3 * Copyright (c) 2008 Loren Merritt
4 * Copyright (c) 2002 Fabrice Bellard
5 * Partly based on libdjbfft by D. J. Bernstein
7 * This file is part of FFmpeg.
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.
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.
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
26 * FFT/IFFT transforms.
31 #include "libavutil/mathematics.h"
32 #include "libavutil/thread.h"
34 #include "fft-internal.h"
37 #include "fft_table.h"
39 static void av_cold fft_lut_init(void)
42 ff_fft_lut_init(ff_fft_offsets_lut, 0, 1 << 17, &n);
45 #else /* FFT_FIXED_32 */
47 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */
48 #if !CONFIG_HARDCODED_TABLES
64 static av_cold void init_ff_cos_tabs(int index)
68 double freq = 2*M_PI/m;
69 FFTSample *tab = FFT_NAME(ff_cos_tabs)[index];
71 tab[i] = FIX15(cos(i*freq));
76 typedef struct CosTabsInitOnce {
81 #define INIT_FF_COS_TABS_FUNC(index, size) \
82 static av_cold void init_ff_cos_tabs_ ## size (void)\
84 init_ff_cos_tabs(index); \
87 INIT_FF_COS_TABS_FUNC(4, 16)
88 INIT_FF_COS_TABS_FUNC(5, 32)
89 INIT_FF_COS_TABS_FUNC(6, 64)
90 INIT_FF_COS_TABS_FUNC(7, 128)
91 INIT_FF_COS_TABS_FUNC(8, 256)
92 INIT_FF_COS_TABS_FUNC(9, 512)
93 INIT_FF_COS_TABS_FUNC(10, 1024)
94 INIT_FF_COS_TABS_FUNC(11, 2048)
95 INIT_FF_COS_TABS_FUNC(12, 4096)
96 INIT_FF_COS_TABS_FUNC(13, 8192)
97 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)
102 static CosTabsInitOnce cos_tabs_init_once[] = {
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 },
124 COSTABLE_CONST FFTSample * const FFT_NAME(ff_cos_tabs)[] = {
125 NULL, NULL, NULL, NULL,
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),
142 #endif /* FFT_FIXED_32 */
144 static void fft_permute_c(FFTContext *s, FFTComplex *z);
145 static void fft_calc_c(FFTContext *s, FFTComplex *z);
147 static int split_radix_permutation(int i, int n, int inverse)
150 if(n <= 2) return i&1;
152 if(!(i&m)) return split_radix_permutation(i, m, inverse)*2;
154 if(inverse == !(i&m)) return split_radix_permutation(i, m, inverse)*4 + 1;
155 else return split_radix_permutation(i, m, inverse)*4 - 1;
158 av_cold void ff_init_ff_cos_tabs(int index)
160 #if (!CONFIG_HARDCODED_TABLES) && (!FFT_FIXED_32)
161 ff_thread_once(&cos_tabs_init_once[index].control, cos_tabs_init_once[index].func);
165 static const int avx_tab[] = {
166 0, 4, 1, 5, 8, 12, 9, 13, 2, 6, 3, 7, 10, 14, 11, 15
169 static int is_second_half_of_fft32(int i, int n)
174 return is_second_half_of_fft32(i, n/2);
176 return is_second_half_of_fft32(i - n/2, n/4);
178 return is_second_half_of_fft32(i - 3*n/4, n/4);
181 static av_cold void fft_perm_avx(FFTContext *s)
184 int n = 1 << s->nbits;
186 for (i = 0; i < n; i += 16) {
188 if (is_second_half_of_fft32(i, n)) {
189 for (k = 0; k < 16; k++)
190 s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] =
194 for (k = 0; k < 16; k++) {
196 j = (j & ~7) | ((j >> 1) & 3) | ((j << 2) & 4);
197 s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] = j;
203 av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
210 if (nbits < 2 || nbits > 17)
