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 1994-2010 *
9 * by the Xiph.Org Foundation http://www.xiph.org/ *
10 * *
11 ********************************************************************
12
13 function: psychoacoustics not including preecho
14 last mod: $Id: psy.c 17077 2010-03-26 06:22:19Z xiphmont $
15
16 ********************************************************************/
17
18 #include <stdlib.h>
19 #include <math.h>
20 #include <string.h>
21 #include "vorbis/codec.h"
22 #include "codec_internal.h"
23
24 #include "masking.h"
25 #include "psy.h"
26 #include "os.h"
27 #include "lpc.h"
28 #include "smallft.h"
29 #include "scales.h"
30 #include "misc.h"
31
32 #define NEGINF -9999.f
33 static const double stereo_threshholds[]={0.0, .5, 1.0, 1.5, 2.5, 4.5, 8.5, 16.5, 9e10};
34 static const double stereo_threshholds_limited[]={0.0, .5, 1.0, 1.5, 2.0, 2.5, 4.5, 8.5, 9e10};
35
_vp_global_look(vorbis_info * vi)36 vorbis_look_psy_global *_vp_global_look(vorbis_info *vi){
37 codec_setup_info *ci=vi->codec_setup;
38 vorbis_info_psy_global *gi=&ci->psy_g_param;
39 vorbis_look_psy_global *look=_ogg_calloc(1,sizeof(*look));
40
41 look->channels=vi->channels;
42
43 look->ampmax=-9999.;
44 look->gi=gi;
45 return(look);
46 }
47
_vp_global_free(vorbis_look_psy_global * look)48 void _vp_global_free(vorbis_look_psy_global *look){
49 if(look){
50 memset(look,0,sizeof(*look));
51 _ogg_free(look);
52 }
53 }
54
_vi_gpsy_free(vorbis_info_psy_global * i)55 void _vi_gpsy_free(vorbis_info_psy_global *i){
56 if(i){
57 memset(i,0,sizeof(*i));
58 _ogg_free(i);
59 }
60 }
61
_vi_psy_free(vorbis_info_psy * i)62 void _vi_psy_free(vorbis_info_psy *i){
63 if(i){
64 memset(i,0,sizeof(*i));
65 _ogg_free(i);
66 }
67 }
68
min_curve(float * c,float * c2)69 static void min_curve(float *c,
70 float *c2){
71 int i;
72 for(i=0;i<EHMER_MAX;i++)if(c2[i]<c[i])c[i]=c2[i];
73 }
max_curve(float * c,float * c2)74 static void max_curve(float *c,
75 float *c2){
76 int i;
77 for(i=0;i<EHMER_MAX;i++)if(c2[i]>c[i])c[i]=c2[i];
78 }
79
attenuate_curve(float * c,float att)80 static void attenuate_curve(float *c,float att){
81 int i;
82 for(i=0;i<EHMER_MAX;i++)
83 c[i]+=att;
84 }
85
setup_tone_curves(float curveatt_dB[P_BANDS],float binHz,int n,float center_boost,float center_decay_rate)86 static float ***setup_tone_curves(float curveatt_dB[P_BANDS],float binHz,int n,
87 float center_boost, float center_decay_rate){
88 int i,j,k,m;
89 float ath[EHMER_MAX];
90 float workc[P_BANDS][P_LEVELS][EHMER_MAX];
91 float athc[P_LEVELS][EHMER_MAX];
92 float *brute_buffer=alloca(n*sizeof(*brute_buffer));
93
94 float ***ret=_ogg_malloc(sizeof(*ret)*P_BANDS);
95
96 memset(workc,0,sizeof(workc));
97
98 for(i=0;i<P_BANDS;i++){
99 /* we add back in the ATH to avoid low level curves falling off to
100 -infinity and unnecessarily cutting off high level curves in the
101 curve limiting (last step). */
102
103 /* A half-band's settings must be valid over the whole band, and
104 it's better to mask too little than too much */
105 int ath_offset=i*4;
106 for(j=0;j<EHMER_MAX;j++){
107 float min=999.;
108 for(k=0;k<4;k++)
109 if(j+k+ath_offset<MAX_ATH){
110 if(min>ATH[j+k+ath_offset])min=ATH[j+k+ath_offset];
111 }else{
112 if(min>ATH[MAX_ATH-1])min=ATH[MAX_ATH-1];
113 }
114 ath[j]=min;
115 }
116
117 /* copy curves into working space, replicate the 50dB curve to 30
118 and 40, replicate the 100dB curve to 110 */
119 for(j=0;j<6;j++)
120 memcpy(workc[i][j+2],tonemasks[i][j],EHMER_MAX*sizeof(*tonemasks[i][j]));
121 memcpy(workc[i][0],tonemasks[i][0],EHMER_MAX*sizeof(*tonemasks[i][0]));
122 memcpy(workc[i][1],tonemasks[i][0],EHMER_MAX*sizeof(*tonemasks[i][0]));
123
124 /* apply centered curve boost/decay */
125 for(j=0;j<P_LEVELS;j++){
126 for(k=0;k<EHMER_MAX;k++){
127 float adj=center_boost+abs(EHMER_OFFSET-k)*center_decay_rate;
128 if(adj<0. && center_boost>0)adj=0.;
129 if(adj>0. && center_boost<0)adj=0.;
130 workc[i][j][k]+=adj;
131 }
132 }
133
134 /* normalize curves so the driving amplitude is 0dB */
135 /* make temp curves with the ATH overlayed */
136 for(j=0;j<P_LEVELS;j++){
137 attenuate_curve(workc[i][j],curveatt_dB[i]+100.-(j<2?2:j)*10.-P_LEVEL_0);
138 memcpy(athc[j],ath,EHMER_MAX*sizeof(**athc));
139 attenuate_curve(athc[j],+100.-j*10.f-P_LEVEL_0);
140 max_curve(athc[j],workc[i][j]);
141 }
142
143 /* Now limit the louder curves.
