vorbis_envelope.c (10672B)
1 /******************************************************************** 2 * * 3 * THIS FILE IS PART OF THE OggVorbis SOFTWARE CODEC SOURCE CODE. * 4 * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS * 5 * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE * 6 * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. * 7 * * 8 * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2009 * 9 * by the Xiph.Org Foundation https://xiph.org/ * 10 * * 11 ******************************************************************** 12 13 function: PCM data envelope analysis 14 15 ********************************************************************/ 16 17 #include <stdlib.h> 18 #include <string.h> 19 #include <stdio.h> 20 #include <math.h> 21 #include <ogg/ogg.h> 22 #include "vorbis/codec.h" 23 #include "codec_internal.h" 24 25 #include "os.h" 26 #include "scales.h" 27 #include "envelope.h" 28 #include "mdct.h" 29 #include "misc.h" 30 31 void _ve_envelope_init(envelope_lookup *e,vorbis_info *vi){ 32 codec_setup_info *ci=vi->codec_setup; 33 vorbis_info_psy_global *gi=&ci->psy_g_param; 34 int ch=vi->channels; 35 int i,j; 36 int n=e->winlength=128; 37 e->searchstep=64; /* not random */ 38 39 e->minenergy=gi->preecho_minenergy; 40 e->ch=ch; 41 e->storage=128; 42 e->cursor=ci->blocksizes[1]/2; 43 e->mdct_win=_ogg_calloc(n,sizeof(*e->mdct_win)); 44 mdct_init(&e->mdct,n); 45 46 for(i=0;i<n;i++){ 47 e->mdct_win[i]=sin(i/(n-1.)*M_PI); 48 e->mdct_win[i]*=e->mdct_win[i]; 49 } 50 51 /* magic follows */ 52 e->band[0].begin=2; e->band[0].end=4; 53 e->band[1].begin=4; e->band[1].end=5; 54 e->band[2].begin=6; e->band[2].end=6; 55 e->band[3].begin=9; e->band[3].end=8; 56 e->band[4].begin=13; e->band[4].end=8; 57 e->band[5].begin=17; e->band[5].end=8; 58 e->band[6].begin=22; e->band[6].end=8; 59 60 for(j=0;j<VE_BANDS;j++){ 61 n=e->band[j].end; 62 e->band[j].window=_ogg_malloc(n*sizeof(*e->band[0].window)); 63 for(i=0;i<n;i++){ 64 e->band[j].window[i]=sin((i+.5)/n*M_PI); 65 e->band[j].total+=e->band[j].window[i]; 66 } 67 e->band[j].total=1./e->band[j].total; 68 } 69 70 e->filter=_ogg_calloc(VE_BANDS*ch,sizeof(*e->filter)); 71 e->mark=_ogg_calloc(e->storage,sizeof(*e->mark)); 72 73 } 74 75 void _ve_envelope_clear(envelope_lookup *e){ 76 int i; 77 mdct_clear(&e->mdct); 78 for(i=0;i<VE_BANDS;i++) 79 _ogg_free(e->band[i].window); 80 _ogg_free(e->mdct_win); 81 _ogg_free(e->filter); 82 _ogg_free(e->mark); 83 memset(e,0,sizeof(*e)); 84 } 85 86 /* fairly straight threshhold-by-band based until we find something 87 that works better and isn't patented. */ 88 89 static int _ve_amp(envelope_lookup *ve, 90 vorbis_info_psy_global *gi, 91 float *data, 92 envelope_band *bands, 93 envelope_filter_state *filters){ 94 long n=ve->winlength; 95 int ret=0; 96 long i,j; 97 float decay; 98 99 /* we want to have a 'minimum bar' for energy, else we're just 100 basing blocks on quantization noise that outweighs the signal 101 itself (for low power signals) */ 102 103 float minV=ve->minenergy; 104 float *vec=alloca(n*sizeof(*vec)); 105 106 /* stretch is used to gradually lengthen the number of windows 107 considered prevoius-to-potential-trigger */ 108 int stretch=max(VE_MINSTRETCH,ve->stretch/2); 109 float penalty=gi->stretch_penalty-(ve->stretch/2-VE_MINSTRETCH); 110 if(penalty<0.f)penalty=0.f; 111 if(penalty>gi->stretch_penalty)penalty=gi->stretch_penalty; 112 113 /*_analysis_output_always("lpcm",seq2,data,n,0,0, 