[chraac]

Refactors the OpenCL kernel argument initialization process by introducing a template helper, significantly reducing boilerplate code and improving readability.

Key Changes

• Introduced cl_set_kernel_args: A variadic template function that automatically iterates through arguments and sets them using clSetKernelArg.
• Type-Specific Setters: Added cl_kernel_arg_setter struct specializations to handle different types:
  • ggml_tensor: Automatically extracts the OpenCL buffer (data_device) and calculates the offset (offset + view_offs), setting both as kernel arguments.
  • Arrays (ne, nb): Automatically unrolls int64_t[4] and size_t[4] arrays into individual scalar arguments (e.g., ne0, ne1, ne2, ne3).
  • Primitives: Direct handling for int, float, cl_ulong.

[chraac]

It might be worth considering a shared header for kernel signatures to prevent mismatches between the host and device.
Knowing the refactoring effort would be high, but it would guarantee consistency.

[lhez]

Currently we use a simple script to read a kernel file and make it into a C++ string in a header file to embed the kernels into executable. Using headers in kernel files will greatly increase the complexity of embedding kernels.

[chraac]

Currently we use a simple script to read a kernel file and make it into a C++ string in a header file to embed the kernels into executable.

Yeah, I saw the embed_kernel.py script that converts the CL source into string literals.
If we introduce a new header here, I think we could extend that script to handle #include directives by simply replacing them with the content of the header file from the same directory.

[chraac]

Heres a safer version of this function that'll check if the parameters match the cl kernel func signiture

[chraac]

@lhez
Also have a thought regarding load_cl_kernels: Consider refactoring this to use a data structure (like an attribute table) to store the program sources and kernel name.
This approach would significantly reduce code duplication and improve maintainability as new kernels are added.