144
145 the idea is this: We don't know what the playback attenuation
146 will be; 0dB SL moves every time the user twiddles the volume
147 knob. So that means we have to use a single 'most pessimal' curve
148 for all masking amplitudes, right? Wrong. The *loudest* sound
149 can be in (we assume) a range of ...+100dB] SL. However, sounds
150 20dB down will be in a range ...+80], 40dB down is from ...+60],
151 etc... */
152
153 for(j=1;j<P_LEVELS;j++){
154 min_curve(athc[j],athc[j-1]);
155 min_curve(workc[i][j],athc[j]);
156 }
157 }
158
159 for(i=0;i<P_BANDS;i++){
160 int hi_curve,lo_curve,bin;
161 ret[i]=_ogg_malloc(sizeof(**ret)*P_LEVELS);
162
163 /* low frequency curves are measured with greater resolution than
164 the MDCT/FFT will actually give us; we want the curve applied
165 to the tone data to be pessimistic and thus apply the minimum
166 masking possible for a given bin. That means that a single bin
167 could span more than one octave and that the curve will be a
168 composite of multiple octaves. It also may mean that a single
169 bin may span > an eighth of an octave and that the eighth
170 octave values may also be composited. */
171
172 /* which octave curves will we be compositing? */
173 bin=floor(fromOC(i*.5)/binHz);
174 lo_curve= ceil(toOC(bin*binHz+1)*2);
175 hi_curve= floor(toOC((bin+1)*binHz)*2);
176 if(lo_curve>i)lo_curve=i;
177 if(lo_curve<0)lo_curve=0;
178 if(hi_curve>=P_BANDS)hi_curve=P_BANDS-1;
179
180 for(m=0;m<P_LEVELS;m++){
181 ret[i][m]=_ogg_malloc(sizeof(***ret)*(EHMER_MAX+2));
182
183 for(j=0;j<n;j++)brute_buffer[j]=999.;
184
185 /* render the curve into bins, then pull values back into curve.
186 The point is that any inherent subsampling aliasing results in
187 a safe minimum */
188 for(k=lo_curve;k<=hi_curve;k++){
189 int l=0;
190
191 for(j=0;j<EHMER_MAX;j++){
192 int lo_bin= fromOC(j*.125+k*.5-2.0625)/binHz;
193 int hi_bin= fromOC(j*.125+k*.5-1.9375)/binHz+1;
194
195 if(lo_bin<0)lo_bin=0;
196 if(lo_bin>n)lo_bin=n;
197 if(lo_bin<l)l=lo_bin;
198 if(hi_bin<0)hi_bin=0;
199 if(hi_bin>n)hi_bin=n;
200
201 for(;l<hi_bin && l<n;l++)
202 if(brute_buffer[l]>workc[k][m][j])
203 brute_buffer[l]=workc[k][m][j];
204 }
205
206 for(;l<n;l++)
207 if(brute_buffer[l]>workc[k][m][EHMER_MAX-1])
208 brute_buffer[l]=workc[k][m][EHMER_MAX-1];
209
210 }
211
212 /* be equally paranoid about being valid up to next half ocatve */
213 if(i+1<P_BANDS){
214 int l=0;
215 k=i+1;
216 for(j=0;j<EHMER_MAX;j++){
217 int lo_bin= fromOC(j*.125+i*.5-2.0625)/binHz;
218 int hi_bin= fromOC(j*.125+i*.5-1.9375)/binHz+1;
219
220 if(lo_bin<0)lo_bin=0;
221 if(lo_bin>n)lo_bin=n;
222 if(lo_bin<l)l=lo_bin;
223 if(hi_bin<0)hi_bin=0;
224 if(hi_bin>n)hi_bin=n;
225
226 for(;l<hi_bin && l<n;l++)
227 if(brute_buffer[l]>workc[k][m][j])
228 brute_buffer[l]=workc[k][m][j];
229 }
230
231 for(;l<n;l++)
232 if(brute_buffer[l]>workc[k][m][EHMER_MAX-1])
233 brute_buffer[l]=workc[k][m][EHMER_MAX-1];
234
235 }
236
237
238 for(j=0;j<EHMER_MAX;j++){
239 int bin=fromOC(j*.125+i*.5-2.)/binHz;
240 if(bin<0){
241 ret[i][m][j+2]=-999.;
242 }else{
243 if(bin>=n){
244 ret[i][m][j+2]=-999.;
245 }else{
246 ret[i][m][j+2]=brute_buffer[bin];
247 }
248 }
249 }
250
251 /* add fenceposts */
252 for(j=0;j<EHMER_OFFSET;j++)
253 if(ret[i][m][j+2]>-200.f)break;
254 ret[i][m][0]=j;
255
256 for(j=EHMER_MAX-1;j>EHMER_OFFSET+1;j--)
257 if(ret[i][m][j+2]>-200.f)
258 break;
259 ret[i][m][1]=j;
260
261 }
262 }
263
264 return(ret);
265 }
266
_vp_psy_init(vorbis_look_psy * p,vorbis_info_psy * vi,vorbis_info_psy_global * gi,int n,long rate)267 void _vp_psy_init(vorbis_look_psy *p,vorbis_info_psy *vi,
268 vorbis_info_psy_global *gi,int n,long rate){
269 long i,j,lo=-99,hi=1;
270 long maxoc;
271 memset(p,0,sizeof(*p));
272
273 p->eighth_octave_lines=gi->eighth_octave_lines;
274 p->shiftoc=rint(log(gi->eighth_octave_lines*8.f)/log(2.f))-1;
275
276 p->firstoc=toOC(.25f*rate*.5/n)*(1<<(p->shiftoc+1))-gi->eighth_octave_lines;
277 maxoc=toOC((n+.25f)*rate*.5/n)*(1<<(p->shiftoc+1))+.5f;
278 p->total_octave_lines=maxoc-p->firstoc+1;
279 p->ath=_ogg_malloc(n*sizeof(*p->ath));
280
281 p->octave=_ogg_malloc(n*sizeof(*p->octave));
282 p->bark=_ogg_malloc(n*sizeof(*p->bark));
