1 | MODULE mppsumtam |
---|
2 | !!====================================================================== |
---|
3 | !! *** MODULE mppsumtam *** |
---|
4 | !! NEMOTAM: Summation of arrays across processors |
---|
5 | !!====================================================================== |
---|
6 | |
---|
7 | !!---------------------------------------------------------------------- |
---|
8 | !! mpp_sum_inter : Interface for depending on nmppsum |
---|
9 | !! mpp_sum_simple : Simple sum |
---|
10 | !! loc_sum_inter : Interface for depending on nmppsum (local version) |
---|
11 | !! loc_sum_simple : Simple sum (local version) |
---|
12 | !! loc_sum_indep : Order independent MPP reproducible sum |
---|
13 | !!---------------------------------------------------------------------- |
---|
14 | !! * Modules used |
---|
15 | USE par_kind, ONLY : & ! Precision variables |
---|
16 | & wp |
---|
17 | USE dom_oce, ONLY : & ! Ocean space and time domain variables |
---|
18 | & nproc |
---|
19 | USE par_oce, ONLY : & ! Ocean parameters |
---|
20 | & jpnij |
---|
21 | #if defined key_mpp_mpi |
---|
22 | USE lib_mpp, ONLY : & ! MPP library |
---|
23 | & mpi_comm_opa, & |
---|
24 | & mpp_sum |
---|
25 | #endif |
---|
26 | USE lib_mpp |
---|
27 | USE mpp_tam |
---|
28 | USE mppsum |
---|
29 | USE in_out_manager |
---|
30 | |
---|
31 | IMPLICIT NONE |
---|
32 | |
---|
33 | !! * Routine accessibility |
---|
34 | PRIVATE |
---|
35 | |
---|
36 | PUBLIC & |
---|
37 | & mpp_sum_inter, & ! Interface for depending on nmppsum |
---|
38 | & mpp_sum_simple, & ! Simple sum |
---|
39 | & loc_sum_inter, & ! Interface for depending on nmppsum |
---|
40 | & loc_sum_simple, & ! Simple sum |
---|
41 | & loc_sum_indep, & ! Order independent local reproducible sum |
---|
42 | & nmppsum ! Summation control |
---|
43 | |
---|
44 | PUBLIC mpp_sum_nfd |
---|
45 | |
---|
46 | INTERFACE mpp_sum_nfd |
---|
47 | MODULE PROCEDURE mpp_sum_nfd_3d,mpp_sum_nfd_4d |
---|
48 | END INTERFACE |
---|
49 | |
---|
50 | !! * Module variables |
---|
51 | INTEGER :: nmppsum = 2 ! Choice of MPP sum for |
---|
52 | ! (1 = simple) |
---|
53 | ! (2 = order independent) |
---|
54 | |
---|
55 | CONTAINS |
---|
56 | |
---|
57 | FUNCTION mpp_sum_inter( pval, kn ) |
---|
58 | !!---------------------------------------------------------------------- |
---|
59 | !! *** ROUTINE mpp_sum_inter *** |
---|
60 | !! |
---|
61 | !! ** Purpose : Summation of arrays across processors |
---|
62 | !! |
---|
63 | !! ** Method : Call either mpp_sum_simple or mpp_sum_indep depending on |
---|
64 | !! nmppsum |
---|
65 | !! |
---|
66 | !! ** Action : This does only work for MPI. |
---|
67 | !! It does not work for SHMEM. |
---|
68 | !! |
---|
69 | !! References : http://www.mpi-forum.org |
---|
70 | !! |
---|
71 | !! History : |
---|
72 | !! ! 07-07 (K. Mogensen) Original code |
---|
73 | !!---------------------------------------------------------------------- |
---|
74 | !! * Function return |
---|
