GCC Code Coverage Report
Directory: ../../../ffmpeg/ Exec Total Coverage
File: src/libavcodec/fft_template.c Lines: 263 283 92.9 %
Date: 2019-11-20 04:07:19 Branches: 60 78 76.9 %

Line Branch Exec Source
1
/*
2
 * FFT/IFFT transforms
3
 * Copyright (c) 2008 Loren Merritt
4
 * Copyright (c) 2002 Fabrice Bellard
5
 * Partly based on libdjbfft by D. J. Bernstein
6
 *
7
 * This file is part of FFmpeg.
8
 *
9
 * FFmpeg is free software; you can redistribute it and/or
10
 * modify it under the terms of the GNU Lesser General Public
11
 * License as published by the Free Software Foundation; either
12
 * version 2.1 of the License, or (at your option) any later version.
13
 *
14
 * FFmpeg is distributed in the hope that it will be useful,
15
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
17
 * Lesser General Public License for more details.
18
 *
19
 * You should have received a copy of the GNU Lesser General Public
20
 * License along with FFmpeg; if not, write to the Free Software
21
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22
 */
23
24
/**
25
 * @file
26
 * FFT/IFFT transforms.
27
 */
28
29
#include <stdlib.h>
30
#include <string.h>
31
#include "libavutil/mathematics.h"
32
#include "libavutil/thread.h"
33
#include "fft.h"
34
#include "fft-internal.h"
35
36
#if FFT_FIXED_32
37
#include "fft_table.h"
38
39
58
static void av_cold fft_lut_init(void)
40
{
41
58
    int n = 0;
42
58
    ff_fft_lut_init(ff_fft_offsets_lut, 0, 1 << 17, &n);
43
58
}
44
45
#else /* FFT_FIXED_32 */
46
47
/* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */
48
#if !CONFIG_HARDCODED_TABLES
49
COSTABLE(16);
50
COSTABLE(32);
51
COSTABLE(64);
52
COSTABLE(128);
53
COSTABLE(256);
54
COSTABLE(512);
55
COSTABLE(1024);
56
COSTABLE(2048);
57
COSTABLE(4096);
58
COSTABLE(8192);
59
COSTABLE(16384);
60
COSTABLE(32768);
61
COSTABLE(65536);
62
COSTABLE(131072);
63
64
2553
static av_cold void init_ff_cos_tabs(int index)
65
{
66
    int i;
67
2553
    int m = 1<<index;
68
2553
    double freq = 2*M_PI/m;
69
2553
    FFTSample *tab = FFT_NAME(ff_cos_tabs)[index];
70
189130
    for(i=0; i<=m/4; i++)
71
186577
        tab[i] = FIX15(cos(i*freq));
72
184024
    for(i=1; i<m/4; i++)
73
181471
        tab[m/2-i] = tab[i];
74
2553
}
75
76
typedef struct CosTabsInitOnce {
77
    void (*func)(void);
78
    AVOnce control;
79
} CosTabsInitOnce;
80
81
#define INIT_FF_COS_TABS_FUNC(index, size)          \
82
static av_cold void init_ff_cos_tabs_ ## size (void)\
83
{                                                   \
84
    init_ff_cos_tabs(index);                        \
85
}
86
87
508
INIT_FF_COS_TABS_FUNC(4, 16)
88
487
INIT_FF_COS_TABS_FUNC(5, 32)
89
409
INIT_FF_COS_TABS_FUNC(6, 64)
90
381
INIT_FF_COS_TABS_FUNC(7, 128)
91
322
INIT_FF_COS_TABS_FUNC(8, 256)
92
294
INIT_FF_COS_TABS_FUNC(9, 512)
