* and certain other details are quite similar, such as the method for storing
* Huffman codes. However, some of the main differences are:
*
- * - LZX preprocesses the data before attempting to compress it.
+ * - LZX preprocesses the data to attempt to make x86 machine code slightly more
+ * compressible before attempting to compress it further.
* - LZX uses a "main" alphabet which combines literals and matches, with the
* match symbols containing a "length header" (giving all or part of the match
* length) and a "position slot" (giving, roughly speaking, the order of
* dynamic Huffman blocks ("aligned offset" and "verbatim").
* - LZX has a minimum match length of 2 rather than 3.
* - In LZX, match offsets 0 through 2 actually represent entries in an LRU
- * queue of match offsets.
+ * queue of match offsets. This is very useful for certain types of files,
+ * such as binary files that have repeating records.
*
* Algorithms
* ==========
*
* The "slow" algorithm to generate LZX-compressed data is roughly as follows:
*
- * 1. Preprocess the input data to translate the targets of x86 call instructions
- * to absolute offsets.
+ * 1. Preprocess the input data to translate the targets of x86 call
+ * instructions to absolute offsets.
*
- * 2. Build the suffix array and inverse suffix array for the input data.
+ * 2. Build the suffix array and inverse suffix array for the input data. The
+ * suffix array contains the indices of all suffixes of the input data,
+ * sorted lexcographically by the corresponding suffixes. The "position" of
+ * a suffix is the index of that suffix in the original string, whereas the
+ * "rank" of a suffix is the index at which that suffix's position is found
+ * in the suffix array.
*
* 3. Build the longest common prefix array corresponding to the suffix array.
*
- * 4. For each suffix rank, find the highest lower suffix rank that has a
- * lower position, the lowest higher suffix rank that has a lower position,
- * and the length of the common prefix shared between each. (Position =
- * index of suffix in original string, rank = index of suffix in suffix
- * array.) This information is later used to link suffix ranks into a
- * doubly-linked list for searching the suffix array.
+ * 4. For each suffix, find the highest lower ranked suffix that has a lower
+ * position, the lowest higher ranked suffix that has a lower position, and
+ * the length of the common prefix shared between each. This information is
+ * later used to link suffix ranks into a doubly-linked list for searching
+ * the suffix array.
*
* 5. Set a default cost model for matches/literals.
*
- * 6. Determine the lowest cost sequence of LZ77 matches ((offset, length) pairs)
- * and literal bytes to divide the input into. Raw match-finding is done by
- * searching the suffix array using a linked list to avoid considering any
- * suffixes that start after the current position. Each run of the
- * match-finder returns the lowest-cost longest match as well as any shorter
- * matches that have even lower costs. Each such run also adds the suffix
- * rank of the current position into the linked list being used to search the
- * suffix array. Parsing, or match-choosing, is solved as a minimum-cost
- * path problem using a forward "optimal parsing" algorithm based on the
- * Deflate encoder from 7-Zip. This algorithm moves forward calculating the
- * minimum cost to reach each byte until either a very long match is found or
- * until a position is found at which no matches start or overlap.
+ * 6. Determine the lowest cost sequence of LZ77 matches ((offset, length)
+ * pairs) and literal bytes to divide the input into. Raw match-finding is
+ * done by searching the suffix array using a linked list to avoid
+ * considering any suffixes that start after the current position. Each run
+ * of the match-finder returns the approximate lowest-cost longest match as
+ * well as any shorter matches that have even lower approximate costs. Each
+ * such run also adds the suffix rank of the current position into the linked
+ * list being used to search the suffix array. Parsing, or match-choosing,
+ * is solved as a minimum-cost path problem using a forward "optimal parsing"
+ * algorithm based on the Deflate encoder from 7-Zip. This algorithm moves
+ * forward calculating the minimum cost to reach each byte until either a
+ * very long match is found or until a position is found at which no matches
+ * start or overlap.
*
* 7. Build the Huffman codes needed to output the matches/literals.
*
* 9. Output the resulting block using the match/literal sequences and the
* Huffman codes that were computed for the block.
*
+ * Note: the algorithm does not yet attempt to split the input into multiple LZX
+ * blocks, instead using a series of blocks of LZX_DIV_BLOCK_SIZE bytes.
+ *
* Fast algorithm
* --------------
*
* The fast algorithm (and the only one available in wimlib v1.5.1 and earlier)
* spends much less time on the main bottlenecks of the compression process ---
- * that is the match finding, match choosing, and block splitting. Matches are
- * found and chosen with hash chains using a greedy parse with one position of
- * look-ahead. No block splitting is done; only compressing the full input into
- * an aligned offset block is considered.
+ * that is, the match finding and match choosing. Matches are found and chosen
+ * with hash chains using a greedy parse with one position of look-ahead. No
+ * block splitting is done; only compressing the full input into an aligned
+ * offset block is considered.
*
* API
* ===
*
- * The old API (retained for backward compatibility) consists of just one function:
+ * The old API (retained for backward compatibility) consists of just one
+ * function:
*
* wimlib_lzx_compress()
*
* wimlib_lzx_alloc_context()
* wimlib_lzx_compress2()
* wimlib_lzx_free_context()
+ * wimlib_lzx_set_default_params()
*
* Both wimlib_lzx_compress() and wimlib_lzx_compress2() are designed to
- * compress an in-memory buffer of up to 32768 bytes. There is no sliding
- * window. This is suitable for the WIM format, which uses fixed-size chunks
- * that are seemingly always 32768 bytes. If needed, the compressor potentially
- * could be extended to support a larger and/or sliding window.
+ * compress an in-memory buffer of up to the window size, which can be any power
+ * of two between 2^15 and 2^21 inclusively. However, by default, the WIM
+ * format uses 2^15, and this is seemingly the only value that is compatible
+ * with WIMGAPI. In any case, the window is not a true "sliding window" since
+ * no data is ever "slid out" of the window. This is needed for the WIM format,
+ * which is designed such that chunks may be randomly accessed.
*
* Both wimlib_lzx_compress() and wimlib_lzx_compress2() return 0 if the data
* could not be compressed to less than the size of the uncompressed data.
* Again, this is suitable for the WIM format, which stores such data chunks
* uncompressed.
*
- * The functions in this API are exported from the library, although this is
- * only in case other programs happen to have uses for it other than WIM
+ * The functions in this LZX compression API are exported from the library,
+ * although with the possible exception of wimlib_lzx_set_default_params(), this
+ * is only in case other programs happen to have uses for it other than WIM
* reading/writing as already handled through the rest of the library.
*
* Acknowledgments
* Acknowledgments to several open-source projects and research papers that made
* it possible to implement this code:
*
- * - divsufsort (author: Yuta Mori), for the suffix array construction code.
+ * - divsufsort (author: Yuta Mori), for the suffix array construction code,
+ * located in a separate directory (divsufsort/).
*
* - "Linear-Time Longest-Common-Prefix Computation in Suffix Arrays and Its
* Applications" (Kasai et al. 2001), for the LCP array computation.
* (match-choosing).
*
* - zlib (author: Jean-loup Gailly and Mark Adler), for the hash table
- * match-finding algorithm.
+ * match-finding algorithm (used in lz77.c).
*
* - lzx-compress (author: Matthew T. Russotto), on which some parts of this
* code were originally based.
#include "wimlib.h"
#include "wimlib/compress.h"
+#include "wimlib/endianness.h"
#include "wimlib/error.h"
#include "wimlib/lzx.h"
#include "wimlib/util.h"
#include <string.h>
#ifdef ENABLE_LZX_DEBUG
-# include <wimlib/decompress.h>
+# include "wimlib/decompress.h"
#endif
#include "divsufsort/divsufsort.h"
-typedef freq_t input_idx_t;
-typedef u32 sym_cost_t;
typedef u32 block_cost_t;
-#define INFINITE_SYM_COST ((sym_cost_t)~0U)
#define INFINITE_BLOCK_COST ((block_cost_t)~0U)
#define LZX_OPTIM_ARRAY_SIZE 4096
-/* Currently, this constant can't simply be changed because the code currently
- * uses a static number of position slots (and may make other assumptions as
- * well). */
-#define LZX_MAX_WINDOW_SIZE 32768
-
-/* This may be WIM-specific */
-#define LZX_DEFAULT_BLOCK_SIZE 32768
+#define LZX_DIV_BLOCK_SIZE 32768
#define LZX_MAX_CACHE_PER_POS 10
/* Codewords for the LZX main, length, and aligned offset Huffman codes */
struct lzx_codewords {
- u16 main[LZX_MAINCODE_NUM_SYMBOLS];
+ u16 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
u16 len[LZX_LENCODE_NUM_SYMBOLS];
u16 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
};
* A 0 length means the codeword has zero frequency.