75 | REAL(wp) mpp_sum_inter |
---|
76 | !! * Arguments |
---|
77 | INTEGER, INTENT(IN) :: & |
---|
78 | & kn |
---|
79 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
80 | & pval |
---|
81 | !! * Local declarations |
---|
82 | |
---|
83 | IF ( nmppsum == 1 ) THEN |
---|
84 | |
---|
85 | mpp_sum_inter = mpp_sum_simple( pval, kn ) |
---|
86 | |
---|
87 | ELSEIF ( nmppsum == 2 ) THEN |
---|
88 | |
---|
89 | mpp_sum_inter = mpp_sum_indep( pval, kn ) |
---|
90 | |
---|
91 | ELSE |
---|
92 | |
---|
93 | CALL ctl_stop( ' mpp_sum: Invalid nmppsum ' ) |
---|
94 | |
---|
95 | ENDIF |
---|
96 | |
---|
97 | END FUNCTION mpp_sum_inter |
---|
98 | |
---|
99 | FUNCTION mpp_sum_simple( pval, kn ) |
---|
100 | !!---------------------------------------------------------------------- |
---|
101 | !! *** ROUTINE mpp_sum_simple *** |
---|
102 | !! |
---|
103 | !! ** Purpose : Summation of arrays across processors |
---|
104 | !! |
---|
105 | !! ** Method : This routine is the naive implementation using |
---|
106 | !! mpi_allreduce for the sum of local data |
---|
107 | !! |
---|
108 | !! ** Action : This does only work for MPI. |
---|
109 | !! It does not work for SHMEM. |
---|
110 | !! |
---|
111 | !! References : http://www.mpi-forum.org |
---|
112 | !! |
---|
113 | !! History : |
---|
114 | !! ! 07-07 (K. Mogensen) Original code |
---|
115 | !!---------------------------------------------------------------------- |
---|
116 | !! * Function return |
---|
117 | REAL(wp) mpp_sum_simple |
---|
118 | !! * Arguments |
---|
119 | INTEGER, INTENT(IN) :: & |
---|
120 | & kn |
---|
121 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
122 | & pval |
---|
123 | !! * Local declarations |
---|
124 | REAL(wp) :: & |
---|
125 | & zres |
---|
126 | INTEGER :: & |
---|
127 | & ierr |
---|
128 | INTEGER :: & |
---|
129 | & ji |
---|
130 | |
---|
131 | ! Compute local sum |
---|
132 | zres = SUM( pval(1:kn) ) |
---|
133 | ! Global sum |
---|
134 | |
---|
135 | CALL mpp_sum( zres ) |
---|
136 | |
---|
137 | mpp_sum_simple = zres |
---|
138 | |
---|
139 | END FUNCTION mpp_sum_simple |
---|
140 | |
---|
141 | FUNCTION loc_sum_inter( pval, kn ) |
---|
142 | !!---------------------------------------------------------------------- |
---|
143 | !! *** ROUTINE loc_sum_inter *** |
---|
144 | !! |
---|
145 | !! ** Purpose : Summation of arrays across processors |
---|
146 | !! |
---|
147 | !! ** Method : Call either loc_sum_simple or loc_sum_indep depending on |
---|
148 | !! nmppsum |
---|
149 | !! |
---|
150 | !! ** Action : |
---|
151 | !! |
---|
152 | !! References : |
---|
153 | !! |
---|
154 | !! History : |
---|
155 | !! ! 07-09 (K. Mogensen) Original code |
---|
156 | !!---------------------------------------------------------------------- |
---|
157 | !! * Function return |
---|
158 | REAL(wp) loc_sum_inter |
---|
159 | !! * Arguments |
---|
160 | INTEGER, INTENT(IN) :: & |
---|
161 | & kn |
---|
162 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
163 | & pval |