93
73
INIT_FF_COS_TABS_FUNC(10, 1024)
94
41
INIT_FF_COS_TABS_FUNC(11, 2048)
95
26
INIT_FF_COS_TABS_FUNC(12, 4096)
96
7
INIT_FF_COS_TABS_FUNC(13, 8192)
97
5
INIT_FF_COS_TABS_FUNC(14, 16384)
98
INIT_FF_COS_TABS_FUNC(15, 32768)
99
INIT_FF_COS_TABS_FUNC(16, 65536)
100
INIT_FF_COS_TABS_FUNC(17, 131072)
101
102
static CosTabsInitOnce cos_tabs_init_once[] = {
103
    { NULL },
104
    { NULL },
105
    { NULL },
106
    { NULL },
107
    { init_ff_cos_tabs_16, AV_ONCE_INIT },
108
    { init_ff_cos_tabs_32, AV_ONCE_INIT },
109
    { init_ff_cos_tabs_64, AV_ONCE_INIT },
110
    { init_ff_cos_tabs_128, AV_ONCE_INIT },
111
    { init_ff_cos_tabs_256, AV_ONCE_INIT },
112
    { init_ff_cos_tabs_512, AV_ONCE_INIT },
113
    { init_ff_cos_tabs_1024, AV_ONCE_INIT },
114
    { init_ff_cos_tabs_2048, AV_ONCE_INIT },
115
    { init_ff_cos_tabs_4096, AV_ONCE_INIT },
116
    { init_ff_cos_tabs_8192, AV_ONCE_INIT },
117
    { init_ff_cos_tabs_16384, AV_ONCE_INIT },
118
    { init_ff_cos_tabs_32768, AV_ONCE_INIT },
119
    { init_ff_cos_tabs_65536, AV_ONCE_INIT },
120
    { init_ff_cos_tabs_131072, AV_ONCE_INIT },
121
};
122
123
#endif
124
COSTABLE_CONST FFTSample * const FFT_NAME(ff_cos_tabs)[] = {
125
    NULL, NULL, NULL, NULL,
126
    FFT_NAME(ff_cos_16),
127
    FFT_NAME(ff_cos_32),
128
    FFT_NAME(ff_cos_64),
129
    FFT_NAME(ff_cos_128),
130
    FFT_NAME(ff_cos_256),
131
    FFT_NAME(ff_cos_512),
132
    FFT_NAME(ff_cos_1024),
133
    FFT_NAME(ff_cos_2048),
134
    FFT_NAME(ff_cos_4096),
135
    FFT_NAME(ff_cos_8192),
136
    FFT_NAME(ff_cos_16384),
137
    FFT_NAME(ff_cos_32768),
138
    FFT_NAME(ff_cos_65536),
139
    FFT_NAME(ff_cos_131072),
140
};
141
142
#endif /* FFT_FIXED_32 */
143
144
static void fft_permute_c(FFTContext *s, FFTComplex *z);
145
static void fft_calc_c(FFTContext *s, FFTComplex *z);
146
147
5243280
static int split_radix_permutation(int i, int n, int inverse)
148
{
149
    int m;
150
5243280
    if(n <= 2) return i&1;
151
4472040
    m = n >> 1;
152
4472040
    if(!(i&m))            return split_radix_permutation(i, m, inverse)*2;
153
2236020
    m >>= 1;
154
2236020
    if(inverse == !(i&m)) return split_radix_permutation(i, m, inverse)*4 + 1;
155
1118010
    else                  return split_radix_permutation(i, m, inverse)*4 - 1;
156
}
157
158
10477
av_cold void ff_init_ff_cos_tabs(int index)
159
{
160
#if (!CONFIG_HARDCODED_TABLES) && (!FFT_FIXED_32)
161
10477
    ff_thread_once(&cos_tabs_init_once[index].control, cos_tabs_init_once[index].func);
162
#endif
163
10477
}
164
165
static const int avx_tab[] = {
166
    0, 4, 1, 5, 8, 12, 9, 13, 2, 6, 3, 7, 10, 14, 11, 15
167
};
168
169
13822
static int is_second_half_of_fft32(int i, int n)
170
{
171
13822
    if (n <= 32)
172
3910
        return i >= 16;
173
9912
    else if (i < n/2)
174
4956
