/********************************************************************
* *
- * THIS FILE IS PART OF THE Ogg Vorbis SOFTWARE CODEC SOURCE CODE. *
- * USE, DISTRIBUTION AND REPRODUCTION OF THIS SOURCE IS GOVERNED BY *
- * THE GNU PUBLIC LICENSE 2, WHICH IS INCLUDED WITH THIS SOURCE. *
- * PLEASE READ THESE TERMS DISTRIBUTING. *
+ * THIS FILE IS PART OF THE OggVorbis SOFTWARE CODEC SOURCE CODE. *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. *
* *
- * THE OggSQUISH SOURCE CODE IS (C) COPYRIGHT 1994-2000 *
- * by Monty <monty@xiph.org> and The XIPHOPHORUS Company *
- * http://www.xiph.org/ *
+ * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2009 *
+ * by the Xiph.Org Foundation http://www.xiph.org/ *
* *
********************************************************************
- function: PCM data envelope analysis and manipulation
- last mod: $Id: envelope.c,v 1.23 2000/10/12 03:12:52 xiphmont Exp $
-
- Preecho calculation.
+ function: PCM data envelope analysis
+ last mod: $Id$
********************************************************************/
#include <math.h>
#include <ogg/ogg.h>
#include "vorbis/codec.h"
+#include "codec_internal.h"
#include "os.h"
#include "scales.h"
#include "envelope.h"
+#include "mdct.h"
#include "misc.h"
-/* We use a Chebyshev bandbass for the preecho trigger bandpass; it's
- close enough for sample rates 32000-48000 Hz (corner frequencies at
- 6k/14k assuming sample rate of 44.1kHz) */
-
-/* Digital filter designed by mkfilter/mkshape/gencode A.J. Fisher
- Command line: /www/usr/fisher/helpers/mkfilter -Ch \
- -6.0000000000e+00 -Bp -o 5 -a 1.3605442177e-01 3.1746031746e-01 -l */
-
-#if 0
-static int cheb_bandpass_stages=10;
-static float cheb_bandpass_gain=5.589612458e+01;
-static float cheb_bandpass_B[]={-1.,0.,5.,0.,-10.,0.,10.,0.,-5.,0.,1};
-static float cheb_bandpass_A[]={
- -0.1917409386,
- 0.0078657069,
- -0.7126903444,
- 0.0266343467,
- -1.4047174730,
- 0.0466964232,
- -1.9032773429,
- 0.0451493360,
- -1.4471447397,
- 0.0303413711};
-#endif
-
-static int cheb_highpass_stages=10;
-static float cheb_highpass_gain= 5.291963434e+01;
-/* z^-stage, z^-stage+1... */
-static float cheb_highpass_B[]={1,-10,45,-120,210,-252,210,-120,45,-10,1};
-static float cheb_highpass_A[]={
- -0.1247628029,
- 0.1334086523,
- -0.3997715614,
- 0.3213011089,
- -1.1131924119,
- 1.7692446626,
- -3.6241199038,
- 4.1950871291,
- -4.2771757867,
- 2.3920318913};
-
void _ve_envelope_init(envelope_lookup *e,vorbis_info *vi){
+ codec_setup_info *ci=vi->codec_setup;
+ vorbis_info_psy_global *gi=&ci->psy_g_param;
int ch=vi->channels;
- int window=vi->envelopesa;
- int i;
- e->winlength=window;
- e->minenergy=fromdB(vi->preecho_minenergy);
- e->iir=calloc(ch,sizeof(IIR_state));
- e->filtered=calloc(ch,sizeof(float *));
+ int i,j;
+ int n=e->winlength=128;
+ e->searchstep=64; /* not random */
