1 |
C $Id$ |
2 |
|
3 |
#include "CPP_EEOPTIONS.h" |
4 |
|
5 |
CStartOfInterface |
6 |
SUBROUTINE INI_THREADING_ENVIRONMENT |
7 |
C /==========================================================\ |
8 |
C | SUBROUTINE INI_THREADING_ENVIRONMENT | |
9 |
C | o Initialise multi-threaded environment. | |
10 |
C |==========================================================| |
11 |
C | Generally we do not start separate threads here but | |
12 |
C | just initialise data structures indicating which of the | |
13 |
C | nSx x nSy blocks a thread is responsible for. | |
14 |
C | The multiple threads are spawned in the top level MAIN | |
15 |
C | routine. | |
16 |
C \==========================================================/ |
17 |
|
18 |
C == Global data == |
19 |
#include "SIZE.h" |
20 |
#include "EEPARAMS.h" |
21 |
#include "EESUPPORT.h" |
22 |
CEndOfInterface |
23 |
|
24 |
C == Local variables == |
25 |
C bXPerThread - Blocks of size sNx per thread. |
26 |
C byPerThread - Blocks of size sNy per thread. |
27 |
C Thid - Thread index. Temporary used in loops |
28 |
C which set per. thread values on a |
29 |
C cartesian grid. |
30 |
C bxLo, bxHi - Work vars. for thread index |
31 |
C byLo, byHi range. bxLo is the lowest i index |
32 |
C that a thread covers, bxHi is the |
33 |
C highest i index. byLo is the lowest |
34 |
C j index, byHi is the highest j index. |
35 |
C I, J - Loop counter |
36 |
C msgBuf - I/O buffer for reporting status information. |
37 |
C myThid - Dummy thread id for use in printed messages |
38 |
C ( this routine "INI_THREADING_ENVIRONMENT" is called before |
39 |
C multi-threading has started.) |
40 |
C threadWest - Temporaries used in calculating neighbor threads. |
41 |
C threadEast |
42 |
C threadSouth |
43 |
C threadNorth |
44 |
INTEGER bxPerThread |
45 |
INTEGER byPerThread |
46 |
INTEGER Thid |
47 |
INTEGER bxLo, bxHi |
48 |
INTEGER byLo, byHi |
49 |
INTEGER I, J |
50 |
CHARACTER*(MAX_LEN_MBUF) msgBuf |
51 |
INTEGER myThid |
52 |
INTEGER threadWest |
53 |
INTEGER threadEast |
54 |
INTEGER threadSouth |
55 |
INTEGER threadNorth |
56 |
INTEGER threadNW |
57 |
INTEGER threadNE |
58 |
INTEGER threadSW |
59 |
INTEGER threadSE |
60 |
|
61 |
#ifdef ALLOW_USE_MPI |
62 |
C elCount - Length in elements of an MPI datatype |
63 |
C elStride - Stride between elements of an MPI datatype. |
64 |
C elLen - Length of each element of the datatype |
65 |
C arrElSize - Size in bytes of an array element |
66 |
C arrElSep - Separation in array elements between consecutive |
67 |
C start locations for an MPI datatype. |
68 |
C mpiRC - MPI return code |
69 |
INTEGER elCount |
70 |
INTEGER elStride |
71 |
