dtoa Benchmark

This is a fork of Milo Yip's dtoa-benchmark with the following changes:

• CMake support
• Fixed reporting of results
• Added {fmt}
• Added Dragonbox
• Added Schubfach
• Removed the use of deprecated strstream
• Disabled Grisu2 implementations since they don't guarantee correctness

Copyright(c) 2014 Milo Yip (miloyip@gmail.com)

Introduction

This benchmark evaluates the performance of conversion from double precision IEEE-754 floating point (double) to ASCII string. The function prototype is:

void dtoa(double value, char* buffer);

The character string result must be convertible to the original value exactly via some correct implementation of strtod, i.e. roundtrip convertible.

Note that dtoa is not a standard function in C and C++.

Procedure

Firstly the program verifies the correctness of implementations.

Then, one case for benchmark is carried out:

1. RandomDigit: Generates 1000 random double values, filtered out +/-inf and nan. Then convert them to limited precision (1 to 17 decimal digits in significand). Finally convert these numbers into ASCII.

Each digit group is run for 100 times. The minimum time duration is measured for 10 trials.

Build and Run

1. Configure: cmake .
2. Build and run benchmark: make run-benchmark

The results in CSV format will be written to the file result/<cpu>_<os>_<compiler>.csv and automatically converted to HTML with the same base name and the .html extension.

Results

The following are results measured on a MacBook Pro (Apple M1 Pro), where dtoa is compiled by Apple clang 17.0.0 (clang-1700.0.13.5) and run on macOS. The speedup is based on sprintf.

Function	Time (ns)	Speedup
ostringstream	875.625	1.00x
sprintf	735.428	1.19x
doubleconv	83.249	10.52x
ryu	37.163	23.56x
schubfach	24.555	35.66x
fmt	22.360	39.16x
dragonbox	20.790	42.12x
null	0.936	935.03x

Notes:

• The null implementation does nothing. It measures the overheads of looping and function call.
• sprintf and ostringstream don't generate the shortest representation, e.g. 0.1 is formatted as 0.10000000000000001.
• ryu, dragonbox and schubfach only produce the output in the exponential format, e.g. 0.1 is formatted as 1E-1 or similar.

Some results of various configurations are located at result. They can be accessed online, with interactivity provided by Google Charts:

• apple-m1-pro_mac64_clang17.0

Implementations

Function	Description
ostringstream	std::ostringstream in C++ standard library with setprecision(17).
sprintf	sprintf in C standard library with "%.17g" format.
doubleconv	C++ implementation extracted from Google's V8 JavaScript Engine with EcmaScriptConverter().ToShortest() (based on Grisu3, fall back to slower bignum algorithm when Grisu3 failed to produce shortest implementation).
fmt	fmt::format_to with format string compilation (implements Dragonbox).
dragonbox	jkj::dragonbox::to_chars with full tables.
schubfach	Schubfach implementation in C++
null	Do nothing.

Notes:

std::to_string is not tested as it does not fulfill the roundtrip requirement (until C++26).

FAQ

1. How to add an implementation?

   You may clone an existing implementation file. And then modify it and add to the CMake config. Note that it will automatically register to the benchmark by macro REGISTER_TEST(name).

   Making a pull request of new implementations is welcome.

2. Why not converting double to std::string?

   It may introduce heap allocation, which is a big overhead. User can easily wrap these low-level functions to return std::string, if needed.

3. Why fast dtoa functions is needed?

   They are a very common operations in writing data in text format. The standard way of sprintf, std::stringstream, often provides poor performance. The author of this benchmark would optimize the sprintf implementation in RapidJSON, thus he creates this project.

Related Benchmarks and Discussions

• Printing Floating-Point Numbers