14280
|
1 /********************************************************************
|
|
2 * *
|
|
3 * THIS FILE IS PART OF THE OggVorbis 'TREMOR' CODEC SOURCE CODE. *
|
|
4 * *
|
|
5 * THE OggVorbis 'TREMOR' SOURCE CODE IS (C) COPYRIGHT 1994-2002 *
|
|
6 * BY THE Xiph.Org FOUNDATION http://www.xiph.org/ *
|
|
7 * ALL REDISTRIBUTION RIGHTS RESERVED. *
|
|
8 * *
|
|
9 ********************************************************************
|
|
10
|
|
11 function: basic shared codebook operations
|
|
12
|
|
13 ********************************************************************/
|
|
14
|
|
15 #include <stdlib.h>
|
|
16 #include <math.h>
|
|
17 #include <string.h>
|
|
18 #include "ogg.h"
|
|
19 #include "os.h"
|
|
20 #include "misc.h"
|
|
21 #include "ivorbiscodec.h"
|
|
22 #include "codebook.h"
|
|
23
|
|
24 /**** pack/unpack helpers ******************************************/
|
|
25 int _ilog(unsigned int v){
|
|
26 int ret=0;
|
|
27 while(v){
|
|
28 ret++;
|
|
29 v>>=1;
|
|
30 }
|
|
31 return(ret);
|
|
32 }
|
|
33
|
|
34 /* 32 bit float (not IEEE; nonnormalized mantissa +
|
|
35 biased exponent) : neeeeeee eeemmmmm mmmmmmmm mmmmmmmm
|
|
36 Why not IEEE? It's just not that important here. */
|
|
37
|
|
38 #define VQ_FEXP 10
|
|
39 #define VQ_FMAN 21
|
|
40 #define VQ_FEXP_BIAS 768 /* bias toward values smaller than 1. */
|
|
41
|
|
42 static ogg_int32_t _float32_unpack(long val,int *point){
|
|
43 long mant=val&0x1fffff;
|
|
44 int sign=val&0x80000000;
|
|
45 long exp =(val&0x7fe00000L)>>VQ_FMAN;
|
|
46
|
|
47 exp-=(VQ_FMAN-1)+VQ_FEXP_BIAS;
|
|
48
|
|
49 if(mant){
|
|
50 while(!(mant&0x40000000)){
|
|
51 mant<<=1;
|
|
52 exp-=1;
|
|
53 }
|
|
54
|
|
55 if(sign)mant= -mant;
|
|
56 }else{
|
|
57 sign=0;
|
|
58 exp=-9999;
|
|
59 }
|
|
60
|
|
61 *point=exp;
|
|
62 return mant;
|
|
63 }
|
|
64
|
|
65 /* given a list of word lengths, generate a list of codewords. Works
|
|
66 for length ordered or unordered, always assigns the lowest valued
|
|
67 codewords first. Extended to handle unused entries (length 0) */
|
|
68 ogg_uint32_t *_make_words(long *l,long n,long sparsecount){
|
|
69 long i,j,count=0;
|
|
70 ogg_uint32_t marker[33];
|
|
71 ogg_uint32_t *r=(ogg_uint32_t *)_ogg_malloc((sparsecount?sparsecount:n)*sizeof(*r));
|
|
72 memset(marker,0,sizeof(marker));
|
|
73
|
|
74 for(i=0;i<n;i++){
|
|
75 long length=l[i];
|
|
76 if(length>0){
|
|
77 ogg_uint32_t entry=marker[length];
|
|
78
|
|
79 /* when we claim a node for an entry, we also claim the nodes
|
|
80 below it (pruning off the imagined tree that may have dangled
|
|
81 from it) as well as blocking the use of any nodes directly
|
|
82 above for leaves */
|
|
83
|
|
84 /* update ourself */
|
|
85 if(length<32 && (entry>>length)){
