#f64 #f32 #float #serialization #formatting #string-conversion #format

float_fast_print

Fast float-to-string conversion based on the Ryū algorithm by Ulf Adams

1 unstable release

Uses old Rust 2015

0.1.2 Jul 30, 2018
0.1.1 Jul 29, 2018
0.1.0 Jul 29, 2018

#1257 in Encoding

Apache-2.0 OR MIT

66KB
853 lines

Overview

A native Rust implementation of the high-performance Ryū float-to-string conversion algorithm by Ulf Adams of Google Germany.

Description

The purpose of this crate is to improve the speed of converting f32 and f64 to string representations, which is particularly important when performing bulk serialisation of JSON, CSV, XLSX, and scientific data formats where all numbers are internally treated as floating point. In some cases, this can be the bottleneck to serialisation.

During development of this code it was noted that the serde_json crate references the dtoa crate, which is both slower and contains unsafe code. This crate uses no unsafe code, and could potentially use #![no_std] as well.

Status

Things are looking promising, with the Ryū bench_write_f32_shortest function significantly outperforming the currently available alternatives:

test tests::bench_dtoa               ... bench:          45 ns/iter (+/- 1)
test tests::bench_f32_debug          ... bench:         108 ns/iter (+/- 0)
test tests::bench_f32_format         ... bench:         109 ns/iter (+/- 0)
test tests::bench_write_f32_shortest ... bench:          25 ns/iter (+/- 0)

The f32 version of the function has been exhaustively tested in version 0.1.1 of the crate, guaranteeing that all possible 2^32 float values can round-trip with no loss of precision when converted using this library:

...
99.930916% complete
99.95419% complete
99.97748% complete
test test_exhaustive_roundtrip ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

A less exhaustive test for the f64 version is available, but obviously can't be exhaustive, as it would take years to run.

Pending Tasks

[ ] Implement 2-digit lookup table and benchmark. [ ] Compare performance to alternatives when the cache has been flushed. [ ] Add JSON-style printing that uses the integer representation for whole numbers.

References

No runtime deps