Migrating CUDA.jl atomic operations to ParallelStencil.jl using Atomix.jl?

Specific Domains GPU
,
1

Hello,

I am currently porting my code from CUDA.jl to ParallelStencil.jl. While I have successfully managed the transition for mixed Float64 and ComplexF64 types, I have run into a challenge replacing CUDA-specific atomic operations, specifically:

CUDA.@atomic F[i, j, 1] += ...

Based on the KernelAbstractions.jl documentation, I am considering using Atomix.jl, but I am unclear on how the integration works in practice. Specifically:

  1. If I initialize my setup with @init_parallel_stencil(CUDA, Float64, 2, inbounds=true), will Atomix.jl automatically utilize the correct CUDA atomic instructions?
  2. Are there specific steps or wrappers required to ensure Atomix.jl works seamlessly within a ParallelStencil kernel?

I have looked through the Atomix.jl documentation, but it is quite sparse. Any guidance or examples would be greatly appreciated!

Best regards

2

ParallelStencil will remain unaware of atomic operations, but the mapping to the correct operations (when available) will happen in the backend of Atomix.

So essentially, there is Atomix.jl/ext/AtomixCUDAExt.jl at main · JuliaConcurrent/Atomix.jl · GitHub and UnsafeAtomics.jl together provide the infrastructure, and the use is transparent to KernelAbstractions and ParallelStencil.

2 Likes
3

Thank you for your help !