*/
struct lzx_lens {
- u8 main[LZX_MAINCODE_NUM_SYMBOLS];
+ u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
u8 len[LZX_LENCODE_NUM_SYMBOLS];
u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
};
* --- generally a high cost, since even if it gets used in the next iteration,
* it probably will not be used very times. */
struct lzx_costs {
- sym_cost_t main[LZX_MAINCODE_NUM_SYMBOLS];
- sym_cost_t len[LZX_LENCODE_NUM_SYMBOLS];
- sym_cost_t aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
+ u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
+ u8 len[LZX_LENCODE_NUM_SYMBOLS];
+ u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
};
/* The LZX main, length, and aligned offset Huffman codes */
/* Tables for tallying symbol frequencies in the three LZX alphabets */
struct lzx_freqs {
- freq_t main[LZX_MAINCODE_NUM_SYMBOLS];
- freq_t len[LZX_LENCODE_NUM_SYMBOLS];
- freq_t aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
+ input_idx_t main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
+ input_idx_t len[LZX_LENCODE_NUM_SYMBOLS];
+ input_idx_t aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
};
/* LZX intermediate match/literal format */
*
* 8-24 position footer. This is the offset of the real formatted
* offset from the position base. This can be at most 17 bits
- * (since lzx_extra_bits[LZX_NUM_POSITION_SLOTS - 1] is 17).
+ * (since lzx_extra_bits[LZX_MAX_POSITION_SLOTS - 1] is 17).
*
* 0-7 length of match, minus 2. This can be at most
* (LZX_MAX_MATCH_LEN - 2) == 255, so it will fit in 8 bits. */
* 0xe8 byte preprocessing is done directly on the data here before
* further compression.
*
- * Note that this compressor does *not* use a sliding window!!!! It's
- * not needed in the WIM format, since every chunk is compressed
+ * Note that this compressor does *not* use a real sliding window!!!!
+ * It's not needed in the WIM format, since every chunk is compressed
* independently. This is by design, to allow random access to the
* chunks.
*
* We reserve a few extra bytes to potentially allow reading off the end
* of the array in the match-finding code for optimization purposes.
*/
- u8 window[LZX_MAX_WINDOW_SIZE + 12];
+ u8 *window;
/* Number of bytes of data to be compressed, which is the number of
* bytes of data in @window that are actually valid. */
input_idx_t window_size;
+ /* Allocated size of the @window. */
+ input_idx_t max_window_size;
+
+ /* Number of symbols in the main alphabet (depends on the
+ * @max_window_size since it determines the maximum allowed offset). */
+ unsigned num_main_syms;
+
/* The current match offset LRU queue. */
struct lzx_lru_queue queue;
* block. */
struct lzx_match *chosen_matches;
- struct raw_match *cached_matches;
- unsigned cached_matches_pos;
- bool matches_cached;
-
/* Information about the LZX blocks the preprocessed input was divided
* into. */
struct lzx_block_spec *block_specs;
* codewords. */
struct lzx_codes zero_codes;
- /* Slow algorithm only: The current cost model. */
+ /* The current cost model. */
struct lzx_costs costs;
- /* Slow algorithm only: Suffix array for window.
- * This is a mapping from suffix rank to suffix position.
- *
- * Suffix rank means the index of the suffix in the sorted list of
- * suffixes, whereas suffix position means the index in the string at
- * which the suffix starts.
- */
+ /* Fast algorithm only: Array of hash table links. */
+ input_idx_t *prev_tab;
+
+ /* Suffix array for window.
+ * This is a mapping from suffix rank to suffix position. */
input_idx_t *SA;
- /* Slow algorithm only: Inverse suffix array for window.
+ /* Inverse suffix array for window.
* This is a mapping from suffix position to suffix rank.
- * In other words, if 0 <= r < window_size, then ISA[SA[r]] == r. */
+ * If 0 <= r < window_size, then ISA[SA[r]] == r. */
input_idx_t *ISA;
- /* Slow algorithm only: Longest Common Prefix array. LCP[i] is the
- * number of initial bytes that the suffixes at positions SA[i - 1] and
- * SA[i] share. LCP[0] is undefined. */
+ /* Longest common prefix array corresponding to the suffix array SA.
+ * LCP[i] is the length of the longest common prefix between the
+ * suffixes with positions SA[i - 1] and SA[i]. LCP[0] is undefined.
+ */
input_idx_t *LCP;
- /* Slow algorithm only: Suffix array links.
+ /* Suffix array links.
*
* During a linear scan of the input string to find matches, this array
* used to keep track of which rank suffixes in the suffix array appear
* list containing only suffixes that appear before that position. */
struct salink *salink;
- /* Slow algorithm only: Position in window of next match to return.
- * This cannot simply be modified, as the match-finder must still be
- * synchronized on the same position. To seek forwards or backwards,
- * use lzx_lz_skip_bytes() or lzx_lz_rewind_matchfinder(), respectively.
- */
+ /* Position in window of next match to return. */
input_idx_t match_window_pos;
- /* Slow algorithm only: The match-finder shall ensure the length of
- * matches does not exceed this position in the input. */
+ /* The match-finder shall ensure the length of matches does not exceed
+ * this position in the input. */
input_idx_t match_window_end;
+ /* Matches found by the match-finder are cached in the following array
+ * to achieve a slight speedup when the same matches are needed on
+ * subsequent passes. This is suboptimal because different matches may
+ * be preferred with different cost models, but seems to be a worthwhile
+ * speedup. */
+ struct raw_match *cached_matches;
+ unsigned cached_matches_pos;
+ bool matches_cached;
+
/* Slow algorithm only: Temporary space used for match-choosing
* algorithm.
*
u32 optimum_end_idx;
};
-/* Returns the LZX position slot that corresponds to a given formatted offset.
- *
- * Logically, this returns the smallest i such that
- * formatted_offset >= lzx_position_base[i].
- *
- * The actual implementation below takes advantage of the regularity of the
- * numbers in the lzx_position_base array to calculate the slot directly from
- * the formatted offset without actually looking at the array.
- */
-static _always_inline_attribute unsigned
-lzx_get_position_slot_raw(unsigned formatted_offset)
-{
-#if 0
- /*
- * Slots 36-49 (formatted_offset >= 262144) can be found by
- * (formatted_offset/131072) + 34 == (formatted_offset >> 17) + 34;
- * however, this check for formatted_offset >= 262144 is commented out
- * because WIM chunks cannot be that large.
- */
- if (formatted_offset >= 262144) {
- return (formatted_offset >> 17) + 34;
- } else
-#endif
- {
- /* Note: this part here only works if:
- *
- * 2 <= formatted_offset < 655360
- *
- * It is < 655360 because the frequency of the position bases
- * increases starting at the 655360 entry, and it is >= 2
- * because the below calculation fails if the most significant
- * bit is lower than the 2's place. */
- LZX_ASSERT(2 <= formatted_offset && formatted_offset < 655360);
- unsigned mssb_idx = bsr32(formatted_offset);
- return (mssb_idx << 1) |
- ((formatted_offset >> (mssb_idx - 1)) & 1);
- }
-}
-
-
/* Returns the LZX position slot that corresponds to a given match offset,
- * taking into account the recent offset queue (and optionally updating it). */
-static _always_inline_attribute unsigned
+ * taking into account the recent offset queue and updating it if the offset is
+ * found in it. */
+static unsigned
lzx_get_position_slot(unsigned offset, struct lzx_lru_queue *queue)
{
unsigned position_slot;
* LRU queue because repeat matches are simply
* swapped to the front. */
swap(queue->R[0], queue->R[i]);
- /* For recent offsets, the position slot is simply the
- * index of the entry in the queue. */
+ /* The resulting position slot is simply the first index
+ * at which the offset was found in the queue. */
return i;
}
}
* a set of tables that map symbols to codewords and codeword lengths. */
static void
lzx_make_huffman_codes(const struct lzx_freqs *freqs,
- struct lzx_codes *codes)
+ struct lzx_codes *codes,
+ unsigned num_main_syms)
{
- make_canonical_huffman_code(LZX_MAINCODE_NUM_SYMBOLS,
+ make_canonical_huffman_code(num_main_syms,
LZX_MAX_MAIN_CODEWORD_LEN,
freqs->main,
codes->lens.main,
}
/* Combine the position slot with the length header into a single symbol
- * that will be encoded with the main tree.
+ * that will be encoded with the main code.