114 totalshift+pos*ve->searchstep);*/ 115 116 /* window and transform */ 117 for(i=0;i<n;i++) 118 vec[i]=data[i]*ve->mdct_win[i]; 119 mdct_forward(&ve->mdct,vec,vec); 120 121 /*_analysis_output_always("mdct",seq2,vec,n/2,0,1,0); */ 122 123 /* near-DC spreading function; this has nothing to do with 124 psychoacoustics, just sidelobe leakage and window size */ 125 { 126 float temp=vec[0]*vec[0]+.7*vec[1]*vec[1]+.2*vec[2]*vec[2]; 127 int ptr=filters->nearptr; 128 129 /* the accumulation is regularly refreshed from scratch to avoid 130 floating point creep */ 131 if(ptr==0){ 132 decay=filters->nearDC_acc=filters->nearDC_partialacc+temp; 133 filters->nearDC_partialacc=temp; 134 }else{ 135 decay=filters->nearDC_acc+=temp; 136 filters->nearDC_partialacc+=temp; 137 } 138 filters->nearDC_acc-=filters->nearDC[ptr]; 139 filters->nearDC[ptr]=temp; 140 141 decay*=(1./(VE_NEARDC+1)); 142 filters->nearptr++; 143 if(filters->nearptr>=VE_NEARDC)filters->nearptr=0; 144 decay=todB(&decay)*.5-15.f; 145 } 146 147 /* perform spreading and limiting, also smooth the spectrum. yes, 148 the MDCT results in all real coefficients, but it still *behaves* 149 like real/imaginary pairs */ 150 for(i=0;i<n/2;i+=2){ 151 float val=vec[i]*vec[i]+vec[i+1]*vec[i+1]; 152 val=todB(&val)*.5f; 153 if(val<decay)val=decay; 154 if(val<minV)val=minV; 155 vec[i>>1]=val; 156 decay-=8.; 157 } 158 159 /*_analysis_output_always("spread",seq2++,vec,n/4,0,0,0);*/ 160 161 /* perform preecho/postecho triggering by band */ 162 for(j=0;j<VE_BANDS;j++){ 163 float acc=0.; 164 float valmax,valmin; 165 166 /* accumulate amplitude */ 167 for(i=0;i<bands[j].end;i++) 168 acc+=vec[i+bands[j].begin]*bands[j].window[i]; 169 170 acc*=bands[j].total; 171 172 /* convert amplitude to delta */ 173 { 174 int p,this=filters[j].ampptr; 175 float postmax,postmin,premax=-99999.f,premin=99999.f; 176 177 p=this; 178 p--; 179 if(p<0)p+=VE_AMP; 180 postmax=max(acc,filters[j].ampbuf[p]); 181 postmin=min(acc,filters[j].ampbuf[p]); 182 183 for(i=0;i<stretch;i++){ 184 p--; 185 if(p<0)p+=VE_AMP; 186 premax=max(premax,filters[j].ampbuf[p]); 187 premin=min(premin,filters[j].ampbuf[p]); 188 } 189 190 valmin=postmin-premin; 191 valmax=postmax-premax; 192 193 /*filters[j].markers[pos]=valmax;*/ 194 filters[j].ampbuf[this]=acc; 195 filters[j].ampptr++; 196 if(filters[j].ampptr>=VE_AMP)filters[j].ampptr=0; 197 } 198 199 /* look at min/max, decide trigger */ 200 if(valmax>gi->preecho_thresh[j]+penalty){ 201 ret|=1; 202 ret|=4; 203 } 204 if(valmin<gi->postecho_thresh[j]-penalty)ret|=2; 205 } 206 207 return(ret); 208 } 209 210 #if 0 211 static int seq=0; 212 static ogg_int64_t totalshift=-1024; 213 #endif 214 215 long _ve_envelope_search(vorbis_dsp_state *v){ 216 vorbis_info *vi=v->vi; 217 codec_setup_info *ci=vi->codec_setup; 218 vorbis_info_psy_global *gi=&ci->psy_g_param; 219 envelope_lookup *ve=((private_state *)(v->backend_state))->ve; 220 long i,j; 221 222 int first=ve->current/ve->searchstep; 223 int last=v->pcm_current/ve->searchstep-VE_WIN; 224 if(first<0)first=0; 225 226 /* make sure we have enough storage to match the PCM */ 227 if(last+VE_WIN+VE_POST>ve->storage){ 228 ve->storage=last+VE_WIN+VE_POST; /* be sure */ 229 ve->mark=_ogg_realloc(ve->mark,ve->storage*sizeof(*ve->mark)); 230 } 231 232 for(j=first;j<last;j++){ 233 int ret=0; 234 235 ve->stretch++; 236 if(ve->stretch>VE_MAXSTRETCH*2) 237 ve->stretch=VE_MAXSTRETCH*2; 238 239 for(i=0;i<ve->ch;i++){ 240 