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 1994-2009 *
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,float *a){
325 int dim=book->dim,i,k,o;
327 encode_aux_threshmatch *tt=book->c->thresh_tree;
329 /* find the quant val of each scalar */
330 for(k=0,o=dim;k<dim;++k){
334 if(val<tt->quantthresh[i]){
335 if(val<tt->quantthresh[i-1]){
337 if(val>=tt->quantthresh[i-1])
342 for(++i;i<tt->threshvals-1;++i)
343 if(val<tt->quantthresh[i])break;
347 best=(best*tt->quantvals)+tt->quantmap[i];
349 /* regular lattices are easy :-) */
351 if(book->c->lengthlist[best]<=0){
352 const static_codebook *c=book->c;
355 float *e=book->valuelist;
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<bestf){
374 float *ptr=book->valuelist+best*dim;
382 static int _encodepart(oggpack_buffer *opb,float *vec, int n,
383 codebook *book,long *acc){
389 int entry=local_book_besterror(book,vec+i*dim);
396 bits+=vorbis_book_encode(book,entry,opb);
403 static long **_01class(vorbis_block *vb,vorbis_look_residue *vl,
406 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
407 vorbis_info_residue0 *info=look->info;
409 /* move all this setup out later */
410 int samples_per_partition=info->grouping;
411 int possible_partitions=info->partitions;
412 int n=info->end-info->begin;
414 int partvals=n/samples_per_partition;
415 long **partword=_vorbis_block_alloc(vb,ch*sizeof(*partword));
416 float scale=100./samples_per_partition;
418 /* we find the partition type for each partition of each
419 channel. We'll go back and do the interleaved encoding in a
420 bit. For now, clarity */
423 partword[i]=_vorbis_block_alloc(vb,n/samples_per_partition*sizeof(*partword[i]));
424 memset(partword[i],0,n/samples_per_partition*sizeof(*partword[i]));
427 for(i=0;i<partvals;i++){
428 int offset=i*samples_per_partition+info->begin;
432 for(k=0;k<samples_per_partition;k++){
433 if(fabs(in[j][offset+k])>max)max=fabs(in[j][offset+k]);
434 ent+=fabs(rint(in[j][offset+k]));
438 for(k=0;k<possible_partitions-1;k++)
439 if(max<=info->classmetric1[k] &&
440 (info->classmetric2[k]<0 || (int)ent<info->classmetric2[k]))
453 sprintf(buffer,"resaux_%d.vqd",look->train_seq);
454 of=fopen(buffer,"a");
455 for(j=0;j<partvals;j++)
456 fprintf(of,"%ld, ",partword[i][j]);
467 /* designed for stereo or other modes where the partition size is an
468 integer multiple of the number of channels encoded in the current
470 static long **_2class(vorbis_block *vb,vorbis_look_residue *vl,float **in,
473 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
474 vorbis_info_residue0 *info=look->info;
476 /* move all this setup out later */
477 int samples_per_partition=info->grouping;
478 int possible_partitions=info->partitions;
479 int n=info->end-info->begin;
481 int partvals=n/samples_per_partition;
482 long **partword=_vorbis_block_alloc(vb,sizeof(*partword));
484 #if defined(TRAIN_RES) || defined (TRAIN_RESAUX)
489 partword[0]=_vorbis_block_alloc(vb,n*ch/samples_per_partition*sizeof(*partword[0]));
490 memset(partword[0],0,n*ch/samples_per_partition*sizeof(*partword[0]));
492 for(i=0,l=info->begin/ch;i<partvals;i++){
495 for(j=0;j<samples_per_partition;j+=ch){
496 if(fabs(in[0][l])>magmax)magmax=fabs(in[0][l]);
498 if(fabs(in[k][l])>angmax)angmax=fabs(in[k][l]);
502 for(j=0;j<possible_partitions-1;j++)
503 if(magmax<=info->classmetric1[j] &&
504 angmax<=info->classmetric2[j])
512 sprintf(buffer,"resaux_%d.vqd",look->train_seq);
513 of=fopen(buffer,"a");
514 for(i=0;i<partvals;i++)
515 fprintf(of,"%ld, ",partword[0][i]);
525 static int _01forward(oggpack_buffer *opb,
526 vorbis_block *vb,vorbis_look_residue *vl,
529 int (*encode)(oggpack_buffer *,float *,int,
532 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
533 vorbis_info_residue0 *info=look->info;
535 /* move all this setup out later */
