[1] | 1 | SUBROUTINE transpose_xy( f_in, work1, f_inv, work2, f_out ) |
---|
| 2 | |
---|
| 3 | !------------------------------------------------------------------------------! |
---|
| 4 | ! Actual revisions: |
---|
| 5 | ! ----------------- |
---|
| 6 | ! |
---|
| 7 | ! |
---|
| 8 | ! Former revisions: |
---|
| 9 | ! ----------------- |
---|
| 10 | ! $Log: transpose.f90,v $ |
---|
| 11 | ! Revision 1.2 2004/04/30 13:12:17 raasch |
---|
| 12 | ! Switched from mpi_alltoallv to the simpler mpi_alltoall, |
---|
| 13 | ! all former transpose-routine files collected in this file, enlarged |
---|
| 14 | ! transposition arrays introduced |
---|
| 15 | ! |
---|
| 16 | ! Revision 1.1 2004/04/30 13:08:16 raasch |
---|
| 17 | ! Initial revision (collection of former routines transpose_xy, transpose_xz, |
---|
| 18 | ! transpose_yx, transpose_yz, transpose_zx, transpose_zy) |
---|
| 19 | ! |
---|
| 20 | ! Revision 1.7 2003/03/16 09:49:53 raasch |
---|
| 21 | ! Two underscores (_) are placed in front of all define-strings |
---|
| 22 | ! |
---|
| 23 | ! Revision 1.6 2001/03/30 07:55:50 raasch |
---|
| 24 | ! Array bound for work2 is explicitely calculated, all comments and identifiers |
---|
| 25 | ! translated into English |
---|
| 26 | ! |
---|
| 27 | ! Revision 1.5 2001/01/22 08:19:14 raasch |
---|
| 28 | ! Module test_variables removed |
---|
| 29 | ! |
---|
| 30 | ! Revision 1.4 1998/07/06 12:34:22 raasch |
---|
| 31 | ! + USE test_variables |
---|
| 32 | ! |
---|
| 33 | ! Revision 1.3 1997/08/13 13:52:42 raasch |
---|
| 34 | ! Zeitmessung von MPI_ALLTOALLV |
---|
| 35 | ! |
---|
| 36 | ! Revision 1.2 1997/08/11 06:27:31 raasch |
---|
| 37 | ! Keine Aenderungen |
---|
| 38 | ! |
---|
| 39 | ! Revision 1.1 1997/07/24 11:25:18 raasch |
---|
| 40 | ! Initial revision |
---|
| 41 | ! |
---|
| 42 | ! |
---|
| 43 | ! Description: |
---|
| 44 | ! ------------ |
---|
| 45 | ! Transposition of input array (f_in) from x to y. For the input array, all |
---|
| 46 | ! elements along x reside on the same PE, while after transposition, all |
---|
| 47 | ! elements along y reside on the same PE. |
---|
| 48 | !------------------------------------------------------------------------------! |
---|
| 49 | |
---|
| 50 | USE cpulog |
---|
| 51 | USE indices |
---|
| 52 | USE interfaces |
---|
| 53 | USE pegrid |
---|
| 54 | USE transpose_indices |
---|
| 55 | |
---|
| 56 | IMPLICIT NONE |
---|
| 57 | |
---|
| 58 | INTEGER :: i, j, k, l, m, ys |
---|
| 59 | |
---|
| 60 | REAL :: f_in(0:nxa,nys_x:nyn_xa,nzb_x:nzt_xa), & |
---|
| 61 | f_inv(nys_x:nyn_xa,nzb_x:nzt_xa,0:nxa), & |
---|
| 62 | f_out(0:nya,nxl_y:nxr_ya,nzb_y:nzt_ya), & |
---|
| 63 | work1(nys_x:nyn_xa,nzb_x:nzt_xa,0:nxa), work2(nnx*nny*nnz) |
---|
| 64 | |
---|
| 65 | #if defined( __parallel ) |
---|
| 66 | |
---|
| 67 | ! |
---|
| 68 | !-- Rearrange indices of input array in order to make data to be send |
---|
| 69 | !-- by MPI contiguous |
---|
| 70 | DO k = nzb_x, nzt_xa |
---|
| 71 | DO j = nys_x, nyn_xa |
---|
| 72 | DO i = 0, nxa |
---|
| 73 | work1(j,k,i) = f_in(i,j,k) |
---|
| 74 | ENDDO |
---|
| 75 | ENDDO |
---|
| 76 | ENDDO |
---|
| 77 | |
---|
| 78 | ! |
---|
| 79 | !-- Move data to different array, because memory location of work1 is |
---|
| 80 | !-- needed further below (work1 = work2) |
---|
| 81 | DO i = 0, nxa |
