Random Numbers

cuTile provides random number generation inside kernels, using a tile-vectorised Philox2x32-7 generator.

Operation	Description
`rand()`	Uniform `Float32` scalar in `(0, 1]` (uses `Random.default_rng()`)
`rand(T)` / `rand(T, dims)` / `rand(dims)`	Uniform scalar or tile
`randn()` / `randn(T)` / `randn(T, dims)`	Standard normal scalar or tile (Box-Muller)
`randexp()` / `randexp(T)` / `randexp(T, dims)`	Standard exponential scalar or tile (`-log(U)`)
`ct.DeviceRNG()`	Open an independent RNG stream
`Random.rand(rng, ...)` / `Random.randn(rng, ...)` / `Random.randexp(rng, ...)`	Explicit-stream variants
`Random.seed!(rng, seed)`	Re-seed a stream

rand supports all of Int{8,16,32,64}, UInt{8,16,32,64}, Float16, BFloat16, Float32 and Float64; randn and randexp cover the four floating-point types. Different DeviceRNG() call sites yield independent streams, all keyed on a per-launch host seed for cross-launch divergence.

The floating-point generator deliberately excludes zero so that randn and randexp can safely take a logarithm. The Float32 conversion can round its largest values to exactly 1.0f0; the narrower and Float64 conversions produce values strictly between zero and one. This differs from Base's usual [0, 1) interval.

julia> function noise(out)
           pid = ct.bid(1)
           t = randn(Float32, (256,))         # default RNG, standard normal
           ct.store(out; index=pid, tile=t)
           return
       end;

julia> out = CUDA.zeros(Float32, 1024);

julia> @cuda backend=cuTile blocks=4 noise(out);

julia> @assert all(isfinite, Array(out))

The same generator is available on the host, for filling CuArrays without writing a kernel; see Host-level Operations.