1 |
C $Header: /u/gcmpack/models/MITgcmUV/eesupp/src/exch_rx_send_put_y.template,v 1.1 2001/05/29 14:06:38 adcroft Exp $ |
2 |
C $Name: $ |
3 |
#include "CPP_EEOPTIONS.h" |
4 |
|
5 |
CBOP |
6 |
C !ROUTINE: EXCH_RX_SEND_PUT_Y |
7 |
|
8 |
C !INTERFACE: |
9 |
SUBROUTINE EXCH_RX_SEND_PUT_Y( array, |
10 |
I myOLw, myOLe, myOLs, myOLn, myNz, |
11 |
I exchWidthX, exchWidthY, |
12 |
I thesimulationMode, thecornerMode, myThid ) |
13 |
IMPLICIT NONE |
14 |
C !DESCRIPTION: |
15 |
C *==========================================================* |
16 |
C | SUBROUTINE SEND_PUT_Y |
17 |
C | o "Send" or "put" Y edges for RX array. |
18 |
C *==========================================================* |
19 |
C | Routine that invokes actual message passing send or |
20 |
C | direct "put" of data to update Y faces of an XY[R] array. |
21 |
C *==========================================================* |
22 |
|
23 |
C !USES: |
24 |
C == Global variables == |
25 |
#include "SIZE.h" |
26 |
#include "EEPARAMS.h" |
27 |
#include "EESUPPORT.h" |
28 |
#include "EXCH.h" |
29 |
|
30 |
C !INPUT/OUTPUT PARAMETERS: |
31 |
C == Routine arguments == |
32 |
C array :: Array with edges to exchange. |
33 |
C myOLw :: West, East, North and South overlap region sizes. |
34 |
C myOLe |
35 |
C myOLn |
36 |
C myOLs |
37 |
C exchWidthX :: Width of data region exchanged. |
38 |
C exchWidthY |
39 |
C theSimulationMode :: Forward or reverse mode exchange ( provides |
40 |
C support for adjoint integration of code. ) |
41 |
C Note - the reverse mode for an assignment |
42 |
C is an accumulation. This means that |
43 |
C put implementations that do leary things |
44 |
C like writing to overlap regions in a |
45 |
C remote process need to be even more |
46 |
C careful. You need to be pretty careful |
47 |
C in forward mode too! |
48 |
C theCornerMode :: Flag indicating whether corner updates are |
49 |
C needed. |
50 |
C myThid :: Thread number of this instance of S/R EXCH... |
51 |
C eBl :: Edge buffer level |
52 |
INTEGER myOLw |
53 |
INTEGER myOLe |
54 |
INTEGER myOLs |
55 |
INTEGER myOLn |
56 |
INTEGER myNz |
57 |
_RX array(1-myOLw:sNx+myOLe, |
58 |
& 1-myOLs:sNy+myOLn, |
59 |
& myNZ, nSx, nSy) |
60 |
INTEGER exchWidthX |
61 |
INTEGER exchWidthY |
62 |
INTEGER theSimulationMode |
63 |
INTEGER theCornerMode |
64 |
INTEGER myThid |
65 |
|
66 |
C !LOCAL VARIABLES: |
67 |
C == Local variables == |
68 |
C I, J, K, jMin, jMax, iB - Loop counters and extents |
69 |
C bi, bj |
70 |
C biS, bjS - South tile indices |
