Typing infrastructure improvements

[PokhodenkoSA]

The basic goals are:

1. Add support for all SYCL types that can be passed as kernel arguments (skip buffers/accessors for now):

• Design dpctl classes corresponding to SYCL kernel argument types dpctl#71 NumPy dtype and cython types to dpctl dtype
• Design support for reduced precision types in dpctl dpctl#72 reduced/half precision types (i.e. BF16)
• Add support for complex numbers as kernel arguments dpctl#85 Add support for complex numbers.
• Add support for SYCL's vector types #1288
• Array as type (w/o extraction)

2. Reduce the multiple places where we have checks to identify the kernel argument type:

• _unpack_argument(): Numba to ctypes: for DPPYKernel
  https://github.com/IntelPython/numba-dppy/blob/2d100c5524e62b1fbd56ca9a96b173a75cf45bec/numba_dppy/compiler.py#L456
• _populate_args(): ctypes to void** and DPCTLKernelArgType*: for submit
  https://github.com/IntelPython/dpctl/blob/609c8fd719ed61670127ad7c4e1dcbe783538ad6/dpctl/_sycl_queue.pyx#L116
• set_kernel_arg(): void** and DPCTLKernelArgType* to set_arg(): for SYCL
  https://github.com/IntelPython/dpctl/blob/0763b213e1c1b5508f36964be5083a1ebea7ccaf/dpctl-capi/source/dpctl_sycl_queue_interface.cpp#L52
• resolve_and_return_dpctl_type(): Numba type to DPCTLKernelArgType: for prange offload
  https://github.com/IntelPython/numba-dppy/blob/3e05b689f0eb7c67cb8d9b3fd0ec8bcfad316969/numba_dppy/dppy_host_fn_call_gen.py#L161

Tasks:

• Evaluate SYCL/DPC++ types (Sec 4.16 of SYCL 2020 Provisional Specification)
• Evaluate the implementation of Ctypes types library
• Evaluate the implementation for CUDA
  • https://github.com/IntelPython/numba/blob/5a1b372956ce881ebaf98c4ddd65681ece90ba0e/numba/cuda/compiler.py#L692
  • https://github.com/IntelPython/numba/blob/5a1b372956ce881ebaf98c4ddd65681ece90ba0e/numba/roc/compiler.py#L373
• Evaluate Numba - ctypes connection
  • https://github.com/IntelPython/numba/blob/5a1b372956ce881ebaf98c4ddd65681ece90ba0e/numba/core/typing/ctypes_utils.py#L62
• Provide a wrapper for SYCL/DPC++ types (Sec 4.16 of SYCL 2020 Provisional Specification) similar to what Ctypes does for C types

Design requirements:

1. Create a typing library inside dpctl similar to сtypes._SimpleCData or even deriving from it to have wrappers for all SYCL types including reduced precision.
2. Add new types to Numba as needed (e.g. for reduced precision types and short vector).
3. Have Numba-dppy specific type inference so that we can infer types as dpctl types.
4. Pass dpctl types to backend C API and do the type casting only once there.

Links:

1. SYCL 2020 Provisional Specification
2. https://github.com/IntelPython/dpctl/projects/1 - board

Problems:

1. How to support vector types? How to define vector types in Python? Vector - fixed size array.
2. DPC++ LLVM IR - For kernel getting vector. Is it possible to generate the same IR for vectors?

[diptorupd]

Reduced precision types are not planned at this point of time. I have opened a new ticket #1288 to track the discussion around SYCL vector type support in our front end API.

The rest of the task list around ctypes is now moot as the kernel is going to be switched to the Numba Dispatcher API and will use native Numba boxing and unboxing and the use of ctypes will be removed.