283 p->vi=vi;
284 p->n=n;
285 p->rate=rate;
286
287 /* AoTuV HF weighting */
288 p->m_val = 1.;
289 if(rate < 26000) p->m_val = 0;
290 else if(rate < 38000) p->m_val = .94; /* 32kHz */
291 else if(rate > 46000) p->m_val = 1.275; /* 48kHz */
292
293 /* set up the lookups for a given blocksize and sample rate */
294
295 for(i=0,j=0;i<MAX_ATH-1;i++){
296 int endpos=rint(fromOC((i+1)*.125-2.)*2*n/rate);
297 float base=ATH[i];
298 if(j<endpos){
299 float delta=(ATH[i+1]-base)/(endpos-j);
300 for(;j<endpos && j<n;j++){
301 p->ath[j]=base+100.;
302 base+=delta;
303 }
304 }
305 }
306
307 for(;j<n;j++){
308 p->ath[j]=p->ath[j-1];
309 }
310
311 for(i=0;i<n;i++){
312 float bark=toBARK(rate/(2*n)*i);
313
314 for(;lo+vi->noisewindowlomin<i &&
315 toBARK(rate/(2*n)*lo)<(bark-vi->noisewindowlo);lo++);
316
317 for(;hi<=n && (hi<i+vi->noisewindowhimin ||
318 toBARK(rate/(2*n)*hi)<(bark+vi->noisewindowhi));hi++);
319
320 p->bark[i]=((lo-1)<<16)+(hi-1);
321
322 }
323
324 for(i=0;i<n;i++)
325 p->octave[i]=toOC((i+.25f)*.5*rate/n)*(1<<(p->shiftoc+1))+.5f;
326
327 p->tonecurves=setup_tone_curves(vi->toneatt,rate*.5/n,n,
328 vi->tone_centerboost,vi->tone_decay);
329
330 /* set up rolling noise median */
331 p->noiseoffset=_ogg_malloc(P_NOISECURVES*sizeof(*p->noiseoffset));
332 for(i=0;i<P_NOISECURVES;i++)
333 p->noiseoffset[i]=_ogg_malloc(n*sizeof(**p->noiseoffset));
334
335 for(i=0;i<n;i++){
336 float halfoc=toOC((i+.5)*rate/(2.*n))*2.;
337 int inthalfoc;
338 float del;
339
340 if(halfoc<0)halfoc=0;
341 if(halfoc>=P_BANDS-1)halfoc=P_BANDS-1;
342 inthalfoc=(int)halfoc;
343 del=halfoc-inthalfoc;
344
345 for(j=0;j<P_NOISECURVES;j++)
346 p->noiseoffset[j][i]=
347 p->vi->noiseoff[j][inthalfoc]*(1.-del) +
348 p->vi->noiseoff[j][inthalfoc+1]*del;
349
350 }
351 #if 0
352 {
353 static int ls=0;
354 _analysis_output_always("noiseoff0",ls,p->noiseoffset[0],n,1,0,0);
355 _analysis_output_always("noiseoff1",ls,p->noiseoffset[1],n,1,0,0);
356 _analysis_output_always("noiseoff2",ls++,p->noiseoffset[2],n,1,0,0);
357 }
358 #endif
359 }
360
_vp_psy_clear(vorbis_look_psy * p)361 void _vp_psy_clear(vorbis_look_psy *p){
362 int i,j;
363 if(p){
364 if(p->ath)_ogg_free(p->ath);
365 if(p->octave)_ogg_free(p->octave);
366 if(p->bark)_ogg_free(p->bark);
367 if(p->tonecurves){
368 for(i=0;i<P_BANDS;i++){
369 for(j=0;j<P_LEVELS;j++){
370 _ogg_free(p->tonecurves[i][j]);
371 }
372 _ogg_free(p->tonecurves[i]);
373 }
374 _ogg_free(p->tonecurves);
375 }
376 if(p->noiseoffset){
377 for(i=0;i<P_NOISECURVES;i++){
378 _ogg_free(p->noiseoffset[i]);
379 }
380 _ogg_free(p->noiseoffset);
381 }
382 memset(p,0,sizeof(*p));
383 }
384 }
385
386 /* octave/(8*eighth_octave_lines) x scale and dB y scale */
seed_curve(float * seed,const float ** curves,float amp,int oc,int n,int linesper,float dBoffset)387 static void seed_curve(float *seed,
388 const float **curves,
389 float amp,
390 int oc, int n,
391 int linesper,float dBoffset){
392 int i,post1;
393 int seedptr;
394 const float *posts,*curve;
395
396 int choice=(int)((amp+dBoffset-P_LEVEL_0)*.1f);
397 choice=max(choice,0);
398 choice=min(choice,P_LEVELS-1);
399 posts=curves[choice];
400 curve=posts+2;
401 post1=(int)posts[1];
402 seedptr=oc+(posts[0]-EHMER_OFFSET)*linesper-(linesper>>1);
403
404 for(i=posts[0];i<post1;i++){
405 if(seedptr>0){
406 float lin=amp+curve[i];
407 if(seed[seedptr]<lin)seed[seedptr]=lin;
408 }
409 seedptr+=linesper;
410 if(seedptr>=n)break;
411 }
412 }
413
seed_loop(vorbis_look_psy * p,const float *** curves,const float * f,const float * flr,float * seed,float specmax)414 static void seed_loop(vorbis_look_psy *p,
415 const float ***curves,
416 const float *f,
417 const float *flr,
418 float *seed,
419 float specmax){
420 vorbis_info_psy *vi=p->vi;
421 long n=p->n,i;
422 float dBoffset=vi->max_curve_dB-specmax;
423
424 /* prime the working vector with peak values */
425
426 for(i=0;i<n;i++){
427 float max=f[i];
428 long oc=p->octave[i];
429 while(i+1<n && p->octave[i+1]==oc){
430 i++;
431 if(f[i]>max)max=f[i];
432 }
433
434 if(max+6.f>flr[i]){
435 oc=oc>>p->shiftoc;
436
437 if(oc>=P_BANDS)oc=P_BANDS-1;
438 if(oc<0)oc=0;
439
440 seed_curve(seed,
441 curves[oc],
442 max,
