Pyccel CUDA Project Roadmap

[bauom]

This roadmap outlines the planned development phases for the Pyccel CUDA project.
Our goal is to enhance Pyccel's capabilities with CUDA support, enabling efficient GPU acceleration for Python code.

The Aim for the Core Development is to have a cuda module inside pyccel that can be only used with pyccel that we can later use it to add any libraries (numba, cupy...)

• Core Development
  • Integrate NVCC for CUDA Compilation in Pyccel #26
  • Add synchronize support in host code #45
  • Basic support for Custom Kernels. #28
  • thread indexing. #43
  • need cuda.syncthreads() #3
  • Basic support for Custom Device function. #41
  • Array Support / Routines. #29
    • Pyccel cuda_ndarrays C stdlib #39
    • Cuda Pyccel Internal library (Host memory) #34
    • Cuda Pyccel Internal library (Device memory) #35
    • Cuda Pyccel Internal library (Unified Memory) #40
    • Cuda Pyccel Internal library (Memory transfer) #36
    • Cuda Pyccel Internal library -- arrays creation routines
      • cuda.full, cuda.ones, cuda.zeros ... #37
      • cuda.arange #38
  • Basic Shared memory. #30
  • add random Cuda API
  • Develop CUDA kernels for key operations. #31
• User interface
• Automatic array transfer when passing CPU arrays to kernels. (user only pyccelizes kernels)
• Auto Optimisation
• The use for tensor ops
  • fusemuladd