71 |
C biN, bjN - North tile indices |
72 |
C eBl - Current exchange buffer level |
73 |
C theProc, theTag, theType, - Variables used in message building |
74 |
C theSize |
75 |
C southCommMode - Working variables holding type |
76 |
C northCommMode of communication a particular |
77 |
C tile face uses. |
78 |
INTEGER I, J, K, jMin, jMax, iMin, iMax, iB |
79 |
INTEGER bi, bj, biS, bjS, biN, bjN |
80 |
INTEGER eBl |
81 |
INTEGER northCommMode |
82 |
INTEGER southCommMode |
83 |
#ifdef ALLOW_USE_MPI |
84 |
INTEGER theProc, theTag, theType, theSize, mpiRc |
85 |
#endif |
86 |
|
87 |
C-- Write data to exchange buffer |
88 |
C Various actions are possible depending on the communication mode |
89 |
C as follows: |
90 |
C Mode Action |
91 |
C -------- --------------------------- |
92 |
C COMM_NONE Do nothing |
93 |
C |
94 |
C COMM_MSG Message passing communication ( e.g. MPI ) |
95 |
C Fill south send buffer from this tile. |
96 |
C Send data with tag identifying tile and direction. |
97 |
C Fill north send buffer from this tile. |
98 |
C Send data with tag identifying tile and direction. |
99 |
C |
100 |
C COMM_PUT "Put" communication ( UMP_, shmemput, etc... ) |
101 |
C Fill south receive buffer of south-neighbor tile |
102 |
C Fill north receive buffer of north-neighbor tile |
103 |
C Sync. memory |
104 |
C Write data-ready Ack for north edge of south-neighbor |
105 |
C tile |
106 |
C Write data-ready Ack for south edge of north-neighbor |
107 |
C tile |
108 |
C Sync. memory |
109 |
CEOP |
110 |
|
111 |
C |
112 |
DO bj=myByLo(myThid),myByHi(myThid) |
113 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
114 |
|
115 |
ebL = exchangeBufLevel(1,bi,bj) |
116 |
southCommMode = _tileCommModeS(bi,bj) |
117 |
northCommMode = _tileCommModeN(bi,bj) |
118 |
biS = _tileBiS(bi,bj) |
119 |
bjS = _tileBjS(bi,bj) |
120 |
biN = _tileBiN(bi,bj) |
121 |
bjN = _tileBjN(bi,bj) |
122 |
iMin = 1 |
123 |
iMax = sNx |
124 |
IF ( theCornerMode .EQ. EXCH_UPDATE_CORNERS ) THEN |
125 |
iMin = 1-exchWidthX |
126 |
iMax = sNx+exchWidthX |
127 |
ENDIF |
128 |
|
129 |
|
130 |
c >>>>>>>>>>>>>>>>>>> <<<<<<<<<<<<<<<<<<< |
131 |
c >>>>>>>>>>>>>>>>>>> FORWARD RUN <<<<<<<<<<<<<<<<<<< |
132 |
c >>>>>>>>>>>>>>>>>>> <<<<<<<<<<<<<<<<<<< |
133 |
|
134 |
C o Send or Put south edge |
135 |
IF ( _theSimulationMode .EQ. FORWARD_SIMULATION ) THEN |
136 |
jMin = 1 |
137 |
jMax = 1+exchWidthY-1 |
138 |
IF ( southCommMode .EQ. COMM_MSG ) THEN |
139 |
iB = 0 |
140 |
DO K=1,myNz |
141 |
DO J=jMin,jMax |
142 |
