This directory contains scripts that can be used to automatically generate wrappers for C headers by NVIDIA, such as CUBLAS or CUDNN. This is done using Clang.jl.

In CUDA.jl, the wrappers need to know whether pointers passed into the library point to CPU or GPU memory (i.e. Ptr or CuPtr). This information is not available from the headers, and instead should be provided by the developer. The specific information is embedded in the TOML files for each component.

Either run wrap.jl directly, or include it using Revise.jl and call the main() function. Be sure to activate the project environment in this folder, which will download CUDA from artifacts (if you want to upgrade the headers, be sure to update the relevant JLLs in the project environment). You can also call main(library) if you want to generate the wrapper for a specific library. The possible values for library are "all" (default), "cuda", "nvml", "cupti", "cublas", "cufft", "curand", "cusparse", "cusolver", "cudnn", "cutensor", "cutensornet" or "custatevec".

For each library, the script performs the following steps:

• generate wrappers with Clang.jl
• rewrite the headers: wrap functions that result status codes with check, add calls to API initializers, etc.

You should always review any changes to the headers! Specifically, verify that pointer arguments are of the correct type, and if they aren't, modify the TOML file and regenerate the wrappers. The following types should be considered as alternatives to plain Ptrs:

• CuPtr: if this pointer is a device pointer
• PtrOrCuPtr: if this pointer can be either a device or host pointer
• Ref: if the pointer represents a scalar or single-value argument on the host
• CuRef: idem, but on the device
• RefOrCuRef: idem, but either on the host or device