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 * NOTE: this compressor currently does not code any delta matches.
35 #include "wimlib/compiler.h"
36 #include "wimlib/compressor_ops.h"
37 #include "wimlib/compress_common.h"
38 #include "wimlib/endianness.h"
39 #include "wimlib/error.h"
40 #include "wimlib/lz_sarray.h"
41 #include "wimlib/lzms.h"
42 #include "wimlib/util.h"
48 #define LZMS_OPTIM_ARRAY_SIZE 1024
50 struct lzms_compressor;
51 struct lzms_adaptive_state {
52 struct lzms_lz_lru_queues lru;
56 u8 lz_repeat_match_state[LZMS_NUM_RECENT_OFFSETS - 1];
58 #define LZ_ADAPTIVE_STATE struct lzms_adaptive_state
59 #define LZ_COMPRESSOR struct lzms_compressor
60 #include "wimlib/lz_optimal.h"
62 /* Stucture used for writing raw bits to the end of the LZMS-compressed data as
63 * a series of 16-bit little endian coding units. */
64 struct lzms_output_bitstream {
65 /* Buffer variable containing zero or more bits that have been logically
66 * written to the bitstream but not yet written to memory. This must be
67 * at least as large as the coding unit size. */
70 /* Number of bits in @bitbuf that are valid. */
71 unsigned num_free_bits;
73 /* Pointer to one past the next position in the compressed data buffer
74 * at which to output a 16-bit coding unit. */
77 /* Maximum number of 16-bit coding units that can still be output to
78 * the compressed data buffer. */
79 size_t num_le16_remaining;
81 /* Set to %true if not all coding units could be output due to
82 * insufficient space. */
86 /* Stucture used for range encoding (raw version). */
87 struct lzms_range_encoder_raw {
89 /* A 33-bit variable that holds the low boundary of the current range.
90 * The 33rd bit is needed to catch carries. */
93 /* Size of the current range. */
96 /* Next 16-bit coding unit to output. */
99 /* Number of 16-bit coding units whose output has been delayed due to
100 * possible carrying. The first such coding unit is @cache; all
101 * subsequent such coding units are 0xffff. */
104 /* Pointer to the next position in the compressed data buffer at which
105 * to output a 16-bit coding unit. */
108 /* Maximum number of 16-bit coding units that can still be output to
109 * the compressed data buffer. */
110 size_t num_le16_remaining;
112 /* %true when the very first coding unit has not yet been output. */
115 /* Set to %true if not all coding units could be output due to
116 * insufficient space. */
120 /* Structure used for range encoding. This wraps around `struct
121 * lzms_range_encoder_raw' to use and maintain probability entries. */
122 struct lzms_range_encoder {
123 /* Pointer to the raw range encoder, which has no persistent knowledge
124 * of probabilities. Multiple lzms_range_encoder's share the same
125 * lzms_range_encoder_raw. */
126 struct lzms_range_encoder_raw *rc;
128 /* Bits recently encoded by this range encoder. This are used as in
129 * index into @prob_entries. */
132 /* Bitmask for @state to prevent its value from exceeding the number of
133 * probability entries. */
136 /* Probability entries being used for this range encoder. */
137 struct lzms_probability_entry prob_entries[LZMS_MAX_NUM_STATES];
140 /* Structure used for Huffman encoding. */
141 struct lzms_huffman_encoder {
143 /* Bitstream to write Huffman-encoded symbols and verbatim bits to.
144 * Multiple lzms_huffman_encoder's share the same lzms_output_bitstream.
146 struct lzms_output_bitstream *os;
148 /* Number of symbols that have been written using this code far. Reset
149 * to 0 whenever the code is rebuilt. */
150 u32 num_syms_written;
152 /* When @num_syms_written reaches this number, the Huffman code must be
156 /* Number of symbols in the represented Huffman code. */
159 /* Running totals of symbol frequencies. These are diluted slightly
160 * whenever the code is rebuilt. */
161 u32 sym_freqs[LZMS_MAX_NUM_SYMS];
163 /* The length, in bits, of each symbol in the Huffman code. */
164 u8 lens[LZMS_MAX_NUM_SYMS];
166 /* The codeword of each symbol in the Huffman code. */
167 u16 codewords[LZMS_MAX_NUM_SYMS];
170 /* State of the LZMS compressor. */
171 struct lzms_compressor {
172 /* Pointer to a buffer holding the preprocessed data to compress. */
175 /* Current position in @buffer. */
178 /* Size of the data in @buffer. */
181 /* Suffix array match-finder. */
182 struct lz_sarray lz_sarray;
184 /* Temporary space to store found matches. */
185 struct raw_match *matches;
188 struct lz_match_chooser mc;
190 /* Maximum block size this compressor instantiation allows. This is the
191 * allocated size of @window. */
