Standard OpenCL wrapper

[vladfaust]

Imagine how https://www.eriksmistad.no/getting-started-with-opencl-and-gpu-computing could look like in Onyx:

import { Program, Kernel, Global } from "std/opencl"

extern @[Kernel] void vector_add(
  @[Global] const int *A, 
  @[Global] const int *B, 
  @[Global] int *C
) {
  // Get the index of the current element to be processed
  int i = get_global_id(0);

  // Do the operation
  C[i] = A[i] + B[i];
}

final program = Program()
program.run<$vector_add>(a, b, c)

Many things are omitted for brevity. However, it can be seen that Onyx annotations may be applied in C context as well, even without the need to use $ (@ is unused in C; it also allows to use Onyx intrinsics within a C context).

Passing a C function prototype as a template argument allows to lens through the prototype metadata during compilation, including annotations.
This can be also done via macros, I suppose, but a compiler is likely to implement kernel compilation natively.

Alternatively, a compiler may allow to load a pre-compiled code like program.load_kernel("mykernel.spirv").
This wouldn't allow type checking, though.

That said, OpenCL binding looks like a big chunk of problems for a compiler to stay compliant. Therefore, I think that a portion of standard library shall be considered optional. Also, OpenCL versioning...