---|
164 | !! * Local declarations |
---|
165 | |
---|
166 | IF ( nmppsum == 1 ) THEN |
---|
167 | |
---|
168 | loc_sum_inter = loc_sum_simple( pval, kn ) |
---|
169 | |
---|
170 | ELSEIF ( nmppsum == 2 ) THEN |
---|
171 | |
---|
172 | loc_sum_inter = loc_sum_indep( pval, kn ) |
---|
173 | |
---|
174 | ELSE |
---|
175 | |
---|
176 | CALL ctl_stop( ' loc_sum: Invalid nmppsum = ' ) |
---|
177 | |
---|
178 | ENDIF |
---|
179 | |
---|
180 | END FUNCTION loc_sum_inter |
---|
181 | |
---|
182 | FUNCTION loc_sum_simple( pval, kn ) |
---|
183 | !!---------------------------------------------------------------------- |
---|
184 | !! *** ROUTINE loc_sum_simple *** |
---|
185 | !! |
---|
186 | !! ** Purpose : Summation of arrays across processors |
---|
187 | !! |
---|
188 | !! ** Method : This routine is the naive implementation for the sum |
---|
189 | !! of local data |
---|
190 | !! |
---|
191 | !! ** Action : |
---|
192 | !! |
---|
193 | !! References : |
---|
194 | !! |
---|
195 | !! History : |
---|
196 | !! ! 07-09 (K. Mogensen) Original code |
---|
197 | !!---------------------------------------------------------------------- |
---|
198 | !! * Function return |
---|
199 | REAL(wp) loc_sum_simple |
---|
200 | !! * Arguments |
---|
201 | INTEGER, INTENT(IN) :: & |
---|
202 | & kn |
---|
203 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
204 | & pval |
---|
205 | !! * Local declarations |
---|
206 | REAL(wp) :: & |
---|
207 | & zloc |
---|
208 | INTEGER :: & |
---|
209 | & ierr |
---|
210 | INTEGER :: & |
---|
211 | & ji |
---|
212 | |
---|
213 | ! Compute local sum |
---|
214 | |
---|
215 | zloc = SUM( pval(1:kn) ) |
---|
216 | |
---|
217 | loc_sum_simple = zloc |
---|
218 | |
---|
219 | END FUNCTION loc_sum_simple |
---|
220 | |
---|
221 | FUNCTION loc_sum_indep( pval, kn ) |
---|
222 | !!---------------------------------------------------------------------- |
---|
223 | !! *** ROUTINE loc_sum_indep *** |
---|
224 | !! |
---|
225 | !! ** Purpose : Sum all elements in the pval array in |
---|
226 | !! an accurate order-independent way. |
---|
227 | !! |
---|
228 | !! ** Method : The code iterates the compensated summation until the |
---|
229 | !! result is guaranteed to be within 4*eps of the true sum. |
---|
230 | !! It then rounds the result to the nearest floating-point |
---|
231 | !! number whose last three bits are zero, thereby |
---|
232 | !! guaranteeing an order-independent result. |
---|
233 | !! |
---|
234 | !! This version is used for local arrays. It is identical |
---|
235 | !! to the MPP version for clarity. |
---|
236 | !! |
---|
237 | !! ** Action : |
---|
238 | !! |
---|
239 | !! References : M. Fisher (ECMWF): IFS code + personal communication |
---|
240 | !! The algorithm is based on Ogita et al. (2005) |
---|
241 | !! SIAM J. Sci. Computing, Vol.26, No.6, pp1955-1988. |
---|
242 | !! This is based in turn on an algorithm |
---|
243 | !! by Knuth (1969, seminumerical algorithms). |
---|
244 | !! |
---|