194 /* Raw range encoder which outputs to the beginning of the compressed
195 * data buffer, proceeding forwards. */
196 struct lzms_range_encoder_raw rc;
198 /* Bitstream which outputs to the end of the compressed data buffer,
199 * proceeding backwards. */
200 struct lzms_output_bitstream os;
202 /* Range encoders. */
203 struct lzms_range_encoder main_range_encoder;
204 struct lzms_range_encoder match_range_encoder;
205 struct lzms_range_encoder lz_match_range_encoder;
206 struct lzms_range_encoder lz_repeat_match_range_encoders[LZMS_NUM_RECENT_OFFSETS - 1];
207 struct lzms_range_encoder delta_match_range_encoder;
208 struct lzms_range_encoder delta_repeat_match_range_encoders[LZMS_NUM_RECENT_OFFSETS - 1];
210 /* Huffman encoders. */
211 struct lzms_huffman_encoder literal_encoder;
212 struct lzms_huffman_encoder lz_offset_encoder;
213 struct lzms_huffman_encoder length_encoder;
214 struct lzms_huffman_encoder delta_power_encoder;
215 struct lzms_huffman_encoder delta_offset_encoder;
217 /* LRU (least-recently-used) queues for match information. */
218 struct lzms_lru_queues lru;
220 /* Used for preprocessing. */
221 s32 last_target_usages[65536];
224 /* Initialize the output bitstream @os to write forwards to the specified
225 * compressed data buffer @out that is @out_limit 16-bit integers long. */
227 lzms_output_bitstream_init(struct lzms_output_bitstream *os,
228 le16 *out, size_t out_limit)
231 os->num_free_bits = 16;
232 os->out = out + out_limit;
233 os->num_le16_remaining = out_limit;
237 /* Write @num_bits bits, contained in the low @num_bits bits of @bits (ordered
238 * from high-order to low-order), to the output bitstream @os. */
240 lzms_output_bitstream_put_bits(struct lzms_output_bitstream *os,
241 u32 bits, unsigned num_bits)
243 bits &= (1U << num_bits) - 1;
245 while (num_bits > os->num_free_bits) {
247 if (unlikely(os->num_le16_remaining == 0)) {
252 unsigned num_fill_bits = os->num_free_bits;
254 os->bitbuf <<= num_fill_bits;
255 os->bitbuf |= bits >> (num_bits - num_fill_bits);
257 *--os->out = cpu_to_le16(os->bitbuf);
258 --os->num_le16_remaining;
260 os->num_free_bits = 16;
261 num_bits -= num_fill_bits;
262 bits &= (1U << num_bits) - 1;
264 os->bitbuf <<= num_bits;
266 os->num_free_bits -= num_bits;
269 /* Flush the output bitstream, ensuring that all bits written to it have been
270 * written to memory. Returns %true if all bits were output successfully, or
271 * %false if an overrun occurred. */
273 lzms_output_bitstream_flush(struct lzms_output_bitstream *os)
275 if (os->num_free_bits != 16)
276 lzms_output_bitstream_put_bits(os, 0, os->num_free_bits + 1);
280 /* Initialize the range encoder @rc to write forwards to the specified
281 * compressed data buffer @out that is @out_limit 16-bit integers long. */
283 lzms_range_encoder_raw_init(struct lzms_range_encoder_raw *rc,
284 le16 *out, size_t out_limit)
287 rc->range = 0xffffffff;
291 rc->num_le16_remaining = out_limit;
297 * Attempt to flush bits from the range encoder.
299 * Note: this is based on the public domain code for LZMA written by Igor
300 * Pavlov. The only differences in this function are that in LZMS the bits must
301 * be output in 16-bit coding units instead of 8-bit coding units, and that in
302 * LZMS the first coding unit is not ignored by the decompressor, so the encoder
303 * cannot output a dummy value to that position.
305 * The basic idea is that we're writing bits from @rc->low to the output.
306 * However, due to carrying, the writing of coding units with value 0xffff, as
307 * well as one prior coding unit, must be delayed until it is determined whether
311 lzms_range_encoder_raw_shift_low(struct lzms_range_encoder_raw *rc)
313 LZMS_DEBUG("low=%"PRIx64", cache=%"PRIx64", cache_size=%u",
314 rc->low, rc->cache, rc->cache_size);
315 if ((u32)(rc->low) < 0xffff0000 ||
316 (u32)(rc->low >> 32) != 0)
318 /* Carry not needed (rc->low < 0xffff0000), or carry occurred
319 * ((rc->low >> 32) != 0, a.k.a. the carry bit is 1). */
322 if (rc->num_le16_remaining == 0) {
326 *rc->out++ = cpu_to_le16(rc->cache +
327 (u16)(rc->low >> 32));
328 --rc->num_le16_remaining;
334 } while (--rc->cache_size != 0);
336 rc->cache = (rc->low >> 16) & 0xffff;
339 rc->low = (rc->low & 0xffff) << 16;
343 lzms_range_encoder_raw_normalize(struct lzms_range_encoder_raw *rc)
345 if (rc->range <= 0xffff) {
347 lzms_range_encoder_raw_shift_low(rc);
352 lzms_range_encoder_raw_flush(struct lzms_range_encoder_raw *rc)
354 for (unsigned i = 0; i < 4; i++)
355 lzms_range_encoder_raw_shift_low(rc);
359 /* Encode the next bit using the range encoder (raw version).
