CUDA project layout

Project with CUDA implementation + Python bindings.

include/
    myproject/
        common.h # common functions for CUDA side
        utils.cuh # CUDA header with source
        main.h # CPP side API, must not contain/include any CUDA code.

src/
    main.cu # implementation for main.h, may contain/include CUDA code. 
    bindings.cpp # pybind code

myproject/
    __init__.py
    main.py # python side API (call the C impl)

setup.py # build C side and install python side

Implement Function in CPP

This is much easier.

We can handle memory allocation in python side, and only pass data pointer to C side.

• define the function API in python side func.py, which imports and calls backend functions.
• write a func.h to define the backend functions.
• implement backend functions in func.cu
• write a bindings.cpp to expose the backend.
• write setup.py to build everything.

Implement Class in CPP

We need to let C side hold data structures and maybe allocate memory! This becomes harder...

For example, we'll need to convert data pointer to custom struct or class.

Header Source Separation

Example