INTEGER elLen |
72 |
INTEGER arrElSize |
73 |
INTEGER arrElSep |
74 |
INTEGER mpiRC |
75 |
#endif /* ALLOW_USE_MPI */ |
76 |
|
77 |
C-- Set default for all threads of having no blocks to |
78 |
C-- work on - except for thread 1. |
79 |
myBxLo(1) = 1 |
80 |
myBxHi(1) = nSx |
81 |
myByLo(1) = 1 |
82 |
myByHi(1) = nSy |
83 |
DO I = 2, MAX_NO_THREADS |
84 |
myBxLo(I) = 0 |
85 |
myBxHi(I) = 1 |
86 |
myByLo(I) = 0 |
87 |
myByHi(I) = 1 |
88 |
ENDDO |
89 |
myThid = 1 |
90 |
|
91 |
C-- If there are multiple threads allocate different range of the |
92 |
C-- nSx*nSy blocks to each thread. |
93 |
C For now handle simple case of no. blocks nSx = n*nTx and |
94 |
C no. blocks nSy = m*nTy ( where m and n are integer ). This |
95 |
C is handled by simply mapping threads to blocks in sequence |
96 |
C with the x thread index moving fastest. |
97 |
C Later code which sets the thread number of neighboring blocks |
98 |
C needs to be consisten with the code here. |
99 |
nThreads = nTx * nTy |
100 |
|
101 |
C-- Initialise the barrier mechanism |
102 |
CALL BARRIER_INIT |
103 |
|
104 |
IF ( nThreads .NE. nTx*nTy ) THEN |
105 |
WRITE(msgBuf,'(A,A,A,I,A,I)') |
106 |
& 'S/R INI_THREADING_ENVIRONMENT:', |
107 |
& ' Total number of threads is not the same as nTx*nTy.', |
108 |
& ' nTx * nTy = ',nTx*nTy,' nThreads = ',nThreads |
109 |
CALL PRINT_ERROR(msgBuf, myThid) |
110 |
eeBootError = .TRUE. |
111 |
STOP 'ABNORMAL END: S/R INI_THREADING_ENVIRONMENT' |
112 |
ENDIF |
113 |
bxPerThread = nSx/nTx |
114 |
IF ( bxPerThread*nTx .NE. nSx ) THEN |
115 |
WRITE(msgBuf,'(A,A)') |
116 |
& 'S/R INI_THREADING_ENVIRONMENT:', |
117 |
& ' Number of blocks in X (nSx) must be exact multiple of threads in X (nTx).' |
118 |
CALL PRINT_ERROR(msgBuf, myThid) |
119 |
eeBootError = .TRUE. |
120 |
STOP 'ABNORMAL END: S/R INI_THREADING_ENVIRONMENT' |
121 |
ENDIF |
122 |
byPerThread = nSy/nTy |
123 |
IF ( byPerThread*nTy .NE. nSy ) THEN |
124 |
WRITE(msgBuf,'(A,A)') |
125 |
& 'S/R INI_THREADING_ENVIRONMENT:', |
126 |
& ' Number of blocks in Y (nSy) must be exact multiple of threads in Y (nTy).' |
127 |
CALL PRINT_ERROR(msgBuf, myThid) |
128 |
eeBootError = .TRUE. |
129 |
STOP 'ABNORMAL END: S/R INI_THREADING_ENVIRONMENT' |
130 |
ENDIF |
131 |
IF ( .NOT. eeBootError ) THEN |
132 |
byLo = 1 |
133 |
DO J=1,nTy |
134 |
byHi = byLo+byPerThread-1 |
135 |
bxLo = 1 |
136 |
DO I=1,nTx |
137 |
Thid = (J-1)*nTx+I |
138 |
bxHi = bxLo+bxPerThread-1 |
139 |
myBxLo(Thid) = bxLo |
140 |
myBxHi(Thid) = bxHi |
141 |
myByLo(Thid) = byLo |
142 |
myByHi(Thid) = byHi |
143 |
bxLo = bxHi+1 |
144 |
ENDDO |
145 |
byLo = byHi+1 |
146 |
ENDDO |
147 |
ENDIF |
148 |
|
149 |
