6 * Copyright (C) 2013 Eric Biggers
8 * This file is part of wimlib, a library for working with WIM files.
10 * wimlib is free software; you can redistribute it and/or modify it under the
11 * terms of the GNU General Public License as published by the Free
12 * Software Foundation; either version 3 of the License, or (at your option)
15 * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
16 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
17 * A PARTICULAR PURPOSE. See the GNU General Public License for more
20 * You should have received a copy of the GNU General Public License
21 * along with wimlib; if not, see http://www.gnu.org/licenses/.
24 /* This a compressor for the LZMS compression format. More details about this
25 * format can be found in lzms-decompress.c.
27 * This is currently an unsophisticated implementation that is fast but does not
28 * attain the best compression ratios allowed by the format.
36 #include "wimlib/compiler.h"
37 #include "wimlib/compressor_ops.h"
38 #include "wimlib/compress_common.h"
39 #include "wimlib/endianness.h"
40 #include "wimlib/error.h"
41 #include "wimlib/lz_hash.h"
42 #include "wimlib/lz_sarray.h"
43 #include "wimlib/lzms.h"
44 #include "wimlib/util.h"
50 #define LZMS_OPTIM_ARRAY_SIZE 1024
52 struct lzms_compressor;
53 struct lzms_adaptive_state {
54 struct lzms_lz_lru_queues lru;
58 u8 lz_repeat_match_state[LZMS_NUM_RECENT_OFFSETS - 1];
60 #define LZ_ADAPTIVE_STATE struct lzms_adaptive_state
61 #define LZ_COMPRESSOR struct lzms_compressor
62 #include "wimlib/lz_optimal.h"
64 /* Stucture used for writing raw bits to the end of the LZMS-compressed data as
65 * a series of 16-bit little endian coding units. */
66 struct lzms_output_bitstream {
67 /* Buffer variable containing zero or more bits that have been logically
68 * written to the bitstream but not yet written to memory. This must be
69 * at least as large as the coding unit size. */
72 /* Number of bits in @bitbuf that are valid. */
73 unsigned num_free_bits;
75 /* Pointer to one past the next position in the compressed data buffer
76 * at which to output a 16-bit coding unit. */
79 /* Maximum number of 16-bit coding units that can still be output to
80 * the compressed data buffer. */
81 size_t num_le16_remaining;
83 /* Set to %true if not all coding units could be output due to
84 * insufficient space. */
88 /* Stucture used for range encoding (raw version). */
89 struct lzms_range_encoder_raw {
91 /* A 33-bit variable that holds the low boundary of the current range.
92 * The 33rd bit is needed to catch carries. */
95 /* Size of the current range. */
98 /* Next 16-bit coding unit to output. */
101 /* Number of 16-bit coding units whose output has been delayed due to
102 * possible carrying. The first such coding unit is @cache; all
103 * subsequent such coding units are 0xffff. */
106 /* Pointer to the next position in the compressed data buffer at which
107 * to output a 16-bit coding unit. */
110 /* Maximum number of 16-bit coding units that can still be output to
111 * the compressed data buffer. */
112 size_t num_le16_remaining;
114 /* %true when the very first coding unit has not yet been output. */
117 /* Set to %true if not all coding units could be output due to
118 * insufficient space. */
122 /* Structure used for range encoding. This wraps around `struct
123 * lzms_range_encoder_raw' to use and maintain probability entries. */
124 struct lzms_range_encoder {
125 /* Pointer to the raw range encoder, which has no persistent knowledge
126 * of probabilities. Multiple lzms_range_encoder's share the same
127 * lzms_range_encoder_raw. */
128 struct lzms_range_encoder_raw *rc;
130 /* Bits recently encoded by this range encoder. This are used as in
131 * index into @prob_entries. */
134 /* Bitmask for @state to prevent its value from exceeding the number of
135 * probability entries. */
138 /* Probability entries being used for this range encoder. */
139 struct lzms_probability_entry prob_entries[LZMS_MAX_NUM_STATES];
142 /* Structure used for Huffman encoding. */
143 struct lzms_huffman_encoder {
145 /* Bitstream to write Huffman-encoded symbols and verbatim bits to.
146 * Multiple lzms_huffman_encoder's share the same lzms_output_bitstream.
