cajeta.math.random.Generator — the random-number surface, built on a counter-based BitGenerator (Philox4x32-10). Counter-based means each output is a stateless function philox(counter, key) of its index and the seed: the stream is reproducible, and element i can be computed independently of all others — embarrassingly parallel, identical on CPU and GPU (the reason for choosing Philox over a sequential generator like PCG64). A Generator holds the 64-bit key (the seed).

package snip.rng;

import cajeta.math.Tensor;
import cajeta.math.random.Generator;

public final class Demo {
    public static void run() {
        Generator g = heap Generator(42);
        Tensor<float32> u #= g.uniform(1000);       // [0, 1)
        Tensor<float32> z #= g.normal(1000);        // N(0, 1)
        Tensor<int64> d #= g.integers(1, 7, 10);    // ten rolls of a die
        return;
    }
}

Methods

Signature
`Generator(uint64 seed)` ⚑	Build a generator keyed by `seed`; the same seed reproduces the same stream
`#Tensor<float32> uniform(int64 n)` ⚑	`n` samples from the uniform distribution over `[0, 1)`: element `i` is the top 24 bits of `philox(i, key)[0]` scaled to `[0, 1)`
`#Tensor<float32> normal(int64 n)`	`n` samples from the standard normal `N(0, 1)` via the Box-Muller transform over counters `2i` and `2i+1`
`#Tensor<int64> integers(int64 low, int64 high, int64 n)`	`n` uniform integers in `[low, high)`: element `i` is `low + (philox(i) mod (high-low))`

⚑ = @EntryPoint

Generator

Methods

See also