---|
| 82 | DO k = nzb_x, nzt_xa |
---|
| 83 | DO j = nys_x, nyn_xa |
---|
| 84 | f_inv(j,k,i) = work1(j,k,i) |
---|
| 85 | ENDDO |
---|
| 86 | ENDDO |
---|
| 87 | ENDDO |
---|
| 88 | |
---|
| 89 | ! |
---|
| 90 | !-- Transpose array |
---|
| 91 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) |
---|
| 92 | CALL MPI_ALLTOALL( f_inv(nys_x,nzb_x,0), sendrecvcount_xy, MPI_REAL, & |
---|
| 93 | work2(1), sendrecvcount_xy, MPI_REAL, & |
---|
| 94 | comm1dy, ierr ) |
---|
| 95 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) |
---|
| 96 | |
---|
| 97 | ! |
---|
| 98 | !-- Reorder transposed array |
---|
| 99 | m = 0 |
---|
| 100 | DO l = 0, pdims(2) - 1 |
---|
| 101 | ys = 0 + l * ( nyn_xa - nys_x + 1 ) |
---|
| 102 | DO i = nxl_y, nxr_ya |
---|
| 103 | DO k = nzb_y, nzt_ya |
---|
| 104 | DO j = ys, ys + nyn_xa - nys_x |
---|
| 105 | m = m + 1 |
---|
| 106 | f_out(j,i,k) = work2(m) |
---|
| 107 | ENDDO |
---|
| 108 | ENDDO |
---|
| 109 | ENDDO |
---|
| 110 | ENDDO |
---|
| 111 | |
---|
| 112 | #endif |
---|
| 113 | |
---|
| 114 | END SUBROUTINE transpose_xy |
---|
| 115 | |
---|
| 116 | |
---|
| 117 | SUBROUTINE transpose_xz( f_in, work1, f_inv, work2, f_out ) |
---|
| 118 | |
---|
| 119 | !------------------------------------------------------------------------------! |
---|
| 120 | ! Description: |
---|
| 121 | ! ------------ |
---|
| 122 | ! Transposition of input array (f_in) from x to z. For the input array, all |
---|
| 123 | ! elements along x reside on the same PE, while after transposition, all |
---|
| 124 | ! elements along z reside on the same PE. |
---|
| 125 | !------------------------------------------------------------------------------! |
---|
| 126 | |
---|
| 127 | USE cpulog |
---|
| 128 | USE indices |
---|
| 129 | USE interfaces |
---|
| 130 | USE pegrid |
---|
| 131 | USE transpose_indices |
---|
| 132 | |
---|
| 133 | IMPLICIT NONE |
---|
| 134 | |
---|
| 135 | INTEGER :: i, j, k, l, m, xs |
---|
| 136 | |
---|
| 137 | REAL :: f_in(0:nxa,nys_x:nyn_xa,nzb_x:nzt_xa), & |
---|
| 138 | f_inv(nxl:nxra,nys:nyna,1:nza), & |
---|
| 139 | f_out(1:nza,nys:nyna,nxl:nxra), & |
---|
| 140 | work1(1:nza,nys:nyna,nxl:nxra), work2(nnx*nny*nnz) |
---|
| 141 | |
---|
| 142 | #if defined( __parallel ) |
---|
| 143 | |
---|
| 144 | ! |
---|
| 145 | !-- If the PE grid is one-dimensional along y, the array has only to be |
---|
| 146 | !-- reordered locally and therefore no transposition has to be done. |
---|
| 147 | IF ( pdims(1) /= 1 ) THEN |
---|
| 148 | ! |
---|
| 149 | !-- Reorder input array for transposition |
---|
| 150 | m = 0 |
---|
| 151 | DO l = 0, pdims(1) - 1 |
---|
| 152 | xs = 0 + l * nnx |
---|
| 153 | DO k = nzb_x, nzt_xa |
---|
| 154 | DO j = nys_x, nyn_xa |
---|
| 155 | DO i = xs, xs + nnx - 1 |
---|
| 156 | m = m + 1 |
---|
| 157 | work2(m) = f_in(i,j,k) |
---|
| 158 | ENDDO |
---|
| 159 | ENDDO |
---|
| 160 | ENDDO |
---|
| 161 | ENDDO |
---|
| 162 | |
---|
| 163 | ! |
---|
| 164 | !-- Transpose array |
---|
| 165 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) |
---|
| 166 | CALL MPI_ALLTOALL( work2(1), sendrecvcount_zx, MPI_REAL, & |
---|
| 167 | f_inv(nxl,nys,1), sendrecvcount_zx, MPI_REAL, & |
---|
| 168 | comm1dx, ierr ) |
---|
| 169 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) |
---|
| 170 | |
---|
| 171 | ! |
---|
