0
0
forked from blender/blender
blender/extern/fmtlib/README.rst
Aras Pranckevicius e2e4c1daaa OBJ: use fmt library instead of sprintf for faster formatting
On Windows/MSVC this gives a minor (~20%) speedup presumably due to a faster float/int formatter. On macOS (Xcode13), this gives a massive speedup, since snprintf that is in system libraries ends up spending almost all the time inside some locale-related mutex lock.

The actual exporter code becomes quite a bit smaller too, since it does not have to do any juggling to support std::string arguments, and the buffer handling code is smaller as well.

Windows (VS2022 release build, Ryzen 5950X 32 threads) timings:
- Blender 3.0 splash scene (2.4GB obj): 4.57s -> 3.86s
- Monkey subdivided level 6 (330MB obj): 1.10s -> 0.99s

macOS (Xcode 13 release build, Apple M1Max) timings:
- Blender 3.0 splash scene (2.4GB obj): 21.03s -> 5.52s
- Monkey subdivided level 6 (330MB obj): 3.28s -> 1.20s

Linux (ThreadRipper 3960X 48 threads) timings:
- Blender 3.0 splash scene (2.4GB obj): 10.10s -> 4.40s
- Monkey subdivided level 6 (330MB obj): 2.16s -> 1.37s

The produced obj/mtl files are identical to before.

Reviewed By: Howard Trickey, Dalai Felinto

Differential Revision: https://developer.blender.org/D13998
2022-03-27 14:25:48 +03:00

19 KiB
Raw Permalink Blame History

Sphinx Warnings

2: WARNING: Duplicate explicit target name: "run".
2: WARNING: Duplicate explicit target name: "run".
2: WARNING: Duplicate explicit target name: "run".
2: WARNING: Duplicate explicit target name: "run".

Note the preview is not accurate and warnings may not indicate real issues.

{fmt}

https://projects.blender.org/glukoz/blender/media/tag/v3.3.17/https://github.com/fmtlib/fmt/workflows/linux/badge.svg https://projects.blender.org/glukoz/blender/media/tag/v3.3.17/https://github.com/fmtlib/fmt/workflows/macos/badge.svg https://projects.blender.org/glukoz/blender/media/tag/v3.3.17/https://github.com/fmtlib/fmt/workflows/windows/badge.svg https://projects.blender.org/glukoz/blender/media/tag/v3.3.17/https://ci.appveyor.com/api/projects/status/ehjkiefde6gucy1v?svg=true fmt is continuously fuzzed at oss-fuzz Ask questions at StackOverflow with the tag fmt

{fmt} is an open-source formatting library providing a fast and safe alternative to C stdio and C++ iostreams.

If you like this project, please consider donating to the BYSOL Foundation that helps victims of political repressions in Belarus: https://bysol.org/en/bs/general/.

Documentation

Q&A: ask questions on StackOverflow with the tag fmt.

Try {fmt} in Compiler Explorer.

Features

  • Simple format API with positional arguments for localization

  • Implementation of C++20 std::format

  • Format string syntax similar to Pythons format

  • Fast IEEE 754 floating-point formatter with correct rounding, shortness and round-trip guarantees

  • Safe printf implementation including the POSIX extension for positional arguments

  • Extensibility: support for user-defined types

  • High performance: faster than common standard library implementations of (s)printf, iostreams, to_string and to_chars, see Speed tests and Converting a hundred million integers to strings per second

  • Small code size both in terms of source code with the minimum configuration consisting of just three files, core.h, format.h and format-inl.h, and compiled code; see Compile time and code bloat

  • Reliability: the library has an extensive set of tests and is continuously fuzzed

  • Safety: the library is fully type safe, errors in format strings can be reported at compile time, automatic memory management prevents buffer overflow errors

  • Ease of use: small self-contained code base, no external dependencies, permissive MIT license

  • Portability with consistent output across platforms and support for older compilers

  • Clean warning-free codebase even on high warning levels such as -Wall -Wextra -pedantic

  • Locale-independence by default

  • Optional header-only configuration enabled with the FMT_HEADER_ONLY macro

See the documentation for more details.

Examples

Print to stdout (run)

#include <fmt/core.h>

int main() {
  fmt::print("Hello, world!\n");
}

Format a string (run)

std::string s = fmt::format("The answer is {}.", 42);
// s == "The answer is 42."

Format a string using positional arguments (run)

std::string s = fmt::format("I'd rather be {1} than {0}.", "right", "happy");
// s == "I'd rather be happy than right."