148 struct lzms_output_bitstream *os;
150 /* Number of symbols that have been written using this code far. Reset
151 * to 0 whenever the code is rebuilt. */
152 u32 num_syms_written;
154 /* When @num_syms_written reaches this number, the Huffman code must be
158 /* Number of symbols in the represented Huffman code. */
161 /* Running totals of symbol frequencies. These are diluted slightly
162 * whenever the code is rebuilt. */
163 u32 sym_freqs[LZMS_MAX_NUM_SYMS];
165 /* The length, in bits, of each symbol in the Huffman code. */
166 u8 lens[LZMS_MAX_NUM_SYMS];
168 /* The codeword of each symbol in the Huffman code. */
169 u16 codewords[LZMS_MAX_NUM_SYMS];
172 /* State of the LZMS compressor. */
173 struct lzms_compressor {
174 /* Pointer to a buffer holding the preprocessed data to compress. */
177 /* Current position in @buffer. */
180 /* Size of the data in @buffer. */
184 /* Temporary array used by lz_analyze_block(); must be at least as long
189 /* Suffix array match-finder. */
190 struct lz_sarray lz_sarray;
192 /* Temporary space to store found matches. */
193 struct raw_match *matches;
196 struct lz_match_chooser mc;
198 /* Maximum block size this compressor instantiation allows. This is the
199 * allocated size of @window. */
202 /* Raw range encoder which outputs to the beginning of the compressed
203 * data buffer, proceeding forwards. */
204 struct lzms_range_encoder_raw rc;
206 /* Bitstream which outputs to the end of the compressed data buffer,
207 * proceeding backwards. */
208 struct lzms_output_bitstream os;
210 /* Range encoders. */
211 struct lzms_range_encoder main_range_encoder;
212 struct lzms_range_encoder match_range_encoder;
213 struct lzms_range_encoder lz_match_range_encoder;
214 struct lzms_range_encoder lz_repeat_match_range_encoders[LZMS_NUM_RECENT_OFFSETS - 1];
215 struct lzms_range_encoder delta_match_range_encoder;
216 struct lzms_range_encoder delta_repeat_match_range_encoders[LZMS_NUM_RECENT_OFFSETS - 1];
218 /* Huffman encoders. */
219 struct lzms_huffman_encoder literal_encoder;
220 struct lzms_huffman_encoder lz_offset_encoder;
221 struct lzms_huffman_encoder length_encoder;
222 struct lzms_huffman_encoder delta_power_encoder;
223 struct lzms_huffman_encoder delta_offset_encoder;
225 /* LRU (least-recently-used) queues for match information. */
226 struct lzms_lru_queues lru;
228 /* Used for preprocessing. */
229 s32 last_target_usages[65536];
232 /* Initialize the output bitstream @os to write forwards to the specified
233 * compressed data buffer @out that is @out_limit 16-bit integers long. */
235 lzms_output_bitstream_init(struct lzms_output_bitstream *os,
236 le16 *out, size_t out_limit)
239 os->num_free_bits = 16;
240 os->out = out + out_limit;
241 os->num_le16_remaining = out_limit;
245 /* Write @num_bits bits, contained in the low @num_bits bits of @bits (ordered
246 * from high-order to low-order), to the output bitstream @os. */
248 lzms_output_bitstream_put_bits(struct lzms_output_bitstream *os,
249 u32 bits, unsigned num_bits)
251 bits &= (1U << num_bits) - 1;
253 while (num_bits > os->num_free_bits) {
255 if (unlikely(os->num_le16_remaining == 0)) {
260 unsigned num_fill_bits = os->num_free_bits;
262 os->bitbuf <<= num_fill_bits;
263 os->bitbuf |= bits >> (num_bits - num_fill_bits);
265 *--os->out = cpu_to_le16(os->bitbuf);
266 --os->num_le16_remaining;
268 os->num_free_bits = 16;
269 num_bits -= num_fill_bits;
270 bits &= (1U << num_bits) - 1;
272 os->bitbuf <<= num_bits;
274 os->num_free_bits -= num_bits;
277 /* Flush the output bitstream, ensuring that all bits written to it have been
278 * written to memory. Returns %true if all bits were output successfully, or
279 * %false if an overrun occurred. */
281 lzms_output_bitstream_flush(struct lzms_output_bitstream *os)
283 if (os->num_free_bits != 16)
284 lzms_output_bitstream_put_bits(os, 0, os->num_free_bits + 1);
288 /* Initialize the range encoder @rc to write forwards to the specified
289 * compressed data buffer @out that is @out_limit 16-bit integers long. */
291 lzms_range_encoder_raw_init(struct lzms_range_encoder_raw *rc,
292 le16 *out, size_t out_limit)
295 rc->range = 0xffffffff;
299 rc->num_le16_remaining = out_limit;
305 * Attempt to flush bits from the range encoder.
307 * Note: this is based on the public domain code for LZMA written by Igor
308 * Pavlov. The only differences in this function are that in LZMS the bits must
309 * be output in 16-bit coding units instead of 8-bit coding units, and that in
310 * LZMS the first coding unit is not ignored by the decompressor, so the encoder
311 * cannot output a dummy value to that position.
313 * The basic idea is that we're writing bits from @rc->low to the output.
314 * However, due to carrying, the writing of coding units with value 0xffff, as
315 * well as one prior coding unit, must be delayed until it is determined whether
319 lzms_range_encoder_raw_shift_low(struct lzms_range_encoder_raw *rc)
321 LZMS_DEBUG("low=%"PRIx64", cache=%"PRIx64", cache_size=%u",
322 rc->low, rc->cache, rc->cache_size);
323 if ((u32)(rc->low) < 0xffff0000 ||
324 (u32)(rc->low >> 32) != 0)
326 /* Carry not needed (rc->low < 0xffff0000), or carry occurred
327 * ((rc->low >> 32) != 0, a.k.a. the carry bit is 1). */
330 if (rc->num_le16_remaining == 0) {
334 *rc->out++ = cpu_to_le16(rc->cache +
335 (u16)(rc->low >> 32));
336 --rc->num_le16_remaining;
342 } while (--rc->cache_size != 0);
344 rc->cache = (rc->low >> 16) & 0xffff;
347 rc->low = (rc->low & 0xffff) << 16;
351 lzms_range_encoder_raw_normalize(struct lzms_range_encoder_raw *rc)
353 if (rc->range <= 0xffff) {
355 lzms_range_encoder_raw_shift_low(rc);
360 lzms_range_encoder_raw_flush(struct lzms_range_encoder_raw *rc)
362 for (unsigned i = 0; i < 4; i++)
363 lzms_range_encoder_raw_shift_low(rc);
367 /* Encode the next bit using the range encoder (raw version).
