[1] | 1 | SUBROUTINE init_pegrid |
---|
| 2 | |
---|
| 3 | !------------------------------------------------------------------------------! |
---|
| 4 | ! Actual revisions: |
---|
| 5 | ! ----------------- |
---|
| 6 | ! |
---|
| 7 | ! |
---|
| 8 | ! Former revisions: |
---|
| 9 | ! ----------------- |
---|
[3] | 10 | ! $Id: init_pegrid.f90 4 2007-02-13 11:33:16Z raasch $ |
---|
| 11 | ! RCS Log replace by Id keyword, revision history cleaned up |
---|
| 12 | ! |
---|
[1] | 13 | ! Revision 1.28 2006/04/26 13:23:32 raasch |
---|
| 14 | ! lcmuk does not understand the !$ comment so a cpp-directive is required |
---|
| 15 | ! |
---|
| 16 | ! Revision 1.1 1997/07/24 11:15:09 raasch |
---|
| 17 | ! Initial revision |
---|
| 18 | ! |
---|
| 19 | ! |
---|
| 20 | ! Description: |
---|
| 21 | ! ------------ |
---|
| 22 | ! Determination of the virtual processor topology (if not prescribed by the |
---|
| 23 | ! user)and computation of the grid point number and array bounds of the local |
---|
| 24 | ! domains. |
---|
| 25 | !------------------------------------------------------------------------------! |
---|
| 26 | |
---|
| 27 | USE control_parameters |
---|
| 28 | USE fft_xy |
---|
| 29 | USE indices |
---|
| 30 | USE pegrid |
---|
| 31 | USE poisfft_mod |
---|
| 32 | USE poisfft_hybrid_mod |
---|
| 33 | USE statistics |
---|
| 34 | USE transpose_indices |
---|
| 35 | |
---|
| 36 | |
---|
| 37 | IMPLICIT NONE |
---|
| 38 | |
---|
| 39 | INTEGER :: gathered_size, i, ind(5), j, k, maximum_grid_level_l, & |
---|
| 40 | mg_switch_to_pe0_level_l, mg_levels_x, mg_levels_y, & |
---|
| 41 | mg_levels_z, nnx_y, nnx_z, nny_x, nny_z, nnz_x, nnz_y, & |
---|
| 42 | numproc_sqr, nx_total, nxl_l, nxr_l, nyn_l, nys_l, nzb_l, & |
---|
| 43 | nzt_l, omp_get_num_threads, subdomain_size |
---|
| 44 | |
---|
| 45 | INTEGER, DIMENSION(:), ALLOCATABLE :: ind_all, nxlf, nxrf, nynf, nysf |
---|
| 46 | |
---|
| 47 | LOGICAL :: found |
---|
| 48 | |
---|
| 49 | ! |
---|
| 50 | !-- Get the number of OpenMP threads |
---|
| 51 | !$OMP PARALLEL |
---|
| 52 | #if defined( __lcmuk ) |
---|
| 53 | threads_per_task = omp_get_num_threads() |
---|
| 54 | #else |
---|
| 55 | !$ threads_per_task = omp_get_num_threads() |
---|
| 56 | #endif |
---|
| 57 | !$OMP END PARALLEL |
---|
| 58 | |
---|
| 59 | |
---|
| 60 | #if defined( __parallel ) |
---|
| 61 | ! |
---|
| 62 | !-- Determine the processor topology or check it, if prescribed by the user |
---|
| 63 | IF ( npex == -1 .AND. npey == -1 ) THEN |
---|
| 64 | |
---|
| 65 | ! |
---|
| 66 | !-- Automatic determination of the topology |
---|
| 67 | !-- The default on SMP- and cluster-hosts is a 1d-decomposition along x |
---|
| 68 | #if defined( __lcmuk ) |
---|
| 69 | host = 'lcmuk' |
---|
| 70 | #endif |
---|
| 71 | IF ( host(1:3) == 'ibm' .OR. host(1:3) == 'nec' .OR. & |
---|
| 72 | host(1:2) == 'lc' .OR. host(1:3) == 'dec' ) THEN |
---|
| 73 | |
---|
| 74 | pdims(1) = numprocs |
---|
| 75 | pdims(2) = 1 |
---|
| 76 | |
---|
| 77 | ELSE |
---|
| 78 | |
---|
| 79 | numproc_sqr = SQRT( REAL( numprocs ) ) |
---|
| 80 | pdims(1) = MAX( numproc_sqr , 1 ) |
---|
| 81 | DO WHILE ( MOD( numprocs , pdims(1) ) /= 0 ) |
---|
| 82 | pdims(1) = pdims(1) - 1 |
---|
| 83 | ENDDO |
---|
| 84 | pdims(2) = numprocs / pdims(1) |
---|
| 85 | |
---|
| 86 | ENDIF |
---|
| 87 | |
---|
| 88 | ELSEIF ( npex /= -1 .AND. npey /= -1 ) THEN |
---|
| 89 | |
---|
| 90 | ! |
---|
| 91 | !-- Prescribed by user. Number of processors on the prescribed topology |
---|
| 92 | !-- must be equal to the number of PEs available to the job |
---|
| 93 | IF ( ( npex * npey ) /= numprocs ) THEN |
---|
| 94 | PRINT*, '+++ init_pegrid:' |
---|
| 95 | PRINT*, ' number of PEs of the prescribed topology (', npex*npey, & |
---|
| 96 | ') does not match the number of PEs available to the ', & |
---|
| 97 | 'job (', numprocs, ')' |
---|
| 98 | CALL local_stop |
---|
| 99 | ENDIF |
---|
| 100 | pdims(1) = npex |
---|
| 101 | pdims(2) = npey |
---|
| 102 | |
---|
| 103 | ELSE |
---|