216 s->revtab = av_malloc(n * sizeof(uint16_t));
220 s->revtab32 = av_malloc(n * sizeof(uint32_t));
224 s->tmp_buf = av_malloc(n * sizeof(FFTComplex));
227 s->inverse = inverse;
228 s->fft_permutation = FF_FFT_PERM_DEFAULT;
230 s->fft_permute = fft_permute_c;
231 s->fft_calc = fft_calc_c;
233 s->imdct_calc = ff_imdct_calc_c;
234 s->imdct_half = ff_imdct_half_c;
235 s->mdct_calc = ff_mdct_calc_c;
240 static AVOnce control = AV_ONCE_INIT;
241 ff_thread_once(&control, fft_lut_init);
243 #else /* FFT_FIXED_32 */
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 if (ARCH_X86) ff_fft_init_x86(s);
249 if (CONFIG_MDCT) s->mdct_calcw = s->mdct_calc;
250 if (HAVE_MIPSFPU) ff_fft_init_mips(s);
252 if (CONFIG_MDCT) s->mdct_calcw = ff_mdct_calcw_c;
253 if (ARCH_ARM) ff_fft_fixed_init_arm(s);
255 for(j=4; j<=nbits; j++) {
256 ff_init_ff_cos_tabs(j);
258 #endif /* FFT_FIXED_32 */
261 if (s->fft_permutation == FF_FFT_PERM_AVX) {
267 if (s->fft_permutation == FF_FFT_PERM_SWAP_LSBS)
268 j = (j&~3) | ((j>>1)&1) | ((j<<1)&2);
269 k = -split_radix_permutation(i, n, s->inverse) & (n-1);
279 av_freep(&s->revtab);
280 av_freep(&s->revtab32);
281 av_freep(&s->tmp_buf);
285 static void fft_permute_c(FFTContext *s, FFTComplex *z)
288 const uint16_t *revtab = s->revtab;
289 const uint32_t *revtab32 = s->revtab32;
291 /* TODO: handle split-radix permute in a more optimal way, probably in-place */
293 for(j=0;j<np;j++) s->tmp_buf[revtab[j]] = z[j];
295 for(j=0;j<np;j++) s->tmp_buf[revtab32[j]] = z[j];
297 memcpy(z, s->tmp_buf, np * sizeof(FFTComplex));
300 av_cold void ff_fft_end(FFTContext *s)
302 av_freep(&s->revtab);
303 av_freep(&s->revtab32);
304 av_freep(&s->tmp_buf);
309 static void fft_calc_c(FFTContext *s, FFTComplex *z) {
311 int nbits, i, n, num_transforms, offset, step;
313 unsigned tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
315 const int fft_size = (1 << s->nbits);
318 num_transforms = (0x2aab >> (16 - s->nbits)) | 1;
320 for (n=0; n<num_transforms; n++){
321 offset = ff_fft_offsets_lut[n] << 2;
324 tmp1 = tmpz[0].re + (unsigned)tmpz[1].re;
325 tmp5 = tmpz[2].re + (unsigned)tmpz[3].re;
326 tmp2 = tmpz[0].im + (unsigned)tmpz[1].im;
327 tmp6 = tmpz[2].im + (unsigned)tmpz[3].im;
328 tmp3 = tmpz[0].re - (unsigned)tmpz[1].re;
329 tmp8 = tmpz[2].im - (unsigned)tmpz[3].im;
330 tmp4 = tmpz[0].im - (unsigned)tmpz[1].im;
331 tmp7 = tmpz[2].re - (unsigned)tmpz[3].re;
333 tmpz[0].re = tmp1 + tmp5;
334 tmpz[2].re = tmp1 - tmp5;
335 tmpz[0].im = tmp2 + tmp6;
336 tmpz[2].im = tmp2 - tmp6;
337 tmpz[1].re = tmp3 + tmp8;
338 tmpz[3].re = tmp3 - tmp8;
339 tmpz[1].im = tmp4 - tmp7;
340 tmpz[3].im = tmp4 + tmp7;
346 num_transforms = (num_transforms >> 1) | 1;
348 for (n=0; n<num_transforms; n++){
349 offset = ff_fft_offsets_lut[n] << 3;
352 tmp1 = tmpz[4].re + (unsigned)tmpz[5].re;
353 tmp3 = tmpz[6].re + (unsigned)tmpz[7].re;
354 tmp2 = tmpz[4].im + (unsigned)tmpz[5].im;
355 tmp4 = tmpz[6].im + (unsigned)tmpz[7].im;
361 tmp1 = tmpz[4].re - (unsigned)tmpz[5].re;
362 tmp2 = tmpz[4].im - (unsigned)tmpz[5].im;
363 tmp3 = tmpz[6].re - (unsigned)tmpz[7].re;
