-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfft5d.c
1278 lines (1142 loc) · 39 KB
/
fft5d.c
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
*
* This source code is part of
*
* G R O M A C S
*
* GROningen MAchine for Chemical Simulations
*
* VERSION 4.5
* Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
* Copyright (c) 1991-2000, University of Groningen, The Netherlands.
* Copyright (c) 2001-2008, The GROMACS development team,
* check out http://www.gromacs.org for more information.
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version 2
* of the License, or (at your option) any later version.
*
* If you want to redistribute modifications, please consider that
* scientific software is very special. Version control is crucial -
* bugs must be traceable. We will be happy to consider code for
* inclusion in the official distribution, but derived work must not
* be called official GROMACS. Details are found in the README & COPYING
* files - if they are missing, get the official version at www.gromacs.org.
*
* To help us fund GROMACS development, we humbly ask that you cite
* the papers on the package - you can find them in the top README file.
*
* For more info, check our website at http://www.gromacs.org
*
* And Hey:
* Groningen Machine for Chemical Simulation
*/
#ifdef HAVE_CONFIG_H
#include <config.h>
#endif
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#ifdef NOGMX
#define GMX_PARALLEL_ENV_INITIALIZED 1
#else
#include "main.h"
#define GMX_PARALLEL_ENV_INITIALIZED gmx_parallel_env_initialized()
#endif
#ifdef GMX_LIB_MPI
#include <mpi.h>
#endif
#ifdef GMX_THREADS
#include "tmpi.h"
#endif
#ifdef GMX_OPENMP
#define FFT5D_THREADS
#endif
#ifdef FFT5D_THREADS
#include <omp.h>
/* requires fftw compiled with openmp */
/* #define FFT5D_FFTW_THREADS (now set by cmake) */
#endif
#include "fft5d.h"
#include <float.h>
#include <math.h>
#include <assert.h>
#ifndef __FLT_EPSILON__
#define __FLT_EPSILON__ FLT_EPSILON
#define __DBL_EPSILON__ DBL_EPSILON
#endif
#ifdef NOGMX
FILE* debug=0;
#endif
#include "gmx_fatal.h"
#ifdef GMX_FFT_FFTW3
#ifdef GMX_THREADS
/* none of the fftw3 calls, except execute(), are thread-safe, so
we need to serialize them with this mutex. */
static tMPI_Thread_mutex_t big_fftw_mutex=TMPI_THREAD_MUTEX_INITIALIZER;
#define FFTW_LOCK tMPI_Thread_mutex_lock(&big_fftw_mutex);
#define FFTW_UNLOCK tMPI_Thread_mutex_unlock(&big_fftw_mutex);
#else /* GMX_THREADS */
#define FFTW_LOCK
#define FFTW_UNLOCK
#endif /* GMX_THREADS */
#endif /* GMX_FFT_FFTW3 */
static double fft5d_fmax(double a, double b){
return (a>b)?a:b;
}
/* largest factor smaller than sqrt */
static int lfactor(int z) {
int i;
for (i=sqrt(z);;i--)
if (z%i==0) return i;
}
/* largest factor */
static int l2factor(int z) {
int i;
if (z==1) return 1;
for (i=z/2;;i--)
if (z%i==0) return i;
}
/* largest prime factor: WARNING: slow recursion, only use for small numbers */
static int lpfactor(int z) {
int f = l2factor(z);
if (f==1) return z;
return fft5d_fmax(lpfactor(f),lpfactor(z/f));
}
#ifndef GMX_MPI
#ifdef HAVE_GETTIMEOFDAY
#include <sys/time.h>
double MPI_Wtime() {
struct timeval tv;
gettimeofday(&tv,0);
return tv.tv_sec+tv.tv_usec*1e-6;
}
#else
double MPI_Wtime() {
return 0.0;
}
#endif
#endif
static int vmax(int* a, int s) {
int i,max=0;
for (i=0;i<s;i++)
{
if (a[i]>max) max=a[i];
}
return max;
}
#ifndef FFT5D_THREADS
static int omp_get_max_threads() { return 1; }
static int omp_get_thread_num() { return 0; }
#endif
/*
copied here from fftgrid, because:
1. function there not publically available
2. not sure whether we keep fftgrid
3. less dependencies for fft5d
Only used for non-fftw case
*/
static void *
gmx_calloc_aligned(size_t size)
{
void *p0,*p;
/*We initialize by zero for Valgrind
For non-divisible case we communicate more than the data.
