| 1 | SUBROUTINE exchange_horiz( ar, nbgp_local) |
|---|
| 2 | |
|---|
| 3 | !--------------------------------------------------------------------------------! |
|---|
| 4 | ! This file is part of PALM. |
|---|
| 5 | ! |
|---|
| 6 | ! PALM is free software: you can redistribute it and/or modify it under the terms |
|---|
| 7 | ! of the GNU General Public License as published by the Free Software Foundation, |
|---|
| 8 | ! either version 3 of the License, or (at your option) any later version. |
|---|
| 9 | ! |
|---|
| 10 | ! PALM is distributed in the hope that it will be useful, but WITHOUT ANY |
|---|
| 11 | ! WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR |
|---|
| 12 | ! A PARTICULAR PURPOSE. See the GNU General Public License for more details. |
|---|
| 13 | ! |
|---|
| 14 | ! You should have received a copy of the GNU General Public License along with |
|---|
| 15 | ! PALM. If not, see <http://www.gnu.org/licenses/>. |
|---|
| 16 | ! |
|---|
| 17 | ! Copyright 1997-2012 Leibniz University Hannover |
|---|
| 18 | !--------------------------------------------------------------------------------! |
|---|
| 19 | ! |
|---|
| 20 | ! Current revisions: |
|---|
| 21 | ! ----------------- |
|---|
| 22 | ! |
|---|
| 23 | ! |
|---|
| 24 | ! Former revisions: |
|---|
| 25 | ! ----------------- |
|---|
| 26 | ! $Id: exchange_horiz.f90 1258 2013-11-08 16:09:09Z fricke $ |
|---|
| 27 | ! |
|---|
| 28 | ! 1257 2013-11-08 15:18:40Z raasch |
|---|
| 29 | ! openacc loop and loop vector clauses removed |
|---|
| 30 | ! |
|---|
| 31 | ! 1128 2013-04-12 06:19:32Z raasch |
|---|
| 32 | ! modifications for asynchronous transfer, |
|---|
| 33 | ! local variables req, wait_stat are global now, and have been moved to module |
|---|
| 34 | ! pegrid |
|---|
| 35 | ! |
|---|
| 36 | ! 1113 2013-03-10 02:48:14Z raasch |
|---|
| 37 | ! GPU-porting for single-core (1PE) mode |
|---|
| 38 | ! |
|---|
| 39 | ! 1036 2012-10-22 13:43:42Z raasch |
|---|
| 40 | ! code put under GPL (PALM 3.9) |
|---|
| 41 | ! |
|---|
| 42 | ! 841 2012-02-28 12:29:49Z maronga |
|---|
| 43 | ! Excluded routine from compilation of namelist_file_check |
|---|
| 44 | ! |
|---|
| 45 | ! 709 2011-03-30 09:31:40Z raasch |
|---|
| 46 | ! formatting adjustments |
|---|
| 47 | ! |
|---|
| 48 | ! 707 2011-03-29 11:39:40Z raasch |
|---|
| 49 | ! grid_level directly used as index for MPI data type arrays, |
|---|
| 50 | ! bc_lr/ns replaced by bc_lr/ns_cyc |
|---|
| 51 | ! |
|---|
| 52 | ! 689 2011-02-20 19:31:12z gryschka |
|---|
| 53 | ! Bugfix for some logical expressions |
|---|
| 54 | ! (syntax was not compatible with all compilers) |
|---|
| 55 | ! |
|---|
| 56 | ! 683 2011-02-09 14:25:15Z raasch |
|---|
| 57 | ! optional synchronous exchange (sendrecv) implemented, code partly reformatted |
|---|
| 58 | ! |
|---|
| 59 | ! 667 2010-12-23 12:06:00Z suehring/gryschka |
|---|
| 60 | ! Dynamic exchange of ghost points with nbgp_local to ensure that no useless |
|---|
| 61 | ! ghost points exchanged in case of multigrid. type_yz(0) and type_xz(0) |
|---|
| 62 | ! used for normal grid, the remaining types used for the several grid levels. |
|---|
| 63 | ! Exchange is done via MPI-Vectors with a dynamic value of ghost points which |
