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 1994-2002 *
9 * by the XIPHOPHORUS Company 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 for(j=0;j<info->partitions;j++){
212 int cascade=oggpack_read(opb,3);
213 if(oggpack_read(opb,1))
214 cascade|=(oggpack_read(opb,5)<<3);
215 info->secondstages[j]=cascade;
217 acc+=icount(cascade);
220 info->booklist[j]=oggpack_read(opb,8);
222 if(info->groupbook>=ci->books)goto errout;
224 if(info->booklist[j]>=ci->books)goto errout;
228 res0_free_info(info);
232 vorbis_look_residue *res0_look(vorbis_dsp_state *vd,
233 vorbis_info_residue *vr){
234 vorbis_info_residue0 *info=(vorbis_info_residue0 *)vr;
235 vorbis_look_residue0 *look=_ogg_calloc(1,sizeof(*look));
236 codec_setup_info *ci=vd->vi->codec_setup;
243 look->parts=info->partitions;
244 look->fullbooks=ci->fullbooks;
245 look->phrasebook=ci->fullbooks+info->groupbook;
246 dim=look->phrasebook->dim;
248 look->partbooks=_ogg_calloc(look->parts,sizeof(*look->partbooks));
250 for(j=0;j<look->parts;j++){
251 int stages=ilog(info->secondstages[j]);
253 if(stages>maxstage)maxstage=stages;
254 look->partbooks[j]=_ogg_calloc(stages,sizeof(*look->partbooks[j]));
255 for(k=0;k<stages;k++)
256 if(info->secondstages[j]&(1<<k)){
257 look->partbooks[j][k]=ci->fullbooks+info->booklist[acc++];
259 look->training_data[k][j]=_ogg_calloc(look->partbooks[j][k]->entries,
260 sizeof(***look->training_data));
266 look->partvals=rint(pow((float)look->parts,(float)dim));
267 look->stages=maxstage;
268 look->decodemap=_ogg_malloc(look->partvals*sizeof(*look->decodemap));
269 for(j=0;j<look->partvals;j++){
271 long mult=look->partvals/look->parts;
272 look->decodemap[j]=_ogg_malloc(dim*sizeof(*look->decodemap[j]));
277 look->decodemap[j][k]=deco;
280 #if defined(TRAIN_RES) || defined (TRAIN_RESAUX)
282 static int train_seq=0;
283 look->train_seq=train_seq++;
289 /* break an abstraction and copy some code for performance purposes */
290 static int local_book_besterror(codebook *book,float *a){
291 int dim=book->dim,i,k,o;
293 encode_aux_threshmatch *tt=book->c->thresh_tree;
295 /* find the quant val of each scalar */
296 for(k=0,o=dim;k<dim;++k){
300 if(val<tt->quantthresh[i]){
301 if(val<tt->quantthresh[i-1]){
303 if(val>=tt->quantthresh[i-1])
308 for(++i;i<tt->threshvals-1;++i)
309 if(val<tt->quantthresh[i])break;
313 best=(best*tt->quantvals)+tt->quantmap[i];
315 /* regular lattices are easy :-) */
317 if(book->c->lengthlist[best]<=0){
318 const static_codebook *c=book->c;
321 float *e=book->valuelist;
323 for(i=0;i<book->entries;i++){
324 if(c->lengthlist[i]>0){
327 float val=(e[j]-a[j]);
330 if(best==-1 || this<bestf){
340 float *ptr=book->valuelist+best*dim;
348 static int _encodepart(oggpack_buffer *opb,float *vec, int n,
349 codebook *book,long *acc){
355 int entry=local_book_besterror(book,vec+i*dim);
361 bits+=vorbis_book_encode(book,entry,opb);
367 static long **_01class(vorbis_block *vb,vorbis_look_residue *vl,
370 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
371 vorbis_info_residue0 *info=look->info;
372 vorbis_info *vi=vb->vd->vi;
373 codec_setup_info *ci=vi->codec_setup;
375 /* move all this setup out later */
376 int samples_per_partition=info->grouping;
377 int possible_partitions=info->partitions;
378 int end=(info->end<vb->pcmend?info->end:vb->pcmend);
379 int n=end-info->begin;
382 int partvals=n/samples_per_partition;