245 | !! History : |
---|
246 | !! ! 07-09 (K. Mogensen) Original code heavily based on IFS. |
---|
247 | !!---------------------------------------------------------------------- |
---|
248 | !! * Function return |
---|
249 | REAL(wp) loc_sum_indep |
---|
250 | !! * Arguments |
---|
251 | INTEGER, INTENT(IN) :: & |
---|
252 | & kn |
---|
253 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
254 | & pval |
---|
255 | !! * Local declarations |
---|
256 | REAL(wp), DIMENSION(3) ::& |
---|
257 | & zbuffl |
---|
258 | REAL(wp), DIMENSION(:), ALLOCATABLE :: & |
---|
259 | & zpsums, & |
---|
260 | & zperrs, & |
---|
261 | & zpcors, & |
---|
262 | & zbuffg, & |
---|
263 | & zp |
---|
264 | REAL(wp) :: & |
---|
265 | & zcorr, & |
---|
266 | & zerr, & |
---|
267 | & zolderr, & |
---|
268 | & zbeta, & |
---|
269 | & zres |
---|
270 | INTEGER :: & |
---|
271 | & jj |
---|
272 | |
---|
273 | ! Check that the the algorithm can work |
---|
274 | |
---|
275 | IF ( ( REAL( 2 * kn ) * EPSILON( zres ) ) >= 1.0 ) THEN |
---|
276 | |
---|
277 | CALL ctl_stop( 'kn is too large to guarantee error bounds' ) |
---|
278 | |
---|
279 | ENDIF |
---|
280 | |
---|
281 | ALLOCATE( & |
---|
282 | & zp(MAX(kn,1)), & |
---|
283 | & zbuffg(1*SIZE(zbuffl)), & |
---|
284 | & zpsums(1), & |
---|
285 | & zperrs(1), & |
---|
286 | & zpcors(1) & |
---|
287 | & ) |
---|
288 | |
---|
289 | zolderr = HUGE(zerr) |
---|
290 | |
---|
291 | ! Copy the input array. This avoids some tricky indexing, at the |
---|
292 | ! expense of some inefficency. |
---|
293 | |
---|
294 | IF ( kn > 0 ) THEN |
---|
295 | |
---|
296 | zp(:) = pval(:) |
---|
297 | |
---|
298 | ELSE |
---|
299 | |
---|
300 | zp(1) = 0.0_wp |
---|
301 | |
---|
302 | ENDIF |
---|
303 | |
---|
304 | k_loop: DO |
---|
305 | |
---|
306 | ! Transform local arrays |
---|
307 | |
---|
308 | IF ( kn > 0 ) THEN |
---|
309 | |
---|
310 | CALL comp_sum ( zp, kn, zcorr, zerr ) |
---|
311 | |
---|
312 | ENDIF |
---|
313 | |
---|
314 | ! Gather partial sums and error bounds to all processors |
---|
315 | |
---|
316 | zbuffl(1) = zp(MAX(kn,1)) |
---|
317 | |
---|
318 | IF ( kn > 0 ) THEN |
---|
319 | |
---|
320 | zbuffl(2) = zerr |
---|
321 | zbuffl(3) = zcorr |
---|
322 | |
---|
323 | ELSE |
---|
324 | |
---|
325 | zbuffl(2) = 0.0_wp |
---|
326 | zbuffl(3) = 0.0_wp |
---|
327 | |
---|
328 | ENDIF |
---|
329 | |
---|
330 | zpsums(1) = zbuffl(1) |
---|
331 | zperrs(1) = zbuffl(2) |
---|
332 | zpcors(1) = zbuffl(3) |
---|
333 | |
---|
334 | ! Transform partial sums |
---|
335 | |
---|
336 | CALL comp_sum( zpsums, 1, zcorr, zerr ) |
---|
337 | zerr = zerr + SUM(zperrs) |
---|
338 | zcorr = zcorr + SUM(zpcors) |
---|
339 | |
---|
340 | ! Calculate final result |
---|
341 | |
---|
342 | zres = zpsums(1) + zcorr |
---|
343 | |
---|
344 | ! Calculate error bound. This is corollary 4.7 from Ogita et al. |
---|
345 | ! (2005) |
---|
346 | |
---|
347 | zbeta = zerr *( REAL( 2*kn, wp ) * EPSILON(zres) ) & |
---|
