Merge pull request #542 from gameblabla/mdec_fix
[pcsx_rearmed.git] / deps / libchdr / src / libchdr_huffman.c
CommitLineData
9c659ffe 1/* license:BSD-3-Clause
2 * copyright-holders:Aaron Giles
3****************************************************************************
4
5 huffman.c
6
7 Static Huffman compression and decompression helpers.
8
9****************************************************************************
10
11 Maximum codelength is officially (alphabetsize - 1). This would be 255 bits
12 (since we use 1 byte values). However, it is also dependent upon the number
13 of samples used, as follows:
14
15 2 bits -> 3..4 samples
16 3 bits -> 5..7 samples
17 4 bits -> 8..12 samples
18 5 bits -> 13..20 samples
19 6 bits -> 21..33 samples
20 7 bits -> 34..54 samples
21 8 bits -> 55..88 samples
22 9 bits -> 89..143 samples
23 10 bits -> 144..232 samples
24 11 bits -> 233..376 samples
25 12 bits -> 377..609 samples
26 13 bits -> 610..986 samples
27 14 bits -> 987..1596 samples
28 15 bits -> 1597..2583 samples
29 16 bits -> 2584..4180 samples -> note that a 4k data size guarantees codelength <= 16 bits
30 17 bits -> 4181..6764 samples
31 18 bits -> 6765..10945 samples
32 19 bits -> 10946..17710 samples
33 20 bits -> 17711..28656 samples
34 21 bits -> 28657..46367 samples
35 22 bits -> 46368..75024 samples
36 23 bits -> 75025..121392 samples
37 24 bits -> 121393..196417 samples
38 25 bits -> 196418..317810 samples
39 26 bits -> 317811..514228 samples
40 27 bits -> 514229..832039 samples
41 28 bits -> 832040..1346268 samples
42 29 bits -> 1346269..2178308 samples
43 30 bits -> 2178309..3524577 samples
44 31 bits -> 3524578..5702886 samples
45 32 bits -> 5702887..9227464 samples
46
47 Looking at it differently, here is where powers of 2 fall into these buckets:
48
49 256 samples -> 11 bits max
50 512 samples -> 12 bits max
51 1k samples -> 14 bits max
52 2k samples -> 15 bits max
53 4k samples -> 16 bits max
54 8k samples -> 18 bits max
55 16k samples -> 19 bits max
56 32k samples -> 21 bits max
57 64k samples -> 22 bits max
58 128k samples -> 24 bits max
59 256k samples -> 25 bits max
60 512k samples -> 27 bits max
61 1M samples -> 28 bits max
62 2M samples -> 29 bits max
63 4M samples -> 31 bits max
64 8M samples -> 32 bits max
65
66****************************************************************************
67
68 Delta-RLE encoding works as follows:
69
70 Starting value is assumed to be 0. All data is encoded as a delta
71 from the previous value, such that final[i] = final[i - 1] + delta.
72 Long runs of 0s are RLE-encoded as follows:
73
74 0x100 = repeat count of 8
75 0x101 = repeat count of 9
76 0x102 = repeat count of 10
77 0x103 = repeat count of 11
78 0x104 = repeat count of 12
79 0x105 = repeat count of 13
80 0x106 = repeat count of 14
81 0x107 = repeat count of 15
82 0x108 = repeat count of 16
83 0x109 = repeat count of 32
84 0x10a = repeat count of 64
85 0x10b = repeat count of 128
86 0x10c = repeat count of 256
87 0x10d = repeat count of 512
88 0x10e = repeat count of 1024
89 0x10f = repeat count of 2048
90
91 Note that repeat counts are reset at the end of a row, so if a 0 run
92 extends to the end of a row, a large repeat count may be used.
93
94 The reason for starting the run counts at 8 is that 0 is expected to
95 be the most common symbol, and is typically encoded in 1 or 2 bits.