| 104 | ! |
---|
| 105 | !-- If the processor topology is prescribed by the user, the number of |
---|
| 106 | !-- PEs must be given in both directions |
---|
| 107 | PRINT*, '+++ init_pegrid:' |
---|
| 108 | PRINT*, ' if the processor topology is prescribed by the user, ', & |
---|
| 109 | 'both values of "npex" and "npey" must be given in the ', & |
---|
| 110 | 'NAMELIST-parameter file' |
---|
| 111 | CALL local_stop |
---|
| 112 | |
---|
| 113 | ENDIF |
---|
| 114 | |
---|
| 115 | ! |
---|
| 116 | !-- The hybrid solver can only be used in case of a 1d-decomposition along x |
---|
| 117 | IF ( pdims(2) /= 1 .AND. psolver == 'poisfft_hybrid' ) THEN |
---|
| 118 | IF ( myid == 0 ) THEN |
---|
| 119 | PRINT*, '*** init_pegrid: psolver = "poisfft_hybrid" can only be' |
---|
| 120 | PRINT*, ' used in case of a 1d-decomposition along x' |
---|
| 121 | ENDIF |
---|
| 122 | ENDIF |
---|
| 123 | |
---|
| 124 | ! |
---|
| 125 | !-- If necessary, set horizontal boundary conditions to non-cyclic |
---|
| 126 | IF ( bc_lr /= 'cyclic' ) cyclic(1) = .FALSE. |
---|
| 127 | IF ( bc_ns /= 'cyclic' ) cyclic(2) = .FALSE. |
---|
| 128 | |
---|
| 129 | ! |
---|
| 130 | !-- Create the virtual processor grid |
---|
| 131 | CALL MPI_CART_CREATE( comm_palm, ndim, pdims, cyclic, reorder, & |
---|
| 132 | comm2d, ierr ) |
---|
| 133 | CALL MPI_COMM_RANK( comm2d, myid, ierr ) |
---|
| 134 | WRITE (myid_char,'(''_'',I4.4)') myid |
---|
| 135 | |
---|
| 136 | CALL MPI_CART_COORDS( comm2d, myid, ndim, pcoord, ierr ) |
---|
| 137 | CALL MPI_CART_SHIFT( comm2d, 0, 1, pleft, pright, ierr ) |
---|
| 138 | CALL MPI_CART_SHIFT( comm2d, 1, 1, psouth, pnorth, ierr ) |
---|
| 139 | |
---|
| 140 | ! |
---|
| 141 | !-- Determine sub-topologies for transpositions |
---|
| 142 | !-- Transposition from z to x: |
---|
| 143 | remain_dims(1) = .TRUE. |
---|
| 144 | remain_dims(2) = .FALSE. |
---|
| 145 | CALL MPI_CART_SUB( comm2d, remain_dims, comm1dx, ierr ) |
---|
| 146 | CALL MPI_COMM_RANK( comm1dx, myidx, ierr ) |
---|
| 147 | ! |
---|
| 148 | !-- Transposition from x to y |
---|
| 149 | remain_dims(1) = .FALSE. |
---|
| 150 | remain_dims(2) = .TRUE. |
---|
| 151 | CALL MPI_CART_SUB( comm2d, remain_dims, comm1dy, ierr ) |
---|
| 152 | CALL MPI_COMM_RANK( comm1dy, myidy, ierr ) |
---|
| 153 | |
---|
| 154 | |
---|
| 155 | ! |
---|
| 156 | !-- Find a grid (used for array d) which will match the transposition demands |
---|
| 157 | IF ( grid_matching == 'strict' ) THEN |
---|
| 158 | |
---|
| 159 | nxa = nx; nya = ny; nza = nz |
---|
| 160 | |
---|
| 161 | ELSE |
---|
| 162 | |
---|
| 163 | found = .FALSE. |
---|
| 164 | xn: DO nxa = nx, 2*nx |
---|
| 165 | ! |
---|
| 166 | !-- Meet conditions for nx |
---|
| 167 | IF ( MOD( nxa+1, pdims(1) ) /= 0 .OR. & |
---|
| 168 | MOD( nxa+1, pdims(2) ) /= 0 ) CYCLE xn |
---|
| 169 | |
---|
| 170 | yn: DO nya = ny, 2*ny |
---|
| 171 | ! |
---|
| 172 | !-- Meet conditions for ny |
---|
| 173 | IF ( MOD( nya+1, pdims(2) ) /= 0 .OR. & |
---|
| 174 | MOD( nya+1, pdims(1) ) /= 0 ) CYCLE yn |
---|
| 175 | |
---|
| 176 | |
---|
| 177 | zn: DO nza = nz, 2*nz |
---|
| 178 | ! |
---|
| 179 | !-- Meet conditions for nz |
---|
| 180 | IF ( ( MOD( nza, pdims(1) ) /= 0 .AND. pdims(1) /= 1 .AND. & |
---|
| 181 | pdims(2) /= 1 ) .OR. & |
---|
| 182 | ( MOD( nza, pdims(2) ) /= 0 .AND. dt_dosp /= 9999999.9 & |
---|
| 183 | ) ) THEN |
---|
| 184 | CYCLE zn |
---|
| 185 | ELSE |
---|
| 186 | found = .TRUE. |
---|
| 187 | EXIT xn |
---|
| 188 | ENDIF |
---|
| 189 | |
---|
| 190 | ENDDO zn |
---|
| 191 | |
---|
| 192 | ENDDO yn |
---|
| 193 | |
---|
| 194 | ENDDO xn |
---|
| 195 | |
---|
| 196 | IF ( .NOT. found ) THEN |
---|
| 197 | IF ( myid == 0 ) THEN |
---|
| 198 | PRINT*,'+++ init_pegrid: no matching grid for transpositions found' |
---|
| 199 | ENDIF |
---|
| 200 | CALL local_stop |
---|
| 201 | ENDIF |
---|
| 202 | |
---|
| 203 | ENDIF |
---|
| 204 | |
---|
| 205 | ! |
---|
| 206 | !-- Calculate array bounds in x-direction for every PE. |
---|
