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 http://www.xiph.org/                  *
10 *                                                                  *
11 ********************************************************************
12
13 function: PCM data envelope analysis
14 last mod: $Id: envelope.c 16227 2009-07-08 06:58:46Z xiphmont $
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
32void _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
76void _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
90static 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 n=ve->winlength;
96  int ret=0;
97  long i,j;
98  float decay;
99
100  /* we want to have a 'minimum bar' for energy, else we're just
101     basing blocks on quantization noise that outweighs the signal
102     itself (for low power signals) */
103
104  float minV=ve->minenergy;
105  float *vec=alloca(n*sizeof(*vec));
106
107  /* stretch is used to gradually lengthen the number of windows
108     considered prevoius-to-potential-trigger */
109  int stretch=max(VE_MINSTRETCH,ve->stretch/2);
110  float penalty=gi->stretch_penalty-(ve->stretch/2-VE_MINSTRETCH);
111  if(penalty<0.f)penalty=0.f;
112  if(penalty>gi->stretch_penalty)penalty=gi->stretch_penalty;
113
114  /*_analysis_output_always("lpcm",seq2,data,n,0,0,
115    totalshift+pos*ve->searchstep);*/
116
117 /* window and transform */
118  for(i=0;i<n;i++)
119    vec[i]=data[i]*ve->mdct_win[i];
120  mdct_forward(&ve->mdct,vec,vec);
121
122  /*_analysis_output_always("mdct",seq2,vec,n/2,0,1,0); */
123
124  /* near-DC spreading function; this has nothing to do with
125     psychoacoustics, just sidelobe leakage and window size */
126  {
127    float temp=vec[0]*vec[0]+.7*vec[1]*vec[1]+.2*vec[2]*vec[2];
128    int ptr=filters->nearptr;
129
130    /* the accumulation is regularly refreshed from scratch to avoid
131       floating point creep */
132    if(ptr==0){
133      decay=filters->nearDC_acc=filters->nearDC_partialacc+temp;
134      filters->nearDC_partialacc=temp;
135    }else{
136      decay=filters->nearDC_acc+=temp;
137      filters->nearDC_partialacc+=temp;
138    }
139    filters->nearDC_acc-=filters->nearDC[ptr];
140    filters->nearDC[ptr]=temp;
141
142    decay*=(1./(VE_NEARDC+1));
143    filters->nearptr++;
144    if(filters->nearptr>=VE_NEARDC)filters->nearptr=0;
145    decay=todB(&decay)*.5-15.f;
146  }
147
148  /* perform spreading and limiting, also smooth the spectrum.  yes,
149     the MDCT results in all real coefficients, but it still *behaves*
150     like real/imaginary pairs */
151  for(i=0;i<n/2;i+=2){
152    float val=vec[i]*vec[i]+vec[i+1]*vec[i+1];
153    val=todB(&val)*.5f;
154    if(val<decay)val=decay;
155    if(val<minV)val=minV;
156    vec[i>>1]=val;
157    decay-=8.;
158  }
159
160  /*_analysis_output_always("spread",seq2++,vec,n/4,0,0,0);*/
161
162  /* perform preecho/postecho triggering by band */
163  for(j=0;j<VE_BANDS;j++){
164    float acc=0.;
165    float valmax,valmin;
166
167    /* accumulate amplitude */
168    for(i=0;i<bands[j].end;i++)
169      acc+=vec[i+bands[j].begin]*bands[j].window[i];
170
171    acc*=bands[j].total;
172
173    /* convert amplitude to delta */
174    {
175      int p,this=filters[j].ampptr;
176      float postmax,postmin,premax=-99999.f,premin=99999.f;
177
178      p=this;
179      p--;
180      if(p<0)p+=VE_AMP;
181      postmax=max(acc,filters[j].ampbuf[p]);
182      postmin=min(acc,filters[j].ampbuf[p]);
183
184      for(i=0;i<stretch;i++){
185        p--;
186        if(p<0)p+=VE_AMP;
187        premax=max(premax,filters[j].ampbuf[p]);
188        premin=min(premin,filters[j].ampbuf[p]);
189      }
190
191      valmin=postmin-premin;
192      valmax=postmax-premax;
193
194      /*filters[j].markers[pos]=valmax;*/
195      filters[j].ampbuf[this]=acc;
196      filters[j].ampptr++;
197      if(filters[j].ampptr>=VE_AMP)filters[j].ampptr=0;
198    }
199
200    /* look at min/max, decide trigger */
201    if(valmax>gi->preecho_thresh[j]+penalty){
202      ret|=1;
203      ret|=4;
204    }
205    if(valmin<gi->postecho_thresh[j]-penalty)ret|=2;
206  }
207
208  return(ret);
209}
210
211#if 0
212static int seq=0;
213static ogg_int64_t totalshift=-1024;
214#endif
215
216long _ve_envelope_search(vorbis_dsp_state *v){
217  vorbis_info *vi=v->vi;
218  codec_setup_info *ci=vi->codec_setup;
219  vorbis_info_psy_global *gi=&ci->psy_g_param;
220  envelope_lookup *ve=((private_state *)(v->backend_state))->ve;
221  long i,j;
222
223  int first=ve->current/ve->searchstep;
224  int last=v->pcm_current/ve->searchstep-VE_WIN;