C-- Set flags saying how each thread is communicating |
150 |
C Notes: |
151 |
C ====== |
152 |
C By default each block communicates with its neighbor using |
153 |
C direct reads and writes from the neighbors overlap regions. |
154 |
C This rule will always applie for the blocks in the interior |
155 |
C of a processes domain, but for the "outside" faces of blocks on |
156 |
C the edges of the processes domain i.e. where bx=1 or nSx or |
157 |
C where by = 1 or nSy. In this section each thread checks to see |
158 |
C whether any of the blocks it is responsible for are "outside" |
159 |
C blocks and if so what communication strategy should be used. |
160 |
C to |
161 |
DO I=1, nThreads |
162 |
|
163 |
C 1. Check for block which is on the west edge. |
164 |
commW(I) = COMM_SHARED |
165 |
IF ( notUsingXPeriodicity .AND. |
166 |
& myBxLo(I) .EQ. 1 .AND. |
167 |
& myPx .EQ. 1 ) THEN |
168 |
commW(I) = COMM_NONE |
169 |
ELSEIF ( myBxLo(I) .EQ. 1 ) THEN |
170 |
#ifdef ALLOW_USE_MPI |
171 |
#ifndef ALWAYS_USE_MPI |
172 |
IF ( usingMPI ) THEN |
173 |
#endif |
174 |
IF ( mpiPidW .NE. MPI_PROC_NULL ) THEN |
175 |
commW(I) = COMM_MPI |
176 |
allMyEdgesAreSharedMemory(I) = .FALSE. |
177 |
ENDIF |
178 |
#ifndef ALWAYS_USE_MPI |
179 |
ENDIF |
180 |
#endif |
181 |
#endif /* ALLOW_USE_MPI */ |
182 |
ENDIF |
183 |
|
184 |
C 2. Check for block which is on the east edge. |
185 |
commE(I) = COMM_SHARED |
186 |
IF ( notUsingXPeriodicity .AND. |
187 |
& myBxHi(I) .EQ. nSx .AND. |
188 |
& myPx .EQ. nPx ) THEN |
189 |
commE(I) = COMM_NONE |
190 |
ELSEIF ( myBxHi(I) .EQ. nSx ) THEN |
191 |
#ifdef ALLOW_USE_MPI |
192 |
#ifndef ALWAYS_USE_MPI |
193 |
IF ( usingMPI ) THEN |
194 |
#endif |
195 |
IF ( mpiPidE .NE. MPI_PROC_NULL ) THEN |
196 |
commE(I) = COMM_MPI |
197 |
allMyEdgesAreSharedMemory(I) = .FALSE. |
198 |
ENDIF |
199 |
#ifndef ALWAYS_USE_MPI |
200 |
ENDIF |
201 |
#endif |
202 |
#endif /* ALLOW_USE_MPI */ |
203 |
ENDIF |
204 |
|
205 |
C 3. Check for block which is southern edge |
206 |
commS(I) = COMM_SHARED |
207 |
IF ( notUsingYPeriodicity .AND. |
208 |
& myByLo(I) .EQ. 1 .AND. |
209 |
& myPy .EQ. 1 ) THEN |
210 |
commS(I) = COMM_NONE |
211 |
ELSEIF ( myByLo(I) .EQ. 1 ) THEN |
212 |
#ifdef ALLOW_USE_MPI |
213 |
#ifndef ALWAYS_USE_MPI |
214 |
IF ( usingMPI ) THEN |
215 |
#endif |
216 |
IF ( mpiPidS .NE. MPI_PROC_NULL ) THEN |
217 |
commS(I) = COMM_MPI |
218 |
allMyEdgesAreSharedMemory(I) = .FALSE. |
219 |
ENDIF |
220 |
#ifndef ALWAYS_USE_MPI |
221 |
ENDIF |
222 |
#endif |
223 |
#endif /* ALLOW_USE_MPI */ |
224 |
ENDIF |
225 |
|
226 |
C 4. Check for block which is on northern edge |
227 |
commN(I) = COMM_SHARED |
228 |
