4 * Copyright (C) 1991-1997, Thomas G. Lane.
5 * This file is part of the Independent JPEG Group's software.
6 * For conditions of distribution and use, see the accompanying README file.
8 * This file contains Huffman entropy decoding routines.
10 * Much of the complexity here has to do with supporting input suspension.
11 * If the data source module demands suspension, we want to be able to back
12 * up to the start of the current MCU. To do this, we copy state variables
13 * into local working storage, and update them back to the permanent
14 * storage only upon successful completion of an MCU.
17 #define JPEG_INTERNALS
20 #include "jdhuff.h" /* Declarations shared with jdphuff.c */
24 * Expanded entropy decoder object for Huffman decoding.
26 * The savable_state subrecord contains fields that change within an MCU,
27 * but must not be updated permanently until we complete the MCU.
31 int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
34 /* This macro is to work around compilers with missing or broken
35 * structure assignment. You'll need to fix this code if you have
36 * such a compiler and you change MAX_COMPS_IN_SCAN.
39 #ifndef NO_STRUCT_ASSIGN
40 #define ASSIGN_STATE(dest,src) ((dest) = (src))
42 #if MAX_COMPS_IN_SCAN == 4
43 #define ASSIGN_STATE(dest,src) \
44 ((dest).last_dc_val[0] = (src).last_dc_val[0], \
45 (dest).last_dc_val[1] = (src).last_dc_val[1], \
46 (dest).last_dc_val[2] = (src).last_dc_val[2], \
47 (dest).last_dc_val[3] = (src).last_dc_val[3])
53 struct jpeg_entropy_decoder pub; /* public fields */
55 /* These fields are loaded into local variables at start of each MCU.
56 * In case of suspension, we exit WITHOUT updating them.
58 bitread_perm_state bitstate; /* Bit buffer at start of MCU */
59 savable_state saved; /* Other state at start of MCU */
61 /* These fields are NOT loaded into local working state. */
62 unsigned int restarts_to_go; /* MCUs left in this restart interval */
64 /* Pointers to derived tables (these workspaces have image lifespan) */
65 d_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
66 d_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
68 /* Precalculated info set up by start_pass for use in decode_mcu: */
70 /* Pointers to derived tables to be used for each block within an MCU */
71 d_derived_tbl * dc_cur_tbls[D_MAX_BLOCKS_IN_MCU];
72 d_derived_tbl * ac_cur_tbls[D_MAX_BLOCKS_IN_MCU];
73 /* Whether we care about the DC and AC coefficient values for each block */
74 boolean dc_needed[D_MAX_BLOCKS_IN_MCU];
75 boolean ac_needed[D_MAX_BLOCKS_IN_MCU];
76 } huff_entropy_decoder;
78 typedef huff_entropy_decoder * huff_entropy_ptr;
82 * Initialize for a Huffman-compressed scan.
86 start_pass_huff_decoder (j_decompress_ptr cinfo)
88 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
89 int ci, blkn, dctbl, actbl;
90 jpeg_component_info * compptr;
92 /* Check that the scan parameters Ss, Se, Ah/Al are OK for sequential JPEG.
93 * This ought to be an error condition, but we make it a warning because
94 * there are some baseline files out there with all zeroes in these bytes.