If we don't initialize the data we communicate uninitialized data*/
p0 = calloc(size+32,1);
if(p0 == NULL)
{
gmx_fatal(FARGS,"Failed to allocated %u bytes of aligned memory.",size+32);
}
p = (void *) (((size_t) p0 + 32) & (~((size_t) 31)));
/* Yeah, yeah, we cannot free this pointer, but who cares... */
return p;
}
/* NxMxK the size of the data
* comm communicator to use for fft5d
* P0 number of processor in 1st axes (can be null for automatic)
* lin is allocated by fft5d because size of array is only known after planning phase
* rlout2 is only used as intermediate buffer - only returned after allocation to reuse for back transform - should not be used by caller
*/
fft5d_plan fft5d_plan_3d(int NG, int MG, int KG, MPI_Comm comm[2], int flags, t_complex** rlin, t_complex** rlout, t_complex** rlout2, t_complex** rlout3)
{
int P[2],bMaster,prank[2],i,t;
int rNG,rMG,rKG;
int *N0=0, *N1=0, *M0=0, *M1=0, *K0=0, *K1=0, *oN0=0, *oN1=0, *oM0=0, *oM1=0, *oK0=0, *oK1=0;
int N[3],M[3],K[3],pN[3],pM[3],pK[3],oM[3],oK[3],*iNin[3]={0},*oNin[3]={0},*iNout[3]={0},*oNout[3]={0};
int C[3],rC[3],nP[2];
int lsize;
t_complex *lin=0,*lout=0,*lout2=0,*lout3=0;
fft5d_plan plan;
int s;
/* comm, prank and P are in the order of the decomposition (plan->cart is in the order of transposes) */
#ifdef GMX_MPI
if (GMX_PARALLEL_ENV_INITIALIZED && comm[0] != MPI_COMM_NULL)
{
MPI_Comm_size(comm[0],&P[0]);
MPI_Comm_rank(comm[0],&prank[0]);
}
else
#endif
{
P[0] = 1;
prank[0] = 0;
}
#ifdef GMX_MPI
if (GMX_PARALLEL_ENV_INITIALIZED && comm[1] != MPI_COMM_NULL)
{
MPI_Comm_size(comm[1],&P[1]);
MPI_Comm_rank(comm[1],&prank[1]);
}
else
#endif
{
P[1] = 1;
prank[1] = 0;
}
bMaster=(prank[0]==0&&prank[1]==0);
if (debug)
{
fprintf(debug,"FFT5D: Using %dx%d processor grid, rank %d,%d\n",
P[0],P[1],prank[0],prank[1]);
}
if (bMaster) {
if (debug)
fprintf(debug,"FFT5D: N: %d, M: %d, K: %d, P: %dx%d, real2complex: %d, backward: %d, order yz: %d, debug %d\n",
NG,MG,KG,P[0],P[1],(flags&FFT5D_REALCOMPLEX)>0,(flags&FFT5D_BACKWARD)>0,(flags&FFT5D_ORDER_YZ)>0,(flags&FFT5D_DEBUG)>0);
/* The check below is not correct, one prime factor 11 or 13 is ok.
if (fft5d_fmax(fft5d_fmax(lpfactor(NG),lpfactor(MG)),lpfactor(KG))>7) {
printf("WARNING: FFT very slow with prime factors larger 7\n");
printf("Change FFT size or in case you cannot change it look at\n");
printf("http://www.fftw.org/fftw3_doc/Generating-your-own-code.html\n");
}
*/
}
if (NG==0 || MG==0 || KG==0) {
if (bMaster) printf("FFT5D: FATAL: Datasize cannot be zero in any dimension\n");
return 0;
}
rNG=NG;rMG=MG;rKG=KG;
if (flags&FFT5D_REALCOMPLEX) {
if (!(flags&FFT5D_BACKWARD)) NG = NG/2+1;
else {
if (!(flags&FFT5D_ORDER_YZ)) MG=MG/2+1;
else KG=KG/2+1;
}
}
/*for transpose we need to know the size for each processor not only our own size*/
N0 = (int*)malloc(P[0]*sizeof(int)); N1 = (int*)malloc(P[1]*sizeof(int));
M0 = (int*)malloc(P[0]*sizeof(int)); M1 = (int*)malloc(P[1]*sizeof(int));