IF ( notUsingYPeriodicity .AND. |
229 |
& myByHi(I) .EQ. nSy .AND. |
230 |
& myPy .EQ. nPy ) THEN |
231 |
commN(I) = COMM_NONE |
232 |
ELSEIF ( myByHi(I) .EQ. nSy ) THEN |
233 |
#ifdef ALLOW_USE_MPI |
234 |
#ifndef ALWAYS_USE_MPI |
235 |
IF ( usingMPI ) THEN |
236 |
#endif |
237 |
IF ( mpiPidN .NE. MPI_PROC_NULL ) THEN |
238 |
commN(I) = COMM_MPI |
239 |
allMyEdgesAreSharedMemory(I) = .FALSE. |
240 |
ENDIF |
241 |
#ifndef ALWAYS_USE_MPI |
242 |
ENDIF |
243 |
#endif |
244 |
#endif /* ALLOW_USE_MPI */ |
245 |
ENDIF |
246 |
ENDDO |
247 |
|
248 |
C-- Print mapping of threads to grid points. |
249 |
WRITE(msgBuf,'(A)') '// ======================================================' |
250 |
CALL PRINT_MESSAGE( msgBuf, standardMessageUnit, |
251 |
& SQUEEZE_RIGHT , 1) |
252 |
|
253 |
WRITE(msgBuf,'(A)') '// Mapping of tiles to threads' |
254 |
CALL PRINT_MESSAGE( msgBuf, standardMessageUnit, |
255 |
& SQUEEZE_RIGHT , 1) |
256 |
|
257 |
WRITE(msgBuf,'(A)') '// ======================================================' |
258 |
CALL PRINT_MESSAGE( msgBuf, standardMessageUnit, |
259 |
& SQUEEZE_RIGHT , 1) |
260 |
|
261 |
DO I=1,nThreads |
262 |
WRITE(msgBuf,'(A,I4,A,4(I4,A1))') |
263 |
& '// -o- Thread',I,', tiles (', |
264 |
& myBxLo(I),':',myBxHi(I),',',myByLo(I),':',myByHi(I),')' |
265 |
CALL PRINT_MESSAGE( msgBuf, standardMessageUnit, SQUEEZE_BOTH , 1) |
266 |
IF ( myBxLo(I) .NE. 1 .OR. |
267 |
& commW(I) .EQ. COMM_SHARED ) THEN |
268 |
WRITE(msgBuf,'(A,A)') '//',' shared memory to west.' |
269 |
ELSEIF ( commW(I) .NE. COMM_NONE ) THEN |
270 |
WRITE(msgBuf,'(A,A)') '//',' messages to west.' |
271 |
ELSE |
272 |
WRITE(msgBuf,'(A,A)') '//',' no communication to west.' |
273 |
ENDIF |
274 |
CALL PRINT_MESSAGE( msgBuf, standardMessageUnit, SQUEEZE_RIGHT , 1) |
275 |
IF ( myBxHi(I) .NE. nSx .OR. |
276 |
& commE(I) .EQ. COMM_SHARED ) THEN |
277 |
WRITE(msgBuf,'(A,A)') '//',' shared memory to east.' |
278 |
ELSEIF ( commE(I) .NE. COMM_NONE ) THEN |
279 |
WRITE(msgBuf,'(A,A)') '//',' messages to east.' |
280 |
ELSE |
281 |
WRITE(msgBuf,'(A,A)') '//',' no communication to east.' |
282 |
ENDIF |
283 |
CALL PRINT_MESSAGE( msgBuf, standardMessageUnit, SQUEEZE_RIGHT , 1) |
284 |
IF ( myByLo(I) .NE. 1 .OR. |
285 |
& commS(I) .EQ. COMM_SHARED ) THEN |
286 |
WRITE(msgBuf,'(A,A)') '//',' shared memory to south.' |
287 |
ELSEIF ( commS(I) .NE. COMM_NONE ) THEN |
288 |
WRITE(msgBuf,'(A,A)') '//',' messages to south.' |
289 |
ELSE |
290 |
WRITE(msgBuf,'(A,A)') '//',' no communication to south.' |
291 |
ENDIF |
292 |
CALL PRINT_MESSAGE( msgBuf, standardMessageUnit, SQUEEZE_RIGHT , 1) |
293 |
IF ( myByHi(I) .NE. nSy .OR. |
294 |
