1 /********************************************************************
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. *
8 * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2010 *
9 * by the Xiph.Org Foundation http://www.xiph.org/ *
11 ********************************************************************
13 function: residue backend 0, 1 and 2 implementation
16 ********************************************************************/
18 /* Slow, slow, slow, simpleminded and did I mention it was slow? The
19 encode/decode loops are coded for clarity and performance is not
20 yet even a nagging little idea lurking in the shadows. Oh and BTW,
27 #include "vorbis/codec.h"
28 #include "codec_internal.h"
34 #if defined(TRAIN_RES) || defined (TRAIN_RESAUX)
39 vorbis_info_residue0 *info;
45 codebook ***partbooks;
54 #if defined(TRAIN_RES) || defined(TRAIN_RESAUX)
56 long *training_data[8][64];
57 float training_max[8][64];
58 float training_min[8][64];
63 } vorbis_look_residue0;
65 void res0_free_info(vorbis_info_residue *i){
66 vorbis_info_residue0 *info=(vorbis_info_residue0 *)i;
68 memset(info,0,sizeof(*info));
73 void res0_free_look(vorbis_look_residue *i){
77 vorbis_look_residue0 *look=(vorbis_look_residue0 *)i;
82 for(j=0;j<look->parts;j++){
83 /*fprintf(stderr,"partition %d: ",j);*/
85 if(look->training_data[k][j]){
88 codebook *statebook=look->partbooks[j][k];
90 /* long and short into the same bucket by current convention */
91 sprintf(buffer,"res_part%d_pass%d.vqd",j,k);
94 for(l=0;l<statebook->entries;l++)
95 fprintf(of,"%d:%ld\n",l,look->training_data[k][j][l]);
99 /*fprintf(stderr,"%d(%.2f|%.2f) ",k,
100 look->training_min[k][j],look->training_max[k][j]);*/
102 _ogg_free(look->training_data[k][j]);
103 look->training_data[k][j]=NULL;
105 /*fprintf(stderr,"\n");*/
108 fprintf(stderr,"min/max residue: %g::%g\n",look->tmin,look->tmax);
110 /*fprintf(stderr,"residue bit usage %f:%f (%f total)\n",
111 (float)look->phrasebits/look->frames,
112 (float)look->postbits/look->frames,
113 (float)(look->postbits+look->phrasebits)/look->frames);*/
117 /*vorbis_info_residue0 *info=look->info;
120 "%ld frames encoded in %ld phrasebits and %ld residue bits "
121 "(%g/frame) \n",look->frames,look->phrasebits,
123 (look->phrasebits+look->resbitsflat)/(float)look->frames);
125 for(j=0;j<look->parts;j++){
127 fprintf(stderr,"\t[%d] == ",j);
128 for(k=0;k<look->stages;k++)
129 if((info->secondstages[j]>>k)&1){
130 fprintf(stderr,"%ld,",look->resbits[j][k]);
131 acc+=look->resbits[j][k];
134 fprintf(stderr,":: (%ld vals) %1.2fbits/sample\n",look->resvals[j],
135 acc?(float)acc/(look->resvals[j]*info->grouping):0);
137 fprintf(stderr,"\n");*/
139 for(j=0;j<look->parts;j++)
140 if(look->partbooks[j])_ogg_free(look->partbooks[j]);
141 _ogg_free(look->partbooks);
142 for(j=0;j<look->partvals;j++)
143 _ogg_free(look->decodemap[j]);
144 _ogg_free(look->decodemap);
146 memset(look,0,sizeof(*look));
151 static int ilog(unsigned int v){
160 static int icount(unsigned int v){
170 void res0_pack(vorbis_info_residue *vr,oggpack_buffer *opb){
171 vorbis_info_residue0 *info=(vorbis_info_residue0 *)vr;
173 oggpack_write(opb,info->begin,24);
174 oggpack_write(opb,info->end,24);
176 oggpack_write(opb,info->grouping-1,24); /* residue vectors to group and
177 code with a partitioned book */
178 oggpack_write(opb,info->partitions-1,6); /* possible partition choices */
179 oggpack_write(opb,info->groupbook,8); /* group huffman book */
181 /* secondstages is a bitmask; as encoding progresses pass by pass, a