96 if (cinfo->Ss != 0 || cinfo->Se != DCTSIZE2-1 ||
97 cinfo->Ah != 0 || cinfo->Al != 0)
98 WARNMS(cinfo, JWRN_NOT_SEQUENTIAL);
100 for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
101 compptr = cinfo->cur_comp_info[ci];
102 dctbl = compptr->dc_tbl_no;
103 actbl = compptr->ac_tbl_no;
104 /* Compute derived values for Huffman tables */
105 /* We may do this more than once for a table, but it's not expensive */
106 jpeg_make_d_derived_tbl(cinfo, TRUE, dctbl,
107 & entropy->dc_derived_tbls[dctbl]);
108 jpeg_make_d_derived_tbl(cinfo, FALSE, actbl,
109 & entropy->ac_derived_tbls[actbl]);
110 /* Initialize DC predictions to 0 */
111 entropy->saved.last_dc_val[ci] = 0;
114 /* Precalculate decoding info for each block in an MCU of this scan */
115 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
116 ci = cinfo->MCU_membership[blkn];
117 compptr = cinfo->cur_comp_info[ci];
118 /* Precalculate which table to use for each block */
119 entropy->dc_cur_tbls[blkn] = entropy->dc_derived_tbls[compptr->dc_tbl_no];
120 entropy->ac_cur_tbls[blkn] = entropy->ac_derived_tbls[compptr->ac_tbl_no];
121 /* Decide whether we really care about the coefficient values */
122 if (compptr->component_needed) {
123 entropy->dc_needed[blkn] = TRUE;
124 /* we don't need the ACs if producing a 1/8th-size image */
125 entropy->ac_needed[blkn] = (compptr->DCT_scaled_size > 1);
127 entropy->dc_needed[blkn] = entropy->ac_needed[blkn] = FALSE;
131 /* Initialize bitread state variables */
132 entropy->bitstate.bits_left = 0;
133 entropy->bitstate.get_buffer = 0; /* unnecessary, but keeps Purify quiet */
134 entropy->pub.insufficient_data = FALSE;
136 /* Initialize restart counter */
137 entropy->restarts_to_go = cinfo->restart_interval;
142 * Compute the derived values for a Huffman table.
143 * This routine also performs some validation checks on the table.
145 * Note this is also used by jdphuff.c.
149 jpeg_make_d_derived_tbl (j_decompress_ptr cinfo, boolean isDC, int tblno,
150 d_derived_tbl ** pdtbl)
154 int p, i, l, si, numsymbols;
157 unsigned int huffcode[257];
160 /* Note that huffsize[] and huffcode[] are filled in code-length order,
161 * paralleling the order of the symbols themselves in htbl->huffval[].
164 /* Find the input Huffman table */
165 if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
166 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
168 isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
170 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
172 /* Allocate a workspace if we haven't already done so. */
174 *pdtbl = (d_derived_tbl *)
175 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
176 SIZEOF(d_derived_tbl));
178 dtbl->pub = htbl; /* fill in back link */
180 /* Figure C.1: make table of Huffman code length for each symbol */
183 for (l = 1; l <= 16; l++) {
184 i = (int) htbl->bits[l];
185 if (i < 0 || p + i > 256) /* protect against table overrun */
186 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
188 huffsize[p++] = (char) l;
193 /* Figure C.2: generate the codes themselves */
194 /* We also validate that the counts represent a legal Huffman code tree. */
199 while (huffsize[p]) {
200 while (((int) huffsize[p]) == si) {
201 huffcode[p++] = code;
204 /* code is now 1 more than the last code used for codelength si; but
205 * it must still fit in si bits, since no code is allowed to be all ones.
207 if (((INT32) code) >= (((INT32) 1) << si))
208 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
213 /* Figure F.15: generate decoding tables for bit-sequential decoding */
216 for (l = 1; l <= 16; l++) {
218 /* valoffset[l] = huffval[] index of 1st symbol of code length l,
219 * minus the minimum code of length l
221 dtbl->valoffset[l] = (INT32) p - (INT32) huffcode[p];
223 dtbl->maxcode[l] = huffcode[p-1]; /* maximum code of length l */
225 dtbl->maxcode[l] = -1; /* -1 if no codes of this length */
228 dtbl->maxcode[17] = 0xFFFFFL; /* ensures jpeg_huff_decode terminates */
230 /* Compute lookahead tables to speed up decoding.
231 * First we set all the table entries to 0, indicating "too long";
232 * then we iterate through the Huffman codes that are short enough and
233 * fill in all the entries that correspond to bit sequences starting
237 MEMZERO(dtbl->look_nbits, SIZEOF(dtbl->look_nbits));
240 for (l = 1; l <= HUFF_LOOKAHEAD; l++) {
241 for (i = 1; i <= (int) htbl->bits[l]; i++, p++) {
242 /* l = current code's length, p = its index in huffcode[] & huffval[]. */
243 /* Generate left-justified code followed by all possible bit sequences */
244 lookbits = huffcode[p] << (HUFF_LOOKAHEAD-l);
245 for (ctr = 1 << (HUFF_LOOKAHEAD-l); ctr > 0; ctr--) {
246 dtbl->look_nbits[lookbits] = l;
247 dtbl->look_sym[lookbits] = htbl->huffval[p];
253 /* Validate symbols as being reasonable.
