/********************************************************************
* *
- * 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 1994-1999 *
- * by 1999 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
- author: Monty <xiphmont@mit.edu>
- modifications by: Monty
- last modification date: Aug 07 1999
-
- Vorbis manipulates the dynamic range of the incoming PCM data
- envelope to minimise time-domain energy leakage from percussive and
- plosive waveforms being quantized in the MDCT domain.
+ function: PCM data envelope analysis
********************************************************************/
#include <string.h>
#include <stdio.h>
#include <math.h>
+#include <ogg/ogg.h>
+#include "vorbis/codec.h"
+#include "codec_internal.h"
-#include "codec.h"
+#include "os.h"
+#include "scales.h"
#include "envelope.h"
+#include "mdct.h"
+#include "misc.h"
+
+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 i,j;
+ int n=e->winlength=128;
+ e->searchstep=64; /* not random */
+
+ e->minenergy=gi->preecho_minenergy;
+ e->ch=ch;
+ e->storage=128;
+ 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];
+ }
-void _ve_envelope_init(envelope_lookup *e,int samples_per){
- int i;
+ /* 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->winlen=samples_per*2;
- e->window=malloc(e->winlen*sizeof(double));
+ e->filter=_ogg_calloc(VE_BANDS*ch,sizeof(*e->filter));
+ e->mark=_ogg_calloc(e->storage,sizeof(*e->mark));
- /* We just use a straight sin^2(x) window for this */
- for(i=0;i<e->winlen;i++){
- double temp=sin((i+.5)/e->winlen*M_PI);
- e->window[i]=temp*temp;
- }
}
void _ve_envelope_clear(envelope_lookup *e){
- if(e->window)free(e->window);
- memset(e,0,sizeof(envelope_lookup));
+ int i;
+ 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));
}
-/* initial and final blocks are special cases. Eg:
- ______
- `--_
- |_______|_`-.___|_______|_______|
-
- ___
- _--' `--_
- |___.-'_|_______|_`-.___|_______|
-
- ___
- _--' `--_
- |_______|___.-'_|_______|_`-.___|
-
- _____
- _--'
- |_______|_______|____.-'|_______|
-
- as we go block by block, we watch the collective metrics span. If we
- span the threshhold (assuming the threshhold is active), we use an
- abbreviated vector */
-
-static int _ve_envelope_generate(double *mult,double *env,double *look,
- int n,int step){
- int i,j,p;
- double m,mo;
-
- for(j=0;j<n;j++)if(mult[j]!=1)break;
- if(j==n)return(0);
-
- n*=step;
- /* first multiplier special case */
- m=ldexp(1,mult[0]-1);
- for(p=0;p<step/2;p++)env[p]=m;
-
- /* mid multipliers normal case */
- for(j=1;p<n-step/2;j++){
- mo=m;
- m=ldexp(1,mult[j]-1);
- if(mo==m)
- for(i=0;i<step;i++,p++)env[p]=m;
- else
- for(i=0;i<step;i++,p++)env[p]=m*look[i]+mo*look[i+step];
+/* 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;
+
+ /* 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) */
+
+ float minV=ve->minenergy;
+ float *vec=alloca(n*sizeof(*vec));
+
+ /* 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_always("lpcm",seq2,data,n,0,0,
+ totalshift+pos*ve->searchstep);*/
+
+ /* window and transform */
+ for(i=0;i<n;i++)
+ vec[i]=data[i]*ve->mdct_win[i];
+ mdct_forward(&ve->mdct,vec,vec);
+
+ /*_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 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;
+
+ decay*=(1./(VE_NEARDC+1));
+ filters->nearptr++;
+ if(filters->nearptr>=VE_NEARDC)filters->nearptr=0;
+ decay=todB(&decay)*.5-15.f;
}
- /* last multiplier special case */
- for(;p<n;p++)env[p]=m;
- return(1);
-}
+ /* 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.;
+ }
-/* right now, we do things simple and dirty (read: our current preecho
- is a joke). Should this prove inadequate, then we'll think of
- something different. The details of the encoding format do not
- depend on the exact behavior, only the format of the bits that come
- out.