        return is_second_half_of_fft32(i, n/2);
175
4956
    else if (i < 3*n/4)
176
2478
        return is_second_half_of_fft32(i - n/2, n/4);
177
    else
178
2478
        return is_second_half_of_fft32(i - 3*n/4, n/4);
179
}
180
181
293
static av_cold void fft_perm_avx(FFTContext *s)
182
{
183
    int i;
184
293
    int n = 1 << s->nbits;
185
186
4203
    for (i = 0; i < n; i += 16) {
187
        int k;
188
3910
        if (is_second_half_of_fft32(i, n)) {
189
22763
            for (k = 0; k < 16; k++)
190
42848
                s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] =
191
21424
                    i + avx_tab[k];
192
193
        } else {
194
43707
            for (k = 0; k < 16; k++) {
195
41136
                int j = i + k;
196
41136
                j = (j & ~7) | ((j >> 1) & 3) | ((j << 2) & 4);
197
41136
                s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] = j;
198
            }
199
        }
200
    }
201
293
}
202
203
3586
av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
204
{
205
    int i, j, n;
206
207
3586
    s->revtab = NULL;
208
3586
    s->revtab32 = NULL;
209
210

3586
    if (nbits < 2 || nbits > 17)
211
        goto fail;
212
3586
    s->nbits = nbits;
213
3586
    n = 1 << nbits;
214
215
3586
    if (nbits <= 16) {
216
3586
        s->revtab = av_malloc(n * sizeof(uint16_t));
217
3586
        if (!s->revtab)
218
            goto fail;
219
    } else {
220
        s->revtab32 = av_malloc(n * sizeof(uint32_t));
221
        if (!s->revtab32)
222
            goto fail;
223
    }
224
3586
    s->tmp_buf = av_malloc(n * sizeof(FFTComplex));
225
3586
    if (!s->tmp_buf)
226
        goto fail;
227
3586
    s->inverse = inverse;
228
3586
    s->fft_permutation = FF_FFT_PERM_DEFAULT;
229
230
3586
    s->fft_permute = fft_permute_c;
231
3586
    s->fft_calc    = fft_calc_c;
232
#if CONFIG_MDCT
233
3586
    s->imdct_calc  = ff_imdct_calc_c;
234
3586
    s->imdct_half  = ff_imdct_half_c;
235
3586
    s->mdct_calc   = ff_mdct_calc_c;
236
#endif
237
238
#if FFT_FIXED_32
239
    {
240
        static AVOnce control = AV_ONCE_INIT;
241
184
        ff_thread_once(&control, fft_lut_init);
242
    }
243
#else /* FFT_FIXED_32 */
244
#if FFT_FLOAT
245
    if (ARCH_AARCH64) ff_fft_init_aarch64(s);
246
    if (ARCH_ARM)     ff_fft_init_arm(s);
247
    if (ARCH_PPC)     ff_fft_init_ppc(s);
248
3362
    if (ARCH_X86)     ff_fft_init_x86(s);
249
3362
    if (CONFIG_MDCT)  s->mdct_calcw = s->mdct_calc;
250
    if (HAVE_MIPSFPU) ff_fft_init_mips(s);
251
#else
252
40
    if (CONFIG_MDCT)  s->mdct_calcw = ff_mdct_calcw_c;
253
    if (ARCH_ARM)     ff_fft_fixed_init_arm(s);
254
#endif
255
13700
    for(j=4; j<=nbits; j++) {
256
10298
        ff_init_ff_cos_tabs(j);
257
    }
258
#endif /* FFT_FIXED_32 */
259
260
261
3586
