An adaptation of Wojciech Mula's AVX-512 base64 encoding/decoding algorithms. Supports all of the major alphabets, and a portable simd path.
Comes as a CLI tool and a library:
Usage: base32-fast [OPTIONS] [INPUT]
Arguments:
[INPUT] The file from which input will be read. If not provided, read from stdin
Options:
-d, --decode Whether to decode the input
-e, --encode Whether to encode the input
--alphabet <ALPHABET> The alphabet to use (rfc4648, rfc4648hex, crockford, geohash, or z; default rfc4648)
-n Whether to not output a newline after the encoded or decoded text
-h, --help Print help
use base32_fast::{Rfc4648, Rfc4648Hex, Crockford, Geohash, Z, b32enc, b32dec};
let encoded = b32enc(&src, &mut dst, Geohash);
let decoded = b32dec(&src, &mut dst, Geohash);base32-fast provides three code paths: A reference scalar implementation, an
implementation using Rust's std::simd, and a handwritten AVX-512
implementation.
If you can use the AVX-512 path, do so. The encoder uses multishift instructions and the decoder abuses integer FMA instructions to maximize speed, and the SIMD implementation can do neither.
-Zbuild-std is required for the portable SIMD path to be fast on some
architectures, especially those with AVX-512.
Benchmarks of each code path per 64-byte block:
test dec::tests::bench_b32dec ... bench: 21.84 ns/iter (+/- 0.59)
test dec::tests::bench_b32dec_simd ... bench: 2.76 ns/iter (+/- 0.01)
test dec::tests::bench_b32dec_avx512 ... bench: 1.16 ns/iter (+/- 0.17)
test enc::tests::bench_b32enc ... bench: 58.35 ns/iter (+/- 0.42)
test enc::tests::bench_b32enc_simd ... bench: 10.42 ns/iter (+/- 0.02)
test enc::tests::bench_b32enc_avx512 ... bench: 1.14 ns/iter (+/- 0.03)
Equivalent benchmarks of the base32 crate for comparison:
test dec::tests::bench_base32_decode ... bench: 39.12 ns/iter (+/- 0.34)
test enc::tests::bench_base32_encode ... bench: 66.15 ns/iter (+/- 1.57)
Benchmarks with 1MB of data:
test dec::tests::bench_b32dec_avx512_bulk ... bench: 199,036.20 ns/iter (+/- 20,178.75)
test dec::tests::bench_b32dec_simd_bulk ... bench: 661,083.40 ns/iter (+/- 4,813.39)
test enc::tests::bench_b32enc_avx512_bulk ... bench: 184,308.08 ns/iter (+/- 2,036.40)
test enc::tests::bench_b32enc_simd_bulk ... bench: 351,228.25 ns/iter (+/- 4,039.00)
Compared to base32:
test dec::tests::bench_base32_decode_bulk ... bench: 8,315,518.80 ns/iter (+/- 77,973.30)
test enc::tests::bench_base32_encode_bulk ... bench: 7,334,881.90 ns/iter (+/- 106,046.64)
