cuberite/libdeflate
1
0
Fork 0
mirror of https://github.com/cuberite/libdeflate.git synced 2025-08-03 09:46:04 -04:00
Code Issues Packages Projects Releases Wiki Activity

deflate_compress: allow shorter blocks

Browse Source 
The MIN_BLOCK_LENGTH of 10000 is too high; shorter blocks are sometimes
worthwhile.  Reduce MIN_BLOCK_LENGTH to 5000, but penalize very short
blocks so they are only chosen when they clearly seem worthwhile.
This commit is contained in:
Eric Biggers 2022-01-04 21:34:22 -08:00
parent 968588d19a
commit 5f4da4b243
1 changed files with 17 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
lib/deflate_compress.c
View File

@ -61,7 +61,7 @@
* reasonable upper bound on the compressed size. It's also needed because our
* block splitting algorithm doesn't work well on very short blocks.
*/
#define MIN_BLOCK_LENGTH 10000
#define MIN_BLOCK_LENGTH 5000
/*
* For the greedy, lazy, lazy2, and near-optimal compressors: This is the soft
@ -2127,6 +2127,8 @@ do_end_block_check(struct block_split_stats *stats, u32 block_length)
* multiplied by 'num_observations'.
*/
u32 total_delta = 0;
u32 num_items;
u32 cutoff;
int i;
for (i = 0; i < NUM_OBSERVATION_TYPES; i++) {
@ -2140,11 +2142,22 @@ do_end_block_check(struct block_split_stats *stats, u32 block_length)
total_delta += delta;
}
num_items = stats->num_observations +
stats->num_new_observations;
cutoff = stats->num_new_observations * 200 / 512 *
stats->num_observations;
/*
* Very short blocks have a lot of overhead for the Huffman
* codes, so only use them if it clearly seems worthwhile.
* (This is an additional penalty, which adds to the smaller
* penalty below which scales more slowly.)
*/
if (block_length < 10000 && num_items < 8192)
cutoff += (u64)cutoff * (8192 - num_items) / 8192;
/* Ready to end the block? */
if (total_delta +
(block_length / 4096) * stats->num_observations >=
stats->num_new_observations * 200 / 512 *
stats->num_observations)
(block_length / 4096) * stats->num_observations >= cutoff)
return true;
}
merge_new_observations(stats);