+
+ e->minenergy=gi->preecho_minenergy;
e->ch=ch;
e->storage=128;
- for(i=0;i<ch;i++){
- IIR_init(e->iir+i,cheb_highpass_stages,cheb_highpass_gain,
- cheb_highpass_A,cheb_highpass_B);
- e->filtered[i]=calloc(e->storage,sizeof(float));
+ e->cursor=ci->blocksizes[1]/2;
+ e->mdct_win=_ogg_calloc(n,sizeof(*e->mdct_win));
+ mdct_init(&e->mdct,n);
+
+ for(i=0;i<n;i++){
+ e->mdct_win[i]=sin(i/(n-1.)*M_PI);
+ e->mdct_win[i]*=e->mdct_win[i];
}
- drft_init(&e->drft,window);
- e->window=malloc(e->winlength*sizeof(float));
- /* We just use a straight sin(x) window for this */
- for(i=0;i<e->winlength;i++)
- e->window[i]=sin((i+.5)/e->winlength*M_PI);
+ /* magic follows */
+ e->band[0].begin=2; e->band[0].end=4;
+ e->band[1].begin=4; e->band[1].end=5;
+ e->band[2].begin=6; e->band[2].end=6;
+ e->band[3].begin=9; e->band[3].end=8;
+ e->band[4].begin=13; e->band[4].end=8;
+ e->band[5].begin=17; e->band[5].end=8;
+ e->band[6].begin=22; e->band[6].end=8;
+
+ for(j=0;j<VE_BANDS;j++){
+ n=e->band[j].end;
+ e->band[j].window=_ogg_malloc(n*sizeof(*e->band[0].window));
+ for(i=0;i<n;i++){
+ e->band[j].window[i]=sin((i+.5)/n*M_PI);
+ e->band[j].total+=e->band[j].window[i];
+ }
+ e->band[j].total=1./e->band[j].total;
+ }
+
+ e->filter=_ogg_calloc(VE_BANDS*ch,sizeof(*e->filter));
+ e->mark=_ogg_calloc(e->storage,sizeof(*e->mark));
+
}
void _ve_envelope_clear(envelope_lookup *e){
int i;
- for(i=0;i<e->ch;i++){
- IIR_clear((e->iir+i));
- free(e->filtered[i]);
- }
- drft_clear(&e->drft);
- free(e->window);
- free(e->filtered);
- memset(e,0,sizeof(envelope_lookup));
+ mdct_clear(&e->mdct);
+ for(i=0;i<VE_BANDS;i++)
+ _ogg_free(e->band[i].window);
+ _ogg_free(e->mdct_win);
+ _ogg_free(e->filter);
+ _ogg_free(e->mark);
+ memset(e,0,sizeof(*e));
}
-static float _ve_deltai(envelope_lookup *ve,IIR_state *iir,
- float *pre,float *post){
- long n2=ve->winlength*2;
+/* fairly straight threshhold-by-band based until we find something
+ that works better and isn't patented. */
+
+static int _ve_amp(envelope_lookup *ve,
+ vorbis_info_psy_global *gi,
+ float *data,
+ envelope_band *bands,
+ envelope_filter_state *filters){
long n=ve->winlength;
+ int ret=0;
+ long i,j;
+ float decay;
- float *workA=alloca(sizeof(float)*n2),A=0.;
- float *workB=alloca(sizeof(float)*n2),B=0.;
- long i;
+ /* we want to have a 'minimum bar' for energy, else we're just
+ basing blocks on quantization noise that outweighs the signal
+ itself (for low power signals) */
- /*_analysis_output("A",granulepos,pre,n,0,0);
- _analysis_output("B",granulepos,post,n,0,0);*/
+ float minV=ve->minenergy;
+ float *vec=alloca(n*sizeof(*vec));
- for(i=0;i<n;i++){
- workA[i]=pre[i]*ve->window[i];
- workB[i]=post[i]*ve->window[i];
- }
+ /* stretch is used to gradually lengthen the number of windows
+ considered prevoius-to-potential-trigger */