K0 = (int*)malloc(P[0]*sizeof(int)); K1 = (int*)malloc(P[1]*sizeof(int));
oN0 = (int*)malloc(P[0]*sizeof(int));oN1 = (int*)malloc(P[1]*sizeof(int));
oM0 = (int*)malloc(P[0]*sizeof(int));oM1 = (int*)malloc(P[1]*sizeof(int));
oK0 = (int*)malloc(P[0]*sizeof(int));oK1 = (int*)malloc(P[1]*sizeof(int));
for (i=0;i<P[0];i++)
{
#define EVENDIST
#ifndef EVENDIST
oN0[i]=i*ceil((double)NG/P[0]);
oM0[i]=i*ceil((double)MG/P[0]);
oK0[i]=i*ceil((double)KG/P[0]);
#else
oN0[i]=(NG*i)/P[0];
oM0[i]=(MG*i)/P[0];
oK0[i]=(KG*i)/P[0];
#endif
}
for (i=0;i<P[1];i++)
{
#ifndef EVENDIST
oN1[i]=i*ceil((double)NG/P[1]);
oM1[i]=i*ceil((double)MG/P[1]);
oK1[i]=i*ceil((double)KG/P[1]);
#else
oN1[i]=(NG*i)/P[1];
oM1[i]=(MG*i)/P[1];
oK1[i]=(KG*i)/P[1];
#endif
}
for (i=0;i<P[0]-1;i++)
{
N0[i]=oN0[i+1]-oN0[i];
M0[i]=oM0[i+1]-oM0[i];
K0[i]=oK0[i+1]-oK0[i];
}
N0[P[0]-1]=NG-oN0[P[0]-1];
M0[P[0]-1]=MG-oM0[P[0]-1];
K0[P[0]-1]=KG-oK0[P[0]-1];
for (i=0;i<P[1]-1;i++)
{
N1[i]=oN1[i+1]-oN1[i];
M1[i]=oM1[i+1]-oM1[i];
K1[i]=oK1[i+1]-oK1[i];
}
N1[P[1]-1]=NG-oN1[P[1]-1];
M1[P[1]-1]=MG-oM1[P[1]-1];
K1[P[1]-1]=KG-oK1[P[1]-1];
/* for step 1-3 the local N,M,K sizes of the transposed system
C: contiguous dimension, and nP: number of processor in subcommunicator
for that step */
pM[0] = M0[prank[0]];
oM[0] = oM0[prank[0]];
pK[0] = K1[prank[1]];
oK[0] = oK1[prank[1]];
C[0] = NG;
rC[0] = rNG;
if (!(flags&FFT5D_ORDER_YZ)) {
N[0] = vmax(N1,P[1]);
M[0] = M0[prank[0]];
K[0] = vmax(K1,P[1]);
pN[0] = N1[prank[1]];
iNout[0] = N1;
oNout[0] = oN1;
nP[0] = P[1];
C[1] = KG;
rC[1] =rKG;
N[1] = vmax(K0,P[0]);
pN[1] = K0[prank[0]];
iNin[1] = K1;
oNin[1] = oK1;
iNout[1] = K0;
oNout[1] = oK0;
M[1] = vmax(M0,P[0]);
pM[1] = M0[prank[0]];
oM[1] = oM0[prank[0]];
K[1] = N1[prank[1]];
pK[1] = N1[prank[1]];
oK[1] = oN1[prank[1]];
nP[1] = P[0];
C[2] = MG;
rC[2] = rMG;
iNin[2] = M0;
oNin[2] = oM0;
M[2] = vmax(K0,P[0]);
pM[2] = K0[prank[0]];
oM[2] = oK0[prank[0]];
K[2] = vmax(N1,P[1]);
pK[2] = N1[prank[1]];
oK[2] = oN1[prank[1]];
free(N0); free(oN0); /*these are not used for this order*/
free(M1); free(oM1); /*the rest is freed in destroy*/
} else {
N[0] = vmax(N0,P[0]);
M[0] = vmax(M0,P[0]);
K[0] = K1[prank[1]];
pN[0] = N0[prank[0]];
iNout[0] = N0;
oNout[0] = oN0;
nP[0] = P[0];
C[1] = MG;
rC[1] =rMG;
N[1] = vmax(M1,P[1]);
pN[1] = M1[prank[1]];
iNin[1] = M0;
oNin[1] = oM0;
iNout[1] = M1;
oNout[1] = oM1;
M[1] = N0[prank[0]];
pM[1] = N0[prank[0]];
oM[1] = oN0[prank[0]];
K[1] = vmax(K1,P[1]);
pK[1] = K1[prank[1]];
oK[1] = oK1[prank[1]];
nP[1] = P[1];
C[2] = KG;
rC[2] = rKG;
iNin[2] = K1;
oNin[2] = oK1;
M[2] = vmax(N0,P[0]);
pM[2] = N0[prank[0]];
oM[2] = oN0[prank[0]];
K[2] = vmax(M1,P[1]);
pK[2] = M1[prank[1]];
oK[2] = oM1[prank[1]];
free(N1); free(oN1); /*these are not used for this order*/
free(K0); free(oK0); /*the rest is freed in destroy*/
}
/*
Difference between x-y-z regarding 2d decomposition is whether they are