536 int samples_per_partition=info->grouping;
537 int possible_partitions=info->partitions;
538 int partitions_per_word=look->phrasebook->dim;
539 int n=info->end-info->begin;
541 int partvals=n/samples_per_partition;
547 for(j=info->begin;j<info->end;j++){
548 if(in[i][j]>look->tmax)look->tmax=in[i][j];
549 if(in[i][j]<look->tmin)look->tmin=in[i][j];
553 memset(resbits,0,sizeof(resbits));
554 memset(resvals,0,sizeof(resvals));
556 /* we code the partition words for each channel, then the residual
557 words for a partition per channel until we've written all the
558 residual words for that partition word. Then write the next
559 partition channel words... */
561 for(s=0;s<look->stages;s++){
563 for(i=0;i<partvals;){
565 /* first we encode a partition codeword for each channel */
568 long val=partword[j][i];
569 for(k=1;k<partitions_per_word;k++){
570 val*=possible_partitions;
572 val+=partword[j][i+k];
576 if(val<look->phrasebook->entries)
577 look->phrasebits+=vorbis_book_encode(look->phrasebook,val,opb);
578 #if 0 /*def TRAIN_RES*/
586 /* now we encode interleaved residual values for the partitions */
587 for(k=0;k<partitions_per_word && i<partvals;k++,i++){
588 long offset=i*samples_per_partition+info->begin;
591 if(s==0)resvals[partword[j][i]]+=samples_per_partition;
592 if(info->secondstages[partword[j][i]]&(1<<s)){
593 codebook *statebook=look->partbooks[partword[j][i]][s];
596 long *accumulator=NULL;
599 accumulator=look->training_data[s][partword[j][i]];
602 float *samples=in[j]+offset;
603 for(l=0;l<samples_per_partition;l++){
604 if(samples[l]<look->training_min[s][partword[j][i]])
605 look->training_min[s][partword[j][i]]=samples[l];
606 if(samples[l]>look->training_max[s][partword[j][i]])
607 look->training_max[s][partword[j][i]]=samples[l];
612 ret=encode(opb,in[j]+offset,samples_per_partition,
613 statebook,accumulator);
616 resbits[partword[j][i]]+=ret;
627 fprintf(stderr,"%d :: ",vb->mode);
628 for(k=0;k<possible_partitions;k++){
629 fprintf(stderr,"%ld/%1.2g, ",resvals[k],(float)resbits[k]/resvals[k]);
631 totalbits+=resbits[k];
634 fprintf(stderr,":: %ld:%1.2g\n",total,(double)totalbits/total);
640 /* a truncated packet here just means 'stop working'; it's not an error */
641 static int _01inverse(vorbis_block *vb,vorbis_look_residue *vl,
643 long (*decodepart)(codebook *, float *,
644 oggpack_buffer *,int)){
647 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
648 vorbis_info_residue0 *info=look->info;
650 /* move all this setup out later */
651 int samples_per_partition=info->grouping;
652 int partitions_per_word=look->phrasebook->dim;
653 int max=vb->pcmend>>1;
654 int end=(info->end<max?info->end:max);
655 int n=end-info->begin;
658 int partvals=n/samples_per_partition;
659 int partwords=(partvals+partitions_per_word-1)/partitions_per_word;
660 int ***partword=alloca(ch*sizeof(*partword));
663 partword[j]=_vorbis_block_alloc(vb,partwords*sizeof(*partword[j]));
665 for(s=0;s<look->stages;s++){
667 /* each loop decodes on partition codeword containing
668 partitions_per_word partitions */
669 for(i=0,l=0;i<partvals;l++){
671 /* fetch the partition word for each channel */
673 int temp=vorbis_book_decode(look->phrasebook,&vb->opb);
675 if(temp==-1 || temp>=info->partvals)goto eopbreak;
676 partword[j][l]=look->decodemap[temp];
677 if(partword[j][l]==NULL)goto errout;
681 /* now we decode residual values for the partitions */
682 for(k=0;k<partitions_per_word && i<partvals;k++,i++)
684 long offset=info->begin+i*samples_per_partition;
685 if(info->secondstages[partword[j][l][k]]&(1<<s)){
686 codebook *stagebook=look->partbooks[partword[j][l][k]][s];
688 if(decodepart(stagebook,in[j]+offset,&vb->opb,
689 samples_per_partition)==-1)goto eopbreak;
702 /* residue 0 and 1 are just slight variants of one another. 0 is
