/
fft_3d.c
514 lines (440 loc) · 14.7 KB
/
fft_3d.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
/* parallel FFT functions - 1998, 1999
Steve Plimpton, MS 1111, Dept 9221, Sandia National Labs
(505) 845-7873
sjplimp@sandia.gov
Copyright (2003) Sandia Corporation. Under the terms of Contract
DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
certain rights in this software. This software is distributed under
the GNU General Public License.
See the README file in the top-level directory of the distribution.
*/
#include "cow-cfg.h"
#if (COW_FFTW && COW_MPI)
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <mpi.h>
#include "pack_3d.h"
#include "remap_3d.h"
#include "fft_3d.h"
#define MIN(A,B) ((A) < (B)) ? (A) : (B)
#define MAX(A,B) ((A) > (B)) ? (A) : (B)
/* ------------------------------------------------------------------- */
/* Data layout for 3d FFTs:
data set of Nfast x Nmid x Nslow elements is owned by P procs
on input, each proc owns a subsection of the elements
on output, each proc will own a (possibly different) subsection
my subsection must not overlap with any other proc's subsection,
i.e. the union of all proc's input (or output) subsections must
exactly tile the global Nfast x Nmid x Nslow data set
when called from C, all subsection indices are
C-style from 0 to N-1 where N = Nfast or Nmid or Nslow
when called from F77, all subsection indices are
F77-style from 1 to N where N = Nfast or Nmid or Nslow
a proc can own 0 elements on input or output
by specifying hi index < lo index
on both input and output, data is stored contiguously on a processor
with a fast-varying, mid-varying, and slow-varying index
*/
/* ------------------------------------------------------------------- */
/* ------------------------------------------------------------------- */
/* Perform 3d FFT */
/* Arguments:
in starting address of input data on this proc
out starting address of where output data for this proc
will be placed (can be same as in)
flag 1 for forward FFT, -1 for inverse FFT
plan plan returned by previous call to fft_3d_create_plan
*/
void fft_3d(FFT_DATA *in, FFT_DATA *out, int flag, struct fft_plan_3d *plan)
{
int i,total,length,num;
double norm;
FFT_DATA *data,*copy;
/* pre-remap to prepare for 1st FFTs if needed
copy = loc for remap result */
if (plan->pre_plan) {
if (plan->pre_target == 0)
copy = out;
else
copy = plan->copy;
remap_3d((double *) in, (double *) copy, (double *) plan->scratch,
plan->pre_plan);
data = copy;
}
else
data = in;
// ---------------------------------------------------------------------------
// 1d FFTs along mid axis
// ---------------------------------------------------------------------------
total = plan->total1;
length = plan->length1;
{
int sign = flag == +1 ? FFTW_FORWARD : FFTW_BACKWARD;
int N = length;
fftw_plan fftplan = fftw_plan_many_dft(1, &N, total/length,
data, NULL,
1, length,
data, NULL,
1, length,
sign, FFTW_ESTIMATE);
fftw_execute(fftplan);
fftw_destroy_plan(fftplan);
}
/* 1st mid-remap to prepare for 2nd FFTs
copy = loc for remap result */
if (plan->mid1_target == 0)
copy = out;
else
copy = plan->copy;
remap_3d((double *) data, (double *) copy, (double *) plan->scratch,
plan->mid1_plan);
data = copy;
// ---------------------------------------------------------------------------