| 207 | !-- The last PE along x may get less grid points than the others |
---|
| 208 | ALLOCATE( nxlf(0:pdims(1)-1), nxrf(0:pdims(1)-1), nynf(0:pdims(2)-1), & |
---|
| 209 | nysf(0:pdims(2)-1), nnx_pe(0:pdims(1)-1), nny_pe(0:pdims(2)-1) ) |
---|
| 210 | |
---|
| 211 | IF ( MOD( nxa+1 , pdims(1) ) /= 0 ) THEN |
---|
| 212 | IF ( myid == 0 ) THEN |
---|
| 213 | PRINT*,'+++ x-direction: gridpoint number (',nx+1,') is not an' |
---|
| 214 | PRINT*,' integral divisor of the number of proces', & |
---|
| 215 | &'sors (', pdims(1),')' |
---|
| 216 | ENDIF |
---|
| 217 | CALL local_stop |
---|
| 218 | ELSE |
---|
| 219 | nnx = ( nxa + 1 ) / pdims(1) |
---|
| 220 | IF ( nnx*pdims(1) - ( nx + 1) > nnx ) THEN |
---|
| 221 | IF ( myid == 0 ) THEN |
---|
| 222 | PRINT*,'+++ x-direction: nx does not match the requirements ', & |
---|
| 223 | 'given by the number of PEs' |
---|
| 224 | PRINT*,' used' |
---|
| 225 | PRINT*,' please use nx = ', nx - ( pdims(1) - ( nnx*pdims(1) & |
---|
| 226 | - ( nx + 1 ) ) ), ' instead of nx =', nx |
---|
| 227 | ENDIF |
---|
| 228 | CALL local_stop |
---|
| 229 | ENDIF |
---|
| 230 | ENDIF |
---|
| 231 | |
---|
| 232 | ! |
---|
| 233 | !-- Left and right array bounds, number of gridpoints |
---|
| 234 | DO i = 0, pdims(1)-1 |
---|
| 235 | nxlf(i) = i * nnx |
---|
| 236 | nxrf(i) = ( i + 1 ) * nnx - 1 |
---|
| 237 | nnx_pe(i) = MIN( nx, nxrf(i) ) - nxlf(i) + 1 |
---|
| 238 | ENDDO |
---|
| 239 | |
---|
| 240 | ! |
---|
| 241 | !-- Calculate array bounds in y-direction for every PE. |
---|
| 242 | IF ( MOD( nya+1 , pdims(2) ) /= 0 ) THEN |
---|
| 243 | IF ( myid == 0 ) THEN |
---|
| 244 | PRINT*,'+++ y-direction: gridpoint number (',ny+1,') is not an' |
---|
| 245 | PRINT*,' integral divisor of the number of proces', & |
---|
| 246 | &'sors (', pdims(2),')' |
---|
| 247 | ENDIF |
---|
| 248 | CALL local_stop |
---|
| 249 | ELSE |
---|
| 250 | nny = ( nya + 1 ) / pdims(2) |
---|
| 251 | IF ( nny*pdims(2) - ( ny + 1) > nny ) THEN |
---|
| 252 | IF ( myid == 0 ) THEN |
---|
| 253 | PRINT*,'+++ x-direction: nx does not match the requirements ', & |
---|
| 254 | 'given by the number of PEs' |
---|
| 255 | PRINT*,' used' |
---|
| 256 | PRINT*,' please use nx = ', nx - ( pdims(1) - ( nnx*pdims(1) & |
---|
| 257 | - ( nx + 1 ) ) ), ' instead of nx =', nx |
---|
| 258 | ENDIF |
---|
| 259 | CALL local_stop |
---|
| 260 | ENDIF |
---|
| 261 | ENDIF |
---|
| 262 | |
---|
| 263 | ! |
---|
| 264 | !-- South and north array bounds |
---|
| 265 | DO j = 0, pdims(2)-1 |
---|
| 266 | nysf(j) = j * nny |
---|
| 267 | nynf(j) = ( j + 1 ) * nny - 1 |
---|
| 268 | nny_pe(j) = MIN( ny, nynf(j) ) - nysf(j) + 1 |
---|
| 269 | ENDDO |
---|
| 270 | |
---|
| 271 | ! |
---|
| 272 | !-- Local array bounds of the respective PEs |
---|
| 273 | nxl = nxlf(pcoord(1)) |
---|
| 274 | nxra = nxrf(pcoord(1)) |
---|
| 275 | nxr = MIN( nx, nxra ) |
---|
| 276 | nys = nysf(pcoord(2)) |
---|
| 277 | nyna = nynf(pcoord(2)) |
---|
| 278 | nyn = MIN( ny, nyna ) |
---|
| 279 | nzb = 0 |
---|
| 280 | nzta = nza |
---|
| 281 | nzt = MIN( nz, nzta ) |
---|
| 282 | nnz = nza |
---|
| 283 | |
---|
| 284 | ! |
---|
| 285 | !-- Calculate array bounds and gridpoint numbers for the transposed arrays |
---|
| 286 | !-- (needed in the pressure solver) |
---|
| 287 | !-- For the transposed arrays, cyclic boundaries as well as top and bottom |
---|
| 288 | !-- boundaries are omitted, because they are obstructive to the transposition |
---|
| 289 | |
---|
| 290 | ! |
---|
| 291 | !-- 1. transposition z --> x |
---|
| 292 | !-- This transposition is not neccessary in case of a 1d-decomposition along x, |
---|
| 293 | !-- except that the uptream-spline method is switched on |
---|
| 294 | IF ( pdims(2) /= 1 .OR. momentum_advec == 'ups-scheme' .OR. & |
---|
| 295 | scalar_advec == 'ups-scheme' ) THEN |
---|
| 296 | |
---|
| 297 | IF ( pdims(2) == 1 .AND. ( momentum_advec == 'ups-scheme' .OR. & |
---|
| 298 | scalar_advec == 'ups-scheme' ) ) THEN |
---|
| 299 | IF ( myid == 0 ) THEN |
---|
