nxl_y, nxr_y_l = nxr_y ) ENDIF CALL cpu_log( log_point_s(7), 'fft_y', 'stop' ) ! !-- Transposition y --> x CALL cpu_log( log_point_s(8), 'transpo invers', 'continue' ) CALL transpose_yx( ar, ar_inv ) IF ( .NOT. temperton_fft_vec ) CALL resort_for_yx( ar_inv, ar ) CALL cpu_log( log_point_s(8), 'transpo invers', 'pause' ) CALL cpu_log( log_point_s(4), 'fft_x', 'continue' ) IF ( temperton_fft_vec ) THEN CALL fft_x( ar, 'backward', ar_inv=ar_inv ) ELSE CALL fft_x( ar, 'backward' ) ENDIF CALL cpu_log( log_point_s(4), 'fft_x', 'stop' ) ! !-- Transposition x --> z CALL cpu_log( log_point_s(8), 'transpo invers', 'continue' ) CALL transpose_xz( ar, ar_inv ) CALL resort_for_xz( ar_inv, ar ) CALL cpu_log( log_point_s(8), 'transpo invers', 'stop' ) #ifndef _OPENACC ELSE ! !-- 2d-domain-decomposition or no decomposition (1 PE run) with !-- overlapping transposition / fft !-- cputime logging must not use barriers, which would prevent overlapping ALLOCATE( f_out_y(0:ny,nxl_y:nxr_y,nzb_y:nzt_y), & f_out_z(0:nx,nys_x:nyn_x,nzb_x:nzt_x) ) ! !-- Transposition z --> x + subsequent fft along x ALLOCATE( f_inv(nys:nyn,nxl:nxr,1:nz) ) CALL resort_for_zx( ar, f_inv ) ! !-- Save original indices and gridpoint counter isave(1) = nz isave(2) = nzb_x isave(3) = nzt_x isave(4) = sendrecvcount_zx ! !-- Set new indices for transformation nblk = nz / pdims(1) nz = pdims(1) nnz_x = 1 nzb_x = 1 + myidx * nnz_x nzt_x = ( myidx + 1 ) * nnz_x sendrecvcount_zx = nnx * nny * nnz_x ALLOCATE( ar1(0:nx,nys_x:nyn_x,nzb_x:nzt_x) ) ALLOCATE( f_in(nys:nyn,nxl:nxr,1:nz) ) DO kk = 1, nblk IF ( kk == 1 ) THEN CALL cpu_log( log_point_s(5), 'transpo forward', 'start', cpu_log_nowait ) ELSE CALL cpu_log( log_point_s(5), 'transpo forward', 'continue', cpu_log_nowait ) ENDIF DO knew = 1, nz ki = kk + nblk * ( knew - 1 ) f_in(:,:,knew) = f_inv(:,:,ki) ENDDO CALL transpose_zx( f_in, ar1(:,:,:)) CALL cpu_log( log_point_s(5), 'transpo forward', 'pause' ) IF ( kk == 1 ) THEN CALL cpu_log( log_point_s(4), 'fft_x', 'start', cpu_log_nowait ) ELSE CALL cpu_log( log_point_s(4), 'fft_x', 'continue', cpu_log_nowait ) ENDIF n = isave(2) + kk - 1 CALL fft_x( ar1(:,:,:), 'forward', ar_2d = f_out_z(:,:,n)) CALL cpu_log( log_point_s(4), 'fft_x', 'pause' ) ENDDO ! !-- Restore original indices/counters nz = isave(1) nzb_x = isave(2) nzt_x = isave(3) sendrecvcount_zx = isave(4) DEALLOCATE( ar1, f_in, f_inv ) ! !-- Transposition x --> y + subsequent fft along y ALLOCATE( f_inv(nys_x:nyn_x,nzb_x:nzt_x,0:nx) ) CALL resort_for_xy( f_out_z, f_inv ) ! !-- Save original indices and gridpoint counter isave(1) = nx isave(2) = nxl_y isave(3) = nxr_y isave(4) = sendrecvcount_xy ! !