fish-shell/fast-float-rust

fast-float

This crate provides a super-fast decimal number parser from strings into floats.

[dependencies]
fast-float = "0.2"

There are no dependencies and the crate can be used in a no_std context by disabling the "std" feature.

Compiler support: rustc 1.37+.

Usage

There's two top-level functions provided: parse() and parse_partial(), both taking either a string or a bytes slice and parsing the input into either f32 or f64:

• parse() treats the whole string as a decimal number and returns an error if there are invalid characters or if the string is empty.
• parse_partial() tries to find the longest substring at the beginning of the given input string that can be parsed as a decimal number and, in the case of success, returns the parsed value along the number of characters processed; an error is returned if the string doesn't start with a decimal number or if it is empty. This function is most useful as a building block when constructing more complex parsers, or when parsing streams of data.

Example:

// Parse the entire string as a decimal number.
let s = "1.23e-02";
let x: f32 = fast_float::parse(s).unwrap();
assert_eq!(x, 0.0123);

// Parse as many characters as possible as a decimal number.
let s = "1.23e-02foo";
let (x, n) = fast_float::parse_partial::<f32, _>(s).unwrap();
assert_eq!(x, 0.0123);
assert_eq!(n, 8);
assert_eq!(&s[n..], "foo");

Details

This crate is a direct port of Daniel Lemire's fast_float C++ library (valuable discussions with Daniel while porting it helped shape the crate and get it to the performance level it's at now), with some Rust-specific tweaks. Please see the original repository for many useful details regarding the algorithm and the implementation.

The parser is locale-independent. The resulting value is the closest floating-point values (using either f32 or f64), using the "round to even" convention for values that would otherwise fall right in-between two values. That is, we provide exact parsing according to the IEEE standard.

Infinity and NaN values can be parsed, along with scientific notation.

Both little-endian and big-endian platforms are equally supported, with extra optimizations enabled on little-endian architectures.

Testing

There are a few ways this crate is tested:

• A suite of explicit tests (taken from the original library) covering lots of edge cases.
• A file-based test suite (taken from the original library; credits to Nigel Tao), ~5M tests.
• All 4B float32 numbers are exhaustively roundtripped via ryu formatter.
• Roundtripping a large quantity of random float64 numbers via ryu formatter.
• Roundtripping float64 numbers and fuzzing random input strings via cargo-fuzz.
• All explicit test suites run on CI; roundtripping and fuzzing are run manually.

Performance

The presented parser seems to beat all of the existing C/C++/Rust float parsers known to us at the moment by a large margin, in all of the datasets we tested it on so far – see detailed benchmarks below (the only exception being the original fast_float C++ library, of course – performance of which is within noise bounds of this crate). On modern machines like Apple M1, parsing throughput can reach up to 1.5 GB/s.

In particular, it is faster than Rust standard library's FromStr::from_str() by a factor of 2-8x (larger factor for longer float strings), and is typically 2-3x faster than the nearest competitors.

While various details regarding the algorithm can be found in the repository for the original C++ library, here are few brief notes:

• The parser is specialized to work lightning-fast on inputs with at most 19 significant digits (which constitutes the so called "fast-path"). We believe that most real-life inputs should fall under this category, and we treat longer inputs as "degenerate" edge cases since it inevitable causes overflows and loss of precision.
• If the significand happens to be longer than 19 digits, the parser falls back to the "slow path", in which case its performance roughly matches that of the top Rust/C++ libraries (and still beats them most of the time, although not by a lot).
• On little-endian systems, there's additional optimizations for numbers with more than 8 digits after the decimal point.

Benchmarks

Below are tables of best timings in nanoseconds for parsing a single number into a 64-bit float.

Intel i7-4771

Intel i7-4771 3.5GHz, macOS, Rust 1.49.

Apple M1

Apple M1, macOS, Rust 1.49.

AMD Rome

AMD Rome, Linux, Rust 1.49.

Parsers

• fast-float - this very crate
• lexical – lexical_core, v0.7 (non-lossy; same performance as lossy)
• from_str – Rust standard library, FromStr trait
• fast_float (C++) – original C++ implementation of 'fast-float' method
• abseil (C++) – Abseil C++ Common Libraries
• netlib (C++) – C++ Network Library
• strtod (C) – C standard library

Datasets

• canada – numbers in canada.txt file
• mesh – numbers in mesh.txt file
• uniform – uniform random numbers from 0 to 1
• iidi – random numbers of format %d%d.%d
• iei – random numbers of format %de%d
• rec32 – reciprocals of random 32-bit integers

Notes

• The two test files referred above can be found in this repository.
• The Rust part of the table (along with a few other benchmarks) can be generated via the benchmark tool that can be found under extras/simple-bench of this repo.
• The C/C++ part of the table (along with a few other benchmarks and parsers) can be generated via a C++ utility that can be found in this repository.

References

• Daniel Lemire, Number Parsing at a Gigabyte per Second, Software: Practice and Experience 51 (8), 2021.

License