443 p->octave[i]-p->firstoc,
444 p->total_octave_lines,
445 p->eighth_octave_lines,
446 dBoffset);
447 }
448 }
449 }
450
seed_chase(float * seeds,int linesper,long n)451 static void seed_chase(float *seeds, int linesper, long n){
452 long *posstack=alloca(n*sizeof(*posstack));
453 float *ampstack=alloca(n*sizeof(*ampstack));
454 long stack=0;
455 long pos=0;
456 long i;
457
458 for(i=0;i<n;i++){
459 if(stack<2){
460 posstack[stack]=i;
461 ampstack[stack++]=seeds[i];
462 }else{
463 while(1){
464 if(seeds[i]<ampstack[stack-1]){
465 posstack[stack]=i;
466 ampstack[stack++]=seeds[i];
467 break;
468 }else{
469 if(i<posstack[stack-1]+linesper){
470 if(stack>1 && ampstack[stack-1]<=ampstack[stack-2] &&
471 i<posstack[stack-2]+linesper){
472 /* we completely overlap, making stack-1 irrelevant. pop it */
473 stack--;
474 continue;
475 }
476 }
477 posstack[stack]=i;
478 ampstack[stack++]=seeds[i];
479 break;
480
481 }
482 }
483 }
484 }
485
486 /* the stack now contains only the positions that are relevant. Scan
487 'em straight through */
488
489 for(i=0;i<stack;i++){
490 long endpos;
491 if(i<stack-1 && ampstack[i+1]>ampstack[i]){
492 endpos=posstack[i+1];
493 }else{
494 endpos=posstack[i]+linesper+1; /* +1 is important, else bin 0 is
495 discarded in short frames */
496 }
497 if(endpos>n)endpos=n;
498 for(;pos<endpos;pos++)
499 seeds[pos]=ampstack[i];
500 }
501
502 /* there. Linear time. I now remember this was on a problem set I
503 had in Grad Skool... I didn't solve it at the time ;-) */
504
505 }
506
507 /* bleaugh, this is more complicated than it needs to be */
508 #include<stdio.h>
max_seeds(vorbis_look_psy * p,float * seed,float * flr)509 static void max_seeds(vorbis_look_psy *p,
510 float *seed,
511 float *flr){
512 long n=p->total_octave_lines;
513 int linesper=p->eighth_octave_lines;
514 long linpos=0;
515 long pos;
516
517 seed_chase(seed,linesper,n); /* for masking */
518
519 pos=p->octave[0]-p->firstoc-(linesper>>1);
520
521 while(linpos+1<p->n){
522 float minV=seed[pos];
523 long end=((p->octave[linpos]+p->octave[linpos+1])>>1)-p->firstoc;
524 if(minV>p->vi->tone_abs_limit)minV=p->vi->tone_abs_limit;
525 while(pos+1<=end){
526 pos++;
527 if((seed[pos]>NEGINF && seed[pos]<minV) || minV==NEGINF)
528 minV=seed[pos];
529 }
530
531 end=pos+p->firstoc;
532 for(;linpos<p->n && p->octave[linpos]<=end;linpos++)
533 if(flr[linpos]<minV)flr[linpos]=minV;
534 }
535
536 {
537 float minV=seed[p->total_octave_lines-1];
538 for(;linpos<p->n;linpos++)
539 if(flr[linpos]<minV)flr[linpos]=minV;
540 }
541
542 }
543
bark_noise_hybridmp(int n,const long * b,const float * f,float * noise,const float offset,const int fixed)544 static void bark_noise_hybridmp(int n,const long *b,
545 const float *f,
546 float *noise,
547 const float offset,
548 const int fixed){
549
550 float *N=alloca(n*sizeof(*N));
551 float *X=alloca(n*sizeof(*N));
552 float *XX=alloca(n*sizeof(*N));
553 float *Y=alloca(n*sizeof(*N));
554 float *XY=alloca(n*sizeof(*N));
555
556 float tN, tX, tXX, tY, tXY;
557 int i;
558
559 int lo, hi;
560 float R=0.f;
561 float A=0.f;
562 float B=0.f;
563 float D=1.f;
564 float w, x, y;
565
566 tN = tX = tXX = tY = tXY = 0.f;
567
568 y = f[0] + offset;
569 if (y < 1.f) y = 1.f;
570
571 w = y * y * .5;
572
573 tN += w;
574 tX += w;
575 tY += w * y;
576
577 N[0] = tN;
578 X[0] = tX;
579 XX[0] = tXX;
580 Y[0] = tY;
581 XY[0] = tXY;
582
583 for (i = 1, x = 1.f; i < n; i++, x += 1.f) {
584
585 y = f[i] + offset;
586 if (y < 1.f) y = 1.f;
587
588 w = y * y;
589
590 tN += w;
591 tX += w * x;
592 tXX += w * x * x;
593 tY += w * y;
594 tXY += w * x * y;
595
596 N[i] = tN;
597 X[i] = tX;
598 XX[i] = tXX;
599 Y[i] = tY;
600 XY[i] = tXY;
601 }
602
603 for (i = 0, x = 0.f;; i++, x += 1.f) {
604
605 lo = b[i] >> 16;
606 if( lo>=0 ) break;
607 hi = b[i] & 0xffff;
608
609 tN = N[hi] + N[-lo];
610 tX = X[hi] - X[-lo];
611 tXX = XX[hi] + XX[-lo];
612 tY = Y[hi] + Y[-lo];
613 tXY = XY[hi] - XY[-lo];
614
615 A = tY * tXX - tX * tXY;
616 B = tN * tXY - tX * tY;
617 D = tN * tXX - tX * tX;
618 R = (A + x * B) / D;
619 if (R < 0.f)
620 R = 0.f;
621
622 noise[i] = R - offset;
623 }
624
625 for ( ;; i++, x += 1.f) {
626
627 lo = b[i] >> 16;