*
* The actual main symbol is offset by LZX_NUM_CHARS because values
* under LZX_NUM_CHARS are used to indicate a literal byte rather than a
/* For aligned offset blocks with at least 3 extra bits, output the
* verbatim bits literally, then the aligned bits encoded using the
- * aligned offset tree. Otherwise, only the verbatim bits need to be
+ * aligned offset code. Otherwise, only the verbatim bits need to be
* output. */
if ((block_type == LZX_BLOCKTYPE_ALIGNED) && (num_extra_bits >= 3)) {
lzx_build_precode(const u8 lens[restrict],
const u8 prev_lens[restrict],
const unsigned num_syms,
- freq_t precode_freqs[restrict LZX_PRECODE_NUM_SYMBOLS],
+ input_idx_t precode_freqs[restrict LZX_PRECODE_NUM_SYMBOLS],
u8 output_syms[restrict num_syms],
u8 precode_lens[restrict LZX_PRECODE_NUM_SYMBOLS],
u16 precode_codewords[restrict LZX_PRECODE_NUM_SYMBOLS],
* literally.
*
* output_syms[] will be filled in with the length symbols that will be
- * output, including RLE codes, not yet encoded using the pre-tree.
+ * output, including RLE codes, not yet encoded using the precode.
*
* cur_run_len keeps track of how many code word lengths are in the
* current run of identical lengths. */
*
* The extra length symbol is encoded as a difference
* from the length of the codeword for the first symbol
- * in the run in the previous tree.
+ * in the run in the previous code.
* */
while (cur_run_len >= 4) {
unsigned additional_bits;
/* Any remaining lengths in the run are outputted without RLE,
* as a difference from the length of that codeword in the
- * previous tree. */
+ * previous code. */
while (cur_run_len > 0) {
signed char delta;
const u8 prev_lens[restrict],
unsigned num_syms)
{
- freq_t precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
+ input_idx_t precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
u8 output_syms[num_syms];
u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
u16 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
}
static void
-lzx_assert_codes_valid(const struct lzx_codes * codes)
+lzx_assert_codes_valid(const struct lzx_codes * codes, unsigned num_main_syms)
{
#ifdef ENABLE_LZX_DEBUG
unsigned i;
- for (i = 0; i < LZX_MAINCODE_NUM_SYMBOLS; i++)
+ for (i = 0; i < num_main_syms; i++)
LZX_ASSERT(codes->lens.main[i] <= LZX_MAX_MAIN_CODEWORD_LEN);
for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
const unsigned tablebits = 10;
u16 decode_table[(1 << tablebits) +
- (2 * max(LZX_MAINCODE_NUM_SYMBOLS, LZX_LENCODE_NUM_SYMBOLS))]
+ (2 * max(num_main_syms, LZX_LENCODE_NUM_SYMBOLS))]
_aligned_attribute(DECODE_TABLE_ALIGNMENT);
LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
- LZX_MAINCODE_NUM_SYMBOLS,
+ num_main_syms,
min(tablebits, LZX_MAINCODE_TABLEBITS),
codes->lens.main,
LZX_MAX_MAIN_CODEWORD_LEN));
static void
lzx_write_compressed_block(int block_type,
unsigned block_size,
+ unsigned max_window_size,
+ unsigned num_main_syms,
struct lzx_match * chosen_matches,
unsigned num_chosen_matches,
const struct lzx_codes * codes,
LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED ||
block_type == LZX_BLOCKTYPE_VERBATIM);
- LZX_ASSERT(block_size <= LZX_MAX_WINDOW_SIZE);
- LZX_ASSERT(num_chosen_matches <= LZX_MAX_WINDOW_SIZE);
- lzx_assert_codes_valid(codes);
+ lzx_assert_codes_valid(codes, num_main_syms);
/* The first three bits indicate the type of block and are one of the
* LZX_BLOCKTYPE_* constants. */
- bitstream_put_bits(ostream, block_type, LZX_BLOCKTYPE_NBITS);
+ bitstream_put_bits(ostream, block_type, 3);
- /* The next bit indicates whether the block size is the default (32768),
- * indicated by a 1 bit, or whether the block size is given by the next
- * 16 bits, indicated by a 0 bit. */
+ /* Output the block size.
+ *
+ * The original LZX format seemed to always encode the block size in 3
+ * bytes. However, the implementation in WIMGAPI, as used in WIM files,
+ * uses the first bit to indicate whether the block is the default size
+ * (32768) or a different size given explicitly by the next 16 bits.
+ *
+ * By default, this compressor uses a window size of 32768 and therefore
+ * follows the WIMGAPI behavior. However, this compressor also supports
+ * window sizes greater than 32768 bytes, which do not appear to be
+ * supported by WIMGAPI. In such cases, we retain the default size bit
+ * to mean a size of 32768 bytes but output non-default block size in 24
+ * bits rather than 16. The compatibility of this behavior is unknown
+ * because WIMs created with chunk size greater than 32768 can seemingly
+ * only be opened by wimlib anyway. */
if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
bitstream_put_bits(ostream, 1, 1);
} else {
bitstream_put_bits(ostream, 0, 1);
- bitstream_put_bits(ostream, block_size, LZX_BLOCKSIZE_NBITS);
+
+ if (max_window_size >= 65536)
+ bitstream_put_bits(ostream, block_size >> 16, 8);
+
+ bitstream_put_bits(ostream, block_size, 16);
}
/* Write out lengths of the main code. Note that the LZX specification
* incorrectly states that the aligned offset code comes after the
- * length code, but in fact it is the very first tree to be written
+ * length code, but in fact it is the very first code to be written
* (before the main code). */
if (block_type == LZX_BLOCKTYPE_ALIGNED)
for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
LZX_DEBUG("Writing main code...");
- /* Write the pre-tree and lengths for the first LZX_NUM_CHARS symbols in
+ /* Write the precode and lengths for the first LZX_NUM_CHARS symbols in
* the main code, which are the codewords for literal bytes. */
lzx_write_compressed_code(ostream,
codes->lens.main,
prev_codes->lens.main,
LZX_NUM_CHARS);
- /* Write the pre-tree and lengths for the rest of the main code, which
+ /* Write the precode and lengths for the rest of the main code, which
* are the codewords for match headers. */
lzx_write_compressed_code(ostream,
codes->lens.main + LZX_NUM_CHARS,
prev_codes->lens.main + LZX_NUM_CHARS,
- LZX_MAINCODE_NUM_SYMBOLS - LZX_NUM_CHARS);
+ num_main_syms - LZX_NUM_CHARS);
LZX_DEBUG("Writing length code...");
- /* Write the pre-tree and lengths for the length code. */
+ /* Write the precode and lengths for the length code. */
lzx_write_compressed_code(ostream,
codes->lens.len,
prev_codes->lens.len,
static void
lzx_write_all_blocks(struct lzx_compressor *ctx, struct output_bitstream *ostream)
{
+
const struct lzx_codes *prev_codes = &ctx->zero_codes;
for (unsigned i = 0; i < ctx->num_blocks; i++) {
const struct lzx_block_spec *spec = &ctx->block_specs[i];
lzx_write_compressed_block(spec->block_type,
spec->block_size,
+ ctx->max_window_size,
+ ctx->num_main_syms,
&ctx->chosen_matches[spec->chosen_matches_start_pos],
spec->num_chosen_matches,
&spec->codes,
prev_codes,
ostream);
+
prev_codes = &spec->codes;
}
}
/* Constructs an LZX match from a literal byte and updates the main code symbol
* frequencies. */
static u32
-lzx_record_literal(u8 literal, void *_freqs)
+lzx_tally_literal(u8 lit, struct lzx_freqs *freqs)
{
- struct lzx_freqs *freqs = _freqs;
-
- freqs->main[literal]++;
-
- return (u32)literal;
+ freqs->main[lit]++;
+ return (u32)lit;
}
/* Constructs an LZX match from an offset and a length, and updates the LRU
* alphabets. The return value is a 32-bit number that provides the match in an
* intermediate representation documented below. */
static u32
-lzx_record_match(unsigned match_offset, unsigned match_len,
- void *_freqs, void *_queue)
+lzx_tally_match(unsigned match_len, unsigned match_offset,
+ struct lzx_freqs *freqs, struct lzx_lru_queue *queue)
{
- struct lzx_freqs *freqs = _freqs;
- struct lzx_lru_queue *queue = _queue;
unsigned position_slot;
unsigned position_footer;
u32 len_header;
freqs->aligned[position_footer & 7]++;
/* Pack the position slot, position footer, and match length into an
- * intermediate representation.
- *
- * bits description
- * ---- -----------------------------------------------------------
- *
- * 31 1 if a match, 0 if a literal.
- *
- * 30-25 position slot. This can be at most 50, so it will fit in 6
- * bits.
