🚀 Base 64

✨ SIMD-accelerated Base64 for Rust ✨

🌟 Features

• 💡 Uses AVX2 instructions for super-fast encoding and decoding
• 🔄 Fallback when AVX2 is unavailable uses any available SIMD

🎯 Project goals

• 🔧 Simple, idiomatic API
• 📦 Sensible defaults
• ⚡ Fast

Installation

cargo add bs64

Usage

use bs64;

fn main() {
  // Encode
  let input = vec![2, 3, 4, 5];
  let output: String = bs64::encode(&input);

  // Decode
  let decoded_output = bs64::decode(output.as_bytes());
}

Benchmarks

Ran using 100k inputs, 10000 iterations on an Intel® Core™ i7-1065G7. Comparisons are made against base64 and data-encoding crates.

cargo run --features "cli" --release -- -b 100000 -i 10000

Encode

name	MB/s
🚀 bs64::encode()	4813.70
🚀 bs64::encode_mut()	6579.17
🚀 bs64 fallback	944.18
data_encoding	858.51
data_encoding mut	873.28
base64	748.02
base64 mut	870.99

Decode

name	MB/s
🚀 bs64::decode()	3899.26
🚀 bs64::decode_mut()	3965.25
🚀 bs64 fallback	837.17
data_encoding	647.33
data_encoding mut	684.01
base64	761.68
base64 mut	805.60

Implementation Details

Code was initially ported from https://github.com/lemire/fastbase64

The simple fallback implementation is based on the chromium implementation from the fastbase64 repo. The use of iterators and chunking the input in the Rust implementation makes it easy for the compiler to vectorise the processing.

The AVX2 implementation is largely untouched compared with the original fastbase64 implementation.

The code is optimised for x86_64, and therefore assumes large-ish caches are available for storing lookup tables. I created a naive implementation that indexed a static array of valid base64 chars - the performance there was only slightly worse than the chromium LUT implementation, so I may add this as an option for low-memory targets (i.e. embedded).

Useful links:

• https://github.com/lemire/fastbase64
• https://www.nickwilcox.com/blog/autovec/
• https://en.wikipedia.org/wiki/Base64

TODO

• Integration tests
• Benchmarking suite
• Comply with MIME, UTF-7, and other Base64 standards
• Regression tests + benchmark in Github Actions
• Change default implementation with feature flags
• Builders for custom configs at runtime