+ int stretch=max(VE_MINSTRETCH,ve->stretch/2);
+ float penalty=gi->stretch_penalty-(ve->stretch/2-VE_MINSTRETCH);
+ if(penalty<0.f)penalty=0.f;
+ if(penalty>gi->stretch_penalty)penalty=gi->stretch_penalty;
- /*_analysis_output("Awin",granulepos,workA,n,0,0);
- _analysis_output("Bwin",granulepos,workB,n,0,0);*/
+ /*_analysis_output_always("lpcm",seq2,data,n,0,0,
+ totalshift+pos*ve->searchstep);*/
- drft_forward(&ve->drft,workA);
- drft_forward(&ve->drft,workB);
+ /* window and transform */
+ for(i=0;i<n;i++)
+ vec[i]=data[i]*ve->mdct_win[i];
+ mdct_forward(&ve->mdct,vec,vec);
- /* we want to have a 'minimum bar' for energy, else we're just
- basing blocks on quantization noise that outweighs the signal
- itself (for low power signals) */
+ /*_analysis_output_always("mdct",seq2,vec,n/2,0,1,0); */
+
+ /* near-DC spreading function; this has nothing to do with
+ psychoacoustics, just sidelobe leakage and window size */
{
- float min=ve->minenergy;
- for(i=0;i<n;i++){
- if(fabs(workA[i])<min)workA[i]=min;
- if(fabs(workB[i])<min)workB[i]=min;
+ float temp=vec[0]*vec[0]+.7*vec[1]*vec[1]+.2*vec[2]*vec[2];
+ int ptr=filters->nearptr;
+
+ /* the accumulation is regularly refreshed from scratch to avoid
+ floating point creep */
+ if(ptr==0){
+ decay=filters->nearDC_acc=filters->nearDC_partialacc+temp;
+ filters->nearDC_partialacc=temp;
+ }else{
+ decay=filters->nearDC_acc+=temp;
+ filters->nearDC_partialacc+=temp;
}
- }
+ filters->nearDC_acc-=filters->nearDC[ptr];
+ filters->nearDC[ptr]=temp;
- /*_analysis_output("Afft",granulepos,workA,n,0,0);
- _analysis_output("Bfft",granulepos,workB,n,0,0);*/
+ decay*=(1./(VE_NEARDC+1));
+ filters->nearptr++;
+ if(filters->nearptr>=VE_NEARDC)filters->nearptr=0;
+ decay=todB(&decay)*.5-15.f;
+ }
- for(i=0;i<n;i++){
- A+=workA[i]*workA[i];
- B+=workB[i]*workB[i];
+ /* perform spreading and limiting, also smooth the spectrum. yes,
+ the MDCT results in all real coefficients, but it still *behaves*
+ like real/imaginary pairs */
+ for(i=0;i<n/2;i+=2){
+ float val=vec[i]*vec[i]+vec[i+1]*vec[i+1];
+ val=todB(&val)*.5f;
+ if(val<decay)val=decay;
+ if(val<minV)val=minV;
+ vec[i>>1]=val;
+ decay-=8.;
}
- A=todB(A);
- B=todB(B);
+ /*_analysis_output_always("spread",seq2++,vec,n/4,0,0,0);*/
+
+ /* perform preecho/postecho triggering by band */
+ for(j=0;j<VE_BANDS;j++){
+ float acc=0.;
+ float valmax,valmin;
+
+ /* accumulate amplitude */
+ for(i=0;i<bands[j].end;i++)
+ acc+=vec[i+bands[j].begin]*bands[j].window[i];
+
+ acc*=bands[j].total;
+
+ /* convert amplitude to delta */
+ {
+ int p,this=filters[j].ampptr;
+ float postmax,postmin,premax=-99999.f,premin=99999.f;
+
+ p=this;
+ p--;
+ if(p<0)p+=VE_AMP;
+ postmax=max(acc,filters[j].ampbuf[p]);
+ postmin=min(acc,filters[j].ampbuf[p]);
+
+ for(i=0;i<stretch;i++){
+ p--;
+ if(p<0)p+=VE_AMP;
+ premax=max(premax,filters[j].ampbuf[p]);