distributed along axis 1, 2 or both
*/
/* int lsize = fmax(N[0]*M[0]*K[0]*nP[0],N[1]*M[1]*K[1]*nP[1]); */
lsize = fft5d_fmax(N[0]*M[0]*K[0]*nP[0],fft5d_fmax(N[1]*M[1]*K[1]*nP[1],C[2]*M[2]*K[2]));
/* int lsize = fmax(C[0]*M[0]*K[0],fmax(C[1]*M[1]*K[1],C[2]*M[2]*K[2])); */
if (!(flags&FFT5D_NOMALLOC)) {
lin = (t_complex*)gmx_calloc_aligned(sizeof(t_complex) * lsize);
lout = (t_complex*)gmx_calloc_aligned(sizeof(t_complex) * lsize);
lout2 = (t_complex*)gmx_calloc_aligned(sizeof(t_complex) * lsize);
lout3 = (t_complex*)gmx_calloc_aligned(sizeof(t_complex) * lsize);
} else {
lin = *rlin;
lout = *rlout;
lout2 = *rlout2;
lout3 = *rlout3;
}
plan = (fft5d_plan)calloc(1,sizeof(struct fft5d_plan_t));
int nthreads = 1;
//FFTW(init_threads)();
nthreads = omp_get_max_threads();
if (debug)
{
fprintf(debug, "Running on %d threads\n",nthreads);
}
#ifdef GMX_FFT_FFTW3 /*if not FFTW - then we don't do a 3d plan but instead use only 1D plans */
/* It is possible to use the 3d plan with OMP threads - but in that case it is not allowed to be called from
* within a parallel region. For now deactivated. If it should be supported it has to made sure that
* that the execute of the 3d plan is in a master/serial block (since it contains it own parallel region)
* and that the 3d plan is faster than the 1d plan.
*/
if ((!(flags&FFT5D_INPLACE)) && (!(P[0]>1 || P[1]>1)) && nthreads==1) { /*don't do 3d plan in parallel or if in_place requested */
int fftwflags=FFTW_DESTROY_INPUT;
fftw_iodim dims[3];
int inNG=NG,outMG=MG,outKG=KG;
FFTW_LOCK;
if (!(flags&FFT5D_NOMEASURE)) fftwflags|=FFTW_MEASURE;
if (flags&FFT5D_REALCOMPLEX) {
if (!(flags&FFT5D_BACKWARD)) { /*input pointer is not complex*/
inNG*=2;
} else { /*output pointer is not complex*/
if (!(flags&FFT5D_ORDER_YZ)) outMG*=2;
else outKG*=2;
}
}
if (!(flags&FFT5D_BACKWARD)) {
dims[0].n = KG;
dims[1].n = MG;
dims[2].n = rNG;
dims[0].is = inNG*MG; /*N M K*/
dims[1].is = inNG;
dims[2].is = 1;
if (!(flags&FFT5D_ORDER_YZ)) {
dims[0].os = MG; /*M K N*/
dims[1].os = 1;
dims[2].os = MG*KG;
} else {
dims[0].os = 1; /*K N M*/
dims[1].os = KG*NG;
dims[2].os = KG;
}
} else {
if (!(flags&FFT5D_ORDER_YZ)) {
dims[0].n = NG;
dims[1].n = KG;
dims[2].n = rMG;
dims[0].is = 1;
dims[1].is = NG*MG;
dims[2].is = NG;
dims[0].os = outMG*KG;
dims[1].os = outMG;
dims[2].os = 1;
} else {
dims[0].n = MG;
dims[1].n = NG;
dims[2].n = rKG;
dims[0].is = NG;
dims[1].is = 1;
dims[2].is = NG*MG;
dims[0].os = outKG*NG;
dims[1].os = outKG;
dims[2].os = 1;
}
}
#ifdef FFT5D_THREADS
#ifdef FFT5D_FFTW_THREADS
FFTW(plan_with_nthreads)(nthreads);
#endif
#endif
if ((flags&FFT5D_REALCOMPLEX) && !(flags&FFT5D_BACKWARD)) {
plan->p3d = FFTW(plan_guru_dft_r2c)(/*rank*/ 3, dims,
/*howmany*/ 0, /*howmany_dims*/0 ,
(real*)lin, (FFTW(complex) *)lout,
/*flags*/ fftwflags);
} else if ((flags&FFT5D_REALCOMPLEX) && (flags&FFT5D_BACKWARD)) {
plan->p3d = FFTW(plan_guru_dft_c2r)(/*rank*/ 3, dims,