| 172 | !-- Reorder transposed array in a way that the z index is in first position |
---|
| 173 | DO i = nxl, nxra |
---|
| 174 | DO j = nys, nyna |
---|
| 175 | DO k = 1, nza |
---|
| 176 | work1(k,j,i) = f_inv(i,j,k) |
---|
| 177 | ENDDO |
---|
| 178 | ENDDO |
---|
| 179 | ENDDO |
---|
| 180 | ELSE |
---|
| 181 | ! |
---|
| 182 | !-- Reorder the array in a way that the z index is in first position |
---|
| 183 | DO i = nxl, nxra |
---|
| 184 | DO j = nys, nyna |
---|
| 185 | DO k = 1, nza |
---|
| 186 | work1(k,j,i) = f_in(i,j,k) |
---|
| 187 | ENDDO |
---|
| 188 | ENDDO |
---|
| 189 | ENDDO |
---|
| 190 | ENDIF |
---|
| 191 | |
---|
| 192 | ! |
---|
| 193 | !-- Move data to output array |
---|
| 194 | DO i = nxl, nxra |
---|
| 195 | DO j = nys, nyna |
---|
| 196 | DO k = 1, nza |
---|
| 197 | f_out(k,j,i) = work1(k,j,i) |
---|
| 198 | ENDDO |
---|
| 199 | ENDDO |
---|
| 200 | ENDDO |
---|
| 201 | |
---|
| 202 | #endif |
---|
| 203 | |
---|
| 204 | END SUBROUTINE transpose_xz |
---|
| 205 | |
---|
| 206 | |
---|
| 207 | SUBROUTINE transpose_yx( f_in, work1, f_inv, work2, f_out ) |
---|
| 208 | |
---|
| 209 | !------------------------------------------------------------------------------! |
---|
| 210 | ! Description: |
---|
| 211 | ! ------------ |
---|
| 212 | ! Transposition of input array (f_in) from y to x. For the input array, all |
---|
| 213 | ! elements along y reside on the same PE, while after transposition, all |
---|
| 214 | ! elements along x reside on the same PE. |
---|
| 215 | !------------------------------------------------------------------------------! |
---|
| 216 | |
---|
| 217 | USE cpulog |
---|
| 218 | USE indices |
---|
| 219 | USE interfaces |
---|
| 220 | USE pegrid |
---|
| 221 | USE transpose_indices |
---|
| 222 | |
---|
| 223 | IMPLICIT NONE |
---|
| 224 | |
---|
| 225 | INTEGER :: i, j, k, l, m, ys |
---|
| 226 | |
---|
| 227 | REAL :: f_in(0:nya,nxl_y:nxr_ya,nzb_y:nzt_ya), & |
---|
| 228 | f_inv(nys_x:nyn_xa,nzb_x:nzt_xa,0:nxa), & |
---|
| 229 | f_out(0:nxa,nys_x:nyn_xa,nzb_x:nzt_xa), & |
---|
| 230 | work1(0:nxa,nys_x:nyn_xa,nzb_x:nzt_xa), work2(nnx*nny*nnz) |
---|
| 231 | |
---|
| 232 | #if defined( __parallel ) |
---|
| 233 | |
---|
| 234 | ! |
---|
| 235 | !-- Reorder input array for transposition |
---|
| 236 | m = 0 |
---|
| 237 | DO l = 0, pdims(2) - 1 |
---|
| 238 | ys = 0 + l * ( nyn_xa - nys_x + 1 ) |
---|
| 239 | DO i = nxl_y, nxr_ya |
---|
| 240 | DO k = nzb_y, nzt_ya |
---|
| 241 | DO j = ys, ys + nyn_xa - nys_x |
---|
| 242 | m = m + 1 |
---|
| 243 | work2(m) = f_in(j,i,k) |
---|
| 244 | ENDDO |
---|
| 245 | ENDDO |
---|
| 246 | ENDDO |
---|
| 247 | ENDDO |
---|
| 248 | |
---|
| 249 | ! |
---|
| 250 | !-- Transpose array |
---|
| 251 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) |
---|
| 252 | CALL MPI_ALLTOALL( work2(1), sendrecvcount_xy, MPI_REAL, & |
---|
| 253 | f_inv(nys_x,nzb_x,0), sendrecvcount_xy, MPI_REAL, & |
---|
| 254 | comm1dy, ierr ) |
---|
| 255 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) |
---|
| 256 | |
---|
| 257 | ! |
---|
| 258 | !-- Reorder transposed array in a way that the x index is in first position |
---|
| 259 | DO i = 0, nxa |
---|
| 260 | DO k = nzb_x, nzt_xa |
---|
| 261 | DO j = nys_x, nyn_xa |
---|
| 262 | work1(i,j,k) = f_inv(j,k,i) |
---|
| 263 | ENDDO |
---|