+ premin=min(premin,filters[j].ampbuf[p]);
+ }
+
+ valmin=postmin-premin;
+ valmax=postmax-premax;
+
+ /*filters[j].markers[pos]=valmax;*/
+ filters[j].ampbuf[this]=acc;
+ filters[j].ampptr++;
+ if(filters[j].ampptr>=VE_AMP)filters[j].ampptr=0;
+ }
+
+ /* look at min/max, decide trigger */
+ if(valmax>gi->preecho_thresh[j]+penalty){
+ ret|=1;
+ ret|=4;
+ }
+ if(valmin<gi->postecho_thresh[j]-penalty)ret|=2;
+ }
- return(B-A);
+ return(ret);
}
-long _ve_envelope_search(vorbis_dsp_state *v,long searchpoint){
+#if 0
+static int seq=0;
+static ogg_int64_t totalshift=-1024;
+#endif
+
+long _ve_envelope_search(vorbis_dsp_state *v){
vorbis_info *vi=v->vi;
- envelope_lookup *ve=v->ve;
+ codec_setup_info *ci=vi->codec_setup;
+ vorbis_info_psy_global *gi=&ci->psy_g_param;
+ envelope_lookup *ve=((private_state *)(v->backend_state))->ve;
long i,j;
-
+
+ int first=ve->current/ve->searchstep;
+ int last=v->pcm_current/ve->searchstep-VE_WIN;
+ if(first<0)first=0;
+
/* make sure we have enough storage to match the PCM */
- if(v->pcm_storage>ve->storage){
- ve->storage=v->pcm_storage;
- for(i=0;i<ve->ch;i++)
- ve->filtered[i]=realloc(ve->filtered[i],ve->storage*sizeof(float));
+ if(last+VE_WIN+VE_POST>ve->storage){
+ ve->storage=last+VE_WIN+VE_POST; /* be sure */
+ ve->mark=_ogg_realloc(ve->mark,ve->storage*sizeof(*ve->mark));
}
- /* catch up the highpass to match the pcm */
- for(i=0;i<ve->ch;i++){
- float *filtered=ve->filtered[i];
- float *pcm=v->pcm[i];
- IIR_state *iir=ve->iir+i;
-
- for(j=ve->current;j<v->pcm_current;j++)
- filtered[j]=IIR_filter(iir,pcm[j]);
- }
- ve->current=v->pcm_current;
+ for(j=first;j<last;j++){
+ int ret=0;
- /* Now search through our cached highpass data for breaking points */
- /* starting point */
- if(v->W)
- j=v->centerW+vi->blocksizes[1]/4-vi->blocksizes[0]/4;
- else
- j=v->centerW;
+ ve->stretch++;
+ if(ve->stretch>VE_MAXSTRETCH*2)
+ ve->stretch=VE_MAXSTRETCH*2;
- while(j+ve->winlength<=v->pcm_current){
for(i=0;i<ve->ch;i++){
- float *filtered=ve->filtered[i]+j;
- IIR_state *iir=ve->iir+i;
- float m=_ve_deltai(ve,iir,filtered-ve->winlength,filtered);
-
- if(m>vi->preecho_thresh){
- /*granulepos++;*/
- return(0);
+ float *pcm=v->pcm[i]+ve->searchstep*(j);
+ ret|=_ve_amp(ve,gi,pcm,ve->band,ve->filter+i*VE_BANDS);
+ }
+
+ ve->mark[j+VE_POST]=0;
+ if(ret&1){
+ ve->mark[j]=1;
+ ve->mark[j+1]=1;
+ }
+
+ if(ret&2){
+ ve->mark[j]=1;
+ if(j>0)ve->mark[j-1]=1;
+ }
+
+ if(ret&4)ve->stretch=-1;
+ }
+
+ ve->current=last*ve->searchstep;
+
+ {
+ long centerW=v->centerW;
+ long testW=
+ centerW+
+ ci->blocksizes[v->W]/4+
+ ci->blocksizes[1]/2+
+ ci->blocksizes[0]/4;
+
+ j=ve->cursor;
+
+ while(j<ve->current-(ve->searchstep)){/* account for postecho
+ working back one window */
+ if(j>=testW)return(1);
+
+ ve->cursor=j;
+
+ if(ve->mark[j/ve->searchstep]){
+ if(j>centerW){