| 300 | PRINT*,'+++ WARNING: init_pegrid: 1d-decomposition along x ', & |
---|
| 301 | &'chosen but nz restrictions may occur' |
---|
| 302 | PRINT*,' since ups-scheme is activated' |
---|
| 303 | ENDIF |
---|
| 304 | ENDIF |
---|
| 305 | nys_x = nys |
---|
| 306 | nyn_xa = nyna |
---|
| 307 | nyn_x = nyn |
---|
| 308 | nny_x = nny |
---|
| 309 | IF ( MOD( nza , pdims(1) ) /= 0 ) THEN |
---|
| 310 | IF ( myid == 0 ) THEN |
---|
| 311 | PRINT*,'+++ transposition z --> x:' |
---|
| 312 | PRINT*,' nz=',nz,' is not an integral divisior of pdims(1)=', & |
---|
| 313 | &pdims(1) |
---|
| 314 | ENDIF |
---|
| 315 | CALL local_stop |
---|
| 316 | ENDIF |
---|
| 317 | nnz_x = nza / pdims(1) |
---|
| 318 | nzb_x = 1 + myidx * nnz_x |
---|
| 319 | nzt_xa = ( myidx + 1 ) * nnz_x |
---|
| 320 | nzt_x = MIN( nzt, nzt_xa ) |
---|
| 321 | |
---|
| 322 | sendrecvcount_zx = nnx * nny * nnz_x |
---|
| 323 | |
---|
| 324 | ENDIF |
---|
| 325 | |
---|
| 326 | ! |
---|
| 327 | !-- 2. transposition x --> y |
---|
| 328 | nnz_y = nnz_x |
---|
| 329 | nzb_y = nzb_x |
---|
| 330 | nzt_ya = nzt_xa |
---|
| 331 | nzt_y = nzt_x |
---|
| 332 | IF ( MOD( nxa+1 , pdims(2) ) /= 0 ) THEN |
---|
| 333 | IF ( myid == 0 ) THEN |
---|
| 334 | PRINT*,'+++ transposition x --> y:' |
---|
| 335 | PRINT*,' nx+1=',nx+1,' is not an integral divisor of ',& |
---|
| 336 | &'pdims(2)=',pdims(2) |
---|
| 337 | ENDIF |
---|
| 338 | CALL local_stop |
---|
| 339 | ENDIF |
---|
| 340 | nnx_y = (nxa+1) / pdims(2) |
---|
| 341 | nxl_y = myidy * nnx_y |
---|
| 342 | nxr_ya = ( myidy + 1 ) * nnx_y - 1 |
---|
| 343 | nxr_y = MIN( nx, nxr_ya ) |
---|
| 344 | |
---|
| 345 | sendrecvcount_xy = nnx_y * nny_x * nnz_y |
---|
| 346 | |
---|
| 347 | ! |
---|
| 348 | !-- 3. transposition y --> z (ELSE: x --> y in case of 1D-decomposition |
---|
| 349 | !-- along x) |
---|
| 350 | IF ( pdims(2) /= 1 .OR. momentum_advec == 'ups-scheme' .OR. & |
---|
| 351 | scalar_advec == 'ups-scheme' ) THEN |
---|
| 352 | ! |
---|
| 353 | !-- y --> z |
---|
| 354 | !-- This transposition is not neccessary in case of a 1d-decomposition |
---|
| 355 | !-- along x, except that the uptream-spline method is switched on |
---|
| 356 | nnx_z = nnx_y |
---|
| 357 | nxl_z = nxl_y |
---|
| 358 | nxr_za = nxr_ya |
---|
| 359 | nxr_z = nxr_y |
---|
| 360 | IF ( MOD( nya+1 , pdims(1) ) /= 0 ) THEN |
---|
| 361 | IF ( myid == 0 ) THEN |
---|
| 362 | PRINT*,'+++ Transposition y --> z:' |
---|
| 363 | PRINT*,' ny+1=',ny+1,' is not an integral divisor of ',& |
---|
| 364 | &'pdims(1)=',pdims(1) |
---|
| 365 | ENDIF |
---|
| 366 | CALL local_stop |
---|
| 367 | ENDIF |
---|
| 368 | nny_z = (nya+1) / pdims(1) |
---|
| 369 | nys_z = myidx * nny_z |
---|
| 370 | nyn_za = ( myidx + 1 ) * nny_z - 1 |
---|
| 371 | nyn_z = MIN( ny, nyn_za ) |
---|
| 372 | |
---|
| 373 | sendrecvcount_yz = nnx_y * nny_z * nnz_y |
---|
| 374 | |
---|
| 375 | ELSE |
---|
| 376 | ! |
---|
| 377 | !-- x --> y. This condition must be fulfilled for a 1D-decomposition along x |
---|
| 378 | IF ( MOD( nya+1 , pdims(1) ) /= 0 ) THEN |
---|
| 379 | IF ( myid == 0 ) THEN |
---|
| 380 | PRINT*,'+++ Transposition x --> y:' |
---|
| 381 | PRINT*,' ny+1=',ny+1,' is not an integral divisor of ',& |
---|
| 382 | &'pdims(1)=',pdims(1) |
---|
| 383 | ENDIF |
---|
| 384 | CALL local_stop |
---|
| 385 | ENDIF |
---|
| 386 | |
---|
| 387 | ENDIF |
---|
| 388 | |
---|
| 389 | ! |
---|
| 390 | !-- Indices for direct transpositions z --> y (used for calculating spectra) |
---|
| 391 | IF ( dt_dosp /= 9999999.9 ) THEN |
---|
| 392 | IF ( MOD( nza, pdims(2) ) /= 0 ) THEN |
---|
| 393 | IF ( myid == 0 ) THEN |
---|
| 394 | PRINT*,'+++ Direct transposition z --> y (needed for spectra):' |
---|
| 395 | PRINT*,' nz=',nz,' is not an integral divisor of ',& |
---|
| 396 | &'pdims(2)=',pdims(2) |
---|
| 397 | ENDIF |
---|
| 398 | CALL local_stop |
---|
| 399 | ELSE |
---|
| 400 | nxl_yd = nxl |
---|
| 401 | nxr_yda = nxra |
---|
| 402 | nxr_yd = nxr |
---|
| 403 | nzb_yd = 1 + myidy * ( nza / pdims(2) ) |
---|
