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
|
/*
* Copyright (c) 2003 Matteo Frigo
* Copyright (c) 2003 Massachusetts Institute of Technology
*
* 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.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
*/
/* $Id: reodft11e-radix2.c,v 1.1 2008/10/17 06:13:18 scuri Exp $ */
/* Do an R{E,O}DFT11 problem of *even* size by a pair of R2HC problems
of half the size, plus some pre/post-processing. Use a trick from:
Zhongde Wang, "On computing the discrete Fourier and cosine transforms,"
IEEE Trans. Acoust. Speech Sig. Proc. ASSP-33 (4), 1341--1344 (1985).
to re-express as a pair of half-size REDFT01 (DCT-III) problems. Our
implementation looks quite a bit different from the algorithm described
in the paper because we combined the paper's pre/post-processing with
the pre/post-processing used to turn REDFT01 into R2HC. (Also, the
paper uses a DCT/DST pair, but we turn the DST into a DCT via the
usual reordering/sign-flip trick. We additionally combined a couple
of the matrices/transformations of the paper into a single pass.)
NOTE: We originally used a simpler method by S. C. Chan and K. L. Ho
that turned out to have numerical problems; see reodft11e-r2hc.c.
(For odd sizes, see reodft11e-r2hc-odd.c.)
*/
#include "reodft.h"
typedef struct {
solver super;
} S;
typedef struct {
plan_rdft super;
plan *cld;
twid *td, *td2;
int is, os;
int n;
int vl;
int ivs, ovs;
rdft_kind kind;
} P;
static void apply_re11(const plan *ego_, R *I, R *O)
{
const P *ego = (const P *) ego_;
int is = ego->is, os = ego->os;
int i, n = ego->n, n2 = n/2;
int iv, vl = ego->vl;
int ivs = ego->ivs, ovs = ego->ovs;
R *W = ego->td->W;
R *W2;
R *buf;
buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
buf[0] = K(2.0) * I[0];
buf[n2] = K(2.0) * I[is * (n - 1)];
for (i = 1; i + i < n2; ++i) {
int k = i + i;
E a, b, a2, b2;
{
E u, v;
u = I[is * (k - 1)];
v = I[is * k];
a = u + v;
b2 = u - v;
}
{
E u, v;
u = I[is * (n - k - 1)];
v = I[is * (n - k)];
b = u + v;
a2 = u - v;
}
{
E wa, wb;
wa = W[2*i];
wb = W[2*i + 1];
{
E apb, amb;
apb = a + b;
amb = a - b;
buf[i] = wa * amb + wb * apb;
buf[n2 - i] = wa * apb - wb * amb;
}
{
E apb, amb;
apb = a2 + b2;
amb = a2 - b2;
buf[n2 + i] = wa * amb + wb * apb;
buf[n - i] = wa * apb - wb * amb;
}
}
}
if (i + i == n2) {
E u, v;
u = I[is * (n2 - 1)];
v = I[is * n2];
buf[i] = K(2.0) * (u + v) * W[2*i];
buf[n - i] = K(2.0) * (u - v) * W[2*i];
}
/* child plan: two r2hc's of size n/2 */
{
plan_rdft *cld = (plan_rdft *) ego->cld;
cld->apply((plan *) cld, buf, buf);
}
W2 = ego->td2->W;
{ /* i == 0 case */
E wa, wb;
E a, b;
wa = W2[0]; /* cos */
wb = W2[1]; /* sin */
a = buf[0];
b = buf[n2];
O[0] = wa * a + wb * b;
O[os * (n - 1)] = wb * a - wa * b;
}
W2 += 2;
for (i = 1; i + i < n2; ++i, W2 += 2) {
int k;
E u, v, u2, v2;
u = buf[i];
v = buf[n2 - i];
u2 = buf[n2 + i];
v2 = buf[n - i];
k = (i + i) - 1;
{
E wa, wb;
E a, b;
wa = W2[0]; /* cos */
wb = W2[1]; /* sin */
a = u - v;
b = v2 - u2;
O[os * k] = wa * a + wb * b;
O[os * (n - 1 - k)] = wb * a - wa * b;
}
++k;
W2 += 2;
{
E wa, wb;
E a, b;
wa = W2[0]; /* cos */
wb = W2[1]; /* sin */
a = u + v;
b = u2 + v2;
O[os * k] = wa * a + wb * b;
