| 1 |
C $Header: /u/gcmpack/models/MITgcmUV/eesupp/src/exch.F,v 1.4.2.2 1998/06/22 02:11:15 cnh Exp $ |
| 2 |
|
| 3 |
#include "CPP_EEOPTIONS.h" |
| 4 |
|
| 5 |
C-- File exch.F: Routines that perform atomic communication operations |
| 6 |
C i.e. communication operations that are completed by |
| 7 |
C time the routines return. |
| 8 |
C Note |
| 9 |
C ==== |
| 10 |
C The code in here although intricate is fairly |
| 11 |
C straightforward. There are four routines, one |
| 12 |
C for each of the data patterns we wish to do overlap |
| 13 |
C updates on - as listed below. Each routine has two |
| 14 |
C parts. The first part does overlap updates to and from |
| 15 |
C remote "processes", that is processes who do not truly |
| 16 |
C share the address space of this process. This part |
| 17 |
C requires a facility like MPI or CRAY shmem get and put |
| 18 |
C operations. It is possible to switch this part on or |
| 19 |
C off using compiler directives. In the case of a simple |
| 20 |
C serial execution nothing happens in this part. This part |
| 21 |
C is also switched off when communication strategies |
| 22 |
C that allow the overlapping of communication and computation |
| 23 |
C are employed. |
| 24 |
C The second part of each routine does the true shared memory |
| 25 |
C overlap copying i.e. copying from one part of array phi |
| 26 |
C to another part. This part is always active, in the case |
| 27 |
C of a single threaded messaging code, however, this part |
| 28 |
C will not do any copies as all edges will be flagged as using |
| 29 |
C for example MPI or SHMPUT, SHMGET. |
| 30 |
C Contents |
| 31 |
C o exch_xy_r4 - Does overlap updates for REAL*4 2d XY fields |
| 32 |
C o exch_xy_r8 - Does overlap updates for REAL*8 2d XY fields |
| 33 |
C o exch_xyz_r4 - Does overlap updates for REAL*4 3d XYZ fields |
| 34 |
C o exch_xyz_r8 - Does overlap updates for REAL*8 3d XYZ fields |
| 35 |
|
| 36 |
CStartOfInterface |
| 37 |
SUBROUTINE EXCH_XY_R4( |
| 38 |
U phi, |
| 39 |
I myThid ) |
| 40 |
C /==========================================================\ |
| 41 |
C | SUBROUTINE EXCH_XY_R4 | |
| 42 |
C | o Handle exchanges for real*4, two-dimensional arrays. | |
| 43 |
C |==========================================================| |
| 44 |
C | Do true shared-memory data transfers and "messaging" | |
| 45 |
C | tranfers for blocking case of data transfers. | |
| 46 |
C | Applications call this routine using | |
| 47 |
C | CALL EXCH..( x, myThid ) | |
| 48 |
C | where x is a two-dimensional array with overlaps. | |
| 49 |
C | This routine does true-shared-memory copies for blocks | |
| 50 |
C | within a thread. It will also do MPI meesaging between | |
| 51 |
C | different processes. | |
| 52 |
C | Note: | |
| 53 |
C | ===== | |
| 54 |
C | If it is used, "asynchronous" messaging in which | |
| 55 |
C | communication overlaps computation is handled elsewhere | |
| 56 |
C | - see recv.F and send.F. | |
| 57 |
C | In practice MPI implementations may not be completely | |
| 58 |
C | thread safe. In principle this code is correct even for | |
| 59 |
C | mixed MPI and multi-threaded execution. | |
| 60 |
C | In multi-thread execution data managed by other threads | |
| 61 |
C | is only automatically visible to another thread if it | |
| 62 |
C | existed before the thread was created. This means that | |
| 63 |
C | all Fortran we declare arrays with overlap regions in | |
| 64 |
C | COMMON blocks. | |
| 65 |
C \==========================================================/ |
| 66 |
|
| 67 |
C === Global data === |
| 68 |
#include "SIZE.h" |
| 69 |
#include "EEPARAMS.h" |
| 70 |
#include "EESUPPORT.h" |
| 71 |
|
| 72 |
C === Routine arguments === |
| 73 |
C phi - Array who's overlap regions are to be exchanged |
| 74 |
C myThid - My thread id. |
| 75 |
Real*4 phi(1-OLx:sNx+OLx,1-OLy:sNy+OLy,nSx,nSy) |
| 76 |
INTEGER myThid |
| 77 |
CEndOfInterface |
| 78 |
|
| 79 |
C === Local variables === |
| 80 |
C bi,bj - Outer loop counters |
| 81 |
C I,J,K - Inner loop counters |
| 82 |
INTEGER bi, bj |
| 83 |
INTEGER I, J |
| 84 |
#ifdef ALLOW_USE_MPI |
| 85 |
C tagSend - Tags used to mark and select messages |
| 86 |
C tagRecv |
| 87 |
C nReqs - MPI request counter. |
| 88 |
C reqIds - Outstanding requests to wait on. |
| 89 |
C mpiRC - MPI return code |
| 90 |
C mpiStatArr - Multi-request status reporting array. |
| 91 |
C toPid - Proc. id sending to |
| 92 |
C fromPid - Proc. id receving from |
| 93 |
C elCount - Number of items to send or receive |
| 94 |
C elType - Datat type of elements |
| 95 |
INTEGER tagSend |
| 96 |
INTEGER tagRecv |
| 97 |
INTEGER nReqs |
| 98 |
INTEGER reqIds(4) |
| 99 |
INTEGER mpiRC |
| 100 |
INTEGER mpiStatArr(MPI_STATUS_SIZE,4) |
| 101 |
INTEGER toPid |
| 102 |
INTEGER fromPid |
| 103 |
INTEGER elCount |
| 104 |
INTEGER elType |
| 105 |
#endif /* ALLOW_USE_MPI */ |
| 106 |
C bE, bW - Block index to east, west, north and |
| 107 |
C bN, bS south. |
| 108 |
INTEGER bW, bE, bS, bN |
| 109 |
|
| 110 |
C-- Do "message-passing" data exchanges |
| 111 |
#ifdef ALLOW_SYNC_COMMUNICATION |
| 112 |
#ifndef ALWAYS_USE_SYNC_COMMUNICATION |
| 113 |
IF ( usingSyncMessages ) THEN |
| 114 |
#endif |
| 115 |
|
| 116 |
C-- MPI based data exchages |
| 117 |
#ifdef ALLOW_USE_MPI |
| 118 |
#ifndef ALWAYS_USE_MPI |
| 119 |
IF ( usingMPI ) THEN |
| 120 |
#endif |
| 121 |
|
| 122 |
C Here use the approach that all the X shifts are completed |
| 123 |
C before the Y shifts start. Then we can expand our Y shifts |
| 124 |
C to include X overlap regions. This means we don't have to |
| 125 |
C transfer corners separately. |
| 126 |
C Notes: |
| 127 |
C ====== |
| 128 |
C 1. If we are only using messages in Y and using "true shared-memory" |
| 129 |
C in X then we rely on the shared-memory code to be implemented so that |
| 130 |
C corner values will be set correctly. The true shared-memory code |
| 131 |
C is in the block below - so be careful when altering it! |
| 132 |
C 2. In order to ensure we grab the right message we need to tag |
| 133 |
C messages. We tag the messages with the direction in which they |
| 134 |
C were sent out. Thus to receive from a processor to our west we |
| 135 |
C request the next message from that processor that was sent out |
| 136 |
C with a tag indicating an easterly send. |
| 137 |
C 3. We could implement a "safe" code here by having everyone |
| 138 |
C sync and then getting a particular thread ( say thread 1 ) to |
| 139 |
C do the MPI_Isend sequentially. In this mode the "communication" |
| 140 |
C thread would loop over nThreads and do MPI_Sendrecv for each |
| 141 |
C one. This could be more efficient on some platforms as the messages |
| 142 |
C might be sent as one large unit rather than by separate threads. |
| 143 |
C It would also be "thread-safe" for MPI implementations that have |
| 144 |
C multi-threading problems with things like request ids. |
| 145 |
C 4. We include overlap regions in both X and Y sends. |
| 146 |
C This means we can interchange the Y block with the X block. |
| 147 |
C Not sure whether this will ever be useful! |
| 148 |
C 5. The generation of a request handle by MPI_Isend and |
| 149 |
C MPI_Irecv ouught to involve some global state within MPI. |