383 long **partword=_vorbis_block_alloc(vb,ch*sizeof(*partword));
384 float scale=100./samples_per_partition;
386 /* we find the partition type for each partition of each
387 channel. We'll go back and do the interleaved encoding in a
388 bit. For now, clarity */
391 partword[i]=_vorbis_block_alloc(vb,n/samples_per_partition*sizeof(*partword[i]));
392 memset(partword[i],0,n/samples_per_partition*sizeof(*partword[i]));
395 for(i=0;i<partvals;i++){
396 int offset=i*samples_per_partition+info->begin;
400 for(k=0;k<samples_per_partition;k++){
401 if(fabs(in[j][offset+k])>max)max=fabs(in[j][offset+k]);
402 ent+=fabs(rint(in[j][offset+k]));
406 for(k=0;k<possible_partitions-1;k++)
407 if(max<=info->classmetric1[k] &&
408 (info->classmetric2[k]<0 || (int)ent<info->classmetric2[k]))
421 sprintf(buffer,"resaux_%d.vqd",look->train_seq);
422 of=fopen(buffer,"a");
423 for(j=0;j<partvals;j++)
424 fprintf(of,"%ld, ",partword[i][j]);
437 /* designed for stereo or other modes where the partition size is an
438 integer multiple of the number of channels encoded in the current
440 static long **_2class(vorbis_block *vb,vorbis_look_residue *vl,float **in,
443 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
444 vorbis_info_residue0 *info=look->info;
446 /* move all this setup out later */
447 int samples_per_partition=info->grouping;
448 int possible_partitions=info->partitions;
449 int end=(info->end<vb->pcmend?info->end:vb->pcmend);
450 int n=end-info->begin;
453 int partvals=n/samples_per_partition;
454 long **partword=_vorbis_block_alloc(vb,sizeof(*partword));
456 #if defined(TRAIN_RES) || defined (TRAIN_RESAUX)
461 partword[0]=_vorbis_block_alloc(vb,n*ch/samples_per_partition*sizeof(*partword[0]));
462 memset(partword[0],0,n*ch/samples_per_partition*sizeof(*partword[0]));
464 for(i=0,l=info->begin/ch;i<partvals;i++){
467 for(j=0;j<samples_per_partition;j+=ch){
468 if(fabs(in[0][l])>magmax)magmax=fabs(in[0][l]);
470 if(fabs(in[k][l])>angmax)angmax=fabs(in[k][l]);
474 for(j=0;j<possible_partitions-1;j++)
475 if(magmax<=info->classmetric1[j] &&
476 angmax<=info->classmetric2[j])
484 sprintf(buffer,"resaux_%d.vqd",look->train_seq);
485 of=fopen(buffer,"a");
486 for(i=0;i<partvals;i++)
487 fprintf(of,"%ld, ",partword[0][i]);
499 static int _01forward(oggpack_buffer *opb,
500 vorbis_block *vb,vorbis_look_residue *vl,
503 int (*encode)(oggpack_buffer *,float *,int,
506 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
507 vorbis_info_residue0 *info=look->info;
509 vorbis_dsp_state *vd=vb->vd;
511 /* move all this setup out later */
512 int samples_per_partition=info->grouping;
513 int possible_partitions=info->partitions;
514 int partitions_per_word=look->phrasebook->dim;
515 int end=(info->end<vb->pcmend?info->end:vb->pcmend);
516 int n=end-info->begin;
519 int partvals=n/samples_per_partition;
525 for(j=info->begin;j<end;j++){
526 if(in[i][j]>look->tmax)look->tmax=in[i][j];
527 if(in[i][j]<look->tmin)look->tmin=in[i][j];
531 memset(resbits,0,sizeof(resbits));
532 memset(resvals,0,sizeof(resvals));
534 /* we code the partition words for each channel, then the residual
535 words for a partition per channel until we've written all the
536 residual words for that partition word. Then write the next
537 partition channel words... */
539 for(s=0;s<look->stages;s++){
541 for(i=0;i<partvals;){
543 /* first we encode a partition codeword for each channel */