- *
- * 8-24 position footer. This is the offset of the real formatted
- * offset from the position base. This can be at most 17 bits
- * (since lzx_extra_bits[LZX_NUM_POSITION_SLOTS - 1] is 17).
- *
- * 0-7 length of match, offset by 2. This can be at most
- * (LZX_MAX_MATCH_LEN - 2) == 255, so it will fit in 8 bits. */
- BUILD_BUG_ON(LZX_NUM_POSITION_SLOTS > 64);
- LZX_ASSERT(lzx_get_num_extra_bits(LZX_NUM_POSITION_SLOTS - 1) <= 17);
- BUILD_BUG_ON(LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1 > 256);
+ * intermediate representation. See `struct lzx_match' for details.
+ */
+ LZX_ASSERT(LZX_MAX_POSITION_SLOTS <= 64);
+ LZX_ASSERT(lzx_get_num_extra_bits(LZX_MAX_POSITION_SLOTS - 1) <= 17);
+ LZX_ASSERT(LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1 <= 256);
+
+ LZX_ASSERT(position_slot <= (1U << (31 - 25)) - 1);
+ LZX_ASSERT(position_footer <= (1U << (25 - 8)) - 1);
+ LZX_ASSERT(adjusted_match_len <= (1U << (8 - 0)) - 1);
return 0x80000000 |
(position_slot << 25) |
(position_footer << 8) |
(adjusted_match_len);
}
+struct lzx_record_ctx {
+ struct lzx_freqs freqs;
+ struct lzx_lru_queue queue;
+ struct lzx_match *matches;
+};
+
+static void
+lzx_record_match(unsigned len, unsigned offset, void *_ctx)
+{
+ struct lzx_record_ctx *ctx = _ctx;
+
+ (ctx->matches++)->data = lzx_tally_match(len, offset, &ctx->freqs, &ctx->queue);
+}
+
+static void
+lzx_record_literal(u8 lit, void *_ctx)
+{
+ struct lzx_record_ctx *ctx = _ctx;
+
+ (ctx->matches++)->data = lzx_tally_literal(lit, &ctx->freqs);
+}
+
/* Returns the cost, in bits, to output a literal byte using the specified cost
* model. */
-static sym_cost_t
+static unsigned
lzx_literal_cost(u8 c, const struct lzx_costs * costs)
{
return costs->main[c];
* codes, return the approximate number of bits it will take to represent this
* match in the compressed output. Take into account the match offset LRU
* queue and optionally update it. */
-static sym_cost_t
+static unsigned
lzx_match_cost(unsigned length, unsigned offset, const struct lzx_costs *costs,
struct lzx_lru_queue *queue)
{
unsigned position_slot;
unsigned len_header, main_symbol;
- sym_cost_t cost = 0;
+ unsigned cost = 0;
position_slot = lzx_get_position_slot(offset, queue);
len_header = min(length - LZX_MIN_MATCH_LEN, LZX_NUM_PRIMARY_LENS);
- main_symbol = (position_slot << 3) | len_header | LZX_NUM_CHARS;
+ main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
/* Account for main symbol. */
cost += costs->main[main_symbol];
}
-/* Very fast heuristic cost evaluation to use in the inner loop of the
- * match-finder. */
-static sym_cost_t
+/* Fast heuristic cost evaluation to use in the inner loop of the match-finder.
+ * Unlike lzx_match_cost() which does a true cost evaluation, this simply
+ * prioritize matches based on their offset. */
+static block_cost_t
lzx_match_cost_fast(unsigned offset, const struct lzx_lru_queue *queue)
{
+ /* It seems well worth it to take the time to give priority to recently
+ * used offsets. */
for (unsigned i = 0; i < LZX_NUM_RECENT_OFFSETS; i++)
if (offset == queue->R[i])
return i;
- BUILD_BUG_ON(LZX_MAX_WINDOW_SIZE >= (sym_cost_t)~0U);
+ BUILD_BUG_ON(LZX_MAX_WINDOW_SIZE >= (block_cost_t)~0U);
return offset;
}
lzx_set_costs(struct lzx_compressor * ctx, const struct lzx_lens * lens)
{
unsigned i;
+ unsigned num_main_syms = ctx->num_main_syms;
/* Main code */
- for (i = 0; i < LZX_MAINCODE_NUM_SYMBOLS; i++) {
+ for (i = 0; i < num_main_syms; i++) {
ctx->costs.main[i] = lens->main[i];
if (ctx->costs.main[i] == 0)
ctx->costs.main[i] = ctx->params.alg_params.slow.main_nostat_cost;
}
}
-/* Rewind the suffix array match-finder to the specified position.
- *
- * This undoes a series of updates by lzx_lz_update_salink(). */
-static void
-lzx_lz_rewind_matchfinder(struct lzx_compressor *ctx,
- const unsigned orig_pos)
-{
- LZX_DEBUG("Rewind match-finder %u => %u", ctx->match_window_pos, orig_pos);
-
- if (ctx->match_window_pos == orig_pos)
- return;
-
- LZX_ASSERT(ctx->match_window_pos > orig_pos);
- LZX_ASSERT(orig_pos == 0);
- ctx->matches_cached = true;
- ctx->cached_matches_pos = 0;
- ctx->match_window_pos = orig_pos;
-}
-
/*
* Use the suffix array match-finder to retrieve a list of LZ matches at the
* current position.
struct raw_match matches[const restrict],
const struct lzx_lru_queue * const restrict queue,
const unsigned min_match_len,
- const uint32_t max_matches_to_consider,
- const uint32_t max_matches_to_return)
+ const u32 max_matches_to_consider,
+ const u32 max_matches_to_return)
{
/* r = Rank of the suffix at the current position. */
const input_idx_t r = ISA[i];
* We keep track of this so that we can ignore shorter matches that do
* not have lower costs than a longer matches already found.
*/
- sym_cost_t best_cost = INFINITE_SYM_COST;
+ block_cost_t best_cost = INFINITE_BLOCK_COST;
/* count_remaining = maximum number of possible matches remaining to be
* considered. */
- uint32_t count_remaining = max_matches_to_consider;
+ u32 count_remaining = max_matches_to_consider;
/* pending = match currently being considered for a specific length. */
struct raw_match pending;
+ block_cost_t pending_cost;
while (lenL >= min_match_len || lenR >= min_match_len)
{
pending.len = lenL;
- pending.offset = (input_idx_t)~0U;
- sym_cost_t pending_cost = INFINITE_SYM_COST;
- sym_cost_t cost;
+ pending_cost = INFINITE_BLOCK_COST;
+ block_cost_t cost;
/* Extend left. */
if (lenL >= min_match_len && lenL >= lenR) {
return nmatches;
}
pending.len = lenL;
- pending.offset = (input_idx_t)~0U;
- pending_cost = INFINITE_SYM_COST;
+ pending_cost = INFINITE_BLOCK_COST;
}
if (lenL < min_match_len || lenL < lenR)
break;
break;
pending.len = lenR;
- pending.offset = (input_idx_t)~0U;
- pending_cost = INFINITE_SYM_COST;
+ pending_cost = INFINITE_BLOCK_COST;
}
R = link[R].next;
}
goto out;
out_save_pending:
- if (pending.offset != (input_idx_t)~0U)
+ if (pending_cost != INFINITE_BLOCK_COST)
matches[nmatches++] = pending;
out:
num_matches = matches[-1].len;
} else {
unsigned min_match_len = LZX_MIN_MATCH_LEN;
- if (min_match_len <= 2 && !ctx->params.alg_params.slow.use_len2_matches)
- min_match_len = 3;
- const uint32_t max_search_depth = ctx->params.alg_params.slow.max_search_depth;
- const uint32_t max_matches_per_pos = ctx->params.alg_params.slow.max_matches_per_pos;
+ if (!ctx->params.alg_params.slow.use_len2_matches)
+ min_match_len = max(min_match_len, 3);
+ const u32 max_search_depth = ctx->params.alg_params.slow.max_search_depth;
+ const u32 max_matches_per_pos = ctx->params.alg_params.slow.max_matches_per_pos;
if (unlikely(max_search_depth == 0 || max_matches_per_pos == 0))
num_matches = 0;
};
}
-#if 0
-static struct raw_match
-lzx_lz_get_greedy_match(struct lzx_compressor * ctx)
