envelope.c (10444B)
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 1994-2002 * 9 * by the XIPHOPHORUS Company http://www.xiph.org/ * 10 * * 11 ******************************************************************** 12 13 function: PCM data envelope analysis 14 last mod: $Id: envelope.c 1919 2005-07-24 14:18:04Z baford $ 15 16 ********************************************************************/ 17 18 #include <stdlib.h> 19 #include <string.h> 20 #include <stdio.h> 21 #include <math.h> 22 #include <ogg/ogg.h> 23 #include "vorbis/codec.h" 24 #include "codec_internal.h" 25 26 #include "os.h" 27 #include "scales.h" 28 #include "envelope.h" 29 #include "mdct.h" 30 #include "misc.h" 31 32 void _ve_envelope_init(envelope_lookup *e,vorbis_info *vi){ 33 codec_setup_info *ci=vi->codec_setup; 34 vorbis_info_psy_global *gi=&ci->psy_g_param; 35 int ch=vi->channels; 36 int i,j; 37 int n=e->winlength=128; 38 e->searchstep=64; /* not random */ 39 40 e->minenergy=gi->preecho_minenergy; 41 e->ch=ch; 42 e->storage=128; 43 e->cursor=ci->blocksizes[1]/2; 44 e->mdct_win=_ogg_calloc(n,sizeof(*e->mdct_win)); 45 mdct_init(&e->mdct,n); 46 47 for(i=0;i<n;i++){ 48 e->mdct_win[i]=sin(i/(n-1.)*M_PI); 49 e->mdct_win[i]*=e->mdct_win[i]; 50 } 51 52 /* magic follows */ 53 e->band[0].begin=2; e->band[0].end=4; 54 e->band[1].begin=4; e->band[1].end=5; 55 e->band[2].begin=6; e->band[2].end=6; 56 e->band[3].begin=9; e->band[3].end=8; 57 e->band[4].begin=13; e->band[4].end=8; 58 e->band[5].begin=17; e->band[5].end=8; 59 e->band[6].begin=22; e->band[6].end=8; 60 61 for(j=0;j<VE_BANDS;j++){ 62 n=e->band[j].end; 63 e->band[j].window=_ogg_malloc(n*sizeof(*e->band[0].window)); 64 for(i=0;i<n;i++){ 65 e->band[j].window[i]=sin((i+.5)/n*M_PI); 66 e->band[j].total+=e->band[j].window[i]; 67 } 68 e->band[j].total=1./e->band[j].total; 69 } 70 71 e->filter=_ogg_calloc(VE_BANDS*ch,sizeof(*e->filter)); 72 e->mark=_ogg_calloc(e->storage,sizeof(*e->mark)); 73 74 } 75 76 void _ve_envelope_clear(envelope_lookup *e){ 77 int i; 78 mdct_clear(&e->mdct); 79 for(i=0;i<VE_BANDS;i++) 80 _ogg_free(e->band[i].window); 81 _ogg_free(e->mdct_win); 82 _ogg_free(e->filter); 83 _ogg_free(e->mark); 84 memset(e,0,sizeof(*e)); 85 } 86 87 /* fairly straight threshhold-by-band based until we find something 88 that works better and isn't patented. */ 89 90 static int _ve_amp(envelope_lookup *ve, 91 vorbis_info_psy_global *gi, 92 float *data, 93 envelope_band *bands, 94 envelope_filter_state *filters, 95 long pos){ 96 long n=ve->winlength; 97 int ret=0; 98 long i,j; 99 float decay; 100 101 /* we want to have a 'minimum bar' for energy, else we're just 102 basing blocks on quantization noise that outweighs the signal 103 itself (for low power signals) */ 104 105 float minV=ve->minenergy; 106 float *vec=alloca(n*sizeof(*vec)); 107 108 /* stretch is used to gradually lengthen the number of windows 109 considered prevoius-to-potential-trigger */ 110 int stretch=max(VE_MINSTRETCH,ve->stretch/2); 111 float penalty=gi->stretch_penalty-(ve->stretch/2-VE_MINSTRETCH); 112 if(penalty<0.f)penalty=0.f; 113 if(penalty>gi->stretch_penalty)penalty=gi->stretch_penalty; 114 115 /*_analysis_output_always("lpcm",seq2,data,n,0,0, 116 totalshift+pos*ve->searchstep);*/ 117 118 /* window and transform */ 119 for(i=0;i<n;i++) 120 vec[i]=data[i]*ve->mdct_win[i]; 121 mdct_forward(&ve->mdct,vec,vec); 122 123 /*_analysis_output_always("mdct",seq2,vec,n/2,0,1,0); */ 124 125 /* near-DC spreading function; this has nothing to do with 126 psychoacoustics, just sidelobe leakage and window size */ 127 { 128 float temp=vec[0]*vec[0]+.7*vec[1]*vec[1]+.2*vec[2]*vec[2]; 129 int ptr=filters->nearptr; 130 131 /* the accumulation is regularly refreshed from scratch to avoid 132 floating point creep */ 133 if(ptr==0){ 134 decay=filters->nearDC_acc=filters->nearDC_partialacc+temp; 135 filters->nearDC_partialacc=temp; 136 }else{ 137 decay=filters->nearDC_acc+=temp; 138 filters->nearDC_partialacc+=temp; 139 } 140 filters->nearDC_acc-=filters->nearDC[ptr]; 141 filters->nearDC[ptr]=temp; 142 143 decay*=(1./(VE_NEARDC+1)); 144 filters->nearptr++; 145 if(filters->nearptr>=VE_NEARDC)filters->nearptr=0; 146 decay=todB(&decay)*.5-15.f; 147 } 148 149 /* perform spreading and limiting, also smooth the spectrum. yes, 150 the MDCT results in all real coefficients, but it still *behaves* 151 like real/imaginary pairs */ 152 for(i=0;i<n/2;i+=2){ 153 float val=vec[i]*vec[i]+vec[i+1]*vec[i+1]; 154 val=todB(&val)*.5f; 155 if(val<decay)val=decay; 156 if(val<minV)val=minV; 157 vec[i>>1]=val; 158 decay-=8.; 159 } 160 161 /*_analysis_output_always("spread",seq2++,vec,n/4,0,0,0);*/ 162 163 /* perform preecho/postecho triggering by band */ 164 for(j=0;j<VE_BANDS;j++){ 165 float acc=0.; 166 float valmax,valmin; 167 168 /* accumulate amplitude */ 169 for(i=0;i<bands[j].end;i++) 170 acc+=vec[i+bands[j].begin]*bands[j].window[i]; 171 172 acc*=bands[j].total; 173 174 /* convert amplitude to delta */ 175 { 176 int p,this=filters[j].ampptr; 177 float postmax,postmin,premax=-99999.f,premin=99999.f; 178 179 p=this; 180 p--; 181 if(p<0)p+=VE_AMP; 182 postmax=max(acc,filters[j].ampbuf[p]); 183 postmin=min(acc,filters[j].ampbuf[p]); 184 185 for(i=0;i<stretch;i++){ 186 p--; 187 if(p<0)p+=VE_AMP; 188 premax=max(premax,filters[j].ampbuf[p]); 189 premin=min(premin,filters[j].ampbuf[p]); 190 } 191 192 valmin=postmin-premin; 193 valmax=postmax-premax; 194 195 /*filters[j].markers[pos]=valmax;*/ 196 filters[j].ampbuf[this]=acc; 197 filters[j].ampptr++; 198 if(filters[j].ampptr>=VE_AMP)filters[j].ampptr=0; 199 } 200 201 /* look at min/max, decide trigger */ 202 if(valmax>gi->preecho_thresh[j]+penalty){ 203 ret|=1; 204 ret|=4; 205 } 206 if(valmin<gi->postecho_thresh[j]-penalty)ret|=2; 207 } 208 209 return(ret); 210 } 211 212 #if 0 213 static int seq=0; 214 static ogg_int64_t totalshift=-1024; 215 #endif 216 217 long _ve_envelope_search(vorbis_dsp_state *v){ 218 vorbis_info *vi=v->vi; 219 codec_setup_info *ci=vi->codec_setup; 220 vorbis_info_psy_global *gi=&ci->psy_g_param; 221 envelope_lookup *ve=((private_state *)(v->backend_state))->ve; 222 long i,j; 223 224 int first=ve->current/ve->searchstep; 225 int last=v->pcm_current/ve->searchstep-VE_WIN; 226 if(first<0)first=0; 227 228 /* make sure we have enough storage to match the PCM */ 229 if(last+VE_WIN+VE_POST>ve->storage){ 230 ve->storage=last+VE_WIN+VE_POST; /* be sure */ 231 ve->mark=_ogg_realloc(ve->mark,ve->storage*sizeof(*ve->mark)); 232 } 233 234 for(j=first;j<last;j++){ 235 int ret=0; 236 237 ve->stretch++; 238 if(ve->stretch>VE_MAXSTRETCH*2) 239 ve->stretch=VE_MAXSTRETCH*2; 240 241 for(i=0;i<ve->ch;i++){ 242 float *pcm=v->pcm[i]+ve->searchstep*(j); 243 ret|=_ve_amp(ve,gi,pcm,ve->band,ve->filter+i*VE_BANDS,j); 