348 | & /(1.0_wp - REAL( 2*kn, wp ) * EPSILON(zres) ) |
---|
349 | |
---|
350 | zerr = EPSILON(zres) * ABS(zres) & |
---|
351 | & +(zbeta + ( 2.0_wp * EPSILON(zres) * EPSILON(zres) * ABS(zres) & |
---|
352 | & +3.0_wp * TINY(zres) ) ) |
---|
353 | |
---|
354 | ! Update the last element of the local array |
---|
355 | |
---|
356 | zp(MAX(kn,1)) = zpsums(nproc+1) |
---|
357 | |
---|
358 | ! Exit if the global error is small enough |
---|
359 | |
---|
360 | IF ( zerr < 4.0_wp * SPACING(zres) ) EXIT k_loop |
---|
361 | |
---|
362 | ! Take appropriate action if ZRES cannot be sufficiently refined. |
---|
363 | |
---|
364 | IF (zerr >= zolderr) THEN |
---|
365 | |
---|
366 | CALL ctl_stop( 'Failed to refine sum', & |
---|
367 | & 'Warning: Possiblity of non-reproducible results' ) |
---|
368 | |
---|
369 | ENDIF |
---|
370 | |
---|
371 | zolderr = zerr |
---|
372 | |
---|
373 | ENDDO k_loop |
---|
374 | |
---|
375 | ! At this stage, we have guaranteed that ZRES less than 4*EPS |
---|
376 | ! away from the exact sum. There are only four floating point |
---|
377 | ! numbers in this range. So, if we find the nearest number that |
---|
378 | ! has its last three bits zero, then we have a reproducible result. |
---|
379 | |
---|
380 | loc_sum_indep = fround(zres) |
---|
381 | |
---|
382 | DEALLOCATE( & |
---|
383 | & zpcors, & |
---|
384 | & zperrs, & |
---|
385 | & zpsums, & |
---|
386 | & zbuffg, & |
---|
387 | & zp & |
---|
388 | & ) |
---|
389 | |
---|
390 | END FUNCTION loc_sum_indep |
---|
391 | |
---|
392 | #if defined key_mpp_mpi |
---|
393 | FUNCTION mpp_sum_nfd_4d( pval, kn1, kn2, kn3, kn4, kcom ) |
---|
394 | !!---------------------------------------------------------------------- |
---|
395 | !! *** ROUTINE mpp_sum_nfd_4d *** |
---|
396 | !! |
---|
397 | !! ** Purpose : Summation of arrays across processors (of kcom group) |
---|
398 | !! |
---|
399 | !! ** Method : Pack 4d array to vector, sum vector and then unpack |
---|
400 | !! |
---|
401 | !! ** Action : This does only work for MPI. |
---|
402 | !! It does not work for SHMEM. |
---|
403 | !! |
---|
404 | !! History : |
---|
405 | !! ! 10-11 (F. Vigilant) Original code |
---|
406 | !!---------------------------------------------------------------------- |
---|
407 | !! * Function return |
---|
408 | REAL(wp), DIMENSION(kn1, kn2, kn3, kn4) :: mpp_sum_nfd_4d |
---|
409 | !! * Arguments |
---|
410 | INTEGER, INTENT(IN) :: kn1, kn2, kn3, kn4 |
---|
411 | REAL(wp), DIMENSION(kn1, kn2, kn3, kn4), INTENT(IN) :: pval |
---|
412 | INTEGER , INTENT( in ), OPTIONAL :: kcom |
---|
413 | !! * Local declarations |
---|
414 | REAL(wp), DIMENSION(:), ALLOCATABLE :: zvec |
---|
415 | LOGICAL, DIMENSION(kn1, kn2, kn3, kn4) :: zmask |
---|
416 | REAL(wp), DIMENSION(kn1, kn2, kn3, kn4) :: zfd |
---|
417 | INTEGER :: zdim |
---|
418 | |
---|
419 | zdim = kn1 * kn2 * kn3 * kn4 |
---|
420 | ALLOCATE( zvec(zdim) ) |
---|
421 | |
---|
422 | zvec = PACK( pval(1:kn1,1:kn2,1:kn3,1:kn4),.TRUE.) |
---|
423 | |
---|
