michael@0: /******************************************************************** michael@0: * * michael@0: * THIS FILE IS PART OF THE OggVorbis SOFTWARE CODEC SOURCE CODE. * michael@0: * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS * michael@0: * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE * michael@0: * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. * michael@0: * * michael@0: * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2009 * michael@0: * by the Xiph.Org Foundation http://www.xiph.org/ * michael@0: * * michael@0: ******************************************************************** michael@0: michael@0: function: basic shared codebook operations michael@0: last mod: $Id: sharedbook.c 19057 2014-01-22 12:32:31Z xiphmont $ michael@0: michael@0: ********************************************************************/ michael@0: michael@0: #include michael@0: #include michael@0: #include michael@0: #include michael@0: #include "os.h" michael@0: #include "misc.h" michael@0: #include "vorbis/codec.h" michael@0: #include "codebook.h" michael@0: #include "scales.h" michael@0: michael@0: /**** pack/unpack helpers ******************************************/ michael@0: int _ilog(unsigned int v){ michael@0: int ret=0; michael@0: while(v){ michael@0: ret++; michael@0: v>>=1; michael@0: } michael@0: return(ret); michael@0: } michael@0: michael@0: /* 32 bit float (not IEEE; nonnormalized mantissa + michael@0: biased exponent) : neeeeeee eeemmmmm mmmmmmmm mmmmmmmm michael@0: Why not IEEE? It's just not that important here. */ michael@0: michael@0: #define VQ_FEXP 10 michael@0: #define VQ_FMAN 21 michael@0: #define VQ_FEXP_BIAS 768 /* bias toward values smaller than 1. */ michael@0: michael@0: /* doesn't currently guard under/overflow */ michael@0: long _float32_pack(float val){ michael@0: int sign=0; michael@0: long exp; michael@0: long mant; michael@0: if(val<0){ michael@0: sign=0x80000000; michael@0: val= -val; michael@0: } michael@0: exp= floor(log(val)/log(2.f)+.001); //+epsilon michael@0: mant=rint(ldexp(val,(VQ_FMAN-1)-exp)); michael@0: exp=(exp+VQ_FEXP_BIAS)<>VQ_FMAN; michael@0: if(sign)mant= -mant; michael@0: return(ldexp(mant,exp-(VQ_FMAN-1)-VQ_FEXP_BIAS)); michael@0: } michael@0: michael@0: /* given a list of word lengths, generate a list of codewords. Works michael@0: for length ordered or unordered, always assigns the lowest valued michael@0: codewords first. Extended to handle unused entries (length 0) */ michael@0: ogg_uint32_t *_make_words(char *l,long n,long sparsecount){ michael@0: long i,j,count=0; michael@0: ogg_uint32_t marker[33]; michael@0: ogg_uint32_t *r=_ogg_malloc((sparsecount?sparsecount:n)*sizeof(*r)); michael@0: memset(marker,0,sizeof(marker)); michael@0: michael@0: for(i=0;i0){ michael@0: ogg_uint32_t entry=marker[length]; michael@0: michael@0: /* when we claim a node for an entry, we also claim the nodes michael@0: below it (pruning off the imagined tree that may have dangled michael@0: from it) as well as blocking the use of any nodes directly michael@0: above for leaves */ michael@0: michael@0: /* update ourself */ michael@0: if(length<32 && (entry>>length)){ michael@0: /* error condition; the lengths must specify an overpopulated tree */ michael@0: _ogg_free(r); michael@0: return(NULL); michael@0: } michael@0: r[count++]=entry; michael@0: michael@0: /* Look to see if the next shorter marker points to the node michael@0: above. if so, update it and repeat. */ michael@0: { michael@0: for(j=length;j>0;j--){ michael@0: michael@0: if(marker[j]&1){ michael@0: /* have to jump branches */ michael@0: if(j==1) michael@0: marker[1]++; michael@0: else michael@0: marker[j]=marker[j-1]<<1; michael@0: break; /* invariant says next upper marker would already michael@0: have been moved if it was on the same path */ michael@0: } michael@0: marker[j]++; michael@0: } michael@0: } michael@0: michael@0: /* prune the tree; the implicit invariant says all the longer michael@0: markers were dangling from our just-taken node. Dangle them michael@0: from our *new* node. */ michael@0: for(j=length+1;j<33;j++) michael@0: if((marker[j]>>1) == entry){ michael@0: entry=marker[j]; michael@0: marker[j]=marker[j-1]<<1; michael@0: }else michael@0: break; michael@0: }else michael@0: if(sparsecount==0)count++; michael@0: } michael@0: michael@0: /* sanity check the huffman tree; an underpopulated tree must be michael@0: rejected. The only exception is the one-node pseudo-nil tree, michael@0: which appears to be underpopulated because the tree doesn't michael@0: really exist; there's only one possible 'codeword' or zero bits, michael@0: but the above tree-gen code doesn't mark that. */ michael@0: if(sparsecount != 1){ michael@0: for(i=1;i<33;i++) michael@0: if(marker[i] & (0xffffffffUL>>(32-i))){ michael@0: _ogg_free(r); michael@0: return(NULL); michael@0: } michael@0: } michael@0: michael@0: /* bitreverse the words because our bitwise packer/unpacker is LSb michael@0: endian */ michael@0: for(i=0,count=0;i>j)&1; michael@0: } michael@0: michael@0: if(sparsecount){ michael@0: if(l[i]) michael@0: r[count++]=temp; michael@0: }else michael@0: r[count++]=temp; michael@0: } michael@0: michael@0: return(r); michael@0: } michael@0: michael@0: /* there might be a straightforward one-line way to do the below michael@0: that's portable and totally safe against roundoff, but I haven't michael@0: thought of it. Therefore, we opt on the side of caution */ michael@0: long _book_maptype1_quantvals(const static_codebook *b){ michael@0: long vals=floor(pow((float)b->entries,1.f/b->dim)); michael@0: michael@0: /* the above *should* be reliable, but we'll not assume that FP is michael@0: ever reliable when bitstream sync is at stake; verify via integer michael@0: means that vals really is the greatest value of dim for which michael@0: vals^b->bim <= b->entries */ michael@0: /* treat the above as an initial guess */ michael@0: while(1){ michael@0: long acc=1; michael@0: long acc1=1; michael@0: int i; michael@0: for(i=0;idim;i++){ michael@0: acc*=vals; michael@0: acc1*=vals+1; michael@0: } michael@0: if(acc<=b->entries && acc1>b->entries){ michael@0: return(vals); michael@0: }else{ michael@0: if(acc>b->entries){ michael@0: vals--; michael@0: }else{ michael@0: vals++; michael@0: } michael@0: } michael@0: } michael@0: } michael@0: michael@0: /* unpack the quantized list of values for encode/decode ***********/ michael@0: /* we need to deal with two map types: in map type 1, the values are michael@0: generated algorithmically (each column of the vector counts through michael@0: the values in the quant vector). in map type 2, all the values came michael@0: in in an explicit list. Both value lists must be unpacked */ michael@0: float *_book_unquantize(const static_codebook *b,int n,int *sparsemap){ michael@0: long j,k,count=0; michael@0: if(b->maptype==1 || b->maptype==2){ michael@0: int quantvals; michael@0: float mindel=_float32_unpack(b->q_min); michael@0: float delta=_float32_unpack(b->q_delta); michael@0: float *r=_ogg_calloc(n*b->dim,sizeof(*r)); michael@0: michael@0: /* maptype 1 and 2 both use a quantized value vector, but michael@0: different sizes */ michael@0: switch(b->maptype){ michael@0: case 1: michael@0: /* most of the time, entries%dimensions == 0, but we need to be michael@0: well defined. We define that the possible vales at each michael@0: scalar is values == entries/dim. If entries%dim != 0, we'll michael@0: have 'too few' values (values*dimentries;j++){ michael@0: if((sparsemap && b->lengthlist[j]) || !sparsemap){ michael@0: float last=0.f; michael@0: int indexdiv=1; michael@0: for(k=0;kdim;k++){ michael@0: int index= (j/indexdiv)%quantvals; michael@0: float val=b->quantlist[index]; michael@0: val=fabs(val)*delta+mindel+last; michael@0: if(b->q_sequencep)last=val; michael@0: if(sparsemap) michael@0: r[sparsemap[count]*b->dim+k]=val; michael@0: else michael@0: r[count*b->dim+k]=val; michael@0: indexdiv*=quantvals; michael@0: } michael@0: count++; michael@0: } michael@0: michael@0: } michael@0: break; michael@0: case 2: michael@0: for(j=0;jentries;j++){ michael@0: if((sparsemap && b->lengthlist[j]) || !sparsemap){ michael@0: float last=0.f; michael@0: michael@0: for(k=0;kdim;k++){ michael@0: float val=b->quantlist[j*b->dim+k]; michael@0: val=fabs(val)*delta+mindel+last; michael@0: if(b->q_sequencep)last=val; michael@0: if(sparsemap) michael@0: r[sparsemap[count]*b->dim+k]=val; michael@0: else michael@0: r[count*b->dim+k]=val; michael@0: } michael@0: count++; michael@0: } michael@0: } michael@0: break; michael@0: } michael@0: michael@0: return(r); michael@0: } michael@0: return(NULL); michael@0: } michael@0: michael@0: void vorbis_staticbook_destroy(static_codebook *b){ michael@0: if(b->allocedp){ michael@0: if(b->quantlist)_ogg_free(b->quantlist); michael@0: if(b->lengthlist)_ogg_free(b->lengthlist); michael@0: memset(b,0,sizeof(*b)); michael@0: _ogg_free(b); michael@0: } /* otherwise, it is in static memory */ michael@0: } michael@0: michael@0: void vorbis_book_clear(codebook *b){ michael@0: /* static book is not cleared; we're likely called on the lookup and michael@0: the static codebook belongs to the info struct */ michael@0: if(b->valuelist)_ogg_free(b->valuelist); michael@0: if(b->codelist)_ogg_free(b->codelist); michael@0: michael@0: if(b->dec_index)_ogg_free(b->dec_index); michael@0: if(b->dec_codelengths)_ogg_free(b->dec_codelengths); michael@0: if(b->dec_firsttable)_ogg_free(b->dec_firsttable); michael@0: michael@0: memset(b,0,sizeof(*b)); michael@0: } michael@0: michael@0: int vorbis_book_init_encode(codebook *c,const static_codebook *s){ michael@0: michael@0: memset(c,0,sizeof(*c)); michael@0: c->c=s; michael@0: c->entries=s->entries; michael@0: c->used_entries=s->entries; michael@0: c->dim=s->dim; michael@0: c->codelist=_make_words(s->lengthlist,s->entries,0); michael@0: //c->valuelist=_book_unquantize(s,s->entries,NULL); michael@0: c->quantvals=_book_maptype1_quantvals(s); michael@0: c->minval=(int)rint(_float32_unpack(s->q_min)); michael@0: c->delta=(int)rint(_float32_unpack(s->q_delta)); michael@0: michael@0: return(0); michael@0: } michael@0: michael@0: static ogg_uint32_t bitreverse(ogg_uint32_t x){ michael@0: x= ((x>>16)&0x0000ffffUL) | ((x<<16)&0xffff0000UL); michael@0: x= ((x>> 8)&0x00ff00ffUL) | ((x<< 8)&0xff00ff00UL); michael@0: x= ((x>> 4)&0x0f0f0f0fUL) | ((x<< 4)&0xf0f0f0f0UL); michael@0: x= ((x>> 2)&0x33333333UL) | ((x<< 2)&0xccccccccUL); michael@0: return((x>> 1)&0x55555555UL) | ((x<< 1)&0xaaaaaaaaUL); michael@0: } michael@0: michael@0: static int sort32a(const void *a,const void *b){ michael@0: return ( **(ogg_uint32_t **)a>**(ogg_uint32_t **)b)- michael@0: ( **(ogg_uint32_t **)a<**(ogg_uint32_t **)b); michael@0: } michael@0: michael@0: /* decode codebook arrangement is more heavily optimized than encode */ michael@0: int vorbis_book_init_decode(codebook *c,const static_codebook *s){ michael@0: int i,j,n=0,tabn; michael@0: int *sortindex; michael@0: memset(c,0,sizeof(*c)); michael@0: michael@0: /* count actually used entries */ michael@0: for(i=0;ientries;i++) michael@0: if(s->lengthlist[i]>0) michael@0: n++; michael@0: michael@0: c->entries=s->entries; michael@0: c->used_entries=n; michael@0: c->dim=s->dim; michael@0: michael@0: if(n>0){ michael@0: michael@0: /* two different remappings go on here. michael@0: michael@0: First, we collapse the likely sparse codebook down only to michael@0: actually represented values/words. This collapsing needs to be michael@0: indexed as map-valueless books are used to encode original entry michael@0: positions as integers. michael@0: michael@0: Second, we reorder all vectors, including the entry index above, michael@0: by sorted bitreversed codeword to allow treeless decode. */ michael@0: michael@0: /* perform sort */ michael@0: ogg_uint32_t *codes=_make_words(s->lengthlist,s->entries,c->used_entries); michael@0: ogg_uint32_t **codep=alloca(sizeof(*codep)*n); michael@0: michael@0: if(codes==NULL)goto err_out; michael@0: michael@0: for(i=0;icodelist=_ogg_malloc(n*sizeof(*c->codelist)); michael@0: /* the index is a reverse index */ michael@0: for(i=0;icodelist[sortindex[i]]=codes[i]; michael@0: _ogg_free(codes); michael@0: michael@0: michael@0: c->valuelist=_book_unquantize(s,n,sortindex); michael@0: c->dec_index=_ogg_malloc(n*sizeof(*c->dec_index)); michael@0: michael@0: for(n=0,i=0;ientries;i++) michael@0: if(s->lengthlist[i]>0) michael@0: c->dec_index[sortindex[n++]]=i; michael@0: michael@0: c->dec_codelengths=_ogg_malloc(n*sizeof(*c->dec_codelengths)); michael@0: for(n=0,i=0;ientries;i++) michael@0: if(s->lengthlist[i]>0) michael@0: c->dec_codelengths[sortindex[n++]]=s->lengthlist[i]; michael@0: michael@0: c->dec_firsttablen=_ilog(c->used_entries)-4; /* this is magic */ michael@0: if(c->dec_firsttablen<5)c->dec_firsttablen=5; michael@0: if(c->dec_firsttablen>8)c->dec_firsttablen=8; michael@0: michael@0: tabn=1<dec_firsttablen; michael@0: c->dec_firsttable=_ogg_calloc(tabn,sizeof(*c->dec_firsttable)); michael@0: c->dec_maxlength=0; michael@0: michael@0: for(i=0;idec_maxlengthdec_codelengths[i]) michael@0: c->dec_maxlength=c->dec_codelengths[i]; michael@0: if(c->dec_codelengths[i]<=c->dec_firsttablen){ michael@0: ogg_uint32_t orig=bitreverse(c->codelist[i]); michael@0: for(j=0;j<(1<<(c->dec_firsttablen-c->dec_codelengths[i]));j++) michael@0: c->dec_firsttable[orig|(j<dec_codelengths[i])]=i+1; michael@0: } michael@0: } michael@0: michael@0: /* now fill in 'unused' entries in the firsttable with hi/lo search michael@0: hints for the non-direct-hits */ michael@0: { michael@0: ogg_uint32_t mask=0xfffffffeUL<<(31-c->dec_firsttablen); michael@0: long lo=0,hi=0; michael@0: michael@0: for(i=0;idec_firsttablen); michael@0: if(c->dec_firsttable[bitreverse(word)]==0){ michael@0: while((lo+1)codelist[lo+1]<=word)lo++; michael@0: while( hi=(c->codelist[hi]&mask))hi++; michael@0: michael@0: /* we only actually have 15 bits per hint to play with here. michael@0: In order to overflow gracefully (nothing breaks, efficiency michael@0: just drops), encode as the difference from the extremes. */ michael@0: { michael@0: unsigned long loval=lo; michael@0: unsigned long hival=n-hi; michael@0: michael@0: if(loval>0x7fff)loval=0x7fff; michael@0: if(hival>0x7fff)hival=0x7fff; michael@0: c->dec_firsttable[bitreverse(word)]= michael@0: 0x80000000UL | (loval<<15) | hival; michael@0: } michael@0: } michael@0: } michael@0: } michael@0: } michael@0: michael@0: return(0); michael@0: err_out: michael@0: vorbis_book_clear(c); michael@0: return(-1); michael@0: } michael@0: michael@0: long vorbis_book_codeword(codebook *book,int entry){ michael@0: if(book->c) /* only use with encode; decode optimizations are michael@0: allowed to break this */ michael@0: return book->codelist[entry]; michael@0: return -1; michael@0: } michael@0: michael@0: long vorbis_book_codelen(codebook *book,int entry){ michael@0: if(book->c) /* only use with encode; decode optimizations are michael@0: allowed to break this */ michael@0: return book->c->lengthlist[entry]; michael@0: return -1; michael@0: } michael@0: michael@0: #ifdef _V_SELFTEST michael@0: michael@0: /* Unit tests of the dequantizer; this stuff will be OK michael@0: cross-platform, I simply want to be sure that special mapping cases michael@0: actually work properly; a bug could go unnoticed for a while */ michael@0: michael@0: #include michael@0: michael@0: /* cases: michael@0: michael@0: no mapping michael@0: full, explicit