-{
- struct raw_match *matches;
-
- if (!lzx_lz_get_matches_caching(ctx, &ctx->queue, &matches))
- return (struct raw_match) {.len = 0};
-
- lzx_lz_skip_bytes(ctx, matches[0].len - 1);
- return matches[0];
-}
-#endif
-
-#if 0
-static struct raw_match
-lzx_lz_get_lazy_match(struct lzx_compressor * ctx)
-{
- unsigned num_matches;
- struct raw_match *matches;
- struct raw_match prev_match;
- struct lzx_lru_queue queue;
-
- if (ctx->optimum_cur_idx != ctx->optimum_end_idx)
- goto retopt;
-
- /* Check for matches at first position. */
- num_matches = lzx_lz_get_matches_caching(ctx, &ctx->queue, &matches);
-
- /* Return literal if no matches were found. */
- if (num_matches == 0)
- return (struct raw_match) { .len = 0 };
-
- /* Immediately choose match if longer than threshold. */
- if (matches[0].len > ctx->params.alg_params.slow.num_fast_bytes)
- goto savecur;
-
- ctx->optimum_cur_idx = ctx->optimum_end_idx = 0;
- for (;;) {
- prev_match = matches[0];
-
- /* Check for matches at next position. */
- num_matches = lzx_lz_get_matches_caching(ctx, &ctx->queue, &matches);
-
- /* Choose previous match if there is not a match at this
- * position. */
- if (num_matches == 0)
- goto saveprev;
-
- /* Choose previous match the longest match at the next position
- * is the same place, just one character shifted over. */
- if (matches[0].offset == prev_match.offset ||
- matches[0].len < prev_match.len)
- goto saveprev;
-
- struct lzx_lru_queue q1 = ctx->queue, q2 = ctx->queue;
- double lazycost = lzx_literal_cost(ctx->window[ctx->match_window_pos - 2],
- &ctx->costs) +
- lzx_match_cost(matches[0].len, matches[0].offset,
- &ctx->costs, &q1);
- double greedycost = lzx_match_cost(prev_match.len, prev_match.offset,
- &ctx->costs, &q2);
- lazycost *= (double)prev_match.len / (1 + matches[0].len);
-
- /* Choose previous match if greedy cost was lower. */
- if (greedycost <= lazycost)
- goto saveprev;
-
- /* Choose literal at the previous position. */
- ctx->optimum[ctx->optimum_end_idx++].next.link = 0;
-
-
- /* Immediately choose match if longer than threshold. */
- if (matches[0].len > ctx->params.alg_params.slow.num_fast_bytes)
- goto savecur;
- }
-
-savecur:
- lzx_lz_skip_bytes(ctx, 1);
- prev_match = matches[0];
-
-saveprev:
- lzx_lz_skip_bytes(ctx, prev_match.len - 2);
- ctx->optimum[ctx->optimum_end_idx].next.link = prev_match.len;
- ctx->optimum[ctx->optimum_end_idx].next.match_offset = prev_match.offset;
- ctx->optimum_end_idx++;
-retopt:
- prev_match.len = ctx->optimum[ctx->optimum_cur_idx].next.link;
- prev_match.offset = ctx->optimum[ctx->optimum_cur_idx].next.match_offset;
- ctx->optimum_cur_idx++;
- return prev_match;
-}
-#endif
-
-
/*
* lzx_lz_get_near_optimal_match() -
*
* model rather than simply assuming that longer is better.
*
* Still, this is not truly an optimal parser because very long matches are
- * taken immediately. This is done to avoid considering many different
- * alternatives that are unlikely to significantly be better.
+ * taken immediately, and the raw match-finder takes some shortcuts. This is
+ * done to avoid considering many different alternatives that are unlikely to
+ * be significantly better.
*
* This algorithm is based on that used in 7-Zip's DEFLATE encoder.
*
* ctx->optimum (internal state; leave uninitialized)
* ctx->optimum_cur_idx (must set to 0 before first call)
* ctx->optimum_end_idx (must set to 0 before first call)
- * ctx->SA (must be built before first call)
- * ctx->ISA (must be built before first call)
- * ctx->salink (must be built before first call)
- * ctx->match_window_pos (must initialize to position of next match to
- * return; subsequent calls return subsequent
- * matches)
- * ctx->match_window_end (must initialize to limit of match-finding region;
- * subsequent calls use the same limit)
+ *
+ * Plus any state used by the raw match-finder.
*
* The return value is a (length, offset) pair specifying the match or literal
* chosen. For literals, the length is less than LZX_MIN_MATCH_LEN and the
* Set default symbol costs.
*/
static void
-lzx_set_default_costs(struct lzx_costs * costs)
+lzx_set_default_costs(struct lzx_costs * costs, unsigned num_main_syms)
{
unsigned i;
costs->main[i] = 8;
/* Match header symbols */
- for (; i < LZX_MAINCODE_NUM_SYMBOLS; i++)
+ for (; i < num_main_syms; i++)
costs->main[i] = 10;
/* Length symbols */
}
/* Find a near-optimal sequence of matches/literals with which to output the
- * specified LZX block, and set its type to that which has the minimum cost to
+ * specified LZX block, then set its type to that which has the minimum cost to
* output. */
static void
lzx_optimize_block(struct lzx_compressor *ctx, struct lzx_block_spec *spec,
unsigned num_passes)
{
- struct lzx_lru_queue orig_queue = ctx->queue;
+ const struct lzx_lru_queue orig_queue = ctx->queue;
struct lzx_freqs freqs;
+ unsigned orig_window_pos = spec->window_pos;
+ unsigned orig_cached_pos = ctx->cached_matches_pos;
+
+ LZX_ASSERT(ctx->match_window_pos == spec->window_pos);
+
ctx->match_window_end = spec->window_pos + spec->block_size;
spec->chosen_matches_start_pos = spec->window_pos;
* computed from the previous pass. */
for (unsigned pass = 0; pass < num_passes; pass++) {
- lzx_lz_rewind_matchfinder(ctx, spec->window_pos);
+ ctx->match_window_pos = orig_window_pos;
+ ctx->cached_matches_pos = orig_cached_pos;
ctx->queue = orig_queue;
spec->num_chosen_matches = 0;
memset(&freqs, 0, sizeof(freqs));
raw_match = lzx_lz_get_near_optimal_match(ctx);
if (raw_match.len >= LZX_MIN_MATCH_LEN) {
- lzx_match.data = lzx_record_match(raw_match.offset, raw_match.len,
- &freqs, &ctx->queue);
+ lzx_match.data = lzx_tally_match(raw_match.len, raw_match.offset,
+ &freqs, &ctx->queue);
i += raw_match.len;
} else {
- lzx_match.data = lzx_record_literal(ctx->window[i], &freqs);
+ lzx_match.data = lzx_tally_literal(ctx->window[i], &freqs);
i += 1;
}
ctx->chosen_matches[spec->chosen_matches_start_pos +
spec->num_chosen_matches++] = lzx_match;
}
- lzx_make_huffman_codes(&freqs, &spec->codes);
+ lzx_make_huffman_codes(&freqs, &spec->codes,
+ ctx->num_main_syms);
if (pass < num_passes - 1)
lzx_set_costs(ctx, &spec->codes.lens);
+ ctx->matches_cached = true;
}
spec->block_type = lzx_choose_verbatim_or_aligned(&freqs, &spec->codes);
+ ctx->matches_cached = false;
}
static void
lzx_optimize_block(ctx, &ctx->block_specs[i], num_passes);
}
-static bool entropy_val_tab_inited = false;
-static double entropy_val_tab[LZX_MAX_WINDOW_SIZE];
-static pthread_mutex_t entropy_val_tab_mutex = PTHREAD_MUTEX_INITIALIZER;
-
-static double entropy_val(unsigned count)
-{
- /*return count * log(count);*/
- return entropy_val_tab[count];
-}
-
-/* Split a LZX block into several if it is advantageous to do so.
