Devices

Bind a generator

TandemRNG uses MLDataDevices in its core interface. New generators are CPU-bound. Bind a generator before allocating or filling arrays on another backend. Install the backend together with KernelAbstractions and GPUArraysCore. For CUDA:

using Pkg
Pkg.add(["CUDA", "KernelAbstractions", "GPUArraysCore", "MLDataDevices"])

using TandemRNG, MLDataDevices
using CUDA, KernelAbstractions, GPUArraysCore
rng = Tandem8x32(42) |> CUDADevice()
values, rng = rand_next(rng, Float32, 1024)
rng = rand_fill!(rng, values)
CUDA.synchronize()
host_values = Array(values)
cpu_rng = rng |> CPUDevice()

Binding preserves the key, position, and cached stream state. Draws and child derivations retain the binding. Allocating draws use the bound backend. Destination fills reject a backend mismatch before mutation, including for empty arrays. Unsupported operations do not fall back to CPU execution.

Tokens identify a backend, not a physical GPU. Select the physical device through the backend, such as CUDA.device!, before allocation and execution. Keep the active device and arrays consistent.

Backend support

Load AMDGPU or Metal in place of CUDA for those backends. Load GPUArrays as well for Metal's optional precompile extension. Supported versions are CUDA 5.8 or later in 5.x, CUDA 6.x, AMDGPU 2.x, and Metal 1.7 or later in 1.x. AMDGPU hardware validation remains open.

GPU fills are asynchronous. Synchronize before timing completion or reading results from the host. CPU fill thread settings do not select GPU threads.

Scalar calls and mutable state

Host scalar calls run on the host even when the generator is GPU-bound. Scalar calls inside a GPU kernel run on the device. A backend token does not move a host call into a kernel.

Stateful(rng) explicitly rebinds its state to the CPU. Use immutable generators for device work. See Reactant for compilation with runtime keys and positions.