703 interleaved, 1 is not */
704 long **res0_class(vorbis_block *vb,vorbis_look_residue *vl,
705 float **in,int *nonzero,int ch){
706 /* we encode only the nonzero parts of a bundle */
712 /*return(_01class(vb,vl,in,used,_interleaved_testhack));*/
713 return(_01class(vb,vl,in,used));
718 int res0_forward(vorbis_block *vb,vorbis_look_residue *vl,
719 float **in,float **out,int *nonzero,int ch,
721 /* we encode only the nonzero parts of a bundle */
722 int i,j,used=0,n=vb->pcmend/2;
731 int ret=_01forward(vb,vl,in,used,partword,
732 _interleaved_encodepart);
738 out[i][j]-=in[used][j];
749 int res0_inverse(vorbis_block *vb,vorbis_look_residue *vl,
750 float **in,int *nonzero,int ch){
756 return(_01inverse(vb,vl,in,used,vorbis_book_decodevs_add));
761 int res1_forward(oggpack_buffer *opb,vorbis_block *vb,vorbis_look_residue *vl,
762 float **in,float **out,int *nonzero,int ch,
764 int i,j,used=0,n=vb->pcmend/2;
774 int ret=_01forward(opb,vb,vl,in,used,partword,_encodepart);
780 out[i][j]-=in[used][j];
790 long **res1_class(vorbis_block *vb,vorbis_look_residue *vl,
791 float **in,int *nonzero,int ch){
797 return(_01class(vb,vl,in,used));
802 int res1_inverse(vorbis_block *vb,vorbis_look_residue *vl,
803 float **in,int *nonzero,int ch){
809 return(_01inverse(vb,vl,in,used,vorbis_book_decodev_add));
814 long **res2_class(vorbis_block *vb,vorbis_look_residue *vl,
815 float **in,int *nonzero,int ch){
818 if(nonzero[i])used++;
820 return(_2class(vb,vl,in,ch));
825 /* res2 is slightly more different; all the channels are interleaved
826 into a single vector and encoded. */
828 int res2_forward(oggpack_buffer *opb,
829 vorbis_block *vb,vorbis_look_residue *vl,
830 float **in,float **out,int *nonzero,int ch,
832 long i,j,k,n=vb->pcmend/2,used=0;
834 /* don't duplicate the code; use a working vector hack for now and
835 reshape ourselves into a single channel res1 */
836 /* ugly; reallocs for each coupling pass :-( */
837 float *work=_vorbis_block_alloc(vb,ch*n*sizeof(*work));
840 if(nonzero[i])used++;
841 for(j=0,k=i;j<n;j++,k+=ch)
846 int ret=_01forward(opb,vb,vl,&work,1,partword,_encodepart);
847 /* update the sofar vector */
852 for(j=0,k=i;j<n;j++,k+=ch)
853 sofar[j]+=pcm[j]-work[k];
863 /* duplicate code here as speed is somewhat more important */
864 int res2_inverse(vorbis_block *vb,vorbis_look_residue *vl,
865 float **in,int *nonzero,int ch){
867 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
868 vorbis_info_residue0 *info=look->info;
870 /* move all this setup out later */
871 int samples_per_partition=info->grouping;
872 int partitions_per_word=look->phrasebook->dim;
873 int max=(vb->pcmend*ch)>>1;
874 int end=(info->end<max?info->end:max);
875 int n=end-info->begin;
878 int partvals=n/samples_per_partition;
879 int partwords=(partvals+partitions_per_word-1)/partitions_per_word;
880 int **partword=_vorbis_block_alloc(vb,partwords*sizeof(*partword));
882 for(i=0;i<ch;i++)if(nonzero[i])break;
883 if(i==ch)return(0); /* no nonzero vectors */
885 for(s=0;s<look->stages;s++){
886 for(i=0,l=0;i<partvals;l++){
889 /* fetch the partition word */
890 int temp=vorbis_book_decode(look->phrasebook,&vb->opb);
891 if(temp==-1 || temp>info->partvals)goto eopbreak;
892 partword[l]=look->decodemap[temp];
893 if(partword[l]==NULL)goto errout;
896 /* now we decode residual values for the partitions */
897 for(k=0;k<partitions_per_word && i<partvals;k++,i++)
898 if(info->secondstages[partword[l][k]]&(1<<s)){
899 codebook *stagebook=look->partbooks[partword[l][k]][s];
902 if(vorbis_book_decodevv_add(stagebook,in,
903 i*samples_per_partition+info->begin,ch,
904 &vb->opb,samples_per_partition)==-1)
917 const vorbis_func_residue residue0_exportbundle={
928 const vorbis_func_residue residue1_exportbundle={
939 const vorbis_func_residue residue2_exportbundle={