speedy-int128 0.1.0

speedy-int128

It's technically a fork of std.int128 with added inline LLVM IR for the LDC compiler to make it faster at handling 128-bit integers. This makes it as fast as Clang, because Clang was actually used as a "donor" of this LLVM IR code via a simple script.

This package also enables access to 128-bit arithmetics for the ancient versions of DMD, GDC and LDC, which don't have the standard std.int128 module yet.

And finally, a oneliner variant is provided for use on programming competition websites.

Example

/+dub.sdl: dependency "speedy-int128" version="~>0.1.0" +/
import speedy.int128; // instead of "std.int128"
import std.stdint, std.stdio, std.range, std.algorithm;

// https://lemire.me/blog/2019/03/19/the-fastest-conventional-random-number-generator-that-can-pass-big-crush/
uint64_t lehmer64() {
  static Int128 g_lehmer64_state = Int128(1L); /* bad seed */
  g_lehmer64_state *= 0xda942042e4dd58b5;
  return g_lehmer64_state.data.hi;
}

void main() {
  1_000_000_000.iota.map!(i => lehmer64).sum.writeln;
}

Install the DUB package manager and run the example in a script-like fashion:

$ dub example.d

Or compile an optimized binary using the LDC compiler:

$ dub build --build release --single --compiler=ldc2 example.d

Performance

Benchmarks are done using the benchmark.d / benchmark.c test programs as part of CI. The optimization options are whatever the DUB tool considers default for producing release builds. Some examples:

<details> <summary>GitHub Actions CI, Linux x86_64, Intel(R) Xeon(R) Platinum 8370C CPU @ 2.80GHz</summary>

https://github.com/ssvb/speedy-int128/actions/runs/3859195372/jobs/6578500703

</details>

<details> <summary>GitHub Actions CI, Linux x86_64, Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz</summary>

https://github.com/ssvb/speedy-int128/actions/runs/3859220724/jobs/6578545848

</details>

<details> <summary>BuildJet CI, Linux aarch64, ARM Neoverse-N1</summary>

https://github.com/ssvb/speedy-int128/actions/runs/3859220721/jobs/6578545846

</details>

Use on programming competition websites

Programming competition websites, such as Codeforces and AtCoder, allow using D language for submitting solutions. But their compilers are typically very old and also installed without any third-party libraries. Needless to say that DUB packages can't be used there in a normal way. Another challenge is that each solution has to be submitted as a single source file with a certain size limit (only 65535 bytes on Codeforces!).

The onelinerizer.rb script can be used to compress the original 42K of D code into a single 16K line by removing comments, extra whitespaces and unittests. The result is speedyint128oneliner.d, which can be copy-pasted into the source code replacing the "import speedy.int128;" line.