- *
- * TODO: This doesn't work very well yet. Should optimal parsing be done
- * before or after splitting? */
-static void
-lzx_block_split(const u32 matches[restrict],
- const input_idx_t n,
- const double epsilon,
- const unsigned max_num_blocks,
- const unsigned min_block_len,
- struct lzx_block_spec block_specs[restrict],
- unsigned * const restrict num_blocks_ret)
-{
- const double block_overhead = 1500;
-
- if (!entropy_val_tab_inited) {
- pthread_mutex_lock(&entropy_val_tab_mutex);
- if (!entropy_val_tab_inited) {
- entropy_val_tab[0] = 0;
- for (input_idx_t i = 1; i < LZX_MAX_WINDOW_SIZE; i++)
- entropy_val_tab[i] = i * log2(i);
- entropy_val_tab_inited = true;
- }
- pthread_mutex_unlock(&entropy_val_tab_mutex);
- }
-
- u16 main_syms[n];
- u8 len_syms[n];
- u8 aligned_syms[n];
- input_idx_t orig_input_indices[n + 1];
-
- LZX_ASSERT(epsilon >= 0);
- LZX_ASSERT(max_num_blocks >= 1);
-
- /* For convenience, extract the main, length, and aligned symbols from
- * the matches. Every position will have a main symbol, but not every
- * position will have a length and aligned symbol. Special values
- * larger than the valid symbols are used to indicate the absense of a
- * symbol. */
- orig_input_indices[0] = 0;
- for (input_idx_t i = 0, orig_input_idx = 0; i < n; i++) {
- u32 match = matches[i];
- u16 main_sym;
- u8 len_sym = LZX_LENCODE_NUM_SYMBOLS;
- u8 aligned_sym = LZX_ALIGNEDCODE_NUM_SYMBOLS;
- if (match & 0x80000000) {
- unsigned match_len_minus_2 = match & 0xff;
- unsigned position_footer = (match >> 8) & 0x1ffff;
- unsigned position_slot = (match >> 25) & 0x3f;
- unsigned len_header;
-
- if (match_len_minus_2 < LZX_NUM_PRIMARY_LENS) {
- len_header = match_len_minus_2;
- } else {
- len_header = LZX_NUM_PRIMARY_LENS;
- len_sym = match_len_minus_2 - LZX_NUM_PRIMARY_LENS;
- }
- main_sym = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
- if (position_slot >= 8)
- aligned_sym = position_footer & 7;
- orig_input_idx += match_len_minus_2 + 2;
- } else {
- main_sym = match;
- orig_input_idx++;
- }
- main_syms[i] = main_sym;
- len_syms[i] = len_sym;
- aligned_syms[i] = aligned_sym;
- orig_input_indices[i + 1] = orig_input_idx;
- }
-
- /* Compute the number of sliding windows that will be used for the
- * entropy calculations. */
- int num_windows = 0;
- unsigned window_len;
- {
- double e = min_block_len;
- do {
- window_len = e;
- num_windows++;
- e *= epsilon + 1;
- } while (window_len < n);
- }
-
- /* Compute the length of each sliding window. */
- unsigned window_lens[num_windows];
- {
- double e = min_block_len;
- unsigned window_idx = 0;
- do {
- window_len = e;
- window_lens[window_idx++] = min(window_len, n);
- e *= epsilon + 1;
- } while (window_len < n);
- }
-
- /* Best estimated compression size, in bits, found so far for the input
- * matches up to each position. */
- unsigned shortest_paths[n + 1];
-
- /* Pointers to follow to get the sequence of blocks that represents the
- * shortest path (in terms of estimated compressed size) up to each
- * position in the input matches. */
- input_idx_t back_ptrs[n + 1];
-
- for (input_idx_t i = 0; i < n + 1; i++) {
- shortest_paths[i] = ~0U;
- back_ptrs[i] = 0;
- }
- shortest_paths[0] = 0;
-
- {
- /* Initialize the per-window symbol and entropy counters */
- input_idx_t mainsym_ctrs[num_windows][LZX_MAINCODE_NUM_SYMBOLS];
- input_idx_t lensym_ctrs[num_windows][LZX_LENCODE_NUM_SYMBOLS + 1];
- input_idx_t alignedsym_ctrs[num_windows][LZX_ALIGNEDCODE_NUM_SYMBOLS + 1];
- ZERO_ARRAY(mainsym_ctrs);
- ZERO_ARRAY(lensym_ctrs);
- ZERO_ARRAY(alignedsym_ctrs);
-
- {
- int start_win_idx = 0;
- for (input_idx_t i = 0; i < n; i++) {
- while (i >= window_lens[start_win_idx])
- start_win_idx++;
- for (int j = start_win_idx; j < num_windows; j++) {
- mainsym_ctrs[j][main_syms[i]]++;
- lensym_ctrs[j][len_syms[i]]++;
- alignedsym_ctrs[j][aligned_syms[i]]++;
- }
- }
- }
-
- double entropy_ctrs[num_windows];
- for (int i = 0; i < num_windows; i++) {
- entropy_ctrs[i] = 0;
- for (unsigned j = 0; j < LZX_MAINCODE_NUM_SYMBOLS; j++)
- entropy_ctrs[i] += entropy_val(mainsym_ctrs[i][j]);
- for (unsigned j = 0; j < LZX_LENCODE_NUM_SYMBOLS; j++)
- entropy_ctrs[i] += entropy_val(lensym_ctrs[i][j]);
- for (unsigned j = 0; j < LZX_ALIGNEDCODE_NUM_SYMBOLS; j++)
- entropy_ctrs[i] += entropy_val(alignedsym_ctrs[i][j]);
- }
-
- /* Slide the windows along the input and compute the shortest
- * path to each position in the matches. */
- int end_window_idx = (int)num_windows - 1;
- for (input_idx_t i = 0; i < n; i++) {
- for (int j = 0; j <= end_window_idx; j++) {
- if (shortest_paths[i] == ~0U)
- continue;
- unsigned num_mainsyms = window_lens[j];
- unsigned num_lensyms = window_lens[j] -
- lensym_ctrs[j][LZX_LENCODE_NUM_SYMBOLS];
- unsigned num_alignedsyms = window_lens[j] -
- alignedsym_ctrs[j][LZX_ALIGNEDCODE_NUM_SYMBOLS];
- unsigned entropy = entropy_val(num_mainsyms) +
- entropy_val(num_lensyms) +
- entropy_val(num_alignedsyms) -
- entropy_ctrs[j];
- unsigned est_csize = entropy + block_overhead;
-
- unsigned end_idx = i + window_lens[j];
- if (est_csize + shortest_paths[i] < shortest_paths[end_idx]) {
- shortest_paths[end_idx] = est_csize + shortest_paths[i];
- back_ptrs[end_idx] = i;
- }
- }
- /* Remove left symbol from windows */
- for (int j = 0; j <= end_window_idx; j++) {
- input_idx_t orig_maincnt = mainsym_ctrs[j][main_syms[i]]--;
- entropy_ctrs[j] -= entropy_val(orig_maincnt);
- entropy_ctrs[j] += entropy_val(orig_maincnt - 1);
-
- input_idx_t orig_lencnt =
- lensym_ctrs[j][len_syms[i]]--;
- if (len_syms[i] != LZX_LENCODE_NUM_SYMBOLS) {
- entropy_ctrs[j] -= entropy_val(orig_lencnt);
- entropy_ctrs[j] += entropy_val(orig_lencnt - 1);
- }
-
- input_idx_t orig_alignedcnt =
- alignedsym_ctrs[j][aligned_syms[i]]--;
- if (aligned_syms[i] != LZX_ALIGNEDCODE_NUM_SYMBOLS) {
- entropy_ctrs[j] -= entropy_val(orig_alignedcnt);
- entropy_ctrs[j] += entropy_val(orig_alignedcnt - 1);
- }
- }
-
- /* Calculate index of longest window remaining */
- while (end_window_idx >= 0 && window_lens[end_window_idx] >= n - i)
- end_window_idx--;
-
- /* Append right symbol to windows */
- for (int j = 0; j <= end_window_idx; j++) {
- input_idx_t orig_maincnt = mainsym_ctrs[j][
- main_syms[i + window_lens[j]]]++;
- entropy_ctrs[j] -= entropy_val(orig_maincnt);
- entropy_ctrs[j] += entropy_val(orig_maincnt + 1);
-
- input_idx_t orig_lencnt =
- lensym_ctrs[j][len_syms[i + window_lens[j]]]++;
- if (len_syms[i + window_lens[j]] != LZX_LENCODE_NUM_SYMBOLS) {
- entropy_ctrs[j] -= entropy_val(orig_lencnt);
- entropy_ctrs[j] += entropy_val(orig_lencnt + 1);
- }
-
- input_idx_t orig_alignedcnt =
- alignedsym_ctrs[j][aligned_syms[i + window_lens[j]]]++;
- if (aligned_syms[i + window_lens[j]] != LZX_ALIGNEDCODE_NUM_SYMBOLS) {
- entropy_ctrs[j] -= entropy_val(orig_alignedcnt);
- entropy_ctrs[j] += entropy_val(orig_alignedcnt + 1);
- }
- }
- }
- }
-
-#if 0
- /* If no cost was computed for the first block (due to it being shorter
- * than all the windows), merge it with the second block. */
- for (input_idx_t i = n; i != 0; i = back_ptrs[i])
- if (back_ptrs[i] != 0 && shortest_paths[back_ptrs[i]] == ~0U)