628 hi = b[i] & 0xffff;
629 if(hi>=n)break;
630
631 tN = N[hi] - N[lo];
632 tX = X[hi] - X[lo];
633 tXX = XX[hi] - XX[lo];
634 tY = Y[hi] - Y[lo];
635 tXY = XY[hi] - XY[lo];
636
637 A = tY * tXX - tX * tXY;
638 B = tN * tXY - tX * tY;
639 D = tN * tXX - tX * tX;
640 R = (A + x * B) / D;
641 if (R < 0.f) R = 0.f;
642
643 noise[i] = R - offset;
644 }
645 for ( ; i < n; i++, x += 1.f) {
646
647 R = (A + x * B) / D;
648 if (R < 0.f) R = 0.f;
649
650 noise[i] = R - offset;
651 }
652
653 if (fixed <= 0) return;
654
655 for (i = 0, x = 0.f;; i++, x += 1.f) {
656 hi = i + fixed / 2;
657 lo = hi - fixed;
658 if(lo>=0)break;
659
660 tN = N[hi] + N[-lo];
661 tX = X[hi] - X[-lo];
662 tXX = XX[hi] + XX[-lo];
663 tY = Y[hi] + Y[-lo];
664 tXY = XY[hi] - XY[-lo];
665
666
667 A = tY * tXX - tX * tXY;
668 B = tN * tXY - tX * tY;
669 D = tN * tXX - tX * tX;
670 R = (A + x * B) / D;
671
672 if (R - offset < noise[i]) noise[i] = R - offset;
673 }
674 for ( ;; i++, x += 1.f) {
675
676 hi = i + fixed / 2;
677 lo = hi - fixed;
678 if(hi>=n)break;
679
680 tN = N[hi] - N[lo];
681 tX = X[hi] - X[lo];
682 tXX = XX[hi] - XX[lo];
683 tY = Y[hi] - Y[lo];
684 tXY = XY[hi] - XY[lo];
685
686 A = tY * tXX - tX * tXY;
687 B = tN * tXY - tX * tY;
688 D = tN * tXX - tX * tX;
689 R = (A + x * B) / D;
690
691 if (R - offset < noise[i]) noise[i] = R - offset;
692 }
693 for ( ; i < n; i++, x += 1.f) {
694 R = (A + x * B) / D;
695 if (R - offset < noise[i]) noise[i] = R - offset;
696 }
697 }
698
_vp_noisemask(vorbis_look_psy * p,float * logmdct,float * logmask)699 void _vp_noisemask(vorbis_look_psy *p,
700 float *logmdct,
701 float *logmask){
702
703 int i,n=p->n;
704 float *work=alloca(n*sizeof(*work));
705
706 bark_noise_hybridmp(n,p->bark,logmdct,logmask,
707 140.,-1);
708
709 for(i=0;i<n;i++)work[i]=logmdct[i]-logmask[i];
710
711 bark_noise_hybridmp(n,p->bark,work,logmask,0.,
712 p->vi->noisewindowfixed);
713
714 for(i=0;i<n;i++)work[i]=logmdct[i]-work[i];
715
716 #if 0
717 {
718 static int seq=0;
719
720 float work2[n];
721 for(i=0;i<n;i++){
722 work2[i]=logmask[i]+work[i];
723 }
724
725 if(seq&1)
726 _analysis_output("median2R",seq/2,work,n,1,0,0);
727 else
728 _analysis_output("median2L",seq/2,work,n,1,0,0);
729
730 if(seq&1)
731 _analysis_output("envelope2R",seq/2,work2,n,1,0,0);
732 else
733 _analysis_output("envelope2L",seq/2,work2,n,1,0,0);
734 seq++;
735 }
736 #endif
737
738 for(i=0;i<n;i++){
739 int dB=logmask[i]+.5;
740 if(dB>=NOISE_COMPAND_LEVELS)dB=NOISE_COMPAND_LEVELS-1;
741 if(dB<0)dB=0;
742 logmask[i]= work[i]+p->vi->noisecompand[dB];
743 }
744
745 }
746
_vp_tonemask(vorbis_look_psy * p,float * logfft,float * logmask,float global_specmax,float local_specmax)747 void _vp_tonemask(vorbis_look_psy *p,
748 float *logfft,
749 float *logmask,
750 float global_specmax,
751 float local_specmax){
752
753 int i,n=p->n;
754
755 float *seed=alloca(sizeof(*seed)*p->total_octave_lines);
756 float att=local_specmax+p->vi->ath_adjatt;
757 for(i=0;i<p->total_octave_lines;i++)seed[i]=NEGINF;
758
759 /* set the ATH (floating below localmax, not global max by a
760 specified att) */
761 if(att<p->vi->ath_maxatt)att=p->vi->ath_maxatt;
762
763 for(i=0;i<n;i++)
764 logmask[i]=p->ath[i]+att;
765
766 /* tone masking */
767 seed_loop(p,(const float ***)p->tonecurves,logfft,logmask,seed,global_specmax);
768 max_seeds(p,seed,logmask);
769
770 }
771
_vp_offset_and_mix(vorbis_look_psy * p,float * noise,float * tone,int offset_select,float * logmask,float * mdct,float * logmdct)772 void _vp_offset_and_mix(vorbis_look_psy *p,
773 float *noise,
774 float *tone,
775 int offset_select,
776 float *logmask,
777 float *mdct,
778 float *logmdct){
779 int i,n=p->n;
780 float de, coeffi, cx;/* AoTuV */
781 float toneatt=p->vi->tone_masteratt[offset_select];
782
783 cx = p->m_val;
784
785 for(i=0;i<n;i++){
786 float val= noise[i]+p->noiseoffset[offset_select][i];
787 if(val>p->vi->noisemaxsupp)val=p->vi->noisemaxsupp;
788 logmask[i]=max(val,tone[i]+toneatt);
789
790
791 /* AoTuV */
792 /** @ M1 **
793 The following codes improve a noise problem.
794 A fundamental idea uses the value of masking and carries out
795 the relative compensation of the MDCT.