/*howmany*/ 0, /*howmany_dims*/0 ,
(FFTW(complex) *)lin, (real*)lout,
/*flags*/ fftwflags);
} else {
plan->p3d = FFTW(plan_guru_dft)(/*rank*/ 3, dims,
/*howmany*/ 0, /*howmany_dims*/0 ,
(FFTW(complex) *)lin, (FFTW(complex) *)lout,
/*sign*/ (flags&FFT5D_BACKWARD)?1:-1, /*flags*/ fftwflags);
}
#ifdef FFT5D_THREADS
#ifdef FFT5D_FFTW_THREADS
FFTW(plan_with_nthreads)(1);
#endif
#endif
FFTW_UNLOCK;
}
if (!plan->p3d) { /* for decomposition and if 3d plan did not work */
#endif /* GMX_FFT_FFTW3 */
for (s=0;s<3;s++) {
if (debug)
{
fprintf(debug,"FFT5D: Plan s %d rC %d M %d pK %d C %d lsize %d\n",
s,rC[s],M[s],pK[s],C[s],lsize);
}
/*TODO:
improvements: make sure that each FFT is aligned. each should best start at a cacheline (true?)
can timers be improved?
*/
plan->p1d[s] = (gmx_fft_t*)malloc(sizeof(gmx_fft_t)*nthreads);
/* Make sure that the init routines are only called by one thread at a time and in order
(later is only important to not confuse valgrind)
*/
#pragma omp parallel for schedule(static),ordered
for(t=0; t<nthreads; t++)
#pragma omp ordered
{
int tsize = ((t+1)*pM[s]*pK[s]/nthreads)-(t*pM[s]*pK[s]/nthreads);
if ((flags&FFT5D_REALCOMPLEX) && ((!(flags&FFT5D_BACKWARD) && s==0) || ((flags&FFT5D_BACKWARD) && s==2))) {
gmx_fft_init_many_1d_real( &plan->p1d[s][t], rC[s], tsize, (flags&FFT5D_NOMEASURE)?GMX_FFT_FLAG_CONSERVATIVE:0 );
} else {
gmx_fft_init_many_1d ( &plan->p1d[s][t], C[s], tsize, (flags&FFT5D_NOMEASURE)?GMX_FFT_FLAG_CONSERVATIVE:0 );
}
}
}
#ifdef GMX_FFT_FFTW3
}
#endif
if ((flags&FFT5D_ORDER_YZ)) { /*plan->cart is in the order of transposes */
plan->cart[0]=comm[0]; plan->cart[1]=comm[1];
} else {
plan->cart[1]=comm[0]; plan->cart[0]=comm[1];
}
#ifdef FFT5D_MPI_TRANSPOSE
FFTW_LOCK
for (s=0;s<2;s++) {
if ((s==0 && !(flags&FFT5D_ORDER_YZ)) || (s==1 && (flags&FFT5D_ORDER_YZ)))
plan->mpip[s] = FFTW(mpi_plan_many_transpose)(nP[s], nP[s], N[s]*K[s]*pM[s]*2, 1, 1, (real*)lin, (real*)lout, plan->cart[s], FFTW_PATIENT);
else
plan->mpip[s] = FFTW(mpi_plan_many_transpose)(nP[s], nP[s], N[s]*pK[s]*M[s]*2, 1, 1, (real*)lin, (real*)lout, plan->cart[s], FFTW_PATIENT);
}
FFTW_UNLOCK
#endif
plan->lin=lin;
plan->lout=lout;
plan->lout2=lout2;
plan->lout3=lout3;
plan->NG=NG;plan->MG=MG;plan->KG=KG;
for (s=0;s<3;s++) {
plan->N[s]=N[s];plan->M[s]=M[s];plan->K[s]=K[s];plan->pN[s]=pN[s];plan->pM[s]=pM[s];plan->pK[s]=pK[s];
plan->oM[s]=oM[s];plan->oK[s]=oK[s];
plan->C[s]=C[s];plan->rC[s]=rC[s];
plan->iNin[s]=iNin[s];plan->oNin[s]=oNin[s];plan->iNout[s]=iNout[s];plan->oNout[s]=oNout[s];
}
for (s=0;s<2;s++) {
plan->P[s]=nP[s];plan->coor[s]=prank[s];
}
/* plan->fftorder=fftorder;
plan->direction=direction;
plan->realcomplex=realcomplex;
*/
plan->flags=flags;
plan->nthreads=nthreads;
*rlin=lin;
*rlout=lout;
*rlout2=lout2;
*rlout3=lout3;
return plan;
}
enum order {
XYZ,
XZY,
YXZ,
YZX,
ZXY,
ZYX
};
/*here x,y,z and N,M,K is in rotated coordinate system!!