-
- Mark Taylor probably has much witter ways of doing this... Let's
- see if simple delta analysis gives us acceptible results for now. */
-
-static void _ve_deltas(double *deltas,double *pcm,int n,double *win,
- int winsize){
- int i,j,p=winsize/2;
- for(j=0;j<n;j++){
- p-=winsize/2;
- for(i=0;i<winsize-1;i++,p++){
- double temp=fabs(win[i]*pcm[p]-win[i+1]*pcm[p+1]);
- if(deltas[j]<temp)deltas[j]=temp;
+ /*_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;
}
- p++;
+
+ /* 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(ret);
}
-void _ve_envelope_multipliers(vorbis_dsp_state *v){
- int step=v->samples_per_envelope_step;
- static int frame=0;
-
- /* we need a 1-1/4 envelope window overlap begin and 1/4 end */
- int dtotal=(v->pcm_current-step/2)/v->samples_per_envelope_step;
- int dcurr=v->envelope_current;
- double *window=v->ve.window;
- int winlen=v->ve.winlen;
- int pch,ech;
- vorbis_info *vi=&v->vi;
-
- if(dtotal>dcurr){
- for(ech=0;ech<vi->envelopech;ech++){
- double *mult=v->multipliers[ech]+dcurr;
- memset(mult,0,sizeof(double)*(dtotal-dcurr));
-
- for(pch=0;pch<vi->channels;pch++){
-
- /* does this channel contribute to the envelope analysis */
- if(vi->envelopemap[pch]==ech){
-
- /* we need a 1/4 envelope window overlap front and back */
- double *pcm=v->pcm[pch]+dcurr*step-step/2;
- _ve_deltas(mult,pcm,dtotal-dcurr,window,winlen);
-
- }
- }
+#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;
+ 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(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));
+ }
+
+ for(j=first;j<last;j++){
+ int ret=0;
+
+ ve->stretch++;
+ if(ve->stretch>VE_MAXSTRETCH*2)
+ ve->stretch=VE_MAXSTRETCH*2;
+
+ for(i=0;i<ve->ch;i++){
+ 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;
}
- v->envelope_current=dtotal;
- frame++;
+
+ if(ret&4)ve->stretch=-1;
}
-}
-/* This readies the multiplier vector for use/coding. Clamp/adjust
- the multipliers to the allowed range and eliminate unneeded
- coefficients */
-
-void _ve_envelope_sparsify(vorbis_block *vb){
- int ch;
- for(ch=0;ch<vb->vd->vi.envelopech;ch++){
- int flag=0;
- double *mult=vb->mult[ch];
- int n=vb->multend;
- double first=mult[0];
- double last=first;
- double clamp;
- int i;
-
- /* are we going to multiply anything? */
-
- for(i=1;i<n;i++){
- if(mult[i]>=last*vb->vd->vi.preecho_thresh){
- flag=1;
- break;
+ 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);
+ }
}
- if(i<n-1 && mult[i+1]>=last*vb->vd->vi.preecho_thresh){
- flag=1;
- break;
- }
- last=mult[i];
+ j+=ve->searchstep;
}
-
- if(flag){
- /* we need to adjust, so we might as well go nuts */
-
- int begin=i;
- clamp=last?last:1;
-
- for(i=0;i<begin;i++)mult[i]=0;
-
- last=1;
- for(;i<n;i++){
- if(mult[i]/last>clamp*vb->vd->vi.preecho_thresh){
- last=mult[i]/vb->vd->vi.preecho_clamp;
-
- mult[i]=floor(log(mult[i]/clamp/vb->vd->vi.preecho_clamp)/log(2))-1;
- if(mult[i]>15)mult[i]=15;
- }else{
- mult[i]=0;
- }
- }
- }else
- memset(mult,0,sizeof(double)*n);
}
+
+ return(-1);
}
-void _ve_envelope_apply(vorbis_block *vb,int multp){
- vorbis_info *vi=&vb->vd->vi;
- double env[vb->multend*vi->envelopesa];
- envelope_lookup *look=&vb->vd->ve;
- int i,j,k;
-
- for(i=0;i<vi->envelopech;i++){
- double *mult=vb->mult[i];
- double last=1.;
-
- /* fill in the multiplier placeholders */
-
- for(j=0;j<vb->multend;j++){
- if(mult[j]){
- last=mult[j];
- }else
- mult[j]=last;
- }
+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;
+ }
- /* compute the envelope curve */
- if(_ve_envelope_generate(mult,env,look->window,vb->multend,
- vi->envelopesa)){
-
- /* apply the envelope curve */
- for(j=0;j<vi->channels;j++){
-
- /* check to see if the generated envelope applies to this channel */
- if(vi->envelopemap[j]==i){
-
- if(multp)
- for(k=0;k<vb->multend*vi->envelopesa;k++)
- vb->pcm[j][k]*=env[k];
- else
- for(k=0;k<vb->multend*vi->envelopesa;k++)
- vb->pcm[j][k]/=env[k];
- }
- }
- }
+ 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;
+}