| 404 | nzt_yda = ( myidy + 1 ) * ( nza / pdims(2) ) |
---|
| 405 | nzt_yd = MIN( nzt, nzt_yda ) |
---|
| 406 | |
---|
| 407 | sendrecvcount_zyd = nnx * nny * ( nza / pdims(2) ) |
---|
| 408 | ENDIF |
---|
| 409 | ENDIF |
---|
| 410 | |
---|
| 411 | ! |
---|
| 412 | !-- Indices for direct transpositions y --> x (they are only possible in case |
---|
| 413 | !-- of a 1d-decomposition along x) |
---|
| 414 | IF ( pdims(2) == 1 ) THEN |
---|
| 415 | nny_x = nny / pdims(1) |
---|
| 416 | nys_x = myid * nny_x |
---|
| 417 | nyn_xa = ( myid + 1 ) * nny_x - 1 |
---|
| 418 | nyn_x = MIN( ny, nyn_xa ) |
---|
| 419 | nzb_x = 1 |
---|
| 420 | nzt_xa = nza |
---|
| 421 | nzt_x = nz |
---|
| 422 | sendrecvcount_xy = nnx * nny_x * nza |
---|
| 423 | ENDIF |
---|
| 424 | |
---|
| 425 | ! |
---|
| 426 | !-- Indices for direct transpositions x --> y (they are only possible in case |
---|
| 427 | !-- of a 1d-decomposition along y) |
---|
| 428 | IF ( pdims(1) == 1 ) THEN |
---|
| 429 | nnx_y = nnx / pdims(2) |
---|
| 430 | nxl_y = myid * nnx_y |
---|
| 431 | nxr_ya = ( myid + 1 ) * nnx_y - 1 |
---|
| 432 | nxr_y = MIN( nx, nxr_ya ) |
---|
| 433 | nzb_y = 1 |
---|
| 434 | nzt_ya = nza |
---|
| 435 | nzt_y = nz |
---|
| 436 | sendrecvcount_xy = nnx_y * nny * nza |
---|
| 437 | ENDIF |
---|
| 438 | |
---|
| 439 | ! |
---|
| 440 | !-- Arrays for storing the array bounds are needed any more |
---|
| 441 | DEALLOCATE( nxlf , nxrf , nynf , nysf ) |
---|
| 442 | |
---|
| 443 | #if defined( __print ) |
---|
| 444 | ! |
---|
| 445 | !-- Control output |
---|
| 446 | IF ( myid == 0 ) THEN |
---|
| 447 | PRINT*, '*** processor topology ***' |
---|
| 448 | PRINT*, ' ' |
---|
| 449 | PRINT*, 'myid pcoord left right south north idx idy nxl: nxr',& |
---|
| 450 | &' nys: nyn' |
---|
| 451 | PRINT*, '------------------------------------------------------------',& |
---|
| 452 | &'-----------' |
---|
| 453 | WRITE (*,1000) 0, pcoord(1), pcoord(2), pleft, pright, psouth, pnorth, & |
---|
| 454 | myidx, myidy, nxl, nxr, nys, nyn |
---|
| 455 | 1000 FORMAT (I4,2X,'(',I3,',',I3,')',3X,I4,2X,I4,3X,I4,2X,I4,2X,I3,1X,I3, & |
---|
| 456 | 2(2X,I4,':',I4)) |
---|
| 457 | |
---|
| 458 | ! |
---|
| 459 | !-- Recieve data from the other PEs |
---|
| 460 | DO i = 1,numprocs-1 |
---|
| 461 | CALL MPI_RECV( ibuf, 12, MPI_INTEGER, i, MPI_ANY_TAG, comm2d, status, & |
---|
| 462 | ierr ) |
---|
| 463 | WRITE (*,1000) i, ( ibuf(j) , j = 1,12 ) |
---|
| 464 | ENDDO |
---|
| 465 | ELSE |
---|
| 466 | |
---|
| 467 | ! |
---|
| 468 | !-- Send data to PE0 |
---|
| 469 | ibuf(1) = pcoord(1); ibuf(2) = pcoord(2); ibuf(3) = pleft |
---|
| 470 | ibuf(4) = pright; ibuf(5) = psouth; ibuf(6) = pnorth; ibuf(7) = myidx |
---|
| 471 | ibuf(8) = myidy; ibuf(9) = nxl; ibuf(10) = nxr; ibuf(11) = nys |
---|
| 472 | ibuf(12) = nyn |
---|
| 473 | CALL MPI_SEND( ibuf, 12, MPI_INTEGER, 0, myid, comm2d, ierr ) |
---|
| 474 | ENDIF |
---|
| 475 | #endif |
---|
| 476 | |
---|
| 477 | #else |
---|
| 478 | |
---|
| 479 | ! |
---|
| 480 | !-- Array bounds when running on a single PE (respectively a non-parallel |
---|
| 481 | !-- machine) |
---|
| 482 | nxl = 0 |
---|
| 483 | nxr = nx |
---|
| 484 | nxra = nx |
---|
| 485 | nnx = nxr - nxl + 1 |
---|
| 486 | nys = 0 |
---|
| 487 | nyn = ny |
---|
| 488 | nyna = ny |
---|
| 489 | nny = nyn - nys + 1 |
---|
| 490 | nzb = 0 |
---|
| 491 | nzt = nz |
---|
| 492 | nzta = nz |
---|
| 493 | nnz = nz |
---|
| 494 | |
---|
| 495 | ! |
---|
| 496 | !-- For non-cyclic boundaries extend array u (v) by one gridpoint |
---|
| 497 | IF ( bc_lr /= 'cyclic' ) uxrp = 1 |
---|
| 498 | IF ( bc_ns /= 'cyclic' ) vynp = 1 |
---|
| 499 | |
---|
| 500 | ! |
---|
| 501 | !-- Array bounds for the pressure solver (in the parallel code, these bounds |
---|
| 502 | !-- are the ones for the transposed arrays) |
---|
| 503 | nys_x = nys |
---|
| 504 | nyn_x = nyn |
---|
| 505 | nyn_xa = nyn |
---|
| 506 | nzb_x = nzb + 1 |
---|
| 507 | nzt_x = nzt |
---|
| 508 | nzt_xa = nzt |
---|
| 509 | |
---|