// 1d FFTs along mid axis
// ---------------------------------------------------------------------------
total = plan->total2;
length = plan->length2;
{
int sign = flag == +1 ? FFTW_FORWARD : FFTW_BACKWARD;
int N = length;
fftw_plan fftplan = fftw_plan_many_dft(1, &N, total/length,
data, NULL,
1, length,
data, NULL,
1, length,
sign, FFTW_ESTIMATE);
fftw_execute(fftplan);
fftw_destroy_plan(fftplan);
}
/* 2nd mid-remap to prepare for 3rd FFTs
copy = loc for remap result */
if (plan->mid2_target == 0)
copy = out;
else
copy = plan->copy;
remap_3d((double *) data, (double *) copy, (double *) plan->scratch,
plan->mid2_plan);
data = copy;
// ---------------------------------------------------------------------------
// 1d FFTs along slow axis
// ---------------------------------------------------------------------------
total = plan->total3;
length = plan->length3;
{
int sign = flag == +1 ? FFTW_FORWARD : FFTW_BACKWARD;
int N = length;
fftw_plan fftplan = fftw_plan_many_dft(1, &N, total/length,
data, NULL,
1, length,
data, NULL,
1, length,
sign, FFTW_ESTIMATE);
fftw_execute(fftplan);
fftw_destroy_plan(fftplan);
}
/* post-remap to put data in output format if needed
destination is always out */
if (plan->post_plan)
remap_3d((double *) data, (double *) out, (double *) plan->scratch,
plan->post_plan);
/* scaling if required */
if (flag == -1 && plan->scaled) {
norm = plan->norm;
num = plan->normnum;
for (i = 0; i < num; i++) {
out[i][0] *= norm;
out[i][1] *= norm;
}
}
}
/* ------------------------------------------------------------------- */
/* Create plan for performing a 3d FFT */
/* Arguments:
comm MPI communicator for the P procs which own the data
nfast,nmid,nslow size of global 3d matrix
in_ilo,in_ihi input bounds of data I own in fast index
in_jlo,in_jhi input bounds of data I own in mid index
in_klo,in_khi input bounds of data I own in slow index
out_ilo,out_ihi output bounds of data I own in fast index
out_jlo,out_jhi output bounds of data I own in mid index
out_klo,out_khi output bounds of data I own in slow index
scaled 0 = no scaling of result, 1 = scaling
permute permutation in storage order of indices on output
0 = no permutation
1 = permute once = mid->fast, slow->mid, fast->slow
2 = permute twice = slow->fast, fast->mid, mid->slow
nbuf returns size of internal storage buffers used by FFT
*/
struct fft_plan_3d *fft_3d_create_plan
(MPI_Comm comm, int nfast, int nmid, int nslow,
int in_ilo, int in_ihi, int in_jlo, int in_jhi,
int in_klo, int in_khi,
int out_ilo, int out_ihi, int out_jlo, int out_jhi,
int out_klo, int out_khi,
int scaled, int permute, int *nbuf)
{
struct fft_plan_3d *plan;
int me,nprocs;
int flag,remapflag;
int first_ilo,first_ihi,first_jlo,first_jhi,first_klo,first_khi;
int second_ilo,second_ihi,second_jlo,second_jhi,second_klo,second_khi;
int third_ilo,third_ihi,third_jlo,third_jhi,third_klo,third_khi;
int out_size,first_size,second_size,third_size,copy_size,scratch_size;
int np1=0,np2=0,ip1,ip2;
MPI_Comm_rank(comm, &me);
MPI_Comm_size(comm, &nprocs);
bifactor(nprocs,&np1,&np2);
ip1 = me % np1;
ip2 = me / np1;
/* allocate memory for plan data struct */
plan = (struct fft_plan_3d *) malloc(sizeof(struct fft_plan_3d));
if (plan == NULL) return NULL;
/* remap from initial distribution to layout needed for 1st set of 1d FFTs