369 * @prob is the chance out of LZMS_PROBABILITY_MAX that the next bit is 0. */
371 lzms_range_encoder_raw_encode_bit(struct lzms_range_encoder_raw *rc, int bit,
374 lzms_range_encoder_raw_normalize(rc);
376 u32 bound = (rc->range >> LZMS_PROBABILITY_BITS) * prob;
385 /* Encode a bit using the specified range encoder. This wraps around
386 * lzms_range_encoder_raw_encode_bit() to handle using and updating the
387 * appropriate probability table. */
389 lzms_range_encode_bit(struct lzms_range_encoder *enc, int bit)
391 struct lzms_probability_entry *prob_entry;
394 /* Load the probability entry corresponding to the current state. */
395 prob_entry = &enc->prob_entries[enc->state];
397 /* Treat the number of zero bits in the most recently encoded
398 * LZMS_PROBABILITY_MAX bits with this probability entry as the chance,
399 * out of LZMS_PROBABILITY_MAX, that the next bit will be a 0. However,
400 * don't allow 0% or 100% probabilities. */
401 prob = prob_entry->num_recent_zero_bits;
404 else if (prob == LZMS_PROBABILITY_MAX)
405 prob = LZMS_PROBABILITY_MAX - 1;
407 /* Encode the next bit. */
408 lzms_range_encoder_raw_encode_bit(enc->rc, bit, prob);
410 /* Update the state based on the newly encoded bit. */
411 enc->state = ((enc->state << 1) | bit) & enc->mask;
413 /* Update the recent bits, including the cached count of 0's. */
414 BUILD_BUG_ON(LZMS_PROBABILITY_MAX > sizeof(prob_entry->recent_bits) * 8);
416 if (prob_entry->recent_bits & (1ULL << (LZMS_PROBABILITY_MAX - 1))) {
417 /* Replacing 1 bit with 0 bit; increment the zero count.
419 prob_entry->num_recent_zero_bits++;
422 if (!(prob_entry->recent_bits & (1ULL << (LZMS_PROBABILITY_MAX - 1)))) {
423 /* Replacing 0 bit with 1 bit; decrement the zero count.
425 prob_entry->num_recent_zero_bits--;
428 prob_entry->recent_bits = (prob_entry->recent_bits << 1) | bit;
431 /* Encode a symbol using the specified Huffman encoder. */
433 lzms_huffman_encode_symbol(struct lzms_huffman_encoder *enc, u32 sym)
435 LZMS_ASSERT(sym < enc->num_syms);
436 lzms_output_bitstream_put_bits(enc->os,
439 ++enc->sym_freqs[sym];
440 if (++enc->num_syms_written == enc->rebuild_freq) {
441 /* Adaptive code needs to be rebuilt. */
442 LZMS_DEBUG("Rebuilding code (num_syms=%u)", enc->num_syms);
443 make_canonical_huffman_code(enc->num_syms,
444 LZMS_MAX_CODEWORD_LEN,
449 /* Dilute the frequencies. */
450 for (unsigned i = 0; i < enc->num_syms; i++) {
451 enc->sym_freqs[i] >>= 1;
452 enc->sym_freqs[i] += 1;
454 enc->num_syms_written = 0;
459 lzms_encode_length(struct lzms_huffman_encoder *enc, u32 length)
462 unsigned num_extra_bits;
465 slot = lzms_get_length_slot(length);
467 num_extra_bits = lzms_extra_length_bits[slot];
469 extra_bits = length - lzms_length_slot_base[slot];
471 lzms_huffman_encode_symbol(enc, slot);
472 lzms_output_bitstream_put_bits(enc->os, extra_bits, num_extra_bits);
476 lzms_encode_offset(struct lzms_huffman_encoder *enc, u32 offset)
479 unsigned num_extra_bits;
482 slot = lzms_get_position_slot(offset);
484 num_extra_bits = lzms_extra_position_bits[slot];
486 extra_bits = offset - lzms_position_slot_base[slot];
488 lzms_huffman_encode_symbol(enc, slot);
489 lzms_output_bitstream_put_bits(enc->os, extra_bits, num_extra_bits);
493 lzms_begin_encode_item(struct lzms_compressor *ctx)
495 ctx->lru.lz.upcoming_offset = 0;
496 ctx->lru.delta.upcoming_offset = 0;
497 ctx->lru.delta.upcoming_power = 0;
501 lzms_end_encode_item(struct lzms_compressor *ctx, u32 length)
503 LZMS_ASSERT(ctx->window_size - ctx->cur_window_pos >= length);
504 ctx->cur_window_pos += length;
505 lzms_update_lru_queues(&ctx->lru);
508 /* Encode a literal byte. */
510 lzms_encode_literal(struct lzms_compressor *ctx, u8 literal)
512 LZMS_DEBUG("Position %u: Encoding literal 0x%02x ('%c')",
513 ctx->cur_window_pos, literal, literal);
515 lzms_begin_encode_item(ctx);
517 /* Main bit: 0 = a literal, not a match. */
518 lzms_range_encode_bit(&ctx->main_range_encoder, 0);
520 /* Encode the literal using the current literal Huffman code. */
521 lzms_huffman_encode_symbol(&ctx->literal_encoder, literal);
523 lzms_end_encode_item(ctx, 1);
526 /* Encode a (length, offset) pair (LZ match). */
528 lzms_encode_lz_match(struct lzms_compressor *ctx, u32 length, u32 offset)
530 int recent_offset_idx;