254 * For AC tables, we make no check, but accept all byte values 0..255.
255 * For DC tables, we require the symbols to be in range 0..15.
256 * (Tighter bounds could be applied depending on the data depth and mode,
257 * but this is sufficient to ensure safe decoding.)
260 for (i = 0; i < numsymbols; i++) {
261 int sym = htbl->huffval[i];
262 if (sym < 0 || sym > 15)
263 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
270 * Out-of-line code for bit fetching (shared with jdphuff.c).
271 * See jdhuff.h for info about usage.
272 * Note: current values of get_buffer and bits_left are passed as parameters,
273 * but are returned in the corresponding fields of the state struct.
275 * On most machines MIN_GET_BITS should be 25 to allow the full 32-bit width
276 * of get_buffer to be used. (On machines with wider words, an even larger
277 * buffer could be used.) However, on some machines 32-bit shifts are
278 * quite slow and take time proportional to the number of places shifted.
279 * (This is true with most PC compilers, for instance.) In this case it may
280 * be a win to set MIN_GET_BITS to the minimum value of 15. This reduces the
281 * average shift distance at the cost of more calls to jpeg_fill_bit_buffer.
285 #define MIN_GET_BITS 15 /* minimum allowable value */
287 #define MIN_GET_BITS (BIT_BUF_SIZE-7)
292 jpeg_fill_bit_buffer (bitread_working_state * state,
293 register bit_buf_type get_buffer, register int bits_left,
295 /* Load up the bit buffer to a depth of at least nbits */
297 /* Copy heavily used state fields into locals (hopefully registers) */
298 register const JOCTET * next_input_byte = state->next_input_byte;
299 register size_t bytes_in_buffer = state->bytes_in_buffer;
300 j_decompress_ptr cinfo = state->cinfo;
302 /* Attempt to load at least MIN_GET_BITS bits into get_buffer. */
303 /* (It is assumed that no request will be for more than that many bits.) */
304 /* We fail to do so only if we hit a marker or are forced to suspend. */
306 if (cinfo->unread_marker == 0) { /* cannot advance past a marker */
307 while (bits_left < MIN_GET_BITS) {
310 /* Attempt to read a byte */
311 if (bytes_in_buffer == 0) {
312 if (! (*cinfo->src->fill_input_buffer) (cinfo))
314 next_input_byte = cinfo->src->next_input_byte;
315 bytes_in_buffer = cinfo->src->bytes_in_buffer;
318 c = GETJOCTET(*next_input_byte++);
320 /* If it's 0xFF, check and discard stuffed zero byte */
322 /* Loop here to discard any padding FF's on terminating marker,
323 * so that we can save a valid unread_marker value. NOTE: we will
324 * accept multiple FF's followed by a 0 as meaning a single FF data
325 * byte. This data pattern is not valid according to the standard.
328 if (bytes_in_buffer == 0) {
329 if (! (*cinfo->src->fill_input_buffer) (cinfo))
331 next_input_byte = cinfo->src->next_input_byte;
332 bytes_in_buffer = cinfo->src->bytes_in_buffer;
335 c = GETJOCTET(*next_input_byte++);
339 /* Found FF/00, which represents an FF data byte */
342 /* Oops, it's actually a marker indicating end of compressed data.
343 * Save the marker code for later use.
344 * Fine point: it might appear that we should save the marker into
345 * bitread working state, not straight into permanent state. But
346 * once we have hit a marker, we cannot need to suspend within the
347 * current MCU, because we will read no more bytes from the data
348 * source. So it is OK to update permanent state right away.
350 cinfo->unread_marker = c;
351 /* See if we need to insert some fake zero bits. */
356 /* OK, load c into get_buffer */
357 get_buffer = (get_buffer << 8) | c;
362 /* We get here if we've read the marker that terminates the compressed
363 * data segment. There should be enough bits in the buffer register
364 * to satisfy the request; if so, no problem.
