sharedbook.c

/********************************************************************
 *                                                                  *
 * THIS FILE IS PART OF THE OggVorbis 'TREMOR' CODEC SOURCE CODE.   *
 *                                                                  *
 * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
 * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
 * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
 *                                                                  *
 * THE OggVorbis 'TREMOR' SOURCE CODE IS (C) COPYRIGHT 1994-2002    *
 * BY THE Xiph.Org FOUNDATION http://www.xiph.org/                  *
 *                                                                  *
 ********************************************************************

 function: basic shared codebook operations

 ********************************************************************/

#include <stdlib.h>
#include <math.h>
#include <string.h>
#include <ogg/ogg.h>
#include "misc.h"
#include "ivorbiscodec.h"
#include "codebook.h"

/**** pack/unpack helpers ******************************************/
int _ilog(unsigned int v){
  int ret=0;
  while(v){
    ret++;
    v>>=1;
  }
  return(ret);
}

/* 32 bit float (not IEEE; nonnormalized mantissa +
   biased exponent) : neeeeeee eeemmmmm mmmmmmmm mmmmmmmm 
   Why not IEEE?  It's just not that important here. */

#define VQ_FEXP 10
#define VQ_FMAN 21
#define VQ_FEXP_BIAS 768 /* bias toward values smaller than 1. */

static ogg_int32_t _float32_unpack(long val,int *point){
  long   mant=val&0x1fffff;
  int    sign=val&0x80000000;
  long   exp =(val&0x7fe00000L)>>VQ_FMAN;

  exp-=(VQ_FMAN-1)+VQ_FEXP_BIAS;

  if(mant){
    while(!(mant&0x40000000)){
      mant<<=1;
      exp-=1;
    }

    if(sign)mant= -mant;
  }else{
    sign=0;
    exp=-9999;
  }

  *point=exp;
  return mant;
}

/* given a list of word lengths, generate a list of codewords.  Works
   for length ordered or unordered, always assigns the lowest valued
   codewords first.  Extended to handle unused entries (length 0) */
ogg_uint32_t *_make_words(long *l,long n,long sparsecount){
  long i,j,count=0;
  ogg_uint32_t marker[33];
  ogg_uint32_t *r=(ogg_uint32_t *)_ogg_malloc((sparsecount?sparsecount:n)*sizeof(*r));
  memset(marker,0,sizeof(marker));

  for(i=0;i<n;i++){
    long length=l[i];
    if(length>0){
      ogg_uint32_t entry=marker[length];
      
      /* when we claim a node for an entry, we also claim the nodes
	 below it (pruning off the imagined tree that may have dangled
	 from it) as well as blocking the use of any nodes directly
	 above for leaves */
      
      /* update ourself */
      if(length<32 && (entry>>length)){
	/* error condition; the lengths must specify an overpopulated tree */
	_ogg_free(r);
	return(NULL);
      }
      r[count++]=entry;
    
      /* Look to see if the next shorter marker points to the node
	 above. if so, update it and repeat.  */
      {
	for(j=length;j>0;j--){
	  
	  if(marker[j]&1){
	    /* have to jump branches */
	    if(j==1)
	      marker[1]++;
	    else
	      marker[j]=marker[j-1]<<1;
	    break; /* invariant says next upper marker would already
		      have been moved if it was on the same path */
	  }
	  marker[j]++;
	}
      }
      
      /* prune the tree; the implicit invariant says all the longer
	 markers were dangling from our just-taken node.  Dangle them
	 from our *new* node. */
      for(j=length+1;j<33;j++)
	if((marker[j]>>1) == entry){
	  entry=marker[j];
	  marker[j]=marker[j-1]<<1;
	}else
	  break;
    }else
      if(sparsecount==0)count++;
  }