|---|
| 64 | ! depend on the advection scheme. Exchange of left and right PEs is 10% faster |
|---|
| 65 | ! with MPI-Vectors than without. |
|---|
| 66 | ! |
|---|
| 67 | ! 75 2007-03-22 09:54:05Z raasch |
|---|
| 68 | ! Special cases for additional gridpoints along x or y in case of non-cyclic |
|---|
| 69 | ! boundary conditions are not regarded any more |
|---|
| 70 | ! |
|---|
| 71 | ! RCS Log replace by Id keyword, revision history cleaned up |
|---|
| 72 | ! |
|---|
| 73 | ! Revision 1.16 2006/02/23 12:19:08 raasch |
|---|
| 74 | ! anz_yz renamed ngp_yz |
|---|
| 75 | ! |
|---|
| 76 | ! Revision 1.1 1997/07/24 11:13:29 raasch |
|---|
| 77 | ! Initial revision |
|---|
| 78 | ! |
|---|
| 79 | ! |
|---|
| 80 | ! Description: |
|---|
| 81 | ! ------------ |
|---|
| 82 | ! Exchange of lateral boundary values (parallel computers) and cyclic |
|---|
| 83 | ! lateral boundary conditions, respectively. |
|---|
| 84 | !------------------------------------------------------------------------------! |
|---|
| 85 | |
|---|
| 86 | USE control_parameters |
|---|
| 87 | USE cpulog |
|---|
| 88 | USE indices |
|---|
| 89 | USE interfaces |
|---|
| 90 | USE pegrid |
|---|
| 91 | |
|---|
| 92 | IMPLICIT NONE |
|---|
| 93 | |
|---|
| 94 | |
|---|
| 95 | INTEGER :: i, j, k, nbgp_local |
|---|
| 96 | REAL, DIMENSION(nzb:nzt+1,nys-nbgp_local:nyn+nbgp_local, & |
|---|
| 97 | nxl-nbgp_local:nxr+nbgp_local) :: ar |
|---|
| 98 | |
|---|
| 99 | #if ! defined( __check ) |
|---|
| 100 | |
|---|
| 101 | CALL cpu_log( log_point_s(2), 'exchange_horiz', 'start' ) |
|---|
| 102 | |
|---|
| 103 | #if defined( __parallel ) |
|---|
| 104 | |
|---|
| 105 | ! |
|---|
| 106 | !-- Exchange in x-direction of lateral boundaries |
|---|
| 107 | IF ( pdims(1) == 1 .OR. mg_switch_to_pe0 ) THEN |
|---|
| 108 | ! |
|---|
| 109 | !-- One-dimensional decomposition along y, boundary values can be exchanged |
|---|
| 110 | !-- within the PE memory |
|---|
| 111 | IF ( bc_lr_cyc ) THEN |
|---|
| 112 | ar(:,:,nxl-nbgp_local:nxl-1) = ar(:,:,nxr-nbgp_local+1:nxr) |
|---|
| 113 | ar(:,:,nxr+1:nxr+nbgp_local) = ar(:,:,nxl:nxl+nbgp_local-1) |
|---|
| 114 | ENDIF |
|---|
| 115 | |
|---|
| 116 | ELSE |
|---|
| 117 | |
|---|
| 118 | IF ( synchronous_exchange ) THEN |
|---|
| 119 | ! |
|---|
| 120 | !-- Send left boundary, receive right one (synchronous) |
|---|
| 121 | CALL MPI_SENDRECV( & |
|---|
| 122 | ar(nzb,nys-nbgp_local,nxl), 1, type_yz(grid_level), pleft, 0, & |
|---|
| 123 | ar(nzb,nys-nbgp_local,nxr+1), 1, type_yz(grid_level), pright, 0, & |
|---|
| 124 | comm2d, status, ierr ) |
|---|
| 125 | ! |
|---|
| 126 | !-- Send right boundary, receive left one (synchronous) |
|---|
| 127 | CALL MPI_SENDRECV( ar(nzb,nys-nbgp_local,nxr+1-nbgp_local), 1, & |
|---|
| 128 | type_yz(grid_level), pright, 1, & |
|---|
| 129 | ar(nzb,nys-nbgp_local,nxl-nbgp_local), 1, & |
|---|
| 130 | type_yz(grid_level), pleft, 1, & |
|---|
| 131 | comm2d, status, ierr ) |
|---|
| 132 | |
|---|
| 133 | ELSE |
|---|
| 134 | |
|---|
| 135 | ! |
|---|
| 136 | !-- In case of background communication switched on, exchange is done |
|---|
| 137 | !-- either along x or along y |
|---|
