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: basic shared codebook operations
16 ********************************************************************/
24 #include "vorbis/codec.h"
28 /**** pack/unpack helpers ******************************************/
29 int _ilog(unsigned int v){
38 /* 32 bit float (not IEEE; nonnormalized mantissa +
39 biased exponent) : neeeeeee eeemmmmm mmmmmmmm mmmmmmmm
40 Why not IEEE? It's just not that important here. */
44 #define VQ_FEXP_BIAS 768 /* bias toward values smaller than 1. */
46 /* doesn't currently guard under/overflow */
47 long _float32_pack(float val){
55 exp= floor(log(val)/log(2.f)+.001); //+epsilon
56 mant=rint(ldexp(val,(VQ_FMAN-1)-exp));
57 exp=(exp+VQ_FEXP_BIAS)<<VQ_FMAN;
59 return(sign|exp|mant);
62 float _float32_unpack(long val){
63 double mant=val&0x1fffff;
64 int sign=val&0x80000000;
65 long exp =(val&0x7fe00000L)>>VQ_FMAN;
67 return(ldexp(mant,exp-(VQ_FMAN-1)-VQ_FEXP_BIAS));
70 /* given a list of word lengths, generate a list of codewords. Works
71 for length ordered or unordered, always assigns the lowest valued
72 codewords first. Extended to handle unused entries (length 0) */
73 ogg_uint32_t *_make_words(long *l,long n,long sparsecount){
75 ogg_uint32_t marker[33];
76 ogg_uint32_t *r=_ogg_malloc((sparsecount?sparsecount:n)*sizeof(*r));
77 memset(marker,0,sizeof(marker));
82 ogg_uint32_t entry=marker[length];
84 /* when we claim a node for an entry, we also claim the nodes
85 below it (pruning off the imagined tree that may have dangled
86 from it) as well as blocking the use of any nodes directly
90 if(length<32 && (entry>>length)){
91 /* error condition; the lengths must specify an overpopulated tree */
97 /* Look to see if the next shorter marker points to the node
98 above. if so, update it and repeat. */
100 for(j=length;j>0;j--){
103 /* have to jump branches */
107 marker[j]=marker[j-1]<<1;
108 break; /* invariant says next upper marker would already
109 have been moved if it was on the same path */
115 /* prune the tree; the implicit invariant says all the longer
116 markers were dangling from our just-taken node. Dangle them
117 from our *new* node. */
118 for(j=length+1;j<33;j++)
119 if((marker[j]>>1) == entry){
121 marker[j]=marker[j-1]<<1;
125 if(sparsecount==0)count++;
128 /* sanity check the huffman tree; an underpopulated tree must be
129 rejected. The only exception is the one-node pseudo-nil tree,
130 which appears to be underpopulated because the tree doesn't
131 really exist; there's only one possible 'codeword' or zero bits,
132 but the above tree-gen code doesn't mark that. */
133 if(sparsecount != 1){
135 if(marker[i] & (0xffffffffUL>>(32-i))){
141 /* bitreverse the words because our bitwise packer/unpacker is LSb
143 for(i=0,count=0;i<n;i++){
147 temp|=(r[count]>>j)&1;
160 /* there might be a straightforward one-line way to do the below
161 that's portable and totally safe against roundoff, but I haven't
162 thought of it. Therefore, we opt on the side of caution */
163 long _book_maptype1_quantvals(const static_codebook *b){
164 long vals=floor(pow((float)b->entries,1.f/b->dim));
166 /* the above *should* be reliable, but we'll not assume that FP is
167 ever reliable when bitstream sync is at stake; verify via integer
168 means that vals really is the greatest value of dim for which
169 vals^b->bim <= b->entries */
170 /* treat the above as an initial guess */
175 for(i=0;i<b->dim;i++){
179 if(acc<=b->entries && acc1>b->entries){
191 /* unpack the quantized list of values for encode/decode ***********/
192 /* we need to deal with two map types: in map type 1, the values are
193 generated algorithmically (each column of the vector counts through
