cuberite/libdeflate
1
0
Fork 0
mirror of https://github.com/cuberite/libdeflate.git synced 2025-09-13 06:15:51 -04:00
Code Issues Packages Projects Releases Wiki Activity

lib/deflate_decompress: build subtables separately

Browse Source 
Further improve build_decode_table() performance by splitting the "fill
direct entries" and "fill subtable pointers and subtables" steps into
separate loops and making some other optimizations.
This commit is contained in:
Eric Biggers 2018-12-25 23:57:43 -06:00
parent 515b7ad15c
commit bfc3f610e1
1 changed files with 93 additions and 69 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

162
lib/deflate_decompress.c
View File

@ -515,6 +515,39 @@ static const u32 offset_decode_results[DEFLATE_NUM_OFFSET_SYMS] = {
#undef ENTRY #undef ENTRY
}; };
/* Advance to the next codeword in the canonical Huffman code */
static forceinline void
next_codeword(unsigned *codeword_p, unsigned *codeword_len_p,
unsigned *stride_p, unsigned len_counts[])
{
unsigned codeword = *codeword_p;
unsigned codeword_len = *codeword_len_p;
unsigned bit;
/*
* Increment the codeword, bit-reversed: find the last (highest order) 0
* bit in the codeword, set it, and clear any later (higher order) bits.
*/
bit = 1U << bsr32(~codeword & ((1U << codeword_len) - 1));
codeword &= bit - 1;
codeword |= bit;
/*
* If there are no more codewords of this length, proceed to the next
* lowest used length. Increasing the length logically appends 0's to
* the codeword, but this is a no-op due to the codeword being
* represented in bit-reversed form.
*/
len_counts[codeword_len]--;
while (len_counts[codeword_len] == 0) {
codeword_len++;
*stride_p <<= 1;
}
*codeword_p = codeword;
*codeword_len_p = codeword_len;
}
/* /*
* Build a table for fast decoding of symbols from a Huffman code. As input, * Build a table for fast decoding of symbols from a Huffman code. As input,
* this function takes the codeword length of each symbol which may be used in * this function takes the codeword length of each symbol which may be used in
@ -562,19 +595,20 @@ build_decode_table(u32 decode_table[],
const unsigned max_codeword_len, const unsigned max_codeword_len,
u16 sorted_syms[]) u16 sorted_syms[])
{ {
const unsigned table_mask = (1U << table_bits) - 1;
unsigned len_counts[DEFLATE_MAX_CODEWORD_LEN + 1]; unsigned len_counts[DEFLATE_MAX_CODEWORD_LEN + 1];
unsigned offsets[DEFLATE_MAX_CODEWORD_LEN + 1]; unsigned offsets[DEFLATE_MAX_CODEWORD_LEN + 1];
unsigned len; unsigned len;
unsigned sym; unsigned sym;
s32 remainder; s32 remainder;
unsigned sym_idx; unsigned sym_idx;
unsigned codeword;
unsigned codeword_len; unsigned codeword_len;
unsigned codeword_reversed = 0; unsigned stride;
unsigned cur_codeword_prefix = -1; unsigned cur_table_end = 1U << table_bits;
unsigned cur_table_start = 0; unsigned subtable_prefix;
unsigned cur_table_bits = table_bits; unsigned subtable_start;
unsigned num_dropped_bits = 0; unsigned subtable_bits;
const unsigned table_mask = (1U << table_bits) - 1;
/* Count how many symbols have each codeword length, including 0. */ /* Count how many symbols have each codeword length, including 0. */
for (len = 0; len <= max_codeword_len; len++) for (len = 0; len <= max_codeword_len; len++)
@ -647,32 +681,53 @@ build_decode_table(u32 decode_table[],
/* Start with the smallest codeword length and the smallest-valued /* Start with the smallest codeword length and the smallest-valued
* symbol which has that codeword length. */ * symbol which has that codeword length. */
sym_idx = offsets[0]; sym_idx = offsets[0];
sym = sorted_syms[sym_idx++];
codeword = 0;
codeword_len = 1; codeword_len = 1;
while (len_counts[codeword_len] == 0) while (len_counts[codeword_len] == 0)
codeword_len++; codeword_len++;
stride = 1U << codeword_len;
for (;;) { /* For each used symbol and its codeword... */ /* For each symbol and its codeword in the main part of the table... */
unsigned sym; do {
u32 entry; u32 entry;
unsigned i; unsigned i;
unsigned end;
unsigned increment;
unsigned bit;
/* Get the next symbol. */ /* Fill in as many copies of the decode table entry as are
sym = sorted_syms[sym_idx]; * needed. The number of entries to fill is a power of 2 and
* depends on the codeword length; it could be as few as 1 or as
* large as half the size of the table. Since the codewords are
* bit-reversed, the indices to fill are those with the codeword
* in its low bits; it's the high bits that vary. */
entry = decode_results[sym] | codeword_len;
i = codeword;
do {
decode_table[i] = entry;
i += stride; /* stride is 1U << codeword_len */
} while (i < cur_table_end);
/* Start a new subtable if the codeword is long enough to /* Advance to the next symbol and codeword */
* require a subtable, *and* the first 'table_bits' bits of the if (sym_idx == num_syms)