361 * @prob is the chance out of LZMS_PROBABILITY_MAX that the next bit is 0. */
363 lzms_range_encoder_raw_encode_bit(struct lzms_range_encoder_raw *rc, int bit,
366 lzms_range_encoder_raw_normalize(rc);
368 u32 bound = (rc->range >> LZMS_PROBABILITY_BITS) * prob;
377 /* Encode a bit using the specified range encoder. This wraps around
378 * lzms_range_encoder_raw_encode_bit() to handle using and updating the
379 * appropriate probability table. */
381 lzms_range_encode_bit(struct lzms_range_encoder *enc, int bit)
383 struct lzms_probability_entry *prob_entry;
386 /* Load the probability entry corresponding to the current state. */
387 prob_entry = &enc->prob_entries[enc->state];
389 /* Treat the number of zero bits in the most recently encoded
390 * LZMS_PROBABILITY_MAX bits with this probability entry as the chance,
391 * out of LZMS_PROBABILITY_MAX, that the next bit will be a 0. However,
392 * don't allow 0% or 100% probabilities. */
393 prob = prob_entry->num_recent_zero_bits;
396 else if (prob == LZMS_PROBABILITY_MAX)
397 prob = LZMS_PROBABILITY_MAX - 1;
399 /* Encode the next bit. */
400 lzms_range_encoder_raw_encode_bit(enc->rc, bit, prob);
402 /* Update the state based on the newly encoded bit. */
403 enc->state = ((enc->state << 1) | bit) & enc->mask;
405 /* Update the recent bits, including the cached count of 0's. */
406 BUILD_BUG_ON(LZMS_PROBABILITY_MAX > sizeof(prob_entry->recent_bits) * 8);
408 if (prob_entry->recent_bits & (1ULL << (LZMS_PROBABILITY_MAX - 1))) {
409 /* Replacing 1 bit with 0 bit; increment the zero count.
411 prob_entry->num_recent_zero_bits++;
414 if (!(prob_entry->recent_bits & (1ULL << (LZMS_PROBABILITY_MAX - 1)))) {
415 /* Replacing 0 bit with 1 bit; decrement the zero count.
417 prob_entry->num_recent_zero_bits--;
420 prob_entry->recent_bits = (prob_entry->recent_bits << 1) | bit;
423 /* Encode a symbol using the specified Huffman encoder. */
425 lzms_huffman_encode_symbol(struct lzms_huffman_encoder *enc, u32 sym)
427 LZMS_ASSERT(sym < enc->num_syms);
428 lzms_output_bitstream_put_bits(enc->os,
431 ++enc->sym_freqs[sym];
432 if (++enc->num_syms_written == enc->rebuild_freq) {
433 /* Adaptive code needs to be rebuilt. */
434 LZMS_DEBUG("Rebuilding code (num_syms=%u)", enc->num_syms);
435 make_canonical_huffman_code(enc->num_syms,
436 LZMS_MAX_CODEWORD_LEN,
441 /* Dilute the frequencies. */
442 for (unsigned i = 0; i < enc->num_syms; i++) {
443 enc->sym_freqs[i] >>= 1;
444 enc->sym_freqs[i] += 1;
446 enc->num_syms_written = 0;
451 lzms_encode_length(struct lzms_huffman_encoder *enc, u32 length)
454 unsigned num_extra_bits;
457 slot = lzms_get_length_slot(length);
459 num_extra_bits = lzms_extra_length_bits[slot];
461 extra_bits = length - lzms_length_slot_base[slot];
463 lzms_huffman_encode_symbol(enc, slot);
464 lzms_output_bitstream_put_bits(enc->os, extra_bits, num_extra_bits);
468 lzms_encode_offset(struct lzms_huffman_encoder *enc, u32 offset)
471 unsigned num_extra_bits;
474 slot = lzms_get_position_slot(offset);
476 num_extra_bits = lzms_extra_position_bits[slot];
478 extra_bits = offset - lzms_position_slot_base[slot];
480 lzms_huffman_encode_symbol(enc, slot);
481 lzms_output_bitstream_put_bits(enc->os, extra_bits, num_extra_bits);
485 lzms_begin_encode_item(struct lzms_compressor *ctx)
487 ctx->lru.lz.upcoming_offset = 0;
488 ctx->lru.delta.upcoming_offset = 0;
489 ctx->lru.delta.upcoming_power = 0;
493 lzms_end_encode_item(struct lzms_compressor *ctx, u32 length)
495 LZMS_ASSERT(ctx->window_size - ctx->cur_window_pos >= length);
496 ctx->cur_window_pos += length;
497 lzms_update_lru_queues(&ctx->lru);
500 /* Encode a literal byte. */
502 lzms_encode_literal(struct lzms_compressor *ctx, u8 literal)
504 LZMS_DEBUG("Position %u: Encoding literal 0x%02x ('%c')",
505 ctx->cur_window_pos, literal, literal);
507 lzms_begin_encode_item(ctx);
509 /* Main bit: 0 = a literal, not a match. */
510 lzms_range_encode_bit(&ctx->main_range_encoder, 0);
512 /* Encode the literal using the current literal Huffman code. */
