/********************************************************************
* *
- * 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 https://xiph.org/ *
* *
********************************************************************
- function: PCM data envelope analysis and manipulation
- author: Monty <xiphmont@mit.edu>
- modifications by: Monty
- last modification date: Jun 17 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 <stdlib.h>
+#include <string.h>
#include <stdio.h>
#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"
+
+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];
+ }
-static typedef struct {
- int divisor;
- double *window;
-} envelope_lookup;
+ /* 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;
+ }
-static double oPI = 3.14159265358979323846;
+ e->filter=_ogg_calloc(VE_BANDS*ch,sizeof(*e->filter));
+ e->mark=_ogg_calloc(e->storage,sizeof(*e->mark));
-envelope_lookup *init_envelope(int length,int divleng){
- envelope_lookup *ret=malloc(sizeof(envelope_lookup));
+}
+
+void _ve_envelope_clear(envelope_lookup *e){
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));
+}
- ret->length=divleng;
- ret->window=malloc(divleng*sizeof(double)*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;
+
+ /* 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;
- /* We just use a straight sin^2(x) window for this */
- for(i=0;i<divleng*2;i++){
- double temp=sin((i+.5)/divleng*oPI);
- ret->window[i]=temp*temp;
+ decay*=(1./(VE_NEARDC+1));
+ filters->nearptr++;
+ if(filters->nearptr>=VE_NEARDC)filters->nearptr=0;
+ decay=todB(&decay)*.5-15.f;
}
-}
-/* right now, we do things simple and dirty. 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.
-
- Using residual from an LPC whitening filter to judge envelope
- energy would probably yield cleaner results, but that's slow.
- Let's see if simple delta analysis gives us acceptible results. */
-
-int analyze_envelope0(double *vector, envelope_lookup *init, int n,
- double *deltas){
-
- int divisor=init->length;
- int divs=n/divisor-1;
- double *win=init->window;
- int i,j,count=0;
-
- double max,spanlo,spanhi;
-
- /* 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 */
-
- /* initial frame */
- max=0;
- for(i=1;i<divisor;i++){
- double temp=abs(vector[i-1]-vector[i]);
- if(max<temp)max=temp;
- }
- for(;i<divisor*2;i++){
- double temp=abs(win[i-1]*vector[i-1]-win[i]*vector[i]);
- if(max<temp)max=temp;
- }
- spanlo=spanhi=deltas[count++]=max;
-
- /* mid frames */
- for(j=divisor;j<n-divisor*2;j+=divisor){
- max=0;
- for(i=1;i<divisor*2;i++){
- double temp=abs(win[i-1]*vector[j+i-1]-win[i]*vector[j+i]);
- if(max<temp)max=temp;
- }
- deltas[count++]=max;
- if(max<spanlo)spanlo=max;
- if(max>spanhi)spanhi=max;
- if(threshhold>1 && spanlo*threshhold<spanhi)
- abbrevflag=1;
- if(abbrevflag && j>n0-divisor/2)break;
+ /* 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.;
}
- /* last frame */
- if(!abbrevflag){
- max=0;
- for(i=1;i<divisor;i++){
- double temp=abs(win[i-1]*vector[j+i-1]-win[i]*vector[j+i]);
- if(max<temp)max=temp;
- }
- for(;i<divisor*2;i++){
- double temp=abs(vector[j+i-1]-vector[j+i]);
- if(max<temp)max=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;
}
- deltas[count++]=max;
- if(max<spanlo)spanlo=max;
- if(max>spanhi)spanhi=max;
- if(threshhold>1 && spanlo*threshhold<spanhi)
- abbrevflag=1;
- }
- if(abbrevflag)return(n0);
- return(n);
-}
-
-/* also decide if we're going with a full sized or abbreviated
- vector. Some encoding tactics might want to use envelope massaging
- fully and discard abbreviated vectors entriely. We make that
- decision here */
-
-int analyze_envelope1(envelope_lookup *init,int n,
- double triggerthresh,double spanthresh,
- double *deltas){
-
- /* Look at the delta values; decide if we need to do any envelope
- manipulation at all on this vector; if so, choose the
- multipliers and placeholders.
-
- '0' is a placeholder. Other values specify a
- multiplier/divisor. Multipliers are used by the decoder, divisors
- in the encoder. The mapped m/d value for each segment is
- 2^(n-1). Placeholders (zeros) take on the value of the last
- non-zero multiplier/divisor. When the placeholder is not
- preceeded by a non-placeholder value in the current vector, it
- assumes the value of the *next* non-zero value. In this way, the
- vector manipulation is local to the current vector and does not
- rely on preceeding vectors.
-
- */
-
- /* scan forward with sliding windows; we start manipulating envelopes
- when the collective deltas span over a threshhold. If in fact we
- begin manipulating, we can manage on a finer scale than the
- original threshhold. first look for the larger threshhold and if
- we span it, manipulate the vector to hold within the smaller span
- threshhold. */
-
- /* scan for the trigger */
-
- int divisor=init->length;
- int divs=n/divisor-1;
- int i,triggerflag=0;
- double spanlo,spanhi;
-
- spanlo=spanhi=deltas[0];
-
- for(i=1;i<divs;i++){
- double max=deltas[i];
- if(max<spanlo)spanlo=max;
- if(max>spanhi)spanhi=max;
- if(spanlo*triggerthresh<spanhi){
- triggerflag=1;
- break;
+ /* 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;
}
- if(triggerflag){
- /* choose divisors/multipliers to fit the vector into the
- specified span. In the decoder, these values are *multipliers*, so */
-
-
+ return(ret);
+}
+#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;
+ }
+ 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);
+ }
+ }
+ j+=ve->searchstep;
+ }
+ }
+ return(-1);
+}
+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(triggerflag);
+ 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;
+}