546 long val=partword[j][i];
547 for(k=1;k<partitions_per_word;k++){
548 val*=possible_partitions;
550 val+=partword[j][i+k];
554 if(val<look->phrasebook->entries)
555 look->phrasebits+=vorbis_book_encode(look->phrasebook,val,opb);
556 #if 0 /*def TRAIN_RES*/
564 /* now we encode interleaved residual values for the partitions */
565 for(k=0;k<partitions_per_word && i<partvals;k++,i++){
566 long offset=i*samples_per_partition+info->begin;
569 if(s==0)resvals[partword[j][i]]+=samples_per_partition;
570 if(info->secondstages[partword[j][i]]&(1<<s)){
571 codebook *statebook=look->partbooks[partword[j][i]][s];
574 long *accumulator=NULL;
577 accumulator=look->training_data[s][partword[j][i]];
580 float *samples=in[j]+offset;
581 for(l=0;l<samples_per_partition;l++){
582 if(samples[l]<look->training_min[s][partword[j][i]])
583 look->training_min[s][partword[j][i]]=samples[l];
584 if(samples[l]>look->training_max[s][partword[j][i]])
585 look->training_max[s][partword[j][i]]=samples[l];
590 ret=encode(opb,in[j]+offset,samples_per_partition,
591 statebook,accumulator);
594 resbits[partword[j][i]]+=ret;
605 fprintf(stderr,"%d :: ",vb->mode);
606 for(k=0;k<possible_partitions;k++){
607 fprintf(stderr,"%ld/%1.2g, ",resvals[k],(float)resbits[k]/resvals[k]);
609 totalbits+=resbits[k];
612 fprintf(stderr,":: %ld:%1.2g\n",total,(double)totalbits/total);
618 /* a truncated packet here just means 'stop working'; it's not an error */
619 static int _01inverse(vorbis_block *vb,vorbis_look_residue *vl,
621 long (*decodepart)(codebook *, float *,
622 oggpack_buffer *,int)){
625 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
626 vorbis_info_residue0 *info=look->info;
628 /* move all this setup out later */
629 int samples_per_partition=info->grouping;
630 int partitions_per_word=look->phrasebook->dim;
631 int end=(info->end<vb->pcmend?info->end:vb->pcmend);
632 int n=end-info->begin;
635 int partvals=n/samples_per_partition;
636 int partwords=(partvals+partitions_per_word-1)/partitions_per_word;
637 int ***partword=alloca(ch*sizeof(*partword));
640 partword[j]=_vorbis_block_alloc(vb,partwords*sizeof(*partword[j]));
642 for(s=0;s<look->stages;s++){
644 /* each loop decodes on partition codeword containing
645 partitions_per_word partitions */
646 for(i=0,l=0;i<partvals;l++){
648 /* fetch the partition word for each channel */
650 int temp=vorbis_book_decode(look->phrasebook,&vb->opb);
652 if(temp==-1)goto eopbreak;
653 partword[j][l]=look->decodemap[temp];
654 if(partword[j][l]==NULL)goto errout;
658 /* now we decode residual values for the partitions */
659 for(k=0;k<partitions_per_word && i<partvals;k++,i++)
661 long offset=info->begin+i*samples_per_partition;
662 if(info->secondstages[partword[j][l][k]]&(1<<s)){
663 codebook *stagebook=look->partbooks[partword[j][l][k]][s];
665 if(decodepart(stagebook,in[j]+offset,&vb->opb,
666 samples_per_partition)==-1)goto eopbreak;
679 /* residue 0 and 1 are just slight variants of one another. 0 is
680 interleaved, 1 is not */
681 long **res0_class(vorbis_block *vb,vorbis_look_residue *vl,
682 float **in,int *nonzero,int ch){
683 /* we encode only the nonzero parts of a bundle */
689 /*return(_01class(vb,vl,in,used,_interleaved_testhack));*/
690 return(_01class(vb,vl,in,used));
695 int res0_forward(vorbis_block *vb,vorbis_look_residue *vl,
