X-Git-Url: https://wimlib.net/git/?p=wimlib;a=blobdiff_plain;f=src%2Flzx_compress.c;h=b6d603d620a84ecba60e9d900f8908d2d7812a19;hp=de8430145f4e7a2867ad60c59b5d17a6b5a4436e;hb=6fbc434a762536773edda6b147795c4b1805a82b;hpb=26396c2c45946ba38c18c6ac0207e8c1f68e4668 diff --git a/src/lzx_compress.c b/src/lzx_compress.c index de843014..b6d603d6 100644 --- a/src/lzx_compress.c +++ b/src/lzx_compress.c @@ -114,10 +114,9 @@ #define LZX_BIT_COST 16 /* - * Consideration of aligned offset costs is disabled for now, due to - * insufficient benefit gained from the time spent. + * Should the compressor take into account the costs of aligned offset symbols? */ -#define LZX_CONSIDER_ALIGNED_COSTS 0 +#define LZX_CONSIDER_ALIGNED_COSTS 1 /* * LZX_MAX_FAST_LEVEL is the maximum compression level at which we use the @@ -153,16 +152,16 @@ #include "wimlib/util.h" /* Matchfinders with 16-bit positions */ -#define pos_t u16 -#define MF_SUFFIX _16 +#define mf_pos_t u16 +#define MF_SUFFIX _16 #include "wimlib/bt_matchfinder.h" #include "wimlib/hc_matchfinder.h" /* Matchfinders with 32-bit positions */ -#undef pos_t +#undef mf_pos_t #undef MF_SUFFIX -#define pos_t u32 -#define MF_SUFFIX _32 +#define mf_pos_t u32 +#define MF_SUFFIX _32 #include "wimlib/bt_matchfinder.h" #include "wimlib/hc_matchfinder.h" @@ -411,8 +410,10 @@ struct lzx_compressor { /* The matches and literals that the parser has chosen for the current * block. The required length of this array is limited by the maximum - * number of matches that can ever be chosen for a single block. */ - struct lzx_sequence chosen_sequences[DIV_ROUND_UP(LZX_DIV_BLOCK_SIZE, LZX_MIN_MATCH_LEN)]; + * number of matches that can ever be chosen for a single block, plus + * one for the special entry at the end. */ + struct lzx_sequence chosen_sequences[ + DIV_ROUND_UP(LZX_DIV_BLOCK_SIZE, LZX_MIN_MATCH_LEN) + 1]; /* Tables for mapping adjusted offsets to offset slots */ @@ -549,7 +550,7 @@ struct lzx_output_bitstream { /* Can the specified number of bits always be added to 'bitbuf' after any * pending 16-bit coding units have been flushed? */ -#define CAN_BUFFER(n) ((n) <= (8 * sizeof(machine_word_t)) - 16) +#define CAN_BUFFER(n) ((n) <= (8 * sizeof(machine_word_t)) - 15) /* * Initialize the output bitstream. @@ -635,7 +636,7 @@ lzx_make_huffman_codes(struct lzx_compressor *c) STATIC_ASSERT(MAIN_CODEWORD_LIMIT >= 9 && MAIN_CODEWORD_LIMIT <= LZX_MAX_MAIN_CODEWORD_LEN); - STATIC_ASSERT(LENGTH_CODEWORD_LIMIT >= 9 && + STATIC_ASSERT(LENGTH_CODEWORD_LIMIT >= 8 && LENGTH_CODEWORD_LIMIT <= LZX_MAX_LEN_CODEWORD_LEN); STATIC_ASSERT(ALIGNED_CODEWORD_LIMIT >= LZX_NUM_ALIGNED_OFFSET_BITS && ALIGNED_CODEWORD_LIMIT <= LZX_MAX_ALIGNED_CODEWORD_LEN); @@ -995,6 +996,8 @@ lzx_write_sequences(struct lzx_output_bitstream *os, int block_type, if (!CAN_BUFFER(MAX_MATCH_BITS)) lzx_flush_bits(os, ALIGNED_CODEWORD_LIMIT); } else { + STATIC_ASSERT(CAN_BUFFER(17)); + lzx_add_bits(os, extra_bits, num_extra_bits); if (!CAN_BUFFER(MAX_MATCH_BITS)) lzx_flush_bits(os, 17); @@ -1557,16 +1560,18 @@ lzx_find_min_cost_path(struct lzx_compressor * const restrict c, u32 offset_data = offset + LZX_OFFSET_ADJUSTMENT; unsigned offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit); + u32 