& commN(I) .EQ. COMM_SHARED ) THEN |
295 |
WRITE(msgBuf,'(A,A)') '//',' shared memory to north.' |
296 |
ELSEIF ( commN(I) .NE. COMM_NONE ) THEN |
297 |
WRITE(msgBuf,'(A,A)') '//',' messages to north.' |
298 |
ELSE |
299 |
WRITE(msgBuf,'(A,A)') '//',' no communication to north.' |
300 |
ENDIF |
301 |
CALL PRINT_MESSAGE( msgBuf, standardMessageUnit, SQUEEZE_RIGHT , 1) |
302 |
ENDDO |
303 |
WRITE(msgBuf,'(A)') ' ' |
304 |
CALL PRINT_MESSAGE( msgBuf, standardMessageUnit, SQUEEZE_RIGHT , 1) |
305 |
|
306 |
#ifdef ALLOW_USE_MPI |
307 |
C-- Create MPI datatypes for communicating thread boundaries if needed |
308 |
C For every thread we define 8 MPI datatypes for use |
309 |
C in indicating regions of data to transfer as follows: |
310 |
C o mpiTypeXFaceThread_xy_r4 |
311 |
C Handles east-west transfer for XY arrays of REAL*4 |
312 |
C o mpiTypeXFaceThread_xy_r8 |
313 |
C Handles east-west transfer for XY arrays of REAL*8 |
314 |
C o mpiTypeYFaceThread_xy_r4 |
315 |
C Handles north-south transfer for XY arrays of REAL*4 |
316 |
C o mpiTypeYFaceThread_xy_r8 |
317 |
C Handles north-south transfer for XY arrays of REAL*8 |
318 |
C o mpiTypeXFaceThread_xyz_r4 |
319 |
C Handles east-west transfer for XYZ arrays of REAL*4 |
320 |
C o mpiTypeXFaceThread_xyz_r8 |
321 |
C Handles east-west transfer for XYZ arrays of REAL*8 |
322 |
C o mpiTypeYFaceThread_xyz_r4 |
323 |
C Handles north-south transfer for XYZ arrays of REAL*4 |
324 |
C o mpiTypeYFaceThread_xyz_r8 |
325 |
C Handles north-south transfer for XYZarrays of REAL*8 |
326 |
#ifndef ALWAYS_USE_MPI |
327 |
IF ( usingMPI ) THEN |
328 |
#endif |
329 |
DO I =1, nThreads |
330 |
|
331 |
C x-face exchanges for xy real*4 data |
332 |
elCount = myByHi(I)-myByLo(I)+1 |
333 |
elLen = 1 |
334 |
arrElSep = (sNx+OLx*2)*(sNy+OLy*2)*nSx |
335 |
arrElSize = 4 |
336 |
elStride = arrElSep*arrElSize |
337 |
CALL MPI_TYPE_HVECTOR(elCount,elLen,elStride,mpiTypeXFaceBlock_xy_r4, |
338 |
O mpiTypeXFaceThread_xy_r4(I), mpiRC ) |
339 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
340 |
eeBootError = .TRUE. |
341 |
WRITE(msgBuf,'(A,I)') |
342 |
& 'S/R INI_THREADS: MPI_TYPE_HVECTOR (mpiTypeXFaceThread_xy_r4)', |
343 |
& mpiRC |
344 |
CALL PRINT_ERROR( msgBuf , myThid) |
345 |
ENDIF |
346 |
CALL MPI_TYPE_COMMIT(mpiTypeXFaceThread_xy_r4(I),mpiRC) |
347 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
348 |
eeBootError = .TRUE. |
349 |
WRITE(msgBuf,'(A,I)') |
350 |
& 'S/R INI_THREADS: MPI_TYPE_COMMIT (mpiTypeXFaceThread_xy_r4)', |
351 |
& mpiRC |
352 |
CALL PRINT_ERROR( msgBuf , myThid) |
353 |
ENDIF |
354 |
|
355 |
C x-face exchanges for xy real*8 data |
356 |
arrElSize = 8 |
