3 * Copyright (c) 2004-2012 Michael Niedermayer <michaelni@gmx.at>
4 * bessel function: Copyright (c) 2006 Xiaogang Zhang
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 * @author Michael Niedermayer <michaelni@gmx.at>
29 #include "libavutil/avassert.h"
30 #include "libavutil/cpu.h"
34 * builds a polyphase filterbank.
35 * @param factor resampling factor
36 * @param scale wanted sum of coefficients for each filter
37 * @param filter_type filter type
38 * @param kaiser_beta kaiser window beta
39 * @return 0 on success, negative on error
41 static int build_filter(ResampleContext *c, void *filter, double factor, int tap_count, int alloc, int phase_count, int scale,
42 int filter_type, double kaiser_beta){
44 int ph_nb = phase_count % 2 ? phase_count : phase_count / 2 + 1;
46 double *tab = av_malloc_array(tap_count+1, sizeof(*tab));
47 double *sin_lut = av_malloc_array(ph_nb, sizeof(*sin_lut));
48 const int center= (tap_count-1)/2;
50 int ret = AVERROR(ENOMEM);
55 av_assert0(tap_count == 1 || tap_count % 2 == 0);
57 /* if upsampling, only need to interpolate, no filter */
62 for (ph = 0; ph < ph_nb; ph++)
63 sin_lut[ph] = sin(M_PI * ph / phase_count) * (center & 1 ? 1 : -1);
65 for(ph = 0; ph < ph_nb; ph++) {
67 for(i=0;i<tap_count;i++) {
68 x = M_PI * ((double)(i - center) - (double)ph / phase_count) * factor;
70 else if (factor == 1.0)
75 case SWR_FILTER_TYPE_CUBIC:{
76 const float d= -0.5; //first order derivative = -0.5
77 x = fabs(((double)(i - center) - (double)ph / phase_count) * factor);
78 if(x<1.0) y= 1 - 3*x*x + 2*x*x*x + d*( -x*x + x*x*x);
79 else y= d*(-4 + 8*x - 5*x*x + x*x*x);
81 case SWR_FILTER_TYPE_BLACKMAN_NUTTALL:
82 w = 2.0*x / (factor*tap_count);
84 y *= 0.3635819 - 0.4891775 * t + 0.1365995 * (2*t*t-1) - 0.0106411 * (4*t*t*t - 3*t);
86 case SWR_FILTER_TYPE_KAISER:
87 w = 2.0*x / (factor*tap_count*M_PI);
88 y *= av_bessel_i0(kaiser_beta*sqrt(FFMAX(1-w*w, 0)));
100 /* normalize so that an uniform color remains the same */
102 case AV_SAMPLE_FMT_S16P:
103 for(i=0;i<tap_count;i++)
104 ((int16_t*)filter)[ph * alloc + i] = av_clip_int16(lrintf(tab[i] * scale / norm));
105 if (phase_count % 2) break;
106 for (i = 0; i < tap_count; i++)
107 ((int16_t*)filter)[(phase_count-ph) * alloc + tap_count-1-i] = ((int16_t*)filter)[ph * alloc + i];
109 case AV_SAMPLE_FMT_S32P:
110 for(i=0;i<tap_count;i++)
111 ((int32_t*)filter)[ph * alloc + i] = av_clipl_int32(llrint(tab[i] * scale / norm));
112 if (phase_count % 2) break;
113 for (i = 0; i < tap_count; i++)
114 ((int32_t*)filter)[(phase_count-ph) * alloc + tap_count-1-i] = ((int32_t*)filter)[ph * alloc + i];
116 case AV_SAMPLE_FMT_FLTP:
117 for(i=0;i<tap_count;i++)
118 ((float*)filter)[ph * alloc + i] = tab[i] * scale / norm;
119 if (phase_count % 2) break;
120 for (i = 0; i < tap_count; i++)
121 ((float*)filter)[(phase_count-ph) * alloc + tap_count-1-i] = ((float*)filter)[ph * alloc + i];
123 case AV_SAMPLE_FMT_DBLP:
124 for(i=0;i<tap_count;i++)
125 ((double*)filter)[ph * alloc + i] = tab[i] * scale / norm;
126 if (phase_count % 2) break;
127 for (i = 0; i < tap_count; i++)
128 ((double*)filter)[(phase_count-ph) * alloc + tap_count-1-i] = ((double*)filter)[ph * alloc + i];
136 double sine[LEN + tap_count];
137 double filtered[LEN];
138 double maxff=-2, minff=2, maxsf=-2, minsf=2;
139 for(i=0; i<LEN; i++){
140 double ss=0, sf=0, ff=0;
141 for(j=0; j<LEN+tap_count; j++)
142 sine[j]= cos(i*j*M_PI/LEN);
143 for(j=0; j<LEN; j++){
146 for(k=0; k<tap_count; k++)
