3 * Copyright (c) 2004-2012 Michael Niedermayer <michaelni@gmx.at>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * @author Michael Niedermayer <michaelni@gmx.at>
28 #include "libavutil/avassert.h"
32 * 0th order modified bessel function of the first kind.
34 static double bessel(double x){
38 static const double inv[100]={
39 1.0/( 1* 1), 1.0/( 2* 2), 1.0/( 3* 3), 1.0/( 4* 4), 1.0/( 5* 5), 1.0/( 6* 6), 1.0/( 7* 7), 1.0/( 8* 8), 1.0/( 9* 9), 1.0/(10*10),
40 1.0/(11*11), 1.0/(12*12), 1.0/(13*13), 1.0/(14*14), 1.0/(15*15), 1.0/(16*16), 1.0/(17*17), 1.0/(18*18), 1.0/(19*19), 1.0/(20*20),
41 1.0/(21*21), 1.0/(22*22), 1.0/(23*23), 1.0/(24*24), 1.0/(25*25), 1.0/(26*26), 1.0/(27*27), 1.0/(28*28), 1.0/(29*29), 1.0/(30*30),
42 1.0/(31*31), 1.0/(32*32), 1.0/(33*33), 1.0/(34*34), 1.0/(35*35), 1.0/(36*36), 1.0/(37*37), 1.0/(38*38), 1.0/(39*39), 1.0/(40*40),
43 1.0/(41*41), 1.0/(42*42), 1.0/(43*43), 1.0/(44*44), 1.0/(45*45), 1.0/(46*46), 1.0/(47*47), 1.0/(48*48), 1.0/(49*49), 1.0/(50*50),
44 1.0/(51*51), 1.0/(52*52), 1.0/(53*53), 1.0/(54*54), 1.0/(55*55), 1.0/(56*56), 1.0/(57*57), 1.0/(58*58), 1.0/(59*59), 1.0/(60*60),
45 1.0/(61*61), 1.0/(62*62), 1.0/(63*63), 1.0/(64*64), 1.0/(65*65), 1.0/(66*66), 1.0/(67*67), 1.0/(68*68), 1.0/(69*69), 1.0/(70*70),
46 1.0/(71*71), 1.0/(72*72), 1.0/(73*73), 1.0/(74*74), 1.0/(75*75), 1.0/(76*76), 1.0/(77*77), 1.0/(78*78), 1.0/(79*79), 1.0/(80*80),
47 1.0/(81*81), 1.0/(82*82), 1.0/(83*83), 1.0/(84*84), 1.0/(85*85), 1.0/(86*86), 1.0/(87*87), 1.0/(88*88), 1.0/(89*89), 1.0/(90*90),
48 1.0/(91*91), 1.0/(92*92), 1.0/(93*93), 1.0/(94*94), 1.0/(95*95), 1.0/(96*96), 1.0/(97*97), 1.0/(98*98), 1.0/(99*99), 1.0/(10000)
54 for(i=1; v != lastv; i+=2){
66 * builds a polyphase filterbank.