182 bitmask of one indicates this partition class has bits to write
184 for(j=0;j<info->partitions;j++){
185 if(ilog(info->secondstages[j])>3){
186 /* yes, this is a minor hack due to not thinking ahead */
187 oggpack_write(opb,info->secondstages[j],3);
188 oggpack_write(opb,1,1);
189 oggpack_write(opb,info->secondstages[j]>>3,5);
191 oggpack_write(opb,info->secondstages[j],4); /* trailing zero */
192 acc+=icount(info->secondstages[j]);
195 oggpack_write(opb,info->booklist[j],8);
199 /* vorbis_info is for range checking */
200 vorbis_info_residue *res0_unpack(vorbis_info *vi,oggpack_buffer *opb){
202 vorbis_info_residue0 *info=_ogg_calloc(1,sizeof(*info));
203 codec_setup_info *ci=vi->codec_setup;
205 info->begin=oggpack_read(opb,24);
206 info->end=oggpack_read(opb,24);
207 info->grouping=oggpack_read(opb,24)+1;
208 info->partitions=oggpack_read(opb,6)+1;
209 info->groupbook=oggpack_read(opb,8);
211 /* check for premature EOP */
212 if(info->groupbook<0)goto errout;
214 for(j=0;j<info->partitions;j++){
215 int cascade=oggpack_read(opb,3);
216 int cflag=oggpack_read(opb,1);
217 if(cflag<0) goto errout;
219 int c=oggpack_read(opb,5);
223 info->secondstages[j]=cascade;
225 acc+=icount(cascade);
228 int book=oggpack_read(opb,8);
229 if(book<0) goto errout;
230 info->booklist[j]=book;
233 if(info->groupbook>=ci->books)goto errout;
235 if(info->booklist[j]>=ci->books)goto errout;
236 if(ci->book_param[info->booklist[j]]->maptype==0)goto errout;
239 /* verify the phrasebook is not specifying an impossible or
240 inconsistent partitioning scheme. */
241 /* modify the phrasebook ranging check from r16327; an early beta
242 encoder had a bug where it used an oversized phrasebook by
243 accident. These files should continue to be playable, but don't
246 int entries = ci->book_param[info->groupbook]->entries;
247 int dim = ci->book_param[info->groupbook]->dim;
250 partvals *= info->partitions;
251 if(partvals > entries) goto errout;
254 info->partvals = partvals;
259 res0_free_info(info);
263 vorbis_look_residue *res0_look(vorbis_dsp_state *vd,
264 vorbis_info_residue *vr){
265 vorbis_info_residue0 *info=(vorbis_info_residue0 *)vr;
266 vorbis_look_residue0 *look=_ogg_calloc(1,sizeof(*look));
267 codec_setup_info *ci=vd->vi->codec_setup;
274 look->parts=info->partitions;
275 look->fullbooks=ci->fullbooks;
276 look->phrasebook=ci->fullbooks+info->groupbook;
277 dim=look->phrasebook->dim;
279 look->partbooks=_ogg_calloc(look->parts,sizeof(*look->partbooks));
281 for(j=0;j<look->parts;j++){
282 int stages=ilog(info->secondstages[j]);
284 if(stages>maxstage)maxstage=stages;
285 look->partbooks[j]=_ogg_calloc(stages,sizeof(*look->partbooks[j]));
286 for(k=0;k<stages;k++)
287 if(info->secondstages[j]&(1<<k)){
288 look->partbooks[j][k]=ci->fullbooks+info->booklist[acc++];
290 look->training_data[k][j]=_ogg_calloc(look->partbooks[j][k]->entries,
291 sizeof(***look->training_data));
299 look->partvals*=look->parts;
301 look->stages=maxstage;
302 look->decodemap=_ogg_malloc(look->partvals*sizeof(*look->decodemap));
303 for(j=0;j<look->partvals;j++){
305 long mult=look->partvals/look->parts;
306 look->decodemap[j]=_ogg_malloc(dim*sizeof(*look->decodemap[j]));
311 look->decodemap[j][k]=deco;
314 #if defined(TRAIN_RES) || defined (TRAIN_RESAUX)
316 static int train_seq=0;
317 look->train_seq=train_seq++;
323 /* break an abstraction and copy some code for performance purposes */
324 static int local_book_besterror(codebook *book,int *a){
327 int minval=book->minval;
329 int qv=book->quantvals;
332 /* assumes integer/centered encoder codebook maptype 1 no more than dim 8 */