| 264 | ENDDO |
---|
| 265 | ENDDO |
---|
| 266 | |
---|
| 267 | ! |
---|
| 268 | !-- Move data to output array |
---|
| 269 | DO k = nzb_x, nzt_xa |
---|
| 270 | DO j = nys_x, nyn_xa |
---|
| 271 | DO i = 0, nxa |
---|
| 272 | f_out(i,j,k) = work1(i,j,k) |
---|
| 273 | ENDDO |
---|
| 274 | ENDDO |
---|
| 275 | ENDDO |
---|
| 276 | |
---|
| 277 | #endif |
---|
| 278 | |
---|
| 279 | END SUBROUTINE transpose_yx |
---|
| 280 | |
---|
| 281 | |
---|
| 282 | SUBROUTINE transpose_yxd( f_in, work1, f_inv, work2, f_out ) |
---|
| 283 | |
---|
| 284 | !------------------------------------------------------------------------------! |
---|
| 285 | ! Description: |
---|
| 286 | ! ------------ |
---|
| 287 | ! Transposition of input array (f_in) from y to x. For the input array, all |
---|
| 288 | ! elements along y reside on the same PE, while after transposition, all |
---|
| 289 | ! elements along x reside on the same PE. |
---|
| 290 | ! This is a direct transposition for arrays with indices in regular order |
---|
| 291 | ! (k,j,i) (cf. transpose_yx). |
---|
| 292 | !------------------------------------------------------------------------------! |
---|
| 293 | |
---|
| 294 | USE cpulog |
---|
| 295 | USE indices |
---|
| 296 | USE interfaces |
---|
| 297 | USE pegrid |
---|
| 298 | USE transpose_indices |
---|
| 299 | |
---|
| 300 | IMPLICIT NONE |
---|
| 301 | |
---|
| 302 | INTEGER :: i, j, k, l, m, recvcount_yx, sendcount_yx, xs |
---|
| 303 | |
---|
| 304 | REAL :: f_in(1:nza,nys:nyna,nxl:nxra), f_inv(nxl:nxra,1:nza,nys:nyna), & |
---|
| 305 | f_out(0:nxa,nys_x:nyn_xa,nzb_x:nzt_xa), & |
---|
| 306 | work1(nxl:nxra,1:nza,nys:nyna), work2(nnx*nny*nnz) |
---|
| 307 | |
---|
| 308 | #if defined( __parallel ) |
---|
| 309 | |
---|
| 310 | ! |
---|
| 311 | !-- Rearrange indices of input array in order to make data to be send |
---|
| 312 | !-- by MPI contiguous |
---|
| 313 | DO k = 1, nza |
---|
| 314 | DO j = nys, nyna |
---|
| 315 | DO i = nxl, nxra |
---|
| 316 | work1(i,k,j) = f_in(k,j,i) |
---|
| 317 | ENDDO |
---|
| 318 | ENDDO |
---|
| 319 | ENDDO |
---|
| 320 | |
---|
| 321 | ! |
---|
| 322 | !-- Move data to different array, because memory location of work1 is |
---|
| 323 | !-- needed further below (work1 = work2) |
---|
| 324 | DO j = nys, nyna |
---|
| 325 | DO k = 1, nza |
---|
| 326 | DO i = nxl, nxra |
---|
| 327 | f_inv(i,k,j) = work1(i,k,j) |
---|
| 328 | ENDDO |
---|
| 329 | ENDDO |
---|
| 330 | ENDDO |
---|
| 331 | |
---|
| 332 | ! |
---|
| 333 | !-- Transpose array |
---|
| 334 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) |
---|
| 335 | CALL MPI_ALLTOALL( f_inv(nxl,1,nys), sendrecvcount_xy, MPI_REAL, & |
---|
| 336 | work2(1), sendrecvcount_xy, MPI_REAL, & |
---|
| 337 | comm1dx, ierr ) |
---|
| 338 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) |
---|
| 339 | |
---|
| 340 | ! |
---|
| 341 | !-- Reorder transposed array |
---|
| 342 | m = 0 |
---|
| 343 | DO l = 0, pdims(1) - 1 |
---|
| 344 | xs = 0 + l * nnx |
---|
| 345 | DO j = nys_x, nyn_xa |
---|
| 346 | DO k = 1, nza |
---|
| 347 | DO i = xs, xs + nnx - 1 |
---|
| 348 | m = m + 1 |
---|
| 349 | f_out(i,j,k) = work2(m) |
---|
| 350 | ENDDO |
---|
| 351 | ENDDO |
---|
| 352 | ENDDO |
---|
| 353 | ENDDO |
---|
| 354 | |
---|
| 355 | #endif |
---|
| 356 | |
---|
| 357 | END SUBROUTINE transpose_yxd |