| 138 | IF ( send_receive == 'lr' .OR. send_receive == 'al' ) THEN |
|---|
| 139 | |
|---|
| 140 | IF ( .NOT. sendrecv_in_background ) THEN |
|---|
| 141 | req(1:4) = 0 |
|---|
| 142 | req_count = 0 |
|---|
| 143 | ENDIF |
|---|
| 144 | ! |
|---|
| 145 | !-- Send left boundary, receive right one (asynchronous) |
|---|
| 146 | CALL MPI_ISEND( ar(nzb,nys-nbgp_local,nxl), 1, type_yz(grid_level), & |
|---|
| 147 | pleft, req_count, comm2d, req(req_count+1), ierr ) |
|---|
| 148 | CALL MPI_IRECV( ar(nzb,nys-nbgp_local,nxr+1), 1, type_yz(grid_level), & |
|---|
| 149 | pright, req_count, comm2d, req(req_count+2), ierr ) |
|---|
| 150 | ! |
|---|
| 151 | !-- Send right boundary, receive left one (asynchronous) |
|---|
| 152 | CALL MPI_ISEND( ar(nzb,nys-nbgp_local,nxr+1-nbgp_local), 1, & |
|---|
| 153 | type_yz(grid_level), pright, req_count+1, comm2d, & |
|---|
| 154 | req(req_count+3), ierr ) |
|---|
| 155 | CALL MPI_IRECV( ar(nzb,nys-nbgp_local,nxl-nbgp_local), 1, & |
|---|
| 156 | type_yz(grid_level), pleft, req_count+1, comm2d, & |
|---|
| 157 | req(req_count+4), ierr ) |
|---|
| 158 | |
|---|
| 159 | IF ( .NOT. sendrecv_in_background ) THEN |
|---|
| 160 | CALL MPI_WAITALL( 4, req, wait_stat, ierr ) |
|---|
| 161 | ELSE |
|---|
| 162 | req_count = req_count + 4 |
|---|
| 163 | ENDIF |
|---|
| 164 | |
|---|
| 165 | ENDIF |
|---|
| 166 | |
|---|
| 167 | ENDIF |
|---|
| 168 | |
|---|
| 169 | ENDIF |
|---|
| 170 | |
|---|
| 171 | |
|---|
| 172 | IF ( pdims(2) == 1 .OR. mg_switch_to_pe0 ) THEN |
|---|
| 173 | ! |
|---|
| 174 | !-- One-dimensional decomposition along x, boundary values can be exchanged |
|---|
| 175 | !-- within the PE memory |
|---|
| 176 | IF ( bc_ns_cyc ) THEN |
|---|
| 177 | ar(:,nys-nbgp_local:nys-1,:) = ar(:,nyn-nbgp_local+1:nyn,:) |
|---|
| 178 | ar(:,nyn+1:nyn+nbgp_local,:) = ar(:,nys:nys+nbgp_local-1,:) |
|---|
| 179 | ENDIF |
|---|
| 180 | |
|---|
| 181 | ELSE |
|---|
| 182 | |
|---|
| 183 | IF ( synchronous_exchange ) THEN |
|---|
| 184 | ! |
|---|
| 185 | !-- Send front boundary, receive rear one (synchronous) |
|---|
| 186 | CALL MPI_SENDRECV( & |
|---|
| 187 | ar(nzb,nys,nxl-nbgp_local), 1, type_xz(grid_level), psouth, 0, & |
|---|
| 188 | ar(nzb,nyn+1,nxl-nbgp_local), 1, type_xz(grid_level), pnorth, 0, & |
|---|
| 189 | comm2d, status, ierr ) |
|---|
| 190 | ! |
|---|
| 191 | !-- Send rear boundary, receive front one (synchronous) |
|---|
| 192 | CALL MPI_SENDRECV( ar(nzb,nyn-nbgp_local+1,nxl-nbgp_local), 1, & |
|---|
| 193 | type_xz(grid_level), pnorth, 1, & |
|---|
| 194 | ar(nzb,nys-nbgp_local,nxl-nbgp_local), 1, & |
|---|
| 195 | type_xz(grid_level), psouth, 1, & |
|---|
| 196 | comm2d, status, ierr ) |
|---|
| 197 | |
|---|
| 198 | ELSE |
|---|
| 199 | |
|---|
| 200 | ! |
|---|
| 201 | !-- In case of background communication switched on, exchange is done |
|---|
| 202 | !-- either along x or along y |
|---|
| 203 | IF ( send_receive == 'lr' .OR. send_receive == 'al' ) THEN |
|---|
| 204 | |
|---|
| 205 | IF ( .NOT. sendrecv_in_background ) THEN |
|---|
| 206 | req(1:4) = 0 |
|---|
| 207 | req_count = 0 |
|---|
| 208 | ENDIF |
|---|
| 209 | |
|---|
| 210 | ! |
|---|
| 211 | !-- Send front boundary, receive rear one (asynchronous) |
|---|