513 lzms_huffman_encode_symbol(&ctx->literal_encoder, literal);
515 lzms_end_encode_item(ctx, 1);
518 /* Encode a (length, offset) pair (LZ match). */
520 lzms_encode_lz_match(struct lzms_compressor *ctx, u32 length, u32 offset)
522 int recent_offset_idx;
524 LZMS_DEBUG("Position %u: Encoding LZ match {length=%u, offset=%u}",
525 ctx->cur_window_pos, length, offset);
527 LZMS_ASSERT(length <= ctx->window_size - ctx->cur_window_pos);
528 LZMS_ASSERT(offset <= ctx->cur_window_pos);
529 LZMS_ASSERT(!memcmp(&ctx->window[ctx->cur_window_pos],
530 &ctx->window[ctx->cur_window_pos - offset],
533 lzms_begin_encode_item(ctx);
535 /* Main bit: 1 = a match, not a literal. */
536 lzms_range_encode_bit(&ctx->main_range_encoder, 1);
538 /* Match bit: 0 = a LZ match, not a delta match. */
539 lzms_range_encode_bit(&ctx->match_range_encoder, 0);
541 /* Determine if the offset can be represented as a recent offset. */
542 for (recent_offset_idx = 0;
543 recent_offset_idx < LZMS_NUM_RECENT_OFFSETS;
545 if (offset == ctx->lru.lz.recent_offsets[recent_offset_idx])
548 if (recent_offset_idx == LZMS_NUM_RECENT_OFFSETS) {
549 /* Explicit offset. */
551 /* LZ match bit: 0 = explicit offset, not a recent offset. */
552 lzms_range_encode_bit(&ctx->lz_match_range_encoder, 0);
554 /* Encode the match offset. */
555 lzms_encode_offset(&ctx->lz_offset_encoder, offset);
561 /* LZ match bit: 1 = recent offset, not an explicit offset. */
562 lzms_range_encode_bit(&ctx->lz_match_range_encoder, 1);
564 /* Encode the recent offset index. A 1 bit is encoded for each
565 * index passed up. This sequence of 1 bits is terminated by a
566 * 0 bit, or automatically when (LZMS_NUM_RECENT_OFFSETS - 1) 1
567 * bits have been encoded. */
568 for (i = 0; i < recent_offset_idx; i++)
569 lzms_range_encode_bit(&ctx->lz_repeat_match_range_encoders[i], 1);
571 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
572 lzms_range_encode_bit(&ctx->lz_repeat_match_range_encoders[i], 0);
574 /* Initial update of the LZ match offset LRU queue. */
575 for (; i < LZMS_NUM_RECENT_OFFSETS; i++)
576 ctx->lru.lz.recent_offsets[i] = ctx->lru.lz.recent_offsets[i + 1];
579 /* Encode the match length. */
580 lzms_encode_length(&ctx->length_encoder, length);
582 /* Save the match offset for later insertion at the front of the LZ
583 * match offset LRU queue. */
584 ctx->lru.lz.upcoming_offset = offset;
586 lzms_end_encode_item(ctx, length);
589 /* Fast heuristic cost evaluation to use in the inner loop of the match-finder.
590 * Unlike lzms_get_lz_match_cost(), which does a true cost evaluation, this
591 * simply prioritize matches based on their offset. */
593 lzms_lz_match_cost_fast(input_idx_t length, input_idx_t offset, const void *_lru)
595 const struct lzms_lz_lru_queues *lru = _lru;
597 for (input_idx_t i = 0; i < LZMS_NUM_RECENT_OFFSETS; i++)
598 if (offset == lru->recent_offsets[i])
604 #define LZMS_COST_SHIFT 5
606 /*#define LZMS_RC_COSTS_USE_FLOATING_POINT*/
609 lzms_rc_costs[LZMS_PROBABILITY_MAX + 1];
611 #ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
616 lzms_do_init_rc_costs(void)
618 /* Fill in a table that maps range coding probabilities needed to code a
619 * bit X (0 or 1) to the number of bits (scaled by a constant factor, to
620 * handle fractional costs) needed to code that bit X.
622 * Consider the range of the range decoder. To eliminate exactly half
623 * the range (logical probability of 0.5), we need exactly 1 bit. For
624 * lower probabilities we need more bits and for higher probabilities we
625 * need fewer bits. In general, a logical probability of N will
626 * eliminate the proportion 1 - N of the range; this information takes
627 * log2(1 / N) bits to encode.
629 * The below loop is simply calculating this number of bits for each
630 * possible probability allowed by the LZMS compression format, but
631 * without using real numbers. To handle fractional probabilities, each
632 * cost is multiplied by (1 << LZMS_COST_SHIFT). These techniques are
633 * based on those used by LZMA.