Print chrono durations (run)

#include <fmt/chrono.h>

int main() {
  using namespace std::literals::chrono_literals;
  fmt::print("Default format: {} {}\n", 42s, 100ms);
  fmt::print("strftime-like format: {:%H:%M:%S}\n", 3h + 15min + 30s);
}

Output:

Default format: 42s 100ms
strftime-like format: 03:15:30

Print a container (run)

#include <vector>
#include <fmt/ranges.h>

int main() {
  std::vector<int> v = {1, 2, 3};
  fmt::print("{}\n", v);
}

Output:

[1, 2, 3]

Check a format string at compile time

std::string s = fmt::format("{:d}", "I am not a number");

This gives a compile-time error in C++20 because d is an invalid format specifier for a string.

Write a file from a single thread

#include <fmt/os.h>

int main() {
  auto out = fmt::output_file("guide.txt");
  out.print("Don't {}", "Panic");
}

This can be 5 to 9 times faster than fprintf.

Print with colors and text styles

#include <fmt/color.h>

int main() {
  fmt::print(fg(fmt::color::crimson) | fmt::emphasis::bold,
             "Hello, {}!\n", "world");
  fmt::print(fg(fmt::color::floral_white) | bg(fmt::color::slate_gray) |
             fmt::emphasis::underline, "Hello, {}!\n", "мир");
  fmt::print(fg(fmt::color::steel_blue) | fmt::emphasis::italic,
             "Hello, {}!\n", "世界");
}

Output on a modern terminal:

https://projects.blender.org/glukoz/blender/media/tag/v3.3.17/https://user-images.githubusercontent.com/576385/88485597-d312f600-cf2b-11ea-9cbe-61f535a86e28.png

Benchmarks

Speed tests

Library

Method

Run Time, s

libc

printf

1.04

libc++

std::ostream

3.05

{fmt} 6.1.1

fmt::print

0.75

Boost Format 1.67

boost::format

7.24

Folly Format

folly::format

2.23

{fmt} is the fastest of the benchmarked methods, ~35% faster than printf.

The above results were generated by building tinyformat_test.cpp on macOS 10.14.6 with clang++ -O3 -DNDEBUG -DSPEED_TEST -DHAVE_FORMAT, and taking the best of three runs. In the test, the format string "%0.10f:%04d:%+g:%s:%p:%c:%%\n" or equivalent is filled 2,000,000 times with output sent to /dev/null; for further details refer to the source.

{fmt} is up to 20-30x faster than std::ostringstream and sprintf on floating-point formatting (dtoa-benchmark) and faster than double-conversion and ryu:

https://projects.blender.org/glukoz/blender/media/tag/v3.3.17/https://user-images.githubusercontent.com/576385/95684665-11719600-0ba8-11eb-8e5b-972ff4e49428.png

Compile time and code bloat

The script bloat-test.py from format-benchmark tests compile time and code bloat for nontrivial projects. It generates 100 translation units and uses printf() or its alternative five times in each to simulate a medium sized project. The resulting executable size and compile time (Apple LLVM version 8.1.0 (clang-802.0.42), macOS Sierra, best of three) is shown in the following tables.

Optimized build (-O3)

Method

Compile Time, s

Executable size, KiB

Stripped size, KiB

printf

2.6

29

26

printf+string

16.4

29

26

iostreams

31.1

59

55

{fmt}

19.0

37

34

Boost Format

91.9

226

203

Folly Format

115.7

101

88

As you can see, {fmt} has 60% less overhead in terms of resulting binary code size compared to iostreams and comes pretty close to printf. Boost Format and Folly Format have the largest overheads.

printf+string is the same as printf but with extra <string> include to measure the overhead of the latter.

Non-optimized build

Method

Compile Time, s

Executable size, KiB

Stripped size, KiB

printf

2.2

33

30

printf+string

16.0

33

30

iostreams

28.3

56

52

{fmt}

18.2

59

50

Boost Format

54.1

365

303

Folly Format

79.9

445

430

libc, lib(std)c++ and libfmt are all linked as shared libraries to compare formatting function overhead only. Boost Format is a header-only library so it doesnt provide any linkage options.

Running the tests

Please refer to Building the library for the instructions on how to build the library and run the unit tests.

Benchmarks reside in a separate repository, format-benchmarks, so to run the benchmarks you first need to clone this repository and generate Makefiles with CMake:

$ git clone --recursive https://github.com/fmtlib/format-benchmark.git
$ cd format-benchmark
$ cmake .

Then you can run the speed test:

$ make speed-test

or the bloat test:

$ make bloat-test

Migrating code

clang-tidy-fmt provides clang tidy checks for converting occurrences of printf and fprintf to fmt::print.