147 sum += filter[ph * tap_count + k] * sine[k+j];
148 filtered[j]= sum / (1<<FILTER_SHIFT);
149 ss+= sine[j + center] * sine[j + center];
150 ff+= filtered[j] * filtered[j];
151 sf+= sine[j + center] * filtered[j];
156 maxff= FFMAX(maxff, ff);
157 minff= FFMIN(minff, ff);
158 maxsf= FFMAX(maxsf, sf);
159 minsf= FFMIN(minsf, sf);
161 av_log(NULL, AV_LOG_ERROR, "i:%4d ss:%f ff:%13.6e-%13.6e sf:%13.6e-%13.6e\n", i, ss, maxff, minff, maxsf, minsf);
176 static void resample_free(ResampleContext **cc){
177 ResampleContext *c = *cc;
180 av_freep(&c->filter_bank);
184 static ResampleContext *resample_init(ResampleContext *c, int out_rate, int in_rate, int filter_size, int phase_shift, int linear,
185 double cutoff0, enum AVSampleFormat format, enum SwrFilterType filter_type, double kaiser_beta,
186 double precision, int cheby, int exact_rational)
188 double cutoff = cutoff0? cutoff0 : 0.97;
189 double factor= FFMIN(out_rate * cutoff / in_rate, 1.0);
190 int phase_count= 1<<phase_shift;
191 int phase_count_compensation = phase_count;
192 int filter_length = FFMAX((int)ceil(filter_size/factor), 1);
194 if (filter_length > 1)
195 filter_length = FFALIGN(filter_length, 2);
197 if (exact_rational) {
198 int phase_count_exact, phase_count_exact_den;
200 av_reduce(&phase_count_exact, &phase_count_exact_den, out_rate, in_rate, INT_MAX);
201 if (phase_count_exact <= phase_count) {
202 phase_count_compensation = phase_count_exact * (phase_count / phase_count_exact);
203 phase_count = phase_count_exact;
207 if (!c || c->phase_count != phase_count || c->linear!=linear || c->factor != factor
208 || c->filter_length != filter_length || c->format != format
209 || c->filter_type != filter_type || c->kaiser_beta != kaiser_beta) {
211 c = av_mallocz(sizeof(*c));
217 c->felem_size= av_get_bytes_per_sample(c->format);
220 case AV_SAMPLE_FMT_S16P:
221 c->filter_shift = 15;
223 case AV_SAMPLE_FMT_S32P:
224 c->filter_shift = 30;
226 case AV_SAMPLE_FMT_FLTP:
227 case AV_SAMPLE_FMT_DBLP:
231 av_log(NULL, AV_LOG_ERROR, "Unsupported sample format\n");
235 if (filter_size/factor > INT32_MAX/256) {
236 av_log(NULL, AV_LOG_ERROR, "Filter length too large\n");
240 c->phase_count = phase_count;
243 c->filter_length = filter_length;
244 c->filter_alloc = FFALIGN(c->filter_length, 8);
245 c->filter_bank = av_calloc(c->filter_alloc, (phase_count+1)*c->felem_size);
246 c->filter_type = filter_type;
247 c->kaiser_beta = kaiser_beta;
248 c->phase_count_compensation = phase_count_compensation;
251 if (build_filter(c, (void*)c->filter_bank, factor, c->filter_length, c->filter_alloc, phase_count, 1<<c->filter_shift, filter_type, kaiser_beta))
253 memcpy(c->filter_bank + (c->filter_alloc*phase_count+1)*c->felem_size, c->filter_bank, (c->filter_alloc-1)*c->felem_size);
254 memcpy(c->filter_bank + (c->filter_alloc*phase_count )*c->felem_size, c->filter_bank + (c->filter_alloc - 1)*c->felem_size, c->felem_size);
257 c->compensation_distance= 0;
258 if(!av_reduce(&c->src_incr, &c->dst_incr, out_rate, in_rate * (int64_t)phase_count, INT32_MAX/2))
260 while (c->dst_incr < (1<<20) && c->src_incr < (1<<20)) {
264 c->ideal_dst_incr = c->dst_incr;
265 c->dst_incr_div = c->dst_incr / c->src_incr;
266 c->dst_incr_mod = c->dst_incr % c->src_incr;
268 c->index= -phase_count*((c->filter_length-1)/2);
271 swri_resample_dsp_init(c);
275 av_freep(&c->filter_bank);
280 static int rebuild_filter_bank_with_compensation(ResampleContext *c)
282 uint8_t *new_filter_bank;
283 int new_src_incr, new_dst_incr;