mapping michael@0: algorithmic mapping michael@0: michael@0: nonsequential michael@0: sequential michael@0: */ michael@0: michael@0: static long full_quantlist1[]={0,1,2,3, 4,5,6,7, 8,3,6,1}; michael@0: static long partial_quantlist1[]={0,7,2}; michael@0: michael@0: /* no mapping */ michael@0: static_codebook test1={ michael@0: 4,16, michael@0: NULL, michael@0: 0, michael@0: 0,0,0,0, michael@0: NULL, michael@0: 0 michael@0: }; michael@0: static float *test1_result=NULL; michael@0: michael@0: /* linear, full mapping, nonsequential */ michael@0: static_codebook test2={ michael@0: 4,3, michael@0: NULL, michael@0: 2, michael@0: -533200896,1611661312,4,0, michael@0: full_quantlist1, michael@0: 0 michael@0: }; michael@0: static float test2_result[]={-3,-2,-1,0, 1,2,3,4, 5,0,3,-2}; michael@0: michael@0: /* linear, full mapping, sequential */ michael@0: static_codebook test3={ michael@0: 4,3, michael@0: NULL, michael@0: 2, michael@0: -533200896,1611661312,4,1, michael@0: full_quantlist1, michael@0: 0 michael@0: }; michael@0: static float test3_result[]={-3,-5,-6,-6, 1,3,6,10, 5,5,8,6}; michael@0: michael@0: /* linear, algorithmic mapping, nonsequential */ michael@0: static_codebook test4={ michael@0: 3,27, michael@0: NULL, michael@0: 1, michael@0: -533200896,1611661312,4,0, michael@0: partial_quantlist1, michael@0: 0 michael@0: }; michael@0: static float test4_result[]={-3,-3,-3, 4,-3,-3, -1,-3,-3, michael@0: -3, 4,-3, 4, 4,-3, -1, 4,-3, michael@0: -3,-1,-3, 4,-1,-3, -1,-1,-3, michael@0: -3,-3, 4, 4,-3, 4, -1,-3, 4, michael@0: -3, 4, 4, 4, 4, 4, -1, 4, 4, michael@0: -3,-1, 4, 4,-1, 4, -1,-1, 4, michael@0: -3,-3,-1, 4,-3,-1, -1,-3,-1, michael@0: -3, 4,-1, 4, 4,-1, -1, 4,-1, michael@0: -3,-1,-1, 4,-1,-1, -1,-1,-1}; michael@0: michael@0: /* linear, algorithmic mapping, sequential */ michael@0: static_codebook test5={ michael@0: 3,27, michael@0: NULL, michael@0: 1, michael@0: -533200896,1611661312,4,1, michael@0: partial_quantlist1, michael@0: 0 michael@0: }; michael@0: static float test5_result[]={-3,-6,-9, 4, 1,-2, -1,-4,-7, michael@0: -3, 1,-2, 4, 8, 5, -1, 3, 0, michael@0: -3,-4,-7, 4, 3, 0, -1,-2,-5, michael@0: -3,-6,-2, 4, 1, 5, -1,-4, 0, michael@0: -3, 1, 5, 4, 8,12, -1, 3, 7, michael@0: -3,-4, 0, 4, 3, 7, -1,-2, 2, michael@0: -3,-6,-7, 4, 1, 0, -1,-4,-5, michael@0: -3, 1, 0, 4, 8, 7, -1, 3, 2, michael@0: -3,-4,-5, 4, 3, 2, -1,-2,-3}; michael@0: michael@0: void run_test(static_codebook *b,float *comp){ michael@0: float *out=_book_unquantize(b,b->entries,NULL); michael@0: int i; michael@0: michael@0: if(comp){ michael@0: if(!out){ michael@0: fprintf(stderr,"_book_unquantize incorrectly returned NULL\n"); michael@0: exit(1); michael@0: } michael@0: michael@0: for(i=0;ientries*b->dim;i++) michael@0: if(fabs(out[i]-comp[i])>.0001){ michael@0: fprintf(stderr,"disagreement in unquantized and reference data:\n" michael@0: "position %d, %g != %g\n",i,out[i],comp[i]); michael@0: exit(1); michael@0: } michael@0: michael@0: }else{ michael@0: if(out){ michael@0: fprintf(stderr,"_book_unquantize returned a value array: \n" michael@0: " correct result should have been NULL\n"); michael@0: exit(1); michael@0: } michael@0: } michael@0: } michael@0: michael@0: int main(){ michael@0: /* run the nine dequant tests, and compare to the hand-rolled results */ michael@0: fprintf(stderr,"Dequant test 1... "); michael@0: run_test(&test1,test1_result); michael@0: fprintf(stderr,"OK\nDequant test 2... "); michael@0: run_test(&test2,test2_result); michael@0: fprintf(stderr,"OK\nDequant test 3... "); michael@0: run_test(&test3,test3_result); michael@0: fprintf(stderr,"OK\nDequant test 4... "); michael@0: run_test(&test4,test4_result); michael@0: fprintf(stderr,"OK\nDequant test 5... "); michael@0: run_test(&test5,test5_result); michael@0: fprintf(stderr,"OK\n\n"); michael@0: michael@0: return(0); michael@0: } michael@0: michael@0: #endif