532 LZMS_DEBUG("Position %u: Encoding LZ match {length=%u, offset=%u}",
533 ctx->cur_window_pos, length, offset);
535 LZMS_ASSERT(length <= ctx->window_size - ctx->cur_window_pos);
536 LZMS_ASSERT(offset <= ctx->cur_window_pos);
537 LZMS_ASSERT(!memcmp(&ctx->window[ctx->cur_window_pos],
538 &ctx->window[ctx->cur_window_pos - offset],
541 lzms_begin_encode_item(ctx);
543 /* Main bit: 1 = a match, not a literal. */
544 lzms_range_encode_bit(&ctx->main_range_encoder, 1);
546 /* Match bit: 0 = a LZ match, not a delta match. */
547 lzms_range_encode_bit(&ctx->match_range_encoder, 0);
549 /* Determine if the offset can be represented as a recent offset. */
550 for (recent_offset_idx = 0;
551 recent_offset_idx < LZMS_NUM_RECENT_OFFSETS;
553 if (offset == ctx->lru.lz.recent_offsets[recent_offset_idx])
556 if (recent_offset_idx == LZMS_NUM_RECENT_OFFSETS) {
557 /* Explicit offset. */
559 /* LZ match bit: 0 = explicit offset, not a recent offset. */
560 lzms_range_encode_bit(&ctx->lz_match_range_encoder, 0);
562 /* Encode the match offset. */
563 lzms_encode_offset(&ctx->lz_offset_encoder, offset);
569 /* LZ match bit: 1 = recent offset, not an explicit offset. */
570 lzms_range_encode_bit(&ctx->lz_match_range_encoder, 1);
572 /* Encode the recent offset index. A 1 bit is encoded for each
573 * index passed up. This sequence of 1 bits is terminated by a
574 * 0 bit, or automatically when (LZMS_NUM_RECENT_OFFSETS - 1) 1
575 * bits have been encoded. */
576 for (i = 0; i < recent_offset_idx; i++)
577 lzms_range_encode_bit(&ctx->lz_repeat_match_range_encoders[i], 1);
579 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
580 lzms_range_encode_bit(&ctx->lz_repeat_match_range_encoders[i], 0);
582 /* Initial update of the LZ match offset LRU queue. */
583 for (; i < LZMS_NUM_RECENT_OFFSETS; i++)
584 ctx->lru.lz.recent_offsets[i] = ctx->lru.lz.recent_offsets[i + 1];
587 /* Encode the match length. */
588 lzms_encode_length(&ctx->length_encoder, length);
590 /* Save the match offset for later insertion at the front of the LZ
591 * match offset LRU queue. */
592 ctx->lru.lz.upcoming_offset = offset;
594 lzms_end_encode_item(ctx, length);
599 lzms_record_literal(u8 literal, void *_ctx)
601 struct lzms_compressor *ctx = _ctx;
603 lzms_encode_literal(ctx, literal);
607 lzms_record_match(unsigned length, unsigned offset, void *_ctx)
609 struct lzms_compressor *ctx = _ctx;
611 lzms_encode_lz_match(ctx, length, offset);
615 lzms_fast_encode(struct lzms_compressor *ctx)
617 static const struct lz_params lzms_lz_params = {
619 .max_match = UINT_MAX,
620 .max_offset = UINT_MAX,
624 .max_lazy_match = 258,
628 lz_analyze_block(ctx->window,
639 /* Fast heuristic cost evaluation to use in the inner loop of the match-finder.
640 * Unlike lzms_get_lz_match_cost(), which does a true cost evaluation, this
641 * simply prioritize matches based on their offset. */
643 lzms_lz_match_cost_fast(input_idx_t length, input_idx_t offset, const void *_lru)
645 const struct lzms_lz_lru_queues *lru = _lru;
647 for (input_idx_t i = 0; i < LZMS_NUM_RECENT_OFFSETS; i++)
648 if (offset == lru->recent_offsets[i])
654 #define LZMS_COST_SHIFT 5
656 /*#define LZMS_RC_COSTS_USE_FLOATING_POINT*/
659 lzms_rc_costs[LZMS_PROBABILITY_MAX + 1];
661 #ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
666 lzms_do_init_rc_costs(void)
668 /* Fill in a table that maps range coding probabilities needed to code a
669 * bit X (0 or 1) to the number of bits (scaled by a constant factor, to
670 * handle fractional costs) needed to code that bit X.
672 * Consider the range of the range decoder. To eliminate exactly half
673 * the range (logical probability of 0.5), we need exactly 1 bit. For
674 * lower probabilities we need more bits and for higher probabilities we
675 * need fewer bits. In general, a logical probability of N will
676 * eliminate the proportion 1 - N of the range; this information takes
677 * log2(1 / N) bits to encode.
679 * The below loop is simply calculating this number of bits for each
680 * possible probability allowed by the LZMS compression format, but
681 * without using real numbers. To handle fractional probabilities, each
682 * cost is multiplied by (1 << LZMS_COST_SHIFT). These techniques are
683 * based on those used by LZMA.