635 * Note that in LZMS, a probability x really means x / 64, and 0 / 64 is
636 * really interpreted as 1 / 64 and 64 / 64 is really interpreted as
639 for (u32 i = 0; i <= LZMS_PROBABILITY_MAX; i++) {
644 else if (prob == LZMS_PROBABILITY_MAX)
645 prob = LZMS_PROBABILITY_MAX - 1;
647 #ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
648 lzms_rc_costs[i] = log2((double)LZMS_PROBABILITY_MAX / prob) *
649 (1 << LZMS_COST_SHIFT);
653 for (u32 j = 0; j < LZMS_COST_SHIFT; j++) {
656 while (w >= (1U << 16)) {
661 lzms_rc_costs[i] = (LZMS_PROBABILITY_BITS << LZMS_COST_SHIFT) -
668 lzms_init_rc_costs(void)
670 static bool done = false;
671 static pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
673 if (unlikely(!done)) {
674 pthread_mutex_lock(&mutex);
676 lzms_do_init_rc_costs();
679 pthread_mutex_unlock(&mutex);
684 * Return the cost to range-encode the specified bit when in the specified
687 * @enc The range encoder to use.
688 * @cur_state Current state, which indicates the probability entry to choose.
689 * Updated by this function.
690 * @bit The bit to encode (0 or 1).
693 lzms_rc_bit_cost(const struct lzms_range_encoder *enc, u8 *cur_state, int bit)
698 prob_zero = enc->prob_entries[*cur_state & enc->mask].num_recent_zero_bits;
700 *cur_state = (*cur_state << 1) | bit;
703 prob_correct = prob_zero;
705 prob_correct = LZMS_PROBABILITY_MAX - prob_zero;
707 return lzms_rc_costs[prob_correct];
711 lzms_huffman_symbol_cost(const struct lzms_huffman_encoder *enc, u32 sym)
713 return enc->lens[sym] << LZMS_COST_SHIFT;
717 lzms_offset_cost(const struct lzms_huffman_encoder *enc, u32 offset)
723 slot = lzms_get_position_slot(offset);
725 cost += lzms_huffman_symbol_cost(enc, slot);
727 num_extra_bits = lzms_extra_position_bits[slot];
729 cost += num_extra_bits << LZMS_COST_SHIFT;
735 lzms_length_cost(const struct lzms_huffman_encoder *enc, u32 length)
741 slot = lzms_get_length_slot(length);
743 cost += lzms_huffman_symbol_cost(enc, slot);
745 num_extra_bits = lzms_extra_length_bits[slot];
747 cost += num_extra_bits << LZMS_COST_SHIFT;
753 lzms_get_matches(struct lzms_compressor *ctx,
754 const struct lzms_adaptive_state *state,
755 struct raw_match **matches_ret)
757 *matches_ret = ctx->matches;
758 return lz_sarray_get_matches(&ctx->lz_sarray,
760 lzms_lz_match_cost_fast,
765 lzms_skip_bytes(struct lzms_compressor *ctx, input_idx_t n)
768 lz_sarray_skip_position(&ctx->lz_sarray);
772 lzms_get_prev_literal_cost(struct lzms_compressor *ctx,
773 struct lzms_adaptive_state *state)
775 u8 literal = ctx->window[lz_sarray_get_pos(&ctx->lz_sarray) - 1];
778 state->lru.upcoming_offset = 0;
779 lzms_update_lz_lru_queues(&state->lru);
781 cost += lzms_rc_bit_cost(&ctx->main_range_encoder,
782 &state->main_state, 0);
784 cost += lzms_huffman_symbol_cost(&ctx->literal_encoder, literal);
790 lzms_get_lz_match_cost(struct lzms_compressor *ctx,
791 struct lzms_adaptive_state *state,
792 input_idx_t length, input_idx_t offset)
795 int recent_offset_idx;
797 cost += lzms_rc_bit_cost(&ctx->main_range_encoder,
798 &state->main_state, 1);
799 cost += lzms_rc_bit_cost(&ctx->match_range_encoder,
800 &state->match_state, 0);
802 for (recent_offset_idx = 0;
803 recent_offset_idx < LZMS_NUM_RECENT_OFFSETS;
805 if (offset == state->lru.recent_offsets[recent_offset_idx])
808 if (recent_offset_idx == LZMS_NUM_RECENT_OFFSETS) {
809 /* Explicit offset. */
810 cost += lzms_rc_bit_cost(&ctx->lz_match_range_encoder,
811 &state->lz_match_state, 0);
813 cost += lzms_offset_cost(&ctx->lz_offset_encoder, offset);
818 cost += lzms_rc_bit_cost(&ctx->lz_match_range_encoder,
819 &state->lz_match_state, 1);
821 for (i = 0; i < recent_offset_idx; i++)
822 cost += lzms_rc_bit_cost(&ctx->lz_repeat_match_range_encoders[i],
823 &state->lz_repeat_match_state[i], 0);
825 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
826 cost += lzms_rc_bit_cost(&ctx->lz_repeat_match_range_encoders[i],
827 &state->lz_repeat_match_state[i], 1);
830 /* Initial update of the LZ match offset LRU queue. */
831 for (; i < LZMS_NUM_RECENT_OFFSETS; i++)
832 state->lru.recent_offsets[i] = state->lru.recent_offsets[i + 1];
835 cost += lzms_length_cost(&ctx->length_encoder, length);
837 state->lru.upcoming_offset = offset;
838 lzms_update_lz_lru_queues(&state->lru);
843 static struct raw_match
844 lzms_get_near_optimal_match(struct lzms_compressor *ctx)
846 struct lzms_adaptive_state initial_state;
848 initial_state.lru = ctx->lru.lz;
849 initial_state.main_state = ctx->main_range_encoder.state;
850 initial_state.match_state = ctx->match_range_encoder.state;
851 initial_state.lz_match_state = ctx->lz_match_range_encoder.state;
852 for (int i = 0; i < LZMS_NUM_RECENT_OFFSETS - 1; i++)
853 initial_state.lz_repeat_match_state[i] =
854 ctx->lz_repeat_match_range_encoders[i].state;
855 return lz_get_near_optimal_match(&ctx->mc,
858 lzms_get_prev_literal_cost,
859 lzms_get_lz_match_cost,
865 * The main loop for the LZMS compressor.