    if (s->fft_permutation == FF_FFT_PERM_AVX) {
262
293
        fft_perm_avx(s);
263
    } else {
264
#define PROCESS_FFT_PERM_SWAP_LSBS(num) do {\
265
    for(i = 0; i < n; i++) {\
266
        int k;\
267
        j = i;\
268
        j = (j & ~3) | ((j >> 1) & 1) | ((j << 1) & 2);\
269
        k = -split_radix_permutation(i, n, s->inverse) & (n - 1);\
270
        s->revtab##num[k] = j;\
271
    } \
272
} while(0);
273
274
#define PROCESS_FFT_PERM_DEFAULT(num) do {\
275
    for(i = 0; i < n; i++) {\
276
        int k;\
277
        j = i;\
278
        k = -split_radix_permutation(i, n, s->inverse) & (n - 1);\
279
        s->revtab##num[k] = j;\
280
    } \
281
} while(0);
282
283
#define SPLIT_RADIX_PERMUTATION(num) do { \
284
    if (s->fft_permutation == FF_FFT_PERM_SWAP_LSBS) {\
285
        PROCESS_FFT_PERM_SWAP_LSBS(num) \
286
    } else {\
287
        PROCESS_FFT_PERM_DEFAULT(num) \
288
    }\
289
} while(0);
290
291
3293
    if (s->revtab)
292

711973
        SPLIT_RADIX_PERMUTATION()
293
3293
    if (s->revtab32)
294
        SPLIT_RADIX_PERMUTATION(32)
295
296
#undef PROCESS_FFT_PERM_DEFAULT
297
#undef PROCESS_FFT_PERM_SWAP_LSBS
298
#undef SPLIT_RADIX_PERMUTATION
299
    }
300
301
3586
    return 0;
302
 fail:
303
    av_freep(&s->revtab);
304
    av_freep(&s->revtab32);
305
    av_freep(&s->tmp_buf);
306
    return -1;
307
}
308
309
48047
static void fft_permute_c(FFTContext *s, FFTComplex *z)
310
{
311
    int j, np;
312
48047
    const uint16_t *revtab = s->revtab;
313
48047
    const uint32_t *revtab32 = s->revtab32;
314
48047
    np = 1 << s->nbits;
315
    /* TODO: handle split-radix permute in a more optimal way, probably in-place */
316
48047
    if (revtab) {
317
12913999
        for(j=0;j<np;j++) s->tmp_buf[revtab[j]] = z[j];
318
    } else
319
        for(j=0;j<np;j++) s->tmp_buf[revtab32[j]] = z[j];
320
321
48047
    memcpy(z, s->tmp_buf, np * sizeof(FFTComplex));
322
48047
}
323
324
3680
av_cold void ff_fft_end(FFTContext *s)
325
{
326
3680
    av_freep(&s->revtab);
327
3680
    av_freep(&s->revtab32);
328
3680
    av_freep(&s->tmp_buf);
329
3680
}
330
331
#if FFT_FIXED_32
332
333
767510
static void fft_calc_c(FFTContext *s, FFTComplex *z) {
334
335
    int nbits, i, n, num_transforms, offset, step;
336
    int n4, n2, n34;
337
    unsigned tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
338
    FFTComplex *tmpz;
339
767510
    const int fft_size = (1 << s->nbits);
340
    int64_t accu;
341
342
767510
    num_transforms = (0x2aab >> (16 - s->nbits)) | 1;
343
344
6540670
    for (n=0; n<num_transforms; n++){
345
5773160
        offset = ff_fft_offsets_lut[n] << 2;
346
5773160
        tmpz = z + offset;
347
348
5773160
        tmp1 = tmpz[0].re + (unsigned)tmpz[1].re;
349
5773160
        tmp5 = tmpz[2].re + (unsigned)tmpz[3].re;
350
5773160