-- Set new indices for transformation nblk = ( ( nx+1 ) / pdims(2) ) - 1 nx = pdims(2) nnx_y = 1 nxl_y = myidy * nnx_y nxr_y = ( myidy + 1 ) * nnx_y - 1 sendrecvcount_xy = nnx_y * ( nyn_x-nys_x+1 ) * ( nzt_x-nzb_x+1 ) ALLOCATE( ar1(0:ny,nxl_y:nxr_y,nzb_y:nzt_y) ) ALLOCATE( f_in(nys_x:nyn_x,nzb_x:nzt_x,0:nx) ) DO ii = 0, nblk CALL cpu_log( log_point_s(5), 'transpo forward', 'continue', cpu_log_nowait ) DO inew = 0, nx-1 iind = ii + ( nblk + 1 ) * inew f_in(:,:,inew) = f_inv(:,:,iind) ENDDO CALL transpose_xy( f_in, ar1(:,:,:) ) CALL cpu_log( log_point_s(5), 'transpo forward', 'pause' ) IF ( ii == 1 ) THEN CALL cpu_log( log_point_s(7), 'fft_y', 'start', cpu_log_nowait ) ELSE CALL cpu_log( log_point_s(7), 'fft_y', 'continue', cpu_log_nowait ) ENDIF nxl_y_bound = isave(2) nxr_y_bound = isave(3) n = isave(2) + ii CALL fft_y( ar1(:,:,:), 'forward', ar_tr = f_out_y, & nxl_y_bound = nxl_y_bound, nxr_y_bound = nxr_y_bound, & nxl_y_l = n, nxr_y_l = n ) CALL cpu_log( log_point_s(7), 'fft_y', 'pause' ) ENDDO ! !-- Restore original indices/counters nx = isave(1) nxl_y = isave(2) nxr_y = isave(3) sendrecvcount_xy = isave(4) DEALLOCATE( ar1, f_in, f_inv ) ! !-- Transposition y --> z + subsequent tridia + resort for z --> y ALLOCATE( f_inv(nxl_y:nxr_y,nzb_y:nzt_y,0:ny) ) CALL resort_for_yz( f_out_y, f_inv ) ! !-- Save original indices and gridpoint counter isave(1) = ny isave(2) = nys_z isave(3) = nyn_z isave(4) = sendrecvcount_yz ! !-- Set new indices for transformation nblk = ( ( ny+1 ) / pdims(1) ) - 1 ny = pdims(1) nny_z = 1 nys_z = myidx * nny_z nyn_z = ( myidx + 1 ) * nny_z - 1 sendrecvcount_yz = ( nxr_y-nxl_y+1 ) * nny_z * ( nzt_y-nzb_y+1 ) ALLOCATE( ar1(nxl_z:nxr_z,nys_z:nyn_z,1:nz) ) ALLOCATE( f_in(nxl_y:nxr_y,nzb_y:nzt_y,0:ny) ) DO jj = 0, nblk ! !-- Forward Fourier Transformation !-- Transposition y --> z CALL cpu_log( log_point_s(5), 'transpo forward', 'continue', cpu_log_nowait ) DO jnew = 0, ny-1 jind = jj + ( nblk + 1 ) * jnew f_in(:,:,jnew) = f_inv(:,:,jind) ENDDO CALL transpose_yz( f_in, ar1(:,:,:) ) IF ( jj == nblk ) THEN CALL cpu_log( log_point_s(5), 'transpo forward', 'stop' ) ELSE CALL cpu_log( log_point_s(5), 'transpo forward', 'pause' ) ENDIF ! !-- Solve the tridiagonal equation system along z CALL cpu_log( log_point_s(6), 'tridia', 'start', cpu_log_nowait ) n = isave(2) + jj CALL tridia_substi_overlap( ar1(:,:,:), n ) CALL cpu_log( log_point_s(6), 'tridia', 'stop' ) ! !-- Inverse Fourier Transformation !-- Transposition z --> y !-- Only one thread should call MPI routines, therefore forward and !-- backward tranpose are in the same section IF ( jj == 0 ) THEN CALL cpu_log( log_point_s(8), 'transpo invers', 'start', cpu_log_nowait ) ELSE CALL cpu_log( log_point_s(8), 'transpo invers', 'continue', cpu_log_nowait ) ENDIF CALL transpose_zy( ar1(:,:,:), f_in ) DO jnew = 0, ny-1 jind = jj + ( nblk + 1 ) * jnew f_inv(:,:,jind) = f_in(:,:,jnew) ENDDO CALL cpu_log( log_point_s(8), 'transpo invers', 'pause' ) ENDDO ! !-- Restore original indices/counters ny = isave(1) nys_z = isave(2) nyn_z = isave(3) sendrecvcount_yz = isave(4) CALL resort_for_zy( f_inv, f_out_y ) DEALLOCATE( ar1, f_in, f_inv ) ! !-- fft along y backward + subsequent transposition y --> x ALLOCATE( f_inv(nys_x:nyn_x,nzb_x:nzt_x,0:nx) ) ! !-- Save original indices and gridpoint counter isave(1) = nx isave(2) = nxl_y isave(3) = nxr_y isave(4) = sendrecvcount_xy ! !