  /* sanity check the huffman tree; an underpopulated tree must be
     rejected. The only exception is the one-node pseudo-nil tree,
     which appears to be underpopulated because the tree doesn't
     really exist; there's only one possible 'codeword' or zero bits,
     but the above tree-gen code doesn't mark that. */
  if(sparsecount != 1){
    for(i=1;i<33;i++)
      if(marker[i] & (0xffffffffUL>>(32-i))){
       _ogg_free(r);
       return(NULL);
      }
  }

  /* bitreverse the words because our bitwise packer/unpacker is LSb
     endian */
  for(i=0,count=0;i<n;i++){
    ogg_uint32_t temp=0;
    for(j=0;j<l[i];j++){
      temp<<=1;
      temp|=(r[count]>>j)&1;
    }

    if(sparsecount){
      if(l[i])
	r[count++]=temp;
    }else
      r[count++]=temp;
  }

  return(r);
}

/* there might be a straightforward one-line way to do the below
   that's portable and totally safe against roundoff, but I haven't
   thought of it.  Therefore, we opt on the side of caution */
long _book_maptype1_quantvals(const static_codebook *b){
  /* get us a starting hint, we'll polish it below */
  int bits=_ilog(b->entries);
  int vals=b->entries>>((bits-1)*(b->dim-1)/b->dim);

  while(1){
    long acc=1;
    long acc1=1;
    int i;
    for(i=0;i<b->dim;i++){
      acc*=vals;
      acc1*=vals+1;
    }
    if(acc<=b->entries && acc1>b->entries){
      return(vals);
    }else{
      if(acc>b->entries){
	vals--;
      }else{
	vals++;
      }
    }
  }
}

/* different than what _book_unquantize does for mainline:
   we repack the book in a fixed point format that shares the same
   binary point.  Upon first use, we can shift point if needed */

/* we need to deal with two map types: in map type 1, the values are
   generated algorithmically (each column of the vector counts through
   the values in the quant vector). in map type 2, all the values came
   in in an explicit list.  Both value lists must be unpacked */

ogg_int32_t *_book_unquantize(const static_codebook *b,int n,int *sparsemap,
			      int *maxpoint){
  long j,k,count=0;
  if(b->maptype==1 || b->maptype==2){
    int quantvals;
    int minpoint,delpoint;
    ogg_int32_t mindel=_float32_unpack(b->q_min,&minpoint);
    ogg_int32_t delta=_float32_unpack(b->q_delta,&delpoint);
    ogg_int32_t *r=(ogg_int32_t *)_ogg_calloc(n*b->dim,sizeof(*r));
    int *rp=(int *)_ogg_calloc(n*b->dim,sizeof(*rp));

    *maxpoint=minpoint;

    /* maptype 1 and 2 both use a quantized value vector, but
       different sizes */
    switch(b->maptype){
    case 1:
      /* most of the time, entries%dimensions == 0, but we need to be
	 well defined.  We define that the possible vales at each
	 scalar is values == entries/dim.  If entries%dim != 0, we'll
	 have 'too few' values (values*dim<entries), which means that
	 we'll have 'left over' entries; left over entries use zeroed
	 values (and are wasted).  So don't generate codebooks like
	 that */
      quantvals=_book_maptype1_quantvals(b);
      for(j=0;j<b->entries;j++){
	if((sparsemap && b->lengthlist[j]) || !sparsemap){
	  ogg_int32_t last=0;
	  int lastpoint=0;
	  int indexdiv=1;
	  for(k=0;k<b->dim;k++){
	    int index= (j/indexdiv)%quantvals;
	    int point=0;
	    int val=VFLOAT_MULTI(delta,delpoint,
				 abs(b->quantlist[index]),&point);

	    val=VFLOAT_ADD(mindel,minpoint,val,point,&point);
	    val=VFLOAT_ADD(last,lastpoint,val,point,&point);
	    
	    if(b->q_sequencep){
	      last=val;	  
	      lastpoint=point;
	    }
	    
	    if(sparsemap){
	      r[sparsemap[count]*b->dim+k]=val;
	      rp[sparsemap[count]*b->dim+k]=point;
	    }else{
	      r[count*b->dim+k]=val;
	      rp[count*b->dim+k]=point;
	    }
	    if(*maxpoint<point)*maxpoint=point;
	    indexdiv*=quantvals;
	  }
	  count++;
	}