194 the values in the quant vector). in map type 2, all the values came
195 in in an explicit list. Both value lists must be unpacked */
196 float *_book_unquantize(const static_codebook *b,int n,int *sparsemap){
198 if(b->maptype==1 || b->maptype==2){
200 float mindel=_float32_unpack(b->q_min);
201 float delta=_float32_unpack(b->q_delta);
202 float *r=_ogg_calloc(n*b->dim,sizeof(*r));
204 /* maptype 1 and 2 both use a quantized value vector, but
208 /* most of the time, entries%dimensions == 0, but we need to be
209 well defined. We define that the possible vales at each
210 scalar is values == entries/dim. If entries%dim != 0, we'll
211 have 'too few' values (values*dim<entries), which means that
212 we'll have 'left over' entries; left over entries use zeroed
213 values (and are wasted). So don't generate codebooks like
215 quantvals=_book_maptype1_quantvals(b);
216 for(j=0;j<b->entries;j++){
217 if((sparsemap && b->lengthlist[j]) || !sparsemap){
220 for(k=0;k<b->dim;k++){
221 int index= (j/indexdiv)%quantvals;
222 float val=b->quantlist[index];
223 val=fabs(val)*delta+mindel+last;
224 if(b->q_sequencep)last=val;
226 r[sparsemap[count]*b->dim+k]=val;
228 r[count*b->dim+k]=val;
237 for(j=0;j<b->entries;j++){
238 if((sparsemap && b->lengthlist[j]) || !sparsemap){
241 for(k=0;k<b->dim;k++){
242 float val=b->quantlist[j*b->dim+k];
243 val=fabs(val)*delta+mindel+last;
244 if(b->q_sequencep)last=val;
246 r[sparsemap[count]*b->dim+k]=val;
248 r[count*b->dim+k]=val;
261 void vorbis_staticbook_clear(static_codebook *b){
263 if(b->quantlist)_ogg_free(b->quantlist);
264 if(b->lengthlist)_ogg_free(b->lengthlist);
266 _ogg_free(b->nearest_tree->ptr0);
267 _ogg_free(b->nearest_tree->ptr1);
268 _ogg_free(b->nearest_tree->p);
269 _ogg_free(b->nearest_tree->q);
270 memset(b->nearest_tree,0,sizeof(*b->nearest_tree));
271 _ogg_free(b->nearest_tree);
274 _ogg_free(b->thresh_tree->quantthresh);
275 _ogg_free(b->thresh_tree->quantmap);
276 memset(b->thresh_tree,0,sizeof(*b->thresh_tree));
277 _ogg_free(b->thresh_tree);
280 memset(b,0,sizeof(*b));
284 void vorbis_staticbook_destroy(static_codebook *b){
286 vorbis_staticbook_clear(b);
291 void vorbis_book_clear(codebook *b){
292 /* static book is not cleared; we're likely called on the lookup and
293 the static codebook belongs to the info struct */
294 if(b->valuelist)_ogg_free(b->valuelist);
295 if(b->codelist)_ogg_free(b->codelist);
297 if(b->dec_index)_ogg_free(b->dec_index);
298 if(b->dec_codelengths)_ogg_free(b->dec_codelengths);
299 if(b->dec_firsttable)_ogg_free(b->dec_firsttable);
301 memset(b,0,sizeof(*b));
304 int vorbis_book_init_encode(codebook *c,const static_codebook *s){
306 memset(c,0,sizeof(*c));
308 c->entries=s->entries;
309 c->used_entries=s->entries;
311 c->codelist=_make_words(s->lengthlist,s->entries,0);
312 c->valuelist=_book_unquantize(s,s->entries,NULL);
317 static ogg_uint32_t bitreverse(ogg_uint32_t x){
318 x= ((x>>16)&0x0000ffffUL) | ((x<<16)&0xffff0000UL);
319 x= ((x>> 8)&0x00ff00ffUL) | ((x<< 8)&0xff00ff00UL);
320 x= ((x>> 4)&0x0f0f0f0fUL) | ((x<< 4)&0xf0f0f0f0UL);
321 x= ((x>> 2)&0x33333333UL) | ((x<< 2)&0xccccccccUL);
322 return((x>> 1)&0x55555555UL) | ((x<< 1)&0xaaaaaaaaUL);
325 static int sort32a(const void *a,const void *b){
326 return ( **(ogg_uint32_t **)a>**(ogg_uint32_t **)b)-
327 ( **(ogg_uint32_t **)a<**(ogg_uint32_t **)b);
330 /* decode codebook arrangement is more heavily optimized than encode */
331 int vorbis_book_init_decode(codebook *c,const static_codebook *s){
334 memset(c,0,sizeof(*c));
336 /* count actually used entries */
337 for(i=0;i<s->entries;i++)
338 if(s->lengthlist[i]>0)
341 c->entries=s->entries;
347 /* two different remappings go on here.
