265f3afc2a
matchfinder: penalize score for overlapping matches
2024-01-09 06:03:56 -08:00
a8d524a96d
matchfinder: replace Score function with DistanceBitCost
2024-01-09 05:40:40 -08:00
578645e154
matchfinder: add MultiHash
2024-01-09 05:06:39 -08:00
24b2bfad2d
matchfinder.M4: add Score function
2024-01-02 13:38:12 -08:00
4a024e3eff
matchfinder.M4: add match chain
2024-01-01 16:13:22 -08:00
3a1c5cd370
Fix typo in comment.
2024-01-01 15:06:21 -08:00
0d2aef37af
matchfinder.M4: factor out extendMatch2
2023-12-30 16:25:51 -08:00
63f3f4372d
matchfinder.M4: add LimitedSearch option
...
Using LimitedSearch, it only checks for overlapping matches in one
place instead of checking at each byte.
This gains about 50% in compression speed while only losing about
2% in compression ratio.
2023-12-30 15:56:13 -08:00
924a0eb0c6
matchfinder.M4: more refactoring
...
Factor out matchEmitter.trim, and make TableBits configurable.
2023-12-28 17:21:34 -08:00
c506503c67
matchfinder: factor out matchEmitter
2023-12-28 17:01:08 -08:00
349ed2fce1
Add matchfinder package.
...
I've been experimenting for a while with a new brotli compressor.
Instead of being a translation of the C implementation,
it's a rewrite in Go, with a modular structure thanks to interfaces.
(A few low-level functions still come from the C version, though.)
The performance is getting to the point where it seems to be worth
adding to the brotli repository.
2023-12-28 16:09:32 -08:00
b7a4cf9ec5
remove Content-Type requirement
...
* https://github.com/golang/go/issues/31753 is fixed
2023-02-25 11:50:08 -05:00
2848168f55
Reader.Reset: recover from errors.
...
When a Reader encounters an error, its internal state may be corrupted.
So Reset needs to be extra thorough, to avoid cascading errors.
Based on PR 37 by Sovianum.
2022-09-23 19:39:05 -07:00
786ec621f6
Reuse ringbuffer in Reader.
...
Fixes #33
2022-05-18 12:06:45 -07:00
f001d275a3
Fix Reader.Reset() docs
2022-05-12 18:31:16 +01:00
34a5640cc1
Add an example for Writer.Reset.
...
It is based on the one for flate.Writer.Reset.
Fixes #31 .
2022-05-03 10:55:39 -07:00
1d750214c2
Optimize log2FloorNonZero with math/bits.
2021-09-22 11:21:01 -07:00
ec682fbe0a
Rest set writer.err=nil
2021-08-25 10:29:47 +08:00
cf8bc3b664
More staticcheck advice.
2021-08-18 18:04:51 -07:00
a61eb82231
Follow some advice from staticcheck.
2021-08-12 12:37:09 -07:00
e073f0d4ed
Merge branch 'issue22'
2021-07-15 12:01:52 -07:00
177b8acd6c
Add test for issue 22.
2021-07-15 12:01:31 -07:00
5376c15dde
Retract v1.0.1
2021-05-26 15:10:20 -07:00
94609f9606
Revert "Faster bit writing."
...
This reverts commit c3da72aa01Fixes #22
2021-04-27 10:48:28 -07:00
47c0dbab12
Simplify control flow in Reader.Reset.
2021-03-01 09:45:40 -08:00
87d8f4575c
Reduce allocations of buffer for reused readers
2021-03-01 09:42:42 -08:00
729edfbcfe
Add documentation link.
...
Fixes #16
2020-08-04 09:53:43 -07:00
1b06c5640c
Check size of split.lengths instead of a nil check
...
Fixes https://github.com/andybalholm/brotli/issues/14
2020-07-17 09:24:47 -07:00
c3da72aa01
Faster bit writing.
...
Replace the functions in write_bits.go with a bitWriter type based on
the compress/flate package.
2020-06-18 18:58:27 -07:00
ef7a42160d
Use a 64-bit store in writeBits.
...
This is an optimization that was present in the C version (behind an
ifdef). It gives a nice speed boost to the lower compression levels.
2020-06-06 14:13:21 -07:00
097c1c5bc9
Use 32-bit loads in isMatch1 and isMatch5.
2020-05-15 11:17:23 -07:00
8f8b18645c
Read multiple bytes in findMatchLengthWithLimit
...
Use 64- or 32-bit loads instead of reading a byte at a time. The
original C source did something like this, in a very C-ish way. It
needed to be simplified to translate it to Go. The exact way this works
was suggested by the assembly code in github.com/golang/snappy.
2020-05-15 10:43:19 -07:00
a01a7b12c9
Reuse buffers and objects using sync.Pool
...
This reduces the amount of garbage generated and relieves pressure on
the GC.
For a workload without reusing the Writer (using Writer.Reset) the number of allocations goes from 31 to 9.
While for a workload when you reuse the Writer the number of allocations goes from 25 to 0.
2020-05-10 10:36:19 +02:00
e2c5f2109f
Use len and cap instead of num_commands_ and cmd_alloc_size_.
2020-05-08 16:48:16 -07:00
4b2775ea5e
Fix some ugly compound literals.
2020-05-08 15:14:47 -07:00
b2497e8d72
Revert "Use sort.Sort to sort Huffman trees."
...
This reverts commit 6b5963335e
2020-05-08 13:52:56 -07:00
6b5963335e
Use sort.Sort to sort Huffman trees.
2020-05-07 17:51:21 -07:00
7c7a5a10ef
Push output directly to dst instead of buffering.
2020-05-07 17:27:37 -07:00
625cbb6f92
Replace storage_size_ with len(storage).
2020-05-07 15:40:58 -07:00
f41712f811
Remove unnecessary parameters from encodeData.
...
The pointers passed to out_size and output were always the same,
so there is no need to have them as parameters.
Based on
00ca10b927b3ee528567
2020-05-06 17:20:27 -07:00
511ca97d30
Benchmark all compression levels.
2020-05-06 16:08:01 -07:00
2c14228f02
Preserve w.commands across Reset.
2020-05-05 17:36:16 -07:00
cb9be97eb7
Reuse more memory when a Writer is Reset.
2020-05-05 17:18:33 -07:00
00ca370ce2
Put full license in bench_test.go
2020-05-05 16:14:45 -07:00
3c3658f2fb
A couple of tweaks
2020-05-04 16:09:27 -07:00
ac459dee64
Fix excessive allocations if sortHuffmanTreeItems.
...
Fixes #11
2020-05-04 15:04:20 -07:00
9fc263f565
Add some benchmarks.
...
These are basically copied from the compress/flate package.
2020-05-04 13:35:04 -07:00
b60f0d972e
Copy negotiateContentEncoding.
...
A little copying is better than a little dependency.
Fixes #8
2019-08-21 08:13:43 -07:00
ed0fd64940
Fix int overflow in test on 32-bit.
...
On 32-bit systems, the random array indexes were sometimes negative as a
result of converting an int64 to int.
Fixes #7 .
2019-07-25 08:52:43 -07:00
71eb68cc46
Preserve compression settings when Writer is reset.
...
Fixes #6
2019-07-04 08:13:24 -07:00
Andy Balholm