x (and N) is mayor (consecutive) dimension, y (M) middle and z (K) major
maxN,maxM,maxK is max size of local data
pN, pM, pK is local size specific to current processor (only different to max if not divisible)
NG, MG, KG is size of global data*/
static void splitaxes(t_complex* lout,const t_complex* lin,
int maxN,int maxM,int maxK, int pN, int pM, int pK,
int P,int NG,int *N, int* oN,int starty,int startz,int endy, int endz)
{
int x,y,z,i;
int in_i,out_i,in_z,out_z,in_y,out_y;
int s_y,e_y;
for (z=startz; z<endz+1; z++) /*3. z l*/
{
if (z==startz) {
s_y=starty;
} else {
s_y=0;
}
if (z==endz) {
e_y=endy;
} else {
e_y=pM;
}
out_z = z*maxN*maxM;
in_z = z*NG*pM;
for (i=0; i<P; i++) /*index cube along long axis*/
{
out_i = out_z + i*maxN*maxM*maxK;
in_i = in_z + oN[i];
for (y=s_y;y<e_y;y++) { /*2. y k*/
out_y = out_i + y*maxN;
in_y = in_i + y*NG;
for (x=0;x<N[i];x++) { /*1. x j*/
lout[out_y+x] = lin[in_y+x]; /*in=z*NG*pM+oN[i]+y*NG+x*/
/*after split important that each processor chunk i has size maxN*maxM*maxK and thus being the same size*/
/*before split data contiguos - thus if different processor get different amount oN is different*/
}
}
}
}
}
/*make axis contiguous again (after AllToAll) and also do local transpose*/
/*transpose mayor and major dimension
variables see above
the major, middle, minor order is only correct for x,y,z (N,M,K) for the input
N,M,K local dimensions
KG global size*/
static void joinAxesTrans13(t_complex* lout,const t_complex* lin,
int maxN,int maxM,int maxK,int pN, int pM, int pK,
int P,int KG, int* K, int* oK,int starty, int startx, int endy, int endx)
{
int i,x,y,z;
int out_i,in_i,out_x,in_x,out_z,in_z;
int s_y,e_y;
for (x=startx;x<endx+1;x++) /*1.j*/
{
if (x==startx) {
s_y=starty;
} else {
s_y=0;
}
if (x==endx) {
e_y=endy;
} else {
e_y=pM;
}
out_x = x*KG*pM;
in_x = x;
for (i=0;i<P;i++) /*index cube along long axis*/
{
out_i = out_x + oK[i];
in_i = in_x + i*maxM*maxN*maxK;
for (z=0;z<K[i];z++) /*3.l*/
{
out_z = out_i + z;
in_z = in_i + z*maxM*maxN;
for (y=s_y;y<e_y;y++) { /*2.k*/
lout[out_z+y*KG] = lin[in_z+y*maxN]; /*out=x*KG*pM+oK[i]+z+y*KG*/
}
}
}
}
}
/*make axis contiguous again (after AllToAll) and also do local transpose
tranpose mayor and middle dimension
variables see above
the minor, middle, major order is only correct for x,y,z (N,M,K) for the input
N,M,K local size
MG, global size*/
static void joinAxesTrans12(t_complex* lout,const t_complex* lin,int maxN,int maxM,int maxK,int pN, int pM, int pK,
int P,int MG, int* M, int* oM, int startx, int startz, int endx, int endz) {
int i,z,y,x;
int out_i,in_i,out_z,in_z,out_x,in_x;
int s_x,e_x;
for (z=startz; z<endz+1; z++)
{
if (z==startz) {
s_x=startx;
} else {
s_x=0;
}
if (z==endz) {
e_x=endx;
} else {
e_x=pN;
}
out_z = z*MG*pN;
in_z = z*maxM*maxN;