349 First, we collapse the likely sparse codebook down only to
350 actually represented values/words. This collapsing needs to be
351 indexed as map-valueless books are used to encode original entry
352 positions as integers.
354 Second, we reorder all vectors, including the entry index above,
355 by sorted bitreversed codeword to allow treeless decode. */
358 ogg_uint32_t *codes=_make_words(s->lengthlist,s->entries,c->used_entries);
359 ogg_uint32_t **codep=alloca(sizeof(*codep)*n);
361 if(codes==NULL)goto err_out;
364 codes[i]=bitreverse(codes[i]);
368 qsort(codep,n,sizeof(*codep),sort32a);
370 sortindex=alloca(n*sizeof(*sortindex));
371 c->codelist=_ogg_malloc(n*sizeof(*c->codelist));
372 /* the index is a reverse index */
374 int position=codep[i]-codes;
375 sortindex[position]=i;
379 c->codelist[sortindex[i]]=codes[i];
383 c->valuelist=_book_unquantize(s,n,sortindex);
384 c->dec_index=_ogg_malloc(n*sizeof(*c->dec_index));
386 for(n=0,i=0;i<s->entries;i++)
387 if(s->lengthlist[i]>0)
388 c->dec_index[sortindex[n++]]=i;
390 c->dec_codelengths=_ogg_malloc(n*sizeof(*c->dec_codelengths));
391 for(n=0,i=0;i<s->entries;i++)
392 if(s->lengthlist[i]>0)
393 c->dec_codelengths[sortindex[n++]]=s->lengthlist[i];
395 c->dec_firsttablen=_ilog(c->used_entries)-4; /* this is magic */
396 if(c->dec_firsttablen<5)c->dec_firsttablen=5;
397 if(c->dec_firsttablen>8)c->dec_firsttablen=8;
399 tabn=1<<c->dec_firsttablen;
400 c->dec_firsttable=_ogg_calloc(tabn,sizeof(*c->dec_firsttable));
404 if(c->dec_maxlength<c->dec_codelengths[i])
405 c->dec_maxlength=c->dec_codelengths[i];
406 if(c->dec_codelengths[i]<=c->dec_firsttablen){
407 ogg_uint32_t orig=bitreverse(c->codelist[i]);
408 for(j=0;j<(1<<(c->dec_firsttablen-c->dec_codelengths[i]));j++)
409 c->dec_firsttable[orig|(j<<c->dec_codelengths[i])]=i+1;
413 /* now fill in 'unused' entries in the firsttable with hi/lo search
414 hints for the non-direct-hits */
416 ogg_uint32_t mask=0xfffffffeUL<<(31-c->dec_firsttablen);
420 ogg_uint32_t word=i<<(32-c->dec_firsttablen);
421 if(c->dec_firsttable[bitreverse(word)]==0){
422 while((lo+1)<n && c->codelist[lo+1]<=word)lo++;
423 while( hi<n && word>=(c->codelist[hi]&mask))hi++;
425 /* we only actually have 15 bits per hint to play with here.