364 tmp4 = tmpz[6].im - (unsigned)tmpz[7].im;
366 tmpz[4].re = tmpz[0].re - tmp5;
367 tmpz[0].re = tmpz[0].re + tmp5;
368 tmpz[4].im = tmpz[0].im - tmp6;
369 tmpz[0].im = tmpz[0].im + tmp6;
370 tmpz[6].re = tmpz[2].re - tmp8;
371 tmpz[2].re = tmpz[2].re + tmp8;
372 tmpz[6].im = tmpz[2].im + tmp7;
373 tmpz[2].im = tmpz[2].im - tmp7;
375 accu = (int64_t)Q31(M_SQRT1_2)*(int)(tmp1 + tmp2);
376 tmp5 = (int32_t)((accu + 0x40000000) >> 31);
377 accu = (int64_t)Q31(M_SQRT1_2)*(int)(tmp3 - tmp4);
378 tmp7 = (int32_t)((accu + 0x40000000) >> 31);
379 accu = (int64_t)Q31(M_SQRT1_2)*(int)(tmp2 - tmp1);
380 tmp6 = (int32_t)((accu + 0x40000000) >> 31);
381 accu = (int64_t)Q31(M_SQRT1_2)*(int)(tmp3 + tmp4);
382 tmp8 = (int32_t)((accu + 0x40000000) >> 31);
388 tmpz[5].re = tmpz[1].re - tmp1;
389 tmpz[1].re = tmpz[1].re + tmp1;
390 tmpz[5].im = tmpz[1].im - tmp2;
391 tmpz[1].im = tmpz[1].im + tmp2;
392 tmpz[7].re = tmpz[3].re - tmp4;
393 tmpz[3].re = tmpz[3].re + tmp4;
394 tmpz[7].im = tmpz[3].im + tmp3;
395 tmpz[3].im = tmpz[3].im - tmp3;
398 step = 1 << ((MAX_LOG2_NFFT-4) - 4);
401 for (nbits=4; nbits<=s->nbits; nbits++){
404 num_transforms = (num_transforms >> 1) | 1;
406 for (n=0; n<num_transforms; n++){
407 const FFTSample *w_re_ptr = ff_w_tab_sr + step;
408 const FFTSample *w_im_ptr = ff_w_tab_sr + MAX_FFT_SIZE/(4*16) - step;
409 offset = ff_fft_offsets_lut[n] << nbits;
412 tmp5 = tmpz[ n2].re + (unsigned)tmpz[n34].re;
413 tmp1 = tmpz[ n2].re - (unsigned)tmpz[n34].re;
414 tmp6 = tmpz[ n2].im + (unsigned)tmpz[n34].im;
415 tmp2 = tmpz[ n2].im - (unsigned)tmpz[n34].im;
417 tmpz[ n2].re = tmpz[ 0].re - tmp5;
418 tmpz[ 0].re = tmpz[ 0].re + tmp5;
419 tmpz[ n2].im = tmpz[ 0].im - tmp6;
420 tmpz[ 0].im = tmpz[ 0].im + tmp6;
421 tmpz[n34].re = tmpz[n4].re - tmp2;
422 tmpz[ n4].re = tmpz[n4].re + tmp2;
423 tmpz[n34].im = tmpz[n4].im + tmp1;
424 tmpz[ n4].im = tmpz[n4].im - tmp1;
426 for (i=1; i<n4; i++){
427 FFTSample w_re = w_re_ptr[0];
428 FFTSample w_im = w_im_ptr[0];
429 accu = (int64_t)w_re*tmpz[ n2+i].re;
430 accu += (int64_t)w_im*tmpz[ n2+i].im;
431 tmp1 = (int32_t)((accu + 0x40000000) >> 31);
432 accu = (int64_t)w_re*tmpz[ n2+i].im;
433 accu -= (int64_t)w_im*tmpz[ n2+i].re;
434 tmp2 = (int32_t)((accu + 0x40000000) >> 31);
435 accu = (int64_t)w_re*tmpz[n34+i].re;
436 accu -= (int64_t)w_im*tmpz[n34+i].im;
437 tmp3 = (int32_t)((accu + 0x40000000) >> 31);
438 accu = (int64_t)w_re*tmpz[n34+i].im;
439 accu += (int64_t)w_im*tmpz[n34+i].re;
440 tmp4 = (int32_t)((accu + 0x40000000) >> 31);
447 tmpz[ n2+i].re = tmpz[ i].re - tmp5;
448 tmpz[ i].re = tmpz[ i].re + tmp5;
449 tmpz[ n2+i].im = tmpz[ i].im - tmp6;
450 tmpz[ i].im = tmpz[ i].im + tmp6;
451 tmpz[n34+i].re = tmpz[n4+i].re - tmp2;
452 tmpz[ n4+i].re = tmpz[n4+i].re + tmp2;
453 tmpz[n34+i].im = tmpz[n4+i].im + tmp1;
454 tmpz[ n4+i].im = tmpz[n4+i].im - tmp1;
465 #else /* FFT_FIXED_32 */
467 #define BUTTERFLIES(a0,a1,a2,a3) {\
469 BF(a2.re, a0.re, a0.re, t5);\
470 BF(a3.im, a1.im, a1.im, t3);\
472 BF(a3.re, a1.re, a1.re, t4);\
473 BF(a2.im, a0.im, a0.im, t6);\
476 // force loading all the inputs before storing any.