| 150 |
C If this process is not thread safe it may be necessary to |
| 151 |
C enable a critical section around the MPI_Isend and |
| 152 |
C MPI_Irecv calls. |
| 153 |
|
| 154 |
C We need a barrier here to synchronise threads otherwise |
| 155 |
C one thread might declare it has is ready to send and receive |
| 156 |
C even though another thread is going to write to the first |
| 157 |
C threads send or receive region. This won't happen, however, |
| 158 |
C if we are careful about which data a thread updates - but at |
| 159 |
C the moment we aren't careful! |
| 160 |
_BARRIER |
| 161 |
|
| 162 |
C-- East-west messaging communication |
| 163 |
nReqs = 0 |
| 164 |
bE = myBxLo(myThid) |
| 165 |
IF ( bE .EQ. 1 .AND. commW(myThid) .EQ. COMM_MPI ) THEN |
| 166 |
C My west face uses MPI. Get ready to receive data from |
| 167 |
C west and post that we are ready to send data to the west. |
| 168 |
C Data we receive has to come with tag of thread we know is |
| 169 |
C sending from our west, data we send is tagged with our thread id |
| 170 |
C and the send direction. |
| 171 |
nReqs = nReqs+1 |
| 172 |
tagSend = mpiTagW*nThreads+myThid |
| 173 |
toPid = mpiPidW |
| 174 |
elCount = 1 |
| 175 |
elType = mpiTypeXFaceThread_xy_r4(myThid) |
| 176 |
C I tagSend,MPI_COMM_WORLD, |
| 177 |
CALL MPI_Isend( |
| 178 |
I phi(1,1-OLy,bE,myByLo(myThid)), |
| 179 |
I elCount, elType, toPid, |
| 180 |
I tagSend,MPI_COMM_WORLD, |
| 181 |
I reqIds(nReqs),mpiRC) |
| 182 |
nReqs = nReqs+1 |
| 183 |
tagRecv = mpiTagE*nThreads+myThrW(myThid) |
| 184 |
fromPid = mpiPidW |
| 185 |
elCount = 1 |
| 186 |
elType = mpiTypeXFaceThread_xy_r4(myThid) |
| 187 |
CALL MPI_Irecv( |
| 188 |
U phi(1-OLx,1-OLy,bE,myByLo(myThid)), |
| 189 |
I elCount, elType, fromPid, |
| 190 |
I tagRecv,MPI_COMM_WORLD, |
| 191 |
I reqIds(nReqs),mpiRC) |
| 192 |
ENDIF |
| 193 |
bW = myBxHi(myThid) |
| 194 |
IF ( bW .EQ. nSx .AND. commE(MyThid) .EQ. COMM_MPI ) THEN |
| 195 |
C My east face uses MPI. Get ready to receive data from |
| 196 |
C east and post that we are ready to send data to the east. |
| 197 |
C Data we receive has to come with tag of thread we know is |
| 198 |
C sending from our west, data we send is tagged with our thread id. |
| 199 |
nReqs = nReqs+1 |
| 200 |
tagSend = mpiTagE*nThreads+myThid |
| 201 |
toPid = mpiPidE |
| 202 |
elCount = 1 |
| 203 |
elType = mpiTypeXFaceThread_xy_r4(myThid) |
| 204 |
CALL MPI_Isend( |
| 205 |
I phi(sNx-OLx+1,1-OLy,bW,myByLo(myThid)), |
| 206 |
I elCount, elType, toPid, |
| 207 |
I tagSend,MPI_COMM_WORLD, |
| 208 |
I reqIds(nReqs),mpiRC) |
| 209 |
nReqs = nReqs+1 |
| 210 |
tagRecv = mpiTagW*nThreads+myThrE(myThid) |
| 211 |
fromPid = mpiPidE |
| 212 |
elCount = 1 |
| 213 |
elType = mpiTypeXFaceThread_xy_r4(myThid) |
| 214 |
CALL MPI_Irecv( |
| 215 |
U phi(sNx+1,1-OLy,bW,myByLo(myThid)), |
| 216 |
I elCount, elType, fromPid, |
| 217 |
I tagRecv,MPI_COMM_WORLD, |
| 218 |
I reqIds(nReqs),mpiRC) |
| 219 |
ENDIF |
| 220 |
C Wait for this threads east-west transactions to finish before |
| 221 |
C posting north-south transactions. We have to do this so that |
| 222 |
C the north-south transactions will send out the correct overlap |
| 223 |
C region values into the corner sections of our neighbors. |
| 224 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 225 |
|
| 226 |
CcnhDebugStarts |
| 227 |
C RETURN |
| 228 |
CcnhDebugEnds |
| 229 |
|
| 230 |
C-- North-south messaging communication |
| 231 |
nReqs = 0 |
| 232 |
bN = myByLo(myTHid) |
| 233 |
IF ( bN .EQ. 1 .AND. commS(myThid) .EQ. COMM_MPI ) THEN |
| 234 |
C My south face uses MPI. Get ready to receive data from |
| 235 |
C south and post that I am ready to send data to the south. |
| 236 |
nReqs = nReqs + 1 |
| 237 |
tagSend = mpiTagS*nThreads+myThid |
| 238 |
toPid = mpiPidS |
| 239 |
elCount = 1 |
| 240 |
elType = mpiTypeYFaceThread_xy_r4(myThid) |
| 241 |
CALL MPI_Isend( |
| 242 |
I phi(1-OLx,1,myBxLo(myThid),bN), |
| 243 |
I elCount, elType, toPid, |
| 244 |
I tagSend,MPI_COMM_WORLD, |
| 245 |
I reqIds(nReqs), mpiRC ) |
| 246 |
nReqs = nReqs + 1 |
| 247 |
tagRecv = mpiTagN*nThreads+myThrS(myThid) |
| 248 |
fromPid = mpiPidS |
| 249 |
elCount = 1 |
| 250 |
elType = mpiTypeYFaceThread_xy_r4(myThid) |
| 251 |
CALL MPI_Irecv( |
| 252 |
I phi(1-OLx,1-OLy,myBxLo(myThid),bN), |
| 253 |
I elCount, elType, fromPid, |
| 254 |
I tagRecv,MPI_COMM_WORLD, |
| 255 |
I reqIds(nReqs), mpiRC ) |
| 256 |
ENDIF |
| 257 |
C |
| 258 |
bS = myByHi(myTHid) |
| 259 |
IF ( bS .EQ. nSy .AND. commN(myThid) .EQ. COMM_MPI ) THEN |
| 260 |
C My north face uses MPI. Get ready to receive data from |
| 261 |
C north and post that I am ready to send data to the north. |
| 262 |
nReqs = nReqs + 1 |
| 263 |
tagSend = mpiTagN*nThreads+myThid |
| 264 |
toPid = mpiPidN |
| 265 |
elCount = 1 |
| 266 |
elType = mpiTypeYFaceThread_xy_r4(myThid) |
| 267 |
CALL MPI_Isend( |
| 268 |
I phi(1-OLx,sNy-OLy+1,myBxLo(myThid),bS), |
| 269 |
I elCount, elType, toPid, |
| 270 |
I tagSend,MPI_COMM_WORLD, |
| 271 |
I reqIds(nReqs), mpiRC ) |
| 272 |
nReqs = nReqs + 1 |
| 273 |
tagRecv = mpiTagS*nThreads+myThrN(myThid) |
| 274 |
fromPid = mpiPidN |
| 275 |
elCount = 1 |
| 276 |
elType = mpiTypeYFaceThread_xy_r4(myThid) |
| 277 |
CALL MPI_Irecv( |
| 278 |
I phi(1-OLx,sNy+1,myBxLo(myThid),bS), |
| 279 |
I elCount, elType, fromPid, |
| 280 |
I tagRecv,MPI_COMM_WORLD, |
| 281 |
I reqIds(nReqs), mpiRC ) |
| 282 |
ENDIF |
| 283 |
C Wait for this threads north-south transactions to finish. |
| 284 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 285 |
|
| 286 |
#ifndef ALWAYS_USE_MPI |
| 287 |
ENDIF |
| 288 |
#endif |
| 289 |
#endif /* ALLOW_USE_MPI */ |
| 290 |
|
| 291 |
#ifndef ALWAYS_USE_SYNC_COMMUNICATION |
| 292 |
ENDIF |
| 293 |
#endif |
| 294 |
#endif /* ALLOW_SYNC_COMMUNICATION */ |
| 295 |
|
| 296 |
C-- Do true shared-memory data exchanges |
| 297 |
C Note this is also doing the edge copies for a regular serial |
| 298 |
C code. |
| 299 |
C-- First make sure all threads have reached here |
| 300 |
_BARRIER |
| 301 |
|
| 302 |
C-- Now do copies |
| 303 |
C Notes: |
| 304 |
C ====== |
| 305 |
C Here we copy from up to and including the overlap |
| 306 |
C regions for both the X shift and the Y shift. This |
| 307 |
C catches the "corners" of the data copied using messages |
| 308 |
C in the "synchronous communication" section above. |
| 309 |
C As coded here we also write to a remote region in |
| 310 |
C one direction. In some situations it may be better only |
| 311 |
C to write to our own overlaps - could be the case with |
| 312 |
C Wildfire replication. This should be chamged once everything |
| 313 |
C is working OK. |
| 314 |
C |
| 315 |
C-- x-axis exchanges |
| 316 |
bE = myBxLo(myThid) |
| 317 |
IF ( bE .NE. 1 .OR. |
| 318 |
& commW(myThid) .EQ. COMM_SHARED ) THEN |
| 319 |
bW = bE-1 |
| 320 |
IF ( bW .LT. 1 ) bW = nSx |
| 321 |
DO bj=myByLo(myThid),myByHi(myThid) |
| 322 |
DO J=1-OLy,sNy+OLy |
| 323 |
DO I=1,OLx |
| 324 |
phi(sNx+I,J,bW,bj)=phi( 1+I-1,J,bE,bj) |
| 325 |
phi(1-I ,J,bE,bj)=phi(sNx-I+1,J,bW,bj) |
| 326 |
ENDDO |
| 327 |
ENDDO |
| 328 |
ENDDO |
| 329 |
ENDIF |
| 330 |
DO bE=myBxLo(myThid)+1,myBxHi(myThid) |
| 331 |
bW = bE-1 |
| 332 |
DO bj=myByLo(myThid),myByHi(myThid) |