base_cost = cur_node->cost; + + #if LZX_CONSIDER_ALIGNED_COSTS + if (offset_data >= 16) + base_cost += c->costs.aligned[offset_data & + LZX_ALIGNED_OFFSET_BITMASK]; + #endif + do { - u32 cost = cur_node->cost + + u32 cost = base_cost + c->costs.match_cost[offset_slot][ next_len - LZX_MIN_MATCH_LEN]; - #if LZX_CONSIDER_ALIGNED_COSTS - if (lzx_extra_offset_bits[offset_slot] >= - LZX_NUM_ALIGNED_OFFSET_BITS) - cost += c->costs.aligned[offset_data & - LZX_ALIGNED_OFFSET_BITMASK]; - #endif if (cost < (cur_node + next_len)->cost) { (cur_node + next_len)->cost = cost; (cur_node + next_len)->item = @@ -1584,8 +1589,7 @@ lzx_find_min_cost_path(struct lzx_compressor * const restrict c, * of coding the literal is integrated into the queue update * code below. */ literal = *in_next++; - cost = cur_node->cost + - c->costs.main[lzx_main_symbol_for_literal(literal)]; + cost = cur_node->cost + c->costs.main[literal]; /* Advance to the next position. */ cur_node++; @@ -1625,17 +1629,17 @@ lzx_find_min_cost_path(struct lzx_compressor * const restrict c, static void lzx_compute_match_costs(struct lzx_compressor *c) { - unsigned num_offset_slots = lzx_get_num_offset_slots(c->window_order); + unsigned num_offset_slots = (c->num_main_syms - LZX_NUM_CHARS) / LZX_NUM_LEN_HEADERS; struct lzx_costs *costs = &c->costs; for (unsigned offset_slot = 0; offset_slot < num_offset_slots; offset_slot++) { u32 extra_cost = (u32)lzx_extra_offset_bits[offset_slot] * LZX_BIT_COST; - unsigned main_symbol = lzx_main_symbol_for_match(offset_slot, 0); + unsigned main_symbol = LZX_NUM_CHARS + (offset_slot * LZX_NUM_LEN_HEADERS); unsigned i; #if LZX_CONSIDER_ALIGNED_COSTS - if (lzx_extra_offset_bits[offset_slot] >= LZX_NUM_ALIGNED_OFFSET_BITS) + if (offset_slot >= 8) extra_cost -= LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST; #endif @@ -1709,15 +1713,21 @@ lzx_update_costs(struct lzx_compressor *c) unsigned i; const struct lzx_lens *lens = &c->codes[c->codes_index].lens; - for (i = 0; i < c->num_main_syms; i++) - c->costs.main[i] = (lens->main[i] ? lens->main[i] : 15) * LZX_BIT_COST; + for (i = 0; i < c->num_main_syms; i++) { + c->costs.main[i] = (lens->main[i] ? lens->main[i] : + MAIN_CODEWORD_LIMIT) * LZX_BIT_COST; + } - for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) - c->costs.len[i] = (lens->len[i] ? lens->len[i] : 15) * LZX_BIT_COST; + for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) { + c->costs.len[i] = (lens->len[i] ? lens->len[i] : + LENGTH_CODEWORD_LIMIT) * LZX_BIT_COST; + } #if LZX_CONSIDER_ALIGNED_COSTS - for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) - c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] : 7) * LZX_BIT_COST; + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) { + c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] : + ALIGNED_CODEWORD_LIMIT) * LZX_BIT_COST; + } #endif lzx_compute_match_costs(c); @@ -1778,9 +1788,9 @@ lzx_compress_near_optimal(struct lzx_compressor *c, const u8 * const in_begin = c->in_buffer; const u8 * in_next = in_begin; const u8 * const in_end = in_begin + c->in_nbytes; - unsigned max_len = LZX_MAX_MATCH_LEN; - unsigned nice_len = min(c->nice_match_length, max_len); - u32 next_hash = 0; + u32 max_len = LZX_MAX_MATCH_LEN; + u32 nice_len = min(c->nice_match_length, max_len); + u32 next_hashes[2] = {}; struct lzx_lru_queue queue; CALL_BT_MF(is_16_bit, c, bt_matchfinder_init); @@ -1796,20 +1806,14 @@ lzx_compress_near_optimal(struct lzx_compressor *c, struct lz_match *cache_ptr = c->match_cache; do { struct lz_match *lz_matchptr; - unsigned best_len; + u32 best_len; /* If approaching the end of the input buffer, adjust * 'max_len' and 'nice_len' accordingly. */ if (unlikely(max_len > in_end - in_next)) { max_len = in_end - in_next; nice_len = min(max_len, nice_len); - - /* This extra check is needed to ensure that we - * never output a length 2 match of the very - * last two bytes with the very first two bytes, - * since such a match has an offset too large to - * be represented. */ - if (unlikely(max_len < 3)) { + if (unlikely(max_len < BT_MATCHFINDER_REQUIRED_NBYTES)) { in_next++; cache_ptr->length = 0; cache_ptr++; @@ -1824,7 +1828,7 @@ lzx_compress_near_optimal(struct lzx_compressor *c, max_len, nice_len, c->max_search_depth, - &next_hash, + next_hashes, &best_len, cache_ptr + 1); in_next++; @@ -1849,7 +1853,7 @@ lzx_compress_near_optimal(struct lzx_compressor *c, if (unlikely(max_len > in_end - in_next)) { max_len = in_end - in_next; nice_len = min(max_len, nice_len); - if (unlikely(max_len < 3)) { + if (unlikely(max_len < BT_MATCHFINDER_REQUIRED_NBYTES)) { in_next++; cache_ptr->length = 0; cache_ptr++; @@ -1862,7 +1866,7 @@ lzx_compress_near_optimal(struct lzx_compressor *c, max_len, nice_len, c->max_search_depth, - &next_hash); + next_hashes); in_next++; cache_ptr->length = 0; cache_ptr++; @@ -2263,8 +2267,8 @@ lzx_create_compressor(size_t max_bufsize, unsigned compression_level, c->impl = lzx_compress_lazy_16; else c->impl = lzx_compress_lazy_32; - c->max_search_depth = (36 * compression_level) / 20; - c->nice_match_length = (72 * compression_level) / 20; + c->max_search_depth = (60 * compression_level) / 20; + c->nice_match_length = (80 * compression_level) / 20; /* lzx_compress_lazy() needs max_search_depth >= 2 because it * halves the max_search_depth when attempting a lazy match, and @@ -2283,7 +2287,7 @@ lzx_create_compressor(size_t max_bufsize, unsigned compression_level, /* Scale nice_match_length and max_search_depth with the * compression level. */ c->max_search_depth = (24 * compression_level) / 50; - c->nice_match_length = (32 * compression_level) / 50; + c->nice_match_length = (48 * compression_level) / 50; /* Set a number of optimization passes appropriate for the * compression level. */ @@ -2344,7 +2348,7 @@ lzx_compress(const void *restrict in, size_t in_nbytes, else memcpy(c->in_buffer, in, in_nbytes); c->in_nbytes = in_nbytes; - lzx_do_e8_preprocessing(c->in_buffer, in_nbytes); + lzx_preprocess(c->in_buffer, in_nbytes); /* Initially, the previous Huffman codeword lengths are all zeroes. */ c->codes_index = 0; @@ -2359,7 +2363,7 @@ lzx_compress(const void *restrict in, size_t in_nbytes, /* Flush the output bitstream and return the compressed size or 0. */ result = lzx_flush_output(&os); if (!result && c->destructive) - lzx_undo_e8_preprocessing(c->in_buffer, c->in_nbytes); + lzx_postprocess(c->in_buffer, c->in_nbytes); return result; }