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 | #endif |
---|
25 | USE lib_mpp |
---|
26 | USE mpp_tam |
---|
27 | USE mppsum |
---|
28 | USE in_out_manager |
---|
29 | |
---|
30 | IMPLICIT NONE |
---|
31 | |
---|
32 | !! * Routine accessibility |
---|
33 | PRIVATE |
---|
34 | |
---|
35 | PUBLIC & |
---|
36 | & mpp_sum_inter, & ! Interface for depending on nmppsum |
---|
37 | & mpp_sum_simple, & ! Simple sum |
---|
38 | & loc_sum_inter, & ! Interface for depending on nmppsum |
---|
39 | & loc_sum_simple, & ! Simple sum |
---|
40 | & loc_sum_indep, & ! Order independent local reproducible sum |
---|
41 | & nmppsum ! Summation control |
---|
42 | |
---|
43 | !! * Module variables |
---|
44 | INTEGER :: nmppsum = 2 ! Choice of MPP sum for |
---|
45 | ! (1 = simple) |
---|
46 | ! (2 = order independent) |
---|
47 | |
---|
48 | CONTAINS |
---|
49 | |
---|
50 | FUNCTION mpp_sum_inter( pval, kn ) |
---|
51 | !!---------------------------------------------------------------------- |
---|
52 | !! *** ROUTINE mpp_sum_inter *** |
---|
53 | !! |
---|
54 | !! ** Purpose : Summation of arrays across processors |
---|
55 | !! |
---|
56 | !! ** Method : Call either mpp_sum_simple or mpp_sum_indep depending on |
---|
57 | !! nmppsum |
---|
58 | !! |
---|
59 | !! ** Action : This does only work for MPI. |
---|
60 | !! It does not work for SHMEM. |
---|
61 | !! |
---|
62 | !! References : http://www.mpi-forum.org |
---|
63 | !! |
---|
64 | !! History : |
---|
65 | !! ! 07-07 (K. Mogensen) Original code |
---|
66 | !!---------------------------------------------------------------------- |
---|
67 | !! * Function return |
---|
68 | REAL(wp) mpp_sum_inter |
---|
69 | !! * Arguments |
---|
70 | INTEGER, INTENT(IN) :: & |
---|
71 | & kn |
---|
72 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
73 | & pval |
---|
74 | !! * Local declarations |
---|
75 | |
---|
76 | IF ( nmppsum == 1 ) THEN |
---|
77 | |
---|
78 | mpp_sum_inter = mpp_sum_simple( pval, kn ) |
---|
79 | |
---|
80 | ELSEIF ( nmppsum == 2 ) THEN |
---|
81 | |
---|
82 | mpp_sum_inter = mpp_sum_indep( pval, kn ) |
---|
83 | |
---|
84 | ELSE |
---|
85 | |
---|
86 | CALL ctl_stop( ' mpp_sum: Invalid nmppsum ' ) |
---|
87 | |
---|
88 | ENDIF |
---|
89 | |
---|
90 | END FUNCTION mpp_sum_inter |
---|
91 | |
---|
92 | FUNCTION mpp_sum_simple( pval, kn ) |
---|
93 | !!---------------------------------------------------------------------- |
---|
94 | !! *** ROUTINE mpp_sum_simple *** |
---|
95 | !! |
---|
96 | !! ** Purpose : Summation of arrays across processors |
---|
97 | !! |
---|
98 | !! ** Method : This routine is the naive implementation using |
---|
99 | !! mpi_allreduce for the sum of local data |
---|
100 | !! |
---|
101 | !! ** Action : This does only work for MPI. |
---|
102 | !! It does not work for SHMEM. |
---|
103 | !! |
---|
104 | !! References : http://www.mpi-forum.org |
---|
105 | !! |
---|
106 | !! History : |
---|
107 | !! ! 07-07 (K. Mogensen) Original code |
---|
108 | !!---------------------------------------------------------------------- |