-- Set new indices for transformation nblk = (( nx+1 ) / pdims(2) ) - 1 nx = pdims(2) nnx_y = 1 nxl_y = myidy * nnx_y nxr_y = ( myidy + 1 ) * nnx_y - 1 sendrecvcount_xy = nnx_y * ( nyn_x-nys_x+1 ) * ( nzt_x-nzb_x+1 ) ALLOCATE( ar1(0:ny,nxl_y:nxr_y,nzb_y:nzt_y) ) ALLOCATE( f_in(nys_x:nyn_x,nzb_x:nzt_x,0:nx) ) DO ii = 0, nblk CALL cpu_log( log_point_s(7), 'fft_y', 'continue', cpu_log_nowait ) n = isave(2) + ii nxl_y_bound = isave(2) nxr_y_bound = isave(3) CALL fft_y( ar1(:,:,:), 'backward', ar_tr = f_out_y, & nxl_y_bound = nxl_y_bound, nxr_y_bound = nxr_y_bound, & nxl_y_l = n, nxr_y_l = n ) IF ( ii == nblk ) THEN CALL cpu_log( log_point_s(7), 'fft_y', 'stop' ) ELSE CALL cpu_log( log_point_s(7), 'fft_y', 'pause' ) ENDIF CALL cpu_log( log_point_s(8), 'transpo invers', 'continue', cpu_log_nowait ) CALL transpose_yx( ar1(:,:,:), f_in ) DO inew = 0, nx-1 iind = ii + (nblk+1) * inew f_inv(:,:,iind) = f_in(:,:,inew) ENDDO CALL cpu_log( log_point_s(8), 'transpo invers', 'pause' ) ENDDO ! !-- Restore original indices/counters nx = isave(1) nxl_y = isave(2) nxr_y = isave(3) sendrecvcount_xy = isave(4) CALL resort_for_yx( f_inv, f_out_z ) DEALLOCATE( ar1, f_in, f_inv ) ! !-- fft along x backward + subsequent final transposition x --> z ALLOCATE( f_inv(nys:nyn,nxl:nxr,1:nz) ) ! !-- Save original indices and gridpoint counter isave(1) = nz isave(2) = nzb_x isave(3) = nzt_x isave(4) = sendrecvcount_zx ! !-- Set new indices for transformation nblk = nz / pdims(1) nz = pdims(1) nnz_x = 1 nzb_x = 1 + myidx * nnz_x nzt_x = ( myidx + 1 ) * nnz_x sendrecvcount_zx = nnx * nny * nnz_x ALLOCATE( ar1(0:nx,nys_x:nyn_x,nzb_x:nzt_x) ) ALLOCATE( f_in(nys:nyn,nxl:nxr,1:nz) ) DO kk = 1, nblk CALL cpu_log( log_point_s(4), 'fft_x', 'continue', cpu_log_nowait ) n = isave(2) + kk - 1 CALL fft_x( ar1(:,:,:), 'backward', f_out_z(:,:,n)) IF ( kk == nblk ) THEN CALL cpu_log( log_point_s(4), 'fft_x', 'stop' ) ELSE CALL cpu_log( log_point_s(4), 'fft_x', 'pause' ) ENDIF CALL cpu_log( log_point_s(8), 'transpo invers', 'continue', cpu_log_nowait ) CALL transpose_xz( ar1(:,:,:), f_in ) DO knew = 1, nz ki = kk + nblk * (knew-1) f_inv(:,:,ki) = f_in(:,:,knew) ENDDO IF ( kk == nblk ) THEN CALL cpu_log( log_point_s(8), 'transpo invers', 'stop' ) ELSE CALL cpu_log( log_point_s(8), 'transpo invers', 'pause' ) ENDIF ENDDO ! !-- Restore original indices/counters nz = isave(1) nzb_x = isave(2) nzt_x = isave(3) sendrecvcount_zx = isave(4) CALL resort_for_xz( f_inv, ar ) DEALLOCATE( ar1, f_in, f_inv ) ENDIF #endif #if !__acc_fft_device !$ACC UPDATE DEVICE(ar) #endif CALL cpu_log( log_point_s(3), 'poisfft', 'stop' ) END SUBROUTINE poisfft !------------------------------------------------------------------------------! ! Description: ! ------------ !> Fourier-transformation along y with subsequent transposition y --> x for !> a 1d-decomposition along x. !> !> @attention The performance of this routine is much faster on the NEC-SX6, !> if the first index of work_ffty_vec is odd. Otherwise !