357 |
elStride = arrElSep*arrElSize |
358 |
CALL MPI_TYPE_HVECTOR(elCount,elLen,elStride,mpiTypeXFaceBlock_xy_r8, |
359 |
O mpiTypeXFaceThread_xy_r8(I), mpiRC ) |
360 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
361 |
eeBootError = .TRUE. |
362 |
WRITE(msgBuf,'(A,I)') |
363 |
& 'S/R INI_THREADS: MPI_TYPE_HVECTOR (mpiTypeXFaceThread_xy_r8)', |
364 |
& mpiRC |
365 |
CALL PRINT_ERROR( msgBuf , myThid) |
366 |
ENDIF |
367 |
CALL MPI_TYPE_COMMIT(mpiTypeXFaceThread_xy_r8(I),mpiRC) |
368 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
369 |
eeBootError = .TRUE. |
370 |
WRITE(msgBuf,'(A,I)') |
371 |
& 'S/R INI_THREADS: MPI_TYPE_COMMIT (mpiTypeXFaceThread_xy_r8)', |
372 |
& mpiRC |
373 |
CALL PRINT_ERROR( msgBuf , myThid) |
374 |
ENDIF |
375 |
|
376 |
C x-face exchanges for xyz real*4 data |
377 |
elCount = myByHi(I)-myByLo(I)+1 |
378 |
elLen = 1 |
379 |
arrElSep = (sNx+OLx*2)*(sNy+OLy*2)*Nz*nSx |
380 |
arrElSize = 4 |
381 |
elStride = arrElSep*arrElSize |
382 |
CALL MPI_TYPE_HVECTOR(elCount,elLen,elStride,mpiTypeXFaceBlock_xyz_r4, |
383 |
O mpiTypeXFaceThread_xyz_r4(I), mpiRC ) |
384 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
385 |
eeBootError = .TRUE. |
386 |
WRITE(msgBuf,'(A,I)') |
387 |
& 'S/R INI_THREADS: MPI_TYPE_HVECTOR (mpiTypeXFaceThread_xyz_r4)', |
388 |
& mpiRC |
389 |
CALL PRINT_ERROR( msgBuf , myThid) |
390 |
ENDIF |
391 |
CALL MPI_TYPE_COMMIT(mpiTypeXFaceThread_xyz_r4(I),mpiRC) |
392 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
393 |
eeBootError = .TRUE. |
394 |
WRITE(msgBuf,'(A,I)') |
395 |
& 'S/R INI_THREADS: MPI_TYPE_COMMIT (mpiTypeXFaceThread_xyz_r4)', |
396 |
& mpiRC |
397 |
CALL PRINT_ERROR( msgBuf , myThid) |
398 |
ENDIF |
399 |
|
400 |
C x-face exchanges for xyz real*8 data |
401 |
arrElSize = 8 |
402 |
elStride = arrElSep*arrElSize |
403 |
CALL MPI_TYPE_HVECTOR(elCount,elLen,elStride,mpiTypeXFaceBlock_xyz_r8, |
404 |
O mpiTypeXFaceThread_xyz_r8(I), mpiRC ) |
405 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
406 |
eeBootError = .TRUE. |
407 |
WRITE(msgBuf,'(A,I)') |
408 |
& 'S/R INI_THREADS: MPI_TYPE_HVECTOR (mpiTypeXFaceThread_xyz_r8)', |
409 |
& mpiRC |
410 |
CALL PRINT_ERROR( msgBuf , myThid) |
411 |
ENDIF |
412 |
CALL MPI_TYPE_COMMIT(mpiTypeXFaceThread_xyz_r8(I),mpiRC) |
413 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
414 |
eeBootError = .TRUE. |
415 |
WRITE(msgBuf,'(A,I)') |
416 |
& 'S/R INI_THREADS: MPI_TYPE_COMMIT (mpiTypeXFaceThread_xyz_r8)', |
417 |
& mpiRC |
418 |
CALL PRINT_ERROR( msgBuf , myThid) |
419 |
ENDIF |
420 |
|
421 |
C y-face exchages for xy real*4 data |
422 |
elCount = myBxHi(I)-myBxLo(I)+1 |
423 |
elLen = 1 |