|
|
86 /* error condition; the lengths must specify an overpopulated tree */
|
|
87 _ogg_free(r);
|
|
88 return(NULL);
|
|
89 }
|
|
90 r[count++]=entry;
|
|
91
|
|
92 /* Look to see if the next shorter marker points to the node
|
|
93 above. if so, update it and repeat. */
|
|
94 {
|
|
95 for(j=length;j>0;j--){
|
|
96
|
|
97 if(marker[j]&1){
|
|
98 /* have to jump branches */
|
|
99 if(j==1)
|
|
100 marker[1]++;
|
|
101 else
|
|
102 marker[j]=marker[j-1]<<1;
|
|
103 break; /* invariant says next upper marker would already
|
|
104 have been moved if it was on the same path */
|
|
105 }
|
|
106 marker[j]++;
|
|
107 }
|
|
108 }
|
|
109
|
|
110 /* prune the tree; the implicit invariant says all the longer
|
|
111 markers were dangling from our just-taken node. Dangle them
|
|
112 from our *new* node. */
|
|
113 for(j=length+1;j<33;j++)
|
|
114 if((marker[j]>>1) == entry){
|
|
115 entry=marker[j];
|
|
116 marker[j]=marker[j-1]<<1;
|
|
117 }else
|
|
118 break;
|
|
119 }else
|
|
120 if(sparsecount==0)count++;
|
|
121 }
|
|
122
|
|
123 /* bitreverse the words because our bitwise packer/unpacker is LSb
|
|
124 endian */
|
|
125 for(i=0,count=0;i<n;i++){
|
|
126 ogg_uint32_t temp=0;
|
|
127 for(j=0;j<l[i];j++){
|
|
128 temp<<=1;
|
|
129 temp|=(r[count]>>j)&1;
|
|
130 }
|
|
131
|
|
132 if(sparsecount){
|
|
133 if(l[i])
|
|
134 r[count++]=temp;
|
|
135 }else
|
|
136 r[count++]=temp;
|
|
137 }
|
|
138
|
|
139 return(r);
|
|
140 }
|
|
141
|
|
142 /* there might be a straightforward one-line way to do the below
|
|
143 that's portable and totally safe against roundoff, but I haven't
|
|
144 thought of it. Therefore, we opt on the side of caution */
|
|
145 long _book_maptype1_quantvals(const static_codebook *b){
|
|
146 /* get us a starting hint, we'll polish it below */
|
|
147 int bits=_ilog(b->entries);
|
|
148 int vals=b->entries>>((bits-1)*(b->dim-1)/b->dim);
|
|
149
|
|
150 while(1){
|
|
151 long acc=1;
|
|
152 long acc1=1;
|
|
153 int i;
|
|
154 for(i=0;i<b->dim;i++){
|
|
155 acc*=vals;
|
|
156 acc1*=vals+1;
|
|
157 }
|
|
158 if(acc<=b->entries && acc1>b->entries){
|
|
159 return(vals);
|
|
160 }else{
|
|
161 if(acc>b->entries){
|
|
162 vals--;
|
|
163 }else{
|
|
164 vals++;
|
|
165 }
|
|
166 }
|
|
167 }
|
|
168 }
|
|
169
|
|
170 /* different than what _book_unquantize does for mainline:
|
|
171 we repack the book in a fixed point format that shares the same
|
|
172 binary point. Upon first use, we can shift point if needed */
|
|
173
|
|
174 /* we need to deal with two map types: in map type 1, the values are
|
|
175 generated algorithmically (each column of the vector counts through
|
|
176 the values in the quant vector). in map type 2, all the values came