---|
109 | !! * Function return |
---|
110 | REAL(wp) mpp_sum_simple |
---|
111 | !! * Arguments |
---|
112 | INTEGER, INTENT(IN) :: & |
---|
113 | & kn |
---|
114 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
115 | & pval |
---|
116 | !! * Local declarations |
---|
117 | REAL(wp) :: & |
---|
118 | & zres |
---|
119 | INTEGER :: & |
---|
120 | & ierr |
---|
121 | INTEGER :: & |
---|
122 | & ji |
---|
123 | |
---|
124 | ! Compute local sum |
---|
125 | zres = SUM( pval(1:kn) ) |
---|
126 | ! Global sum |
---|
127 | |
---|
128 | CALL mpp_sum( zres ) |
---|
129 | |
---|
130 | mpp_sum_simple = zres |
---|
131 | |
---|
132 | END FUNCTION mpp_sum_simple |
---|
133 | |
---|
134 | FUNCTION loc_sum_inter( pval, kn ) |
---|
135 | !!---------------------------------------------------------------------- |
---|
136 | !! *** ROUTINE loc_sum_inter *** |
---|
137 | !! |
---|
138 | !! ** Purpose : Summation of arrays across processors |
---|
139 | !! |
---|
140 | !! ** Method : Call either loc_sum_simple or loc_sum_indep depending on |
---|
141 | !! nmppsum |
---|
142 | !! |
---|
143 | !! ** Action : |
---|
144 | !! |
---|
145 | !! References : |
---|
146 | !! |
---|
147 | !! History : |
---|
148 | !! ! 07-09 (K. Mogensen) Original code |
---|
149 | !!---------------------------------------------------------------------- |
---|
150 | !! * Function return |
---|
151 | REAL(wp) loc_sum_inter |
---|
152 | !! * Arguments |
---|
153 | INTEGER, INTENT(IN) :: & |
---|
154 | & kn |
---|
155 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
156 | & pval |
---|
157 | !! * Local declarations |
---|
158 | |
---|
159 | IF ( nmppsum == 1 ) THEN |
---|
160 | |
---|
161 | loc_sum_inter = loc_sum_simple( pval, kn ) |
---|
162 | |
---|
163 | ELSEIF ( nmppsum == 2 ) THEN |
---|
164 | |
---|
165 | loc_sum_inter = loc_sum_indep( pval, kn ) |
---|
166 | |
---|
167 | ELSE |
---|
168 | |
---|
169 | CALL ctl_stop( ' loc_sum: Invalid nmppsum = ' ) |
---|
170 | |
---|
171 | ENDIF |
---|
172 | |
---|
173 | END FUNCTION loc_sum_inter |
---|
174 | |
---|
175 | FUNCTION loc_sum_simple( pval, kn ) |
---|
176 | !!---------------------------------------------------------------------- |
---|
177 | !! *** ROUTINE loc_sum_simple *** |
---|
178 | !! |
---|
179 | !! ** Purpose : Summation of arrays across processors |
---|
180 | !! |
---|
181 | !! ** Method : This routine is the naive implementation for the sum |
---|
182 | !! of local data |
---|
183 | !! |
---|
184 | !! ** Action : |
---|
185 | !! |
---|
186 | !! References : |
---|
187 | !! |
---|
188 | !! History : |
---|
189 | !! ! 07-09 (K. Mogensen) Original code |
---|
190 | !!---------------------------------------------------------------------- |
---|
191 | !! * Function return |
---|
192 | REAL(wp) loc_sum_simple |
---|
193 | !! * Arguments |
---|
194 | INTEGER, INTENT(IN) :: & |
---|
195 | & kn |
---|
196 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
197 | & pval |
---|
198 | !! * Local declarations |
---|
199 | REAL(wp) :: & |
