mir-random

Professional Random Number Generators

Documentation: http://docs.random.dlang.io/latest/index.html

Example (3 seconds)

#!/usr/bin/env dub
/+ dub.json:
{
    "name": "test_random",
    "dependencies": {"mir-random": "~>0.4.0"}
}
+/

import std.range, std.stdio;

import mir.random;
import mir.random.variable: NormalVariable;
import mir.random.algorithm: range;

void main()
{
    auto sample = NormalVariable!double(0, 1).range.take(10).array;

    sample[$.randIndex].writeln; // prints random element from the sample
}

Example (10 seconds)

#!/usr/bin/env dub
/+ dub.json:
{
    "name": "test_random",
    "dependencies": {"mir-random": "~>0.4.0"}
}
+/

import std.range, std.stdio;

import mir.random;
import mir.random.variable: NormalVariable;
import mir.random.algorithm: range;


void main(){
    auto rng = Random(unpredictableSeed);        // Engines are allocated on stack or global
    auto sample = range!rng                      // Engines can passed to algorithms by alias or by pointer
        (NormalVariable!double(0, 1))            // Random variables are passed by value
        .take(10)                                // Fix sample length to 10 elements (Input Range API)
        .array;                                  // Allocates memory and performs computation

    sample[$.randIndex].writeln;                 // prints random element from the sample
}

Comparison with Phobos

• Does not depend on DRuntime (Better C concept)

random (new implementation and API)

• Mir Random rand!float/rand!double/rand!real generates saturated real random numbers in (-1, 1). For example, rand!real can produce more than 2^78 unique numbers. In other hand, std.random.uniform01!real produces less than 2^31 unique numbers with default Engine.
• Mir Random fixes Phobos integer underflow bugs.
• Additional optimization was added for enumerated types.
• Random nd-array (ndslice) generation.
• Bounded integer generation in randIndex uses Daniel Lemire's fast alternative to modulo reduction. The throughput increase measured for randIndex!uint on an x86-64 processor compiled with LDC 1.6.0 was 1.40x for Mt19937_64 and 1.73x for Xoroshiro128Plus. The throughput increase measured for randIndex!ulong was 2.36x for Mt19937_64 and 4.25x for Xoroshiro128Plus.

random.variable (new)

• Uniform
• Exponential
• Gamma
• Normal
• Cauchy
• ...

random.ndvariable (new)

• Simplex
• Sphere
• Multivariate Normal
• ...

random.algorithm (new)

• Ndslice and range API adaptors

random.engine.* (fixed, reworked, new)

• opCall API instead of range interface is used (similar to C++)
• No default and copy constructors are allowed for generators.
• unpredictableSeed has not state, returns size_t
• Any unsigned generators are allowed.
• min property was removed. Any integer generator can normalize its minimum down to zero.
• Mt19937: +100% performance for initialization. (merged to Phobos)
• Mt19937: +54% performance for generation. (merged to Phobos)
• Mt19937: fixed to be more CPU cache friendly. (merged to Phobos)
• 64-bit Mt19937 initialization is fixed (merged to Phobos)
• 64-bit Mt19937 is default for 64-bit targets
• Permuted Congruential Generators (new)
• SplitMix generators (new)
• Fixed XorshiftEngine's support for non-uint word sizes & allow various shift directions
• XorshiftStar Generators (new)
• Xoroshiro128Plus generator (new)
• Xoshiro256StarStar & Xoshiro128StarStar_32 generators (new)