O[os * (n - 1 - k)] = wb * a - wa * b;
}
}
if (i + i == n2) {
int k = (i + i) - 1;
E wa, wb;
E a, b;
wa = W2[0]; /* cos */
wb = W2[1]; /* sin */
a = buf[i];
b = buf[n2 + i];
O[os * k] = wa * a - wb * b;
O[os * (n - 1 - k)] = wb * a + wa * b;
}
}
X(ifree)(buf);
}
#if 0
/* This version of apply_re11 uses REDFT01 child plans, more similar
to the original paper by Z. Wang. We keep it around for reference
(it is simpler) and because it may become more efficient if we
ever implement REDFT01 codelets. */
static void apply_re11(const plan *ego_, R *I, R *O)
{
const P *ego = (const P *) ego_;
int is = ego->is, os = ego->os;
int i, n = ego->n;
int iv, vl = ego->vl;
int ivs = ego->ivs, ovs = ego->ovs;
R *W;
R *buf;
buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
buf[0] = K(2.0) * I[0];
buf[n/2] = K(2.0) * I[is * (n - 1)];
for (i = 1; i + i < n; ++i) {
int k = i + i;
E a, b;
a = I[is * (k - 1)];
b = I[is * k];
buf[i] = a + b;
buf[n - i] = a - b;
}
/* child plan: two redft01's (DCT-III) */
{
plan_rdft *cld = (plan_rdft *) ego->cld;
cld->apply((plan *) cld, buf, buf);
}
W = ego->td2->W;
for (i = 0; i + 1 < n/2; ++i, W += 2) {
{
E wa, wb;
E a, b;
wa = W[0]; /* cos */
wb = W[1]; /* sin */
a = buf[i];
b = buf[n/2 + i];
O[os * i] = wa * a + wb * b;
O[os * (n - 1 - i)] = wb * a - wa * b;
}
++i;
W += 2;
{
E wa, wb;
E a, b;
wa = W[0]; /* cos */
wb = W[1]; /* sin */
a = buf[i];
b = buf[n/2 + i];
O[os * i] = wa * a - wb * b;
O[os * (n - 1 - i)] = wb * a + wa * b;
}
}
if (i < n/2) {
E wa, wb;
E a, b;
wa = W[0]; /* cos */
wb = W[1]; /* sin */
a = buf[i];
b = buf[n/2 + i];
O[os * i] = wa * a + wb * b;
O[os * (n - 1 - i)] = wb * a - wa * b;
}
}
X(ifree)(buf);
}
#endif /* 0 */
/* like for rodft01, rodft11 is obtained from redft11 by
reversing the input and flipping the sign of every other output. */
static void apply_ro11(const plan *ego_, R *I, R *O)
{
const P *ego = (const P *) ego_;
int is = ego->is, os = ego->os;
int i, n = ego->n, n2 = n/2;
int iv, vl = ego->vl;
int ivs = ego->ivs, ovs = ego->ovs;
R *W = ego->td->W;
R *W2;
R *buf;
buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
buf[0] = K(2.0) * I[is * (n - 1)];
buf[n2] = K(2.0) * I[0];
for (i = 1; i + i < n2; ++i) {
int k = i + i;
E a, b, a2, b2;
{
E u, v;
u = I[is * (n - k)];
v = I[is * (n - 1 - k)];
a = u + v;
b2 = u - v;
}
{
E u, v;
u = I[is * (k)];
v = I[is * (k - 1)];
b = u + v;
a2 = u - v;
}
{
E wa, wb;
wa = W[2*i];
wb = W[2*i + 1];
{
E apb, amb;
apb = a + b;
amb = a - b;
buf[i] = wa * amb + wb * apb;
buf[n2 - i] = wa * apb - wb * amb;
}
{
E apb, amb;
apb = a2 + b2;
amb = a2 - b2;
buf[n2 + i] = wa * amb + wb * apb;
buf[n - i] = wa * apb - wb * amb;
}
}
}
if (i + i == n2) {
E u, v;
u = I[is * n2];
v = I[is * (n2 - 1)];
buf[i] = K(2.0) * (u + v) * W[2*i];
buf[n - i] = K(2.0) * (u - v) * W[2*i];
}
/* child plan: two r2hc's of size n/2 */
{
plan_rdft *cld = (plan_rdft *) ego->cld;
cld->apply((plan *) cld, buf, buf);
}
W2 = ego->td2->W;
{ /* i == 0 case */
E wa, wb;
E a, b;
wa = W2[0]; /* cos */
wb = W2[1]; /* sin */
a = buf[0];
b = buf[n2];
O[0] = wa * a + wb * b;
O[os * (n - 1)] = wa * b - wb * a;
}
W2 += 2;
for (i = 1; i + i < n2; ++i, W2 += 2) {
int k;