---|
200 | & zloc |
---|
201 | INTEGER :: & |
---|
202 | & ierr |
---|
203 | INTEGER :: & |
---|
204 | & ji |
---|
205 | |
---|
206 | ! Compute local sum |
---|
207 | |
---|
208 | zloc = SUM( pval(1:kn) ) |
---|
209 | |
---|
210 | loc_sum_simple = zloc |
---|
211 | |
---|
212 | END FUNCTION loc_sum_simple |
---|
213 | |
---|
214 | FUNCTION loc_sum_indep( pval, kn ) |
---|
215 | !!---------------------------------------------------------------------- |
---|
216 | !! *** ROUTINE loc_sum_indep *** |
---|
217 | !! |
---|
218 | !! ** Purpose : Sum all elements in the pval array in |
---|
219 | !! an accurate order-independent way. |
---|
220 | !! |
---|
221 | !! ** Method : The code iterates the compensated summation until the |
---|
222 | !! result is guaranteed to be within 4*eps of the true sum. |
---|
223 | !! It then rounds the result to the nearest floating-point |
---|
224 | !! number whose last three bits are zero, thereby |
---|
225 | !! guaranteeing an order-independent result. |
---|
226 | !! |
---|
227 | !! This version is used for local arrays. It is identical |
---|
228 | !! to the MPP version for clarity. |
---|
229 | !! |
---|
230 | !! ** Action : |
---|
231 | !! |
---|
232 | !! References : M. Fisher (ECMWF): IFS code + personal communication |
---|
233 | !! The algorithm is based on Ogita et al. (2005) |
---|
234 | !! SIAM J. Sci. Computing, Vol.26, No.6, pp1955-1988. |
---|
235 | !! This is based in turn on an algorithm |
---|
236 | !! by Knuth (1969, seminumerical algorithms). |
---|
237 | !! |
---|
238 | !! History : |
---|
239 | !! ! 07-09 (K. Mogensen) Original code heavily based on IFS. |
---|
240 | !!---------------------------------------------------------------------- |
---|
241 | !! * Function return |
---|
242 | REAL(wp) loc_sum_indep |
---|
243 | !! * Arguments |
---|
244 | INTEGER, INTENT(IN) :: & |
---|
245 | & kn |
---|
246 | REAL(wp), DIMENSION(kn), INTENT(IN) :: & |
---|
247 | & pval |
---|
248 | !! * Local declarations |
---|
249 | REAL(wp), DIMENSION(3) ::& |
---|
250 | & zbuffl |
---|
251 | REAL(wp), DIMENSION(:), ALLOCATABLE :: & |
---|
252 | & zpsums, & |
---|
253 | & zperrs, & |
---|
254 | & zpcors, & |
---|
255 | & zbuffg, & |
---|
256 | & zp |
---|
257 | REAL(wp) :: & |
---|
258 | & zcorr, & |
---|
259 | & zerr, & |
---|
260 | & zolderr, & |
---|
261 | & zbeta, & |
---|
262 | & zres |
---|
263 | INTEGER :: & |
---|
264 | & jj |
---|
265 | |
---|
266 | ! Check that the the algorithm can work |
---|
267 | |
---|
268 | IF ( ( REAL( 2 * kn ) * EPSILON( zres ) ) >= 1.0 ) THEN |
---|
269 | |
---|
270 | CALL ctl_stop( 'kn is too large to guarantee error bounds' ) |
---|
271 | |
---|
272 | ENDIF |
---|
273 | |
---|
274 | ALLOCATE( & |
---|
275 | & zp(MAX(kn,1)), & |
---|
276 | & zbuffg(1*SIZE(zbuffl)), & |
---|
277 | & zpsums(1), & |
---|
278 | & zperrs(1), & |
---|
279 | & zpcors(1) & |
---|
280 | & ) |
---|
281 | |
---|
282 | zolderr = HUGE(zerr) |
---|
283 | |
---|
284 | ! Copy the input array. This avoids some tricky indexing, at the |
---|