for (i=0; i<P; i++) /*index cube along long axis*/
{
out_i = out_z + oM[i];
in_i = in_z + i*maxM*maxN*maxK;
for (x=s_x;x<e_x;x++) {
out_x = out_i + x*MG;
in_x = in_i + x;
for (y=0;y<M[i];y++) {
lout[out_x+y] = lin[in_x+y*maxN]; /*out=z*MG*pN+oM[i]+x*MG+y*/
}
}
}
}
}
static void rotate(int x[]) {
int t=x[0];
/* x[0]=x[2];
x[2]=x[1];
x[1]=t;*/
x[0]=x[1];
x[1]=x[2];
x[2]=t;
}
/*compute the offset to compare or print transposed local data in original input coordinates
xs matrix dimension size, xl dimension length, xc decomposition offset
s: step in computation = number of transposes*/
static void compute_offsets(fft5d_plan plan, int xs[], int xl[], int xc[], int NG[], int s) {
/* int direction = plan->direction;
int fftorder = plan->fftorder;*/
int o;
int pos[3],i;
int *pM=plan->pM, *pK=plan->pK, *oM=plan->oM, *oK=plan->oK,
*C=plan->C, *rC=plan->rC;
NG[0]=plan->NG;NG[1]=plan->MG;NG[2]=plan->KG;
if (!(plan->flags&FFT5D_ORDER_YZ)) {
switch (s) {
case 0: o=XYZ; break;
case 1: o=ZYX; break;
case 2: o=YZX; break;
default: assert(0);
}
} else {
switch (s) {
case 0: o=XYZ; break;
case 1: o=YXZ; break;
case 2: o=ZXY; break;
default: assert(0);
}
}
switch (o) {
case XYZ:pos[0]=1;pos[1]=2;pos[2]=3;break;
case XZY:pos[0]=1;pos[1]=3;pos[2]=2;break;
case YXZ:pos[0]=2;pos[1]=1;pos[2]=3;break;
case YZX:pos[0]=3;pos[1]=1;pos[2]=2;break;
case ZXY:pos[0]=2;pos[1]=3;pos[2]=1;break;
case ZYX:pos[0]=3;pos[1]=2;pos[2]=1;break;
}
/*if (debug) printf("pos: %d %d %d\n",pos[0],pos[1],pos[2]);*/
/*xs, xl give dimension size and data length in local transposed coordinate system
for 0(/1/2): x(/y/z) in original coordinate system*/
for (i=0;i<3;i++) {
switch (pos[i]) {
case 1: xs[i]=1; xc[i]=0; xl[i]=C[s];break;
case 2: xs[i]=C[s]; xc[i]=oM[s]; xl[i]=pM[s];break;
case 3: xs[i]=C[s]*pM[s];xc[i]=oK[s]; xl[i]=pK[s];break;
}
}
/*input order is different for test program to match FFTW order
(important for complex to real)*/
if (plan->flags&FFT5D_BACKWARD) {
rotate(xs);
rotate(xl);
rotate(xc);
rotate(NG);
if (plan->flags&FFT5D_ORDER_YZ) {
rotate(xs);
rotate(xl);
rotate(xc);
rotate(NG);
}
}
if ((plan->flags&FFT5D_REALCOMPLEX) && ((!((plan->flags&FFT5D_BACKWARD)) && s==0) || ((plan->flags&FFT5D_BACKWARD) && s==2))) {
xl[0] = rC[s];
}
}
static void print_localdata(const t_complex* lin, const char* txt, int s, fft5d_plan plan) {
int x,y,z,l,t;
int *coor = plan->coor;
int xs[3],xl[3],xc[3],NG[3];
int ll=(plan->flags&FFT5D_REALCOMPLEX)?1:2;
compute_offsets(plan,xs,xl,xc,NG,s);
fprintf(debug,txt,coor[0],coor[1],s);
/*printf("xs: %d %d %d, xl: %d %d %d\n",xs[0],xs[1],xs[2],xl[0],xl[1],xl[2]);*/
for(z=0;z<xl[2];z++) {
for(y=0;y<xl[1];y++) {
fprintf(debug,"%d %d: ",coor[0],coor[1]);
for (x=0;x<xl[0];x++) {
for (l=0;l<ll;l++) {
fprintf(debug,"%f ",((real*)lin)[(z*xs[2]+y*xs[1])*2+(x*xs[0])*ll+l]);
}
fprintf(debug,",");
}