96
97***************************************************************************/
98
99#include <stdlib.h>
100#include <assert.h>
101#include <stdio.h>
102#include <string.h>
103
2ff0b512 104#include <libchdr/huffman.h>
9c659ffe 105
106#define MAX(x,y) ((x) > (y) ? (x) : (y))
107
108/***************************************************************************
109 * MACROS
110 ***************************************************************************
111 */
112
113#define MAKE_LOOKUP(code,bits) (((code) << 5) | ((bits) & 0x1f))
114
115/***************************************************************************
116 * IMPLEMENTATION
117 ***************************************************************************
118 */
119
120/*-------------------------------------------------
121 * huffman_context_base - create an encoding/
122 * decoding context
123 *-------------------------------------------------
124 */
125
126struct huffman_decoder* create_huffman_decoder(int numcodes, int maxbits)
127{
dcd8e4fd 128 struct huffman_decoder* decoder = NULL;
129
9c659ffe 130 /* limit to 24 bits */
131 if (maxbits > 24)
132 return NULL;
133
dcd8e4fd 134 decoder = (struct huffman_decoder*)malloc(sizeof(struct huffman_decoder));
9c659ffe 135 decoder->numcodes = numcodes;
136 decoder->maxbits = maxbits;
137 decoder->lookup = (lookup_value*)malloc(sizeof(lookup_value) * (1 << maxbits));
138 decoder->huffnode = (struct node_t*)malloc(sizeof(struct node_t) * numcodes);
139 decoder->datahisto = NULL;
140 decoder->prevdata = 0;
141 decoder->rleremaining = 0;
142 return decoder;
143}
144
2ff0b512 145void delete_huffman_decoder(struct huffman_decoder* decoder)
146{
147 if (decoder != NULL)
148 {
149 if (decoder->lookup != NULL)
150 free(decoder->lookup);
151 if (decoder->huffnode != NULL)
152 free(decoder->huffnode);
153 free(decoder);
154 }
155}
156
9c659ffe 157/*-------------------------------------------------
158 * decode_one - decode a single code from the
159 * huffman stream
160 *-------------------------------------------------
161 */
162
163uint32_t huffman_decode_one(struct huffman_decoder* decoder, struct bitstream* bitbuf)
164{
165 /* peek ahead to get maxbits worth of data */
166 uint32_t bits = bitstream_peek(bitbuf, decoder->maxbits);
167
168 /* look it up, then remove the actual number of bits for this code */
169 lookup_value lookup = decoder->lookup[bits];
170 bitstream_remove(bitbuf, lookup & 0x1f);
171
172 /* return the value */
173 return lookup >> 5;
174}
175
176/*-------------------------------------------------
177 * import_tree_rle - import an RLE-encoded
178 * huffman tree from a source data stream
179 *-------------------------------------------------
180 */
181
182enum huffman_error huffman_import_tree_rle(struct huffman_decoder* decoder, struct bitstream* bitbuf)
183{
dcd8e4fd 184 int numbits, curnode;
185 enum huffman_error error;
186
9c659ffe 187 /* bits per entry depends on the maxbits */
9c659ffe 188 if (decoder->maxbits >= 16)
189 numbits = 5;
190 else if (decoder->maxbits >= 8)
191 numbits = 4;
192 else
193 numbits = 3;
194
195 /* loop until we read all the nodes */
9c659ffe 196 for (curnode = 0; curnode < decoder->numcodes; )
197 {
198 /* a non-one value is just raw */
199 int nodebits = bitstream_read(bitbuf, numbits);
200 if (nodebits != 1)
201 decoder->huffnode[curnode++].numbits = nodebits;
202
203 /* a one value is an escape code */
204 else
205 {
206 /* a double 1 is just a single 1 */
207 nodebits = bitstream_read(bitbuf, numbits);
208 if (nodebits == 1)
209 decoder->huffnode[curnode++].numbits = nodebits;
210
211 /* otherwise, we need one for value for the repeat count */
212 else
213 {
214 int repcount = bitstream_read(bitbuf, numbits) + 3;
215 while (repcount--)
