fast_io

fast_io is an extremely fast C++20 input/output library aiming to replace <iostream> and <cstdio>.

It is header only + MIT licensed for easy inclusion in any project. It requires a capable C++20 compiler supporting concepts.

Old Repository was archived due to bloat with git through time; old commits were here.

Examples

Print to stdout

#include <fast_io.h>

int main() {
    println("Hello, fast_io world!");
}

Load whole file through memory mapping

#include <fast_io.h>

int main() {
    fast_io::native_file_loader file_data("text.txt");
    // file_data satisfies std::ranges::contiguous_range
}

Other examples

For other up to date examples please look in the examples/ folder.

Deprecated examples are here but they might not work anymore since fast_io keeps changing.

This i18n repo stores the i18n source files.
https://bitbucket.org/ejsvifq_mabmip/fast_io_i18n

F.A.Q.

What does “fast” mean in fast_io?

It doesn’t necessary mean it will be faster than everything else (or it would be named fastest_io; of course {fmt} trolls would claim that) it means it is significantly faster than <iostream> and <cstdio> in all cases.

What are the supported platforms, compilers, etc.

See support.md for a list of supported “stuff” + platform specific usage notes.

Other questions

You can directly ask questions on the Discord server.
QQ group: 1076846339

Documentation

See either https://ewindy.gitee.io/fast_io_rst/index.html or https://gitee.com/qabeowjbtkwb/fast_io/wikis.

Features

  • As close to system call as possible.

  • Unicode support (UTF-8, UTF-16, UTF-32) + codecvt for GB18030 and UTF-EBCDIC (no libc correctly deals with it)

  • RAII for C FILE*, POSIX fd and win32 HANDLE

  • Interop with <cstdio> and <iostream>

  • No easily misused stuff like std::endl

  • Static I/O manipulators instead of format string nonsense.

  • No locale by default. It is optional.

  • No internal iomanip states (since it creates security issues)

  • Consistent error handling; when available, exceptions as the only error reporting mechanism (no std::error_code, std::system_error or useless bounds checking)

  • Freestanding mode

  • Addresss Sanitizer special code for protecting memory safety issues.

  • Supports POSIX iconv. You can use fast_io for your coding conversion.

  • Binary serialization for trivially copyable types and standard containers

  • Plays well with C++ containers (e.g. std::vector<fast_io::obuf_file> is valid)

  • Basic/Lua/Python/etc format (print, scan). No support for C and C++ since they are security hazards.

  • Provide APIs to expose the internal implementation of FILE* and C++ streams.

  • Native Handle Interface

  • Extremely easy to support custom devices

  • C and C++ style compile time open mode parsing.

  • No traits_type and EOF

  • Dynamic Type Support

  • Multi Process

  • Memory mapping

  • debugging IO (optionally with GUI)

  • Round-trip floating point algorithm

  • Hash algorithms support: intrinsic SHA-1, intrinsic HMAC-SHA1, intrinsic SHA-256, intrinsic HMAC-SHA256, SHA-512, HMAC-SHA512 aswell as non crypto hash algorithms, such as Jenkins Hash.

  • zlib compression/decompression

  • Filesystem

  • OpenSSL BIO, Qt QFile, MFC CFile support

Possible Improvements

  1. Interfaces for cloud computing algorithms like MapReduce
  2. This being in the standard library.

Post C++20

  1. Module support
  2. Coroutine support for async IO
  3. Improve and refactor code once Zero-overhead deterministic exceptions are added to the standard

Benchmarks

  1. I/O 10M integers

Goal: Print out ten million integers from 0 to 10M to file. Then reopen that file to scan back.

All benchmarks are in benchmarks/0000.10m_size_t/unit.

Notice: I modified libstdc++’s BUFSIZ 1048576 due to BUFSIZE is too small (512 bytes) for MinGW-W64, or it performs horribly.

