FFmpeg: libswresample/resample.c Source File

00001 /*
00002  * audio resampling
00003  * Copyright (c) 2004-2012 Michael Niedermayer <michaelni@gmx.at>
00004  *
00005  * This file is part of FFmpeg.
00006  *
00007  * FFmpeg is free software; you can redistribute it and/or
00008  * modify it under the terms of the GNU Lesser General Public
00009  * License as published by the Free Software Foundation; either
00010  * version 2.1 of the License, or (at your option) any later version.
00011  *
00012  * FFmpeg is distributed in the hope that it will be useful,
00013  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00014  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00015  * Lesser General Public License for more details.
00016  *
00017  * You should have received a copy of the GNU Lesser General Public
00018  * License along with FFmpeg; if not, write to the Free Software
00019  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00020  */
00021 
00028 #include "libavutil/log.h"
00029 #include "libavutil/avassert.h"
00030 #include "swresample_internal.h"
00031 
00032 
00033 typedef struct ResampleContext {
00034     const AVClass *av_class;
00035     uint8_t *filter_bank;
00036     int filter_length;
00037     int filter_alloc;
00038     int ideal_dst_incr;
00039     int dst_incr;
00040     int index;
00041     int frac;
00042     int src_incr;
00043     int compensation_distance;
00044     int phase_shift;
00045     int phase_mask;
00046     int linear;
00047     enum SwrFilterType filter_type;
00048     int kaiser_beta;
00049     double factor;
00050     enum AVSampleFormat format;
00051     int felem_size;
00052     int filter_shift;
00053 } ResampleContext;
00054 
00058 static double bessel(double x){
00059     double v=1;
00060     double lastv=0;
00061     double t=1;
00062     int i;
00063     static const double inv[100]={
00064  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),
00065  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),
00066  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),
00067  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),
00068  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),
00069  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),
00070  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),
00071  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),
00072  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),
00073  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)
00074     };
00075 
00076     x= x*x/4;
00077     for(i=0; v != lastv; i++){
00078         lastv=v;
00079         t *= x*inv[i];
00080         v += t;
00081         av_assert2(i<99);
00082     }
00083     return v;
00084 }
00085 
00094 static int build_filter(ResampleContext *c, void *filter, double factor, int tap_count, int alloc, int phase_count, int scale,
00095                         int filter_type, int kaiser_beta){
00096     int ph, i;
00097     double x, y, w;
00098     double *tab = av_malloc(tap_count * sizeof(*tab));
00099     const int center= (tap_count-1)/2;
00100 
00101     if (!tab)
00102         return AVERROR(ENOMEM);
00103 
00104     /* if upsampling, only need to interpolate, no filter */
00105     if (factor > 1.0)
00106         factor = 1.0;
00107 
00108     for(ph=0;ph<phase_count;ph++) {
00109         double norm = 0;
00110         for(i=0;i<tap_count;i++) {
00111             x = M_PI * ((double)(i - center) - (double)ph / phase_count) * factor;
00112             if (x == 0) y = 1.0;
00113             else        y = sin(x) / x;
00114             switch(filter_type){
00115             case SWR_FILTER_TYPE_CUBIC:{
00116                 const float d= -0.5; //first order derivative = -0.5
00117                 x = fabs(((double)(i - center) - (double)ph / phase_count) * factor);
00118                 if(x<1.0) y= 1 - 3*x*x + 2*x*x*x + d*(            -x*x + x*x*x);
00119                 else      y=                       d*(-4 + 8*x - 5*x*x + x*x*x);
00120                 break;}
00121             case SWR_FILTER_TYPE_BLACKMAN_NUTTALL:
00122                 w = 2.0*x / (factor*tap_count) + M_PI;