796 However, this code is not perfect and all noise problems cannot be solved.
797 by Aoyumi @ 2004/04/18
798 */
799
800 if(offset_select == 1) {
801 coeffi = -17.2; /* coeffi is a -17.2dB threshold */
802 val = val - logmdct[i]; /* val == mdct line value relative to floor in dB */
803
804 if(val > coeffi){
805 /* mdct value is > -17.2 dB below floor */
806
807 de = 1.0-((val-coeffi)*0.005*cx);
808 /* pro-rated attenuation:
809 -0.00 dB boost if mdct value is -17.2dB (relative to floor)
810 -0.77 dB boost if mdct value is 0dB (relative to floor)
811 -1.64 dB boost if mdct value is +17.2dB (relative to floor)
812 etc... */
813
814 if(de < 0) de = 0.0001;
815 }else
816 /* mdct value is <= -17.2 dB below floor */
817
818 de = 1.0-((val-coeffi)*0.0003*cx);
819 /* pro-rated attenuation:
820 +0.00 dB atten if mdct value is -17.2dB (relative to floor)
821 +0.45 dB atten if mdct value is -34.4dB (relative to floor)
822 etc... */
823
824 mdct[i] *= de;
825
826 }
827 }
828 }
829
_vp_ampmax_decay(float amp,vorbis_dsp_state * vd)830 float _vp_ampmax_decay(float amp,vorbis_dsp_state *vd){
831 vorbis_info *vi=vd->vi;
832 codec_setup_info *ci=vi->codec_setup;
833 vorbis_info_psy_global *gi=&ci->psy_g_param;
834
835 int n=ci->blocksizes[vd->W]/2;
836 float secs=(float)n/vi->rate;
837
838 amp+=secs*gi->ampmax_att_per_sec;
839 if(amp<-9999)amp=-9999;
840 return(amp);
841 }
842
843 static float FLOOR1_fromdB_LOOKUP[256]={
844 1.0649863e-07F, 1.1341951e-07F, 1.2079015e-07F, 1.2863978e-07F,
845 1.3699951e-07F, 1.4590251e-07F, 1.5538408e-07F, 1.6548181e-07F,
846 1.7623575e-07F, 1.8768855e-07F, 1.9988561e-07F, 2.128753e-07F,
847 2.2670913e-07F, 2.4144197e-07F, 2.5713223e-07F, 2.7384213e-07F,
848 2.9163793e-07F, 3.1059021e-07F, 3.3077411e-07F, 3.5226968e-07F,
849 3.7516214e-07F, 3.9954229e-07F, 4.2550680e-07F, 4.5315863e-07F,
850 4.8260743e-07F, 5.1396998e-07F, 5.4737065e-07F, 5.8294187e-07F,
851 6.2082472e-07F, 6.6116941e-07F, 7.0413592e-07F, 7.4989464e-07F,
852 7.9862701e-07F, 8.5052630e-07F, 9.0579828e-07F, 9.6466216e-07F,
853 1.0273513e-06F, 1.0941144e-06F, 1.1652161e-06F, 1.2409384e-06F,
854 1.3215816e-06F, 1.4074654e-06F, 1.4989305e-06F, 1.5963394e-06F,
855 1.7000785e-06F, 1.8105592e-06F, 1.9282195e-06F, 2.0535261e-06F,
856 2.1869758e-06F, 2.3290978e-06F, 2.4804557e-06F, 2.6416497e-06F,
857 2.8133190e-06F, 2.9961443e-06F, 3.1908506e-06F, 3.3982101e-06F,
858 3.6190449e-06F, 3.8542308e-06F, 4.1047004e-06F, 4.3714470e-06F,
859 4.6555282e-06F, 4.9580707e-06F, 5.2802740e-06F, 5.6234160e-06F,
860 5.9888572e-06F, 6.3780469e-06F, 6.7925283e-06F, 7.2339451e-06F,
861 7.7040476e-06F, 8.2047000e-06F, 8.7378876e-06F, 9.3057248e-06F,
862 9.9104632e-06F, 1.0554501e-05F, 1.1240392e-05F, 1.1970856e-05F,
863 1.2748789e-05F, 1.3577278e-05F, 1.4459606e-05F, 1.5399272e-05F,
864 1.6400004e-05F, 1.7465768e-05F, 1.8600792e-05F, 1.9809576e-05F,
865 2.1096914e-05F, 2.2467911e-05F, 2.3928002e-05F, 2.5482978e-05F,
866 2.7139006e-05F, 2.8902651e-05F, 3.0780908e-05F, 3.2781225e-05F,
867 3.4911534e-05F, 3.7180282e-05F, 3.9596466e-05F, 4.2169667e-05F,
868 4.4910090e-05F, 4.7828601e-05F, 5.0936773e-05F, 5.4246931e-05F,
869 5.7772202e-05F, 6.1526565e-05F, 6.5524908e-05F, 6.9783085e-05F,
870 7.4317983e-05F, 7.9147585e-05F, 8.4291040e-05F, 8.9768747e-05F,
871 9.5602426e-05F, 0.00010181521F, 0.00010843174F, 0.00011547824F,
872 0.00012298267F, 0.00013097477F, 0.00013948625F, 0.00014855085F,
873 0.00015820453F, 0.00016848555F, 0.00017943469F, 0.00019109536F,
874 0.00020351382F, 0.00021673929F, 0.00023082423F, 0.00024582449F,
875 0.00026179955F, 0.00027881276F, 0.00029693158F, 0.00031622787F,
876 0.00033677814F, 0.00035866388F, 0.00038197188F, 0.00040679456F,
877 0.00043323036F, 0.00046138411F, 0.00049136745F, 0.00052329927F,
878 0.00055730621F, 0.00059352311F, 0.00063209358F, 0.00067317058F,
879 0.00071691700F, 0.00076350630F, 0.00081312324F, 0.00086596457F,
880 0.00092223983F, 0.00098217216F, 0.0010459992F, 0.0011139742F,
881 0.0011863665F, 0.0012634633F, 0.0013455702F, 0.0014330129F,
882 0.0015261382F, 0.0016253153F, 0.0017309374F, 0.0018434235F,
883 0.0019632195F, 0.0020908006F, 0.0022266726F, 0.0023713743F,
884 0.0025254795F, 0.0026895994F, 0.0028643847F, 0.0030505286F,
885 0.0032487691F, 0.0034598925F, 0.0036847358F, 0.0039241906F,
886 0.0041792066F, 0.0044507950F, 0.0047400328F, 0.0050480668F,
887 0.0053761186F, 0.0057254891F, 0.0060975636F, 0.0064938176F,
888 0.0069158225F, 0.0073652516F, 0.0078438871F, 0.0083536271F,
889 0.0088964928F, 0.009474637F, 0.010090352F, 0.010746080F,
890 0.011444421F, 0.012188144F, 0.012980198F, 0.013823725F,
891 0.014722068F, 0.015678791F, 0.016697687F, 0.017782797F,
892 0.018938423F, 0.020169149F, 0.021479854F, 0.022875735F,
893 0.024362330F, 0.025945531F, 0.027631618F, 0.029427276F,
894 0.031339626F, 0.033376252F, 0.035545228F, 0.037855157F,
