blob: dad6eba76c8e5fa98ae1aac6323443d7b87d1847 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
|
# Optimizations
This document tracks which optimizations have been done after the initial implementation passed corpus tests and a good amount of fuzzing.
## Introducing more unsafe code:
These optimizations introduced more unsafe code. These should yield significant improvements, or else they are not really worth it.
### Optimizing bitreader with byteorder which uses ptr::copy_nonoverlapping
* Reverse bitreader_reversed::get_bits was identified by linux perf tool using about 36% of the whole time
* Benchmark: decode enwik9
* Before: about 14.7 seconds
* After: about 12.2 seconds with about 25% of the time used for get_bits()
### Optimizing decodebuffer::repeat with ptr::copy_nonoverlapping
* decodebuffer::repeate was identified by linux perf tool using about 28% of the whole time
* Benchmark: decode enwik9
* Before: about 9.9 seconds
* After: about 9.4 seconds
### Use custom ringbuffer in the decodebuffer
The decode buffer must be able to do two things efficiently
* Collect bytes from the front
* Copy bytes from the contents to the end
The stdlibs VecDequeu and Vec can each do one but not the other efficiently. So a custom implementation of a ringbuffer was written.
## Introducing NO additional unsafe code
These are just nice to have
### Even better bitreaders
Studying this material lead to a big improvement in bitreader speed
* https://fgiesen.wordpress.com/2018/02/19/reading-bits-in-far-too-many-ways-part-1/
* https://fgiesen.wordpress.com/2018/02/20/reading-bits-in-far-too-many-ways-part-2/
|
