Backends

JustPIC uses KernelAbstractions backends directly. The CPU path is JustPIC.CPU; CUDA, AMDGPU, and Metal backends are activated by loading their respective packages:

Hardware	Package	Backend
CPU	—	`JustPIC.CPU`
Nvidia GPU	`CUDA`	`CUDA.CUDABackend`
AMD GPU	`AMDGPU`	`AMDGPU.ROCBackend`
Apple GPU	`Metal`	`Metal.MetalBackend`

Allocate backend arrays with TA(backend) and pass the same backend to init_particles:

julia

using JustPIC
using CUDA

backend = CUDA.CUDABackend   # or JustPIC.CPU, AMDGPU.ROCBackend, Metal.MetalBackend
T = TA(backend)(zeros(Float32, 32, 32))
particles = init_particles(backend, 16, 24, 8, grid_vx, grid_vy)

The high-level particle, interpolation, and advection routines then dispatch on the arrays stored in particles:

julia

V = TA(backend).(V_cpu)
advection!(particles, RungeKutta2(), V, dt)

Allocating every array through TA(backend) keeps a script backend-generic: switching hardware means changing the backend line only. Use Float32 grids and fields when targeting Metal, which does not support Float64.

For mixed CPU/GPU simulations, keep the JustPIC backend on the CPU and copy only the fields needed by GPU calculations. See Mixed CPU and GPU.

Backends ​

Backends