float *pcm=v->pcm[i]+ve->searchstep*(j); 241 ret|=_ve_amp(ve,gi,pcm,ve->band,ve->filter+i*VE_BANDS); 242 } 243 244 ve->mark[j+VE_POST]=0; 245 if(ret&1){ 246 ve->mark[j]=1; 247 ve->mark[j+1]=1; 248 } 249 250 if(ret&2){ 251 ve->mark[j]=1; 252 if(j>0)ve->mark[j-1]=1; 253 } 254 255 if(ret&4)ve->stretch=-1; 256 } 257 258 ve->current=last*ve->searchstep; 259 260 { 261 long centerW=v->centerW; 262 long testW= 263 centerW+ 264 ci->blocksizes[v->W]/4+ 265 ci->blocksizes[1]/2+ 266 ci->blocksizes[0]/4; 267 268 j=ve->cursor; 269 270 while(j<ve->current-(ve->searchstep)){/* account for postecho 271 working back one window */ 272 if(j>=testW)return(1); 273 274 ve->cursor=j; 275 276 if(ve->mark[j/ve->searchstep]){ 277 if(j>centerW){ 278 279 #if 0 280 if(j>ve->curmark){ 281 float *marker=alloca(v->pcm_current*sizeof(*marker)); 282 int l,m; 283 memset(marker,0,sizeof(*marker)*v->pcm_current); 284 fprintf(stderr,"mark! seq=%d, cursor:%fs time:%fs\n", 285 seq, 286 (totalshift+ve->cursor)/44100., 287 (totalshift+j)/44100.); 288 _analysis_output_always("pcmL",seq,v->pcm[0],v->pcm_current,0,0,totalshift); 289 _analysis_output_always("pcmR",seq,v->pcm[1],v->pcm_current,0,0,totalshift); 290 291 _analysis_output_always("markL",seq,v->pcm[0],j,0,0,totalshift); 292 _analysis_output_always("markR",seq,v->pcm[1],j,0,0,totalshift); 293 294 for(m=0;m<VE_BANDS;m++){ 295 char buf[80]; 296 sprintf(buf,"delL%d",m); 297 for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m].markers[l]*.1; 298 _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift); 299 } 300 301 for(m=0;m<VE_BANDS;m++){ 302 char buf[80]; 303 sprintf(buf,"delR%d",m); 304 for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m+VE_BANDS].markers[l]*.1; 305 _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift); 306 } 307 308 for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->mark[l]*.4; 309 _analysis_output_always("mark",seq,marker,v->pcm_current,0,0,totalshift); 310 311 312 seq++; 313 314 } 315 #endif 316 317 ve->curmark=j; 318 if(j>=testW)return(1); 319 return(0); 320 } 321 } 322 j+=ve->searchstep; 323 } 324 } 325 326 return(-1); 327 } 328 329 int _ve_envelope_mark(vorbis_dsp_state *v){ 330 envelope_lookup *ve=((private_state *)(v->backend_state))->ve; 331 vorbis_info *vi=v->vi; 332 codec_setup_info *ci=vi->codec_setup; 333 long centerW=v->centerW; 334 long beginW=centerW-ci->blocksizes[v->W]/4; 335 long endW=centerW+ci->blocksizes[v->W]/4; 336 if(v->W){ 337 beginW-=ci->blocksizes[v->lW]/4; 338 endW+=ci->blocksizes[v->nW]/4; 339 }else{ 340 beginW-=ci->blocksizes[0]/4; 341 endW+=ci->blocksizes[0]/4; 342 } 343 344 if(ve->curmark>=beginW && ve->curmark<endW)return(1); 345 { 346 long first=beginW/ve->searchstep; 347 long last=endW/ve->searchstep; 348 long i; 349 for(i=first;i<last;i++) 350 if(ve->mark[i])return(1); 351 } 352 return(0); 353 } 354 355 void _ve_envelope_shift(envelope_lookup *e,long shift){ 356 int smallsize=e->current/e->searchstep+VE_POST; /* adjust for placing marks 357 ahead of ve->current */ 358 int smallshift=shift/e->searchstep; 359 360 memmove(e->mark,e->mark+smallshift,(smallsize-smallshift)*sizeof(*e->mark)); 361 362 #if 0 363 for(i=0;i<VE_BANDS*e->ch;i++) 364 memmove(e->filter[i].markers, 365 e->filter[i].markers+smallshift, 366 (1024-smallshift)*sizeof(*(*e->filter).markers)); 367 totalshift+=shift; 368 #endif 369 370 e->current-=shift; 371 if(e->curmark>=0) 372 e->curmark-=shift; 373 e->cursor-=shift; 374 }