* codeword don't match the prefix for the previous subtable if return true;
* any. */ sym = sorted_syms[sym_idx++];
if (codeword_len > table_bits && next_codeword(&codeword, &codeword_len, &stride, len_counts);
(codeword_reversed & table_mask) != cur_codeword_prefix) { } while (codeword_len <= table_bits);
cur_codeword_prefix = (codeword_reversed & table_mask); /* The codeword length has exceeded table_bits, so we're done filling
* direct entries. Start filling subtable pointers and subtables. */
stride >>= table_bits;
goto new_subtable;
cur_table_start += 1U << cur_table_bits; for (;;) {
u32 entry;
unsigned i;
/* Start a new subtable if the first 'table_bits' bits of the
* codeword don't match the prefix for the current subtable. */
if ((codeword & table_mask) != subtable_prefix) {
new_subtable:
subtable_prefix = (codeword & table_mask);
subtable_start = cur_table_end;
/* Calculate the subtable length. If the codeword /* Calculate the subtable length. If the codeword
* length exceeds 'table_bits' by n, the subtable needs * length exceeds 'table_bits' by n, the subtable needs
* at least 2**n entries. But it may need more; if * at least 2**n entries. But it may need more; if
@ -684,74 +739,43 @@ build_decode_table(u32 decode_table[],
* subtable, since the only case where we may have an * subtable, since the only case where we may have an
* incomplete code is a single codeword of length 1, * incomplete code is a single codeword of length 1,
* and that never requires any subtables. */ * and that never requires any subtables. */
cur_table_bits = codeword_len - table_bits; subtable_bits = codeword_len - table_bits;
remainder = (s32)1 << cur_table_bits; remainder = 1U << subtable_bits;
for (;;) { for (;;) {
remainder -= len_counts[table_bits + remainder -= len_counts[table_bits +
cur_table_bits]; subtable_bits];
if (remainder <= 0) if (remainder <= 0)
break; break;
cur_table_bits++; subtable_bits++;
remainder <<= 1; remainder <<= 1;
} }
cur_table_end = subtable_start + (1U << subtable_bits);
/* Create the entry that points from the main table to /* Create the entry that points from the main table to
* the subtable. This entry contains the index of the * the subtable. This entry contains the index of the
* start of the subtable and the number of bits with * start of the subtable and the number of bits with
* which the subtable is indexed (the log base 2 of the * which the subtable is indexed (the log base 2 of the
* number of entries it contains). */ * number of entries it contains). */
decode_table[cur_codeword_prefix] = decode_table[subtable_prefix] =
HUFFDEC_SUBTABLE_POINTER | HUFFDEC_SUBTABLE_POINTER |
HUFFDEC_RESULT_ENTRY(cur_table_start) | HUFFDEC_RESULT_ENTRY(subtable_start) |
cur_table_bits; subtable_bits;
/* Now that we're filling a subtable, we need to drop
* the first 'table_bits' bits of the codewords. */
num_dropped_bits = table_bits;
} }
/* Create the decode table entry, which packs the decode result /* Fill the subtable entries */
* and the codeword length (minus 'table_bits' for subtables) entry = decode_results[sym] | (codeword_len - table_bits);
* together. */ i = subtable_start + (codeword >> table_bits);
entry = decode_results[sym] | (codeword_len - num_dropped_bits);
/* Fill in as many copies of the decode table entry as are
* needed. The number of entries to fill is a power of 2 and
* depends on the codeword length; it could be as few as 1 or as
* large as half the size of the table. Since the codewords are
* bit-reversed, the indices to fill are those with the codeword
* in its low bits; it's the high bits that vary. */
i = cur_table_start + (codeword_reversed >> num_dropped_bits);
end = cur_table_start + (1U << cur_table_bits);
increment = 1U << (codeword_len - num_dropped_bits);
do { do {
decode_table[i] = entry; decode_table[i] = entry;
i += increment; /* stride is 1U << (codeword_len - table_bits) */
} while (i < end); i += stride;
} while (i < cur_table_end);
/* Advance to the next symbol and codeword */ /* Advance to the next symbol and codeword */
if (sym_idx == num_syms)
if (++sym_idx == num_syms)
return true; return true;
/* sym = sorted_syms[sym_idx++];
* Increment the codeword, bit-reversed: find the last (highest next_codeword(&codeword, &codeword_len, &stride, len_counts);
* order) 0 bit in the codeword, set it, and clear any later
* (higher order) bits.
*/
bit = 1U << bsr32(~codeword_reversed &
((1U << codeword_len) - 1));
codeword_reversed &= bit - 1;
codeword_reversed |= bit;
/*
* If there are no more codewords of this length, proceed to the
* next lowest used length. Increasing the length logically
* appends 0's to the codeword, but this is a no-op due to the
* codeword being represented in bit-reversed form.
*/
len_counts[codeword_len]--;
while (len_counts[codeword_len] == 0)
codeword_len++;
} }
} }