Platform Windows MinGW-W64 GCC 11.0.0 MSVCRT + libstdc++
Method Output time Input time Comment
stdio.h(fprintf/fscanf) 2.412987s 5.607791s
fstream 0.462012s 1.192s
fstream with rdbuf.sputc trick 0.33895s 1.170173s
fast_io::i/obuf_file 0.04903s 0.080996s
fast_io::i/obuf_file_mutex 0.146064s 0.113155s thread safe
c_locale_i/obuf_file (“C”) 0.065988s 0.086012s imbued with locale, locale “C”
c_locale_i/obuf_file local 0.153995s imbued with locale, locale “”
fmt::format_int+obuf_file 0.122999s
fmt::format_int+ofstream 0.209055s
fmt::format+ofstream 0.548s fmt makes things slower
fmt::print 0.663996s fmt makes things slower
std::to_chars+obuf_file 0.12s
std::to_chars+ofstream 0.192s
fast_io::c_file_unlocked 0.098999s 0.126003s I hacked MSVCRT’s FILE* implementation
fast_io::c_file 0.298988s 0.318001s Thread Safe. I hacked MSVCRET’s FILE* implementation
fast_io::filebuf_file 0.048999s 0.081s I hacked libstdc++’s streambuf/filebuf implementation
fast_io::iobuf_utf8_file_char16 0.124s 0.112001s UTF-16=>UTF-8 with SSE
fast_io::iobuf_utf8_file_char32 0.110999s 0.111011s UTF-32=>UTF-8 with SSE
std::wofstream 2.64s 3.843735s wofstream with std::locale codecvt. Extremely slow tbh.
fast_io::wfilebuf_io_observer 2.415692s 2.497704s wofstream with std::locale codecvt. This proves fstream can never get fixed.
Rust language 0.483s RUST IS SLOW. Also Rust does not deal with locale. Think how bad it is.
Rust itoa library 0.4.6 > 0.165s I ignored the \n part for it to ensure no bias.

Rust language is 10x slower than fast_io. + binary bloat and itoa library is still extremely slow and usable for me. It is at least 3x slower than fast_io.

Run the same test on MSVC 19.26.28805.

Platform Windows MSVC 19.26.28805 Install fmtlib wastes time of my life
Method Output time Input time Comment
stdio.h(fprintf/fscanf) 1.5353597s 1.4157233s
fstream 3.6350262s 3.8420339s
fstream with rdbuf.sputc trick 3.3735902s 3.8145566s
fast_io::i/obuf_file 0.0631433s 0.1030554s
fast_io::i/obuf_file_mutex 0.2190659s 0.2485886s thread safe
std::to_chars+obuf_file 0.1641641s
std::to_chars+ofstream 0.5461922s
fast_io::c_file_unlocked 0.1102575s 0.2399757s I hacked Universal CRT’s FILE* implementation
fast_io::c_file 0.2034755s 0.2621148s Thread Safe. I hacked UCRT’s FILE* implementation
fast_io::filebuf_file 0.126661s 0.2378803s I hacked MSVC STL’s streambuf/filebuf implementation

Run the same test on GCC 11. glibc + libstdc++

Platform Linux GCC 11.0.0 glibc + libstdc++
Method Output time Input time Comment
stdio.h(fprintf/fscanf) 0.532792935s 0.591907111s
fstream with rdbuf.sputc trick 0.318896068s 0.429406415s
fast_io::i/obuf_file 0.050300857s 0.065372395s
fast_io::i/obuf_file_mutex 0.05290654s 0.083040518s thread safe
c_locale_i/obuf_file (“C”) 0.051939052s 0.065820056s imbued with locale, locale “C”
c_locale_i/obuf_file local 0.162406082s imbued with locale, locale “”
std::to_chars+obuf_file 0.115453587s
fmt::format_int+obuf_file 0.1183587s
fmt::format_int+ofstream 0.195914384s
fmt::format+ofstream 0.633590975s fmt makes things slower
fmt::print 0.495270371s fmt makes things slower
boost::iostreams 0.400906063s 0.444717051s Using boost iostreams does not make your code faster
fast_io::c_file_unlocked 0.060076723s 0.073299716s I hacked glibc’s FILE* implementation
fast_io::c_file 0.092490191s 0.104545535s Thread Safe. I hacked glibc’s FILE* implementation
fast_io::filebuf_file 0.052251608s 0.06655806s I hacked libstdc++’s streambuf/filebuf implementation

You can see fast_io can also boost the performance of existing facilities for 10x! Yes, it can even improve FILE* and fstream’s performance for 10x depending on platforms since I use concepts to abstract them all. fmtlib actually slows down I/O performance.

  1. Output 10M double in round-trip mode with Ryu algorithm

We only perform this test for MSVC since only msvc’s charconv implements it. Yes. fast_io defeats msvc’s charconv for over 20% for running the same algorithm.

All benchmarks are in benchmarks/0001.10m_double/charconv.

Run the same test on MSVC 19.26.28805.

Platform Windows MSVC 19.26.28805
Method Output time Comment
i/obuf_file 0.4653818s
charconv + obuf_file 0.6011s
  1. Raw I/O Performance

All benchmarks are in benchmarks/0014.file_io/file_io.

Output 100000000x “Hello World\n”

Notice: I modified libstdc++’s std::filebuf’s BUFSIZ to 1048576 due to BUFSIZE is too small (512 bytes) for MinGW-W64 or it performs horribly.