| 333 |
DO J=1-OLy,sNy+OLy |
| 334 |
DO I=1,OLx |
| 335 |
phi(sNx+I,J,bW,bj)=phi( 1+I-1,J,bE,bj) |
| 336 |
phi(1-I ,J,bE,bj)=phi(sNx-I+1,J,bW,bj) |
| 337 |
ENDDO |
| 338 |
ENDDO |
| 339 |
ENDDO |
| 340 |
ENDDO |
| 341 |
|
| 342 |
C Need to ensure all threads have completed x-axis transfers before |
| 343 |
C we can do y-axis exchanges. |
| 344 |
_BARRIER |
| 345 |
|
| 346 |
C-- y-axis exchanges |
| 347 |
bN = myByLo(myThid) |
| 348 |
IF ( bN .NE. 1 .OR. |
| 349 |
& commS(myThid) .EQ. COMM_SHARED ) THEN |
| 350 |
bS = bN - 1 |
| 351 |
IF ( bS .LT. 1 ) bS = nSy |
| 352 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
| 353 |
DO J=1,OLy |
| 354 |
DO I=1-OLx,sNx+OLx |
| 355 |
phi(I,1-J ,bi,bN)=phi(I,sNy-J+1,bi,bS) |
| 356 |
phi(I,sNy+J ,bi,bS)=phi(I,1+J-1 ,bi,bN) |
| 357 |
ENDDO |
| 358 |
ENDDO |
| 359 |
ENDDO |
| 360 |
ENDIF |
| 361 |
DO bN=myByLo(myThid)+1,myByHi(myThid) |
| 362 |
bS = bN - 1 |
| 363 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
| 364 |
DO J=1,OLy |
| 365 |
DO I=1-OLx,sNx+OLx |
| 366 |
phi(I,1-J ,bi,bN)=phi(I,sNy-J+1,bi,bS) |
| 367 |
phi(I,sNy+J ,bi,bS)=phi(I,1+J-1 ,bi,bN) |
| 368 |
ENDDO |
| 369 |
ENDDO |
| 370 |
ENDDO |
| 371 |
ENDDO |
| 372 |
|
| 373 |
_BARRIER |
| 374 |
|
| 375 |
RETURN |
| 376 |
END |
| 377 |
|
| 378 |
CStartOfInterface |
| 379 |
SUBROUTINE EXCH_XYZ_R4( |
| 380 |
U phi, |
| 381 |
I myThid ) |
| 382 |
C /==========================================================\ |
| 383 |
C | SUBROUTINE EXCH_XYZ_R4 | |
| 384 |
C | o Handle exchanges for real*4, three-dimensional arrays. | |
| 385 |
C |==========================================================| |
| 386 |
C | Do true shared-memory data transfers and "messaging" | |
| 387 |
C | tranfers for blocking case of data transfers. | |
| 388 |
C | Applications call this routine using | |
| 389 |
C | CALL EXCH..( x, myThid ) | |
| 390 |
C | where x is a three-dimensional array with overlaps. | |
| 391 |
C | This routine does true-shared-memory copies for blocks | |
| 392 |
C | within a thread. It will also do MPI meesaging between | |
| 393 |
C | different processes. | |
| 394 |
C | Note: | |
| 395 |
C | ===== | |
| 396 |
C | If it is used, "asynchronous" messaging in which | |
| 397 |
C | communication overlaps computation is handled elsewhere | |
| 398 |
C | - see recv.F and send.F. | |
| 399 |
C | In practice MPI implementations may not be completely | |
| 400 |
C | thread safe. In principle this code is correct even for | |
| 401 |
C | mixed MPI and multi-threaded execution. | |
| 402 |
C | In multi-thread execution data managed by other threads | |
| 403 |
C | is only automatically visible to another thread if it | |
| 404 |
C | existed before the thread was created. This means that | |
| 405 |
C | for Fortran we declare arrays with overlap regions in | |
| 406 |
C | COMMON blocks. | |
| 407 |
C \==========================================================/ |
| 408 |
|
| 409 |
C === Global data === |
| 410 |
#include "SIZE.h" |
| 411 |
#include "EEPARAMS.h" |
| 412 |
#include "EESUPPORT.h" |
| 413 |
|
| 414 |
C === Routine arguments === |
| 415 |
C phi - Array who's overlap regions are to be exchanged |
| 416 |
C myThid - My thread id. |
| 417 |
Real*4 phi(1-OLx:sNx+OLx,1-OLy:sNy+OLy,1:Nz,nSx,nSy) |
| 418 |
INTEGER myThid |
| 419 |
CEndOfInterface |
| 420 |
|
| 421 |
C === Local variables === |
| 422 |
C bi,bj - Outer loop counters |
| 423 |
C I,J,K - Inner loop counters |
| 424 |
INTEGER bi, bj |
| 425 |
INTEGER I, J, K |
| 426 |
#ifdef ALLOW_USE_MPI |
| 427 |
C tagSend - Tags used to mark and select messages |
| 428 |
C tagRecv |
| 429 |
C nReqs - MPI request counter. |
| 430 |
C reqIds - Outstanding requests to wait on. |
| 431 |
C mpiRC - MPI return code |
| 432 |
C mpiStatArr - Multi-request status reporting array. |
| 433 |
C toPid - Proc. id sending to |
| 434 |
C fromPid - Proc. id receving from |
| 435 |
C elCount - Number of items to send or receive |
| 436 |
C elType - Data type of elements |
| 437 |
INTEGER tagSend |
| 438 |
INTEGER tagRecv |
| 439 |
INTEGER nReqs |
| 440 |
INTEGER reqIds(4) |
| 441 |
INTEGER mpiRC |
| 442 |
INTEGER mpiStatArr(MPI_STATUS_SIZE,4) |
| 443 |
INTEGER toPid |
| 444 |
INTEGER fromPid |
| 445 |
INTEGER elCount |
| 446 |
INTEGER elType |
| 447 |
#endif /* ALLOW_USE_MPI */ |
| 448 |
C bE, bW - Block index to east, west, north and |
| 449 |
C bN, bS south. |
| 450 |
INTEGER bW, bE, bS, bN |
| 451 |
|
| 452 |
C-- Do "message-passing" data exchanges |
| 453 |
#ifdef ALLOW_SYNC_COMMUNICATION |
| 454 |
#ifndef ALWAYS_USE_SYNC_COMMUNICATION |
| 455 |
IF ( usingSyncMessages ) THEN |
| 456 |
#endif |
| 457 |
|
| 458 |
C-- MPI based data exchages |
| 459 |
#ifdef ALLOW_USE_MPI |
| 460 |
#ifndef ALWAYS_USE_MPI |
| 461 |
IF ( usingMPI ) THEN |
| 462 |
#endif |
| 463 |
|
| 464 |
C Here use the approach that all the X shifts are completed |
| 465 |
C before the Y shifts start. Then we can expand our Y shifts |
| 466 |
C to include X overlap regions. This means we don't have to |
| 467 |
C transfer corners separately. |
| 468 |
C Notes: |
| 469 |
C ====== |
| 470 |
C 1. If we are only using messages in Y and using "true shared-memory" |
| 471 |
C in X then we rely on the shared-memory code to be implemented so that |
| 472 |
C corner values will be set correctly. The true shared-memory code |
| 473 |
C is in the block below - so be careful when altering it! |
| 474 |
C 2. In order to ensure we grab the right message we need to tag |
| 475 |
C messages. We tag the messages with the direction in which they |
| 476 |
C were sent out. Thus to receive from a processor to our west we |
| 477 |
C request the next message from that processor that was sent out |
| 478 |
C with a tag indicating an easterly send. |
| 479 |
C 3. We could implement a "safe" code here by having everyone |
| 480 |
C sync and then getting a particular thread ( say thread 1 ) to |
| 481 |
C do the MPI_Isend sequentially. In this mode the "communication" |
| 482 |
C thread would loop over nThreads and do MPI_Sendrecv for each |
| 483 |
C one. This could be more efficient on some platforms as the messages |
| 484 |
C might be sent as one large unit rather than by separate threads. |
| 485 |
C It would also be "thread-safe" for MPI implementations that have |
| 486 |
C multi-threading problems with things like request ids. |
| 487 |
C 4. We include overlap regions in both X and Y sends. |
| 488 |
C This means we can interchange the Y block with the X block. |
| 489 |
C Not sure whether this will ever be useful! |
| 490 |
C 5. The generation of a request handle by MPI_Isend and |
| 491 |
C MPI_Irecv ouught to involve some global state within MPI. |
| 492 |
C If this process is not thread safe it may be necessary to |
| 493 |
C enable a critical section around the MPI_Isend and |
| 494 |
C MPI_Irecv calls. |
| 495 |
|
| 496 |
C We need a barrier here to synchronise threads otherwise |
| 497 |
C one thread might declare it has is ready to send and receive |
| 498 |
C even though another thread is going to write to the first |
| 499 |
C threads send or receive region. This won't happen, however, |
| 500 |
C if we are careful about which data a thread updates - but at |
| 501 |
C the moment we aren't careful! |
| 502 |