216 decoder->huffnode[curnode++].numbits = nodebits;
217 }
218 }
219 }
220
221 /* make sure we ended up with the right number */
222 if (curnode != decoder->numcodes)
223 return HUFFERR_INVALID_DATA;
224
225 /* assign canonical codes for all nodes based on their code lengths */
dcd8e4fd 226 error = huffman_assign_canonical_codes(decoder);
9c659ffe 227 if (error != HUFFERR_NONE)
228 return error;
229
230 /* build the lookup table */
231 huffman_build_lookup_table(decoder);
232
233 /* determine final input length and report errors */
234 return bitstream_overflow(bitbuf) ? HUFFERR_INPUT_BUFFER_TOO_SMALL : HUFFERR_NONE;
235}
236
237
238/*-------------------------------------------------
239 * import_tree_huffman - import a huffman-encoded
240 * huffman tree from a source data stream
241 *-------------------------------------------------
242 */
243
244enum huffman_error huffman_import_tree_huffman(struct huffman_decoder* decoder, struct bitstream* bitbuf)
245{
dcd8e4fd 246 int start;
247 int last = 0;
248 int count = 0;
249 int index;
250 int curcode;
251 uint8_t rlefullbits = 0;
252 uint32_t temp;
253 enum huffman_error error;
9c659ffe 254 /* start by parsing the lengths for the small tree */
255 struct huffman_decoder* smallhuff = create_huffman_decoder(24, 6);
256 smallhuff->huffnode[0].numbits = bitstream_read(bitbuf, 3);
dcd8e4fd 257 start = bitstream_read(bitbuf, 3) + 1;
258 for (index = 1; index < 24; index++)
9c659ffe 259 {
260 if (index < start || count == 7)
261 smallhuff->huffnode[index].numbits = 0;
262 else
263 {
264 count = bitstream_read(bitbuf, 3);
265 smallhuff->huffnode[index].numbits = (count == 7) ? 0 : count;
266 }
267 }
268
269 /* then regenerate the tree */
dcd8e4fd 270 error = huffman_assign_canonical_codes(smallhuff);
9c659ffe 271 if (error != HUFFERR_NONE)
272 return error;
273 huffman_build_lookup_table(smallhuff);
274
275 /* determine the maximum length of an RLE count */
dcd8e4fd 276 temp = decoder->numcodes - 9;
9c659ffe 277 while (temp != 0)
278 temp >>= 1, rlefullbits++;
279
280 /* now process the rest of the data */
9c659ffe 281 for (curcode = 0; curcode < decoder->numcodes; )
282 {
283 int value = huffman_decode_one(smallhuff, bitbuf);
284 if (value != 0)
285 decoder->huffnode[curcode++].numbits = last = value - 1;
286 else
287 {
288 int count = bitstream_read(bitbuf, 3) + 2;
289 if (count == 7+2)
290 count += bitstream_read(bitbuf, rlefullbits);
291 for ( ; count != 0 && curcode < decoder->numcodes; count--)
292 decoder->huffnode[curcode++].numbits = last;
293 }
294 }
295
296 /* make sure we ended up with the right number */
297 if (curcode != decoder->numcodes)
298 return HUFFERR_INVALID_DATA;
299
300 /* assign canonical codes for all nodes based on their code lengths */
301 error = huffman_assign_canonical_codes(decoder);
302 if (error != HUFFERR_NONE)
303 return error;
304
305 /* build the lookup table */
306 huffman_build_lookup_table(decoder);
307
308 /* determine final input length and report errors */
309 return bitstream_overflow(bitbuf) ? HUFFERR_INPUT_BUFFER_TOO_SMALL : HUFFERR_NONE;
310}
311
312/*-------------------------------------------------
313 * compute_tree_from_histo - common backend for
314 * computing a tree based on the data histogram
315 *-------------------------------------------------
316 */
317
318enum huffman_error huffman_compute_tree_from_histo(struct huffman_decoder* decoder)
319{
dcd8e4fd 320 int i;
321 uint32_t lowerweight;
322 uint32_t upperweight;
9c659ffe 323 /* compute the number of data items in the histogram */
324 uint32_t sdatacount = 0;
dcd8e4fd 325 for (i = 0; i < decoder->numcodes; i++)
9c659ffe 326 sdatacount += decoder->datahisto[i];
327
328 /* binary search to achieve the optimum encoding */