---|
| 358 | |
---|
| 359 | |
---|
| 360 | SUBROUTINE transpose_yz( f_in, work1, f_inv, work2, f_out ) |
---|
| 361 | |
---|
| 362 | !------------------------------------------------------------------------------! |
---|
| 363 | ! Description: |
---|
| 364 | ! ------------ |
---|
| 365 | ! Transposition of input array (f_in) from y to z. For the input array, all |
---|
| 366 | ! elements along y reside on the same PE, while after transposition, all |
---|
| 367 | ! elements along z reside on the same PE. |
---|
| 368 | !------------------------------------------------------------------------------! |
---|
| 369 | |
---|
| 370 | USE cpulog |
---|
| 371 | USE indices |
---|
| 372 | USE interfaces |
---|
| 373 | USE pegrid |
---|
| 374 | USE transpose_indices |
---|
| 375 | |
---|
| 376 | IMPLICIT NONE |
---|
| 377 | |
---|
| 378 | INTEGER :: i, j, k, l, m, zs |
---|
| 379 | |
---|
| 380 | REAL :: f_in(0:nya,nxl_y:nxr_ya,nzb_y:nzt_ya), & |
---|
| 381 | f_inv(nxl_y:nxr_ya,nzb_y:nzt_ya,0:nya), & |
---|
| 382 | f_out(nxl_z:nxr_za,nys_z:nyn_za,1:nza), & |
---|
| 383 | work1(nxl_y:nxr_ya,nzb_y:nzt_ya,0:nya), work2(nnx*nny*nnz) |
---|
| 384 | |
---|
| 385 | #if defined( __parallel ) |
---|
| 386 | |
---|
| 387 | ! |
---|
| 388 | !-- Rearrange indices of input array in order to make data to be send |
---|
| 389 | !-- by MPI contiguous |
---|
| 390 | DO k = nzb_y, nzt_ya |
---|
| 391 | DO i = nxl_y, nxr_ya |
---|
| 392 | DO j = 0, nya |
---|
| 393 | work1(i,k,j) = f_in(j,i,k) |
---|
| 394 | ENDDO |
---|
| 395 | ENDDO |
---|
| 396 | ENDDO |
---|
| 397 | |
---|
| 398 | ! |
---|
| 399 | !-- Move data to different array, because memory location of work1 is |
---|
| 400 | !-- needed further below (work1 = work2). |
---|
| 401 | !-- If the PE grid is one-dimensional along y, only local reordering |
---|
| 402 | !-- of the data is necessary and no transposition has to be done. |
---|
| 403 | IF ( pdims(1) == 1 ) THEN |
---|
| 404 | DO j = 0, nya |
---|
| 405 | DO k = nzb_y, nzt_ya |
---|
| 406 | DO i = nxl_y, nxr_ya |
---|
| 407 | f_out(i,j,k) = work1(i,k,j) |
---|
| 408 | ENDDO |
---|
| 409 | ENDDO |
---|
| 410 | ENDDO |
---|
| 411 | RETURN |
---|
| 412 | ELSE |
---|
| 413 | DO j = 0, nya |
---|
| 414 | DO k = nzb_y, nzt_ya |
---|
| 415 | DO i = nxl_y, nxr_ya |
---|
| 416 | f_inv(i,k,j) = work1(i,k,j) |
---|
| 417 | ENDDO |
---|
| 418 | ENDDO |
---|
| 419 | ENDDO |
---|
| 420 | ENDIF |
---|
| 421 | |
---|
| 422 | ! |
---|
| 423 | !-- Transpose array |
---|
| 424 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) |
---|
| 425 | CALL MPI_ALLTOALL( f_inv(nxl_y,nzb_y,0), sendrecvcount_yz, MPI_REAL, & |
---|
| 426 | work2(1), sendrecvcount_yz, MPI_REAL, & |
---|
| 427 | comm1dx, ierr ) |
---|
| 428 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) |
---|
| 429 | |
---|
| 430 | ! |
---|
| 431 | !-- Reorder transposed array |
---|
| 432 | m = 0 |
---|
| 433 | DO l = 0, pdims(1) - 1 |
---|
| 434 | zs = 1 + l * ( nzt_ya - nzb_y + 1 ) |
---|
| 435 | DO j = nys_z, nyn_za |
---|
| 436 | DO k = zs, zs + nzt_ya - nzb_y |
---|
| 437 | DO i = nxl_z, nxr_za |
---|
| 438 | m = m + 1 |
---|
| 439 | f_out(i,j,k) = work2(m) |
---|
| 440 | ENDDO |
---|
| 441 | ENDDO |
---|
| 442 | ENDDO |
---|
| 443 | ENDDO |
---|
| 444 | |
---|
| 445 | #endif |
---|
| 446 | |
---|