685 * Note that in LZMS, a probability x really means x / 64, and 0 / 64 is
686 * really interpreted as 1 / 64 and 64 / 64 is really interpreted as
689 for (u32 i = 0; i <= LZMS_PROBABILITY_MAX; i++) {
694 else if (prob == LZMS_PROBABILITY_MAX)
695 prob = LZMS_PROBABILITY_MAX - 1;
697 #ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
698 lzms_rc_costs[i] = log2((double)LZMS_PROBABILITY_MAX / prob) *
699 (1 << LZMS_COST_SHIFT);
703 for (u32 j = 0; j < LZMS_COST_SHIFT; j++) {
706 while (w >= (1U << 16)) {
711 lzms_rc_costs[i] = (LZMS_PROBABILITY_BITS << LZMS_COST_SHIFT) -
718 lzms_init_rc_costs(void)
720 static bool done = false;
721 static pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
723 if (unlikely(!done)) {
724 pthread_mutex_lock(&mutex);
726 lzms_do_init_rc_costs();
729 pthread_mutex_unlock(&mutex);
734 * Return the cost to range-encode the specified bit when in the specified
737 * @enc The range encoder to use.
738 * @cur_state Current state, which indicates the probability entry to choose.
739 * Updated by this function.
740 * @bit The bit to encode (0 or 1).
743 lzms_rc_bit_cost(const struct lzms_range_encoder *enc, u8 *cur_state, int bit)
748 prob_zero = enc->prob_entries[*cur_state & enc->mask].num_recent_zero_bits;
750 *cur_state = (*cur_state << 1) | bit;
753 prob_correct = prob_zero;
755 prob_correct = LZMS_PROBABILITY_MAX - prob_zero;
757 return lzms_rc_costs[prob_correct];
761 lzms_huffman_symbol_cost(const struct lzms_huffman_encoder *enc, u32 sym)
763 return enc->lens[sym] << LZMS_COST_SHIFT;
767 lzms_offset_cost(const struct lzms_huffman_encoder *enc, u32 offset)
773 slot = lzms_get_position_slot(offset);
775 cost += lzms_huffman_symbol_cost(enc, slot);
777 num_extra_bits = lzms_extra_position_bits[slot];
779 cost += num_extra_bits << LZMS_COST_SHIFT;
785 lzms_length_cost(const struct lzms_huffman_encoder *enc, u32 length)
791 slot = lzms_get_length_slot(length);
793 cost += lzms_huffman_symbol_cost(enc, slot);
795 num_extra_bits = lzms_extra_length_bits[slot];
797 cost += num_extra_bits << LZMS_COST_SHIFT;
803 lzms_get_matches(struct lzms_compressor *ctx,
804 const struct lzms_adaptive_state *state,
805 struct raw_match **matches_ret)
807 *matches_ret = ctx->matches;
808 return lz_sarray_get_matches(&ctx->lz_sarray,
810 lzms_lz_match_cost_fast,
815 lzms_skip_bytes(struct lzms_compressor *ctx, input_idx_t n)
818 lz_sarray_skip_position(&ctx->lz_sarray);
822 lzms_get_prev_literal_cost(struct lzms_compressor *ctx,
823 struct lzms_adaptive_state *state)
825 u8 literal = ctx->window[lz_sarray_get_pos(&ctx->lz_sarray) - 1];
828 state->lru.upcoming_offset = 0;
829 lzms_update_lz_lru_queues(&state->lru);
831 cost += lzms_rc_bit_cost(&ctx->main_range_encoder,
832 &state->main_state, 0);
834 cost += lzms_huffman_symbol_cost(&ctx->literal_encoder, literal);
840 lzms_get_lz_match_cost(struct lzms_compressor *ctx,
841 struct lzms_adaptive_state *state,
842 input_idx_t length, input_idx_t offset)
845 int recent_offset_idx;
847 cost += lzms_rc_bit_cost(&ctx->main_range_encoder,
848 &state->main_state, 1);
849 cost += lzms_rc_bit_cost(&ctx->match_range_encoder,
850 &state->match_state, 0);
852 for (recent_offset_idx = 0;
853 recent_offset_idx < LZMS_NUM_RECENT_OFFSETS;
855 if (offset == state->lru.recent_offsets[recent_offset_idx])
858 if (recent_offset_idx == LZMS_NUM_RECENT_OFFSETS) {
859 /* Explicit offset. */
860 cost += lzms_rc_bit_cost(&ctx->lz_match_range_encoder,
861 &state->lz_match_state, 0);
863 cost += lzms_offset_cost(&ctx->lz_offset_encoder, offset);
868 cost += lzms_rc_bit_cost(&ctx->lz_match_range_encoder,
869 &state->lz_match_state, 1);
871 for (i = 0; i < recent_offset_idx; i++)
872 cost += lzms_rc_bit_cost(&ctx->lz_repeat_match_range_encoders[i],
873 &state->lz_repeat_match_state[i], 0);
875 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
876 cost += lzms_rc_bit_cost(&ctx->lz_repeat_match_range_encoders[i],
877 &state->lz_repeat_match_state[i], 1);
880 /* Initial update of the LZ match offset LRU queue. */
881 for (; i < LZMS_NUM_RECENT_OFFSETS; i++)
882 state->lru.recent_offsets[i] = state->lru.recent_offsets[i + 1];
885 cost += lzms_length_cost(&ctx->length_encoder, length);
887 state->lru.upcoming_offset = offset;
888 lzms_update_lz_lru_queues(&state->lru);
893 static struct raw_match
894 lzms_get_near_optimal_match(struct lzms_compressor *ctx)
896 struct lzms_adaptive_state initial_state;
898 initial_state.lru = ctx->lru.lz;
899 initial_state.main_state = ctx->main_range_encoder.state;
900 initial_state.match_state = ctx->match_range_encoder.state;
901 initial_state.lz_match_state = ctx->lz_match_range_encoder.state;
902 for (int i = 0; i < LZMS_NUM_RECENT_OFFSETS - 1; i++)
903 initial_state.lz_repeat_match_state[i] =
904 ctx->lz_repeat_match_range_encoders[i].state;
905 return lz_get_near_optimal_match(&ctx->mc,
908 lzms_get_prev_literal_cost,
909 lzms_get_lz_match_cost,
915 * The main loop for the LZMS compressor.
