fast_float number parsing library: 4x faster than strtod

The fast_float library provides fast header-only implementations for the C++ from_chars functions for float and double types. These functions convert ASCII strings representing decimal values (e.g., 1.3e10) into binary types. We provide exact rounding (including round to even). In our experience, these fast_float functions many times faster than comparable number-parsing functions from existing C++ standard libraries.

Specifically, fast_float provides the following two functions with a C++17-like syntax (the library itself only requires C++11):

from_chars_result from_chars(const char* first, const char* last, float& value, ...);
from_chars_result from_chars(const char* first, const char* last, double& value, ...);

The return type (from_chars_result) is defined as the struct:

struct from_chars_result {
    const char* ptr;
    std::errc ec;
};

It parses the character sequence [first,last) for a number. It parses floating-point numbers expecting a locale-independent format equivalent to what is used by std::strtod in the default ("C") locale. The resulting floating-point value is the closest floating-point values (using either float or double), 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.

Given a successful parse, the pointer (ptr) in the returned value is set to point right after the parsed number, and the value referenced is set to the parsed value. In case of error, the returned ec contains a representative error, otherwise the default (std::errc()) value is stored.

The implementation does not throw and does not allocate memory (e.g., with new or malloc).

It will parse infinity and nan values.

Example:

#include "fast_float/fast_float.h"
#include <iostream>
 
int main() {
    const std::string input =  "3.1416 xyz ";
    double result;
    auto answer = fast_float::from_chars(input.data(), input.data()+input.size(), result);
    if(answer.ec != std::errc()) { std::cerr << "parsing failure\n"; return EXIT_FAILURE; }
    std::cout << "parsed the number " << result << std::endl;
    return EXIT_SUCCESS;
}

Like the C++17 standard, the fast_float::from_chars functions take an optional last argument of the type fast_float::chars_format. It is a bitset value: we check whether fmt & fast_float::chars_format::fixed and fmt & fast_float::chars_format::scientific are set to determine whether we allow the fixed point and scientific notation respectively. The default is fast_float::chars_format::general which allows both fixed and scientific.

The library seeks to follow the C++17 (see 20.19.3.(7.1)) specification. In particular, it forbids leading spaces and the leading '+' sign.

We support Visual Studio, macOS, Linux, freeBSD. We support big and little endian. We support 32-bit and 64-bit systems.

Using commas as decimal separator

The C++ standard stipulate that from_chars has to be locale-independent. In particular, the decimal separator has to be the period (.). However, some users still want to use the fast_float library with in a locale-dependent manner. Using a separate function called from_chars_advanced, we allow the users to pass a parse_options instance which contains a custom decimal separator (e.g., the comma). You may use it as follows.

#include "fast_float/fast_float.h"
#include <iostream>
 
int main() {
    const std::string input =  "3,1416 xyz ";
    double result;
    fast_float::parse_options options{fast_float::chars_format::general, ','};
    auto answer = fast_float::from_chars_advanced(input.data(), input.data()+input.size(), result, options);
    if((answer.ec != std::errc()) || ((result != 3.1416))) { std::cerr << "parsing failure\n"; return EXIT_FAILURE; }
    std::cout << "parsed the number " << result << std::endl;
    return EXIT_SUCCESS;
}

Reference

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

Other programming languages

• There is an R binding called rcppfastfloat.
• There is a Rust port of the fast_float library called fast-float-rust.
• There is a Java port of the fast_float library called FastDoubleParser.
• There is a C# port of the fast_float library called csFastFloat.

Relation With Other Work

The fast_float library provides a performance similar to that of the fast_double_parser library but using an updated algorithm reworked from the ground up, and while offering an API more in line with the expectations of C++ programmers. The fast_double_parser library is part of the Microsoft LightGBM machine-learning framework.

Users

The fast_float library is used by Apache Arrow where it multiplied the number parsing speed by two or three times. It is also used by Yandex ClickHouse and by Google Jsonnet.

How fast is it?

It can parse random floating-point numbers at a speed of 1 GB/s on some systems. We find that it is often twice as fast as the best available competitor, and many times faster than many standard-library implementations.

$ ./build/benchmarks/benchmark 
# parsing random integers in the range [0,1)
volume = 2.09808 MB 
netlib                                  :   271.18 MB/s (+/- 1.2 %)    12.93 Mfloat/s  
doubleconversion                        :   225.35 MB/s (+/- 1.2 %)    10.74 Mfloat/s  
strtod                                  :   190.94 MB/s (+/- 1.6 %)     9.10 Mfloat/s  
abseil                                  :   430.45 MB/s (+/- 2.2 %)    20.52 Mfloat/s  
fastfloat                               :  1042.38 MB/s (+/- 9.9 %)    49.68 Mfloat/s  

See https://github.com/lemire/simple_fastfloat_benchmark for our benchmarking code.

Video

Using as a CMake dependency

This library is header-only by design. The CMake file provides the fast_float target which is merely a pointer to the include directory.

If you drop the fast_float repository in your CMake project, you should be able to use it in this manner:

add_subdirectory(fast_float)
target_link_libraries(myprogram PUBLIC fast_float)

Or you may want to retrieve the dependency automatically if you have a sufficiently recent version of CMake (3.11 or better at least):

FetchContent_Declare(
  fast_float
  GIT_REPOSITORY https://github.com/lemire/fast_float.git
  GIT_TAG tags/v1.1.2
  GIT_SHALLOW TRUE)

FetchContent_MakeAvailable(fast_float)
target_link_libraries(myprogram PUBLIC fast_float)

You should change the GIT_TAG line so that you recover the version you wish to use.

Using as single header

The script script/amalgamate.py may be used to generate a single header version of the library if so desired. Just run the script from the root directory of this repository. You can customize the license type and output file if desired as described in the command line help.

You may directly download automatically generated single-header files:

https://github.com/fastfloat/fast_float/releases/download/v1.1.2/fast_float.h

Credit

Though this work is inspired by many different people, this work benefited especially from exchanges with Michael Eisel, who motivated the original research with his key insights, and with Nigel Tao who provided invaluable feedback. Rémy Oudompheng first implemented a fast path we use in the case of long digits.

The library includes code adapted from Google Wuffs (written by Nigel Tao) which was originally published under the Apache 2.0 license.

License

^{Licensed under either of Apache License, Version 2.0 or MIT license at your option.}
_{Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this repository by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.}