366 if (nbits > bits_left) {
367 /* Uh-oh. Report corrupted data to user and stuff zeroes into
368 * the data stream, so that we can produce some kind of image.
369 * We use a nonvolatile flag to ensure that only one warning message
370 * appears per data segment.
372 if (! cinfo->entropy->insufficient_data) {
373 WARNMS(cinfo, JWRN_HIT_MARKER);
374 cinfo->entropy->insufficient_data = TRUE;
376 /* Fill the buffer with zero bits */
377 get_buffer <<= MIN_GET_BITS - bits_left;
378 bits_left = MIN_GET_BITS;
382 /* Unload the local registers */
383 state->next_input_byte = next_input_byte;
384 state->bytes_in_buffer = bytes_in_buffer;
385 state->get_buffer = get_buffer;
386 state->bits_left = bits_left;
393 * Out-of-line code for Huffman code decoding.
394 * See jdhuff.h for info about usage.
398 jpeg_huff_decode (bitread_working_state * state,
399 register bit_buf_type get_buffer, register int bits_left,
400 d_derived_tbl * htbl, int min_bits)
402 register int l = min_bits;
405 /* HUFF_DECODE has determined that the code is at least min_bits */
406 /* bits long, so fetch that many bits in one swoop. */
408 CHECK_BIT_BUFFER(*state, l, return -1);
411 /* Collect the rest of the Huffman code one bit at a time. */
412 /* This is per Figure F.16 in the JPEG spec. */
414 while (code > htbl->maxcode[l]) {
416 CHECK_BIT_BUFFER(*state, 1, return -1);
421 /* Unload the local registers */
422 state->get_buffer = get_buffer;
423 state->bits_left = bits_left;
425 /* With garbage input we may reach the sentinel value l = 17. */
428 WARNMS(state->cinfo, JWRN_HUFF_BAD_CODE);
429 return 0; /* fake a zero as the safest result */
432 return htbl->pub->huffval[ (int) (code + htbl->valoffset[l]) ];
437 * Figure F.12: extend sign bit.
438 * On some machines, a shift and add will be faster than a table lookup.
443 #define HUFF_EXTEND(x,s) ((x) < (1<<((s)-1)) ? (x) + (((-1)<<(s)) + 1) : (x))
447 #define HUFF_EXTEND(x,s) ((x) < extend_test[s] ? (x) + extend_offset[s] : (x))
449 static const int extend_test[16] = /* entry n is 2**(n-1) */
450 { 0, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
451 0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000 };
453 static const int extend_offset[16] = /* entry n is (-1 << n) + 1 */
454 { 0, ((-1)<<1) + 1, ((-1)<<2) + 1, ((-1)<<3) + 1, ((-1)<<4) + 1,
455 ((-1)<<5) + 1, ((-1)<<6) + 1, ((-1)<<7) + 1, ((-1)<<8) + 1,
456 ((-1)<<9) + 1, ((-1)<<10) + 1, ((-1)<<11) + 1, ((-1)<<12) + 1,
457 ((-1)<<13) + 1, ((-1)<<14) + 1, ((-1)<<15) + 1 };
459 #endif /* AVOID_TABLES */
463 * Check for a restart marker & resynchronize decoder.
464 * Returns FALSE if must suspend.
468 process_restart (j_decompress_ptr cinfo)
470 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
473 /* Throw away any unused bits remaining in bit buffer; */
474 /* include any full bytes in next_marker's count of discarded bytes */
475 cinfo->marker->discarded_bytes += entropy->bitstate.bits_left / 8;
476 entropy->bitstate.bits_left = 0;
478 /* Advance past the RSTn marker */
479 if (! (*cinfo->marker->read_restart_marker) (cinfo))
482 /* Re-initialize DC predictions to 0 */
483 for (ci = 0; ci < cinfo->comps_in_scan; ci++)
484 entropy->saved.last_dc_val[ci] = 0;
486 /* Reset restart counter */
487 entropy->restarts_to_go = cinfo->restart_interval;
489 /* Reset out-of-data flag, unless read_restart_marker left us smack up
490 * against a marker. In that case we will end up treating the next data
491 * segment as empty, and we can avoid producing bogus output pixels by
492 * leaving the flag set.