| 212 | CALL MPI_ISEND( ar(nzb,nys,nxl-nbgp_local), 1, type_xz(grid_level), & |
|---|
| 213 | psouth, req_count, comm2d, req(req_count+1), ierr ) |
|---|
| 214 | CALL MPI_IRECV( ar(nzb,nyn+1,nxl-nbgp_local), 1, type_xz(grid_level), & |
|---|
| 215 | pnorth, req_count, comm2d, req(req_count+2), ierr ) |
|---|
| 216 | ! |
|---|
| 217 | !-- Send rear boundary, receive front one (asynchronous) |
|---|
| 218 | CALL MPI_ISEND( ar(nzb,nyn-nbgp_local+1,nxl-nbgp_local), 1, & |
|---|
| 219 | type_xz(grid_level), pnorth, req_count+1, comm2d, & |
|---|
| 220 | req(req_count+3), ierr ) |
|---|
| 221 | CALL MPI_IRECV( ar(nzb,nys-nbgp_local,nxl-nbgp_local), 1, & |
|---|
| 222 | type_xz(grid_level), psouth, req_count+1, comm2d, & |
|---|
| 223 | req(req_count+4), ierr ) |
|---|
| 224 | |
|---|
| 225 | IF ( .NOT. sendrecv_in_background ) THEN |
|---|
| 226 | CALL MPI_WAITALL( 4, req, wait_stat, ierr ) |
|---|
| 227 | ELSE |
|---|
| 228 | req_count = req_count + 4 |
|---|
| 229 | ENDIF |
|---|
| 230 | |
|---|
| 231 | ENDIF |
|---|
| 232 | |
|---|
| 233 | ENDIF |
|---|
| 234 | |
|---|
| 235 | ENDIF |
|---|
| 236 | |
|---|
| 237 | #else |
|---|
| 238 | |
|---|
| 239 | ! |
|---|
| 240 | !-- Lateral boundary conditions in the non-parallel case. |
|---|
| 241 | !-- Case dependent, because in GPU mode still not all arrays are on device. This |
|---|
| 242 | !-- workaround has to be removed later. Also, since PGI compiler 12.5 has problems |
|---|
| 243 | !-- with array syntax, explicit loops are used. |
|---|
| 244 | IF ( bc_lr == 'cyclic' ) THEN |
|---|
| 245 | IF ( on_device ) THEN |
|---|
| 246 | !$acc kernels present( ar ) |
|---|
| 247 | !$acc loop independent |
|---|
| 248 | DO i = 0, nbgp_local-1 |
|---|
| 249 | DO j = nys-nbgp_local, nyn+nbgp_local |
|---|
| 250 | DO k = nzb, nzt+1 |
|---|
| 251 | ar(k,j,nxl-nbgp_local+i) = ar(k,j,nxr-nbgp_local+1+i) |
|---|
| 252 | ar(k,j,nxr+1+i) = ar(k,j,nxl+i) |
|---|
| 253 | ENDDO |
|---|
| 254 | ENDDO |
|---|
| 255 | ENDDO |
|---|
| 256 | !$acc end kernels |
|---|
| 257 | ELSE |
|---|
| 258 | ar(:,:,nxl-nbgp_local:nxl-1) = ar(:,:,nxr-nbgp_local+1:nxr) |
|---|
| 259 | ar(:,:,nxr+1:nxr+nbgp_local) = ar(:,:,nxl:nxl+nbgp_local-1) |
|---|
| 260 | ENDIF |
|---|
| 261 | ENDIF |
|---|
| 262 | |
|---|
| 263 | IF ( bc_ns == 'cyclic' ) THEN |
|---|
| 264 | IF ( on_device ) THEN |
|---|
| 265 | !$acc kernels present( ar ) |
|---|
| 266 | DO i = nxl-nbgp_local, nxr+nbgp_local |
|---|
| 267 | !$acc loop independent |
|---|
| 268 | DO j = 0, nbgp_local-1 |
|---|
| 269 | !$acc loop independent |
|---|
| 270 | DO k = nzb, nzt+1 |
|---|
| 271 | ar(k,nys-nbgp_local+j,i) = ar(k,nyn-nbgp_local+1+j,i) |
|---|
| 272 | ar(k,nyn+1+j,i) = ar(k,nys+j,i) |
|---|
| 273 | ENDDO |
|---|
| 274 | ENDDO |
|---|
| 275 | ENDDO |
|---|
| 276 | !$acc end kernels |
|---|
| 277 | ELSE |
|---|
| 278 | ar(:,nys-nbgp_local:nys-1,:) = ar(:,nyn-nbgp_local+1:nyn,:) |
|---|
| 279 | ar(:,nyn+1:nyn+nbgp_local,:) = ar(:,nys:nys+nbgp_local-1,:) |
|---|
| 280 | ENDIF |
|---|
| 281 | ENDIF |
|---|
| 282 | |
|---|
| 283 | #endif |
|---|
| 284 | CALL cpu_log( log_point_s(2), 'exchange_horiz', 'stop' ) |
|---|
| 285 | |
|---|
| 286 | #endif |
|---|
| 287 | END SUBROUTINE exchange_horiz |
|---|