Projects using this library

  • 0 A.D.: a free, open-source, cross-platform real-time strategy game

  • 2GIS: free business listings with a city map

  • AMPL/MP: an open-source library for mathematical programming

  • Aseprite: animated sprite editor & pixel art tool

  • AvioBook: a comprehensive aircraft operations suite

  • Blizzard Battle.net: an online gaming platform

  • Celestia: real-time 3D visualization of space

  • Ceph: a scalable distributed storage system

  • ccache: a compiler cache

  • ClickHouse: analytical database management system

  • CUAUV: Cornell Universitys autonomous underwater vehicle

  • Drake: a planning, control, and analysis toolbox for nonlinear dynamical systems (MIT)

  • Envoy: C++ L7 proxy and communication bus (Lyft)

  • FiveM: a modification framework for GTA V

  • fmtlog: a performant fmtlib-style logging library with latency in nanoseconds

  • Folly: Facebook open-source library

  • Grand Mountain Adventure: A beautiful open-world ski & snowboarding game

  • HarpyWar/pvpgn: Player vs Player Gaming Network with tweaks

  • KBEngine: an open-source MMOG server engine

  • Keypirinha: a semantic launcher for Windows

  • Kodi (formerly xbmc): home theater software

  • Knuth: high-performance Bitcoin full-node

  • Microsoft Verona: research programming language for concurrent ownership

  • MongoDB: distributed document database

  • MongoDB Smasher: a small tool to generate randomized datasets

  • OpenSpace: an open-source astrovisualization framework

  • PenUltima Online (POL): an MMO server, compatible with most Ultima Online clients

  • PyTorch: an open-source machine learning library

  • quasardb: a distributed, high-performance, associative database

  • Quill: asynchronous low-latency logging library

  • QKW: generalizing aliasing to simplify navigation, and executing complex multi-line terminal command sequences

  • redis-cerberus: a Redis cluster proxy

  • redpanda: a 10x faster Kafka® replacement for mission critical systems written in C++

  • rpclib: a modern C++ msgpack-RPC server and client library

  • Salesforce Analytics Cloud: business intelligence software

  • Scylla: a Cassandra-compatible NoSQL data store that can handle 1 million transactions per second on a single server

  • Seastar: an advanced, open-source C++ framework for high-performance server applications on modern hardware

  • spdlog: super fast C++ logging library

  • Stellar: financial platform

  • Touch Surgery: surgery simulator

  • TrinityCore: open-source MMORPG framework

  • Windows Terminal: the new Windows terminal

More…

If you are aware of other projects using this library, please let me know by email or by submitting an issue.

Motivation

So why yet another formatting library?

There are plenty of methods for doing this task, from standard ones like the printf family of function and iostreams to Boost Format and FastFormat libraries. The reason for creating a new library is that every existing solution that I found either had serious issues or didnt provide all the features I needed.

printf

The good thing about printf is that it is pretty fast and readily available being a part of the C standard library. The main drawback is that it doesnt support user-defined types. printf also has safety issues although they are somewhat mitigated with __attribute__ ((format (printf, …)) in GCC. There is a POSIX extension that adds positional arguments required for i18n to printf but it is not a part of C99 and may not be available on some platforms.

iostreams

The main issue with iostreams is best illustrated with an example:

std::cout << std::setprecision(2) << std::fixed << 1.23456 << "\n";

which is a lot of typing compared to printf:

printf("%.2f\n", 1.23456);

Matthew Wilson, the author of FastFormat, called this “chevron hell”. iostreams dont support positional arguments by design.

The good part is that iostreams support user-defined types and are safe although error handling is awkward.

Boost Format

This is a very powerful library which supports both printf-like format strings and positional arguments. Its main drawback is performance. According to various benchmarks, it is much slower than other methods considered here. Boost Format also has excessive build times and severe code bloat issues (see Benchmarks).

FastFormat

This is an interesting library which is fast, safe and has positional arguments. However, it has significant limitations, citing its author:

Three features that have no hope of being accommodated within the current design are:

  • Leading zeros (or any other non-space padding)

  • Octal/hexadecimal encoding

  • Runtime width/alignment specification

It is also quite big and has a heavy dependency, STLSoft, which might be too restrictive for using it in some projects.

Boost Spirit.Karma

This is not really a formatting library but I decided to include it here for completeness. As iostreams, it suffers from the problem of mixing verbatim text with arguments. The library is pretty fast, but slower on integer formatting than fmt::format_to with format string compilation on Karmas own benchmark, see Converting a hundred million integers to strings per second.

License

{fmt} is distributed under the MIT license.

Documentation License

The Format String Syntax section in the documentation is based on the one from Python string module documentation. For this reason the documentation is distributed under the Python Software Foundation license available in doc/python-license.txt. It only applies if you distribute the documentation of {fmt}.

Maintainers

The {fmt} library is maintained by Victor Zverovich (vitaut) and Jonathan Müller (foonathan) with contributions from many other people. See Contributors and Releases for some of the names. Let us know if your contribution is not listed or mentioned incorrectly and well make it right.