424 |
arrElSep = (sNx+OLx*2)*(sNy+OLy*2) |
425 |
arrElSize = 4 |
426 |
elStride = arrElSep*arrElSize |
427 |
CALL MPI_TYPE_HVECTOR(elCount,elLen,elStride,mpiTypeYFaceBlock_xy_r4, |
428 |
O mpiTypeYFaceThread_xy_r4(I), mpiRC ) |
429 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
430 |
eeBootError = .TRUE. |
431 |
WRITE(msgBuf,'(A,I)') |
432 |
& 'S/R INI_THREADS: MPI_TYPE_HVECTOR (mpiTypeYFaceThread_xy_r4)', |
433 |
& mpiRC |
434 |
CALL PRINT_ERROR( msgBuf , myThid) |
435 |
ENDIF |
436 |
CALL MPI_TYPE_COMMIT(mpiTypeYFaceThread_xy_r4(I),mpiRC) |
437 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
438 |
eeBootError = .TRUE. |
439 |
WRITE(msgBuf,'(A,I)') |
440 |
& 'S/R INI_THREADS: MPI_TYPE_COMMIT (mpiTypeYFaceThread_xy_r4)', |
441 |
& mpiRC |
442 |
CALL PRINT_ERROR( msgBuf , myThid) |
443 |
ENDIF |
444 |
|
445 |
C y-face exchages for xy real*8 data |
446 |
arrElSize = 8 |
447 |
elStride = arrElSep*arrElSize |
448 |
CALL MPI_TYPE_HVECTOR(elCount,elLen,elStride,mpiTypeYFaceBlock_xy_r8, |
449 |
O mpiTypeYFaceThread_xy_r8(I), mpiRC ) |
450 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
451 |
eeBootError = .TRUE. |
452 |
WRITE(msgBuf,'(A,I)') |
453 |
& 'S/R INI_THREADS: MPI_TYPE_HVECTOR (mpiTypeYFaceThread_xy_r8)', |
454 |
& mpiRC |
455 |
CALL PRINT_ERROR( msgBuf , myThid) |
456 |
ENDIF |
457 |
CALL MPI_TYPE_COMMIT(mpiTypeYFaceThread_xy_r8(I),mpiRC) |
458 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
459 |
eeBootError = .TRUE. |
460 |
WRITE(msgBuf,'(A,I)') |
461 |
& 'S/R INI_THREADS: MPI_TYPE_COMMIT (mpiTypeYFaceThread_xy_r8)', |
462 |
& mpiRC |
463 |
CALL PRINT_ERROR( msgBuf , myThid) |
464 |
ENDIF |
465 |
|
466 |
C y-face exchages for xyz real*4 data |
467 |
elCount = myBxHi(I)-myBxLo(I)+1 |
468 |
elLen = 1 |
469 |
arrElSep = (sNx+OLx*2)*(sNy+OLy*2)*Nz |
470 |
arrElSize = 4 |
471 |
elStride = arrElSep*arrElSize |
472 |
CALL MPI_TYPE_HVECTOR(elCount,elLen,elStride,mpiTypeYFaceBlock_xyz_r4, |
473 |
O mpiTypeYFaceThread_xyz_r4(I), mpiRC ) |
474 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
475 |
eeBootError = .TRUE. |
476 |
WRITE(msgBuf,'(A,I)') |
477 |
& 'S/R INI_THREADS: MPI_TYPE_HVECTOR (mpiTypeYFaceThread_xyz_r4)', |
478 |
& mpiRC |
479 |
CALL PRINT_ERROR( msgBuf , myThid) |
480 |
ENDIF |
481 |
CALL MPI_TYPE_COMMIT(mpiTypeYFaceThread_xyz_r4(I),mpiRC) |
482 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
483 |
eeBootError = .TRUE. |
484 |
WRITE(msgBuf,'(A,I)') |
485 |
& 'S/R INI_THREADS: MPI_TYPE_COMMIT (mpiTypeYFaceThread_xyz_r4)', |
486 |
& mpiRC |
487 |
CALL PRINT_ERROR( msgBuf , myThid) |
488 |
ENDIF |
489 |
|
490 |
C y-face exchages for xy real*8 data |
