JustPIC.jlParticle-in-Cell advection on CPUs and GPUs

What is JustPIC.jl? ​

Particle-in-Cell (PIC) advection for large-scale, multi-XPU simulations. JustPIC runs the same code on CPU, CUDA, AMDGPU, and Metal through KernelAbstractions: the compute backend is chosen by the array type, not by rewriting the algorithm.

Installation ​

JustPIC is registered in the General registry:

using Pkg
Pkg.add("JustPIC")

To run on a GPU, additionally install the matching backend package (CUDA, AMDGPU, or Metal) and select its backend — see Backends.

Getting started ​

Choose a backend, build a staggered grid, and initialize a particle container:

julia> using JustPIC

julia> backend = JustPIC.CPU;

julia> nxcell, max_xcell, min_xcell = 6, 12, 3;

julia> n = 8;  # vertices per dimension

julia> xv = yv = LinRange(0.0, 1.0, n);

julia> dx = dy = xv[2] - xv[1];

julia> xc = yc = LinRange(dx / 2, 1 - dx / 2, n - 1);

julia> grid_vx = xv, LinRange(first(yc) - dy, last(yc) + dy, length(yc) + 2);

julia> grid_vy = LinRange(first(xc) - dx, last(xc) + dx, length(xc) + 2), yv;

julia> particles = init_particles(backend, nxcell, max_xcell, min_xcell, grid_vx, grid_vy);

julia> sum(particles.index.data)  # number of active markers across the grid
392

particles can then be advected through a velocity field and used for grid-to-particle and particle-to-grid interpolation. See Particles and Interpolations for the full workflow, or the runnable 2D field advection example.

Manual ​

• Particles and Particle lifecycle — the Particles container and the interpolate/advect/move/inject time step

• Grid layouts — vertex, centroid, and staggered velocity grids

• Backends — running on CPU, CUDA, AMDGPU, and Metal

• CellArrays — the cell-local storage layout

• Interpolations, PIC/FLIP, and Velocity interpolation

• Particle fields — per-particle data carried with the particles

• Marker chain — 2D free-surface / topography tracking

• Marker surface — 3D free-surface / topography tracking

• Examples: 2D, 2D with MPI, 3D, periodic, and PIC/FLIP advection

• I/O — checkpointing and restart

• Mixed CPU/GPU — particles on the CPU alongside a GPU solver

• Input validation and Support matrix

• Public API

• Performance — benchmark history on CPU and GPU backends

Funding ​

The development of this package is supported by the GPU4GEO PASC project.