> memory bank conflicts may occur (especially if the index is a !> multiple of 128). That's why work_ffty_vec is dimensioned as !> 0:ny+1. !> Of course, this will not work if users are using an odd number !> of gridpoints along y. !------------------------------------------------------------------------------! SUBROUTINE ffty_tr_yx( f_in, f_out ) USE control_parameters, & ONLY: loop_optimization USE cpulog, & ONLY: cpu_log, log_point_s USE kinds USE pegrid IMPLICIT NONE INTEGER(iwp) :: i !< INTEGER(iwp) :: iend !< INTEGER(iwp) :: iouter !< INTEGER(iwp) :: ir !< INTEGER(iwp) :: j !< INTEGER(iwp) :: k !< INTEGER(iwp), PARAMETER :: stridex = 4 !< REAL(wp), DIMENSION(1:nz,0:ny,nxl:nxr) :: f_in !< REAL(wp), DIMENSION(nnx,1:nz,nys_x:nyn_x,pdims(1)) :: f_out !< REAL(wp), DIMENSION(nxl:nxr,1:nz,0:ny) :: work !< REAL(wp), DIMENSION(:,:), ALLOCATABLE :: work_ffty !< REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: work_ffty_vec !< ! !-- Carry out the FFT along y, where all data are present due to the !-- 1d-decomposition along x. Resort the data in a way that x becomes !-- the first index. CALL cpu_log( log_point_s(7), 'fft_y_1d', 'start' ) IF ( loop_optimization == 'vector' ) THEN ALLOCATE( work_ffty_vec(0:ny+1,1:nz,nxl:nxr) ) ! !-- Code optimized for vector processors !$OMP PARALLEL PRIVATE ( i, j, k ) !$OMP DO DO i = nxl, nxr DO j = 0, ny DO k = 1, nz work_ffty_vec(j,k,i) = f_in(k,j,i) ENDDO ENDDO CALL fft_y_m( work_ffty_vec(:,:,i), ny+1, 'forward' ) ENDDO !$OMP DO DO k = 1, nz DO j = 0, ny DO i = nxl, nxr work(i,k,j) = work_ffty_vec(j,k,i) ENDDO ENDDO ENDDO !$OMP END PARALLEL DEALLOCATE( work_ffty_vec ) ELSE ! !-- Cache optimized code. ALLOCATE( work_ffty(0:ny,stridex) ) ! !-- The i-(x-)direction is split into a strided outer loop and an inner !-- loop for better cache performance !$OMP PARALLEL PRIVATE (i,iend,iouter,ir,j,k,work_ffty) !$OMP DO DO iouter = nxl, nxr, stridex iend = MIN( iouter+stridex-1, nxr ) ! Upper bound for inner i loop DO k = 1, nz DO i = iouter, iend ir = i-iouter+1 ! counter within a stride DO j = 0, ny work_ffty(j,ir) = f_in(k,j,i) ENDDO ! !-- FFT along y CALL fft_y_1d( work_ffty(:,ir), 'forward' ) ENDDO ! !-- Resort DO j = 0, ny DO i = iouter, iend work(i,k,j) = work_ffty(j,i-iouter+1) ENDDO ENDDO ENDDO ENDDO !$OMP END PARALLEL DEALLOCATE( work_ffty ) ENDIF CALL cpu_log( log_point_s(7), 'fft_y_1d', 'pause' ) ! !-- Transpose array #if defined( __parallel ) CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) IF ( collective_wait ) CALL MPI_BARRIER( comm2d, ierr ) CALL MPI_ALLTOALL( work(nxl,1,0), sendrecvcount_xy, MPI_REAL, & f_out(1,1,nys_x,1), sendrecvcount_xy, MPI_REAL, & comm1dx, ierr ) CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) #endif END SUBROUTINE ffty_tr_yx !------------------------------------------------------------------------------! ! Description: ! ------------ !> Transposition x --> y with a subsequent backward Fourier transformation for !> a 1d-decomposition along x !