MemoryAccess package for abstract optimized memory operations

Package Announcements
1

I’ve developed MemoryAccess.jl, a package providing comprehensive abstractions for GPU memory access patterns:

  • Vectorized memory loads and stores using LLVMPtr
  • Warp-level shuffle operations for arbitrary data structures, leveraging CUDA backend semantics
  • Memory fences with configurable scopes and memory ordering, built on UnsafeAtomics semantics

Registration PR

The initial release targets the CUDA backend, though the abstraction layer is designed to accommodate other GPU backends (ROCm, Metal, oneAPI) in future versions.

I believe MemoryAccess would complement the existing JuliaGPU ecosystem well. Would the maintainers be open to transferring the repository to the JuliaGPU organization?

2 Likes