|
|
177 in in an explicit list. Both value lists must be unpacked */
|
|
178
|
|
179 ogg_int32_t *_book_unquantize(const static_codebook *b,int n,int *sparsemap,
|
|
180 int *maxpoint){
|
|
181 long j,k,count=0;
|
|
182 if(b->maptype==1 || b->maptype==2){
|
|
183 int quantvals;
|
|
184 int minpoint,delpoint;
|
|
185 ogg_int32_t mindel=_float32_unpack(b->q_min,&minpoint);
|
|
186 ogg_int32_t delta=_float32_unpack(b->q_delta,&delpoint);
|
|
187 ogg_int32_t *r=(ogg_int32_t *)_ogg_calloc(n*b->dim,sizeof(*r));
|
|
188 int *rp=(int *)_ogg_calloc(n*b->dim,sizeof(*rp));
|
|
189
|
|
190 *maxpoint=minpoint;
|
|
191
|
|
192 /* maptype 1 and 2 both use a quantized value vector, but
|
|
193 different sizes */
|
|
194 switch(b->maptype){
|
|
195 case 1:
|
|
196 /* most of the time, entries%dimensions == 0, but we need to be
|
|
197 well defined. We define that the possible vales at each
|
|
198 scalar is values == entries/dim. If entries%dim != 0, we'll
|
|
199 have 'too few' values (values*dim<entries), which means that
|
|
200 we'll have 'left over' entries; left over entries use zeroed
|
|
201 values (and are wasted). So don't generate codebooks like
|
|
202 that */
|
|
203 quantvals=_book_maptype1_quantvals(b);
|
|
204 for(j=0;j<b->entries;j++){
|
|
205 if((sparsemap && b->lengthlist[j]) || !sparsemap){
|
|
206 ogg_int32_t last=0;
|
|
207 int lastpoint=0;
|
|
208 int indexdiv=1;
|
|
209 for(k=0;k<b->dim;k++){
|
|
210 int index= (j/indexdiv)%quantvals;
|
|
211 int point;
|
|
212 int val=VFLOAT_MULTI(delta,delpoint,
|
|
213 abs(b->quantlist[index]),&point);
|
|
214
|
|
215 val=VFLOAT_ADD(mindel,minpoint,val,point,&point);
|
|
216 val=VFLOAT_ADD(last,lastpoint,val,point,&point);
|
|
217
|
|
218 if(b->q_sequencep){
|
|
219 last=val;
|
|
220 lastpoint=point;
|
|
221 }
|
|
222
|
|
223 if(sparsemap){
|
|
224 r[sparsemap[count]*b->dim+k]=val;
|
|
225 rp[sparsemap[count]*b->dim+k]=point;
|
|
226 }else{
|
|
227 r[count*b->dim+k]=val;
|
|
228 rp[count*b->dim+k]=point;
|
|
229 }
|
|
230 if(*maxpoint<point)*maxpoint=point;
|
|
231 indexdiv*=quantvals;
|
|
232 }
|
|
233 count++;
|
|
234 }
|
|
235
|
|
236 }
|
|
237 break;
|
|
238 case 2:
|
|
239 for(j=0;j<b->entries;j++){
|
|
240 if((sparsemap && b->lengthlist[j]) || !sparsemap){
|
|
241 ogg_int32_t last=0;
|
|
242 int lastpoint=0;
|
|
243
|
|
244 for(k=0;k<b->dim;k++){
|
|
245 int point;
|
|
246 int val=VFLOAT_MULTI(delta,delpoint,
|
|
247 abs(b->quantlist[j*b->dim+k]),&point);
|
|
248
|
|
249 val=VFLOAT_ADD(mindel,minpoint,val,point,&point);
|
|
250 val=VFLOAT_ADD(last,lastpoint,val,point,&point);
|
|
251
|
|
252 if(b->q_sequencep){
|
|
253 last=val;
|
|
254 lastpoint=point;
|
|
255 }
|
|
256
|
|
257 if(sparsemap){