491 |
arrElSize = 8 |
492 |
elStride = arrElSep*arrElSize |
493 |
CALL MPI_TYPE_HVECTOR(elCount,elLen,elStride,mpiTypeYFaceBlock_xyz_r8, |
494 |
O mpiTypeYFaceThread_xyz_r8(I), mpiRC ) |
495 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
496 |
eeBootError = .TRUE. |
497 |
WRITE(msgBuf,'(A,I)') |
498 |
& 'S/R INI_THREADS: MPI_TYPE_HVECTOR (mpiTypeYFaceThread_xyz_r8)', |
499 |
& mpiRC |
500 |
CALL PRINT_ERROR( msgBuf , myThid) |
501 |
ENDIF |
502 |
CALL MPI_TYPE_COMMIT(mpiTypeYFaceThread_xyz_r8(I),mpiRC) |
503 |
IF ( mpiRC .NE. MPI_SUCCESS ) THEN |
504 |
eeBootError = .TRUE. |
505 |
WRITE(msgBuf,'(A,I)') |
506 |
& 'S/R INI_THREADS: MPI_TYPE_COMMIT (mpiTypeYFaceThread_xyz_r8)', |
507 |
& mpiRC |
508 |
CALL PRINT_ERROR( msgBuf , myThid) |
509 |
ENDIF |
510 |
|
511 |
ENDDO |
512 |
|
513 |
#ifndef ALWAYS_USE_MPI |
514 |
ENDIF |
515 |
#endif |
516 |
#endif /* ALLOW_USE_MPI */ |
517 |
|
518 |
C-- Calculate the thread numbers of the threads I might want to "message" |
519 |
C Notes: |
520 |
C 1. This code needs to be consistent with the code that maps threads to |
521 |
C blocks earlier in this routine in which threads are arranged |
522 |
C 13 14 15 16 /|\ |
523 |
C 9 10 11 12 | nTy |
524 |
C 5 6 7 8 | |
525 |
C 1 2 3 4 \|/ |
526 |
C <---- nTx ---> |
527 |
C on equally sized collections of sNx x sNy sub-blocks. |
528 |
DO I = 1, nThreads |
529 |
C Find thread to west, east, south, north using wrap around for |
530 |
C threads managing "outside" blocks. |
531 |
threadWest = I-1 |
532 |
IF ( myBxLo(I) .EQ. 1 ) threadWest = I+nTx-1 |
533 |
threadEast = I+1 |
534 |
IF ( myBxHi(I) .EQ. nSx ) threadEast = I-nTx+1 |
535 |
threadSouth = I-nTx |
536 |
IF ( myByLo(I) .EQ. 1 ) threadSouth = I+nTx*(nTy-1) |
537 |
threadNorth = I+nTx |
538 |
IF ( myByHi(I) .EQ. nSy ) threadNorth = I-nTx*(nTy-1) |
539 |
C Find thread to NW, NE, SW, SE - again with wrap around. |
540 |
threadNW = threadWest+nTx |
541 |
IF ( myByHi(threadWest) .EQ. nSy ) threadNW = threadWest-nTx*(nTy-1) |
542 |
threadNE = threadEast+nTx |
543 |
IF ( myByHi(threadEast) .EQ. nSy ) threadNE = threadEast-nTx*(nTy-1) |
544 |
threadSW = threadWest-nTx |
545 |
IF ( myByHi(threadWest) .EQ. 1 ) threadSW = threadWest+nTx*(nTy-1) |
546 |
threadSE = threadEast-nTx |
547 |
IF ( myByHi(threadEast) .EQ. 1 ) threadSE = threadEast+nTx*(nTy-1) |
548 |
myThrW(I) = threadWest |
549 |
myThrE(I) = threadEast |
550 |
myThrN(I) = threadNorth |
551 |
myThrS(I) = threadSouth |
552 |
myThrNW(I) = threadNW |
553 |
myThrNE(I) = threadNE |
554 |
myThrSW(I) = threadSW |
555 |
myThrSE(I) = threadSE |
556 |
ENDDO |
557 |
|
558 |
RETURN |
559 |
END |
560 |
|
561 |
C $Id: $ |