477 // this is slightly slower for small data, but avoids store->load aliasing
478 // for addresses separated by large powers of 2.
479 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
480 FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
482 BF(a2.re, a0.re, r0, t5);\
483 BF(a3.im, a1.im, i1, t3);\
485 BF(a3.re, a1.re, r1, t4);\
486 BF(a2.im, a0.im, i0, t6);\
489 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
490 CMUL(t1, t2, a2.re, a2.im, wre, -wim);\
491 CMUL(t5, t6, a3.re, a3.im, wre, wim);\
492 BUTTERFLIES(a0,a1,a2,a3)\
495 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
500 BUTTERFLIES(a0,a1,a2,a3)\
503 /* z[0...8n-1], w[1...2n-1] */
505 static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\
507 FFTDouble t1, t2, t3, t4, t5, t6;\
511 const FFTSample *wim = wre+o1;\
514 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
515 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
520 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
521 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
527 #define BUTTERFLIES BUTTERFLIES_BIG
530 #define DECL_FFT(n,n2,n4)\
531 static void fft##n(FFTComplex *z)\
536 pass(z,FFT_NAME(ff_cos_##n),n4/2);\
539 static void fft4(FFTComplex *z)
541 FFTDouble t1, t2, t3, t4, t5, t6, t7, t8;
543 BF(t3, t1, z[0].re, z[1].re);
544 BF(t8, t6, z[3].re, z[2].re);
545 BF(z[2].re, z[0].re, t1, t6);
546 BF(t4, t2, z[0].im, z[1].im);
547 BF(t7, t5, z[2].im, z[3].im);
548 BF(z[3].im, z[1].im, t4, t8);
549 BF(z[3].re, z[1].re, t3, t7);
550 BF(z[2].im, z[0].im, t2, t5);
553 static void fft8(FFTComplex *z)
555 FFTDouble t1, t2, t3, t4, t5, t6;
559 BF(t1, z[5].re, z[4].re, -z[5].re);
560 BF(t2, z[5].im, z[4].im, -z[5].im);
561 BF(t5, z[7].re, z[6].re, -z[7].re);
562 BF(t6, z[7].im, z[6].im, -z[7].im);
564 BUTTERFLIES(z[0],z[2],z[4],z[6]);
565 TRANSFORM(z[1],z[3],z[5],z[7],sqrthalf,sqrthalf);
569 static void fft16(FFTComplex *z)
571 FFTDouble t1, t2, t3, t4, t5, t6;
572 FFTSample cos_16_1 = FFT_NAME(ff_cos_16)[1];
573 FFTSample cos_16_3 = FFT_NAME(ff_cos_16)[3];
579 TRANSFORM_ZERO(z[0],z[4],z[8],z[12]);
580 TRANSFORM(z[2],z[6],z[10],z[14],sqrthalf,sqrthalf);
581 TRANSFORM(z[1],z[5],z[9],z[13],cos_16_1,cos_16_3);
582 TRANSFORM(z[3],z[7],z[11],z[15],cos_16_3,cos_16_1);
591 DECL_FFT(512,256,128)
593 #define pass pass_big
595 DECL_FFT(1024,512,256)
596 DECL_FFT(2048,1024,512)
597 DECL_FFT(4096,2048,1024)
598 DECL_FFT(8192,4096,2048)
599 DECL_FFT(16384,8192,4096)
600 DECL_FFT(32768,16384,8192)
601 DECL_FFT(65536,32768,16384)
602 DECL_FFT(131072,65536,32768)
604 static void (* const fft_dispatch[])(FFTComplex*) = {
605 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
606 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536, fft131072
609 static void fft_calc_c(FFTContext *s, FFTComplex *z)
611 fft_dispatch[s->nbits-2](z);
613 #endif /* FFT_FIXED_32 */