- back_ptrs[i] = 0;
-#endif
-
- /* Calculate number of blocks */
- input_idx_t num_blocks = 0;
- for (input_idx_t i = n; i != 0; i = back_ptrs[i])
- num_blocks++;
-
- while (num_blocks > max_num_blocks) {
- LZX_DEBUG("Joining blocks to bring total under max_num_blucks=%u",
- max_num_blocks);
- back_ptrs[n] = back_ptrs[back_ptrs[n]];
- num_blocks--;
- }
-
- LZX_ASSERT(num_blocks != 0);
-
- /* fill in the 'struct lzx_block_spec' for each block */
- for (input_idx_t i = n, j = num_blocks - 1; i != 0; i = back_ptrs[i], j--) {
-
- block_specs[j].chosen_matches_start_pos = back_ptrs[i];
- block_specs[j].num_chosen_matches = i - back_ptrs[i];
- block_specs[j].window_pos = orig_input_indices[back_ptrs[i]];
- block_specs[j].block_size = orig_input_indices[i] -
- orig_input_indices[back_ptrs[i]];
- /*block_specs[j].est_csize = (shortest_paths[i] -*/
- /*shortest_paths[back_ptrs[i]]) / 8;*/
-
- LZX_DEBUG("block match_indices [%u, %u) est_csize %u bits\n",
- back_ptrs[i], i,
- shortest_paths[i] - shortest_paths[back_ptrs[i]]);
-
- struct lzx_freqs freqs = {};
-
- for (input_idx_t k = back_ptrs[i]; k < i; k++) {
- freqs.main[main_syms[k]]++;
- if (len_syms[k] != LZX_LENCODE_NUM_SYMBOLS)
- freqs.len[len_syms[k]]++;
- if (aligned_syms[k] != LZX_LENCODE_NUM_SYMBOLS)
- freqs.aligned[aligned_syms[k]]++;
- }
- lzx_make_huffman_codes(&freqs, &block_specs[j].codes);
-
- block_specs[j].block_type = lzx_choose_verbatim_or_aligned(&freqs,
- &block_specs[j].codes);
- }
- *num_blocks_ret = num_blocks;
-}
-
-
/* Initialize the suffix array match-finder for the specified input. */
static void
lzx_lz_init_matchfinder(const u8 T[const restrict],
/* Compute SA (Suffix Array). */
{
- saidx_t sa[n];
/* ISA and link are used as temporary space. */
- BUILD_BUG_ON(LZX_MAX_WINDOW_SIZE * sizeof(ISA[0]) < 256 * sizeof(saidx_t));
- BUILD_BUG_ON(LZX_MAX_WINDOW_SIZE * 2 * sizeof(link[0]) < 256 * 256 * sizeof(saidx_t));
- divsufsort(T, sa, n, (saidx_t*)ISA, (saidx_t*)link);
- for (input_idx_t i = 0; i < n; i++)
- SA[i] = sa[i];
+ BUILD_BUG_ON(LZX_MIN_WINDOW_SIZE * sizeof(ISA[0]) < 256 * sizeof(saidx_t));
+ BUILD_BUG_ON(LZX_MIN_WINDOW_SIZE * 2 * sizeof(link[0]) < 256 * 256 * sizeof(saidx_t));
+
+ if (sizeof(input_idx_t) == sizeof(saidx_t)) {
+ divsufsort(T, SA, n, (saidx_t*)ISA, (saidx_t*)link);
+ } else {
+ saidx_t sa[n];
+ divsufsort(T, sa, n, (saidx_t*)ISA, (saidx_t*)link);
+ for (input_idx_t i = 0; i < n; i++)
+ SA[i] = sa[i];
+ }
}
#ifdef ENABLE_LZX_DEBUG
}
link[r].next = t;
link[r].lcpnext = min(l, max_match_len);
- LZX_ASSERT(t == (input_idx_t)~0 || l <= n - SA[t]);
+ LZX_ASSERT(t == (input_idx_t)~0U || l <= n - SA[t]);
LZX_ASSERT(l <= n - SA[r]);
LZX_ASSERT(memcmp(&T[SA[r]], &T[SA[t]], l) == 0);
}
ctx->match_window_pos = 0;
/* Set up a default cost model. */
- lzx_set_default_costs(&ctx->costs);
+ lzx_set_default_costs(&ctx->costs, ctx->num_main_syms);
- /* Initially assume that the entire input will be one LZX block. */
- ctx->block_specs[0].block_type = LZX_BLOCKTYPE_ALIGNED;
- ctx->block_specs[0].window_pos = 0;
- ctx->block_specs[0].block_size = ctx->window_size;
- ctx->num_blocks = 1;
+ ctx->num_blocks = DIV_ROUND_UP(ctx->window_size, LZX_DIV_BLOCK_SIZE);
+ for (unsigned i = 0; i < ctx->num_blocks; i++) {
+ unsigned pos = LZX_DIV_BLOCK_SIZE * i;
+ ctx->block_specs[i].window_pos = pos;
+ ctx->block_specs[i].block_size = min(ctx->window_size - pos, LZX_DIV_BLOCK_SIZE);
+ }
- /* Perform near-optimal LZ parsing. */
+ /* Determine sequence of matches/literals to output for each block. */
lzx_optimize_blocks(ctx);
-
- /* Possibly divide up the LZX block. */
- const unsigned max_num_blocks = 1U << ctx->params.alg_params.slow.num_split_passes;
- if (max_num_blocks > 1) {
- const double epsilon = 0.2;
- const unsigned min_block_len = 500;
-
- lzx_block_split((const u32*)ctx->chosen_matches,
- ctx->block_specs[0].num_chosen_matches,
- epsilon, max_num_blocks, min_block_len,
- ctx->block_specs, &ctx->num_blocks);
- }
}
/*
* ctx->window[]
* ctx->window_size
*
- * Working space:
- * ctx->queue
- *
* Output --- the block specification and the corresponding match/literal data:
*
* ctx->block_specs[]
static void
lzx_prepare_block_fast(struct lzx_compressor * ctx)
{
- unsigned num_matches;
- struct lzx_freqs freqs;
+ struct lzx_record_ctx record_ctx;
struct lzx_block_spec *spec;
/* Parameters to hash chain LZ match finder
* aren't worth choosing when using greedy or lazy parsing. */
.min_match = 3,
.max_match = LZX_MAX_MATCH_LEN,
+ .max_offset = LZX_MAX_WINDOW_SIZE,
.good_match = LZX_MAX_MATCH_LEN,
.nice_match = LZX_MAX_MATCH_LEN,
.max_chain_len = LZX_MAX_MATCH_LEN,
};
/* Initialize symbol frequencies and match offset LRU queue. */
- memset(&freqs, 0, sizeof(struct lzx_freqs));
- lzx_lru_queue_init(&ctx->queue);
+ memset(&record_ctx.freqs, 0, sizeof(struct lzx_freqs));
+ lzx_lru_queue_init(&record_ctx.queue);
+ record_ctx.matches = ctx->chosen_matches;
/* Determine series of matches/literals to output. */
- num_matches = lz_analyze_block(ctx->window,
- ctx->window_size,
- (u32*)ctx->chosen_matches,
- lzx_record_match,
- lzx_record_literal,
- &freqs,
- &ctx->queue,
- &freqs,
- &lzx_lz_params);
-
+ lz_analyze_block(ctx->window,
+ ctx->window_size,
+ lzx_record_match,
+ lzx_record_literal,
+ &record_ctx,
+ &lzx_lz_params,
+ ctx->prev_tab);
/* Set up block specification. */
spec = &ctx->block_specs[0];
spec->block_type = LZX_BLOCKTYPE_ALIGNED;
spec->window_pos = 0;
spec->block_size = ctx->window_size;
- spec->num_chosen_matches = num_matches;
+ spec->num_chosen_matches = (record_ctx.matches - ctx->chosen_matches);
spec->chosen_matches_start_pos = 0;
- lzx_make_huffman_codes(&freqs, &spec->codes);
+ lzx_make_huffman_codes(&record_ctx.freqs, &spec->codes,
+ ctx->num_main_syms);
ctx->num_blocks = 1;
}
{
struct lzx_compressor *ctx = (struct lzx_compressor*)lzx_ctx;
struct output_bitstream ostream;
- unsigned compressed_len;
+ input_idx_t compressed_len;
if (uncompressed_len < 100) {
LZX_DEBUG("Too small to bother compressing.");
return 0;
}
- if (uncompressed_len > 32768) {
- LZX_DEBUG("Only up to 32768 bytes of uncompressed data are supported.");
+ if (uncompressed_len > ctx->max_window_size) {
+ LZX_DEBUG("Can't compress %u bytes using window of %u bytes!",
+ uncompressed_len, ctx->max_window_size);
return 0;
}
- wimlib_assert(lzx_ctx != NULL);
-
LZX_DEBUG("Attempting to compress %u bytes...", uncompressed_len);
/* The input data must be preprocessed. To avoid changing the original
lzx_write_all_blocks(ctx, &ostream);
LZX_DEBUG("Flushing bitstream...");
- if (flush_output_bitstream(&ostream)) {
- /* If the bitstream cannot be flushed, then the output space was
- * exhausted. */
+ compressed_len = flush_output_bitstream(&ostream);
+ if (compressed_len == ~(input_idx_t)0) {
LZX_DEBUG("Data did not compress to less than original length!");
return 0;
}
- /* Compute the length of the compressed data. */
- compressed_len = ostream.bit_output - (u8*)compressed_data;
-
LZX_DEBUG("Done: compressed %u => %u bytes.",
uncompressed_len, compressed_len);
/* Verify that we really get the same thing back when decompressing.