_BARRIER |
| 503 |
|
| 504 |
C-- East-west messaging communication |
| 505 |
nReqs = 0 |
| 506 |
bE = myBxLo(myThid) |
| 507 |
IF ( bE .EQ. 1 .AND. commW(myThid) .EQ. COMM_MPI ) THEN |
| 508 |
C My west face uses MPI. Get ready to receive data from |
| 509 |
C west and post that we are ready to send data to the west. |
| 510 |
C Data we receive has to come with tag of thread we know is |
| 511 |
C sending from our west, data we send is tagged with our thread id |
| 512 |
C and the send direction. |
| 513 |
nReqs = nReqs+1 |
| 514 |
tagSend = mpiTagW*nThreads+myThid |
| 515 |
toPid = mpiPidW |
| 516 |
elCount = 1 |
| 517 |
elType = mpiTypeXFaceThread_xyz_r4(myThid) |
| 518 |
CALL MPI_Isend( |
| 519 |
I phi(1,1-OLy,1,bE,myByLo(myThid)), |
| 520 |
I elCount, elType, toPid, |
| 521 |
I tagSend,MPI_COMM_WORLD, |
| 522 |
I reqIds(nReqs),mpiRC) |
| 523 |
nReqs = nReqs+1 |
| 524 |
tagRecv = mpiTagE*nThreads+myThrW(myThid) |
| 525 |
fromPid = mpiPidW |
| 526 |
elCount = 1 |
| 527 |
elType = mpiTypeXFaceThread_xyz_r4(myThid) |
| 528 |
CALL MPI_Irecv( |
| 529 |
U phi(1-OLx,1-OLy,1,bE,myByLo(myThid)), |
| 530 |
I elCount, elType, fromPid, |
| 531 |
I tagRecv,MPI_COMM_WORLD, |
| 532 |
I reqIds(nReqs),mpiRC) |
| 533 |
ENDIF |
| 534 |
bW = myBxHi(myThid) |
| 535 |
IF ( bW .EQ. nSx .AND. commE(MyThid) .EQ. COMM_MPI ) THEN |
| 536 |
C My east face uses MPI. Get ready to receive data from |
| 537 |
C east and post that we are ready to send data to the east. |
| 538 |
C Data we receive has to come with tag of thread we know is |
| 539 |
C sending from our west, data we send is tagged with our thread id. |
| 540 |
nReqs = nReqs+1 |
| 541 |
tagSend = mpiTagE*nThreads+myThid |
| 542 |
toPid = mpiPidE |
| 543 |
elCount = 1 |
| 544 |
elType = mpiTypeXFaceThread_xyz_r4(myThid) |
| 545 |
CALL MPI_Isend( |
| 546 |
I phi(sNx-OLx+1,1-OLy,1,bW,myByLo(myThid)), |
| 547 |
I elCount, elType, toPid, |
| 548 |
I tagSend,MPI_COMM_WORLD, |
| 549 |
I reqIds(nReqs),mpiRC) |
| 550 |
nReqs = nReqs+1 |
| 551 |
tagRecv = mpiTagW*nThreads+myThrE(myThid) |
| 552 |
fromPid = mpiPidE |
| 553 |
elCount = 1 |
| 554 |
elType = mpiTypeXFaceThread_xyz_r4(myThid) |
| 555 |
CALL MPI_Irecv( |
| 556 |
U phi(sNx+1,1-OLy,1,bW,myByLo(myThid)), |
| 557 |
I elCount, elType, fromPid, |
| 558 |
I tagRecv,MPI_COMM_WORLD, |
| 559 |
I reqIds(nReqs),mpiRC) |
| 560 |
ENDIF |
| 561 |
C Wait for this threads east-west transactions to finish before |
| 562 |
C posting north-south transactions. We have to do this so that |
| 563 |
C the north-south transactions will send out the correct overlap |
| 564 |
C region values into the corner sections of our neighbors. |
| 565 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 566 |
CcnhDebugStarts |
| 567 |
C RETURN |
| 568 |
CcnhDebugEnds |
| 569 |
|
| 570 |
|
| 571 |
C-- North-south messaging communication |
| 572 |
nReqs = 0 |
| 573 |
bN = myByLo(myTHid) |
| 574 |
IF ( bN .EQ. 1 .AND. commS(myThid) .EQ. COMM_MPI ) THEN |
| 575 |
C My south face uses MPI. Get ready to receive data from |
| 576 |
C south and post that I am ready to send data to the south. |
| 577 |
nReqs = nReqs + 1 |
| 578 |
tagSend = mpiTagS*nThreads+myThid |
| 579 |
toPid = mpiPidS |
| 580 |
elCount = 1 |
| 581 |
elType = mpiTypeYFaceThread_xyz_r4(myThid) |
| 582 |
CALL MPI_Isend( |
| 583 |
I phi(1-OLx,1,1,myBxLo(myThid),bN), |
| 584 |
I elCount, elType, toPid, |
| 585 |
I tagSend,MPI_COMM_WORLD, |
| 586 |
I reqIds(nReqs), mpiRC ) |
| 587 |
nReqs = nReqs + 1 |
| 588 |
tagRecv = mpiTagN*nThreads+myThrS(myThid) |
| 589 |
fromPid = mpiPidS |
| 590 |
elCount = 1 |
| 591 |
elType = mpiTypeYFaceThread_xyz_r4(myThid) |
| 592 |
CALL MPI_Irecv( |
| 593 |
I phi(1-OLx,1-OLy,1,myBxLo(myThid),bN), |
| 594 |
I elCount, elType, fromPid, |
| 595 |
I tagRecv,MPI_COMM_WORLD, |
| 596 |
I reqIds(nReqs), mpiRC ) |
| 597 |
ENDIF |
| 598 |
C |
| 599 |
bS = myByHi(myTHid) |
| 600 |
IF ( bN .EQ. 1 .AND. commS(myThid) .EQ. COMM_MPI ) THEN |
| 601 |
C My north face uses MPI. Get ready to receive data from |
| 602 |
C north and post that I am ready to send data to the north. |
| 603 |
nReqs = nReqs + 1 |
| 604 |
tagSend = mpiTagN*nThreads+myThid |
| 605 |
toPid = mpiPidN |
| 606 |
elCount = 1 |
| 607 |
elType = mpiTypeYFaceThread_xyz_r4(myThid) |
| 608 |
CALL MPI_Isend( |
| 609 |
I phi(1-OLx,sNy-OLy+1,1,myBxLo(myThid),bS), |
| 610 |
I elCount, elType, toPid, |
| 611 |
I tagSend,MPI_COMM_WORLD, |
| 612 |
I reqIds(nReqs), mpiRC ) |
| 613 |
nReqs = nReqs + 1 |
| 614 |
tagRecv = mpiTagS*nThreads+myThrN(myThid) |
| 615 |
fromPid = mpiPidN |
| 616 |
elCount = 1 |
| 617 |
elType = mpiTypeYFaceThread_xyz_r4(myThid) |
| 618 |
CALL MPI_Irecv( |
| 619 |
I phi(1-OLx,sNy+1,1,myBxLo(myThid),bS), |
| 620 |
I elCount, elType, fromPid, |
| 621 |
I tagRecv,MPI_COMM_WORLD, |
| 622 |
I reqIds(nReqs), mpiRC ) |
| 623 |
ENDIF |
| 624 |
C Wait for this threads north-south transactions to finish. |
| 625 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 626 |
|
| 627 |
#ifndef ALWAYS_USE_MPI |
| 628 |
ENDIF |
| 629 |
#endif |
| 630 |
#endif /* ALLOW_USE_MPI */ |
| 631 |
|
| 632 |
#ifndef ALWAYS_USE_SYNC_COMMUNICATION |
| 633 |
ENDIF |
| 634 |
#endif |
| 635 |
#endif /* ALLOW_SYNC_COMMUNICATION */ |
| 636 |
|
| 637 |
C-- Do true shared-memory data exchanges |
| 638 |
C Note: This is also doing the overlap copies |
| 639 |
C for a single threaded code. |
| 640 |
C-- First make sure all threads have reached here |
| 641 |
_BARRIER |
| 642 |
|
| 643 |
C-- Now do copies |
| 644 |
C Notes: |
| 645 |
C ====== |
| 646 |
C Here we copy from up to and including the overlap |
| 647 |
C regions for both the X shift and the Y shift. This |
| 648 |
C catches the "corners" of the data copied using messages |
| 649 |
C in the "synchronous communication" section above. |
| 650 |
C-- x-axis exchanges |
| 651 |
bE = myBxLo(myThid) |
| 652 |
IF ( bE .NE. 1 .OR. |
| 653 |
& commW(myThid) .EQ. COMM_SHARED ) THEN |
| 654 |
bW = bE-1 |
| 655 |
IF ( bW .LT. 1 ) bW = nSx |
| 656 |
DO bj=myByLo(myThid),myByHi(myThid) |
| 657 |
DO K=1,Nz |
| 658 |
DO J=1-OLy,sNy+OLy |
| 659 |
DO I=1,OLx |
| 660 |
phi(sNx+I,J,K,bW,bj)=phi( 1+I-1,J,K,bE,bj) |
| 661 |
phi(1-I ,J,K,bE,bj)=phi(sNx-I+1,J,K,bW,bj) |
| 662 |
ENDDO |
| 663 |
ENDDO |
| 664 |
ENDDO |
| 665 |
ENDDO |
| 666 |
ENDIF |
| 667 |
DO bE=myBxLo(myThid)+1,myBxHi(myThid) |
| 668 |
bW = bE-1 |
| 669 |
DO bj=myByLo(myThid),myByHi(myThid) |
| 670 |
DO K=1,Nz |
| 671 |
DO J=1-OLy,sNy+OLy |
| 672 |
DO I=1,OLx |
| 673 |
phi(sNx+I,J,K,bW,bj)=phi( 1+I-1,J,K,bE,bj) |
| 674 |
phi(1-I ,J,K,bE,bj)=phi(sNx-I+1,J,K,bW,bj) |
| 675 |
ENDDO |
| 676 |
ENDDO |
| 677 |
ENDDO |
| 678 |
ENDDO |
| 679 |
ENDDO |
| 680 |
|
| 681 |
C Need to ensure all threads have completed x-axis transfers before |
| 682 |
C we can do y-axis exchanges. |
| 683 |
_BARRIER |
| 684 |
|
| 685 |
C-- y-axis exchanges |
| 686 |
bN = myByLo(myThid) |
| 687 |
IF ( bN .NE. 1 .OR. |
| 688 |
& commS(myThid) .EQ. COMM_SHARED ) THEN |
| 689 |
bS = bN - 1 |
| 690 |
IF ( bS .LT. 1 ) bS = nSy |
| 691 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
| 692 |
DO K=1,Nz |
| 693 |
DO J=1,OLy |
| 694 |
DO I=1-OLx,sNx+OLx |
| 695 |
phi(I,1-J ,K,bi,bN)=phi(I,sNy-J+1,K,bi,bS) |
| 696 |
phi(I,sNy+J ,K,bi,bS)=phi(I,1+J-1 ,K,bi,bN) |
| 697 |