E u, v, u2, v2;
u = buf[i];
v = buf[n2 - i];
u2 = buf[n2 + i];
v2 = buf[n - i];
k = (i + i) - 1;
{
E wa, wb;
E a, b;
wa = W2[0]; /* cos */
wb = W2[1]; /* sin */
a = v - u;
b = u2 - v2;
O[os * k] = wa * a + wb * b;
O[os * (n - 1 - k)] = wa * b - wb * a;
}
++k;
W2 += 2;
{
E wa, wb;
E a, b;
wa = W2[0]; /* cos */
wb = W2[1]; /* sin */
a = u + v;
b = u2 + v2;
O[os * k] = wa * a + wb * b;
O[os * (n - 1 - k)] = wa * b - wb * a;
}
}
if (i + i == n2) {
int k = (i + i) - 1;
E wa, wb;
E a, b;
wa = W2[0]; /* cos */
wb = W2[1]; /* sin */
a = buf[i];
b = buf[n2 + i];
O[os * k] = wb * b - wa * a;
O[os * (n - 1 - k)] = wa * b + wb * a;
}
}
X(ifree)(buf);
}
static void awake(plan *ego_, int flg)
{
P *ego = (P *) ego_;
static const tw_instr reodft010e_tw[] = {
{ TW_COS, 0, 1 },
{ TW_SIN, 0, 1 },
{ TW_NEXT, 1, 0 }
};
static const tw_instr reodft11e_tw[] = {
{ TW_COS, 1, 1 },
{ TW_SIN, 1, 1 },
{ TW_NEXT, 2, 0 }
};
AWAKE(ego->cld, flg);
X(twiddle_awake)(flg, &ego->td, reodft010e_tw, 2*ego->n, 1, ego->n/4+1);
X(twiddle_awake)(flg, &ego->td2, reodft11e_tw, 8*ego->n, 1, ego->n);
}
static void destroy(plan *ego_)
{
P *ego = (P *) ego_;
X(plan_destroy_internal)(ego->cld);
}
static void print(const plan *ego_, printer *p)
{
const P *ego = (const P *) ego_;
p->print(p, "(%se-radix2-r2hc-%d%v%(%p%))",
X(rdft_kind_str)(ego->kind), ego->n, ego->vl, ego->cld);
}
static int applicable0(const solver *ego_, const problem *p_)
{
UNUSED(ego_);
if (RDFTP(p_)) {
const problem_rdft *p = (const problem_rdft *) p_;
return (1
&& p->sz->rnk == 1
&& p->vecsz->rnk <= 1
&& p->sz->dims[0].n % 2 == 0
&& (p->kind[0] == REDFT11 || p->kind[0] == RODFT11)
);
}
return 0;
}
static int applicable(const solver *ego, const problem *p, const planner *plnr)
{
return (!NO_UGLYP(plnr) && applicable0(ego, p));
}
static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
{
P *pln;
const problem_rdft *p;
plan *cld;
R *buf;
int n;
opcnt ops;
static const plan_adt padt = {
X(rdft_solve), awake, print, destroy
};
if (!applicable(ego_, p_, plnr))
return (plan *)0;
p = (const problem_rdft *) p_;
n = p->sz->dims[0].n;
buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
cld = X(mkplan_d)(plnr, X(mkproblem_rdft_1_d)(X(mktensor_1d)(n/2, 1, 1),
X(mktensor_1d)(2, n/2, n/2),
buf, buf, R2HC));
X(ifree)(buf);
if (!cld)
return (plan *)0;
pln = MKPLAN_RDFT(P, &padt, p->kind[0]==REDFT11 ? apply_re11:apply_ro11);
pln->n = n;
pln->is = p->sz->dims[0].is;
pln->os = p->sz->dims[0].os;
pln->cld = cld;
pln->td = pln->td2 = 0;
pln->kind = p->kind[0];
X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs);
X(ops_zero)(&ops);
ops.add = 2 + (n/2 - 1)/2 * 20;
ops.mul = 6 + (n/2 - 1)/2 * 16;
ops.other = 4*n + 2 + (n/2 - 1)/2 * 6;
if ((n/2) % 2 == 0) {
ops.add += 4;
ops.mul += 8;
ops.other += 4;
}
X(ops_zero)(&pln->super.super.ops);
X(ops_madd2)(pln->vl, &ops, &pln->super.super.ops);
X(ops_madd2)(pln->vl, &cld->ops, &pln->super.super.ops);
return &(pln->super.super);
}
/* constructor */
static solver *mksolver(void)
{
static const solver_adt sadt = { mkplan };
S *slv = MKSOLVER(S, &sadt);
return &(slv->super);
}
void X(reodft11e_radix2_r2hc_register)(planner *p)
{
REGISTER_SOLVER(p, mksolver());
}
|