omlins/ParallelStencil.jl

What it solves

ParallelStencil.jl provides a way for domain scientists to write high-performance parallel stencil computations without needing to write low-level code for different hardware architectures. It bridges the gap between high-level Julia code and the performance of CUDA C or HIP, allowing researchers to deploy the same code across CPUs and various GPUs (Nvidia, AMD, Apple Metal) with minimal effort.

How it works

The library uses a set of macros, primarily @parallel, to parallelize and optimize functions. It leverages existing backends like CUDA.jl, AMDGPU.jl, Metal.jl, and KernelAbstractions.jl for GPUs, and Base.Threads and Polyester.jl for CPUs.

Key features include:

• Math-close notation: Specialized submodules (FiniteDifferences1D, 2D, 3D) provide macros like @inn and @d2_xi that allow users to write finite difference equations in a syntax that closely resembles mathematical notation.
• Memory Optimization: The memopt=true argument in @parallel enables automatic optimization of registers and shared memory.
• Hardware Agnosticism: Users can initialize the backend once and switch hardware targets at runtime using KernelAbstractions.jl without redefining kernels.
• Distributed Scaling: Through interoperability with ImplicitGlobalGrid.jl, it supports massively scalable applications across thousands of GPUs/CPUs, including the ability to hide communication overhead behind computation.
• Automatic Differentiation: It supports the automatic differentiation of parallel kernels via Enzyme.jl.

Who it’s for

It is designed for domain scientists and researchers performing high-performance numerical simulations (e.g., heat diffusion, poro-hydro-mechanical applications) who need GPU acceleration and multi-node scaling but prefer to avoid the complexity of writing architecture-specific kernel code.

Highlights

• High Performance: Can achieve up to 70% of theoretical GPU performance bounds, significantly outperforming standard GPU array programming.
• Cross-Platform: Single codebase for CPU, Nvidia GPU, AMD GPU, and Apple Metal.
• Massive Scalability: Supports weak scaling on thousands of GPUs when paired with ImplicitGlobalGrid.jl.
• Developer Friendly: Provides math-like syntax for stencil operations and runtime hardware selection for interactive prototyping.