| 447 | END SUBROUTINE transpose_yz |
---|
| 448 | |
---|
| 449 | |
---|
| 450 | SUBROUTINE transpose_zx( f_in, work1, f_inv, work2, f_out ) |
---|
| 451 | |
---|
| 452 | !------------------------------------------------------------------------------! |
---|
| 453 | ! Description: |
---|
| 454 | ! ------------ |
---|
| 455 | ! Transposition of input array (f_in) from z to x. For the input array, all |
---|
| 456 | ! elements along z reside on the same PE, while after transposition, all |
---|
| 457 | ! elements along x reside on the same PE. |
---|
| 458 | !------------------------------------------------------------------------------! |
---|
| 459 | |
---|
| 460 | USE cpulog |
---|
| 461 | USE indices |
---|
| 462 | USE interfaces |
---|
| 463 | USE pegrid |
---|
| 464 | USE transpose_indices |
---|
| 465 | |
---|
| 466 | IMPLICIT NONE |
---|
| 467 | |
---|
| 468 | INTEGER :: i, j, k, l, m, xs |
---|
| 469 | |
---|
| 470 | REAL :: f_in(1:nza,nys:nyna,nxl:nxra), f_inv(nxl:nxra,nys:nyna,1:nza), & |
---|
| 471 | f_out(0:nxa,nys_x:nyn_xa,nzb_x:nzt_xa), & |
---|
| 472 | work1(nxl:nxra,nys:nyna,1:nza), work2(nnx*nny*nnz) |
---|
| 473 | |
---|
| 474 | #if defined( __parallel ) |
---|
| 475 | |
---|
| 476 | ! |
---|
| 477 | !-- Rearrange indices of input array in order to make data to be send |
---|
| 478 | !-- by MPI contiguous |
---|
| 479 | DO i = nxl, nxra |
---|
| 480 | DO j = nys, nyna |
---|
| 481 | DO k = 1,nza |
---|
| 482 | work1(i,j,k) = f_in(k,j,i) |
---|
| 483 | ENDDO |
---|
| 484 | ENDDO |
---|
| 485 | ENDDO |
---|
| 486 | |
---|
| 487 | ! |
---|
| 488 | !-- Move data to different array, because memory location of work1 is |
---|
| 489 | !-- needed further below (work1 = work2). |
---|
| 490 | !-- If the PE grid is one-dimensional along y, only local reordering |
---|
| 491 | !-- of the data is necessary and no transposition has to be done. |
---|
| 492 | IF ( pdims(1) == 1 ) THEN |
---|
| 493 | DO k = 1, nza |
---|
| 494 | DO j = nys, nyna |
---|
| 495 | DO i = nxl, nxra |
---|
| 496 | f_out(i,j,k) = work1(i,j,k) |
---|
| 497 | ENDDO |
---|
| 498 | ENDDO |
---|
| 499 | ENDDO |
---|
| 500 | RETURN |
---|
| 501 | ELSE |
---|
| 502 | DO k = 1, nza |
---|
| 503 | DO j = nys, nyna |
---|
| 504 | DO i = nxl, nxra |
---|
| 505 | f_inv(i,j,k) = work1(i,j,k) |
---|
| 506 | ENDDO |
---|
| 507 | ENDDO |
---|
| 508 | ENDDO |
---|
| 509 | ENDIF |
---|
| 510 | |
---|
| 511 | ! |
---|
| 512 | !-- Transpose array |
---|
| 513 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) |
---|
| 514 | CALL MPI_ALLTOALL( f_inv(nxl,nys,1), sendrecvcount_zx, MPI_REAL, & |
---|
| 515 | work2(1), sendrecvcount_zx, MPI_REAL, & |
---|
| 516 | comm1dx, ierr ) |
---|
| 517 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) |
---|
| 518 | |
---|
| 519 | ! |
---|
| 520 | !-- Reorder transposed array |
---|
| 521 | m = 0 |
---|
| 522 | DO l = 0, pdims(1) - 1 |
---|
| 523 | xs = 0 + l * nnx |
---|
| 524 | DO k = nzb_x, nzt_xa |
---|
| 525 | DO j = nys_x, nyn_xa |
---|
| 526 | DO i = xs, xs + nnx - 1 |
---|
| 527 | m = m + 1 |
---|
| 528 | f_out(i,j,k) = work2(m) |
---|
| 529 | ENDDO |
---|
| 530 | ENDDO |
---|
| 531 | ENDDO |
---|
| 532 | ENDDO |
---|
| 533 | |
---|
| 534 | #endif |
---|
| 535 | |
---|
| 536 | END SUBROUTINE transpose_zx |
---|
| 537 | |
---|
| 538 | |
---|