DO I=iMin,iMax |
143 |
iB = iB + 1 |
144 |
southSendBuf_RX(iB,eBl,bi,bj) = array(I,J,K,bi,bj) |
145 |
ENDDO |
146 |
ENDDO |
147 |
ENDDO |
148 |
C Send the data |
149 |
#ifdef ALLOW_USE_MPI |
150 |
#ifndef ALWAYS_USE_MPI |
151 |
IF ( usingMPI ) THEN |
152 |
#endif |
153 |
theProc = tilePidS(bi,bj) |
154 |
theTag = _tileTagSendS(bi,bj) |
155 |
theSize = iB |
156 |
theType = MPI_DOUBLE_PRECISION |
157 |
exchNreqsY(1,bi,bj) = exchNreqsY(1,bi,bj)+1 |
158 |
CALL MPI_Isend(southSendBuf_RX(1,eBl,bi,bj), theSize, theType, |
159 |
& theProc, theTag, MPI_COMM_MODEL, |
160 |
& exchReqIdY(exchNreqsY(1,bi,bj),1,bi,bj), mpiRc) |
161 |
#ifndef ALWAYS_USE_MPI |
162 |
ENDIF |
163 |
#endif |
164 |
#endif /* ALLOW_USE_MPI */ |
165 |
northRecvAck(eBl,biS,bjS) = 1. |
166 |
ELSEIF ( southCommMode .EQ. COMM_PUT ) THEN |
167 |
iB = 0 |
168 |
DO K=1,myNz |
169 |
DO J=jMin,jMax |
170 |
DO I=iMin,iMax |
171 |
iB = iB + 1 |
172 |
northRecvBuf_RX(iB,eBl,biS,bjS) = array(I,J,K,bi,bj) |
173 |
ENDDO |
174 |
ENDDO |
175 |
ENDDO |
176 |
ELSEIF ( southCommMode .NE. COMM_NONE |
177 |
& .AND. southCommMode .NE. COMM_GET ) THEN |
178 |
STOP ' S/R EXCH: Invalid commS mode.' |
179 |
ENDIF |
180 |
|
181 |
C o Send or Put north edge |
182 |
jMin = sNy-exchWidthY+1 |
183 |
jMax = sNy |
184 |
IF ( northCommMode .EQ. COMM_MSG ) THEN |
185 |
iB = 0 |
186 |
DO K=1,myNz |
187 |
DO J=jMin,jMax |
188 |
DO I=iMin,iMax |
189 |
iB = iB + 1 |
190 |
northSendBuf_RX(iB,eBl,bi,bj) = array(I,J,K,bi,bj) |
191 |
ENDDO |
192 |
ENDDO |
193 |
ENDDO |
194 |
C Send the data |
195 |
#ifdef ALLOW_USE_MPI |
196 |
#ifndef ALWAYS_USE_MPI |
197 |
IF ( usingMPI ) THEN |
198 |
#endif |
199 |
theProc = tilePidN(bi,bj) |
200 |
theTag = _tileTagSendN(bi,bj) |
201 |
theSize = iB |
202 |
theType = MPI_DOUBLE_PRECISION |
203 |
exchNreqsY(1,bi,bj) = exchNreqsY(1,bi,bj)+1 |
204 |
CALL MPI_Isend(northSendBuf_RX(1,eBl,bi,bj), theSize, theType, |
205 |
& theProc, theTag, MPI_COMM_MODEL, |
206 |
& exchReqIdY(exchNreqsY(1,bi,bj),1,bi,bj), mpiRc ) |
207 |
#ifndef ALWAYS_USE_MPI |
208 |
ENDIF |
209 |
#endif |
210 |
#endif /* ALLOW_USE_MPI */ |
211 |
southRecvAck(eBl,biN,bjN) = 1. |
212 |
ELSEIF ( northCommMode .EQ. COMM_PUT ) THEN |
213 |
iB = 0 |
214 |
DO K=1,myNz |
215 |
DO J=jMin,jMax |
216 |
DO I=iMin,iMax |
217 |
iB = iB + 1 |
218 |
southRecvBuf_RX(iB,eBl,biN,bjN) = array(I,J,K,bi,bj) |
219 |
ENDDO |
220 |
ENDDO |
221 |
ENDDO |
222 |
ELSEIF ( northCommMode .NE. COMM_NONE |
223 |
& .AND. northCommMode .NE. COMM_GET ) THEN |
224 |
STOP ' S/R EXCH: Invalid commN mode.' |