ENDDO |
| 698 |
ENDDO |
| 699 |
ENDDO |
| 700 |
ENDDO |
| 701 |
ENDIF |
| 702 |
DO bN=myByLo(myThid)+1,myByHi(myThid) |
| 703 |
bS = bN - 1 |
| 704 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
| 705 |
DO K=1,Nz |
| 706 |
DO J=1,OLy |
| 707 |
DO I=1-OLx,sNx+OLx |
| 708 |
phi(I,1-J ,K,bi,bN)=phi(I,sNy-J+1,K,bi,bS) |
| 709 |
phi(I,sNy+J ,K,bi,bS)=phi(I,1+J-1 ,K,bi,bN) |
| 710 |
ENDDO |
| 711 |
ENDDO |
| 712 |
ENDDO |
| 713 |
ENDDO |
| 714 |
ENDDO |
| 715 |
|
| 716 |
_BARRIER |
| 717 |
|
| 718 |
RETURN |
| 719 |
END |
| 720 |
|
| 721 |
CStartOfInterface |
| 722 |
SUBROUTINE EXCH_XY_R8( |
| 723 |
U phi, |
| 724 |
I myThid ) |
| 725 |
C /==========================================================\ |
| 726 |
C | SUBROUTINE EXCH_XY_R8 | |
| 727 |
C | o Handle exchanges for real*8, two-dimensional arrays. | |
| 728 |
C |==========================================================| |
| 729 |
C | Do true shared-memory data transfers and "messaging" | |
| 730 |
C | tranfers for blocking case of data transfers. | |
| 731 |
C | Applications call this routine using | |
| 732 |
C | CALL EXCH..( x, myThid ) | |
| 733 |
C | where x is a two-dimensional array with overlaps. | |
| 734 |
C | This routine does true-shared-memory copies for blocks | |
| 735 |
C | within a thread. It will also do MPI meesaging between | |
| 736 |
C | different processes. | |
| 737 |
C | Note: | |
| 738 |
C | ===== | |
| 739 |
C | If it is used, "asynchronous" messaging in which | |
| 740 |
C | communication overlaps computation is handled elsewhere | |
| 741 |
C | - see recv.F and send.F. | |
| 742 |
C | In practice MPI implementations may not be completely | |
| 743 |
C | thread safe. In principle this code is correct even for | |
| 744 |
C | mixed MPI and multi-threaded execution. | |
| 745 |
C | In multi-thread execution data managed by other threads | |
| 746 |
C | is only automatically visible to another thread if it | |
| 747 |
C | existed before the thread was created. This means that | |
| 748 |
C | all Fortran we declare arrays with overlap regions in | |
| 749 |
C | COMMON blocks. | |
| 750 |
C \==========================================================/ |
| 751 |
|
| 752 |
C === Global data === |
| 753 |
#include "SIZE.h" |
| 754 |
#include "EEPARAMS.h" |
| 755 |
#include "EESUPPORT.h" |
| 756 |
|
| 757 |
C === Routine arguments === |
| 758 |
C phi - Array who's overlap regions are to be exchanged |
| 759 |
C myThid - My thread id. |
| 760 |
Real*8 phi(1-OLx:sNx+OLx,1-OLy:sNy+OLy,nSx,nSy) |
| 761 |
INTEGER myThid |
| 762 |
CEndOfInterface |
| 763 |
|
| 764 |
C === Local variables === |
| 765 |
C bi,bj - Outer loop counters |
| 766 |
C I,J,K - Inner loop counters |
| 767 |
INTEGER bi, bj |
| 768 |
INTEGER I, J |
| 769 |
#ifdef ALLOW_USE_MPI |
| 770 |
C tagSend - Tags used to mark and select messages |
| 771 |
C tagRecv |
| 772 |
C nReqs - MPI request counter. |
| 773 |
C reqIds - Outstanding requests to wait on. |
| 774 |
C mpiRC - MPI return code |
| 775 |
C mpiStatArr - Multi-request status reporting array. |
| 776 |
C toPid - Proc. id sending to |
| 777 |
C fromPid - Proc. id receving from |
| 778 |
C elCount - Number of items to send or receive |
| 779 |
C elType - Datat type of elements |
| 780 |
INTEGER tagSend |
| 781 |
INTEGER tagRecv |
| 782 |
INTEGER nReqs |
| 783 |
INTEGER reqIds(4) |
| 784 |
INTEGER mpiRC |
| 785 |
INTEGER mpiStatArr(MPI_STATUS_SIZE,4) |
| 786 |
INTEGER toPid |
| 787 |
INTEGER fromPid |
| 788 |
INTEGER elCount |
| 789 |
INTEGER elType |
| 790 |
#endif /* ALLOW_USE_MPI */ |
| 791 |
C bE, bW - Block index to east, west, north and |
| 792 |
C bN, bS south. |
| 793 |
INTEGER bW, bE, bS, bN |
| 794 |
|
| 795 |
C-- Do "message-passing" data exchanges |
| 796 |
#ifdef ALLOW_SYNC_COMMUNICATION |
| 797 |
#ifndef ALWAYS_USE_SYNC_COMMUNICATION |
| 798 |
IF ( usingSyncMessages ) THEN |
| 799 |
#endif |
| 800 |
|
| 801 |
C-- MPI based data exchages |
| 802 |
#ifdef ALLOW_USE_MPI |
| 803 |
#ifndef ALWAYS_USE_MPI |
| 804 |
IF ( usingMPI ) THEN |
| 805 |
#endif |
| 806 |
|
| 807 |
C Here use the approach that all the X shifts are completed |
| 808 |
C before the Y shifts start. Then we can expand our Y shifts |
| 809 |
C to include X overlap regions. This means we don't have to |
| 810 |
C transfer corners separately. |
| 811 |
C Notes: |
| 812 |
C ====== |
| 813 |
C 1. If we are only using messages in Y and using "true shared-memory" |
| 814 |
C in X then we rely on the shared-memory code to be implemented so that |
| 815 |
C corner values will be set correctly. The true shared-memory code |
| 816 |
C is in the block below - so be careful when altering it! |
| 817 |
C 2. In order to ensure we grab the right message we need to tag |
| 818 |
C messages. We tag the messages with the direction in which they |
| 819 |
C were sent out. Thus to receive from a processor to our west we |
| 820 |
C request the next message from that processor that was sent out |
| 821 |
C with a tag indicating an easterly send. |
| 822 |
C 3. We could implement a "safe" code here by having everyone |
| 823 |
C sync and then getting a particular thread ( say thread 1 ) to |
| 824 |
C do the MPI_Isend sequentially. In this mode the "communication" |
| 825 |
C thread would loop over nThreads and do MPI_Sendrecv for each |
| 826 |
C one. This could be more efficient on some platforms as the messages |
| 827 |
C might be sent as one large unit rather than by separate threads. |
| 828 |
C It would also be "thread-safe" for MPI implementations that have |
| 829 |
C multi-threading problems with things like request ids. |
| 830 |
C 4. We include overlap regions in both X and Y sends. |
| 831 |
C This means we can interchange the Y block with the X block. |
| 832 |
C Not sure whether this will ever be useful! |
| 833 |
C 5. The generation of a request handle by MPI_Isend and |
| 834 |
C MPI_Irecv ouught to involve some global state within MPI. |
| 835 |
C If this process is not thread safe it may be necessary to |
| 836 |
C enable a critical section around the MPI_Isend and |
| 837 |
C MPI_Irecv calls. |
| 838 |
|
| 839 |
C We need a barrier here to synchronise threads otherwise |
| 840 |
C one thread might declare it has is ready to send and receive |
| 841 |
C even though another thread is going to write to the first |
| 842 |
C threads send or receive region. This won't happen, however, |
| 843 |
C if we are careful about which data a thread updates - but at |
| 844 |
C the moment we aren't careful! |
| 845 |
_BARRIER |
| 846 |
|
| 847 |
C-- East-west messaging communication |
| 848 |
nReqs = 0 |
| 849 |
bE = myBxLo(myThid) |
| 850 |
IF ( bE .EQ. 1 .AND. commW(myThid) .EQ. COMM_MPI ) THEN |
| 851 |
C My west face uses MPI. Get ready to receive data from |
| 852 |
C west and post that we are ready to send data to the west. |
| 853 |
C Data we receive has to come with tag of thread we know is |
| 854 |
C sending from our west, data we send is tagged with our thread id |
| 855 |
C and the send direction. |
| 856 |
nReqs = nReqs+1 |
| 857 |
tagSend = mpiTagW*nThreads+myThid |
| 858 |
toPid = mpiPidW |
| 859 |
elCount = 1 |
| 860 |
elType = mpiTypeXFaceThread_xy_R8(myThid) |
| 861 |
C I tagSend,MPI_COMM_WORLD, |
| 862 |
CALL MPI_Isend( |
| 863 |