67 * @param factor resampling factor
68 * @param scale wanted sum of coefficients for each filter
69 * @param filter_type filter type
70 * @param kaiser_beta kaiser window beta
71 * @return 0 on success, negative on error
73 static int build_filter(ResampleContext *c, void *filter, double factor, int tap_count, int alloc, int phase_count, int scale,
74 int filter_type, int kaiser_beta){
77 double *tab = av_malloc_array(tap_count, sizeof(*tab));
78 const int center= (tap_count-1)/2;
81 return AVERROR(ENOMEM);
83 /* if upsampling, only need to interpolate, no filter */
87 for(ph=0;ph<phase_count;ph++) {
89 for(i=0;i<tap_count;i++) {
90 x = M_PI * ((double)(i - center) - (double)ph / phase_count) * factor;
94 case SWR_FILTER_TYPE_CUBIC:{
95 const float d= -0.5; //first order derivative = -0.5
96 x = fabs(((double)(i - center) - (double)ph / phase_count) * factor);
97 if(x<1.0) y= 1 - 3*x*x + 2*x*x*x + d*( -x*x + x*x*x);
98 else y= d*(-4 + 8*x - 5*x*x + x*x*x);
100 case SWR_FILTER_TYPE_BLACKMAN_NUTTALL:
101 w = 2.0*x / (factor*tap_count) + M_PI;
102 y *= 0.3635819 - 0.4891775 * cos(w) + 0.1365995 * cos(2*w) - 0.0106411 * cos(3*w);
104 case SWR_FILTER_TYPE_KAISER:
105 w = 2.0*x / (factor*tap_count*M_PI);
106 y *= bessel(kaiser_beta*sqrt(FFMAX(1-w*w, 0)));
116 /* normalize so that an uniform color remains the same */
118 case AV_SAMPLE_FMT_S16P:
119 for(i=0;i<tap_count;i++)
120 ((int16_t*)filter)[ph * alloc + i] = av_clip(lrintf(tab[i] * scale / norm), INT16_MIN, INT16_MAX);
122 case AV_SAMPLE_FMT_S32P:
123 for(i=0;i<tap_count;i++)
124 ((int32_t*)filter)[ph * alloc + i] = av_clipl_int32(llrint(tab[i] * scale / norm));
126 case AV_SAMPLE_FMT_FLTP:
127 for(i=0;i<tap_count;i++)
128 ((float*)filter)[ph * alloc + i] = tab[i] * scale / norm;
130 case AV_SAMPLE_FMT_DBLP:
131 for(i=0;i<tap_count;i++)
132 ((double*)filter)[ph * alloc + i] = tab[i] * scale / norm;
140 double sine[LEN + tap_count];
141 double filtered[LEN];
142 double maxff=-2, minff=2, maxsf=-2, minsf=2;
143 for(i=0; i<LEN; i++){
144 double ss=0, sf=0, ff=0;
145 for(j=0; j<LEN+tap_count; j++)
146 sine[j]= cos(i*j*M_PI/LEN);
147 for(j=0; j<LEN; j++){
150 for(k=0; k<tap_count; k++)
151 sum += filter[ph * tap_count + k] * sine[k+j];
152 filtered[j]= sum / (1<<FILTER_SHIFT);
153 ss+= sine[j + center] * sine[j + center];
154 ff+= filtered[j] * filtered[j];
155 sf+= sine[j + center] * filtered[j];
160 maxff= FFMAX(maxff, ff);
161 minff= FFMIN(minff, ff);
162 maxsf= FFMAX(maxsf, sf);
163 minsf= FFMIN(minsf, sf);
165 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);
177 static ResampleContext *resample_init(ResampleContext *c, int out_rate, int in_rate, int filter_size, int phase_shift, int linear,
178 double cutoff0, enum AVSampleFormat format, enum SwrFilterType filter_type, int kaiser_beta,
179 double precision, int cheby)
181 double cutoff = cutoff0? cutoff0 : 0.97;
182 double factor= FFMIN(out_rate * cutoff / in_rate, 1.0);
183 int phase_count= 1<<phase_shift;
185 if (!c || c->phase_shift != phase_shift || c->linear!=linear || c->factor != factor
186 || c->filter_length != FFMAX((int)ceil(filter_size/factor), 1) || c->format != format
187 || c->filter_type != filter_type || c->kaiser_beta != kaiser_beta) {
188 c = av_mallocz(sizeof(*c));