| 510 | nxl_y = nxl |
---|
| 511 | nxr_y = nxr |
---|
| 512 | nxr_ya = nxr |
---|
| 513 | nzb_y = nzb + 1 |
---|
| 514 | nzt_y = nzt |
---|
| 515 | nzt_ya = nzt |
---|
| 516 | |
---|
| 517 | nxl_z = nxl |
---|
| 518 | nxr_z = nxr |
---|
| 519 | nxr_za = nxr |
---|
| 520 | nys_z = nys |
---|
| 521 | nyn_z = nyn |
---|
| 522 | nyn_za = nyn |
---|
| 523 | |
---|
| 524 | #endif |
---|
| 525 | |
---|
| 526 | ! |
---|
| 527 | !-- Calculate number of grid levels necessary for the multigrid poisson solver |
---|
| 528 | !-- as well as the gridpoint indices on each level |
---|
| 529 | IF ( psolver == 'multigrid' ) THEN |
---|
| 530 | |
---|
| 531 | ! |
---|
| 532 | !-- First calculate number of possible grid levels for the subdomains |
---|
| 533 | mg_levels_x = 1 |
---|
| 534 | mg_levels_y = 1 |
---|
| 535 | mg_levels_z = 1 |
---|
| 536 | |
---|
| 537 | i = nnx |
---|
| 538 | DO WHILE ( MOD( i, 2 ) == 0 .AND. i /= 2 ) |
---|
| 539 | i = i / 2 |
---|
| 540 | mg_levels_x = mg_levels_x + 1 |
---|
| 541 | ENDDO |
---|
| 542 | |
---|
| 543 | j = nny |
---|
| 544 | DO WHILE ( MOD( j, 2 ) == 0 .AND. j /= 2 ) |
---|
| 545 | j = j / 2 |
---|
| 546 | mg_levels_y = mg_levels_y + 1 |
---|
| 547 | ENDDO |
---|
| 548 | |
---|
| 549 | k = nnz |
---|
| 550 | DO WHILE ( MOD( k, 2 ) == 0 .AND. k /= 2 ) |
---|
| 551 | k = k / 2 |
---|
| 552 | mg_levels_z = mg_levels_z + 1 |
---|
| 553 | ENDDO |
---|
| 554 | |
---|
| 555 | maximum_grid_level = MIN( mg_levels_x, mg_levels_y, mg_levels_z ) |
---|
| 556 | |
---|
| 557 | ! |
---|
| 558 | !-- Find out, if the total domain allows more levels. These additional |
---|
| 559 | !-- levels are processed on PE0 only. |
---|
| 560 | IF ( numprocs > 1 ) THEN |
---|
| 561 | IF ( mg_levels_z > MIN( mg_levels_x, mg_levels_y ) ) THEN |
---|
| 562 | mg_switch_to_pe0_level_l = maximum_grid_level |
---|
| 563 | |
---|
| 564 | mg_levels_x = 1 |
---|
| 565 | mg_levels_y = 1 |
---|
| 566 | |
---|
| 567 | i = nx+1 |
---|
| 568 | DO WHILE ( MOD( i, 2 ) == 0 .AND. i /= 2 ) |
---|
| 569 | i = i / 2 |
---|
| 570 | mg_levels_x = mg_levels_x + 1 |
---|
| 571 | ENDDO |
---|
| 572 | |
---|
| 573 | j = ny+1 |
---|
| 574 | DO WHILE ( MOD( j, 2 ) == 0 .AND. j /= 2 ) |
---|
| 575 | j = j / 2 |
---|
| 576 | mg_levels_y = mg_levels_y + 1 |
---|
| 577 | ENDDO |
---|
| 578 | |
---|
| 579 | maximum_grid_level_l = MIN( mg_levels_x, mg_levels_y, mg_levels_z ) |
---|
| 580 | |
---|
| 581 | IF ( maximum_grid_level_l > mg_switch_to_pe0_level_l ) THEN |
---|
| 582 | mg_switch_to_pe0_level_l = maximum_grid_level_l - & |
---|
| 583 | mg_switch_to_pe0_level_l + 1 |
---|
| 584 | ELSE |
---|
| 585 | mg_switch_to_pe0_level_l = 0 |
---|
| 586 | ENDIF |
---|
| 587 | ELSE |
---|
| 588 | mg_switch_to_pe0_level_l = 0 |
---|
| 589 | maximum_grid_level_l = maximum_grid_level |
---|
| 590 | ENDIF |
---|
| 591 | |
---|
| 592 | ! |
---|
| 593 | !-- Use switch level calculated above only if it is not pre-defined |
---|
| 594 | !-- by user |
---|
| 595 | IF ( mg_switch_to_pe0_level == 0 ) THEN |
---|
| 596 | |
---|
| 597 | IF ( mg_switch_to_pe0_level_l /= 0 ) THEN |
---|
| 598 | mg_switch_to_pe0_level = mg_switch_to_pe0_level_l |
---|
| 599 | maximum_grid_level = maximum_grid_level_l |
---|
| 600 | ENDIF |
---|
| 601 | |
---|
| 602 | ELSE |
---|
| 603 | ! |
---|
| 604 | !-- Check pre-defined value and reset to default, if neccessary |
---|
| 605 | IF ( mg_switch_to_pe0_level < mg_switch_to_pe0_level_l .OR. & |
---|
| 606 | mg_switch_to_pe0_level >= maximum_grid_level_l ) THEN |
---|
| 607 | IF ( myid == 0 ) THEN |
---|
| 608 | PRINT*, '+++ WARNING init_pegrid: mg_switch_to_pe0_level ', & |
---|
| 609 | 'out of range and reset to default (=0)' |
---|
| 610 | ENDIF |
---|
| 611 | mg_switch_to_pe0_level = 0 |
---|
| 612 | ELSE |
---|
| 613 | ! |
---|
| 614 | !-- Use the largest number of possible levels anyway and recalculate |
---|
| 615 | !-- the switch level to this largest number of possible values |
---|
| 616 | maximum_grid_level = maximum_grid_level_l |
---|
| 617 | |
---|
| 618 | ENDIF |