|
|
258 r[sparsemap[count]*b->dim+k]=val;
|
|
259 rp[sparsemap[count]*b->dim+k]=point;
|
|
260 }else{
|
|
261 r[count*b->dim+k]=val;
|
|
262 rp[count*b->dim+k]=point;
|
|
263 }
|
|
264 if(*maxpoint<point)*maxpoint=point;
|
|
265 }
|
|
266 count++;
|
|
267 }
|
|
268 }
|
|
269 break;
|
|
270 }
|
|
271
|
|
272 for(j=0;j<n*b->dim;j++)
|
|
273 if(rp[j]<*maxpoint)
|
|
274 r[j]>>=*maxpoint-rp[j];
|
|
275
|
|
276 _ogg_free(rp);
|
|
277 return(r);
|
|
278 }
|
|
279 return(NULL);
|
|
280 }
|
|
281
|
|
282 void vorbis_staticbook_clear(static_codebook *b){
|
|
283 if(b->quantlist)_ogg_free(b->quantlist);
|
|
284 if(b->lengthlist)_ogg_free(b->lengthlist);
|
|
285 memset(b,0,sizeof(*b));
|
|
286
|
|
287 }
|
|
288
|
|
289 void vorbis_staticbook_destroy(static_codebook *b){
|
|
290 vorbis_staticbook_clear(b);
|
|
291 _ogg_free(b);
|
|
292 }
|
|
293
|
|
294 void vorbis_book_clear(codebook *b){
|
|
295 /* static book is not cleared; we're likely called on the lookup and
|
|
296 the static codebook belongs to the info struct */
|
|
297 if(b->valuelist)_ogg_free(b->valuelist);
|
|
298 if(b->codelist)_ogg_free(b->codelist);
|
|
299
|
|
300 if(b->dec_index)_ogg_free(b->dec_index);
|
|
301 if(b->dec_codelengths)_ogg_free(b->dec_codelengths);
|
|
302 if(b->dec_firsttable)_ogg_free(b->dec_firsttable);
|
|
303
|
|
304 memset(b,0,sizeof(*b));
|
|
305 }
|
|
306
|
|
307 static ogg_uint32_t bitreverse(ogg_uint32_t x){
|
|
308 x= ((x>>16)&0x0000ffffUL) | ((x<<16)&0xffff0000UL);
|
|
309 x= ((x>> 8)&0x00ff00ffUL) | ((x<< 8)&0xff00ff00UL);
|
|
310 x= ((x>> 4)&0x0f0f0f0fUL) | ((x<< 4)&0xf0f0f0f0UL);
|
|
311 x= ((x>> 2)&0x33333333UL) | ((x<< 2)&0xccccccccUL);
|
|
312 return((x>> 1)&0x55555555UL) | ((x<< 1)&0xaaaaaaaaUL);
|
|
313 }
|
|
314
|
|
315 static int sort32a(const void *a,const void *b){
|
|
316 return ( (**(ogg_uint32_t **)a>**(ogg_uint32_t **)b)<<1)-1;
|
|
317 }
|
|
318
|
|
319 /* decode codebook arrangement is more heavily optimized than encode */
|
|
320 int vorbis_book_init_decode(codebook *c,const static_codebook *s){
|
|
321 int i,j,n=0,tabn;
|
|
322 int *sortindex;
|
|
323 memset(c,0,sizeof(*c));
|
|
324
|
|
325 /* count actually used entries */
|
|
326 for(i=0;i<s->entries;i++)
|
|
327 if(s->lengthlist[i]>0)
|
|
328 n++;
|
|
329
|
|
330 c->entries=s->entries;
|
|
331 c->used_entries=n;
|
|
332 c->dim=s->dim;
|
|
333
|
|
334 c->q_min=s->q_min;
|
|
335 c->q_delta=s->q_delta;
|
|
336
|
|
337 /* two different remappings go on here.
|
|
338
|
|
339 First, we collapse the likely sparse codebook down only to
|
|
340 actually represented values/words. This collapsing needs to be
|
|
341 indexed as map-valueless books are used to encode original entry
|
|
342 positions as integers.