------------------------------------------------------------------------------! SUBROUTINE tr_xy_ffty( f_in, f_out ) USE control_parameters, & ONLY: loop_optimization USE cpulog, & ONLY: cpu_log, log_point_s USE kinds USE pegrid IMPLICIT NONE INTEGER(iwp) :: i !< INTEGER(iwp) :: iend !< INTEGER(iwp) :: iouter !< INTEGER(iwp) :: ir !< INTEGER(iwp) :: j !< INTEGER(iwp) :: k !< INTEGER(iwp), PARAMETER :: stridex = 4 !< REAL(wp), DIMENSION(nnx,1:nz,nys_x:nyn_x,pdims(1)) :: f_in !< REAL(wp), DIMENSION(1:nz,0:ny,nxl:nxr) :: f_out !< REAL(wp), DIMENSION(nxl:nxr,1:nz,0:ny) :: work !< REAL(wp), DIMENSION(:,:), ALLOCATABLE :: work_ffty !< REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: work_ffty_vec !< ! !-- Transpose array #if defined( __parallel ) CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) IF ( collective_wait ) CALL MPI_BARRIER( comm2d, ierr ) CALL MPI_ALLTOALL( f_in(1,1,nys_x,1), sendrecvcount_xy, MPI_REAL, & work(nxl,1,0), sendrecvcount_xy, MPI_REAL, & comm1dx, ierr ) CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) #endif ! !-- Resort the data in a way that y becomes the first index and carry out the !-- backward fft along y. CALL cpu_log( log_point_s(7), 'fft_y_1d', 'continue' ) IF ( loop_optimization == 'vector' ) THEN ALLOCATE( work_ffty_vec(0:ny+1,1:nz,nxl:nxr) ) ! !-- Code optimized for vector processors !$OMP PARALLEL PRIVATE ( i, j, k ) !$OMP DO DO k = 1, nz DO j = 0, ny DO i = nxl, nxr work_ffty_vec(j,k,i) = work(i,k,j) ENDDO ENDDO ENDDO !$OMP DO DO i = nxl, nxr CALL fft_y_m( work_ffty_vec(:,:,i), ny+1, 'backward' ) DO j = 0, ny DO k = 1, nz f_out(k,j,i) = work_ffty_vec(j,k,i) ENDDO ENDDO ENDDO !$OMP END PARALLEL DEALLOCATE( work_ffty_vec ) ELSE ! !-- Cache optimized code. ALLOCATE( work_ffty(0:ny,stridex) ) ! !-- The i-(x-)direction is split into a strided outer loop and an inner !-- loop for better cache performance !$OMP PARALLEL PRIVATE ( i, iend, iouter, ir, j, k, work_ffty ) !$OMP DO DO iouter = nxl, nxr, stridex iend = MIN( iouter+stridex-1, nxr ) ! Upper bound for inner i loop DO k = 1, nz ! !-- Resort DO j = 0, ny DO i = iouter, iend work_ffty(j,i-iouter+1) = work(i,k,j) ENDDO ENDDO DO i = iouter, iend ! !-- FFT along y ir = i-iouter+1 ! counter within a stride CALL fft_y_1d( work_ffty(:,ir), 'backward' ) DO j = 0, ny f_out(k,j,i) = work_ffty(j,ir) ENDDO ENDDO ENDDO ENDDO !$OMP END PARALLEL DEALLOCATE( work_ffty ) ENDIF CALL cpu_log( log_point_s(7), 'fft_y_1d', 'stop' ) END SUBROUTINE tr_xy_ffty !------------------------------------------------------------------------------! ! Description: ! ------------ !> FFT along x, solution of the tridiagonal system and backward FFT for !> a 1d-decomposition along x !> !> @warning this subroutine may still not work for hybrid parallelization !> with OpenMP (for possible necessary changes see the original !> routine poisfft_hybrid, developed by Klaus Ketelsen, May 2002) !