869 * - This uses near-optimal LZ parsing backed by a suffix-array match-finder.
870 * More details can be found in the corresponding files (lz_optimal.h,
873 * - This does not output any delta matches. It would take a specialized
874 * algorithm to find them, then more code in lz_optimal.h and here to handle
875 * evaluating and outputting them.
877 * - The costs of literals and matches are estimated using the range encoder
878 * states and the semi-adaptive Huffman codes. Except for range encoding
879 * states, costs are assumed to be constant throughout a single run of the
880 * parsing algorithm, which can parse up to LZMS_OPTIM_ARRAY_SIZE bytes of
881 * data. This introduces a source of inaccuracy because the probabilities and
882 * Huffman codes can change over this part of the data.
885 lzms_encode(struct lzms_compressor *ctx)
887 struct raw_match match;
889 /* Load window into suffix array match-finder. */
890 lz_sarray_load_window(&ctx->lz_sarray, ctx->window, ctx->window_size);
892 /* Reset the match-chooser. */
893 lz_match_chooser_begin(&ctx->mc);
895 while (ctx->cur_window_pos != ctx->window_size) {
896 match = lzms_get_near_optimal_match(ctx);
898 lzms_encode_literal(ctx, ctx->window[ctx->cur_window_pos]);
900 lzms_encode_lz_match(ctx, match.len, match.offset);
905 lzms_init_range_encoder(struct lzms_range_encoder *enc,
906 struct lzms_range_encoder_raw *rc, u32 num_states)
910 enc->mask = num_states - 1;
911 for (u32 i = 0; i < num_states; i++) {
912 enc->prob_entries[i].num_recent_zero_bits = LZMS_INITIAL_PROBABILITY;
913 enc->prob_entries[i].recent_bits = LZMS_INITIAL_RECENT_BITS;
918 lzms_init_huffman_encoder(struct lzms_huffman_encoder *enc,
919 struct lzms_output_bitstream *os,
921 unsigned rebuild_freq)
924 enc->num_syms_written = 0;
925 enc->rebuild_freq = rebuild_freq;
926 enc->num_syms = num_syms;
927 for (unsigned i = 0; i < num_syms; i++)
928 enc->sym_freqs[i] = 1;
930 make_canonical_huffman_code(enc->num_syms,
931 LZMS_MAX_CODEWORD_LEN,
937 /* Initialize the LZMS compressor. */
939 lzms_init_compressor(struct lzms_compressor *ctx, const u8 *udata, u32 ulen,
940 le16 *cdata, u32 clen16)
942 unsigned num_position_slots;
944 /* Copy the uncompressed data into the @ctx->window buffer. */
945 memcpy(ctx->window, udata, ulen);
946 ctx->cur_window_pos = 0;
947 ctx->window_size = ulen;
949 /* Initialize the raw range encoder (writing forwards). */
950 lzms_range_encoder_raw_init(&ctx->rc, cdata, clen16);
952 /* Initialize the output bitstream for Huffman symbols and verbatim bits
953 * (writing backwards). */
954 lzms_output_bitstream_init(&ctx->os, cdata, clen16);
956 /* Calculate the number of position slots needed for this compressed
958 num_position_slots = lzms_get_position_slot(ulen - 1) + 1;
960 LZMS_DEBUG("Using %u position slots", num_position_slots);
962 /* Initialize Huffman encoders for each alphabet used in the compressed
964 lzms_init_huffman_encoder(&ctx->literal_encoder, &ctx->os,
965 LZMS_NUM_LITERAL_SYMS,
966 LZMS_LITERAL_CODE_REBUILD_FREQ);
968 lzms_init_huffman_encoder(&ctx->lz_offset_encoder, &ctx->os,
970 LZMS_LZ_OFFSET_CODE_REBUILD_FREQ);
972 lzms_init_huffman_encoder(&ctx->length_encoder, &ctx->os,
974 LZMS_LENGTH_CODE_REBUILD_FREQ);