        tmp2 = tmpz[0].im + (unsigned)tmpz[1].im;
351
5773160
        tmp6 = tmpz[2].im + (unsigned)tmpz[3].im;
352
5773160
        tmp3 = tmpz[0].re - (unsigned)tmpz[1].re;
353
5773160
        tmp8 = tmpz[2].im - (unsigned)tmpz[3].im;
354
5773160
        tmp4 = tmpz[0].im - (unsigned)tmpz[1].im;
355
5773160
        tmp7 = tmpz[2].re - (unsigned)tmpz[3].re;
356
357
5773160
        tmpz[0].re = tmp1 + tmp5;
358
5773160
        tmpz[2].re = tmp1 - tmp5;
359
5773160
        tmpz[0].im = tmp2 + tmp6;
360
5773160
        tmpz[2].im = tmp2 - tmp6;
361
5773160
        tmpz[1].re = tmp3 + tmp8;
362
5773160
        tmpz[3].re = tmp3 - tmp8;
363
5773160
        tmpz[1].im = tmp4 - tmp7;
364
5773160
        tmpz[3].im = tmp4 + tmp7;
365
    }
366
367
767510
    if (fft_size < 8)
368
1
        return;
369
370
767509
    num_transforms = (num_transforms >> 1) | 1;
371
372
4020228
    for (n=0; n<num_transforms; n++){
373
3252719
        offset = ff_fft_offsets_lut[n] << 3;
374
3252719
        tmpz = z + offset;
375
376
3252719
        tmp1 = tmpz[4].re + (unsigned)tmpz[5].re;
377
3252719
        tmp3 = tmpz[6].re + (unsigned)tmpz[7].re;
378
3252719
        tmp2 = tmpz[4].im + (unsigned)tmpz[5].im;
379
3252719
        tmp4 = tmpz[6].im + (unsigned)tmpz[7].im;
380
3252719
        tmp5 = tmp1 + tmp3;
381
3252719
        tmp7 = tmp1 - tmp3;
382
3252719
        tmp6 = tmp2 + tmp4;
383
3252719
        tmp8 = tmp2 - tmp4;
384
385
3252719
        tmp1 = tmpz[4].re - (unsigned)tmpz[5].re;
386
3252719
        tmp2 = tmpz[4].im - (unsigned)tmpz[5].im;
387
3252719
        tmp3 = tmpz[6].re - (unsigned)tmpz[7].re;
388
3252719
        tmp4 = tmpz[6].im - (unsigned)tmpz[7].im;
389
390
3252719
        tmpz[4].re = tmpz[0].re - tmp5;
391
3252719
        tmpz[0].re = tmpz[0].re + tmp5;
392
3252719
        tmpz[4].im = tmpz[0].im - tmp6;
393
3252719
        tmpz[0].im = tmpz[0].im + tmp6;
394
3252719
        tmpz[6].re = tmpz[2].re - tmp8;
395
3252719
        tmpz[2].re = tmpz[2].re + tmp8;
396
3252719
        tmpz[6].im = tmpz[2].im + tmp7;
397
3252719
        tmpz[2].im = tmpz[2].im - tmp7;
398
399
3252719
        accu = (int64_t)Q31(M_SQRT1_2)*(int)(tmp1 + tmp2);
400
3252719
        tmp5 = (int32_t)((accu + 0x40000000) >> 31);
401
3252719
        accu = (int64_t)Q31(M_SQRT1_2)*(int)(tmp3 - tmp4);
402
3252719
        tmp7 = (int32_t)((accu + 0x40000000) >> 31);
403
3252719
        accu = (int64_t)Q31(M_SQRT1_2)*(int)(tmp2 - tmp1);
404
3252719
        tmp6 = (int32_t)((accu + 0x40000000) >> 31);
405
3252719
        accu = (int64_t)Q31(M_SQRT1_2)*(int)(tmp3 + tmp4);
406
3252719
        tmp8 = (int32_t)((accu + 0x40000000) >> 31);
407
3252719
        tmp1 = tmp5 + tmp7;
408
3252719
        tmp3 = tmp5 - tmp7;
409
3252719
        tmp2 = tmp6 + tmp8;
410
3252719
        tmp4 = tmp6 - tmp8;
411