919 * - This uses near-optimal LZ parsing backed by a suffix-array match-finder.
920 * More details can be found in the corresponding files (lz_optimal.h,
923 * - This does not output any delta matches. It would take a specialized
924 * algorithm to find them, then more code in lz_optimal.h and here to handle
925 * evaluating and outputting them.
927 * - The costs of literals and matches are estimated using the range encoder
928 * states and the semi-adaptive Huffman codes. Except for range encoding
929 * states, costs are assumed to be constant throughout a single run of the
930 * parsing algorithm, which can parse up to LZMS_OPTIM_ARRAY_SIZE bytes of
931 * data. This introduces a source of inaccuracy because the probabilities and
932 * Huffman codes can change over this part of the data.
935 lzms_normal_encode(struct lzms_compressor *ctx)
937 struct raw_match match;
939 /* Load window into suffix array match-finder. */
940 lz_sarray_load_window(&ctx->lz_sarray, ctx->window, ctx->window_size);
942 /* Reset the match-chooser. */
943 lz_match_chooser_begin(&ctx->mc);
945 while (ctx->cur_window_pos != ctx->window_size) {
946 match = lzms_get_near_optimal_match(ctx);
948 lzms_encode_literal(ctx, ctx->window[ctx->cur_window_pos]);
950 lzms_encode_lz_match(ctx, match.len, match.offset);
955 lzms_init_range_encoder(struct lzms_range_encoder *enc,
956 struct lzms_range_encoder_raw *rc, u32 num_states)
960 enc->mask = num_states - 1;
961 for (u32 i = 0; i < num_states; i++) {
962 enc->prob_entries[i].num_recent_zero_bits = LZMS_INITIAL_PROBABILITY;
963 enc->prob_entries[i].recent_bits = LZMS_INITIAL_RECENT_BITS;
968 lzms_init_huffman_encoder(struct lzms_huffman_encoder *enc,
969 struct lzms_output_bitstream *os,
971 unsigned rebuild_freq)
974 enc->num_syms_written = 0;
975 enc->rebuild_freq = rebuild_freq;
976 enc->num_syms = num_syms;
977 for (unsigned i = 0; i < num_syms; i++)
978 enc->sym_freqs[i] = 1;
980 make_canonical_huffman_code(enc->num_syms,
981 LZMS_MAX_CODEWORD_LEN,
987 /* Initialize the LZMS compressor. */
989 lzms_init_compressor(struct lzms_compressor *ctx, const u8 *udata, u32 ulen,
990 le16 *cdata, u32 clen16)
992 unsigned num_position_slots;
994 /* Copy the uncompressed data into the @ctx->window buffer. */
995 memcpy(ctx->window, udata, ulen);
996 ctx->cur_window_pos = 0;
997 ctx->window_size = ulen;
999 /* Initialize the raw range encoder (writing forwards). */
1000 lzms_range_encoder_raw_init(&ctx->rc, cdata, clen16);
1002 /* Initialize the output bitstream for Huffman symbols and verbatim bits
1003 * (writing backwards). */
1004 lzms_output_bitstream_init(&ctx->os, cdata, clen16);
1006 /* Calculate the number of position slots needed for this compressed
1008 num_position_slots = lzms_get_position_slot(ulen - 1) + 1;
1010 LZMS_DEBUG("Using %u position slots", num_position_slots);
1012 /* Initialize Huffman encoders for each alphabet used in the compressed
1013 * representation. */
1014 lzms_init_huffman_encoder(&ctx->literal_encoder, &ctx->os,
1015 LZMS_NUM_LITERAL_SYMS,
1016 LZMS_LITERAL_CODE_REBUILD_FREQ);
1018 lzms_init_huffman_encoder(&ctx->lz_offset_encoder, &ctx->os,
1020 LZMS_LZ_OFFSET_CODE_REBUILD_FREQ);
1022 lzms_init_huffman_encoder(&ctx->length_encoder, &ctx->os,
1024 LZMS_LENGTH_CODE_REBUILD_FREQ);