895 0.040315199F, 0.042935108F, 0.045725273F, 0.048696758F,
896 0.051861348F, 0.055231591F, 0.058820850F, 0.062643361F,
897 0.066714279F, 0.071049749F, 0.075666962F, 0.080584227F,
898 0.085821044F, 0.091398179F, 0.097337747F, 0.10366330F,
899 0.11039993F, 0.11757434F, 0.12521498F, 0.13335215F,
900 0.14201813F, 0.15124727F, 0.16107617F, 0.17154380F,
901 0.18269168F, 0.19456402F, 0.20720788F, 0.22067342F,
902 0.23501402F, 0.25028656F, 0.26655159F, 0.28387361F,
903 0.30232132F, 0.32196786F, 0.34289114F, 0.36517414F,
904 0.38890521F, 0.41417847F, 0.44109412F, 0.46975890F,
905 0.50028648F, 0.53279791F, 0.56742212F, 0.60429640F,
906 0.64356699F, 0.68538959F, 0.72993007F, 0.77736504F,
907 0.82788260F, 0.88168307F, 0.9389798F, 1.F,
908 };
909
910 /* this is for per-channel noise normalization */
apsort(const void * a,const void * b)911 static int apsort(const void *a, const void *b){
912 float f1=**(float**)a;
913 float f2=**(float**)b;
914 return (f1<f2)-(f1>f2);
915 }
916
flag_lossless(int limit,float prepoint,float postpoint,float * mdct,float * floor,int * flag,int i,int jn)917 static void flag_lossless(int limit, float prepoint, float postpoint, float *mdct,
918 float *floor, int *flag, int i, int jn){
919 int j;
920 for(j=0;j<jn;j++){
921 float point = j>=limit-i ? postpoint : prepoint;
922 float r = fabs(mdct[j])/floor[j];
923 if(r<point)
924 flag[j]=0;
925 else
926 flag[j]=1;
927 }
928 }
929
930 /* Overload/Side effect: On input, the *q vector holds either the
931 quantized energy (for elements with the flag set) or the absolute
932 values of the *r vector (for elements with flag unset). On output,
933 *q holds the quantized energy for all elements */
noise_normalize(vorbis_look_psy * p,int limit,float * r,float * q,float * f,int * flags,float acc,int i,int n,int * out)934 static float noise_normalize(vorbis_look_psy *p, int limit, float *r, float *q, float *f, int *flags, float acc, int i, int n, int *out){
935
936 vorbis_info_psy *vi=p->vi;
937 float **sort = alloca(n*sizeof(*sort));
938 int j,count=0;
939 int start = (vi->normal_p ? vi->normal_start-i : n);
940 if(start>n)start=n;
941
942 /* force classic behavior where only energy in the current band is considered */
943 acc=0.f;
944
945 /* still responsible for populating *out where noise norm not in
946 effect. There's no need to [re]populate *q in these areas */
947 for(j=0;j<start;j++){
948 if(!flags || !flags[j]){ /* lossless coupling already quantized.
949 Don't touch; requantizing based on
950 energy would be incorrect. */
951 float ve = q[j]/f[j];
952 if(r[j]<0)
953 out[j] = -rint(sqrt(ve));
954 else
955 out[j] = rint(sqrt(ve));
956 }
957 }
958
959 /* sort magnitudes for noise norm portion of partition */
960 for(;j<n;j++){
961 if(!flags || !flags[j]){ /* can't noise norm elements that have
962 already been loslessly coupled; we can
963 only account for their energy error */
964 float ve = q[j]/f[j];
965 /* Despite all the new, more capable coupling code, for now we
966 implement noise norm as it has been up to this point. Only
967 consider promotions to unit magnitude from 0. In addition
968 the only energy error counted is quantizations to zero. */
969 /* also-- the original point code only applied noise norm at > pointlimit */
970 if(ve<.25f && (!flags || j>=limit-i)){
971 acc += ve;
972 sort[count++]=q+j; /* q is fabs(r) for unflagged element */
973 }else{
974 /* For now: no acc adjustment for nonzero quantization. populate *out and q as this value is final. */
975 if(r[j]<0)
976 out[j] = -rint(sqrt(ve));
977 else
978 out[j] = rint(sqrt(ve));
979 q[j] = out[j]*out[j]*f[j];
980 }
981 }/* else{
982 again, no energy adjustment for error in nonzero quant-- for now
983 }*/
984 }
985
986 if(count){
987 /* noise norm to do */
988 qsort(sort,count,sizeof(*sort),apsort);
989 for(j=0;j<count;j++){
990 int k=sort[j]-q;
991 if(acc>=vi->normal_thresh){
992 out[k]=unitnorm(r[k]);
993 acc-=1.f;
994 q[k]=f[k];
995 }else{
996 out[k]=0;
997 q[k]=0.f;
998 }
999 }
1000 }
1001
1002 return acc;
1003 }
1004
1005 /* Noise normalization, quantization and coupling are not wholly
1006 seperable processes in depth>1 coupling. */
_vp_couple_quantize_normalize(int blobno,vorbis_info_psy_global * g,vorbis_look_psy * p,vorbis_info_mapping0 * vi,float ** mdct,int ** iwork,int * nonzero,int sliding_lowpass,int ch)1007 void _vp_couple_quantize_normalize(int blobno,
1008 vorbis_info_psy_global *g,
1009 vorbis_look_psy *p,
1010 vorbis_info_mapping0 *vi,
1011 float **mdct,
1012 int **iwork,
1013 int *nonzero,
1014 int sliding_lowpass,
1015 int ch){
1016
1017 int i;
1018 int n = p->n;
1019 int partition=(p->vi->normal_p ? p->vi->normal_partition : 16);
1020 int limit = g->coupling_pointlimit[p->vi->blockflag][blobno];
1021 float prepoint=stereo_threshholds[g->coupling_prepointamp[blobno]];