fprintf(debug,"\n");
}
}
}
void fft5d_execute(fft5d_plan plan,fft5d_time times) {
t_complex *lin = plan->lin;
t_complex *lout = plan->lout;
t_complex *lout2 = plan->lout2;
t_complex *lout3 = plan->lout3;
t_complex *fftout,*joinin;
gmx_fft_t **p1d=plan->p1d;
#ifdef FFT5D_MPI_TRANSPOSE
FFTW(plan) *mpip=plan->mpip;
#endif
#ifdef GMX_MPI
MPI_Comm *cart=plan->cart;
#endif
double time_fft=0,time_local=0,time_mpi[2]={0},time=0;
int *N=plan->N,*M=plan->M,*K=plan->K,*pN=plan->pN,*pM=plan->pM,*pK=plan->pK,
*C=plan->C,*P=plan->P,**iNin=plan->iNin,**oNin=plan->oNin,**iNout=plan->iNout,**oNout=plan->oNout;
int s=0,t,tstart,tend,bParallelDim;
#ifdef GMX_FFT_FFTW3
if (plan->p3d) {
#pragma omp master
{
#ifdef NOGMX
if (times!=0)
{
time=MPI_Wtime();
}
#endif
FFTW(execute)(plan->p3d);
#ifdef NOGMX
if (times!=0)
{
times->fft+=MPI_Wtime()-time;
}
#endif
}
return;
}
#endif
s=0;
t = omp_get_thread_num();
/*lin: x,y,z*/
#pragma omp master
{
if (plan->flags&FFT5D_DEBUG) print_localdata(lin, "%d %d: copy in lin\n", s, plan);
}
for (s=0;s<2;s++) { /*loop over first two FFT steps (corner rotations)*/
#ifdef GMX_MPI
if (GMX_PARALLEL_ENV_INITIALIZED && cart[s] !=0 && P[s]>1 )
{
bParallelDim = 1;
}
else
#endif
{
bParallelDim = 0;
}
/* ---------- START FFT ------------ */
#ifdef NOGMX
#pragma omp master
{
if (times!=0)
{
time=MPI_Wtime();
}
}
#endif
if (bParallelDim) {
fftout = lout;
}
else
{
if (s==0)
{
fftout = lout3;
} else
{
fftout = lout2;
}
}
tstart = (t*pM[s]*pK[s]/plan->nthreads)*C[s];
if ((plan->flags&FFT5D_REALCOMPLEX) && !(plan->flags&FFT5D_BACKWARD) && s==0)
{
gmx_fft_many_1d_real(p1d[s][t],(plan->flags&FFT5D_BACKWARD)?GMX_FFT_COMPLEX_TO_REAL:GMX_FFT_REAL_TO_COMPLEX,lin+tstart,fftout+tstart);
} else
{
gmx_fft_many_1d( p1d[s][t],(plan->flags&FFT5D_BACKWARD)?GMX_FFT_BACKWARD:GMX_FFT_FORWARD, lin+tstart,fftout+tstart);
}
#pragma omp master
{
#ifdef NOGMX
if (times!=0)
{
time_fft+=MPI_Wtime()-time;
}
#endif
if (plan->flags&FFT5D_DEBUG) print_localdata(lout, "%d %d: FFT %d\n", s, plan);
}
/* ---------- END FFT ------------ */
/* ---------- START SPLIT + TRANSPOSE------------ (if parallel in in this dimension)*/
if (bParallelDim) {
#ifdef NOGMX
#pragma omp master
{
if (times!=0)
{
time=MPI_Wtime();
}
}
#endif
/*prepare for A
llToAll
1. (most outer) axes (x) is split into P[s] parts of size N[s]
for sending*/
if (pM[s]>0)
{
tend = ((t+1)*pM[s]*pK[s]/plan->nthreads);
tstart/=C[s];
splitaxes(lout2,lout,N[s],M[s],K[s], pN[s],pM[s],pK[s],P[s],C[s],iNout[s],oNout[s],tstart%pM[s],tstart/pM[s],tend%pM[s],tend/pM[s]);
}
#pragma omp barrier /*barrier required before AllToAll (all input has to be their) - before timing to make timing more acurate*/
#ifdef NOGMX
#pragma omp master
{
if (times!=0)
{
time_local+=MPI_Wtime()-time;
}
}
#endif
/* ---------- END SPLIT , START TRANSPOSE------------ */