494 if (cinfo->unread_marker == 0)
495 entropy->pub.insufficient_data = FALSE;
502 * Decode and return one MCU's worth of Huffman-compressed coefficients.
503 * The coefficients are reordered from zigzag order into natural array order,
504 * but are not dequantized.
506 * The i'th block of the MCU is stored into the block pointed to by
507 * MCU_data[i]. WE ASSUME THIS AREA HAS BEEN ZEROED BY THE CALLER.
508 * (Wholesale zeroing is usually a little faster than retail...)
510 * Returns FALSE if data source requested suspension. In that case no
511 * changes have been made to permanent state. (Exception: some output
512 * coefficients may already have been assigned. This is harmless for
513 * this module, since we'll just re-assign them on the next call.)
517 decode_mcu (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
519 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
524 /* Process restart marker if needed; may have to suspend */
525 if (cinfo->restart_interval) {
526 if (entropy->restarts_to_go == 0)
527 if (! process_restart(cinfo))
531 /* If we've run out of data, just leave the MCU set to zeroes.
532 * This way, we return uniform gray for the remainder of the segment.
534 if (! entropy->pub.insufficient_data) {
536 /* Load up working state */
537 BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
538 ASSIGN_STATE(state, entropy->saved);
540 /* Outer loop handles each block in the MCU */
542 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
543 JBLOCKROW block = MCU_data[blkn];
544 d_derived_tbl * dctbl = entropy->dc_cur_tbls[blkn];
545 d_derived_tbl * actbl = entropy->ac_cur_tbls[blkn];
546 register int s, k, r;
548 /* Decode a single block's worth of coefficients */
550 /* Section F.2.2.1: decode the DC coefficient difference */
551 HUFF_DECODE(s, br_state, dctbl, return FALSE, label1);
553 CHECK_BIT_BUFFER(br_state, s, return FALSE);
555 s = HUFF_EXTEND(r, s);
558 if (entropy->dc_needed[blkn]) {
559 /* Convert DC difference to actual value, update last_dc_val */
560 int ci = cinfo->MCU_membership[blkn];
561 s += state.last_dc_val[ci];
562 state.last_dc_val[ci] = s;
563 /* Output the DC coefficient (assumes jpeg_natural_order[0] = 0) */
564 (*block)[0] = (JCOEF) s;
567 if (entropy->ac_needed[blkn]) {
569 /* Section F.2.2.2: decode the AC coefficients */
570 /* Since zeroes are skipped, output area must be cleared beforehand */
571 for (k = 1; k < DCTSIZE2; k++) {
572 HUFF_DECODE(s, br_state, actbl, return FALSE, label2);
579 CHECK_BIT_BUFFER(br_state, s, return FALSE);
581 s = HUFF_EXTEND(r, s);
582 /* Output coefficient in natural (dezigzagged) order.
583 * Note: the extra entries in jpeg_natural_order[] will save us
584 * if k >= DCTSIZE2, which could happen if the data is corrupted.
586 (*block)[jpeg_natural_order[k]] = (JCOEF) s;
596 /* Section F.2.2.2: decode the AC coefficients */
597 /* In this path we just discard the values */
598 for (k = 1; k < DCTSIZE2; k++) {
599 HUFF_DECODE(s, br_state, actbl, return FALSE, label3);
606 CHECK_BIT_BUFFER(br_state, s, return FALSE);
618 /* Completed MCU, so update state */
619 BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
620 ASSIGN_STATE(entropy->saved, state);
623 /* Account for restart interval (no-op if not using restarts) */
624 entropy->restarts_to_go--;
631 * Module initialization routine for Huffman entropy decoding.
635 jinit_huff_decoder (j_decompress_ptr cinfo)
637 huff_entropy_ptr entropy;
640 entropy = (huff_entropy_ptr)
641 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
642 SIZEOF(huff_entropy_decoder));
643 cinfo->entropy = (struct jpeg_entropy_decoder *) entropy;
644 entropy->pub.start_pass = start_pass_huff_decoder;
645 entropy->pub.decode_mcu = decode_mcu;
647 /* Mark tables unallocated */
648 for (i = 0; i < NUM_HUFF_TBLS; i++) {
649 entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;