1022 float postpoint=stereo_threshholds[g->coupling_postpointamp[blobno]];
1023 float de=0.1*p->m_val; /* a blend of the AoTuV M2 and M3 code here and below */
1024
1025 /* mdct is our raw mdct output, floor not removed. */
1026 /* inout passes in the ifloor, passes back quantized result */
1027
1028 /* unquantized energy (negative indicates amplitude has negative sign) */
1029 float **raw = alloca(ch*sizeof(*raw));
1030
1031 /* dual pupose; quantized energy (if flag set), othersize fabs(raw) */
1032 float **quant = alloca(ch*sizeof(*quant));
1033
1034 /* floor energy */
1035 float **floor = alloca(ch*sizeof(*floor));
1036
1037 /* flags indicating raw/quantized status of elements in raw vector */
1038 int **flag = alloca(ch*sizeof(*flag));
1039
1040 /* non-zero flag working vector */
1041 int *nz = alloca(ch*sizeof(*nz));
1042
1043 /* energy surplus/defecit tracking */
1044 float *acc = alloca((ch+vi->coupling_steps)*sizeof(*acc));
1045
1046 /* The threshold of a stereo is changed with the size of n */
1047 if(n > 1000)
1048 postpoint=stereo_threshholds_limited[g->coupling_postpointamp[blobno]];
1049
1050 raw[0] = alloca(ch*partition*sizeof(**raw));
1051 quant[0] = alloca(ch*partition*sizeof(**quant));
1052 floor[0] = alloca(ch*partition*sizeof(**floor));
1053 flag[0] = alloca(ch*partition*sizeof(**flag));
1054
1055 for(i=1;i<ch;i++){
1056 raw[i] = &raw[0][partition*i];
1057 quant[i] = &quant[0][partition*i];
1058 floor[i] = &floor[0][partition*i];
1059 flag[i] = &flag[0][partition*i];
1060 }
1061 for(i=0;i<ch+vi->coupling_steps;i++)
1062 acc[i]=0.f;
1063
1064 for(i=0;i<n;i+=partition){
1065 int k,j,jn = partition > n-i ? n-i : partition;
1066 int step,track = 0;
1067
1068 memcpy(nz,nonzero,sizeof(*nz)*ch);
1069
1070 /* prefill */
1071 memset(flag[0],0,ch*partition*sizeof(**flag));
1072 for(k=0;k<ch;k++){
1073 int *iout = &iwork[k][i];
1074 if(nz[k]){
1075
1076 for(j=0;j<jn;j++)
1077 floor[k][j] = FLOOR1_fromdB_LOOKUP[iout[j]];
1078
1079 flag_lossless(limit,prepoint,postpoint,&mdct[k][i],floor[k],flag[k],i,jn);
1080
1081 for(j=0;j<jn;j++){
1082 quant[k][j] = raw[k][j] = mdct[k][i+j]*mdct[k][i+j];
1083 if(mdct[k][i+j]<0.f) raw[k][j]*=-1.f;
1084 floor[k][j]*=floor[k][j];
1085 }
1086
1087 acc[track]=noise_normalize(p,limit,raw[k],quant[k],floor[k],NULL,acc[track],i,jn,iout);
1088
1089 }else{
1090 for(j=0;j<jn;j++){
1091 floor[k][j] = 1e-10f;
1092 raw[k][j] = 0.f;
1093 quant[k][j] = 0.f;
1094 flag[k][j] = 0;
1095 iout[j]=0;
1096 }
1097 acc[track]=0.f;
1098 }
1099 track++;
1100 }
1101
1102 /* coupling */
1103 for(step=0;step<vi->coupling_steps;step++){
1104 int Mi = vi->coupling_mag[step];
1105 int Ai = vi->coupling_ang[step];
1106 int *iM = &iwork[Mi][i];
1107 int *iA = &iwork[Ai][i];
1108 float *reM = raw[Mi];
1109 float *reA = raw[Ai];
1110 float *qeM = quant[Mi];
1111 float *qeA = quant[Ai];
1112 float *floorM = floor[Mi];
1113 float *floorA = floor[Ai];
1114 int *fM = flag[Mi];
1115 int *fA = flag[Ai];
1116
1117 if(nz[Mi] || nz[Ai]){
1118 nz[Mi] = nz[Ai] = 1;
1119
1120 for(j=0;j<jn;j++){
1121
1122 if(j<sliding_lowpass-i){
1123 if(fM[j] || fA[j]){
1124 /* lossless coupling */
1125
1126 reM[j] = fabs(reM[j])+fabs(reA[j]);
1127 qeM[j] = qeM[j]+qeA[j];
1128 fM[j]=fA[j]=1;
1129
1130 /* couple iM/iA */
1131 {
1132 int A = iM[j];
1133 int B = iA[j];
1134
1135 if(abs(A)>abs(B)){
1136 iA[j]=(A>0?A-B:B-A);
1137 }else{
1138 iA[j]=(B>0?A-B:B-A);
1139 iM[j]=B;
1140 }
1141
1142 /* collapse two equivalent tuples to one */
1143 if(iA[j]>=abs(iM[j])*2){
1144 iA[j]= -iA[j];
1145 iM[j]= -iM[j];
1146 }
1147
1148 }
1149
1150 }else{
1151 /* lossy (point) coupling */
1152 if(j<limit-i){
1153 /* dipole */
1154 reM[j] += reA[j];
1155 qeM[j] = fabs(reM[j]);
1156 }else{
1157 /* AoTuV */
1158 /** @ M2 **
1159 The boost problem by the combination of noise normalization and point stereo is eased.
1160 However, this is a temporary patch.
1161 by Aoyumi @ 2004/04/18
1162 */
1163 float derate = (1.0 - de*((float)(j-limit+i) / (float)(n-limit)));
1164
1165 /* elliptical */
1166 if(reM[j]+reA[j]<0){
1167 reM[j] = - (qeM[j] = (fabs(reM[j])+fabs(reA[j]))*derate*derate);
1168 }else{
1169 reM[j] = (qeM[j] = (fabs(reM[j])+fabs(reA[j]))*derate*derate);
1170 }
1171 }
1172 reA[j]=qeA[j]=0.f;
1173 fA[j]=1;
1174 iA[j]=0;
1175 }
1176 }
1177 floorM[j]=floorA[j]=floorM[j]+floorA[j];
1178 }
1179 /* normalize the resulting mag vector */
1180 acc[track]=noise_normalize(p,limit,raw[Mi],quant[Mi],floor[Mi],flag[Mi],acc[track],i,jn,iM);
1181 track++;
1182 }
1183 }
1184 }
1185
1186 for(i=0;i<vi->coupling_steps;i++){
1187 /* make sure coupling a zero and a nonzero channel results in two
1188 nonzero channels. */
1189 if(nonzero[vi->coupling_mag[i]] ||
1190 nonzero[vi->coupling_ang[i]]){
1191 nonzero[vi->coupling_mag[i]]=1;
1192 nonzero[vi->coupling_ang[i]]=1;
1193 }
1194 }
1195 }
1196