+
+#if 0
+ if(j>ve->curmark){
+ float *marker=alloca(v->pcm_current*sizeof(*marker));
+ int l,m;
+ memset(marker,0,sizeof(*marker)*v->pcm_current);
+ fprintf(stderr,"mark! seq=%d, cursor:%fs time:%fs\n",
+ seq,
+ (totalshift+ve->cursor)/44100.,
+ (totalshift+j)/44100.);
+ _analysis_output_always("pcmL",seq,v->pcm[0],v->pcm_current,0,0,totalshift);
+ _analysis_output_always("pcmR",seq,v->pcm[1],v->pcm_current,0,0,totalshift);
+
+ _analysis_output_always("markL",seq,v->pcm[0],j,0,0,totalshift);
+ _analysis_output_always("markR",seq,v->pcm[1],j,0,0,totalshift);
+
+ for(m=0;m<VE_BANDS;m++){
+ char buf[80];
+ sprintf(buf,"delL%d",m);
+ for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m].markers[l]*.1;
+ _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift);
+ }
+
+ for(m=0;m<VE_BANDS;m++){
+ char buf[80];
+ sprintf(buf,"delR%d",m);
+ for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m+VE_BANDS].markers[l]*.1;
+ _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift);
+ }
+
+ for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->mark[l]*.4;
+ _analysis_output_always("mark",seq,marker,v->pcm_current,0,0,totalshift);
+
+
+ seq++;
+
+ }
+#endif
+
+ ve->curmark=j;
+ if(j>=testW)return(1);
+ return(0);
+ }
}
- /*granulepos++;*/
+ j+=ve->searchstep;
}
-
- j+=vi->blocksizes[0]/2;
- if(j>=searchpoint)return(1);
}
-
+
return(-1);
}
-void _ve_envelope_shift(envelope_lookup *e,long shift){
- int i;
- for(i=0;i<e->ch;i++)
- memmove(e->filtered[i],e->filtered[i]+shift,(e->current-shift)*
- sizeof(float));
- e->current-=shift;
+int _ve_envelope_mark(vorbis_dsp_state *v){
+ envelope_lookup *ve=((private_state *)(v->backend_state))->ve;
+ vorbis_info *vi=v->vi;
+ codec_setup_info *ci=vi->codec_setup;
+ long centerW=v->centerW;
+ long beginW=centerW-ci->blocksizes[v->W]/4;
+ long endW=centerW+ci->blocksizes[v->W]/4;
+ if(v->W){
+ beginW-=ci->blocksizes[v->lW]/4;
+ endW+=ci->blocksizes[v->nW]/4;
+ }else{
+ beginW-=ci->blocksizes[0]/4;
+ endW+=ci->blocksizes[0]/4;
+ }
+
+ if(ve->curmark>=beginW && ve->curmark<endW)return(1);
+ {
+ long first=beginW/ve->searchstep;
+ long last=endW/ve->searchstep;
+ long i;
+ for(i=first;i<last;i++)
+ if(ve->mark[i])return(1);
+ }
+ return(0);
}
+void _ve_envelope_shift(envelope_lookup *e,long shift){
+ int smallsize=e->current/e->searchstep+VE_POST; /* adjust for placing marks
+ ahead of ve->current */
+ int smallshift=shift/e->searchstep;
+ memmove(e->mark,e->mark+smallshift,(smallsize-smallshift)*sizeof(*e->mark));
+
+#if 0
+ for(i=0;i<VE_BANDS*e->ch;i++)
+ memmove(e->filter[i].markers,
+ e->filter[i].markers+smallshift,
+ (1024-smallshift)*sizeof(*(*e->filter).markers));
+ totalshift+=shift;
+#endif
+
+ e->current-=shift;
+ if(e->curmark>=0)
+ e->curmark-=shift;
+ e->cursor-=shift;
+}
projects / platform / upstream / libvorbis.git / blobdiff