976 lzms_init_huffman_encoder(&ctx->delta_offset_encoder, &ctx->os,
978 LZMS_DELTA_OFFSET_CODE_REBUILD_FREQ);
980 lzms_init_huffman_encoder(&ctx->delta_power_encoder, &ctx->os,
981 LZMS_NUM_DELTA_POWER_SYMS,
982 LZMS_DELTA_POWER_CODE_REBUILD_FREQ);
984 /* Initialize range encoders, all of which wrap around the same
985 * lzms_range_encoder_raw. */
986 lzms_init_range_encoder(&ctx->main_range_encoder,
987 &ctx->rc, LZMS_NUM_MAIN_STATES);
989 lzms_init_range_encoder(&ctx->match_range_encoder,
990 &ctx->rc, LZMS_NUM_MATCH_STATES);
992 lzms_init_range_encoder(&ctx->lz_match_range_encoder,
993 &ctx->rc, LZMS_NUM_LZ_MATCH_STATES);
995 for (size_t i = 0; i < ARRAY_LEN(ctx->lz_repeat_match_range_encoders); i++)
996 lzms_init_range_encoder(&ctx->lz_repeat_match_range_encoders[i],
997 &ctx->rc, LZMS_NUM_LZ_REPEAT_MATCH_STATES);
999 lzms_init_range_encoder(&ctx->delta_match_range_encoder,
1000 &ctx->rc, LZMS_NUM_DELTA_MATCH_STATES);
1002 for (size_t i = 0; i < ARRAY_LEN(ctx->delta_repeat_match_range_encoders); i++)
1003 lzms_init_range_encoder(&ctx->delta_repeat_match_range_encoders[i],
1004 &ctx->rc, LZMS_NUM_DELTA_REPEAT_MATCH_STATES);
1006 /* Initialize LRU match information. */
1007 lzms_init_lru_queues(&ctx->lru);
1010 /* Flush the output streams, prepare the final compressed data, and return its
1013 * A return value of 0 indicates that the data could not be compressed to fit in
1014 * the available space. */
1016 lzms_finalize(struct lzms_compressor *ctx, u8 *cdata, size_t csize_avail)
1018 size_t num_forwards_bytes;
1019 size_t num_backwards_bytes;
1020 size_t compressed_size;
1022 /* Flush both the forwards and backwards streams, and make sure they
1023 * didn't cross each other and start overwriting each other's data. */
1024 if (!lzms_output_bitstream_flush(&ctx->os)) {
1025 LZMS_DEBUG("Backwards bitstream overrun.");
1029 if (!lzms_range_encoder_raw_flush(&ctx->rc)) {
1030 LZMS_DEBUG("Forwards bitstream overrun.");
1034 if (ctx->rc.out > ctx->os.out) {
1035 LZMS_DEBUG("Two bitstreams crossed.");
1039 /* Now the compressed buffer contains the data output by the forwards
1040 * bitstream, then empty space, then data output by the backwards
1041 * bitstream. Move the data output by the backwards bitstream to be
1042 * adjacent to the data output by the forward bitstream, and calculate
1043 * the compressed size that this results in. */
1044 num_forwards_bytes = (u8*)ctx->rc.out - (u8*)cdata;
1045 num_backwards_bytes = ((u8*)cdata + csize_avail) - (u8*)ctx->os.out;
1047 memmove(cdata + num_forwards_bytes, ctx->os.out, num_backwards_bytes);
1049 compressed_size = num_forwards_bytes + num_backwards_bytes;
1050 LZMS_DEBUG("num_forwards_bytes=%zu, num_backwards_bytes=%zu, "
1051 "compressed_size=%zu",
1052 num_forwards_bytes, num_backwards_bytes, compressed_size);
1053 LZMS_ASSERT(compressed_size % 2 == 0);
1054 return compressed_size;
1058 lzms_compress(const void *uncompressed_data, size_t uncompressed_size,
1059 void *compressed_data, size_t compressed_size_avail, void *_ctx)
1061 struct lzms_compressor *ctx = _ctx;
1062 size_t compressed_size;
1064 LZMS_DEBUG("uncompressed_size=%zu, compressed_size_avail=%zu",
1065 uncompressed_size, compressed_size_avail);
1067 /* Make sure the uncompressed size is compatible with this compressor.