I phi(1,1-OLy,bE,myByLo(myThid)), |
| 864 |
I elCount, elType, toPid, |
| 865 |
I tagSend,MPI_COMM_WORLD, |
| 866 |
I reqIds(nReqs),mpiRC) |
| 867 |
nReqs = nReqs+1 |
| 868 |
tagRecv = mpiTagE*nThreads+myThrW(myThid) |
| 869 |
fromPid = mpiPidW |
| 870 |
elCount = 1 |
| 871 |
elType = mpiTypeXFaceThread_xy_R8(myThid) |
| 872 |
CALL MPI_Irecv( |
| 873 |
U phi(1-OLx,1-OLy,bE,myByLo(myThid)), |
| 874 |
I elCount, elType, fromPid, |
| 875 |
I tagRecv,MPI_COMM_WORLD, |
| 876 |
I reqIds(nReqs),mpiRC) |
| 877 |
ENDIF |
| 878 |
bW = myBxHi(myThid) |
| 879 |
IF ( bW .EQ. nSx .AND. commE(MyThid) .EQ. COMM_MPI ) THEN |
| 880 |
C My east face uses MPI. Get ready to receive data from |
| 881 |
C east and post that we are ready to send data to the east. |
| 882 |
C Data we receive has to come with tag of thread we know is |
| 883 |
C sending from our west, data we send is tagged with our thread id. |
| 884 |
nReqs = nReqs+1 |
| 885 |
tagSend = mpiTagE*nThreads+myThid |
| 886 |
toPid = mpiPidE |
| 887 |
elCount = 1 |
| 888 |
elType = mpiTypeXFaceThread_xy_R8(myThid) |
| 889 |
CALL MPI_Isend( |
| 890 |
I phi(sNx-OLx+1,1-OLy,bW,myByLo(myThid)), |
| 891 |
I elCount, elType, toPid, |
| 892 |
I tagSend,MPI_COMM_WORLD, |
| 893 |
I reqIds(nReqs),mpiRC) |
| 894 |
nReqs = nReqs+1 |
| 895 |
tagRecv = mpiTagW*nThreads+myThrE(myThid) |
| 896 |
fromPid = mpiPidE |
| 897 |
elCount = 1 |
| 898 |
elType = mpiTypeXFaceThread_xy_R8(myThid) |
| 899 |
CALL MPI_Irecv( |
| 900 |
U phi(sNx+1,1-OLy,bW,myByLo(myThid)), |
| 901 |
I elCount, elType, fromPid, |
| 902 |
I tagRecv,MPI_COMM_WORLD, |
| 903 |
I reqIds(nReqs),mpiRC) |
| 904 |
ENDIF |
| 905 |
C Wait for this threads east-west transactions to finish before |
| 906 |
C posting north-south transactions. We have to do this so that |
| 907 |
C the north-south transactions will send out the correct overlap |
| 908 |
C region values into the corner sections of our neighbors. |
| 909 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 910 |
|
| 911 |
CcnhDebugStarts |
| 912 |
C RETURN |
| 913 |
CcnhDebugEnds |
| 914 |
|
| 915 |
|
| 916 |
C-- North-south messaging communication |
| 917 |
nReqs = 0 |
| 918 |
bN = myByLo(myTHid) |
| 919 |
IF ( bN .EQ. 1 .AND. commS(myThid) .EQ. COMM_MPI ) THEN |
| 920 |
C My south face uses MPI. Get ready to receive data from |
| 921 |
C south and post that I am ready to send data to the south. |
| 922 |
nReqs = nReqs + 1 |
| 923 |
tagSend = mpiTagS*nThreads+myThid |
| 924 |
toPid = mpiPidS |
| 925 |
elCount = 1 |
| 926 |
elType = mpiTypeYFaceThread_xy_R8(myThid) |
| 927 |
CALL MPI_Isend( |
| 928 |
I phi(1-OLx,1,myBxLo(myThid),bN), |
| 929 |
I elCount, elType, toPid, |
| 930 |
I tagSend,MPI_COMM_WORLD, |
| 931 |
I reqIds(nReqs), mpiRC ) |
| 932 |
nReqs = nReqs + 1 |
| 933 |
tagRecv = mpiTagN*nThreads+myThrS(myThid) |
| 934 |
fromPid = mpiPidS |
| 935 |
elCount = 1 |
| 936 |
elType = mpiTypeYFaceThread_xy_R8(myThid) |
| 937 |
CALL MPI_Irecv( |
| 938 |
I phi(1-OLx,1-OLy,myBxLo(myThid),bN), |
| 939 |
I elCount, elType, fromPid, |
| 940 |
I tagRecv,MPI_COMM_WORLD, |
| 941 |
I reqIds(nReqs), mpiRC ) |
| 942 |
ENDIF |
| 943 |
C |
| 944 |
bS = myByHi(myTHid) |
| 945 |
IF ( bS .EQ. nSy .AND. commN(myThid) .EQ. COMM_MPI ) THEN |
| 946 |
C My north face uses MPI. Get ready to receive data from |
| 947 |
C north and post that I am ready to send data to the north. |
| 948 |
nReqs = nReqs + 1 |
| 949 |
tagSend = mpiTagN*nThreads+myThid |
| 950 |
toPid = mpiPidN |
| 951 |
elCount = 1 |
| 952 |
elType = mpiTypeYFaceThread_xy_R8(myThid) |
| 953 |
CALL MPI_Isend( |
| 954 |
I phi(1-OLx,sNy-OLy+1,myBxLo(myThid),bS), |
| 955 |
I elCount, elType, toPid, |
| 956 |
I tagSend,MPI_COMM_WORLD, |
| 957 |
I reqIds(nReqs), mpiRC ) |
| 958 |
nReqs = nReqs + 1 |
| 959 |
tagRecv = mpiTagS*nThreads+myThrN(myThid) |
| 960 |
fromPid = mpiPidN |
| 961 |
elCount = 1 |
| 962 |
elType = mpiTypeYFaceThread_xy_R8(myThid) |
| 963 |
CALL MPI_Irecv( |
| 964 |
I phi(1-OLx,sNy+1,myBxLo(myThid),bS), |
| 965 |
I elCount, elType, fromPid, |
| 966 |
I tagRecv,MPI_COMM_WORLD, |
| 967 |
I reqIds(nReqs), mpiRC ) |
| 968 |
ENDIF |
| 969 |
C Wait for this threads north-south transactions to finish. |
| 970 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 971 |
|
| 972 |
#ifndef ALWAYS_USE_MPI |
| 973 |
ENDIF |
| 974 |
#endif |
| 975 |
#endif /* ALLOW_USE_MPI */ |
| 976 |
|
| 977 |
#ifndef ALWAYS_USE_SYNC_COMMUNICATION |
| 978 |
ENDIF |
| 979 |
#endif |
| 980 |
#endif /* ALLOW_SYNC_COMMUNICATION */ |
| 981 |
|
| 982 |
C-- Do true shared-memory data exchanges |
| 983 |
C Note: This section also does the overlap copies for serial |
| 984 |
C code overlap copies. |
| 985 |
C-- First make sure all threads have reached here |
| 986 |
_BARRIER |
| 987 |
|
| 988 |
C-- Now do copies |
| 989 |
C Notes: |
| 990 |
C ====== |
| 991 |
C Here we copy from up to and including the overlap |
| 992 |
C regions for both the X shift and the Y shift. This |
| 993 |
C catches the "corners" of the data copied using messages |
| 994 |
C in the "synchronous communication" section above. |
| 995 |
C-- x-axis exchanges |
| 996 |
bE = myBxLo(myThid) |
| 997 |
IF ( bE .NE. 1 .OR. |
| 998 |
& commW(myThid) .EQ. COMM_SHARED ) THEN |
| 999 |
bW = bE-1 |
| 1000 |
IF ( bW .LT. 1 ) bW = nSx |
| 1001 |
DO bj=myByLo(myThid),myByHi(myThid) |
| 1002 |
DO J=1-OLy,sNy+OLy |
| 1003 |
DO I=1,OLx |
| 1004 |
phi(sNx+I,J,bW,bj)=phi( 1+I-1,J,bE,bj) |
| 1005 |
phi(1-I ,J,bE,bj)=phi(sNx-I+1,J,bW,bj) |
| 1006 |
ENDDO |
| 1007 |
ENDDO |
| 1008 |
ENDDO |
| 1009 |
ENDIF |
| 1010 |
DO bE=myBxLo(myThid)+1,myBxHi(myThid) |
| 1011 |
bW = bE-1 |
| 1012 |
DO bj=myByLo(myThid),myByHi(myThid) |
| 1013 |
DO J=1-OLy,sNy+OLy |
| 1014 |
DO I=1,OLx |
| 1015 |
phi(sNx+I,J,bW,bj)=phi( 1+I-1,J,bE,bj) |
| 1016 |
phi(1-I ,J,bE,bj)=phi(sNx-I+1,J,bW,bj) |
| 1017 |
ENDDO |
| 1018 |
ENDDO |
| 1019 |
ENDDO |
| 1020 |
ENDDO |
| 1021 |
|
| 1022 |
C Need to ensure all threads have completed x-axis transfers before |
| 1023 |
C we can do y-axis exchanges. |
| 1024 |
_BARRIER |
| 1025 |
|
| 1026 |
C-- y-axis exchanges |
| 1027 |
bN = myByLo(myThid) |
| 1028 |
IF ( bN .NE. 1 .OR. |
| 1029 |
& commS(myThid) .EQ. COMM_SHARED ) THEN |
| 1030 |
bS = bN - 1 |
| 1031 |
IF ( bS .LT. 1 ) bS = nSy |
| 1032 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
| 1033 |
DO J=1,OLy |
| 1034 |
DO I=1-OLx,sNx+OLx |
| 1035 |
phi(I,1-J ,bi,bN)=phi(I,sNy-J+1,bi,bS) |
| 1036 |
phi(I,sNy+J ,bi,bS)=phi(I,1+J-1 ,bi,bN) |
| 1037 |
ENDDO |
| 1038 |
ENDDO |
| 1039 |
ENDDO |
| 1040 |
ENDIF |
| 1041 |
DO bN=myByLo(myThid)+1,myByHi(myThid) |
| 1042 |
bS = bN - 1 |
| 1043 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
| 1044 |
DO J=1,OLy |
| 1045 |
DO I=1-OLx,sNx+OLx |
| 1046 |
phi(I,1-J ,bi,bN)=phi(I,sNy-J+1,bi,bS) |
| 1047 |
phi(I,sNy+J ,bi,bS)=phi(I,1+J-1 ,bi,bN) |
| 1048 |
ENDDO |
| 1049 |
ENDDO |
| 1050 |
ENDDO |
| 1051 |
ENDDO |
| 1052 |
|
| 1053 |
_BARRIER |
| 1054 |
|
| 1055 |
RETURN |
| 1056 |
END |
| 1057 |
|
| 1058 |
CStartOfInterface |
| 1059 |
SUBROUTINE EXCH_XYZ_R8( |
| 1060 |
U phi, |
| 1061 |
I myThid ) |
| 1062 |
C /==========================================================\ |
| 1063 |
C | SUBROUTINE EXCH_XYZ_R8 | |
| 1064 |
C | o Handle exchanges for real*8, three-dimensional arrays. | |
