Running on GPU

To use this package on a GPU, you need to use KernelAbstractions.jl to define the arrays and run the computations. The KernelAbstractions.jl package introduces a macro-based programming model that simplifies vendor-specific GPU programming by abstracting away its complexities. This allows hardware-independent kernels to be written that can be compiled and executed on different device backends without altering the high-level code or compromising performance.

This will load an extension from FiniteDiffWENO5GPU.jl:

using KernelAbstractions
using FiniteDiffWENO5

That way, the two functions WENOScheme() and WENO_step!() will automatically dispatch to the array types if the argument backend is provided. See the KernelAbstractions.jl documentation for more details on how to set up the GPU backend.

Simplex-constrained material fractions use the same phase ordering and boundary semantics on every backend:

scheme = MultiphaseWENOScheme(phases, backend;
    boundary = ntuple(_ -> PeriodicBC(), 4), stag = true)
WENO_step!(phases, velocity, scheme, Δt, Δx, Δy, backend)

# Chmy fields
scheme = MultiphaseWENOScheme(phases, grid;
    boundary = ntuple(_ -> PeriodicBC(), 4), stag = true)
WENO_step!(phases, velocity, scheme, Δt, Δx, Δy, grid, arch)

For Chmy schemes, tangential arrays carried by PrescribedInflowBC are copied to the same backend as the phase fields during construction. It is therefore valid to supply ordinary host arrays when constructing a GPU-backed Chmy scheme; the scheme retains the backend copies used by its kernels.

KernelAbstractions CPU tests verify the backend-generic multiphase kernels. Actual GPU execution requires a device-enabled downstream test job.

Scalar transport on the GPU

Just like on the CPU, WENOScheme takes independent form and stag keywords, and both backends dispatch on the same form tag:

weno = WENOScheme(u, backend; form = :conservative, boundary = (PeriodicBC(), PeriodicBC()), stag = true)
velocity = (; x = ones(backend, nx + 1))
WENO_step!(u, velocity, weno, Δt, Δx, backend)

See Getting Started for a full worked example (the same call pattern applies; just construct arrays on the desired backend and pass it as an extra argument to WENOScheme and WENO_step!).