412
3252719
        tmpz[5].re = tmpz[1].re - tmp1;
413
3252719
        tmpz[1].re = tmpz[1].re + tmp1;
414
3252719
        tmpz[5].im = tmpz[1].im - tmp2;
415
3252719
        tmpz[1].im = tmpz[1].im + tmp2;
416
3252719
        tmpz[7].re = tmpz[3].re - tmp4;
417
3252719
        tmpz[3].re = tmpz[3].re + tmp4;
418
3252719
        tmpz[7].im = tmpz[3].im + tmp3;
419
3252719
        tmpz[3].im = tmpz[3].im - tmp3;
420
    }
421
422
767509
    step = 1 << ((MAX_LOG2_NFFT-4) - 4);
423
767509
    n4 = 4;
424
425
2424638
    for (nbits=4; nbits<=s->nbits; nbits++){
426
1657129
        n2  = 2*n4;
427
1657129
        n34 = 3*n4;
428
1657129
        num_transforms = (num_transforms >> 1) | 1;
429
430
4159954
        for (n=0; n<num_transforms; n++){
431
2502825
            const FFTSample *w_re_ptr = ff_w_tab_sr + step;
432
2502825
            const FFTSample *w_im_ptr = ff_w_tab_sr + MAX_FFT_SIZE/(4*16) - step;
433
2502825
            offset = ff_fft_offsets_lut[n] << nbits;
434
2502825
            tmpz = z + offset;
435
436
2502825
            tmp5 = tmpz[ n2].re + (unsigned)tmpz[n34].re;
437
2502825
            tmp1 = tmpz[ n2].re - (unsigned)tmpz[n34].re;
438
2502825
            tmp6 = tmpz[ n2].im + (unsigned)tmpz[n34].im;
439
2502825
            tmp2 = tmpz[ n2].im - (unsigned)tmpz[n34].im;
440
441
2502825
            tmpz[ n2].re = tmpz[ 0].re - tmp5;
442
2502825
            tmpz[  0].re = tmpz[ 0].re + tmp5;
443
2502825
            tmpz[ n2].im = tmpz[ 0].im - tmp6;
444
2502825
            tmpz[  0].im = tmpz[ 0].im + tmp6;
445
2502825
            tmpz[n34].re = tmpz[n4].re - tmp2;
446
2502825
            tmpz[ n4].re = tmpz[n4].re + tmp2;
447
2502825
            tmpz[n34].im = tmpz[n4].im + tmp1;
448
2502825
            tmpz[ n4].im = tmpz[n4].im - tmp1;
449
450
21313368
            for (i=1; i<n4; i++){
451
18810543
                FFTSample w_re = w_re_ptr[0];
452
18810543
                FFTSample w_im = w_im_ptr[0];
453
18810543
                accu  = (int64_t)w_re*tmpz[ n2+i].re;
454
18810543
                accu += (int64_t)w_im*tmpz[ n2+i].im;
455
18810543
                tmp1 = (int32_t)((accu + 0x40000000) >> 31);
456
18810543
                accu  = (int64_t)w_re*tmpz[ n2+i].im;
457
18810543
                accu -= (int64_t)w_im*tmpz[ n2+i].re;
458
18810543
                tmp2 = (int32_t)((accu + 0x40000000) >> 31);
459
18810543
                accu  = (int64_t)w_re*tmpz[n34+i].re;
460
18810543
                accu -= (int64_t)w_im*tmpz[n34+i].im;
461
18810543
                tmp3 = (int32_t)((accu + 0x40000000) >> 31);
462
18810543
                accu  = (int64_t)w_re*tmpz[n34+i].im;
463
18810543
                accu += (int64_t)w_im*tmpz[n34+i].re;
464
18810543
                tmp4 = (int32_t)((accu + 0x40000000) >> 31);
465
466
18810543