333 int p[8]={0,0,0,0,0,0,0,0};
336 for(i=0,o=dim;i<dim;i++){
337 int v = (a[--o]-minval+(del>>1))/del;
338 int m = (v<ze ? ((ze-v)<<1)-1 : ((v-ze)<<1));
339 index = index*qv+ (m<0?0:(m>=qv?qv-1:m));
343 for(i=0,o=dim;i<dim;i++){
344 int v = a[--o]-minval;
345 int m = (v<ze ? ((ze-v)<<1)-1 : ((v-ze)<<1));
346 index = index*qv+ (m<0?0:(m>=qv?qv-1:m));
351 if(book->c->lengthlist[index]<=0){
352 const static_codebook *c=book->c;
354 /* assumes integer/centered encoder codebook maptype 1 no more than dim 8 */
355 int e[8]={0,0,0,0,0,0,0,0};
356 int maxval = book->minval + book->delta*(book->quantvals-1);
357 for(i=0;i<book->entries;i++){
358 if(c->lengthlist[i]>0){
361 float val=(e[j]-a[j]);
364 if(best==-1 || this<best){
365 memcpy(p,e,sizeof(p));
370 /* assumes the value patterning created by the tools in vq/ */
388 static int _encodepart(oggpack_buffer *opb,int *vec, int n,
389 codebook *book,long *acc){
395 int entry=local_book_besterror(book,vec+i*dim);
402 bits+=vorbis_book_encode(book,entry,opb);
409 static long **_01class(vorbis_block *vb,vorbis_look_residue *vl,
412 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
413 vorbis_info_residue0 *info=look->info;
415 /* move all this setup out later */
416 int samples_per_partition=info->grouping;
417 int possible_partitions=info->partitions;
418 int n=info->end-info->begin;
420 int partvals=n/samples_per_partition;
421 long **partword=_vorbis_block_alloc(vb,ch*sizeof(*partword));
422 float scale=100./samples_per_partition;
424 /* we find the partition type for each partition of each
425 channel. We'll go back and do the interleaved encoding in a
426 bit. For now, clarity */
429 partword[i]=_vorbis_block_alloc(vb,n/samples_per_partition*sizeof(*partword[i]));
430 memset(partword[i],0,n/samples_per_partition*sizeof(*partword[i]));
433 for(i=0;i<partvals;i++){
434 int offset=i*samples_per_partition+info->begin;
438 for(k=0;k<samples_per_partition;k++){
439 if(abs(in[j][offset+k])>max)max=abs(in[j][offset+k]);
440 ent+=abs(in[j][offset+k]);
444 for(k=0;k<possible_partitions-1;k++)
445 if(max<=info->classmetric1[k] &&
446 (info->classmetric2[k]<0 || ent<info->classmetric2[k]))
459 sprintf(buffer,"resaux_%d.vqd",look->train_seq);
460 of=fopen(buffer,"a");
461 for(j=0;j<partvals;j++)
462 fprintf(of,"%ld, ",partword[i][j]);
473 /* designed for stereo or other modes where the partition size is an
474 integer multiple of the number of channels encoded in the current
476 static long **_2class(vorbis_block *vb,vorbis_look_residue *vl,int **in,
479 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
480 vorbis_info_residue0 *info=look->info;
482 /* move all this setup out later */
483 int samples_per_partition=info->grouping;
484 int possible_partitions=info->partitions;
485 int n=info->end-info->begin;
487 int partvals=n/samples_per_partition;
488 long **partword=_vorbis_block_alloc(vb,sizeof(*partword));
490 #if defined(TRAIN_RES) || defined (TRAIN_RESAUX)
495 partword[0]=_vorbis_block_alloc(vb,n*ch/samples_per_partition*sizeof(*partword[0]));
496 memset(partword[0],0,n*ch/samples_per_partition*sizeof(*partword[0]));
498 for(i=0,l=info->begin/ch;i<partvals;i++){
501 for(j=0;j<samples_per_partition;j+=ch){
502 if(abs(in[0][l])>magmax)magmax=abs(in[0][l]);
504 if(abs(in[k][l])>angmax)angmax=abs(in[k][l]);
508 for(j=0;j<possible_partitions-1;j++)
509 if(magmax<=info->classmetric1[j] &&
510 angmax<=info->classmetric2[j])
518 sprintf(buffer,"resaux_%d.vqd",look->train_seq);
519 of=fopen(buffer,"a");
520 for(i=0;i<partvals;i++)