284 int phase_count = c->phase_count_compensation;
287 if (phase_count == c->phase_count)
290 av_assert0(!c->frac && !c->dst_incr_mod);
292 new_filter_bank = av_calloc(c->filter_alloc, (phase_count + 1) * c->felem_size);
293 if (!new_filter_bank)
294 return AVERROR(ENOMEM);
296 ret = build_filter(c, new_filter_bank, c->factor, c->filter_length, c->filter_alloc,
297 phase_count, 1 << c->filter_shift, c->filter_type, c->kaiser_beta);
299 av_freep(&new_filter_bank);
302 memcpy(new_filter_bank + (c->filter_alloc*phase_count+1)*c->felem_size, new_filter_bank, (c->filter_alloc-1)*c->felem_size);
303 memcpy(new_filter_bank + (c->filter_alloc*phase_count )*c->felem_size, new_filter_bank + (c->filter_alloc - 1)*c->felem_size, c->felem_size);
305 if (!av_reduce(&new_src_incr, &new_dst_incr, c->src_incr,
306 c->dst_incr * (int64_t)(phase_count/c->phase_count), INT32_MAX/2))
308 av_freep(&new_filter_bank);
309 return AVERROR(EINVAL);
312 c->src_incr = new_src_incr;
313 c->dst_incr = new_dst_incr;
314 while (c->dst_incr < (1<<20) && c->src_incr < (1<<20)) {
318 c->ideal_dst_incr = c->dst_incr;
319 c->dst_incr_div = c->dst_incr / c->src_incr;
320 c->dst_incr_mod = c->dst_incr % c->src_incr;
321 c->index *= phase_count / c->phase_count;
322 c->phase_count = phase_count;
323 av_freep(&c->filter_bank);
324 c->filter_bank = new_filter_bank;
328 static int set_compensation(ResampleContext *c, int sample_delta, int compensation_distance){
331 if (compensation_distance && sample_delta) {
332 ret = rebuild_filter_bank_with_compensation(c);
337 c->compensation_distance= compensation_distance;
338 if (compensation_distance)
339 c->dst_incr = c->ideal_dst_incr - c->ideal_dst_incr * (int64_t)sample_delta / compensation_distance;
341 c->dst_incr = c->ideal_dst_incr;
343 c->dst_incr_div = c->dst_incr / c->src_incr;
344 c->dst_incr_mod = c->dst_incr % c->src_incr;
349 static int multiple_resample(ResampleContext *c, AudioData *dst, int dst_size, AudioData *src, int src_size, int *consumed){
351 int64_t max_src_size = (INT64_MAX/2 / c->phase_count) / c->src_incr;
353 if (c->compensation_distance)
354 dst_size = FFMIN(dst_size, c->compensation_distance);
355 src_size = FFMIN(src_size, max_src_size);
359 if (c->filter_length == 1 && c->phase_count == 1) {
360 int64_t index2= (1LL<<32)*c->frac/c->src_incr + (1LL<<32)*c->index;
361 int64_t incr= (1LL<<32) * c->dst_incr / c->src_incr;
362 int new_size = (src_size * (int64_t)c->src_incr - c->frac + c->dst_incr - 1) / c->dst_incr;
364 dst_size = FFMAX(FFMIN(dst_size, new_size), 0);
366 for (i = 0; i < dst->ch_count; i++) {
367 c->dsp.resample_one(dst->ch[i], src->ch[i], dst_size, index2, incr);
368 if (i+1 == dst->ch_count) {
369 c->index += dst_size * c->dst_incr_div;
370 c->index += (c->frac + dst_size * (int64_t)c->dst_incr_mod) / c->src_incr;
371 av_assert2(c->index >= 0);
372 *consumed = c->index;
373 c->frac = (c->frac + dst_size * (int64_t)c->dst_incr_mod) % c->src_incr;
379 int64_t end_index = (1LL + src_size - c->filter_length) * c->phase_count;
380 int64_t delta_frac = (end_index - c->index) * c->src_incr - c->frac;
381 int delta_n = (delta_frac + c->dst_incr - 1) / c->dst_incr;
382 int (*resample_func)(struct ResampleContext *c, void *dst,
383 const void *src, int n, int update_ctx);
385 dst_size = FFMAX(FFMIN(dst_size, delta_n), 0);
387 /* resample_linear and resample_common should have same behavior
388 * when frac and dst_incr_mod are zero */
389 resample_func = (c->linear && (c->frac || c->dst_incr_mod)) ?