                tmp5 = tmp1 + tmp3;
467
18810543
                tmp1 = tmp1 - tmp3;
468
18810543
                tmp6 = tmp2 + tmp4;
469
18810543
                tmp2 = tmp2 - tmp4;
470
471
18810543
                tmpz[ n2+i].re = tmpz[   i].re - tmp5;
472
18810543
                tmpz[    i].re = tmpz[   i].re + tmp5;
473
18810543
                tmpz[ n2+i].im = tmpz[   i].im - tmp6;
474
18810543
                tmpz[    i].im = tmpz[   i].im + tmp6;
475
18810543
                tmpz[n34+i].re = tmpz[n4+i].re - tmp2;
476
18810543
                tmpz[ n4+i].re = tmpz[n4+i].re + tmp2;
477
18810543
                tmpz[n34+i].im = tmpz[n4+i].im + tmp1;
478
18810543
                tmpz[ n4+i].im = tmpz[n4+i].im - tmp1;
479
480
18810543
                w_re_ptr += step;
481
18810543
                w_im_ptr -= step;
482
            }
483
        }
484
1657129
        step >>= 1;
485
1657129
        n4   <<= 1;
486
    }
487
}
488
489
#else /* FFT_FIXED_32 */
490
491
#define BUTTERFLIES(a0,a1,a2,a3) {\
492
    BF(t3, t5, t5, t1);\
493
    BF(a2.re, a0.re, a0.re, t5);\
494
    BF(a3.im, a1.im, a1.im, t3);\
495
    BF(t4, t6, t2, t6);\
496
    BF(a3.re, a1.re, a1.re, t4);\
497
    BF(a2.im, a0.im, a0.im, t6);\
498
}
499
500
// force loading all the inputs before storing any.
501
// this is slightly slower for small data, but avoids store->load aliasing
502
// for addresses separated by large powers of 2.
503
#define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
504
    FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
505
    BF(t3, t5, t5, t1);\
506
    BF(a2.re, a0.re, r0, t5);\
507
    BF(a3.im, a1.im, i1, t3);\
508
    BF(t4, t6, t2, t6);\
509
    BF(a3.re, a1.re, r1, t4);\
510
    BF(a2.im, a0.im, i0, t6);\
511
}
512
513
#define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
514
    CMUL(t1, t2, a2.re, a2.im, wre, -wim);\
515
    CMUL(t5, t6, a3.re, a3.im, wre,  wim);\
516
    BUTTERFLIES(a0,a1,a2,a3)\
517
}
518
519
#define TRANSFORM_ZERO(a0,a1,a2,a3) {\
520
    t1 = a2.re;\
521
    t2 = a2.im;\
522
    t5 = a3.re;\
523
    t6 = a3.im;\
524
    BUTTERFLIES(a0,a1,a2,a3)\
525
}
526
527
/* z[0...8n-1], w[1...2n-1] */
528
#define PASS(name)\
529
static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\
530
{\
531
    FFTDouble t1, t2, t3, t4, t5, t6;\
532
    int o1 = 2*n;\
533
    int o2 = 4*n;\
534
    int o3 = 6*n;\
535
    const FFTSample *wim = wre+o1;\
536
    n--;\
537
\
538
    TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
539
    TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
540
    do {\
541
        z += 2;\
542
        wre += 2;\
543
        wim -= 2;\
544
        TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
545
        TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
546
    } while(--n);\
547
}