| 539 | SUBROUTINE transpose_zy( f_in, work1, f_inv, work2, f_out ) |
---|
| 540 | |
---|
| 541 | !------------------------------------------------------------------------------! |
---|
| 542 | ! Description: |
---|
| 543 | ! ------------ |
---|
| 544 | ! Transposition of input array (f_in) from z to y. For the input array, all |
---|
| 545 | ! elements along z reside on the same PE, while after transposition, all |
---|
| 546 | ! elements along y reside on the same PE. |
---|
| 547 | !------------------------------------------------------------------------------! |
---|
| 548 | |
---|
| 549 | USE cpulog |
---|
| 550 | USE indices |
---|
| 551 | USE interfaces |
---|
| 552 | USE pegrid |
---|
| 553 | USE transpose_indices |
---|
| 554 | |
---|
| 555 | IMPLICIT NONE |
---|
| 556 | |
---|
| 557 | INTEGER :: i, j, k, l, m, zs |
---|
| 558 | |
---|
| 559 | REAL :: f_in(nxl_z:nxr_za,nys_z:nyn_za,1:nza), & |
---|
| 560 | f_inv(nxl_y:nxr_ya,nzb_y:nzt_ya,0:nya), & |
---|
| 561 | f_out(0:nya,nxl_y:nxr_ya,nzb_y:nzt_ya), & |
---|
| 562 | work1(0:nya,nxl_y:nxr_ya,nzb_y:nzt_ya), work2(nnx*nny*nnz) |
---|
| 563 | |
---|
| 564 | #if defined( __parallel ) |
---|
| 565 | |
---|
| 566 | ! |
---|
| 567 | !-- If the PE grid is one-dimensional along y, the array has only to be |
---|
| 568 | !-- reordered locally and therefore no transposition has to be done. |
---|
| 569 | IF ( pdims(1) /= 1 ) THEN |
---|
| 570 | ! |
---|
| 571 | !-- Reorder input array for transposition |
---|
| 572 | m = 0 |
---|
| 573 | DO l = 0, pdims(1) - 1 |
---|
| 574 | zs = 1 + l * ( nzt_ya - nzb_y + 1 ) |
---|
| 575 | DO j = nys_z, nyn_za |
---|
| 576 | DO k = zs, zs + nzt_ya - nzb_y |
---|
| 577 | DO i = nxl_z, nxr_za |
---|
| 578 | m = m + 1 |
---|
| 579 | work2(m) = f_in(i,j,k) |
---|
| 580 | ENDDO |
---|
| 581 | ENDDO |
---|
| 582 | ENDDO |
---|
| 583 | ENDDO |
---|
| 584 | |
---|
| 585 | ! |
---|
| 586 | !-- Transpose array |
---|
| 587 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) |
---|
| 588 | CALL MPI_ALLTOALL( work2(1), sendrecvcount_yz, MPI_REAL, & |
---|
| 589 | f_inv(nxl_y,nzb_y,0), sendrecvcount_yz, MPI_REAL, & |
---|
| 590 | comm1dx, ierr ) |
---|
| 591 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) |
---|
| 592 | |
---|
| 593 | ! |
---|
| 594 | !-- Reorder transposed array in a way that the y index is in first position |
---|
| 595 | DO k = nzb_y, nzt_ya |
---|
| 596 | DO i = nxl_y, nxr_ya |
---|
| 597 | DO j = 0, nya |
---|
| 598 | work1(j,i,k) = f_inv(i,k,j) |
---|
| 599 | ENDDO |
---|
| 600 | ENDDO |
---|
| 601 | ENDDO |
---|
| 602 | ELSE |
---|
| 603 | ! |
---|
| 604 | !-- Reorder the array in a way that the y index is in first position |
---|
| 605 | DO k = nzb_y, nzt_ya |
---|
| 606 | DO i = nxl_y, nxr_ya |
---|
| 607 | DO j = 0, nya |
---|
| 608 | work1(j,i,k) = f_in(i,j,k) |
---|
| 609 | ENDDO |
---|
| 610 | ENDDO |
---|
| 611 | ENDDO |
---|
| 612 | ENDIF |
---|
| 613 | |
---|
| 614 | ! |
---|
| 615 | !-- Move data to output array |
---|
| 616 | DO k = nzb_y, nzt_ya |
---|
| 617 | DO i = nxl_y, nxr_ya |
---|
| 618 | DO j = 0, nya |
---|
| 619 | f_out(j,i,k) = work1(j,i,k) |
---|
| 620 | ENDDO |
---|
| 621 | ENDDO |
---|
| 622 | ENDDO |
---|
| 623 | |
---|
| 624 | #endif |
---|
| 625 | |
---|
| 626 | END SUBROUTINE transpose_zy |
---|
| 627 | |
---|
| 628 | |
---|
| 629 | SUBROUTINE transpose_zyd( f_in, work1, f_inv, work2, f_out ) |