- * TODO: Disable this check by default on the slow algorithm. */
+ * Although this could be disabled by default in all cases, it only
+ * takes around 2-3% of the running time of the slow algorithm to do the
+ * verification. */
if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_SLOW
#if defined(ENABLE_LZX_DEBUG) || defined(ENABLE_VERIFY_COMPRESSION)
|| 1
#endif
)
{
- u8 buf[uncompressed_len];
- int ret;
+ /* The decompression buffer can be any temporary space that's no
+ * longer needed. */
+ u8 *buf = (u8*)(ctx->SA ? ctx->SA : ctx->prev_tab);
- ret = wimlib_lzx_decompress(compressed_data, compressed_len,
- buf, uncompressed_len);
- if (ret) {
+ if (wimlib_lzx_decompress2(compressed_data, compressed_len,
+ buf, uncompressed_len, ctx->max_window_size))
+ {
ERROR("Failed to decompress data we "
"compressed using LZX algorithm");
wimlib_assert(0);
LZX_DEBUG("Invalid aligned_nostat_cost!");
return false;
}
-
- if (params->alg_params.slow.num_split_passes > 31) {
- LZX_DEBUG("Invalid num_split_passes!");
- return false;
- }
}
return true;
}
+/* API function documented in wimlib.h */
WIMLIBAPI int
wimlib_lzx_set_default_params(const struct wimlib_lzx_params * params)
{
/* API function documented in wimlib.h */
WIMLIBAPI int
-wimlib_lzx_alloc_context(const struct wimlib_lzx_params *params,
+wimlib_lzx_alloc_context(u32 window_size,
+ const struct wimlib_lzx_params *params,
struct wimlib_lzx_context **ctx_pp)
{
LZX_DEBUG("Allocating LZX context...");
+ if (!lzx_window_size_valid(window_size))
+ return WIMLIB_ERR_INVALID_PARAM;
+
struct lzx_compressor *ctx;
static const struct wimlib_lzx_params fast_default = {
.use_len2_matches = 1,
.num_fast_bytes = 32,
.num_optim_passes = 2,
- .num_split_passes = 0,
.max_search_depth = 50,
.max_matches_per_pos = 3,
.main_nostat_cost = 15,
if (ctx_pp) {
ctx = *(struct lzx_compressor**)ctx_pp;
- if (ctx && lzx_params_compatible(&ctx->params, params))
+ if (ctx &&
+ lzx_params_compatible(&ctx->params, params) &&
+ ctx->max_window_size == window_size)
return 0;
} else {
LZX_DEBUG("Check parameters only.");
LZX_DEBUG("Allocating memory.");
- ctx = MALLOC(sizeof(struct lzx_compressor));
+ ctx = CALLOC(1, sizeof(struct lzx_compressor));
if (ctx == NULL)
goto err;
- size_t block_specs_length;
+ ctx->num_main_syms = lzx_get_num_main_syms(window_size);
+ ctx->max_window_size = window_size;
+ ctx->window = MALLOC(window_size + 12);
+ if (ctx->window == NULL)
+ goto err;
+
+ if (params->algorithm == WIMLIB_LZX_ALGORITHM_FAST) {
+ ctx->prev_tab = MALLOC(window_size * sizeof(ctx->prev_tab[0]));
+ if (ctx->prev_tab == NULL)
+ goto err;
+ }
- if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW)
- block_specs_length = 1U << params->alg_params.slow.num_split_passes;
- else
- block_specs_length = 1U;
+ size_t block_specs_length = DIV_ROUND_UP(window_size, LZX_DIV_BLOCK_SIZE);
ctx->block_specs = MALLOC(block_specs_length * sizeof(ctx->block_specs[0]));
if (ctx->block_specs == NULL)
- goto err_free_ctx;
+ goto err;
if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
- ctx->SA = MALLOC(3U * LZX_MAX_WINDOW_SIZE * sizeof(ctx->SA[0]));
+ ctx->SA = MALLOC(3U * window_size * sizeof(ctx->SA[0]));
if (ctx->SA == NULL)
- goto err_free_block_specs;
- ctx->ISA = ctx->SA + LZX_MAX_WINDOW_SIZE;
- ctx->LCP = ctx->ISA + LZX_MAX_WINDOW_SIZE;
- ctx->salink = MALLOC(LZX_MAX_WINDOW_SIZE * sizeof(ctx->salink[0]));
+ goto err;
+ ctx->ISA = ctx->SA + window_size;
+ ctx->LCP = ctx->ISA + window_size;
+
+ ctx->salink = MALLOC(window_size * sizeof(ctx->salink[0]));
if (ctx->salink == NULL)
- goto err_free_SA;
- } else {
- ctx->SA = NULL;
- ctx->ISA = NULL;
- ctx->LCP = NULL;
- ctx->salink = NULL;
+ goto err;
}
if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
ctx->optimum = MALLOC((LZX_OPTIM_ARRAY_SIZE + LZX_MAX_MATCH_LEN) *
sizeof(ctx->optimum[0]));
if (ctx->optimum == NULL)
- goto err_free_salink;
- } else {
- ctx->optimum = NULL;
+ goto err;
}
if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
- uint32_t cache_per_pos;
+ u32 cache_per_pos;
cache_per_pos = params->alg_params.slow.max_matches_per_pos;
if (cache_per_pos > LZX_MAX_CACHE_PER_POS)
cache_per_pos = LZX_MAX_CACHE_PER_POS;
- ctx->cached_matches = MALLOC(LZX_MAX_WINDOW_SIZE * (cache_per_pos + 1) *
+ ctx->cached_matches = MALLOC(window_size * (cache_per_pos + 1) *
sizeof(ctx->cached_matches[0]));
if (ctx->cached_matches == NULL)
- goto err_free_optimum;
- } else {
- ctx->cached_matches = NULL;
+ goto err;
}
- ctx->chosen_matches = MALLOC(LZX_MAX_WINDOW_SIZE *
- sizeof(ctx->chosen_matches[0]));
+ ctx->chosen_matches = MALLOC(window_size * sizeof(ctx->chosen_matches[0]));
if (ctx->chosen_matches == NULL)
- goto err_free_cached_matches;
+ goto err;
memcpy(&ctx->params, params, sizeof(struct wimlib_lzx_params));
memset(&ctx->zero_codes, 0, sizeof(ctx->zero_codes));
*ctx_pp = (struct wimlib_lzx_context*)ctx;
return 0;
-err_free_cached_matches:
- FREE(ctx->cached_matches);
-err_free_optimum:
- FREE(ctx->optimum);
-err_free_salink:
- FREE(ctx->salink);
-err_free_SA:
- FREE(ctx->SA);
-err_free_block_specs:
- FREE(ctx->block_specs);
-err_free_ctx:
- FREE(ctx);
err:
+ wimlib_lzx_free_context((struct wimlib_lzx_context*)ctx);
LZX_DEBUG("Ran out of memory.");
return WIMLIB_ERR_NOMEM;
}
struct lzx_compressor *ctx = (struct lzx_compressor*)_ctx;
if (ctx) {
- FREE(ctx->cached_matches);
FREE(ctx->chosen_matches);
+ FREE(ctx->cached_matches);
FREE(ctx->optimum);
- FREE(ctx->SA);
FREE(ctx->salink);
+ FREE(ctx->SA);
FREE(ctx->block_specs);
+ FREE(ctx->prev_tab);
+ FREE(ctx->window);
FREE(ctx);
}
}
struct wimlib_lzx_context *ctx = NULL;
unsigned compressed_len;
- ret = wimlib_lzx_alloc_context(NULL, &ctx);
+ ret = wimlib_lzx_alloc_context(32768, NULL, &ctx);
if (ret) {
wimlib_assert(ret != WIMLIB_ERR_INVALID_PARAM);
WARNING("Couldn't allocate LZX compression context: %"TS"",
git://wimlib.net / wimlib / blobdiff