1026 lzms_init_huffman_encoder(&ctx->delta_offset_encoder, &ctx->os,
1028 LZMS_DELTA_OFFSET_CODE_REBUILD_FREQ);
1030 lzms_init_huffman_encoder(&ctx->delta_power_encoder, &ctx->os,
1031 LZMS_NUM_DELTA_POWER_SYMS,
1032 LZMS_DELTA_POWER_CODE_REBUILD_FREQ);
1034 /* Initialize range encoders, all of which wrap around the same
1035 * lzms_range_encoder_raw. */
1036 lzms_init_range_encoder(&ctx->main_range_encoder,
1037 &ctx->rc, LZMS_NUM_MAIN_STATES);
1039 lzms_init_range_encoder(&ctx->match_range_encoder,
1040 &ctx->rc, LZMS_NUM_MATCH_STATES);
1042 lzms_init_range_encoder(&ctx->lz_match_range_encoder,
1043 &ctx->rc, LZMS_NUM_LZ_MATCH_STATES);
1045 for (size_t i = 0; i < ARRAY_LEN(ctx->lz_repeat_match_range_encoders); i++)
1046 lzms_init_range_encoder(&ctx->lz_repeat_match_range_encoders[i],
1047 &ctx->rc, LZMS_NUM_LZ_REPEAT_MATCH_STATES);
1049 lzms_init_range_encoder(&ctx->delta_match_range_encoder,
1050 &ctx->rc, LZMS_NUM_DELTA_MATCH_STATES);
1052 for (size_t i = 0; i < ARRAY_LEN(ctx->delta_repeat_match_range_encoders); i++)
1053 lzms_init_range_encoder(&ctx->delta_repeat_match_range_encoders[i],
1054 &ctx->rc, LZMS_NUM_DELTA_REPEAT_MATCH_STATES);
1056 /* Initialize LRU match information. */
1057 lzms_init_lru_queues(&ctx->lru);
1060 /* Flush the output streams, prepare the final compressed data, and return its
1063 * A return value of 0 indicates that the data could not be compressed to fit in
1064 * the available space. */
1066 lzms_finalize(struct lzms_compressor *ctx, u8 *cdata, size_t csize_avail)
1068 size_t num_forwards_bytes;
1069 size_t num_backwards_bytes;
1070 size_t compressed_size;
1072 /* Flush both the forwards and backwards streams, and make sure they
1073 * didn't cross each other and start overwriting each other's data. */
1074 if (!lzms_output_bitstream_flush(&ctx->os)) {
1075 LZMS_DEBUG("Backwards bitstream overrun.");
1079 if (!lzms_range_encoder_raw_flush(&ctx->rc)) {
1080 LZMS_DEBUG("Forwards bitstream overrun.");
1084 if (ctx->rc.out > ctx->os.out) {
1085 LZMS_DEBUG("Two bitstreams crossed.");
1089 /* Now the compressed buffer contains the data output by the forwards
1090 * bitstream, then empty space, then data output by the backwards
1091 * bitstream. Move the data output by the backwards bitstream to be
1092 * adjacent to the data output by the forward bitstream, and calculate
1093 * the compressed size that this results in. */
1094 num_forwards_bytes = (u8*)ctx->rc.out - (u8*)cdata;
1095 num_backwards_bytes = ((u8*)cdata + csize_avail) - (u8*)ctx->os.out;
1097 memmove(cdata + num_forwards_bytes, ctx->os.out, num_backwards_bytes);
1099 compressed_size = num_forwards_bytes + num_backwards_bytes;
1100 LZMS_DEBUG("num_forwards_bytes=%zu, num_backwards_bytes=%zu, "
1101 "compressed_size=%zu",
1102 num_forwards_bytes, num_backwards_bytes, compressed_size);
1103 LZMS_ASSERT(compressed_size % 2 == 0);
1104 return compressed_size;
1108 lzms_compress(const void *uncompressed_data, size_t uncompressed_size,
1109 void *compressed_data, size_t compressed_size_avail, void *_ctx)
1111 struct lzms_compressor *ctx = _ctx;
1112 size_t compressed_size;
1114 LZMS_DEBUG("uncompressed_size=%zu, compressed_size_avail=%zu",
1115 uncompressed_size, compressed_size_avail);
1117 /* Make sure the uncompressed size is compatible with this compressor.