244 } 245 246 ve->mark[j+VE_POST]=0; 247 if(ret&1){ 248 ve->mark[j]=1; 249 ve->mark[j+1]=1; 250 } 251 252 if(ret&2){ 253 ve->mark[j]=1; 254 if(j>0)ve->mark[j-1]=1; 255 } 256 257 if(ret&4)ve->stretch=-1; 258 } 259 260 ve->current=last*ve->searchstep; 261 262 { 263 long centerW=v->centerW; 264 long testW= 265 centerW+ 266 ci->blocksizes[v->W]/4+ 267 ci->blocksizes[1]/2+ 268 ci->blocksizes[0]/4; 269 270 j=ve->cursor; 271 272 while(j<ve->current-(ve->searchstep)){/* account for postecho 273 working back one window */ 274 if(j>=testW)return(1); 275 276 ve->cursor=j; 277 278 if(ve->mark[j/ve->searchstep]){ 279 if(j>centerW){ 280 281 #if 0 282 if(j>ve->curmark){ 283 float *marker=alloca(v->pcm_current*sizeof(*marker)); 284 int l,m; 285 memset(marker,0,sizeof(*marker)*v->pcm_current); 286 fprintf(stderr,"mark! seq=%d, cursor:%fs time:%fs\n", 287 seq, 288 (totalshift+ve->cursor)/44100., 289 (totalshift+j)/44100.); 290 _analysis_output_always("pcmL",seq,v->pcm[0],v->pcm_current,0,0,totalshift); 291 _analysis_output_always("pcmR",seq,v->pcm[1],v->pcm_current,0,0,totalshift); 292 293 _analysis_output_always("markL",seq,v->pcm[0],j,0,0,totalshift); 294 _analysis_output_always("markR",seq,v->pcm[1],j,0,0,totalshift); 295 296 for(m=0;m<VE_BANDS;m++){ 297 char buf[80]; 298 sprintf(buf,"delL%d",m); 299 for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m].markers[l]*.1; 300 _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift); 301 } 302 303 for(m=0;m<VE_BANDS;m++){ 304 char buf[80]; 305 sprintf(buf,"delR%d",m); 306 for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m+VE_BANDS].markers[l]*.1; 307 _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift); 308 } 309 310 for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->mark[l]*.4; 311 _analysis_output_always("mark",seq,marker,v->pcm_current,0,0,totalshift); 312 313 314 seq++; 315 316 } 317 #endif 318 319 ve->curmark=j; 320 if(j>=testW)return(1); 321 return(0); 322 } 323 } 324 j+=ve->searchstep; 325 } 326 } 327 328 return(-1); 329 } 330 331 int _ve_envelope_mark(vorbis_dsp_state *v){ 332 envelope_lookup *ve=((private_state *)(v->backend_state))->ve; 333 vorbis_info *vi=v->vi; 334 codec_setup_info *ci=vi->codec_setup; 335 long centerW=v->centerW; 336 long beginW=centerW-ci->blocksizes[v->W]/4; 337 long endW=centerW+ci->blocksizes[v->W]/4; 338 if(v->W){ 339 beginW-=ci->blocksizes[v->lW]/4; 340 endW+=ci->blocksizes[v->nW]/4; 341 }else{ 342 beginW-=ci->blocksizes[0]/4; 343 endW+=ci->blocksizes[0]/4; 344 } 345 346 if(ve->curmark>=beginW && ve->curmark<endW)return(1); 347 { 348 long first=beginW/ve->searchstep; 349 long last=endW/ve->searchstep; 350 long i; 351 for(i=first;i<last;i++) 352 if(ve->mark[i])return(1); 353 } 354 return(0); 355 } 356 357 void _ve_envelope_shift(envelope_lookup *e,long shift){ 358 int smallsize=e->current/e->searchstep+VE_POST; /* adjust for placing marks 359 ahead of ve->current */ 360 int smallshift=shift/e->searchstep; 361 362 memmove(e->mark,e->mark+smallshift,(smallsize-smallshift)*sizeof(*e->mark)); 363 364 #if 0 365 for(i=0;i<VE_BANDS*e->ch;i++) 366 memmove(e->filter[i].markers, 367 e->filter[i].markers+smallshift, 368 (1024-smallshift)*sizeof(*(*e->filter).markers)); 369 totalshift+=shift; 370 #endif 371 372 e->current-=shift; 373 if(e->curmark>=0) 374 e->curmark-=shift; 375 e->cursor-=shift; 376 } 377 378 379 380 381 382