1069 if (uncompressed_size > ctx->max_block_size) {
1070 LZMS_DEBUG("Can't compress %zu bytes: LZMS context "
1071 "only supports %u bytes",
1072 uncompressed_size, ctx->max_block_size);
1076 /* Don't bother compressing extremely small inputs. */
1077 if (uncompressed_size < 4) {
1078 LZMS_DEBUG("Input too small to bother compressing.");
1082 /* Cap the available compressed size to a 32-bit integer and round it
1083 * down to the nearest multiple of 2. */
1084 if (compressed_size_avail > UINT32_MAX)
1085 compressed_size_avail = UINT32_MAX;
1086 if (compressed_size_avail & 1)
1087 compressed_size_avail--;
1089 /* Initialize the compressor structures. */
1090 lzms_init_compressor(ctx, uncompressed_data, uncompressed_size,
1091 compressed_data, compressed_size_avail / 2);
1093 /* Preprocess the uncompressed data. */
1094 lzms_x86_filter(ctx->window, ctx->window_size,
1095 ctx->last_target_usages, false);
1097 /* Compute and encode a literal/match sequence that decompresses to the
1098 * preprocessed data. */
1101 /* Get and return the compressed data size. */
1102 compressed_size = lzms_finalize(ctx, compressed_data,
1103 compressed_size_avail);
1105 if (compressed_size == 0) {
1106 LZMS_DEBUG("Data did not compress to requested size or less.");
1110 LZMS_DEBUG("Compressed %zu => %zu bytes",
1111 uncompressed_size, compressed_size);
1113 #if defined(ENABLE_VERIFY_COMPRESSION) || defined(ENABLE_LZMS_DEBUG)
1114 /* Verify that we really get the same thing back when decompressing. */
1116 struct wimlib_decompressor *decompressor;
1118 LZMS_DEBUG("Verifying LZMS compression.");
1120 if (0 == wimlib_create_decompressor(WIMLIB_COMPRESSION_TYPE_LZMS,
1121 ctx->max_block_size,
1126 ret = wimlib_decompress(compressed_data,
1131 wimlib_free_decompressor(decompressor);
1134 ERROR("Failed to decompress data we "
1135 "compressed using LZMS algorithm");
1139 if (memcmp(uncompressed_data, ctx->window,
1142 ERROR("Data we compressed using LZMS algorithm "
1143 "didn't decompress to original");
1148 WARNING("Failed to create decompressor for "
1149 "data verification!");
1152 #endif /* ENABLE_LZMS_DEBUG || ENABLE_VERIFY_COMPRESSION */
1154 return compressed_size;
1158 lzms_free_compressor(void *_ctx)
1160 struct lzms_compressor *ctx = _ctx;
1165 lz_sarray_destroy(&ctx->lz_sarray);
1166 lz_match_chooser_destroy(&ctx->mc);
1171 static const struct wimlib_lzms_compressor_params lzms_default = {
1172 .hdr = sizeof(struct wimlib_lzms_compressor_params),
1173 .min_match_length = 2,
1174 .max_match_length = UINT32_MAX,
1175 .nice_match_length = 32,
1176 .max_search_depth = 50,
1177 .max_matches_per_pos = 3,
1178 .optim_array_length = 1024,
1181 static const struct wimlib_lzms_compressor_params *
1182 lzms_get_params(const struct wimlib_compressor_params_header *_params)
1184 const struct wimlib_lzms_compressor_params *params =
1185 (const struct wimlib_lzms_compressor_params*)_params;
1188 params = &lzms_default;
1194 lzms_create_compressor(size_t max_block_size,
1195 const struct wimlib_compressor_params_header *_params,
1198 struct lzms_compressor *ctx;
1199 const struct wimlib_lzms_compressor_params *params = lzms_get_params(_params);
1201 if (max_block_size == 0 || max_block_size >= INT32_MAX) {
1202 LZMS_DEBUG("Invalid max_block_size (%u)", max_block_size);
1203 return WIMLIB_ERR_INVALID_PARAM;
1206 ctx = CALLOC(1, sizeof(struct lzms_compressor));
1210 ctx->window = MALLOC(max_block_size);
1211 if (ctx->window == NULL)
1214 ctx->matches = MALLOC(min(params->max_match_length -
1215 params->min_match_length + 1,
1216 params->max_matches_per_pos) *
1217 sizeof(ctx->matches[0]));
1218 if (ctx->matches == NULL)
1221 if (!lz_sarray_init(&ctx->lz_sarray, max_block_size,
1222 params->min_match_length,
1223 params->max_match_length,
1224 params->max_search_depth,
1225 params->max_matches_per_pos))
1228 if (!lz_match_chooser_init(&ctx->mc,
1229 params->optim_array_length,
1230 params->nice_match_length,
1231 params->max_match_length))
1234 /* Initialize position and length slot data if not done already. */
1237 /* Initialize range encoding cost table if not done already. */
1238 lzms_init_rc_costs();
1240 ctx->max_block_size = max_block_size;
1246 lzms_free_compressor(ctx);
1247 return WIMLIB_ERR_NOMEM;
1251 lzms_get_needed_memory(size_t max_block_size,
1252 const struct wimlib_compressor_params_header *_params)
1254 const struct wimlib_lzms_compressor_params *params = lzms_get_params(_params);
1258 size += max_block_size;
1259 size += sizeof(struct lzms_compressor);
1260 size += lz_sarray_get_needed_memory(max_block_size);
1261 size += lz_match_chooser_get_needed_memory(params->optim_array_length,
1262 params->nice_match_length,
1263 params->max_match_length);
1264 size += min(params->max_match_length -
1265 params->min_match_length + 1,
1266 params->max_matches_per_pos) *
1267 sizeof(((struct lzms_compressor*)0)->matches[0]);
1272 lzms_params_valid(const struct wimlib_compressor_params_header *_params)
1274 const struct wimlib_lzms_compressor_params *params =
1275 (const struct wimlib_lzms_compressor_params*)_params;
1277 if (params->hdr.size != sizeof(*params) ||
1278 params->max_match_length < params->min_match_length ||
1279 params->min_match_length < 2 ||
1280 params->optim_array_length == 0 ||
1281 min(params->max_match_length, params->nice_match_length) > 65536)
1287 const struct compressor_ops lzms_compressor_ops = {
1288 .params_valid = lzms_params_valid,
1289 .get_needed_memory = lzms_get_needed_memory,
1290 .create_compressor = lzms_create_compressor,
1291 .compress = lzms_compress,
1292 .free_compressor = lzms_free_compressor,