285 | ! expense of some inefficency. |
---|
286 | |
---|
287 | IF ( kn > 0 ) THEN |
---|
288 | |
---|
289 | zp(:) = pval(:) |
---|
290 | |
---|
291 | ELSE |
---|
292 | |
---|
293 | zp(1) = 0.0_wp |
---|
294 | |
---|
295 | ENDIF |
---|
296 | |
---|
297 | k_loop: DO |
---|
298 | |
---|
299 | ! Transform local arrays |
---|
300 | |
---|
301 | IF ( kn > 0 ) THEN |
---|
302 | |
---|
303 | CALL comp_sum ( zp, kn, zcorr, zerr ) |
---|
304 | |
---|
305 | ENDIF |
---|
306 | |
---|
307 | ! Gather partial sums and error bounds to all processors |
---|
308 | |
---|
309 | zbuffl(1) = zp(MAX(kn,1)) |
---|
310 | |
---|
311 | IF ( kn > 0 ) THEN |
---|
312 | |
---|
313 | zbuffl(2) = zerr |
---|
314 | zbuffl(3) = zcorr |
---|
315 | |
---|
316 | ELSE |
---|
317 | |
---|
318 | zbuffl(2) = 0.0_wp |
---|
319 | zbuffl(3) = 0.0_wp |
---|
320 | |
---|
321 | ENDIF |
---|
322 | |
---|
323 | zpsums(1) = zbuffl(1) |
---|
324 | zperrs(1) = zbuffl(2) |
---|
325 | zpcors(1) = zbuffl(3) |
---|
326 | |
---|
327 | ! Transform partial sums |
---|
328 | |
---|
329 | CALL comp_sum( zpsums, 1, zcorr, zerr ) |
---|
330 | zerr = zerr + SUM(zperrs) |
---|
331 | zcorr = zcorr + SUM(zpcors) |
---|
332 | |
---|
333 | ! Calculate final result |
---|
334 | |
---|
335 | zres = zpsums(1) + zcorr |
---|
336 | |
---|
337 | ! Calculate error bound. This is corollary 4.7 from Ogita et al. |
---|
338 | ! (2005) |
---|
339 | |
---|
340 | zbeta = zerr *( REAL( 2*kn, wp ) * EPSILON(zres) ) & |
---|
341 | & /(1.0_wp - REAL( 2*kn, wp ) * EPSILON(zres) ) |
---|
342 | |
---|
343 | zerr = EPSILON(zres) * ABS(zres) & |
---|
344 | & +(zbeta + ( 2.0_wp * EPSILON(zres) * EPSILON(zres) * ABS(zres) & |
---|
345 | & +3.0_wp * TINY(zres) ) ) |
---|
346 | |
---|
347 | ! Update the last element of the local array |
---|
348 | |
---|
349 | zp(MAX(kn,1)) = zpsums(nproc+1) |
---|
350 | |
---|
351 | ! Exit if the global error is small enough |
---|
352 | |
---|
353 | IF ( zerr < 4.0_wp * SPACING(zres) ) EXIT k_loop |
---|
354 | |
---|
355 | ! Take appropriate action if ZRES cannot be sufficiently refined. |
---|
356 | |
---|
357 | IF (zerr >= zolderr) THEN |
---|
358 | |
---|
359 | CALL ctl_stop( 'Failed to refine sum', & |
---|
360 | & 'Warning: Possiblity of non-reproducible results' ) |
---|
361 | |
---|
362 | ENDIF |
---|
363 | |
---|
364 | zolderr = zerr |
---|
365 | |
---|
366 | ENDDO k_loop |
---|
367 | |
---|
368 | ! At this stage, we have guaranteed that ZRES less than 4*EPS |
---|
369 | ! away from the exact sum. There are only four floating point |
---|
370 | ! numbers in this range. So, if we find the nearest number that |
---|
371 | ! has its last three bits zero, then we have a reproducible result. |
---|
372 | |
---|
373 | loc_sum_indep = fround(zres) |
---|
374 | |
---|
375 | DEALLOCATE( & |
---|
376 | & zpcors, & |
---|
377 | & zperrs, & |
---|
378 | & zpsums, & |
---|
379 | & zbuffg, & |
---|
380 | & zp & |
---|
381 | & ) |
---|
382 | |
---|
383 | END FUNCTION loc_sum_indep |
---|
384 | |
---|
385 | |
---|
386 | END MODULE mppsumtam |
---|