225 |
ENDIF |
226 |
|
227 |
c >>>>>>>>>>>>>>>>>>> <<<<<<<<<<<<<<<<<<< |
228 |
c >>>>>>>>>>>>>>>>>>> ADJOINT RUN <<<<<<<<<<<<<<<<<<< |
229 |
c >>>>>>>>>>>>>>>>>>> <<<<<<<<<<<<<<<<<<< |
230 |
|
231 |
ELSEIF ( _theSimulationMode .EQ. REVERSE_SIMULATION ) THEN |
232 |
jMin = 1-exchWidthY |
233 |
jMax = 0 |
234 |
IF ( southCommMode .EQ. COMM_MSG ) THEN |
235 |
iB = 0 |
236 |
DO K=1,myNz |
237 |
DO J=jMin,jMax |
238 |
DO I=iMin,iMax |
239 |
iB = iB + 1 |
240 |
southSendBuf_RX(iB,eBl,bi,bj) = array(I,J,K,bi,bj) |
241 |
array(I,J,K,bi,bj) = 0.0 |
242 |
ENDDO |
243 |
ENDDO |
244 |
ENDDO |
245 |
C Send the data |
246 |
#ifdef ALLOW_USE_MPI |
247 |
#ifndef ALWAYS_USE_MPI |
248 |
IF ( usingMPI ) THEN |
249 |
#endif |
250 |
theProc = tilePidS(bi,bj) |
251 |
theTag = _tileTagSendS(bi,bj) |
252 |
theSize = iB |
253 |
theType = MPI_DOUBLE_PRECISION |
254 |
exchNreqsY(1,bi,bj) = exchNreqsY(1,bi,bj)+1 |
255 |
CALL MPI_Isend(southSendBuf_RX(1,eBl,bi,bj), theSize, theType, |
256 |
& theProc, theTag, MPI_COMM_MODEL, |
257 |
& exchReqIdY(exchNreqsY(1,bi,bj),1,bi,bj), mpiRc) |
258 |
#ifndef ALWAYS_USE_MPI |
259 |
ENDIF |
260 |
#endif |
261 |
#endif /* ALLOW_USE_MPI */ |
262 |
northRecvAck(eBl,biS,bjS) = 1. |
263 |
ELSEIF ( southCommMode .EQ. COMM_PUT ) THEN |
264 |
iB = 0 |
265 |
DO K=1,myNz |
266 |
DO J=jMin,jMax |
267 |
DO I=iMin,iMax |
268 |
iB = iB + 1 |
269 |
northRecvBuf_RX(iB,eBl,biS,bjS) = array(I,J,K,bi,bj) |
270 |
array(I,J,K,bi,bj) = 0.0 |
271 |
ENDDO |
272 |
ENDDO |
273 |
ENDDO |
274 |
ELSEIF ( southCommMode .NE. COMM_NONE |
275 |
& .AND. southCommMode .NE. COMM_GET ) THEN |
276 |
STOP ' S/R EXCH: Invalid commS mode.' |
277 |
ENDIF |
278 |
|
279 |
C o Send or Put north edge |
280 |
jMin = sNy+1 |
281 |
jMax = sNy+exchWidthY |
282 |
IF ( northCommMode .EQ. COMM_MSG ) THEN |
283 |
iB = 0 |
284 |
DO K=1,myNz |
285 |
DO J=jMin,jMax |
286 |
DO I=iMin,iMax |
287 |
iB = iB + 1 |
288 |
northSendBuf_RX(iB,eBl,bi,bj) = array(I,J,K,bi,bj) |
289 |
array(I,J,K,bi,bj) = 0.0 |
290 |
ENDDO |
291 |
ENDDO |
292 |
ENDDO |
293 |
C Send the data |
294 |
#ifdef ALLOW_USE_MPI |
295 |
#ifndef ALWAYS_USE_MPI |
296 |
IF ( usingMPI ) THEN |
297 |
#endif |
298 |
theProc = tilePidN(bi,bj) |
299 |
theTag = _tileTagSendN(bi,bj) |
300 |
theSize = iB |
301 |
theType = MPI_DOUBLE_PRECISION |
302 |
exchNreqsY(1,bi,bj) = exchNreqsY(1,bi,bj)+1 |
303 |
CALL MPI_Isend(northSendBuf_RX(1,eBl,bi,bj), theSize, theType, |
304 |
& theProc, theTag, MPI_COMM_MODEL, |
305 |
& exchReqIdY(exchNreqsY(1,bi,bj),1,bi,bj), mpiRc ) |