not needed if all procs own entire fast axis initially
first indices = distribution after 1st set of FFTs */
if (in_ilo == 0 && in_ihi == nfast-1)
flag = 0;
else
flag = 1;
MPI_Allreduce(&flag,&remapflag,1,MPI_INT,MPI_MAX,comm);
if (remapflag == 0) {
first_ilo = in_ilo;
first_ihi = in_ihi;
first_jlo = in_jlo;
first_jhi = in_jhi;
first_klo = in_klo;
first_khi = in_khi;
plan->pre_plan = NULL;
}
else {
first_ilo = 0;
first_ihi = nfast - 1;
first_jlo = ip1*nmid/np1;
first_jhi = (ip1+1)*nmid/np1 - 1;
first_klo = ip2*nslow/np2;
first_khi = (ip2+1)*nslow/np2 - 1;
plan->pre_plan =
remap_3d_create_plan(comm,in_ilo,in_ihi,in_jlo,in_jhi,in_klo,in_khi,
first_ilo,first_ihi,first_jlo,first_jhi,
first_klo,first_khi,
FFT_PRECISION,0,0,2);
if (plan->pre_plan == NULL) return NULL;
}
/* 1d FFTs along fast axis */
plan->length1 = nfast;
plan->total1 = nfast * (first_jhi-first_jlo+1) * (first_khi-first_klo+1);
/* remap from 1st to 2nd FFT
choose which axis is split over np1 vs np2 to minimize communication
second indices = distribution after 2nd set of FFTs */
second_ilo = ip1*nfast/np1;
second_ihi = (ip1+1)*nfast/np1 - 1;
second_jlo = 0;
second_jhi = nmid - 1;
second_klo = ip2*nslow/np2;
second_khi = (ip2+1)*nslow/np2 - 1;
plan->mid1_plan =
remap_3d_create_plan(comm,
first_ilo,first_ihi,first_jlo,first_jhi,
first_klo,first_khi,
second_ilo,second_ihi,second_jlo,second_jhi,
second_klo,second_khi,
FFT_PRECISION,1,0,2);
if (plan->mid1_plan == NULL) return NULL;
/* 1d FFTs along mid axis */
plan->length2 = nmid;
plan->total2 = (second_ihi-second_ilo+1) * nmid * (second_khi-second_klo+1);
/* remap from 2nd to 3rd FFT
if final distribution is permute=2 with all procs owning entire slow axis
then this remapping goes directly to final distribution
third indices = distribution after 3rd set of FFTs */
if (permute == 2 && out_klo == 0 && out_khi == nslow-1)
flag = 0;
else
flag = 1;
MPI_Allreduce(&flag,&remapflag,1,MPI_INT,MPI_MAX,comm);
if (remapflag == 0) {
third_ilo = out_ilo;
third_ihi = out_ihi;
third_jlo = out_jlo;
third_jhi = out_jhi;
third_klo = out_klo;
third_khi = out_khi;
}
else {
third_ilo = ip1*nfast/np1;
third_ihi = (ip1+1)*nfast/np1 - 1;
third_jlo = ip2*nmid/np2;
third_jhi = (ip2+1)*nmid/np2 - 1;
third_klo = 0;
third_khi = nslow - 1;
}
plan->mid2_plan =
remap_3d_create_plan(comm,
second_jlo,second_jhi,second_klo,second_khi,
second_ilo,second_ihi,
third_jlo,third_jhi,third_klo,third_khi,
third_ilo,third_ihi,
FFT_PRECISION,1,0,2);
if (plan->mid2_plan == NULL) return NULL;
/* 1d FFTs along slow axis */
plan->length3 = nslow;
plan->total3 = (third_ihi-third_ilo+1) * (third_jhi-third_jlo+1) * nslow;
/* remap from 3rd FFT to final distribution
not needed if permute = 2 and third indices = out indices on all procs */
if (permute == 2 &&
out_ilo == third_ilo && out_ihi == third_ihi &&
out_jlo == third_jlo && out_jhi == third_jhi &&
out_klo == third_klo && out_khi == third_khi)
flag = 0;
else
flag = 1;
MPI_Allreduce(&flag,&remapflag,1,MPI_INT,MPI_MAX,comm);
if (remapflag == 0)
plan->post_plan = NULL;
else {
plan->post_plan =
remap_3d_create_plan(comm,
third_klo,third_khi,third_ilo,third_ihi,
third_jlo,third_jhi,
out_klo,out_khi,out_ilo,out_ihi,
out_jlo,out_jhi,
FFT_PRECISION,(permute+1)%3,0,2);
if (plan->post_plan == NULL) return NULL;
}