424 | CALL mpp_sum( zvec, zdim, kcom ) |
---|
425 | |
---|
426 | zmask(:,:,:,:) = .TRUE. |
---|
427 | mpp_sum_nfd_4d = UNPACK( zvec, zmask, zfd ) |
---|
428 | |
---|
429 | END FUNCTION mpp_sum_nfd_4d |
---|
430 | |
---|
431 | FUNCTION mpp_sum_nfd_3d( pval, kn1, kn2, kn3, kcom ) |
---|
432 | !!---------------------------------------------------------------------- |
---|
433 | !! *** ROUTINE mpp_sum_nfd_3d *** |
---|
434 | !! |
---|
435 | !! ** Purpose : Summation of arrays across processors (of kcom group) |
---|
436 | !! |
---|
437 | !! ** Method : Pack 3d array to vector, sum vector and then unpack |
---|
438 | !! |
---|
439 | !! ** Action : This does only work for MPI. |
---|
440 | !! It does not work for SHMEM. |
---|
441 | !! |
---|
442 | !! History : |
---|
443 | !! ! 10-11 (F. Vigilant) Original code |
---|
444 | !!---------------------------------------------------------------------- |
---|
445 | !! * Function return |
---|
446 | REAL(wp), DIMENSION(kn1, kn2, kn3) :: mpp_sum_nfd_3d |
---|
447 | !! * Arguments |
---|
448 | INTEGER, INTENT(IN) ::kn1, kn2, kn3 |
---|
449 | REAL(wp), DIMENSION(kn1, kn2, kn3), INTENT(IN) :: pval |
---|
450 | INTEGER , INTENT( in ), OPTIONAL :: kcom |
---|
451 | !! * Local declarations |
---|
452 | REAL(wp), DIMENSION(:), ALLOCATABLE :: zvec |
---|
453 | LOGICAL, DIMENSION(kn1, kn2, kn3) :: zmask |
---|
454 | REAL(wp), DIMENSION(kn1, kn2, kn3) :: zfd |
---|
455 | INTEGER :: zdim |
---|
456 | |
---|
457 | zdim = kn1 * kn2 * kn3 |
---|
458 | ALLOCATE( zvec(zdim) ) |
---|
459 | |
---|
460 | zvec = PACK( pval(1:kn1,1:kn2,1:kn3),.TRUE.) |
---|
461 | |
---|
462 | CALL mpp_sum( zvec, zdim, kcom ) |
---|
463 | |
---|
464 | zmask(:,:,:) = .TRUE. |
---|
465 | mpp_sum_nfd_3d = UNPACK( zvec, zmask, zfd ) |
---|
466 | |
---|
467 | END FUNCTION mpp_sum_nfd_3d |
---|
468 | #else |
---|
469 | FUNCTION mpp_sum_nfd_4d( pval, kn1, kn2, kn3, kn4, kcom ) |
---|
470 | !! * Function return |
---|
471 | REAL(wp), DIMENSION(kn1, kn2, kn3, kn4) :: mpp_sum_nfd_4d |
---|
472 | !! * Arguments |
---|
473 | INTEGER, INTENT(IN) :: kn1, kn2, kn3, kn4 |
---|
474 | REAL(wp), DIMENSION(kn1, kn2, kn3, kn4), INTENT(IN) :: pval |
---|
475 | INTEGER , INTENT( in ), OPTIONAL :: kcom |
---|
476 | mpp_sum_nfd_4d = 0.0_wp |
---|
477 | WRITE(*,*) 'mpp_sum_nfd_4d: You should not have seen this print! error?' |
---|
478 | END FUNCTION mpp_sum_nfd_4d |
---|
479 | FUNCTION mpp_sum_nfd_3d( pval, kn1, kn2, kn3, kcom ) |
---|
480 | !! * Function return |
---|
481 | REAL(wp), DIMENSION(kn1, kn2, kn3) :: mpp_sum_nfd_3d |
---|
482 | !! * Arguments |
---|
483 | INTEGER, INTENT(IN) :: kn1, kn2, kn3 |
---|
484 | REAL(wp), DIMENSION(kn1, kn2, kn3), INTENT(IN) :: pval |
---|
485 | INTEGER , INTENT( in ), OPTIONAL :: kcom |
---|
486 | mpp_sum_nfd_3d = 0.0_wp |
---|
487 | WRITE(*,*) 'mpp_sum_nfd_3d: You should not have seen this print! error?' |
---|
488 | END FUNCTION mpp_sum_nfd_3d |
---|
489 | #endif |
---|
490 | |
---|
491 | END MODULE mppsumtam |
---|