194 c->felem_size= av_get_bytes_per_sample(c->format);
197 case AV_SAMPLE_FMT_S16P:
198 c->filter_shift = 15;
200 case AV_SAMPLE_FMT_S32P:
201 c->filter_shift = 30;
203 case AV_SAMPLE_FMT_FLTP:
204 case AV_SAMPLE_FMT_DBLP:
208 av_log(NULL, AV_LOG_ERROR, "Unsupported sample format\n");
212 if (filter_size/factor > INT32_MAX/256) {
213 av_log(NULL, AV_LOG_ERROR, "Filter length too large\n");
217 c->phase_shift = phase_shift;
218 c->phase_mask = phase_count - 1;
221 c->filter_length = FFMAX((int)ceil(filter_size/factor), 1);
222 c->filter_alloc = FFALIGN(c->filter_length, 8);
223 c->filter_bank = av_calloc(c->filter_alloc, (phase_count+1)*c->felem_size);
224 c->filter_type = filter_type;
225 c->kaiser_beta = kaiser_beta;
228 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))
230 memcpy(c->filter_bank + (c->filter_alloc*phase_count+1)*c->felem_size, c->filter_bank, (c->filter_alloc-1)*c->felem_size);
231 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);
234 c->compensation_distance= 0;
235 if(!av_reduce(&c->src_incr, &c->dst_incr, out_rate, in_rate * (int64_t)phase_count, INT32_MAX/2))
237 c->ideal_dst_incr = c->dst_incr;
238 c->dst_incr_div = c->dst_incr / c->src_incr;
239 c->dst_incr_mod = c->dst_incr % c->src_incr;
241 c->index= -phase_count*((c->filter_length-1)/2);
244 swri_resample_dsp_init(c);
248 av_freep(&c->filter_bank);
253 static void resample_free(ResampleContext **c){
256 av_freep(&(*c)->filter_bank);
260 static int set_compensation(ResampleContext *c, int sample_delta, int compensation_distance){
261 c->compensation_distance= compensation_distance;
262 if (compensation_distance)
263 c->dst_incr = c->ideal_dst_incr - c->ideal_dst_incr * (int64_t)sample_delta / compensation_distance;
265 c->dst_incr = c->ideal_dst_incr;
267 c->dst_incr_div = c->dst_incr / c->src_incr;
268 c->dst_incr_mod = c->dst_incr % c->src_incr;
273 static int swri_resample(ResampleContext *c,
274 uint8_t *dst, const uint8_t *src, int *consumed,
275 int src_size, int dst_size, int update_ctx)
277 if (c->filter_length == 1 && c->phase_shift == 0) {
280 int64_t index2= (1LL<<32)*c->frac/c->src_incr + (1LL<<32)*index;
281 int64_t incr= (1LL<<32) * c->dst_incr / c->src_incr;
282 int new_size = (src_size * (int64_t)c->src_incr - frac + c->dst_incr - 1) / c->dst_incr;
284 dst_size= FFMIN(dst_size, new_size);
285 c->dsp.resample_one(dst, src, dst_size, index2, incr);
287 index += dst_size * c->dst_incr_div;
288 index += (frac + dst_size * (int64_t)c->dst_incr_mod) / c->src_incr;
289 av_assert2(index >= 0);
292 c->frac = (frac + dst_size * (int64_t)c->dst_incr_mod) % c->src_incr;
296 int64_t end_index = (1LL + src_size - c->filter_length) << c->phase_shift;
297 int64_t delta_frac = (end_index - c->index) * c->src_incr - c->frac;
298 int delta_n = (delta_frac + c->dst_incr - 1) / c->dst_incr;
300 dst_size = FFMIN(dst_size, delta_n);
302 *consumed = c->dsp.resample(c, dst, src, dst_size, update_ctx);
311 static int multiple_resample(ResampleContext *c, AudioData *dst, int dst_size, AudioData *src, int src_size, int *consumed){
313 int av_unused mm_flags = av_get_cpu_flags();
314 int need_emms = c->format == AV_SAMPLE_FMT_S16P && ARCH_X86_32 &&
315 (mm_flags & (AV_CPU_FLAG_MMX2 | AV_CPU_FLAG_SSE2)) == AV_CPU_FLAG_MMX2;