521 fprintf(of,"%ld, ",partword[0][i]);
531 static int _01forward(oggpack_buffer *opb,
532 vorbis_block *vb,vorbis_look_residue *vl,
535 int (*encode)(oggpack_buffer *,int *,int,
538 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
539 vorbis_info_residue0 *info=look->info;
541 /* move all this setup out later */
542 int samples_per_partition=info->grouping;
543 int possible_partitions=info->partitions;
544 int partitions_per_word=look->phrasebook->dim;
545 int n=info->end-info->begin;
547 int partvals=n/samples_per_partition;
553 for(j=info->begin;j<info->end;j++){
554 if(in[i][j]>look->tmax)look->tmax=in[i][j];
555 if(in[i][j]<look->tmin)look->tmin=in[i][j];
559 memset(resbits,0,sizeof(resbits));
560 memset(resvals,0,sizeof(resvals));
562 /* we code the partition words for each channel, then the residual
563 words for a partition per channel until we've written all the
564 residual words for that partition word. Then write the next
565 partition channel words... */
567 for(s=0;s<look->stages;s++){
569 for(i=0;i<partvals;){
571 /* first we encode a partition codeword for each channel */
574 long val=partword[j][i];
575 for(k=1;k<partitions_per_word;k++){
576 val*=possible_partitions;
578 val+=partword[j][i+k];
582 if(val<look->phrasebook->entries)
583 look->phrasebits+=vorbis_book_encode(look->phrasebook,val,opb);
584 #if 0 /*def TRAIN_RES*/
592 /* now we encode interleaved residual values for the partitions */
593 for(k=0;k<partitions_per_word && i<partvals;k++,i++){
594 long offset=i*samples_per_partition+info->begin;
597 if(s==0)resvals[partword[j][i]]+=samples_per_partition;
598 if(info->secondstages[partword[j][i]]&(1<<s)){
599 codebook *statebook=look->partbooks[partword[j][i]][s];
602 long *accumulator=NULL;
605 accumulator=look->training_data[s][partword[j][i]];
608 float *samples=in[j]+offset;
609 for(l=0;l<samples_per_partition;l++){
610 if(samples[l]<look->training_min[s][partword[j][i]])
611 look->training_min[s][partword[j][i]]=samples[l];
612 if(samples[l]>look->training_max[s][partword[j][i]])
613 look->training_max[s][partword[j][i]]=samples[l];
618 ret=encode(opb,in[j]+offset,samples_per_partition,
619 statebook,accumulator);
622 resbits[partword[j][i]]+=ret;
633 fprintf(stderr,"%d :: ",vb->mode);
634 for(k=0;k<possible_partitions;k++){
635 fprintf(stderr,"%ld/%1.2g, ",resvals[k],(float)resbits[k]/resvals[k]);
637 totalbits+=resbits[k];
640 fprintf(stderr,":: %ld:%1.2g\n",total,(double)totalbits/total);
646 /* a truncated packet here just means 'stop working'; it's not an error */
647 static int _01inverse(vorbis_block *vb,vorbis_look_residue *vl,
649 long (*decodepart)(codebook *, float *,
650 oggpack_buffer *,int)){
653 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
654 vorbis_info_residue0 *info=look->info;
656 /* move all this setup out later */
657 int samples_per_partition=info->grouping;
658 int partitions_per_word=look->phrasebook->dim;
659 int max=vb->pcmend>>1;
660 int end=(info->end<max?info->end:max);
661 int n=end-info->begin;
664 int partvals=n/samples_per_partition;
665 int partwords=(partvals+partitions_per_word-1)/partitions_per_word;
666 int ***partword=alloca(ch*sizeof(*partword));
669 partword[j]=_vorbis_block_alloc(vb,partwords*sizeof(*partword[j]));
671 for(s=0;s<look->stages;s++){
673 /* each loop decodes on partition codeword containing
674 partitions_per_word partitions */
675 for(i=0,l=0;i<partvals;l++){
677 /* fetch the partition word for each channel */
679 int temp=vorbis_book_decode(look->phrasebook,&vb->opb);