426 In order to overflow gracefully (nothing breaks, efficiency
427 just drops), encode as the difference from the extremes. */
429 unsigned long loval=lo;
430 unsigned long hival=n-hi;
432 if(loval>0x7fff)loval=0x7fff;
433 if(hival>0x7fff)hival=0x7fff;
434 c->dec_firsttable[bitreverse(word)]=
435 0x80000000UL | (loval<<15) | hival;
444 vorbis_book_clear(c);
448 static float _dist(int el,float *ref, float *b,int step){
452 float val=(ref[i]-b[i*step]);
458 int _best(codebook *book, float *a, int step){
459 encode_aux_threshmatch *tt=book->c->thresh_tree;
462 encode_aux_nearestmatch *nt=book->c->nearest_tree;
463 encode_aux_pigeonhole *pt=book->c->pigeon_tree;
470 /* do we have a threshhold encode hint? */
473 /* find the quant val of each scalar */
474 for(k=0,o=step*(dim-1);k<dim;k++,o-=step){
477 if(a[o]<tt->quantthresh[i]){
480 if(a[o]>=tt->quantthresh[i-1])
485 for(i++;i<tt->threshvals-1;i++)
486 if(a[o]<tt->quantthresh[i])break;
490 index=(index*tt->quantvals)+tt->quantmap[i];
492 /* regular lattices are easy :-) */
493 if(book->c->lengthlist[index]>0) /* is this unused? If so, we'll
494 use a decision tree after all
500 /* do we have a pigeonhole encode hint? */
502 const static_codebook *c=book->c;
507 /* dealing with sequentialness is a pain in the ass */
512 for(k=0,o=0;k<dim;k++,o+=step){
513 pv=(int)((a[o]-qlast-pt->min)/pt->del);
514 if(pv<0 || pv>=pt->mapentries)break;
515 entry+=pt->pigeonmap[pv]*mul;
517 qlast+=pv*pt->del+pt->min;
520 for(k=0,o=step*(dim-1);k<dim;k++,o-=step){
521 int pv=(int)((a[o]-pt->min)/pt->del);
522 if(pv<0 || pv>=pt->mapentries)break;
523 entry=entry*pt->quantvals+pt->pigeonmap[pv];
527 /* must be within the pigeonholable range; if we quant outside (or
528 in an entry that we define no list for), brute force it */
529 if(k==dim && pt->fitlength[entry]){
530 /* search the abbreviated list */
531 long *list=pt->fitlist+pt->fitmap[entry];
532 for(i=0;i<pt->fitlength[entry];i++){
533 float this=_dist(dim,book->valuelist+list[i]*dim,a,step);
534 if(besti==-1 || this<best){
545 /* optimized using the decision tree */
548 float *p=book->valuelist+nt->p[ptr];
549 float *q=book->valuelist+nt->q[ptr];
551 for(k=0,o=0;k<dim;k++,o+=step)
552 c+=(p[k]-q[k])*(a[o]-(p[k]+q[k])*.5);
564 /* brute force it! */
566 const static_codebook *c=book->c;
569 float *e=book->valuelist;
570 for(i=0;i<book->entries;i++){
571 if(c->lengthlist[i]>0){
572 float this=_dist(dim,e,a,step);
573 if(besti==-1 || this<best){
581 /*if(savebest!=-1 && savebest!=besti){
582 fprintf(stderr,"brute force/pigeonhole disagreement:\n"
584 for(i=0;i<dim*step;i+=step)fprintf(stderr,"%g,",a[i]);
586 "pigeonhole (entry %d, err %g):",savebest,saverr);
587 for(i=0;i<dim;i++)fprintf(stderr,"%g,",
588 (book->valuelist+savebest*dim)[i]);
590 "bruteforce (entry %d, err %g):",besti,best);
591 for(i=0;i<dim;i++)fprintf(stderr,"%g,",
592 (book->valuelist+besti*dim)[i]);
593 fprintf(stderr,"\n");