---|
| 619 | ENDIF |
---|
| 620 | |
---|
| 621 | ENDIF |
---|
| 622 | |
---|
| 623 | ALLOCATE( grid_level_count(maximum_grid_level), & |
---|
| 624 | nxl_mg(maximum_grid_level), nxr_mg(maximum_grid_level), & |
---|
| 625 | nyn_mg(maximum_grid_level), nys_mg(maximum_grid_level), & |
---|
| 626 | nzt_mg(maximum_grid_level) ) |
---|
| 627 | |
---|
| 628 | grid_level_count = 0 |
---|
| 629 | nxl_l = nxl; nxr_l = nxr; nys_l = nys; nyn_l = nyn; nzt_l = nzt |
---|
| 630 | |
---|
| 631 | DO i = maximum_grid_level, 1 , -1 |
---|
| 632 | |
---|
| 633 | IF ( i == mg_switch_to_pe0_level ) THEN |
---|
| 634 | #if defined( __parallel ) |
---|
| 635 | ! |
---|
| 636 | !-- Save the grid size of the subdomain at the switch level, because |
---|
| 637 | !-- it is needed in poismg. |
---|
| 638 | !-- Array bounds of the local subdomain grids are gathered on PE0 |
---|
| 639 | ind(1) = nxl_l; ind(2) = nxr_l |
---|
| 640 | ind(3) = nys_l; ind(4) = nyn_l |
---|
| 641 | ind(5) = nzt_l |
---|
| 642 | ALLOCATE( ind_all(5*numprocs), mg_loc_ind(5,0:numprocs-1) ) |
---|
| 643 | CALL MPI_ALLGATHER( ind, 5, MPI_INTEGER, ind_all, 5, & |
---|
| 644 | MPI_INTEGER, comm2d, ierr ) |
---|
| 645 | DO j = 0, numprocs-1 |
---|
| 646 | DO k = 1, 5 |
---|
| 647 | mg_loc_ind(k,j) = ind_all(k+j*5) |
---|
| 648 | ENDDO |
---|
| 649 | ENDDO |
---|
| 650 | DEALLOCATE( ind_all ) |
---|
| 651 | ! |
---|
| 652 | !-- Calculate the grid size of the total domain gathered on PE0 |
---|
| 653 | nxr_l = ( nxr_l-nxl_l+1 ) * pdims(1) - 1 |
---|
| 654 | nxl_l = 0 |
---|
| 655 | nyn_l = ( nyn_l-nys_l+1 ) * pdims(2) - 1 |
---|
| 656 | nys_l = 0 |
---|
| 657 | ! |
---|
| 658 | !-- The size of this gathered array must not be larger than the |
---|
| 659 | !-- array tend, which is used in the multigrid scheme as a temporary |
---|
| 660 | !-- array |
---|
| 661 | subdomain_size = ( nxr - nxl + 3 ) * ( nyn - nys + 3 ) * & |
---|
| 662 | ( nzt - nzb + 2 ) |
---|
| 663 | gathered_size = ( nxr_l - nxl_l + 3 ) * ( nyn_l - nys_l + 3 ) * & |
---|
| 664 | ( nzt_l - nzb + 2 ) |
---|
| 665 | |
---|
| 666 | IF ( gathered_size > subdomain_size ) THEN |
---|
| 667 | IF ( myid == 0 ) THEN |
---|
| 668 | PRINT*, '+++ init_pegrid: not enough memory for storing ', & |
---|
| 669 | 'gathered multigrid data on PE0' |
---|
| 670 | ENDIF |
---|
| 671 | CALL local_stop |
---|
| 672 | ENDIF |
---|
| 673 | #else |
---|
| 674 | PRINT*, '+++ init_pegrid: multigrid gather/scatter impossible ', & |
---|
| 675 | 'in non parallel mode' |
---|
| 676 | CALL local_stop |
---|
| 677 | #endif |
---|
| 678 | ENDIF |
---|
| 679 | |
---|
| 680 | nxl_mg(i) = nxl_l |
---|
| 681 | nxr_mg(i) = nxr_l |
---|
| 682 | nys_mg(i) = nys_l |
---|
| 683 | nyn_mg(i) = nyn_l |
---|
| 684 | nzt_mg(i) = nzt_l |
---|
| 685 | |
---|
| 686 | nxl_l = nxl_l / 2 |
---|
| 687 | nxr_l = nxr_l / 2 |
---|
| 688 | nys_l = nys_l / 2 |
---|
| 689 | nyn_l = nyn_l / 2 |
---|
| 690 | nzt_l = nzt_l / 2 |
---|
| 691 | ENDDO |
---|
| 692 | |
---|
| 693 | ELSE |
---|
| 694 | |
---|
| 695 | maximum_grid_level = 1 |
---|
| 696 | |
---|
| 697 | ENDIF |
---|
| 698 | |
---|
| 699 | grid_level = maximum_grid_level |
---|
| 700 | |
---|
| 701 | #if defined( __parallel ) |
---|
| 702 | ! |
---|
| 703 | !-- Gridpoint number for the exchange of ghost points (y-line for 2D-arrays) |
---|
| 704 | ngp_y = nyn - nys + 1 |
---|
| 705 | |
---|
| 706 | ! |
---|
| 707 | !-- Define a new MPI derived datatype for the exchange of ghost points in |
---|
| 708 | !-- y-direction for 2D-arrays (line) |
---|
| 709 | CALL MPI_TYPE_VECTOR( nxr-nxl+3, 1, ngp_y+2, MPI_REAL, type_x, ierr ) |
---|
| 710 | CALL MPI_TYPE_COMMIT( type_x, ierr ) |
---|
| 711 | CALL MPI_TYPE_VECTOR( nxr-nxl+3, 1, ngp_y+2, MPI_INTEGER, type_x_int, ierr ) |
---|
| 712 | CALL MPI_TYPE_COMMIT( type_x_int, ierr ) |
---|
| 713 | |
---|
| 714 | ! |
---|
| 715 | !-- Calculate gridpoint numbers for the exchange of ghost points along x |
---|
| 716 | !-- (yz-plane for 3D-arrays) and define MPI derived data type(s) for the |
---|