------------------------------------------------------------------------------! SUBROUTINE fftx_tri_fftx( ar ) USE control_parameters, & ONLY: loop_optimization USE cpulog, & ONLY: cpu_log, log_point_s USE grid_variables, & ONLY: ddx2, ddy2 USE kinds USE pegrid IMPLICIT NONE INTEGER(iwp) :: i !< INTEGER(iwp) :: j !< INTEGER(iwp) :: k !< INTEGER(iwp) :: m !< INTEGER(iwp) :: n !< !$ INTEGER(iwp) :: omp_get_thread_num !< INTEGER(iwp) :: tn !< REAL(wp), DIMENSION(0:nx) :: work_fftx !< REAL(wp), DIMENSION(0:nx,1:nz) :: work_trix !< REAL(wp), DIMENSION(nnx,1:nz,nys_x:nyn_x,pdims(1)) :: ar !< REAL(wp), DIMENSION(:,:,:,:), ALLOCATABLE :: tri !< CALL cpu_log( log_point_s(33), 'fft_x_1d + tridia', 'start' ) ALLOCATE( tri(5,0:nx,0:nz-1,0:threads_per_task-1) ) tn = 0 ! Default thread number in case of one thread !$OMP PARALLEL DO PRIVATE ( i, j, k, m, n, tn, work_fftx, work_trix ) DO j = nys_x, nyn_x !$ tn = omp_get_thread_num() IF ( loop_optimization == 'vector' ) THEN ! !-- Code optimized for vector processors DO k = 1, nz m = 0 DO n = 1, pdims(1) DO i = 1, nnx work_trix(m,k) = ar(i,k,j,n) m = m + 1 ENDDO ENDDO ENDDO CALL fft_x_m( work_trix, 'forward' ) ELSE ! !-- Cache optimized code DO k = 1, nz m = 0 DO n = 1, pdims(1) DO i = 1, nnx work_fftx(m) = ar(i,k,j,n) m = m + 1 ENDDO ENDDO CALL fft_x_1d( work_fftx, 'forward' ) DO i = 0, nx work_trix(i,k) = work_fftx(i) ENDDO ENDDO ENDIF ! !-- Solve the linear equation system CALL tridia_1dd( ddx2, ddy2, nx, ny, j, work_trix, tri(:,:,:,tn) ) IF ( loop_optimization == 'vector' ) THEN ! !-- Code optimized for vector processors CALL fft_x_m( work_trix, 'backward' ) DO k = 1, nz m = 0 DO n = 1, pdims(1) DO i = 1, nnx ar(i,k,j,n) = work_trix(m,k) m = m + 1 ENDDO ENDDO ENDDO ELSE ! !-- Cache optimized code DO k = 1, nz DO i = 0, nx work_fftx(i) = work_trix(i,k) ENDDO CALL fft_x_1d( work_fftx, 'backward' ) m = 0 DO n = 1, pdims(1) DO i = 1, nnx ar(i,k,j,n) = work_fftx(m) m = m + 1 ENDDO ENDDO ENDDO ENDIF ENDDO DEALLOCATE( tri ) CALL cpu_log( log_point_s(33), 'fft_x_1d + tridia', 'stop' ) END SUBROUTINE fftx_tri_fftx !------------------------------------------------------------------------------! ! Description: ! ------------ !> Fourier-transformation along x with subsequent transposition x --> y for !> a 1d-decomposition along y. !> !> @attention NEC-branch of this routine may significantly profit from !> further optimizations. So far, performance is much worse than !> for routine ffty_tr_yx (more than three times slower). !------------------------------------------------------------------------------! SUBROUTINE fftx_tr_xy( f_in, f_out ) USE control_parameters, & ONLY: loop_optimization USE cpulog, & ONLY: cpu_log, log_point_s USE kinds USE pegrid IMPLICIT NONE INTEGER(iwp) :: i !< INTEGER(iwp) :: j !< INTEGER(iwp) :: k !< REAL(wp), DIMENSION(0:nx,1:nz,nys:nyn) :: work_fftx !< REAL(wp), DIMENSION(1:nz,nys:nyn,0:nx) :: f_in !< REAL(wp), DIMENSION(nny,1:nz,nxl_y:nxr_y,pdims(2)) :: f_out !< REAL(wp), DIMENSION(nys:nyn,1:nz,0:nx) :: work !< ! !-- Carry out the FFT along x, where all data are present due to the !-- 1d-decomposition along y. Resort the data in a way that y becomes !