599 long vorbis_book_codeword(codebook *book,int entry){
600 if(book->c) /* only use with encode; decode optimizations are
601 allowed to break this */
602 return book->codelist[entry];
606 long vorbis_book_codelen(codebook *book,int entry){
607 if(book->c) /* only use with encode; decode optimizations are
608 allowed to break this */
609 return book->c->lengthlist[entry];
615 /* Unit tests of the dequantizer; this stuff will be OK
616 cross-platform, I simply want to be sure that special mapping cases
617 actually work properly; a bug could go unnoticed for a while */
624 full, explicit mapping
631 static long full_quantlist1[]={0,1,2,3, 4,5,6,7, 8,3,6,1};
632 static long partial_quantlist1[]={0,7,2};
635 static_codebook test1={
644 static float *test1_result=NULL;
646 /* linear, full mapping, nonsequential */
647 static_codebook test2={
651 -533200896,1611661312,4,0,
656 static float test2_result[]={-3,-2,-1,0, 1,2,3,4, 5,0,3,-2};
658 /* linear, full mapping, sequential */
659 static_codebook test3={
663 -533200896,1611661312,4,1,
668 static float test3_result[]={-3,-5,-6,-6, 1,3,6,10, 5,5,8,6};
670 /* linear, algorithmic mapping, nonsequential */
671 static_codebook test4={
675 -533200896,1611661312,4,0,
680 static float test4_result[]={-3,-3,-3, 4,-3,-3, -1,-3,-3,
681 -3, 4,-3, 4, 4,-3, -1, 4,-3,
682 -3,-1,-3, 4,-1,-3, -1,-1,-3,
683 -3,-3, 4, 4,-3, 4, -1,-3, 4,
684 -3, 4, 4, 4, 4, 4, -1, 4, 4,
685 -3,-1, 4, 4,-1, 4, -1,-1, 4,
686 -3,-3,-1, 4,-3,-1, -1,-3,-1,
687 -3, 4,-1, 4, 4,-1, -1, 4,-1,
688 -3,-1,-1, 4,-1,-1, -1,-1,-1};
690 /* linear, algorithmic mapping, sequential */
691 static_codebook test5={
695 -533200896,1611661312,4,1,
700 static float test5_result[]={-3,-6,-9, 4, 1,-2, -1,-4,-7,
701 -3, 1,-2, 4, 8, 5, -1, 3, 0,
702 -3,-4,-7, 4, 3, 0, -1,-2,-5,
703 -3,-6,-2, 4, 1, 5, -1,-4, 0,
704 -3, 1, 5, 4, 8,12, -1, 3, 7,
705 -3,-4, 0, 4, 3, 7, -1,-2, 2,
706 -3,-6,-7, 4, 1, 0, -1,-4,-5,
707 -3, 1, 0, 4, 8, 7, -1, 3, 2,
708 -3,-4,-5, 4, 3, 2, -1,-2,-3};
710 void run_test(static_codebook *b,float *comp){
711 float *out=_book_unquantize(b,b->entries,NULL);
716 fprintf(stderr,"_book_unquantize incorrectly returned NULL\n");
720 for(i=0;i<b->entries*b->dim;i++)
721 if(fabs(out[i]-comp[i])>.0001){
722 fprintf(stderr,"disagreement in unquantized and reference data:\n"
723 "position %d, %g != %g\n",i,out[i],comp[i]);
729 fprintf(stderr,"_book_unquantize returned a value array: \n"
730 " correct result should have been NULL\n");
737 /* run the nine dequant tests, and compare to the hand-rolled results */
738 fprintf(stderr,"Dequant test 1... ");
739 run_test(&test1,test1_result);
740 fprintf(stderr,"OK\nDequant test 2... ");
741 run_test(&test2,test2_result);
742 fprintf(stderr,"OK\nDequant test 3... ");
743 run_test(&test3,test3_result);
744 fprintf(stderr,"OK\nDequant test 4... ");
745 run_test(&test4,test4_result);
746 fprintf(stderr,"OK\nDequant test 5... ");
747 run_test(&test5,test5_result);
748 fprintf(stderr,"OK\n\n");