---|
| 630 | |
---|
| 631 | !------------------------------------------------------------------------------! |
---|
| 632 | ! Description: |
---|
| 633 | ! ------------ |
---|
| 634 | ! Transposition of input array (f_in) from z to y. For the input array, all |
---|
| 635 | ! elements along z reside on the same PE, while after transposition, all |
---|
| 636 | ! elements along y reside on the same PE. |
---|
| 637 | ! This is a direct transposition for arrays with indices in regular order |
---|
| 638 | ! (k,j,i) (cf. transpose_zy). |
---|
| 639 | !------------------------------------------------------------------------------! |
---|
| 640 | |
---|
| 641 | USE cpulog |
---|
| 642 | USE indices |
---|
| 643 | USE interfaces |
---|
| 644 | USE pegrid |
---|
| 645 | USE transpose_indices |
---|
| 646 | |
---|
| 647 | IMPLICIT NONE |
---|
| 648 | |
---|
| 649 | INTEGER :: i, j, k, l, m, ys |
---|
| 650 | |
---|
| 651 | REAL :: f_in(1:nza,nys:nyna,nxl:nxra), f_inv(nys:nyna,nxl:nxra,1:nza), & |
---|
| 652 | f_out(0:nya,nxl_yd:nxr_yda,nzb_yd:nzt_yda), & |
---|
| 653 | work1(nys:nyna,nxl:nxra,1:nza), work2(nnx*nny*nnz) |
---|
| 654 | |
---|
| 655 | #if defined( __parallel ) |
---|
| 656 | |
---|
| 657 | ! |
---|
| 658 | !-- Rearrange indices of input array in order to make data to be send |
---|
| 659 | !-- by MPI contiguous |
---|
| 660 | DO i = nxl, nxra |
---|
| 661 | DO j = nys, nyna |
---|
| 662 | DO k = 1, nza |
---|
| 663 | work1(j,i,k) = f_in(k,j,i) |
---|
| 664 | ENDDO |
---|
| 665 | ENDDO |
---|
| 666 | ENDDO |
---|
| 667 | |
---|
| 668 | ! |
---|
| 669 | !-- Move data to different array, because memory location of work1 is |
---|
| 670 | !-- needed further below (work1 = work2). |
---|
| 671 | !-- If the PE grid is one-dimensional along x, only local reordering |
---|
| 672 | !-- of the data is necessary and no transposition has to be done. |
---|
| 673 | IF ( pdims(2) == 1 ) THEN |
---|
| 674 | DO k = 1, nza |
---|
| 675 | DO i = nxl, nxra |
---|
| 676 | DO j = nys, nyna |
---|
| 677 | f_out(j,i,k) = work1(j,i,k) |
---|
| 678 | ENDDO |
---|
| 679 | ENDDO |
---|
| 680 | ENDDO |
---|
| 681 | RETURN |
---|
| 682 | ELSE |
---|
| 683 | DO k = 1, nza |
---|
| 684 | DO i = nxl, nxra |
---|
| 685 | DO j = nys, nyna |
---|
| 686 | f_inv(j,i,k) = work1(j,i,k) |
---|
| 687 | ENDDO |
---|
| 688 | ENDDO |
---|
| 689 | ENDDO |
---|
| 690 | ENDIF |
---|
| 691 | |
---|
| 692 | ! |
---|
| 693 | !-- Transpose array |
---|
| 694 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'start' ) |
---|
| 695 | CALL MPI_ALLTOALL( f_inv(nys,nxl,1), sendrecvcount_zyd, MPI_REAL, & |
---|
| 696 | work2(1), sendrecvcount_zyd, MPI_REAL, & |
---|
| 697 | comm1dy, ierr ) |
---|
| 698 | CALL cpu_log( log_point_s(32), 'mpi_alltoall', 'stop' ) |
---|
| 699 | |
---|
| 700 | ! |
---|
| 701 | !-- Reorder transposed array |
---|
| 702 | m = 0 |
---|
| 703 | DO l = 0, pdims(2) - 1 |
---|
| 704 | ys = 0 + l * nny |
---|
| 705 | DO k = nzb_yd, nzt_yda |
---|
| 706 | DO i = nxl_yd, nxr_yda |
---|
| 707 | DO j = ys, ys + nny - 1 |
---|
| 708 | m = m + 1 |
---|
| 709 | f_out(j,i,k) = work2(m) |
---|
| 710 | ENDDO |
---|
| 711 | ENDDO |
---|
| 712 | ENDDO |
---|
| 713 | ENDDO |
---|
| 714 | |
---|
| 715 | #endif |
---|
| 716 | |
---|
| 717 | END SUBROUTINE transpose_zyd |
---|