548
549
25874510
PASS(pass)
550
#if !CONFIG_SMALL
551
#undef BUTTERFLIES
552
#define BUTTERFLIES BUTTERFLIES_BIG
553
3066827
PASS(pass_big)
554
#endif
555
556
#define DECL_FFT(n,n2,n4)\
557
static void fft##n(FFTComplex *z)\
558
{\
559
    fft##n2(z);\
560
    fft##n4(z+n4*2);\
561
    fft##n4(z+n4*3);\
562
    pass(z,FFT_NAME(ff_cos_##n),n4/2);\
563
}
564
565
19004237
static void fft4(FFTComplex *z)
566
{
567
    FFTDouble t1, t2, t3, t4, t5, t6, t7, t8;
568
569
19004237
    BF(t3, t1, z[0].re, z[1].re);
570
19004237
    BF(t8, t6, z[3].re, z[2].re);
571
19004237
    BF(z[2].re, z[0].re, t1, t6);
572
19004237
    BF(t4, t2, z[0].im, z[1].im);
573
19004237
    BF(t7, t5, z[2].im, z[3].im);
574
19004237
    BF(z[3].im, z[1].im, t4, t8);
575
19004237
    BF(z[3].re, z[1].re, t3, t7);
576
19004237
    BF(z[2].im, z[0].im, t2, t5);
577
19004237
}
578
579
9672333
static void fft8(FFTComplex *z)
580
{
581
    FFTDouble t1, t2, t3, t4, t5, t6;
582
583
9672333
    fft4(z);
584
585
9672333
    BF(t1, z[5].re, z[4].re, -z[5].re);
586
9672333
    BF(t2, z[5].im, z[4].im, -z[5].im);
587
9672333
    BF(t5, z[7].re, z[6].re, -z[7].re);
588
9672333
    BF(t6, z[7].im, z[6].im, -z[7].im);
589
590
9672333
    BUTTERFLIES(z[0],z[2],z[4],z[6]);
591
9672333
    TRANSFORM(z[1],z[3],z[5],z[7],sqrthalf,sqrthalf);
592
9672333
}
593
594
#if !CONFIG_SMALL
595
4154839
static void fft16(FFTComplex *z)
596
{
597
    FFTDouble t1, t2, t3, t4, t5, t6;
598
4154839
    FFTSample cos_16_1 = FFT_NAME(ff_cos_16)[1];
599
4154839
    FFTSample cos_16_3 = FFT_NAME(ff_cos_16)[3];
600
601
4154839
    fft8(z);
602
4154839
    fft4(z+8);
603
4154839
    fft4(z+12);
604
605
4154839
    TRANSFORM_ZERO(z[0],z[4],z[8],z[12]);
606
4154839
    TRANSFORM(z[2],z[6],z[10],z[14],sqrthalf,sqrthalf);
607
4154839
    TRANSFORM(z[1],z[5],z[9],z[13],cos_16_1,cos_16_3);
608
4154839
    TRANSFORM(z[3],z[7],z[11],z[15],cos_16_3,cos_16_1);
609
4154839
}
610
#else
611
DECL_FFT(16,8,4)
612
#endif
613
2604921
DECL_FFT(32,16,8)
614
635379
DECL_FFT(64,32,16)
615
328319
DECL_FFT(128,64,32)
616
114812
DECL_FFT(256,128,64)
617
81399
DECL_FFT(512,256,128)
618
#if !CONFIG_SMALL
619
#define pass pass_big
620
#endif
621
8565
DECL_FFT(1024,512,256)
622
2973
DECL_FFT(2048,1024,512)
623
1010
DECL_FFT(4096,2048,1024)
624
689
DECL_FFT(8192,4096,2048)
625
DECL_FFT(16384,8192,4096)
626
DECL_FFT(32768,16384,8192)
627
DECL_FFT(65536,32768,16384)
628
DECL_FFT(131072,65536,32768)
629
630
static void (* const fft_dispatch[])(FFTComplex*) = {
631
    fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
632
    fft2048, fft4096, fft8192, fft16384, fft32768, fft65536, fft131072
633
};
634
635
3138425
static void fft_calc_c(FFTContext *s, FFTComplex *z)
636
{
637
3138425
    fft_dispatch[s->nbits-2](z);
638
3138425
}
639
#endif /* FFT_FIXED_32 */