GPU

FourierFlows allows you to easily construct and run problems on GPUs.

Upon calling

using FourierFlows

FourierFlows.jl will check whether any CUDA enabled device is present. If such a device is found then FourierFlows.jl makes sure that CUDA related packages are loaded and also it will overload all methods to work with GPU() device as their argument (instead of the standard CPU() device).

It's easy to construct a grid that lives on the GPU. Calling:

dev = GPU()
n, L = 16, 2.0
grid = OneDGrid(dev, n, L)

OneDimensionalGrid
  ├─────────── Device: GPU
  ├──────── FloatType: Float64
  ├────────── size Lx: 2.0
  ├──── resolution nx: 16
  ├── grid spacing dx: 0.125
  ├─────────── domain: x ∈ [-1.0, 0.875]
  └─ aliased fraction: 0.3333333333333333

gives out a grid whose arrays are CuArrays. (Calling OneDGrid(n, L) defaults to CPU, i.e., OneDGrid(CPU(), n, L).)

When we construct the Params, Vars, and Equation for our problem we need to make sure that we create arrays on the appropriate device, i.e., Arrays for CPU or CuArrays for the GPU. Function ArrayType is useful in constructing appropriately chosen arrays.

ArrayType(::Device)
ArrayType(::Device, T, dim)

The FourierFlows.Problem constructor then takes an optional positional argument dev::Device. If not provided anything, the default values for dev=CPU().

problem = Problem(equation, stepper, dt, grid, vars, params, GPU())

The FourierFlows.Diffusion module is written in a way such that switching from CPU to GPU is only a matter of calling FourierFlows.Diffusion.Problem() with dev=GPU(). All physics modules in GeophysicalFlows.jl can also seamlessly run on a GPU with dev=GPU() argument.

Selecting GPU device

FourierFlows.jl can only utilize a single GPU. If your machine has more than one GPU available, then functionality within CUDA.jl package enables the user to choose the GPU device that FourierFlows.jl should use. The user is referred to the CUDA.jl Documentation; in particular, CUDA.devices and CUDA.CuDevice.