696 float **in,float **out,int *nonzero,int ch,
698 /* we encode only the nonzero parts of a bundle */
699 int i,j,used=0,n=vb->pcmend/2;
708 int ret=_01forward(vb,vl,in,used,partword,
709 _interleaved_encodepart);
715 out[i][j]-=in[used][j];
726 int res0_inverse(vorbis_block *vb,vorbis_look_residue *vl,
727 float **in,int *nonzero,int ch){
733 return(_01inverse(vb,vl,in,used,vorbis_book_decodevs_add));
738 int res1_forward(oggpack_buffer *opb,vorbis_block *vb,vorbis_look_residue *vl,
739 float **in,float **out,int *nonzero,int ch,
741 int i,j,used=0,n=vb->pcmend/2;
751 int ret=_01forward(opb,vb,vl,in,used,partword,_encodepart);
757 out[i][j]-=in[used][j];
767 long **res1_class(vorbis_block *vb,vorbis_look_residue *vl,
768 float **in,int *nonzero,int ch){
774 return(_01class(vb,vl,in,used));
779 int res1_inverse(vorbis_block *vb,vorbis_look_residue *vl,
780 float **in,int *nonzero,int ch){
786 return(_01inverse(vb,vl,in,used,vorbis_book_decodev_add));
791 long **res2_class(vorbis_block *vb,vorbis_look_residue *vl,
792 float **in,int *nonzero,int ch){
795 if(nonzero[i])used++;
797 return(_2class(vb,vl,in,ch));
802 /* res2 is slightly more different; all the channels are interleaved
803 into a single vector and encoded. */
805 int res2_forward(oggpack_buffer *opb,
806 vorbis_block *vb,vorbis_look_residue *vl,
807 float **in,float **out,int *nonzero,int ch,
809 long i,j,k,n=vb->pcmend/2,used=0;
811 /* don't duplicate the code; use a working vector hack for now and
812 reshape ourselves into a single channel res1 */
813 /* ugly; reallocs for each coupling pass :-( */
814 float *work=_vorbis_block_alloc(vb,ch*n*sizeof(*work));
817 if(nonzero[i])used++;
818 for(j=0,k=i;j<n;j++,k+=ch)
823 int ret=_01forward(opb,vb,vl,&work,1,partword,_encodepart);
824 /* update the sofar vector */
829 for(j=0,k=i;j<n;j++,k+=ch)
830 sofar[j]+=pcm[j]-work[k];
840 /* duplicate code here as speed is somewhat more important */
841 int res2_inverse(vorbis_block *vb,vorbis_look_residue *vl,
842 float **in,int *nonzero,int ch){
844 vorbis_look_residue0 *look=(vorbis_look_residue0 *)vl;
845 vorbis_info_residue0 *info=look->info;
847 /* move all this setup out later */
848 int samples_per_partition=info->grouping;
849 int partitions_per_word=look->phrasebook->dim;
850 int end=(info->end<vb->pcmend?info->end:vb->pcmend);
851 int n=end-info->begin;
854 int partvals=n/samples_per_partition;
855 int partwords=(partvals+partitions_per_word-1)/partitions_per_word;
856 int **partword=_vorbis_block_alloc(vb,partwords*sizeof(*partword));
859 for(i=0;i<ch;i++)if(nonzero[i])break;
860 if(i==ch)return(0); /* no nonzero vectors */
862 for(s=0;s<look->stages;s++){
863 for(i=0,l=0;i<partvals;l++){
866 /* fetch the partition word */
867 int temp=vorbis_book_decode(look->phrasebook,&vb->opb);
868 if(temp==-1)goto eopbreak;
869 partword[l]=look->decodemap[temp];
870 if(partword[l]==NULL)goto errout;
873 /* now we decode residual values for the partitions */
874 for(k=0;k<partitions_per_word && i<partvals;k++,i++)
875 if(info->secondstages[partword[l][k]]&(1<<s)){
876 codebook *stagebook=look->partbooks[partword[l][k]][s];
879 if(vorbis_book_decodevv_add(stagebook,in,
880 i*samples_per_partition+info->begin,ch,
881 &vb->opb,samples_per_partition)==-1)
894 vorbis_func_residue residue0_exportbundle={
905 vorbis_func_residue residue1_exportbundle={
916 vorbis_func_residue residue2_exportbundle={