390 c->dsp.resample_linear : c->dsp.resample_common;
391 for (i = 0; i < dst->ch_count; i++)
392 *consumed = resample_func(c, dst->ch[i], src->ch[i], dst_size, i+1 == dst->ch_count);
396 if (c->compensation_distance) {
397 c->compensation_distance -= dst_size;
398 if (!c->compensation_distance) {
399 c->dst_incr = c->ideal_dst_incr;
400 c->dst_incr_div = c->dst_incr / c->src_incr;
401 c->dst_incr_mod = c->dst_incr % c->src_incr;
408 static int64_t get_delay(struct SwrContext *s, int64_t base){
409 ResampleContext *c = s->resample;
410 int64_t num = s->in_buffer_count - (c->filter_length-1)/2;
411 num *= c->phase_count;
415 return av_rescale(num, base, s->in_sample_rate*(int64_t)c->src_incr * c->phase_count);
418 static int64_t get_out_samples(struct SwrContext *s, int in_samples) {
419 ResampleContext *c = s->resample;
420 // The + 2 are added to allow implementations to be slightly inaccurate, they should not be needed currently.
421 // They also make it easier to proof that changes and optimizations do not
422 // break the upper bound.
423 int64_t num = s->in_buffer_count + 2LL + in_samples;
424 num *= c->phase_count;
426 num = av_rescale_rnd(num, s->out_sample_rate, ((int64_t)s->in_sample_rate) * c->phase_count, AV_ROUND_UP) + 2;
428 if (c->compensation_distance) {
430 return AVERROR(EINVAL);
432 num = FFMAX(num, (num * c->ideal_dst_incr - 1) / c->dst_incr + 1);
437 static int resample_flush(struct SwrContext *s) {
438 ResampleContext *c = s->resample;
439 AudioData *a= &s->in_buffer;
441 int reflection = (FFMIN(s->in_buffer_count, c->filter_length) + 1) / 2;
443 if((ret = swri_realloc_audio(a, s->in_buffer_index + s->in_buffer_count + reflection)) < 0)
445 av_assert0(a->planar);
446 for(i=0; i<a->ch_count; i++){
447 for(j=0; j<reflection; j++){
448 memcpy(a->ch[i] + (s->in_buffer_index+s->in_buffer_count+j )*a->bps,
449 a->ch[i] + (s->in_buffer_index+s->in_buffer_count-j-1)*a->bps, a->bps);
452 s->in_buffer_count += reflection;
456 // in fact the whole handle multiple ridiculously small buffers might need more thinking...
457 static int invert_initial_buffer(ResampleContext *c, AudioData *dst, const AudioData *src,
458 int in_count, int *out_idx, int *out_sz)
460 int n, ch, num = FFMIN(in_count + *out_sz, c->filter_length + 1), res;
465 if ((res = swri_realloc_audio(dst, c->filter_length * 2 + 1)) < 0)
469 for (n = *out_sz; n < num; n++) {
470 for (ch = 0; ch < src->ch_count; ch++) {
471 memcpy(dst->ch[ch] + ((c->filter_length + n) * c->felem_size),
472 src->ch[ch] + ((n - *out_sz) * c->felem_size), c->felem_size);
476 // if not enough data is in, return and wait for more
477 if (num < c->filter_length + 1) {
479 *out_idx = c->filter_length;
484 for (n = 1; n <= c->filter_length; n++) {
485 for (ch = 0; ch < src->ch_count; ch++) {
486 memcpy(dst->ch[ch] + ((c->filter_length - n) * c->felem_size),
487 dst->ch[ch] + ((c->filter_length + n) * c->felem_size),
493 *out_idx = c->filter_length;
494 while (c->index < 0) {
496 c->index += c->phase_count;
498 *out_sz = FFMAX(*out_sz + c->filter_length,
499 1 + c->filter_length * 2) - *out_idx;
501 return FFMAX(res, 0);
504 struct Resampler const swri_resampler={
511 invert_initial_buffer,