225  if(first<0)first=0;
226
227  /* make sure we have enough storage to match the PCM */
228  if(last+VE_WIN+VE_POST>ve->storage){
229    ve->storage=last+VE_WIN+VE_POST; /* be sure */
230    ve->mark=_ogg_realloc(ve->mark,ve->storage*sizeof(*ve->mark));
231  }
232
233  for(j=first;j<last;j++){
234    int ret=0;
235
236    ve->stretch++;
237    if(ve->stretch>VE_MAXSTRETCH*2)
238      ve->stretch=VE_MAXSTRETCH*2;
239
240    for(i=0;i<ve->ch;i++){
241      float *pcm=v->pcm[i]+ve->searchstep*(j);
242      ret|=_ve_amp(ve,gi,pcm,ve->band,ve->filter+i*VE_BANDS);
243    }
244
245    ve->mark[j+VE_POST]=0;
246    if(ret&1){
247      ve->mark[j]=1;
248      ve->mark[j+1]=1;
249    }
250
251    if(ret&2){
252      ve->mark[j]=1;
253      if(j>0)ve->mark[j-1]=1;
254    }
255
256    if(ret&4)ve->stretch=-1;
257  }
258
259  ve->current=last*ve->searchstep;
260
261  {
262    long centerW=v->centerW;
263    long testW=
264      centerW+
265      ci->blocksizes[v->W]/4+
266      ci->blocksizes[1]/2+
267      ci->blocksizes[0]/4;
268
269    j=ve->cursor;
270
271    while(j<ve->current-(ve->searchstep)){/* account for postecho
272                                             working back one window */
273      if(j>=testW)return(1);
274
275      ve->cursor=j;
276
277      if(ve->mark[j/ve->searchstep]){
278        if(j>centerW){
279
280#if 0
281          if(j>ve->curmark){
282            float *marker=alloca(v->pcm_current*sizeof(*marker));
283            int l,m;
284            memset(marker,0,sizeof(*marker)*v->pcm_current);
285            fprintf(stderr,"mark! seq=%d, cursor:%fs time:%fs\n",
286                    seq,
287                    (totalshift+ve->cursor)/44100.,
288                    (totalshift+j)/44100.);
289            _analysis_output_always("pcmL",seq,v->pcm[0],v->pcm_current,0,0,totalshift);
290            _analysis_output_always("pcmR",seq,v->pcm[1],v->pcm_current,0,0,totalshift);
291
292            _analysis_output_always("markL",seq,v->pcm[0],j,0,0,totalshift);
293            _analysis_output_always("markR",seq,v->pcm[1],j,0,0,totalshift);
294
295            for(m=0;m<VE_BANDS;m++){
296              char buf[80];
297              sprintf(buf,"delL%d",m);
298              for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m].markers[l]*.1;
299              _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift);
300            }
301
302            for(m=0;m<VE_BANDS;m++){
303              char buf[80];
304              sprintf(buf,"delR%d",m);
305              for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m+VE_BANDS].markers[l]*.1;
306              _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift);
307            }
308
309            for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->mark[l]*.4;
310            _analysis_output_always("mark",seq,marker,v->pcm_current,0,0,totalshift);
311
312
313            seq++;
314
315          }
316#endif
317
318          ve->curmark=j;
319          if(j>=testW)return(1);
320          return(0);
321        }
322      }
323      j+=ve->searchstep;
324    }
325  }
326
327  return(-1);
328}
329
330int _ve_envelope_mark(vorbis_dsp_state *v){
331  envelope_lookup *ve=((private_state *)(v->backend_state))->ve;
332  vorbis_info *vi=v->vi;
333  codec_setup_info *ci=vi->codec_setup;
334  long centerW=v->centerW;
335  long beginW=centerW-ci->blocksizes[v->W]/4;
336  long endW=centerW+ci->blocksizes[v->W]/4;
337  if(v->W){
338    beginW-=ci->blocksizes[v->lW]/4;
339    endW+=ci->blocksizes[v->nW]/4;
340  }else{
341    beginW-=ci->blocksizes[0]/4;
342    endW+=ci->blocksizes[0]/4;
343  }
344
345  if(ve->curmark>=beginW && ve->curmark<endW)return(1);
346  {
347    long first=beginW/ve->searchstep;
348    long last=endW/ve->searchstep;
349    long i;
350    for(i=first;i<last;i++)
351      if(ve->mark[i])return(1);
352  }
353  return(0);
354}
355
356void _ve_envelope_shift(envelope_lookup *e,long shift){
357  int smallsize=e->current/e->searchstep+VE_POST; /* adjust for placing marks
358                                                     ahead of ve->current */
359  int smallshift=shift/e->searchstep;
360
361  memmove(e->mark,e->mark+smallshift,(smallsize-smallshift)*sizeof(*e->mark));
362
363#if 0
364  for(i=0;i<VE_BANDS*e->ch;i++)
365    memmove(e->filter[i].markers,
366            e->filter[i].markers+smallshift,
367            (1024-smallshift)*sizeof(*(*e->filter).markers));
368  totalshift+=shift;
369#endif
370
371  e->current-=shift;
372  if(e->curmark>=0)
373    e->curmark-=shift;
374  e->cursor-=shift;
375}
376