      }
      break;
    case 2:
      for(j=0;j<b->entries;j++){
	if((sparsemap && b->lengthlist[j]) || !sparsemap){
	  ogg_int32_t last=0;
	  int         lastpoint=0;

	  for(k=0;k<b->dim;k++){
	    int point=0;
	    int val=VFLOAT_MULTI(delta,delpoint,
				 abs(b->quantlist[j*b->dim+k]),&point);

	    val=VFLOAT_ADD(mindel,minpoint,val,point,&point);
	    val=VFLOAT_ADD(last,lastpoint,val,point,&point);
	    
	    if(b->q_sequencep){
	      last=val;	  
	      lastpoint=point;
	    }

	    if(sparsemap){
	      r[sparsemap[count]*b->dim+k]=val;
	      rp[sparsemap[count]*b->dim+k]=point;
	    }else{
	      r[count*b->dim+k]=val;
	      rp[count*b->dim+k]=point;
	    }
	    if(*maxpoint<point)*maxpoint=point;
	  }
	  count++;
	}
      }
      break;
    }

    for(j=0;j<n*b->dim;j++)
      if(rp[j]<*maxpoint)
	r[j]>>=*maxpoint-rp[j];
	    
    _ogg_free(rp);
    return(r);
  }
  return(NULL);
}

void vorbis_staticbook_destroy(static_codebook *b){
  if(b->quantlist)_ogg_free(b->quantlist);
  if(b->lengthlist)_ogg_free(b->lengthlist);
  memset(b,0,sizeof(*b));
  _ogg_free(b);
}

void vorbis_book_clear(codebook *b){
  /* static book is not cleared; we're likely called on the lookup and
     the static codebook belongs to the info struct */
  if(b->valuelist)_ogg_free(b->valuelist);
  if(b->codelist)_ogg_free(b->codelist);

  if(b->dec_index)_ogg_free(b->dec_index);
  if(b->dec_codelengths)_ogg_free(b->dec_codelengths);
  if(b->dec_firsttable)_ogg_free(b->dec_firsttable);

  memset(b,0,sizeof(*b));
}

static ogg_uint32_t bitreverse(ogg_uint32_t x){
  x=    ((x>>16)&0x0000ffffUL) | ((x<<16)&0xffff0000UL);
  x=    ((x>> 8)&0x00ff00ffUL) | ((x<< 8)&0xff00ff00UL);
  x=    ((x>> 4)&0x0f0f0f0fUL) | ((x<< 4)&0xf0f0f0f0UL);
  x=    ((x>> 2)&0x33333333UL) | ((x<< 2)&0xccccccccUL);
  return((x>> 1)&0x55555555UL) | ((x<< 1)&0xaaaaaaaaUL);
}

static int sort32a(const void *a,const void *b){
  return (**(ogg_uint32_t **)a>**(ogg_uint32_t **)b)-
    (**(ogg_uint32_t **)a<**(ogg_uint32_t **)b);
}

/* decode codebook arrangement is more heavily optimized than encode */
int vorbis_book_init_decode(codebook *c,const static_codebook *s){
  int i,j,n=0,tabn;
  int *sortindex;
  memset(c,0,sizeof(*c));
  
  /* count actually used entries */
  for(i=0;i<s->entries;i++)
    if(s->lengthlist[i]>0)
      n++;

  c->entries=s->entries;
  c->used_entries=n;
  c->dim=s->dim;

  if(n>0){
    /* two different remappings go on here.  
       
       First, we collapse the likely sparse codebook down only to
       actually represented values/words.  This collapsing needs to be
       indexed as map-valueless books are used to encode original entry
       positions as integers.
       