00123                 y *= 0.3635819 - 0.4891775 * cos(w) + 0.1365995 * cos(2*w) - 0.0106411 * cos(3*w);
00124                 break;
00125             case SWR_FILTER_TYPE_KAISER:
00126                 w = 2.0*x / (factor*tap_count*M_PI);
00127                 y *= bessel(kaiser_beta*sqrt(FFMAX(1-w*w, 0)));
00128                 break;
00129             default:
00130                 av_assert0(0);
00131             }
00132 
00133             tab[i] = y;
00134             norm += y;
00135         }
00136 
00137         /* normalize so that an uniform color remains the same */
00138         switch(c->format){
00139         case AV_SAMPLE_FMT_S16P:
00140             for(i=0;i<tap_count;i++)
00141                 ((int16_t*)filter)[ph * alloc + i] = av_clip(lrintf(tab[i] * scale / norm), INT16_MIN, INT16_MAX);
00142             break;
00143         case AV_SAMPLE_FMT_S32P:
00144             for(i=0;i<tap_count;i++)
00145                 ((int32_t*)filter)[ph * alloc + i] = av_clip(lrintf(tab[i] * scale / norm), INT32_MIN, INT32_MAX);
00146             break;
00147         case AV_SAMPLE_FMT_FLTP:
00148             for(i=0;i<tap_count;i++)
00149                 ((float*)filter)[ph * alloc + i] = tab[i] * scale / norm;
00150             break;
00151         case AV_SAMPLE_FMT_DBLP:
00152             for(i=0;i<tap_count;i++)
00153                 ((double*)filter)[ph * alloc + i] = tab[i] * scale / norm;
00154             break;
00155         }
00156     }
00157 #if 0
00158     {
00159 #define LEN 1024
00160         int j,k;
00161         double sine[LEN + tap_count];
00162         double filtered[LEN];
00163         double maxff=-2, minff=2, maxsf=-2, minsf=2;
00164         for(i=0; i<LEN; i++){
00165             double ss=0, sf=0, ff=0;
00166             for(j=0; j<LEN+tap_count; j++)
00167                 sine[j]= cos(i*j*M_PI/LEN);
00168             for(j=0; j<LEN; j++){
00169                 double sum=0;
00170                 ph=0;
00171                 for(k=0; k<tap_count; k++)
00172                     sum += filter[ph * tap_count + k] * sine[k+j];
00173                 filtered[j]= sum / (1<<FILTER_SHIFT);
00174                 ss+= sine[j + center] * sine[j + center];
00175                 ff+= filtered[j] * filtered[j];
00176                 sf+= sine[j + center] * filtered[j];
00177             }
00178             ss= sqrt(2*ss/LEN);
00179             ff= sqrt(2*ff/LEN);
00180             sf= 2*sf/LEN;
00181             maxff= FFMAX(maxff, ff);
00182             minff= FFMIN(minff, ff);
00183             maxsf= FFMAX(maxsf, sf);
00184             minsf= FFMIN(minsf, sf);
00185             if(i%11==0){
00186                 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);
00187                 minff=minsf= 2;
00188                 maxff=maxsf= -2;
00189             }
00190         }
00191     }
00192 #endif
00193 
00194     av_free(tab);
00195     return 0;
00196 }
00197 
00198 ResampleContext *swri_resample_init(ResampleContext *c, int out_rate, int in_rate, int filter_size, int phase_shift, int linear,
00199                                     double cutoff, enum AVSampleFormat format, enum SwrFilterType filter_type, int kaiser_beta){
00200     double factor= FFMIN(out_rate * cutoff / in_rate, 1.0);
00201     int phase_count= 1<<phase_shift;
00202 
00203     if (!c || c->phase_shift != phase_shift || c->linear!=linear || c->factor != factor
00204            || c->filter_length != FFMAX((int)ceil(filter_size/factor), 1) || c->format != format
00205            || c->filter_type != filter_type || c->kaiser_beta != kaiser_beta) {
00206         c = av_mallocz(sizeof(*c));
00207         if (!c)
00208             return NULL;
00209 
00210         c->format= format;
00211 
00212         c->felem_size= av_get_bytes_per_sample(c->format);
00213 
00214         switch(c->format){
00215         case AV_SAMPLE_FMT_S16P:
00216             c->filter_shift = 15;
00217             break;
00218         case AV_SAMPLE_FMT_S32P:
00219             c->filter_shift = 30;
00220             break;
00221         case AV_SAMPLE_FMT_FLTP:
00222         case AV_SAMPLE_FMT_DBLP:
00223             c->filter_shift = 0;
00224             break;
00225         default:
00226             av_log(NULL, AV_LOG_ERROR, "Unsupported sample format\n");
00227             return NULL;
00228         }
00229 
00230         c->phase_shift   = phase_shift;
00231         c->phase_mask    = phase_count - 1;
00232         c->linear        = linear;
00233         c->factor        = factor;
00234         c->filter_length = FFMAX((int)ceil(filter_size/factor), 1);
00235         c->filter_alloc  = FFALIGN(c->filter_length, 8);
00236         c->filter_bank   = av_mallocz(c->filter_alloc*(phase_count+1)*c->felem_size);
00237         c->filter_type   = filter_type;
00238         c->kaiser_beta   = kaiser_beta;
00239         if (!c->filter_bank)
00240             goto error;
00241         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))
00242             goto error;
00243         memcpy(c->filter_bank + (c->filter_alloc*phase_count+1)*c->felem_size, c->filter_bank, (c->filter_alloc-1)*c->felem_size);
00244         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);
00245     }
00246 
00247     c->compensation_distance= 0;
00248     if(!av_reduce(&c->src_incr, &c->dst_incr, out_rate, in_rate * (int64_t)phase_count, INT32_MAX/2))
00249         goto error;
00250     c->ideal_dst_incr= c->dst_incr;
00251 
00252     c->index= -phase_count*((c->filter_length-1)/2);
00253     c->frac= 0;
00254 
00255     return c;
00256 error:
00257     av_free(c->filter_bank);
00258     av_free(c);
00259     return NULL;
00260 }
00261 
00262 void swri_resample_free(ResampleContext **c){
00263     if(!*c)
00264         return;
00265     av_freep(&(*c)->filter_bank);
00266     av_freep(c);
00267 }
00268 
00269 int swr_set_compensation(struct SwrContext *s, int sample_delta, int compensation_distance){
00270     ResampleContext *c;
00271     int ret;
00272 
00273     if (!s || compensation_distance < 0)
00274         return AVERROR(EINVAL);
00275     if (!compensation_distance && sample_delta)
00276         return AVERROR(EINVAL);
00277     if (!s->resample) {
00278         s->flags |= SWR_FLAG_RESAMPLE;
00279         ret = swr_init(s);
00280         if (ret < 0)
00281             return ret;
00282     }
00283     c= s->resample;
00284     c->compensation_distance= compensation_distance;
00285     if (compensation_distance)
00286         c->dst_incr = c->ideal_dst_incr - c->ideal_dst_incr * (int64_t)sample_delta / compensation_distance;
00287     else
00288         c->dst_incr = c->ideal_dst_incr;
00289     return 0;
00290 }
00291 
00292 #define RENAME(N) N ## _int16
00293 #define FILTER_SHIFT 15
00294 #define DELEM  int16_t
00295 #define FELEM  int16_t
00296 #define FELEM2 int32_t
00297 #define FELEML int64_t
00298 #define FELEM_MAX INT16_MAX
00299 #define FELEM_MIN INT16_MIN
00300 #define OUT(d, v) v = (v + (1<<(FILTER_SHIFT-1)))>>FILTER_SHIFT;\
00301                   d = (unsigned)(v + 32768) > 65535 ? (v>>31) ^ 32767 : v
00302 #include "resample_template.c"
00303 
00304 #undef RENAME
00305 #undef FELEM
00306 #undef FELEM2
00307 #undef DELEM
00308 #undef FELEML
00309 #undef OUT
00310 #undef FELEM_MIN
00311 #undef FELEM_MAX
00312 #undef FILTER_SHIFT
00313 
00314 
00315 #define RENAME(N) N ## _int32
00316 #define FILTER_SHIFT 30
00317 #define DELEM  int32_t
00318 #define FELEM  int32_t
00319 #define FELEM2 int64_t
00320 #define FELEML int64_t
00321 #define FELEM_MAX INT32_MAX
00322 #define FELEM_MIN INT32_MIN
00323 #define OUT(d, v) v = (v + (1<<(FILTER_SHIFT-1)))>>FILTER_SHIFT;\
00324                   d = (uint64_t)(v + 0x80000000) > 0xFFFFFFFF ? (v>>63) ^ 0x7FFFFFFF : v
00325 #include "resample_template.c"
00326 
00327 #undef RENAME
00328 #undef FELEM
00329 #undef FELEM2
00330 #undef DELEM
00331 #undef FELEML
00332 #undef OUT
00333 #undef FELEM_MIN
00334 #undef FELEM_MAX
00335 #undef FILTER_SHIFT
00336 
00337 
00338 #define RENAME(N) N ## _float
00339 #define FILTER_SHIFT 0
00340 #define DELEM  float
00341 #define FELEM  float
00342 #define FELEM2 float
00343 #define FELEML float
00344 #define OUT(d, v) d = v
00345 #include "resample_template.c"
00346 
00347 #undef RENAME
00348 #undef FELEM
00349 #undef FELEM2
00350 #undef DELEM
00351 #undef FELEML
00352 #undef OUT
00353 #undef FELEM_MIN
00354 #undef FELEM_MAX
00355 #undef FILTER_SHIFT