1119 if (uncompressed_size > ctx->max_block_size) {
1120 LZMS_DEBUG("Can't compress %zu bytes: LZMS context "
1121 "only supports %u bytes",
1122 uncompressed_size, ctx->max_block_size);
1126 /* Don't bother compressing extremely small inputs. */
1127 if (uncompressed_size < 4) {
1128 LZMS_DEBUG("Input too small to bother compressing.");
1132 /* Cap the available compressed size to a 32-bit integer and round it
1133 * down to the nearest multiple of 2. */
1134 if (compressed_size_avail > UINT32_MAX)
1135 compressed_size_avail = UINT32_MAX;
1136 if (compressed_size_avail & 1)
1137 compressed_size_avail--;
1139 /* Initialize the compressor structures. */
1140 lzms_init_compressor(ctx, uncompressed_data, uncompressed_size,
1141 compressed_data, compressed_size_avail / 2);
1143 /* Preprocess the uncompressed data. */
1144 lzms_x86_filter(ctx->window, ctx->window_size,
1145 ctx->last_target_usages, false);
1147 /* Compute and encode a literal/match sequence that decompresses to the
1148 * preprocessed data. */
1150 lzms_normal_encode(ctx);
1152 lzms_fast_encode(ctx);
1155 /* Get and return the compressed data size. */
1156 compressed_size = lzms_finalize(ctx, compressed_data,
1157 compressed_size_avail);
1159 if (compressed_size == 0) {
1160 LZMS_DEBUG("Data did not compress to requested size or less.");
1164 LZMS_DEBUG("Compressed %zu => %zu bytes",
1165 uncompressed_size, compressed_size);
1167 #if defined(ENABLE_VERIFY_COMPRESSION) || defined(ENABLE_LZMS_DEBUG)
1168 /* Verify that we really get the same thing back when decompressing. */
1170 struct wimlib_decompressor *decompressor;
1172 LZMS_DEBUG("Verifying LZMS compression.");
1174 if (0 == wimlib_create_decompressor(WIMLIB_COMPRESSION_TYPE_LZMS,
1175 ctx->max_block_size,
1180 ret = wimlib_decompress(compressed_data,
1185 wimlib_free_decompressor(decompressor);
1188 ERROR("Failed to decompress data we "
1189 "compressed using LZMS algorithm");
1193 if (memcmp(uncompressed_data, ctx->window,
1196 ERROR("Data we compressed using LZMS algorithm "
1197 "didn't decompress to original");
1202 WARNING("Failed to create decompressor for "
1203 "data verification!");
1206 #endif /* ENABLE_LZMS_DEBUG || ENABLE_VERIFY_COMPRESSION */
1208 return compressed_size;
1212 lzms_free_compressor(void *_ctx)
1214 struct lzms_compressor *ctx = _ctx;
1219 FREE(ctx->prev_tab);
1222 lz_sarray_destroy(&ctx->lz_sarray);
1223 lz_match_chooser_destroy(&ctx->mc);
1228 static const struct wimlib_lzms_compressor_params lzms_default = {
1229 .hdr = sizeof(struct wimlib_lzms_compressor_params),
1230 .min_match_length = 2,
1231 .max_match_length = UINT32_MAX,
1232 .nice_match_length = 32,
1233 .max_search_depth = 50,
1234 .max_matches_per_pos = 3,
1235 .optim_array_length = 1024,
1238 static const struct wimlib_lzms_compressor_params *
1239 lzms_get_params(const struct wimlib_compressor_params_header *_params)
1241 const struct wimlib_lzms_compressor_params *params =
1242 (const struct wimlib_lzms_compressor_params*)_params;
1245 params = &lzms_default;
1251 lzms_create_compressor(size_t max_block_size,
1252 const struct wimlib_compressor_params_header *_params,
1255 struct lzms_compressor *ctx;
1256 const struct wimlib_lzms_compressor_params *params = lzms_get_params(_params);
1258 if (max_block_size == 0 || max_block_size >= INT32_MAX) {
1259 LZMS_DEBUG("Invalid max_block_size (%u)", max_block_size);
1260 return WIMLIB_ERR_INVALID_PARAM;
1263 ctx = CALLOC(1, sizeof(struct lzms_compressor));
1267 ctx->window = MALLOC(max_block_size);
1268 if (ctx->window == NULL)
1272 ctx->prev_tab = MALLOC(max_block_size * sizeof(ctx->prev_tab[0]));
1273 if (ctx->prev_tab == NULL)
1277 ctx->matches = MALLOC(min(params->max_match_length -
1278 params->min_match_length + 1,
1279 params->max_matches_per_pos) *
1280 sizeof(ctx->matches[0]));
1281 if (ctx->matches == NULL)
1284 if (!lz_sarray_init(&ctx->lz_sarray, max_block_size,
1285 params->min_match_length,
1286 params->max_match_length,
1287 params->max_search_depth,
1288 params->max_matches_per_pos))
1291 if (!lz_match_chooser_init(&ctx->mc,
1292 params->optim_array_length,
1293 params->nice_match_length,
1294 params->max_match_length))
1297 /* Initialize position and length slot data if not done already. */
1300 /* Initialize range encoding cost table if not done already. */
1301 lzms_init_rc_costs();
1303 ctx->max_block_size = max_block_size;
1309 lzms_free_compressor(ctx);
1310 return WIMLIB_ERR_NOMEM;
1314 lzms_get_needed_memory(size_t max_block_size,
1315 const struct wimlib_compressor_params_header *_params)
1317 const struct wimlib_lzms_compressor_params *params = lzms_get_params(_params);
1321 size += max_block_size;
1322 size += sizeof(struct lzms_compressor);
1323 size += lz_sarray_get_needed_memory(max_block_size);
1324 size += lz_match_chooser_get_needed_memory(params->optim_array_length,
1325 params->nice_match_length,
1326 params->max_match_length);
1327 size += min(params->max_match_length -
1328 params->min_match_length + 1,
1329 params->max_matches_per_pos) *
1330 sizeof(((struct lzms_compressor*)0)->matches[0]);
1335 lzms_params_valid(const struct wimlib_compressor_params_header *_params)
1337 const struct wimlib_lzms_compressor_params *params =
1338 (const struct wimlib_lzms_compressor_params*)_params;
1340 if (params->hdr.size != sizeof(*params) ||
1341 params->max_match_length < params->min_match_length ||
1342 params->min_match_length < 2 ||
1343 params->optim_array_length == 0 ||
1344 min(params->max_match_length, params->nice_match_length) > 65536)
1350 const struct compressor_ops lzms_compressor_ops = {
1351 .params_valid = lzms_params_valid,
1352 .get_needed_memory = lzms_get_needed_memory,
1353 .create_compressor = lzms_create_compressor,
1354 .compress = lzms_compress,
1355 .free_compressor = lzms_free_compressor,