       Second, we reorder all vectors, including the entry index above,
       by sorted bitreversed codeword to allow treeless decode. */
    
    /* perform sort */
    ogg_uint32_t *codes=_make_words(s->lengthlist,s->entries,c->used_entries);
    ogg_uint32_t **codep=(ogg_uint32_t **)alloca(sizeof(*codep)*n);
    
    if(codes==NULL)goto err_out;

    for(i=0;i<n;i++){
      codes[i]=bitreverse(codes[i]);
      codep[i]=codes+i;
    }

    qsort(codep,n,sizeof(*codep),sort32a);

    sortindex=(int *)alloca(n*sizeof(*sortindex));
    c->codelist=(ogg_uint32_t *)_ogg_malloc(n*sizeof(*c->codelist));
    /* the index is a reverse index */
    for(i=0;i<n;i++){
      int position=codep[i]-codes;
      sortindex[position]=i;
    }

    for(i=0;i<n;i++)
      c->codelist[sortindex[i]]=codes[i];
    _ogg_free(codes);
    
    
    
    c->valuelist=_book_unquantize(s,n,sortindex,&c->binarypoint);
    c->dec_index=(int *)_ogg_malloc(n*sizeof(*c->dec_index));
    
    for(n=0,i=0;i<s->entries;i++)
      if(s->lengthlist[i]>0)
	c->dec_index[sortindex[n++]]=i;
    
    c->dec_codelengths=(char *)_ogg_malloc(n*sizeof(*c->dec_codelengths));
    for(n=0,i=0;i<s->entries;i++)
      if(s->lengthlist[i]>0)
	c->dec_codelengths[sortindex[n++]]=s->lengthlist[i];
    
    c->dec_firsttablen=_ilog(c->used_entries)-4; /* this is magic */
    if(c->dec_firsttablen<5)c->dec_firsttablen=5;
    if(c->dec_firsttablen>8)c->dec_firsttablen=8;
    
    tabn=1<<c->dec_firsttablen;
    c->dec_firsttable=(ogg_uint32_t *)_ogg_calloc(tabn,sizeof(*c->dec_firsttable));
    c->dec_maxlength=0;
    
    for(i=0;i<n;i++){
      if(c->dec_maxlength<c->dec_codelengths[i])
	c->dec_maxlength=c->dec_codelengths[i];
      if(c->dec_codelengths[i]<=c->dec_firsttablen){
	ogg_uint32_t orig=bitreverse(c->codelist[i]);
	for(j=0;j<(1<<(c->dec_firsttablen-c->dec_codelengths[i]));j++)
	  c->dec_firsttable[orig|(j<<c->dec_codelengths[i])]=i+1;
      }
    }
    
    /* now fill in 'unused' entries in the firsttable with hi/lo search
       hints for the non-direct-hits */
    {
      ogg_uint32_t mask=0xfffffffeUL<<(31-c->dec_firsttablen);
      long lo=0,hi=0;
      
      for(i=0;i<tabn;i++){
	ogg_uint32_t word=i<<(32-c->dec_firsttablen);
	if(c->dec_firsttable[bitreverse(word)]==0){
	  while((lo+1)<n && c->codelist[lo+1]<=word)lo++;
	  while(    hi<n && word>=(c->codelist[hi]&mask))hi++;
	  
	  /* we only actually have 15 bits per hint to play with here.
	     In order to overflow gracefully (nothing breaks, efficiency
	     just drops), encode as the difference from the extremes. */
	  {
	    unsigned long loval=lo;
	    unsigned long hival=n-hi;
	    
	    if(loval>0x7fff)loval=0x7fff;
	    if(hival>0x7fff)hival=0x7fff;
	    c->dec_firsttable[bitreverse(word)]=
	      0x80000000UL | (loval<<15) | hival;
	  }
	}
      }
    }
  }

  return(0);
 err_out:
  vorbis_book_clear(c);
  return(-1);
}