| 1065 |
C |==========================================================| |
| 1066 |
C | Do true shared-memory data transfers and "messaging" | |
| 1067 |
C | tranfers for blocking case of data transfers. | |
| 1068 |
C | Applications call this routine using | |
| 1069 |
C | CALL EXCH..( x, myThid ) | |
| 1070 |
C | where x is a three-dimensional array with overlaps. | |
| 1071 |
C | This routine does true-shared-memory copies for blocks | |
| 1072 |
C | within a thread. It will also do MPI meesaging between | |
| 1073 |
C | different processes. | |
| 1074 |
C | Note: | |
| 1075 |
C | ===== | |
| 1076 |
C | If it is used, "asynchronous" messaging in which | |
| 1077 |
C | communication overlaps computation is handled elsewhere | |
| 1078 |
C | - see recv.F and send.F. | |
| 1079 |
C | In practice MPI implementations may not be completely | |
| 1080 |
C | thread safe. In principle this code is correct even for | |
| 1081 |
C | mixed MPI and multi-threaded execution. | |
| 1082 |
C | In multi-thread execution data managed by other threads | |
| 1083 |
C | is only automatically visible to another thread if it | |
| 1084 |
C | existed before the thread was created. This means that | |
| 1085 |
C | for Fortran we declare arrays with overlap regions in | |
| 1086 |
C | COMMON blocks. | |
| 1087 |
C \==========================================================/ |
| 1088 |
|
| 1089 |
C === Global data === |
| 1090 |
#include "SIZE.h" |
| 1091 |
#include "EEPARAMS.h" |
| 1092 |
#include "EESUPPORT.h" |
| 1093 |
|
| 1094 |
C === Routine arguments === |
| 1095 |
C phi - Array who's overlap regions are to be exchanged |
| 1096 |
C myThid - My thread id. |
| 1097 |
Real*8 phi(1-OLx:sNx+OLx,1-OLy:sNy+OLy,1:Nz,nSx,nSy) |
| 1098 |
INTEGER myThid |
| 1099 |
CEndOfInterface |
| 1100 |
|
| 1101 |
C === Local variables === |
| 1102 |
C bi,bj - Outer loop counters |
| 1103 |
C I,J,K - Inner loop counters |
| 1104 |
INTEGER bi, bj |
| 1105 |
INTEGER I, J, K |
| 1106 |
#ifdef ALLOW_USE_MPI |
| 1107 |
C tagSend - Tags used to mark and select messages |
| 1108 |
C tagRecv |
| 1109 |
C nReqs - MPI request counter. |
| 1110 |
C reqIds - Outstanding requests to wait on. |
| 1111 |
C mpiRC - MPI return code |
| 1112 |
C mpiStatArr - Multi-request status reporting array. |
| 1113 |
C toPid - Proc. id sending to |
| 1114 |
C fromPid - Proc. id receving from |
| 1115 |
C elCount - Number of items to send or receive |
| 1116 |
C elType - Datat type of elements |
| 1117 |
INTEGER tagSend |
| 1118 |
INTEGER tagRecv |
| 1119 |
INTEGER nReqs |
| 1120 |
INTEGER reqIds(4) |
| 1121 |
INTEGER mpiRC |
| 1122 |
INTEGER mpiStatArr(MPI_STATUS_SIZE,4) |
| 1123 |
INTEGER toPid |
| 1124 |
INTEGER fromPid |
| 1125 |
INTEGER elCount |
| 1126 |
INTEGER elType |
| 1127 |
#endif /* ALLOW_USE_MPI */ |
| 1128 |
C bE, bW - Block index to east, west, north and |
| 1129 |
C bN, bS south. |
| 1130 |
INTEGER bW, bE, bS, bN |
| 1131 |
|
| 1132 |
C-- Do "message-passing" data exchanges |
| 1133 |
#ifdef ALLOW_SYNC_COMMUNICATION |
| 1134 |
#ifndef ALWAYS_USE_SYNC_COMMUNICATION |
| 1135 |
IF ( usingSyncMessages ) THEN |
| 1136 |
#endif |
| 1137 |
|
| 1138 |
C-- MPI based data exchages |
| 1139 |
#ifdef ALLOW_USE_MPI |
| 1140 |
#ifndef ALWAYS_USE_MPI |
| 1141 |
IF ( usingMPI ) THEN |
| 1142 |
#endif |
| 1143 |
|
| 1144 |
C Here use the approach that all the X shifts are completed |
| 1145 |
C before the Y shifts start. Then we can expand our Y shifts |
| 1146 |
C to include X overlap regions. This means we don't have to |
| 1147 |
C transfer corners separately. |
| 1148 |
C Notes: |
| 1149 |
C ====== |
| 1150 |
C 1. If we are only using messages in Y and using "true shared-memory" |
| 1151 |
C in X then we rely on the shared-memory code to be implemented so that |
| 1152 |
C corner values will be set correctly. The true shared-memory code |
| 1153 |
C is in the block below - so be careful when altering it! |
| 1154 |
C 2. In order to ensure we grab the right message we need to tag |
| 1155 |
C messages. We tag the messages with the direction in which they |
| 1156 |
C were sent out. Thus to receive from a processor to our west we |
| 1157 |
C request the next message from that processor that was sent out |
| 1158 |
C with a tag indicating an easterly send. |
| 1159 |
C 3. We could implement a "safe" code here by having everyone |
| 1160 |
C sync and then getting a particular thread ( say thread 1 ) to |
| 1161 |
C do the MPI_Isend sequentially. In this mode the "communication" |
| 1162 |
C thread would loop over nThreads and do MPI_Sendrecv for each |
| 1163 |
C one. This could be more efficient on some platforms as the messages |
| 1164 |
C might be sent as one large unit rather than by separate threads. |
| 1165 |
C It would also be "thread-safe" for MPI implementations that have |
| 1166 |
C multi-threading problems with things like request ids. |
| 1167 |
C 4. We include overlap regions in both X and Y sends. |
| 1168 |
C This means we can interchange the Y block with the X block. |
| 1169 |
C Not sure whether this will ever be useful! |
| 1170 |
C 5. The generation of a request handle by MPI_Isend and |
| 1171 |
C MPI_Irecv ouught to involve some global state within MPI. |
| 1172 |
C If this process is not thread safe it may be necessary to |
| 1173 |
C enable a critical section around the MPI_Isend and |
| 1174 |
C MPI_Irecv calls. |
| 1175 |
|
| 1176 |
C We need a barrier here to synchronise threads otherwise |
| 1177 |
C one thread might declare it has is ready to send and receive |
| 1178 |
C even though another thread is going to write to the first |
| 1179 |
C threads send or receive region. This won't happen, however, |
| 1180 |
C if we are careful about which data a thread updates - but at |
| 1181 |
C the moment we aren't careful! |
| 1182 |
_BARRIER |
| 1183 |
|
| 1184 |
C-- East-west messaging communication |
| 1185 |
bE = myBxLo(myThid) |
| 1186 |
bW = myBxHi(myThid) |
| 1187 |
nReqs = 0 |
| 1188 |
IF ( bE .EQ. 1 .AND. commW(myThid) .EQ. COMM_MPI ) THEN |
| 1189 |
C My west face uses MPI. Get ready to receive data from |
| 1190 |
C west and post that we are ready to send data to the west. |
| 1191 |
C Data we receive has to come with tag of thread we know is |
| 1192 |
C sending from our west, data we send is tagged with our thread id |
| 1193 |
C and the send direction. |
| 1194 |
nReqs = nReqs+1 |
| 1195 |
tagSend = mpiTagW*nThreads+myThid |
| 1196 |
toPid = mpiPidW |
| 1197 |
elCount = 1 |
| 1198 |
elType = mpiTypeXFaceThread_xyz_R8(myThid) |
| 1199 |
CALL MPI_Isend( |
| 1200 |
I phi(1,1-OLy,1,bE,myByLo(myThid)), |
| 1201 |
I elCount, elType, toPid, |
| 1202 |
I tagSend,MPI_COMM_WORLD, |
| 1203 |
I reqIds(nReqs),mpiRC) |
| 1204 |
ENDIF |
| 1205 |
IF ( bW .EQ. nSx .AND. commE(MyThid) .EQ. COMM_MPI ) THEN |
| 1206 |
nReqs = nReqs+1 |
| 1207 |
tagRecv = mpiTagW*nThreads+myThrE(myThid) |
| 1208 |
fromPid = mpiPidE |
| 1209 |
elCount = 1 |
| 1210 |
elType = mpiTypeXFaceThread_xyz_R8(myThid) |
| 1211 |
CALL MPI_Irecv( |
| 1212 |
U phi(sNx+1,1-OLy,1,bW,myByLo(myThid)), |
| 1213 |
I elCount, elType, fromPid, |
| 1214 |
I tagRecv,MPI_COMM_WORLD, |
| 1215 |
I reqIds(nReqs),mpiRC) |
| 1216 |
ENDIF |
| 1217 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 1218 |