681 if(temp==-1 || temp>=info->partvals)goto eopbreak;
682 partword[j][l]=look->decodemap[temp];
683 if(partword[j][l]==NULL)goto errout;
687 /* now we decode residual values for the partitions */
688 for(k=0;k<partitions_per_word && i<partvals;k++,i++)
690 long offset=info->begin+i*samples_per_partition;
691 if(info->secondstages[partword[j][l][k]]&(1<<s)){
692 codebook *stagebook=look->partbooks[partword[j][l][k]][s];
694 if(decodepart(stagebook,in[j]+offset,&vb->opb,
695 samples_per_partition)==-1)goto eopbreak;
707 int res0_inverse(vorbis_block *vb,vorbis_look_residue *vl,
708 float **in,int *nonzero,int ch){
714 return(_01inverse(vb,vl,in,used,vorbis_book_decodevs_add));
719 int res1_forward(oggpack_buffer *opb,vorbis_block *vb,vorbis_look_residue *vl,
720 int **in,int *nonzero,int ch, long **partword){
727 return _01forward(opb,vb,vl,in,used,partword,_encodepart);
733 long **res1_class(vorbis_block *vb,vorbis_look_residue *vl,
734 int **in,int *nonzero,int ch){
740 return(_01class(vb,vl,in,used));
745 int res1_inverse(vorbis_block *vb,vorbis_look_residue *vl,
746 float **in,int *nonzero,int ch){
752 return(_01inverse(vb,vl,in,used,vorbis_book_decodev_add));
757 long **res2_class(vorbis_block *vb,vorbis_look_residue *vl,
758 int **in,int *nonzero,int ch){
761 if(nonzero[i])used++;
763 return(_2class(vb,vl,in,ch));
768 /* res2 is slightly more different; all the channels are interleaved
769 into a single vector and encoded. */
771 int res2_forward(oggpack_buffer *opb,
772 vorbis_block *vb,vorbis_look_residue *vl,
773 int **in,int *nonzero,int ch, long **partword){
774 long i,j,k,n=vb->pcmend/2,used=0;
776 /* don't duplicate the code; use a working vector hack for now and
777 reshape ourselves into a single channel res1 */
778 /* ugly; reallocs for each coupling pass :-( */
779 int *work=_vorbis_block_alloc(vb,ch*n*sizeof(*work));
782 if(nonzero[i])used++;
783 for(j=0,k=i;j<n;j++,k+=ch)
788 return _01forward(opb,vb,vl,&work,1,partword,_encodepart);
794 /* duplicate code here as speed is somewhat more important */
795 int res2_inverse(vorbis_block *vb,vorbis_look_residue *vl,
796 float **in,int *nonzero,int ch){
798 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
799 vorbis_info_residue0 *info=look->info;
801 /* move all this setup out later */
802 int samples_per_partition=info->grouping;
803 int partitions_per_word=look->phrasebook->dim;
804 int max=(vb->pcmend*ch)>>1;
805 int end=(info->end<max?info->end:max);
806 int n=end-info->begin;
809 int partvals=n/samples_per_partition;
810 int partwords=(partvals+partitions_per_word-1)/partitions_per_word;
811 int **partword=_vorbis_block_alloc(vb,partwords*sizeof(*partword));
813 for(i=0;i<ch;i++)if(nonzero[i])break;
814 if(i==ch)return(0); /* no nonzero vectors */
816 for(s=0;s<look->stages;s++){
817 for(i=0,l=0;i<partvals;l++){
820 /* fetch the partition word */
821 int temp=vorbis_book_decode(look->phrasebook,&vb->opb);
822 if(temp==-1 || temp>info->partvals)goto eopbreak;
823 partword[l]=look->decodemap[temp];
824 if(partword[l]==NULL)goto errout;
827 /* now we decode residual values for the partitions */
828 for(k=0;k<partitions_per_word && i<partvals;k++,i++)
829 if(info->secondstages[partword[l][k]]&(1<<s)){
830 codebook *stagebook=look->partbooks[partword[l][k]][s];
833 if(vorbis_book_decodevv_add(stagebook,in,
834 i*samples_per_partition+info->begin,ch,
835 &vb->opb,samples_per_partition)==-1)
848 const vorbis_func_residue residue0_exportbundle={
859 const vorbis_func_residue residue1_exportbundle={
870 const vorbis_func_residue residue2_exportbundle={