Platform Windows MinGW-W64 GCC 11.0.0 MSVCRT + libstdc++
Method Output time Comment
fwrite 2.524001s
fstream 1.013001s
fast_io::obuf_file 0.437998s
fast_io::obuf_file_mutex 1.371s Thread safe
fast_io::c_file_unlocked 1.164997s I hacked MSVCRT’s FILE* implementation
fast_io::c_file 3.337945s Thread Safe. I hacked MSVCRT’s FILE* implementation. Need further optimization
fast_io::filebuf_file 0.467001s I hacked libstdc++’s std::filebuf implementation
Platform Linux GCC 11.0.0 glibc + libstdc++
Method Output time Comment
fwrite 1.457288317s
fstream 1.249783346s
fast_io::obuf_file 0.494827134s
fast_io::obuf_file_mutex 0.497138826s Thread safe
fast_io::c_file_unlocked 0.687976666s I hacked glibc’s FILE* implementation
fast_io::c_file 0.910792697s Thread Safe. I hacked glibc’s FILE* implementation
fast_io::filebuf_file 0.526955039s I hacked libstdc++’s std::filebuf implementation
Platform Windows MSVC 19.26.28805 UCRT + MSVC STL
Method Output time Comment
fwrite 3.3139122s
fstream 1.7184119s
fast_io::obuf_file 0.7996034s
fast_io::obuf_file_mutex 2.2949221s Thread safe. It looks like std::mutex is horribly slow for MSVC STL.
fast_io::c_file_unlocked 1.2103924s I hacked UCRT’s FILE* implementation
fast_io::c_file 2.3604295s Thread Safe. I hacked UCRT’s FILE* implementation
fast_io::filebuf_file 1.2805368s I hacked MSVC STL’s std::filebuf implementation
  1. Binary Size
    Just use the benchmark in benchmarks/0014.file_io/file_io.
    Dude, you should avoid stream as plague tbh. It is not healthy.
Platform Windows MinGW-W64 GCC 11.0.0 MSVCRT + libstdc++ + static compile
Method Binary Size Comment
fstream 925KB Use fstream is not good for your health since std::locale bloats your binary.
fast_io::obuf_file 155KB
fast_io::c_file_unlocked 157KB I hacked MSVCRT’s FILE* implementation
fast_io::c_file 157KB Thread Safe. I hacked MSVCRT’s FILE* implementation
fast_io::filebuf_file 933KB I hacked libstdc++’s std::filebuf implementation. C++ stream sucks
  1. Code Convert

Generate 100000000 🚄 emojis by using the program in benchmarks/0020.utf/fill_nc.cc

Benchmarks are in examples/0043.iconv
Universal iconv. (UTF-8 to GB18030 as an example)
iconv test:

Platform Windows MinGW-W64 GCC 11.0.0 MSVCRT + libstdc++
Method Elapsed time Comment
iconv command 1.529s
universal.cc 1.293s use POSIX libiconv

UTF8->UTF16LE

Benchmarks are in examples/0022.utf

iconv test:

Platform Windows MinGW-W64 GCC 11.0.0 MSVCRT + libstdc++
Method Elapsed time Comment
iconv command 0.967s GNU iconv. No BOM which sucks
utf8_file_to_utf16_file.cc 0.498s I use the SSE algorithms provided by the utf-utils project.

UTF8->UTF32LE

Benchmarks are in examples/0022.utf

iconv test:

Platform Windows MinGW-W64 GCC 11.0.0 MSVCRT + libstdc++
Method Elapsed time Comment
iconv command 0.844s GNU iconv. No BOM which sucks
utf8_file_to_utf32_file.cc 0.442s I use the SSE algorithms provided by the utf-utils project.

Credits

This project is made possible by referencing other open-source projects. (I don’t take their code directly, they are reimplemented by myself to fit the purpose of this library or it might have integration issues.)

Project Url
Grisu-Exact https://github.com/jk-jeon/Grisu-Exact
Ryu https://github.com/ulfjack/ryu
SHA-Intrinsics https://github.com/noloader/SHA-Intrinsics
SHA1 https://github.com/vog/sha1
UTF-utils https://github.com/BobSteagall/utf_utils
jenkins-hash-java https://github.com/vkandy/jenkins-hash-java
md5 https://github.com/JieweiWei/md5
ReactOS https://github.com/reactos/reactos
dirent_h https://github.com/win32ports/dirent_h
GNU C library https://www.gnu.org/software/libc/
GNU Newlib https://sourceware.org/newlib/
Dragonbox https://github.com/jk-jeon/dragonbox
JEAIII https://github.com/jeaiii/itoa
Crypto++ https://github.com/weidai11/cryptopp

GitHub

View Github