| 717 | !-- exchange of ghost points in y-direction (xz-plane). |
---|
| 718 | !-- Do these calculations for the model grid and (if necessary) also |
---|
| 719 | !-- for the coarser grid levels used in the multigrid method |
---|
| 720 | ALLOCATE ( ngp_yz(maximum_grid_level), type_xz(maximum_grid_level) ) |
---|
| 721 | |
---|
| 722 | nxl_l = nxl; nxr_l = nxr; nys_l = nys; nyn_l = nyn; nzb_l = nzb; nzt_l = nzt |
---|
| 723 | |
---|
| 724 | DO i = maximum_grid_level, 1 , -1 |
---|
| 725 | ngp_yz(i) = (nzt_l - nzb_l + 2) * (nyn_l - nys_l + 3) |
---|
| 726 | |
---|
| 727 | CALL MPI_TYPE_VECTOR( nxr_l-nxl_l+3, nzt_l-nzb_l+2, ngp_yz(i), & |
---|
| 728 | MPI_REAL, type_xz(i), ierr ) |
---|
| 729 | CALL MPI_TYPE_COMMIT( type_xz(i), ierr ) |
---|
| 730 | |
---|
| 731 | nxl_l = nxl_l / 2 |
---|
| 732 | nxr_l = nxr_l / 2 |
---|
| 733 | nys_l = nys_l / 2 |
---|
| 734 | nyn_l = nyn_l / 2 |
---|
| 735 | nzt_l = nzt_l / 2 |
---|
| 736 | ENDDO |
---|
| 737 | #endif |
---|
| 738 | |
---|
| 739 | #if defined( __parallel ) |
---|
| 740 | ! |
---|
| 741 | !-- Setting of flags for inflow/outflow conditions in case of non-cyclic |
---|
| 742 | !-- horizontal boundary conditions. Set variables for extending array u (v) |
---|
| 743 | !-- by one gridpoint on the left/rightmost (northest/southest) processor |
---|
| 744 | IF ( pleft == MPI_PROC_NULL ) THEN |
---|
| 745 | IF ( bc_lr == 'dirichlet/neumann' ) THEN |
---|
| 746 | inflow_l = .TRUE. |
---|
| 747 | ELSEIF ( bc_lr == 'neumann/dirichlet' ) THEN |
---|
| 748 | outflow_l = .TRUE. |
---|
| 749 | ENDIF |
---|
| 750 | ENDIF |
---|
| 751 | |
---|
| 752 | IF ( pright == MPI_PROC_NULL ) THEN |
---|
| 753 | IF ( bc_lr == 'dirichlet/neumann' ) THEN |
---|
| 754 | outflow_r = .TRUE. |
---|
| 755 | ELSEIF ( bc_lr == 'neumann/dirichlet' ) THEN |
---|
| 756 | inflow_r = .TRUE. |
---|
| 757 | ENDIF |
---|
| 758 | uxrp = 1 |
---|
| 759 | ENDIF |
---|
| 760 | |
---|
| 761 | IF ( psouth == MPI_PROC_NULL ) THEN |
---|
| 762 | IF ( bc_ns == 'dirichlet/neumann' ) THEN |
---|
| 763 | outflow_s = .TRUE. |
---|
| 764 | ELSEIF ( bc_ns == 'neumann/dirichlet' ) THEN |
---|
| 765 | inflow_s = .TRUE. |
---|
| 766 | ENDIF |
---|
| 767 | ENDIF |
---|
| 768 | |
---|
| 769 | IF ( pnorth == MPI_PROC_NULL ) THEN |
---|
| 770 | IF ( bc_ns == 'dirichlet/neumann' ) THEN |
---|
| 771 | inflow_n = .TRUE. |
---|
| 772 | ELSEIF ( bc_ns == 'neumann/dirichlet' ) THEN |
---|
| 773 | outflow_n = .TRUE. |
---|
| 774 | ENDIF |
---|
| 775 | vynp = 1 |
---|
| 776 | ENDIF |
---|
| 777 | |
---|
| 778 | ! |
---|
| 779 | !-- Additional MPI derived data type for the exchange of ghost points along x |
---|
| 780 | !-- needed in case of non-cyclic boundary conditions along y on the northmost |
---|
| 781 | !-- processors (for the exchange of the enlarged v array) |
---|
| 782 | IF ( bc_ns /= 'cyclic' .AND. pnorth == MPI_PROC_NULL ) THEN |
---|
| 783 | ngp_yz_p = ( nzt - nzb + 2 ) * ( nyn + vynp - nys + 3 ) |
---|
| 784 | CALL MPI_TYPE_VECTOR( nxr-nxl+3, nzt-nzb+2, ngp_yz_p, & |
---|
| 785 | MPI_REAL, type_xz_p, ierr ) |
---|
| 786 | CALL MPI_TYPE_COMMIT( type_xz_p, ierr ) |
---|
| 787 | ENDIF |
---|
| 788 | #else |
---|
| 789 | IF ( bc_lr == 'dirichlet/neumann' ) THEN |
---|
| 790 | inflow_l = .TRUE. |
---|
| 791 | outflow_r = .TRUE. |
---|
| 792 | uxrp = 1 |
---|
| 793 | ELSEIF ( bc_lr == 'neumann/dirichlet' ) THEN |
---|
| 794 | outflow_l = .TRUE. |
---|
| 795 | inflow_r = .TRUE. |
---|
| 796 | ENDIF |
---|
| 797 | |
---|
| 798 | IF ( bc_ns == 'dirichlet/neumann' ) THEN |
---|
| 799 | inflow_n = .TRUE. |
---|
| 800 | outflow_s = .TRUE. |
---|
| 801 | ELSEIF ( bc_ns == 'neumann/dirichlet' ) THEN |
---|
| 802 | outflow_n = .TRUE. |
---|
| 803 | inflow_s = .TRUE. |
---|
| 804 | vynp = 1 |
---|
| 805 | ENDIF |
---|
| 806 | #endif |
---|
| 807 | |
---|
| 808 | IF ( psolver == 'poisfft_hybrid' ) THEN |
---|
| 809 | CALL poisfft_hybrid_ini |
---|
| 810 | ELSE |
---|
| 811 | CALL poisfft_init |
---|
| 812 | ENDIF |
---|
| 813 | |
---|
| 814 | END SUBROUTINE init_pegrid |
---|