-- the first index. CALL cpu_log( log_point_s(4), 'fft_x_1d', 'start' ) IF ( loop_optimization == 'vector' ) THEN ! !-- Code for vector processors !$OMP PARALLEL PRIVATE ( i, j, k ) !$OMP DO DO i = 0, nx DO j = nys, nyn DO k = 1, nz work_fftx(i,k,j) = f_in(k,j,i) ENDDO ENDDO ENDDO !$OMP DO DO j = nys, nyn CALL fft_x_m( work_fftx(:,:,j), 'forward' ) DO k = 1, nz DO i = 0, nx work(j,k,i) = work_fftx(i,k,j) ENDDO ENDDO ENDDO !$OMP END PARALLEL ELSE ! !-- Cache optimized code (there might be still a potential for better !-- optimization). !$OMP PARALLEL PRIVATE (i,j,k) !$OMP DO DO i = 0, nx DO j = nys, nyn DO k = 1, nz work_fftx(i,k,j) = f_in(k,j,i) ENDDO ENDDO ENDDO !$OMP DO DO j = nys, nyn DO k = 1, nz CALL fft_x_1d( work_fftx(0:nx,k,j), 'forward' ) DO i = 0, nx work(j,k,i) = work_fftx(i,k,j) ENDDO ENDDO ENDDO !$OMP END PARALLEL ENDIF CALL cpu_log( log_point_s(4), 'fft_x_1d', 'pause' ) ! !-- Transpose array #if defined( __parallel ) CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) IF ( collective_wait ) CALL MPI_BARRIER( comm2d, ierr ) CALL MPI_ALLTOALL( work(nys,1,0), sendrecvcount_xy, MPI_REAL, & f_out(1,1,nxl_y,1), sendrecvcount_xy, MPI_REAL, & comm1dy, ierr ) CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) #endif END SUBROUTINE fftx_tr_xy !------------------------------------------------------------------------------! ! Description: ! ------------ !> Transposition y --> x with a subsequent backward Fourier transformation for !> a 1d-decomposition along x. !------------------------------------------------------------------------------! SUBROUTINE tr_yx_fftx( f_in, f_out ) USE control_parameters, & ONLY: loop_optimization USE cpulog, & ONLY: cpu_log, log_point_s USE kinds USE pegrid IMPLICIT NONE INTEGER(iwp) :: i !< INTEGER(iwp) :: j !< INTEGER(iwp) :: k !< REAL(wp), DIMENSION(0:nx,1:nz,nys:nyn) :: work_fftx !< REAL(wp), DIMENSION(nny,1:nz,nxl_y:nxr_y,pdims(2)) :: f_in !< REAL(wp), DIMENSION(1:nz,nys:nyn,0:nx) :: f_out !< REAL(wp), DIMENSION(nys:nyn,1:nz,0:nx) :: work !< ! !-- Transpose array #if defined( __parallel ) CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) IF ( collective_wait ) CALL MPI_BARRIER( comm2d, ierr ) CALL MPI_ALLTOALL( f_in(1,1,nxl_y,1), sendrecvcount_xy, MPI_REAL, & work(nys,1,0), sendrecvcount_xy, MPI_REAL, & comm1dy, ierr ) CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) #endif ! !-- Carry out the FFT along x, where all data are present due to the !-- 1d-decomposition along y. Resort the data in a way that y becomes !-- the first index. CALL cpu_log( log_point_s(4), 'fft_x_1d', 'continue' ) IF ( loop_optimization == 'vector' ) THEN ! !-- Code optimized for vector processors !$OMP PARALLEL PRIVATE ( i, j, k ) !$OMP DO DO j = nys, nyn DO k = 1, nz DO i = 0, nx work_fftx(i,k,j) = work(j,k,i) ENDDO ENDDO CALL fft_x_m( work_fftx(:,:,j), 'backward' ) ENDDO !$OMP DO DO i = 0, nx DO j = nys, nyn DO k = 1, nz f_out(k,j,i) = work_fftx(i,k,j) ENDDO ENDDO ENDDO !$OMP END PARALLEL ELSE ! !