/* configure plan memory pointers and allocate work space
out_size = amount of memory given to FFT by user
first/second/third_size = amount of memory needed after pre,mid1,mid2 remaps
copy_size = amount needed internally for extra copy of data
scratch_size = amount needed internally for remap scratch space
for each remap:
use out space for result if big enough, else require copy buffer
accumulate largest required remap scratch space */
out_size = (out_ihi-out_ilo+1) * (out_jhi-out_jlo+1) * (out_khi-out_klo+1);
first_size = (first_ihi-first_ilo+1) * (first_jhi-first_jlo+1) *
(first_khi-first_klo+1);
second_size = (second_ihi-second_ilo+1) * (second_jhi-second_jlo+1) *
(second_khi-second_klo+1);
third_size = (third_ihi-third_ilo+1) * (third_jhi-third_jlo+1) *
(third_khi-third_klo+1);
copy_size = 0;
scratch_size = 0;
if (plan->pre_plan) {
if (first_size <= out_size)
plan->pre_target = 0;
else {
plan->pre_target = 1;
copy_size = MAX(copy_size,first_size);
}
scratch_size = MAX(scratch_size,first_size);
}
if (plan->mid1_plan) {
if (second_size <= out_size)
plan->mid1_target = 0;
else {
plan->mid1_target = 1;
copy_size = MAX(copy_size,second_size);
}
scratch_size = MAX(scratch_size,second_size);
}
if (plan->mid2_plan) {
if (third_size <= out_size)
plan->mid2_target = 0;
else {
plan->mid2_target = 1;
copy_size = MAX(copy_size,third_size);
}
scratch_size = MAX(scratch_size,third_size);
}
if (plan->post_plan)
scratch_size = MAX(scratch_size,out_size);
*nbuf = copy_size + scratch_size;
if (copy_size) {
plan->copy = (FFT_DATA *) malloc(copy_size*sizeof(FFT_DATA));
if (plan->copy == NULL) return NULL;
}
else plan->copy = NULL;
if (scratch_size) {
plan->scratch = (FFT_DATA *) malloc(scratch_size*sizeof(FFT_DATA));
if (plan->scratch == NULL) return NULL;
}
else plan->scratch = NULL;
if (scaled == 0)
plan->scaled = 0;
else {
plan->scaled = 1;
plan->norm = 1.0/(nfast*nmid*nslow);
plan->normnum = (out_ihi-out_ilo+1) * (out_jhi-out_jlo+1) *
(out_khi-out_klo+1);
}
return plan;
}
void fft_3d_destroy_plan(struct fft_plan_3d *plan)
{
if (plan->pre_plan) remap_3d_destroy_plan(plan->pre_plan);
if (plan->mid1_plan) remap_3d_destroy_plan(plan->mid1_plan);
if (plan->mid2_plan) remap_3d_destroy_plan(plan->mid2_plan);
if (plan->post_plan) remap_3d_destroy_plan(plan->post_plan);
if (plan->copy) free(plan->copy);
if (plan->scratch) free(plan->scratch);
free(plan);
}
void factor(int n, int *num, int *list)
{
if (n == 1) {
return;
}
else if (n % 2 == 0) {
*list = 2;
(*num)++;
factor(n/2,num,list+1);
}
else if (n % 3 == 0) {
*list = 3;
(*num)++;
factor(n/3,num,list+1);
}
else if (n % 5 == 0) {
*list = 5;
(*num)++;
factor(n/5,num,list+1);
}
else if (n % 7 == 0) {
*list = 7;
(*num)++;
factor(n/7,num,list+1);
}
else if (n % 11 == 0) {
*list = 11;
(*num)++;
factor(n/11,num,list+1);
}
else if (n % 13 == 0) {
*list = 13;
(*num)++;
factor(n/13,num,list+1);
}
else {
*list = n;
(*num)++;
return;
}
}
/* ------------------------------------------------------------------- */
/* divide n into 2 factors of as equal size as possible */
void bifactor(int n, int *factor1, int *factor2)
{
int n1,n2,facmax;
facmax = sqrt((double) n);
for (n1 = facmax; n1 > 0; n1--) {
n2 = n/n1;
if (n1*n2 == n) {
*factor1 = n1;
*factor2 = n2;
return;
}
}
}
#else
void __fft_3d_stub() { }
#endif // (COW_FFTW && COW_MPI)