316 int64_t max_src_size = (INT64_MAX >> (c->phase_shift+1)) / c->src_incr;
318 if (c->compensation_distance)
319 dst_size = FFMIN(dst_size, c->compensation_distance);
320 src_size = FFMIN(src_size, max_src_size);
322 for(i=0; i<dst->ch_count; i++){
323 ret= swri_resample(c, dst->ch[i], src->ch[i],
324 consumed, src_size, dst_size, i+1==dst->ch_count);
329 if (c->compensation_distance) {
330 c->compensation_distance -= ret;
331 if (!c->compensation_distance) {
332 c->dst_incr = c->ideal_dst_incr;
333 c->dst_incr_div = c->dst_incr / c->src_incr;
334 c->dst_incr_mod = c->dst_incr % c->src_incr;
341 static int64_t get_delay(struct SwrContext *s, int64_t base){
342 ResampleContext *c = s->resample;
343 int64_t num = s->in_buffer_count - (c->filter_length-1)/2;
344 num *= 1 << c->phase_shift;
348 return av_rescale(num, base, s->in_sample_rate*(int64_t)c->src_incr << c->phase_shift);
351 static int64_t get_out_samples(struct SwrContext *s, int in_samples) {
352 ResampleContext *c = s->resample;
353 // The + 2 are added to allow implementations to be slightly inaccurate, they should not be needed currently.
354 // They also make it easier to proof that changes and optimizations do not
355 // break the upper bound.
356 int64_t num = s->in_buffer_count + 2LL + in_samples;
357 num *= 1 << c->phase_shift;
359 num = av_rescale_rnd(num, s->out_sample_rate, ((int64_t)s->in_sample_rate) << c->phase_shift, AV_ROUND_UP) + 2;
361 if (c->compensation_distance) {
363 return AVERROR(EINVAL);
365 num = FFMAX(num, (num * c->ideal_dst_incr - 1) / c->dst_incr + 1);
370 static int resample_flush(struct SwrContext *s) {
371 AudioData *a= &s->in_buffer;
373 if((ret = swri_realloc_audio(a, s->in_buffer_index + 2*s->in_buffer_count)) < 0)
375 av_assert0(a->planar);
376 for(i=0; i<a->ch_count; i++){
377 for(j=0; j<s->in_buffer_count; j++){
378 memcpy(a->ch[i] + (s->in_buffer_index+s->in_buffer_count+j )*a->bps,
379 a->ch[i] + (s->in_buffer_index+s->in_buffer_count-j-1)*a->bps, a->bps);
382 s->in_buffer_count += (s->in_buffer_count+1)/2;
386 // in fact the whole handle multiple ridiculously small buffers might need more thinking...
387 static int invert_initial_buffer(ResampleContext *c, AudioData *dst, const AudioData *src,
388 int in_count, int *out_idx, int *out_sz)
390 int n, ch, num = FFMIN(in_count + *out_sz, c->filter_length + 1), res;
395 if ((res = swri_realloc_audio(dst, c->filter_length * 2 + 1)) < 0)
399 for (n = *out_sz; n < num; n++) {
400 for (ch = 0; ch < src->ch_count; ch++) {
401 memcpy(dst->ch[ch] + ((c->filter_length + n) * c->felem_size),
402 src->ch[ch] + ((n - *out_sz) * c->felem_size), c->felem_size);
406 // if not enough data is in, return and wait for more
407 if (num < c->filter_length + 1) {
409 *out_idx = c->filter_length;
414 for (n = 1; n <= c->filter_length; n++) {
415 for (ch = 0; ch < src->ch_count; ch++) {
416 memcpy(dst->ch[ch] + ((c->filter_length - n) * c->felem_size),
417 dst->ch[ch] + ((c->filter_length + n) * c->felem_size),
423 *out_idx = c->filter_length + (c->index >> c->phase_shift);
424 *out_sz = FFMAX(*out_sz + c->filter_length,
425 1 + c->filter_length * 2) - *out_idx;
426 c->index &= c->phase_mask;
428 return FFMAX(res, 0);
431 struct Resampler const swri_resampler={
438 invert_initial_buffer,