306 |
#ifndef ALWAYS_USE_MPI |
307 |
ENDIF |
308 |
#endif |
309 |
#endif /* ALLOW_USE_MPI */ |
310 |
southRecvAck(eBl,biN,bjN) = 1. |
311 |
ELSEIF ( northCommMode .EQ. COMM_PUT ) THEN |
312 |
iB = 0 |
313 |
DO K=1,myNz |
314 |
DO J=jMin,jMax |
315 |
DO I=iMin,iMax |
316 |
iB = iB + 1 |
317 |
southRecvBuf_RX(iB,eBl,biN,bjN) = array(I,J,K,bi,bj) |
318 |
array(I,J,K,bi,bj) = 0.0 |
319 |
ENDDO |
320 |
ENDDO |
321 |
ENDDO |
322 |
ELSEIF ( northCommMode .NE. COMM_NONE |
323 |
& .AND. northCommMode .NE. COMM_GET ) THEN |
324 |
STOP ' S/R EXCH: Invalid commN mode.' |
325 |
ENDIF |
326 |
endif |
327 |
ENDDO |
328 |
ENDDO |
329 |
|
330 |
C-- Signal completetion ( making sure system-wide memory state is |
331 |
C-- consistent ). |
332 |
|
333 |
C ** NOTE ** We are relying on being able to produce strong-ordered |
334 |
C memory semantics here. In other words we assume that there is a |
335 |
C mechanism which can ensure that by the time the Ack is seen the |
336 |
C overlap region data that will be exchanged is up to date. |
337 |
IF ( exchNeedsMemSync ) CALL MEMSYNC |
338 |
|
339 |
DO bj=myByLo(myThid),myByHi(myThid) |
340 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
341 |
ebL = exchangeBufLevel(1,bi,bj) |
342 |
biS = _tileBiS(bi,bj) |
343 |
bjS = _tileBjS(bi,bj) |
344 |
biN = _tileBiN(bi,bj) |
345 |
bjN = _tileBjN(bi,bj) |
346 |
southCommMode = _tileCommModeS(bi,bj) |
347 |
northCommMode = _tileCommModeN(bi,bj) |
348 |
IF ( southCommMode .EQ. COMM_PUT ) northRecvAck(eBl,biS,bjS) = 1. |
349 |
IF ( northCommMode .EQ. COMM_PUT ) southRecvAck(eBl,biN,bjN) = 1. |
350 |
IF ( southCommMode .EQ. COMM_GET ) northRecvAck(eBl,biS,bjS) = 1. |
351 |
IF ( northCommMode .EQ. COMM_GET ) southRecvAck(eBl,biN,bjN) = 1. |
352 |
ENDDO |
353 |
ENDDO |
354 |
|
355 |
C-- Make sure "ack" setting is seen system-wide. |
356 |
C Here strong-ordering is not an issue but we want to make |
357 |
C sure that processes that might spin on the above Ack settings |
358 |
C will see the setting. |
359 |
C ** NOTE ** On some machines we wont spin on the Ack setting |
360 |
C ( particularly the T90 ), instead we will use s system barrier. |
361 |
C On the T90 the system barrier is very fast and switches out the |
362 |
C thread while it waits. On most machines the system barrier |
363 |
C is much too slow and if we own the machine and have one thread |
364 |
C per process preemption is not a problem. |
365 |
IF ( exchNeedsMemSync ) CALL MEMSYNC |
366 |
|
367 |
RETURN |
368 |
END |