Crates.io | bintext |
lib.rs | bintext |
version | 0.1.3 |
source | src |
created_at | 2020-08-23 14:42:52.930171 |
updated_at | 2020-08-23 15:22:06.38501 |
description | Encode and decodes binary encoded text into aligned binary blobs using SIMD |
homepage | |
repository | https://github.com/lassade/bintext |
max_upload_size | |
id | 279835 |
size | 39,647 |
Encode and decodes binary encoded text into aligned binary blobs using SIMD
Binary text encoding and decoding with support for SIMD (AVX2 and SSSE3) with good fallback performance.
The main idea of this crate is to have a zero copy binary deserialization for text formats.
Alignment can't be guaranteed in a text format, so no matter what the data will need
to be re-aligned while decoding, if the required alignment is N
maximum amount
of offset need to move the bytes is less than N
thus by providing an start padding
in the binary encoded text of N - 1
it's possible to align the data up to N
.
Quick note this crate will only accept padding equal or grater than N
, because
it's a bit cheap to do this way.
// Padding of 8 (suppose it was read form a file)
let hex = "--------a1f7d5e8d14f0f76".to_string();
unsafe {
// Decode with padding of 8 and alignment of 8
let slice = bintext::hex::decode_aligned(&mut hex, 8, 8).unwrap();
// Data is aligned so you can safely do this:
let slice: &[u64] = std::slice::from_raw_parts(
slice.as_ptr() as *const _,
slice.len() / std::mem::size_of::<u64>()
);
}
There is a lot of other crates that already does the job, but every single one have some kind of performance downside:
hex
bad performance, but smaller code size since doesn't rely on LUTs;base16
, no SIMD, but have faster decoding and overall good encoding;faster-hex
fastest encode using SSSE3 and AVX2, decode only has an AVX2.
SSE has the best SIMD coverage of all instructions sets, so it should be a must.This crate provides implementations that covers all their competitors
performance weaknesses. Just run cargo bench
.
All these crates doesn't seem to provide functions to decode aligned data, so after decoding you will need and extra alignment step.