00356 
00357 
00358 #define RENAME(N) N ## _double
00359 #define FILTER_SHIFT 0
00360 #define DELEM  double
00361 #define FELEM  double
00362 #define FELEM2 double
00363 #define FELEML double
00364 #define OUT(d, v) d = v
00365 #include "resample_template.c"
00366 
00367 #undef RENAME
00368 #undef FELEM
00369 #undef FELEM2
00370 #undef DELEM
00371 #undef FELEML
00372 #undef OUT
00373 #undef FELEM_MIN
00374 #undef FELEM_MAX
00375 #undef FILTER_SHIFT
00376 
00377 // XXX FIXME the whole C loop should be written in asm so this x86 specific code here isnt needed
00378 #if HAVE_MMXEXT_INLINE
00379 #include "x86/resample_mmx.h"
00380 #define COMMON_CORE COMMON_CORE_INT16_MMX2
00381 #define RENAME(N) N ## _int16_mmx2
00382 #define FILTER_SHIFT 15
00383 #define DELEM  int16_t
00384 #define FELEM  int16_t
00385 #define FELEM2 int32_t
00386 #define FELEML int64_t
00387 #define FELEM_MAX INT16_MAX
00388 #define FELEM_MIN INT16_MIN
00389 #define OUT(d, v) v = (v + (1<<(FILTER_SHIFT-1)))>>FILTER_SHIFT;\
00390                   d = (unsigned)(v + 32768) > 65535 ? (v>>31) ^ 32767 : v
00391 #include "resample_template.c"
00392 
00393 #undef COMMON_CORE
00394 #undef RENAME
00395 #undef FELEM
00396 #undef FELEM2
00397 #undef DELEM
00398 #undef FELEML
00399 #undef OUT
00400 #undef FELEM_MIN
00401 #undef FELEM_MAX
00402 #undef FILTER_SHIFT
00403 
00404 #if HAVE_SSSE3_INLINE
00405 #define COMMON_CORE COMMON_CORE_INT16_SSSE3
00406 #define RENAME(N) N ## _int16_ssse3
00407 #define FILTER_SHIFT 15
00408 #define DELEM  int16_t
00409 #define FELEM  int16_t
00410 #define FELEM2 int32_t
00411 #define FELEML int64_t
00412 #define FELEM_MAX INT16_MAX
00413 #define FELEM_MIN INT16_MIN
00414 #define OUT(d, v) v = (v + (1<<(FILTER_SHIFT-1)))>>FILTER_SHIFT;\
00415                   d = (unsigned)(v + 32768) > 65535 ? (v>>31) ^ 32767 : v
00416 #include "resample_template.c"
00417 #endif
00418 #endif // HAVE_MMXEXT_INLINE
00419 
00420 int swri_multiple_resample(ResampleContext *c, AudioData *dst, int dst_size, AudioData *src, int src_size, int *consumed){
00421     int i, ret= -1;
00422     int av_unused mm_flags = av_get_cpu_flags();
00423     int need_emms= 0;
00424 
00425     for(i=0; i<dst->ch_count; i++){
00426 #if HAVE_MMXEXT_INLINE
00427 #if HAVE_SSSE3_INLINE
00428              if(c->format == AV_SAMPLE_FMT_S16P && (mm_flags&AV_CPU_FLAG_SSSE3)) ret= swri_resample_int16_ssse3(c, (int16_t*)dst->ch[i], (const int16_t*)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
00429         else
00430 #endif
00431              if(c->format == AV_SAMPLE_FMT_S16P && (mm_flags&AV_CPU_FLAG_MMX2 )){
00432                  ret= swri_resample_int16_mmx2 (c, (int16_t*)dst->ch[i], (const int16_t*)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
00433                  need_emms= 1;
00434              } else
00435 #endif
00436              if(c->format == AV_SAMPLE_FMT_S16P) ret= swri_resample_int16(c, (int16_t*)dst->ch[i], (const int16_t*)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
00437         else if(c->format == AV_SAMPLE_FMT_S32P) ret= swri_resample_int32(c, (int32_t*)dst->ch[i], (const int32_t*)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
00438         else if(c->format == AV_SAMPLE_FMT_FLTP) ret= swri_resample_float(c, (float  *)dst->ch[i], (const float  *)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
00439         else if(c->format == AV_SAMPLE_FMT_DBLP) ret= swri_resample_double(c,(double *)dst->ch[i], (const double *)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
00440     }
00441     if(need_emms)
00442         emms_c();
00443     return ret;
00444 }
00445 
00446 int64_t swr_get_delay(struct SwrContext *s, int64_t base){
00447     ResampleContext *c = s->resample;
00448     if(c){
00449         int64_t num = s->in_buffer_count - (c->filter_length-1)/2;
00450         num <<= c->phase_shift;
00451         num -= c->index;
00452         num *= c->src_incr;
00453         num -= c->frac;
00454 
00455         return av_rescale(num, base, s->in_sample_rate*(int64_t)c->src_incr << c->phase_shift);
00456     }else{
00457         return (s->in_buffer_count*base + (s->in_sample_rate>>1))/ s->in_sample_rate;
00458     }
00459 }