dcd8e4fd 329 lowerweight = 0;
330 upperweight = sdatacount * 2;
9c659ffe 331 while (1)
332 {
333 /* build a tree using the current weight */
334 uint32_t curweight = (upperweight + lowerweight) / 2;
335 int curmaxbits = huffman_build_tree(decoder, sdatacount, curweight);
336
337 /* apply binary search here */
338 if (curmaxbits <= decoder->maxbits)
339 {
340 lowerweight = curweight;
341
342 /* early out if it worked with the raw weights, or if we're done searching */
343 if (curweight == sdatacount || (upperweight - lowerweight) <= 1)
344 break;
345 }
346 else
347 upperweight = curweight;
348 }
349
350 /* assign canonical codes for all nodes based on their code lengths */
351 return huffman_assign_canonical_codes(decoder);
352}
353
354/***************************************************************************
355 * INTERNAL FUNCTIONS
356 ***************************************************************************
357 */
358
359/*-------------------------------------------------
360 * tree_node_compare - compare two tree nodes
361 * by weight
362 *-------------------------------------------------
363 */
364
365static int huffman_tree_node_compare(const void *item1, const void *item2)
366{
367 const struct node_t *node1 = *(const struct node_t **)item1;
368 const struct node_t *node2 = *(const struct node_t **)item2;
369 if (node2->weight != node1->weight)
370 return node2->weight - node1->weight;
371 if (node2->bits - node1->bits == 0)
372 fprintf(stderr, "identical node sort keys, should not happen!\n");
373 return (int)node1->bits - (int)node2->bits;
374}
375
376/*-------------------------------------------------
377 * build_tree - build a huffman tree based on the
378 * data distribution
379 *-------------------------------------------------
380 */
381
382int huffman_build_tree(struct huffman_decoder* decoder, uint32_t totaldata, uint32_t totalweight)
383{
dcd8e4fd 384 int curcode;
385 int nextalloc;
386 int listitems = 0;
387 int maxbits = 0;
9c659ffe 388 /* make a list of all non-zero nodes */
389 struct node_t** list = (struct node_t**)malloc(sizeof(struct node_t*) * decoder->numcodes * 2);
9c659ffe 390 memset(decoder->huffnode, 0, decoder->numcodes * sizeof(decoder->huffnode[0]));
dcd8e4fd 391 for (curcode = 0; curcode < decoder->numcodes; curcode++)
9c659ffe 392 if (decoder->datahisto[curcode] != 0)
393 {
394 list[listitems++] = &decoder->huffnode[curcode];
395 decoder->huffnode[curcode].count = decoder->datahisto[curcode];
396 decoder->huffnode[curcode].bits = curcode;
397
398 /* scale the weight by the current effective length, ensuring we don't go to 0 */
399 decoder->huffnode[curcode].weight = ((uint64_t)decoder->datahisto[curcode]) * ((uint64_t)totalweight) / ((uint64_t)totaldata);
400 if (decoder->huffnode[curcode].weight == 0)
401 decoder->huffnode[curcode].weight = 1;
402 }
403
404#if 0
405 fprintf(stderr, "Pre-sort:\n");
406 for (int i = 0; i < listitems; i++) {
407 fprintf(stderr, "weight: %d code: %d\n", list[i]->m_weight, list[i]->m_bits);
408 }
409#endif
410
411 /* sort the list by weight, largest weight first */
412 qsort(&list[0], listitems, sizeof(list[0]), huffman_tree_node_compare);
413
414#if 0
415 fprintf(stderr, "Post-sort:\n");
416 for (int i = 0; i < listitems; i++) {
417 fprintf(stderr, "weight: %d code: %d\n", list[i]->m_weight, list[i]->m_bits);
418 }
419 fprintf(stderr, "===================\n");
420#endif
421
422 /* now build the tree */
dcd8e4fd 423 nextalloc = decoder->numcodes;
9c659ffe 424 while (listitems > 1)
425 {
2ff0b512 426 int curitem;
9c659ffe 427 /* remove lowest two items */
428 struct node_t* node1 = &(*list[--listitems]);
429 struct node_t* node0 = &(*list[--listitems]);
430
431 /* create new node */