-- Cache optimized code (there might be still a potential for better !-- optimization). !$OMP PARALLEL PRIVATE (i,j,k) !$OMP DO DO j = nys, nyn DO k = 1, nz DO i = 0, nx work_fftx(i,k,j) = work(j,k,i) ENDDO CALL fft_x_1d( work_fftx(0:nx,k,j), 'backward' ) ENDDO ENDDO !$OMP DO DO i = 0, nx DO j = nys, nyn DO k = 1, nz f_out(k,j,i) = work_fftx(i,k,j) ENDDO ENDDO ENDDO !$OMP END PARALLEL ENDIF CALL cpu_log( log_point_s(4), 'fft_x_1d', 'stop' ) END SUBROUTINE tr_yx_fftx !------------------------------------------------------------------------------! ! Description: ! ------------ !> FFT along y, solution of the tridiagonal system and backward FFT for !> a 1d-decomposition along y. !> !> @warning this subroutine may still not work for hybrid parallelization !> with OpenMP (for possible necessary changes see the original !> routine poisfft_hybrid, developed by Klaus Ketelsen, May 2002) !------------------------------------------------------------------------------! SUBROUTINE ffty_tri_ffty( ar ) USE control_parameters, & ONLY: loop_optimization USE cpulog, & ONLY: cpu_log, log_point_s USE grid_variables, & ONLY: ddx2, ddy2 USE kinds USE pegrid IMPLICIT NONE INTEGER(iwp) :: i !< INTEGER(iwp) :: j !< INTEGER(iwp) :: k !< INTEGER(iwp) :: m !< INTEGER(iwp) :: n !< !$ INTEGER(iwp) :: omp_get_thread_num !< INTEGER(iwp) :: tn !< REAL(wp), DIMENSION(0:ny) :: work_ffty !< REAL(wp), DIMENSION(0:ny,1:nz) :: work_triy !< REAL(wp), DIMENSION(nny,1:nz,nxl_y:nxr_y,pdims(2)) :: ar !< REAL(wp), DIMENSION(:,:,:,:), ALLOCATABLE :: tri !< CALL cpu_log( log_point_s(39), 'fft_y_1d + tridia', 'start' ) ALLOCATE( tri(5,0:ny,0:nz-1,0:threads_per_task-1) ) tn = 0 ! Default thread number in case of one thread !$OMP PARALLEL DO PRIVATE ( i, j, k, m, n, tn, work_ffty, work_triy ) DO i = nxl_y, nxr_y !$ tn = omp_get_thread_num() IF ( loop_optimization == 'vector' ) THEN ! !-- Code optimized for vector processors DO k = 1, nz m = 0 DO n = 1, pdims(2) DO j = 1, nny work_triy(m,k) = ar(j,k,i,n) m = m + 1 ENDDO ENDDO ENDDO CALL fft_y_m( work_triy, ny, 'forward' ) ELSE ! !-- Cache optimized code DO k = 1, nz m = 0 DO n = 1, pdims(2) DO j = 1, nny work_ffty(m) = ar(j,k,i,n) m = m + 1 ENDDO ENDDO CALL fft_y_1d( work_ffty, 'forward' ) DO j = 0, ny work_triy(j,k) = work_ffty(j) ENDDO ENDDO ENDIF ! !-- Solve the linear equation system CALL tridia_1dd( ddy2, ddx2, ny, nx, i, work_triy, tri(:,:,:,tn) ) IF ( loop_optimization == 'vector' ) THEN ! !-- Code optimized for vector processors CALL fft_y_m( work_triy, ny, 'backward' ) DO k = 1, nz m = 0 DO n = 1, pdims(2) DO j = 1, nny ar(j,k,i,n) = work_triy(m,k) m = m + 1 ENDDO ENDDO ENDDO ELSE ! !-- Cache optimized code DO k = 1, nz DO j = 0, ny work_ffty(j) = work_triy(j,k) ENDDO CALL fft_y_1d( work_ffty, 'backward' ) m = 0 DO n = 1, pdims(2) DO j = 1, nny ar(j,k,i,n) = work_ffty(m) m = m + 1 ENDDO ENDDO ENDDO ENDIF ENDDO DEALLOCATE( tri ) CALL cpu_log( log_point_s(39), 'fft_y_1d + tridia', 'stop' ) END SUBROUTINE ffty_tri_ffty END MODULE poisfft_mod