|
| 1219 |
nReqs = 0 |
| 1220 |
IF ( bW .EQ. nSx .AND. commE(MyThid) .EQ. COMM_MPI ) THEN |
| 1221 |
C My east face uses MPI. Get ready to receive data from |
| 1222 |
C east and post that we are ready to send data to the east. |
| 1223 |
C Data we receive has to come with tag of thread we know is |
| 1224 |
C sending from our west, data we send is tagged with our thread id. |
| 1225 |
nReqs = nReqs+1 |
| 1226 |
tagSend = mpiTagE*nThreads+myThid |
| 1227 |
toPid = mpiPidE |
| 1228 |
elCount = 1 |
| 1229 |
elType = mpiTypeXFaceThread_xyz_R8(myThid) |
| 1230 |
CALL MPI_Isend( |
| 1231 |
I phi(sNx-OLx+1,1-OLy,1,bW,myByLo(myThid)), |
| 1232 |
I elCount, elType, toPid, |
| 1233 |
I tagSend,MPI_COMM_WORLD, |
| 1234 |
I reqIds(nReqs),mpiRC) |
| 1235 |
ENDIF |
| 1236 |
IF ( bE .EQ. 1 .AND. commW(myThid) .EQ. COMM_MPI ) THEN |
| 1237 |
nReqs = nReqs+1 |
| 1238 |
tagRecv = mpiTagE*nThreads+myThrW(myThid) |
| 1239 |
fromPid = mpiPidW |
| 1240 |
elCount = 1 |
| 1241 |
elType = mpiTypeXFaceThread_xyz_R8(myThid) |
| 1242 |
CALL MPI_Irecv( |
| 1243 |
U phi(1-OLx,1-OLy,1,bE,myByLo(myThid)), |
| 1244 |
I elCount, elType, fromPid, |
| 1245 |
I tagRecv,MPI_COMM_WORLD, |
| 1246 |
I reqIds(nReqs),mpiRC) |
| 1247 |
ENDIF |
| 1248 |
|
| 1249 |
C Wait for this threads east-west transactions to finish before |
| 1250 |
C posting north-south transactions. We have to do this so that |
| 1251 |
C the north-south transactions will send out the correct overlap |
| 1252 |
C region values into the corner sections of our neighbors. |
| 1253 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 1254 |
|
| 1255 |
CcnhDebugStarts |
| 1256 |
C RETURN |
| 1257 |
CcnhDebugEnds |
| 1258 |
|
| 1259 |
|
| 1260 |
C-- North-south messaging communication |
| 1261 |
nReqs = 0 |
| 1262 |
bN = myByLo(myTHid) |
| 1263 |
bS = myByHi(myTHid) |
| 1264 |
|
| 1265 |
IF ( bN .EQ. 1 .AND. commS(myThid) .EQ. COMM_MPI ) THEN |
| 1266 |
C My south face uses MPI. Get ready to receive data from |
| 1267 |
C south and post that I am ready to send data to the south. |
| 1268 |
nReqs = nReqs + 1 |
| 1269 |
tagSend = mpiTagS*nThreads+myThid |
| 1270 |
toPid = mpiPidS |
| 1271 |
elCount = 1 |
| 1272 |
elType = mpiTypeYFaceThread_xyz_R8(myThid) |
| 1273 |
CALL MPI_Isend( |
| 1274 |
I phi(1-OLx,1,1,myBxLo(myThid),bN), |
| 1275 |
I elCount, elType, toPid, |
| 1276 |
I tagSend,MPI_COMM_WORLD, |
| 1277 |
I reqIds(nReqs), mpiRC ) |
| 1278 |
ENDIF |
| 1279 |
IF ( bS .EQ. nSy .AND. commN(myThid) .EQ. COMM_MPI ) THEN |
| 1280 |
nReqs = nReqs + 1 |
| 1281 |
tagRecv = mpiTagS*nThreads+myThrN(myThid) |
| 1282 |
fromPid = mpiPidN |
| 1283 |
elCount = 1 |
| 1284 |
elType = mpiTypeYFaceThread_xyz_R8(myThid) |
| 1285 |
CALL MPI_Irecv( |
| 1286 |
I phi(1-OLx,sNy+1,1,myBxLo(myThid),bS), |
| 1287 |
I elCount, elType, fromPid, |
| 1288 |
I tagRecv,MPI_COMM_WORLD, |
| 1289 |
I reqIds(nReqs), mpiRC ) |
| 1290 |
ENDIF |
| 1291 |
C Wait for this threads north-south transactions to finish. |
| 1292 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 1293 |
C |
| 1294 |
nReqs = 0 |
| 1295 |
IF ( bS .EQ. nSy .AND. commN(myThid) .EQ. COMM_MPI ) THEN |
| 1296 |
C My north face uses MPI. Get ready to receive data from |
| 1297 |
C north and post that I am ready to send data to the north. |
| 1298 |
nReqs = nReqs + 1 |
| 1299 |
tagSend = mpiTagN*nThreads+myThid |
| 1300 |
toPid = mpiPidN |
| 1301 |
elCount = 1 |
| 1302 |
elType = mpiTypeYFaceThread_xyz_R8(myThid) |
| 1303 |
CALL MPI_Isend( |
| 1304 |
I phi(1-OLx,sNy-OLy+1,1,myBxLo(myThid),bS), |
| 1305 |
I elCount, elType, toPid, |
| 1306 |
I tagSend,MPI_COMM_WORLD, |
| 1307 |
I reqIds(nReqs), mpiRC ) |
| 1308 |
ENDIF |
| 1309 |
IF ( bN .EQ. 1 .AND. commS(myThid) .EQ. COMM_MPI ) THEN |
| 1310 |
nReqs = nReqs + 1 |
| 1311 |
tagRecv = mpiTagN*nThreads+myThrS(myThid) |
| 1312 |
fromPid = mpiPidS |
| 1313 |
elCount = 1 |
| 1314 |
elType = mpiTypeYFaceThread_xyz_R8(myThid) |
| 1315 |
CALL MPI_Irecv( |
| 1316 |
I phi(1-OLx,1-OLy,1,myBxLo(myThid),bN), |
| 1317 |
I elCount, elType, fromPid, |
| 1318 |
I tagRecv,MPI_COMM_WORLD, |
| 1319 |
I reqIds(nReqs), mpiRC ) |
| 1320 |
ENDIF |
| 1321 |
C Wait for this threads north-south transactions to finish. |
| 1322 |
CALL MPI_Waitall( nReqs, reqIds, mpiStatArr, mpiRC ) |
| 1323 |
|
| 1324 |
#ifndef ALWAYS_USE_MPI |
| 1325 |
ENDIF |
| 1326 |
#endif |
| 1327 |
#endif /* ALLOW_USE_MPI */ |
| 1328 |
|
| 1329 |
#ifndef ALWAYS_USE_SYNC_COMMUNICATION |
| 1330 |
ENDIF |
| 1331 |
#endif |
| 1332 |
#endif /* ALLOW_SYNC_COMMUNICATION */ |
| 1333 |
|
| 1334 |
C-- Do true shared-memory data exchanges |
| 1335 |
C Note: This section also does the overlap copies for serial |
| 1336 |
C code overlap copies. |
| 1337 |
C-- First make sure all threads have reached here |
| 1338 |
_BARRIER |
| 1339 |
|
| 1340 |
C-- Now do copies |
| 1341 |
C Notes: |
| 1342 |
C ====== |
| 1343 |
C Here we copy from up to and including the overlap |
| 1344 |
C regions for both the X shift and the Y shift. This |
| 1345 |
C catches the "corners" of the data copied using messages |
| 1346 |
C in the "synchronous communication" section above. |
| 1347 |
C-- x-axis exchanges |
| 1348 |
bE = myBxLo(myThid) |
| 1349 |
IF ( bE .NE. 1 .OR. |
| 1350 |
& commW(myThid) .EQ. COMM_SHARED ) THEN |
| 1351 |
bW = bE-1 |
| 1352 |
IF ( bW .LT. 1 ) bW = nSx |
| 1353 |
DO bj=myByLo(myThid),myByHi(myThid) |
| 1354 |
DO K=1,Nz |
| 1355 |
DO J=1-OLy,sNy+OLy |
| 1356 |
DO I=1,OLx |
| 1357 |
phi(sNx+I,J,K,bW,bj)=phi( 1+I-1,J,K,bE,bj) |
| 1358 |
phi(1-I ,J,K,bE,bj)=phi(sNx-I+1,J,K,bW,bj) |
| 1359 |
ENDDO |
| 1360 |
ENDDO |
| 1361 |
ENDDO |
| 1362 |
ENDDO |
| 1363 |
ENDIF |
| 1364 |
DO bE=myBxLo(myThid)+1,myBxHi(myThid) |
| 1365 |
bW = bE-1 |
| 1366 |
DO bj=myByLo(myThid),myByHi(myThid) |
| 1367 |
DO K=1,Nz |
| 1368 |
DO J=1-OLy,sNy+OLy |
| 1369 |
DO I=1,OLx |
| 1370 |
phi(sNx+I,J,K,bW,bj)=phi( 1+I-1,J,K,bE,bj) |
| 1371 |
phi(1-I ,J,K,bE,bj)=phi(sNx-I+1,J,K,bW,bj) |
| 1372 |
ENDDO |
| 1373 |
ENDDO |
| 1374 |
ENDDO |
| 1375 |
ENDDO |
| 1376 |
ENDDO |
| 1377 |
|
| 1378 |
C Need to ensure all threads have completed x-axis transfers before |
| 1379 |
C we can do y-axis exchanges. |
| 1380 |
_BARRIER |
| 1381 |
|
| 1382 |
C-- y-axis exchanges |
| 1383 |
bN = myByLo(myThid) |
| 1384 |
IF ( bN .NE. 1 .OR. |
| 1385 |
& commS(myThid) .EQ. COMM_SHARED ) THEN |
| 1386 |
bS = bN - 1 |
| 1387 |
IF ( bS .LT. 1 ) bS = nSy |
| 1388 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
| 1389 |
DO K=1,Nz |
| 1390 |
DO J=1,OLy |
| 1391 |
DO I=1-OLx,sNx+OLx |
| 1392 |
phi(I,1-J ,K,bi,bN)=phi(I,sNy-J+1,K,bi,bS) |
| 1393 |
phi(I,sNy+J ,K,bi,bS)=phi(I,1+J-1 ,K,bi,bN) |
| 1394 |
ENDDO |
| 1395 |
ENDDO |
| 1396 |
ENDDO |
| 1397 |
ENDDO |
| 1398 |
ENDIF |
| 1399 |
DO bN=myByLo(myThid)+1,myByHi(myThid) |
| 1400 |
bS = bN - 1 |
| 1401 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
| 1402 |
DO K=1,Nz |
| 1403 |
DO J=1,OLy |
| 1404 |
DO I=1-OLx,sNx+OLx |
| 1405 |
phi(I,1-J ,K,bi,bN)=phi(I,sNy-J+1,K,bi,bS) |
| 1406 |
phi(I,sNy+J ,K,bi,bS)=phi(I,1+J-1 ,K,bi,bN) |
| 1407 |
ENDDO |
| 1408 |
ENDDO |
| 1409 |
ENDDO |
| 1410 |
ENDDO |
| 1411 |
ENDDO |
| 1412 |
|
| 1413 |
_BARRIER |
| 1414 |
|
| 1415 |
RETURN |
| 1416 |
END |
Parent Directory
| 
Revision Log
| 
Revision Graph