432 struct node_t* newnode = &decoder->huffnode[nextalloc++];
433 newnode->parent = NULL;
434 node0->parent = node1->parent = newnode;
435 newnode->weight = node0->weight + node1->weight;
436
437 /* insert into list at appropriate location */
9c659ffe 438 for (curitem = 0; curitem < listitems; curitem++)
439 if (newnode->weight > list[curitem]->weight)
440 {
441 memmove(&list[curitem+1], &list[curitem], (listitems - curitem) * sizeof(list[0]));
442 break;
443 }
444 list[curitem] = newnode;
445 listitems++;
446 }
447
448 /* compute the number of bits in each code, and fill in another histogram */
dcd8e4fd 449 for (curcode = 0; curcode < decoder->numcodes; curcode++)
9c659ffe 450 {
2ff0b512 451 struct node_t *curnode;
9c659ffe 452 struct node_t* node = &decoder->huffnode[curcode];
453 node->numbits = 0;
454 node->bits = 0;
455
456 /* if we have a non-zero weight, compute the number of bits */
457 if (node->weight > 0)
458 {
459 /* determine the number of bits for this node */
2ff0b512 460 for (curnode = node; curnode->parent != NULL; curnode = curnode->parent)
9c659ffe 461 node->numbits++;
462 if (node->numbits == 0)
463 node->numbits = 1;
464
465 /* keep track of the max */
466 maxbits = MAX(maxbits, ((int)node->numbits));
467 }
468 }
469 return maxbits;
470}
471
472/*-------------------------------------------------
473 * assign_canonical_codes - assign canonical codes
474 * to all the nodes based on the number of bits
475 * in each
476 *-------------------------------------------------
477 */
478
479enum huffman_error huffman_assign_canonical_codes(struct huffman_decoder* decoder)
480{
dcd8e4fd 481 int curcode, codelen;
482 uint32_t curstart = 0;
9c659ffe 483 /* build up a histogram of bit lengths */
484 uint32_t bithisto[33] = { 0 };
dcd8e4fd 485 for (curcode = 0; curcode < decoder->numcodes; curcode++)
9c659ffe 486 {
487 struct node_t* node = &decoder->huffnode[curcode];
488 if (node->numbits > decoder->maxbits)
489 return HUFFERR_INTERNAL_INCONSISTENCY;
490 if (node->numbits <= 32)
491 bithisto[node->numbits]++;
492 }
493
494 /* for each code length, determine the starting code number */
dcd8e4fd 495 for (codelen = 32; codelen > 0; codelen--)
9c659ffe 496 {
497 uint32_t nextstart = (curstart + bithisto[codelen]) >> 1;
498 if (codelen != 1 && nextstart * 2 != (curstart + bithisto[codelen]))
499 return HUFFERR_INTERNAL_INCONSISTENCY;
500 bithisto[codelen] = curstart;
501 curstart = nextstart;
502 }
503
504 /* now assign canonical codes */
dcd8e4fd 505 for (curcode = 0; curcode < decoder->numcodes; curcode++)
9c659ffe 506 {
507 struct node_t* node = &decoder->huffnode[curcode];
508 if (node->numbits > 0)
509 node->bits = bithisto[node->numbits]++;
510 }
511 return HUFFERR_NONE;
512}
513
514/*-------------------------------------------------
515 * build_lookup_table - build a lookup table for
516 * fast decoding
517 *-------------------------------------------------
518 */
519
520void huffman_build_lookup_table(struct huffman_decoder* decoder)
521{
dcd8e4fd 522 int curcode;
9c659ffe 523 /* iterate over all codes */
dcd8e4fd 524 for (curcode = 0; curcode < decoder->numcodes; curcode++)
9c659ffe 525 {
526 /* process all nodes which have non-zero bits */
527 struct node_t* node = &decoder->huffnode[curcode];
528 if (node->numbits > 0)
529 {
2ff0b512 530 int shift;
531 lookup_value *dest;
532 lookup_value *destend;
9c659ffe 533 /* set up the entry */
534 lookup_value value = MAKE_LOOKUP(curcode, node->numbits);
535
536 /* fill all matching entries */
2ff0b512 537 shift = decoder->maxbits - node->numbits;
538 dest = &decoder->lookup[node->bits << shift];
539 destend = &decoder->lookup[((node->bits + 1) << shift) - 1];
9c659ffe 540 while (dest <= destend)
541 *dest++ = value;
542 }
543 }
544}