|
|
343
|
|
344 Second, we reorder all vectors, including the entry index above,
|
|
345 by sorted bitreversed codeword to allow treeless decode. */
|
|
346
|
|
347 {
|
|
348 /* perform sort */
|
|
349 ogg_uint32_t *codes=_make_words(s->lengthlist,s->entries,c->used_entries);
|
|
350 ogg_uint32_t **codep=(ogg_uint32_t **)alloca(sizeof(*codep)*n);
|
|
351
|
|
352 if(codes==NULL)goto err_out;
|
|
353
|
|
354 for(i=0;i<n;i++){
|
|
355 codes[i]=bitreverse(codes[i]);
|
|
356 codep[i]=codes+i;
|
|
357 }
|
|
358
|
|
359 qsort(codep,n,sizeof(*codep),sort32a);
|
|
360
|
|
361 sortindex=(int *)alloca(n*sizeof(*sortindex));
|
|
362 c->codelist=(ogg_uint32_t *)_ogg_malloc(n*sizeof(*c->codelist));
|
|
363 /* the index is a reverse index */
|
|
364 for(i=0;i<n;i++){
|
|
365 int position=codep[i]-codes;
|
|
366 sortindex[position]=i;
|
|
367 }
|
|
368
|
|
369 for(i=0;i<n;i++)
|
|
370 c->codelist[sortindex[i]]=codes[i];
|
|
371 _ogg_free(codes);
|
|
372 }
|
|
373
|
|
374
|
|
375 c->valuelist=_book_unquantize(s,n,sortindex,&c->binarypoint);
|
|
376 c->dec_index=(int *)_ogg_malloc(n*sizeof(*c->dec_index));
|
|
377
|
|
378 for(n=0,i=0;i<s->entries;i++)
|
|
379 if(s->lengthlist[i]>0)
|
|
380 c->dec_index[sortindex[n++]]=i;
|
|
381
|
|
382 c->dec_codelengths=(char *)_ogg_malloc(n*sizeof(*c->dec_codelengths));
|
|
383 for(n=0,i=0;i<s->entries;i++)
|
|
384 if(s->lengthlist[i]>0)
|
|
385 c->dec_codelengths[sortindex[n++]]=s->lengthlist[i];
|
|
386
|
|
387 c->dec_firsttablen=_ilog(c->used_entries)-4; /* this is magic */
|
|
388 if(c->dec_firsttablen<5)c->dec_firsttablen=5;
|
|
389 if(c->dec_firsttablen>8)c->dec_firsttablen=8;
|
|
390
|
|
391 tabn=1<<c->dec_firsttablen;
|
|
392 c->dec_firsttable=(ogg_uint32_t *)_ogg_calloc(tabn,sizeof(*c->dec_firsttable));
|
|
393 c->dec_maxlength=0;
|
|
394
|
|
395 for(i=0;i<n;i++){
|
|
396 if(c->dec_maxlength<c->dec_codelengths[i])
|
|
397 c->dec_maxlength=c->dec_codelengths[i];
|
|
398 if(c->dec_codelengths[i]<=c->dec_firsttablen){
|
|
399 ogg_uint32_t orig=bitreverse(c->codelist[i]);
|
|
400 for(j=0;j<(1<<(c->dec_firsttablen-c->dec_codelengths[i]));j++)
|
|
401 c->dec_firsttable[orig|(j<<c->dec_codelengths[i])]=i+1;
|
|
402 }
|
|
403 }
|
|
404
|
|
405 /* now fill in 'unused' entries in the firsttable with hi/lo search
|
|
406 hints for the non-direct-hits */
|
|
407 {
|
|
408 ogg_uint32_t mask=0xfffffffeUL<<(31-c->dec_firsttablen);
|
|
409 long lo=0,hi=0;
|
|
410
|
|
411 for(i=0;i<tabn;i++){
|
|
412 ogg_uint32_t word=i<<(32-c->dec_firsttablen);
|
|
413 if(c->dec_firsttable[bitreverse(word)]==0){
|
|
414 while((lo+1)<n && c->codelist[lo+1]<=word)lo++;
|
|
415 while( hi<n && word>=(c->codelist[hi]&mask))hi++;
|
|
416
|
|
417 /* we only actually have 15 bits per hint to play with here.
|
|
418 In order to overflow gracefully (nothing breaks, efficiency
|
|
419 just drops), encode as the difference from the extremes. */
|
|
420 {
|
|
421 unsigned long loval=lo;
|
|
422 unsigned long hival=n-hi;
|
|
423
|
|
424 if(loval>0x7fff)loval=0x7fff;
|
|
425 if(hival>0x7fff)hival=0x7fff;
|
|
426 c->dec_firsttable[bitreverse(word)]=
|
|
427 0x80000000UL | (loval<<15) | hival;
|
|
428 }
|
|
429 }
|
|
430 }
|
|
431 }
|
|
432
|
|
433
|
|
434 return(0);
|
|
435 err_out:
|
|
436 vorbis_book_clear(c);
|
|
437 return(-1);
|
|
438 }
|
|
439
|