Design dpctl classes corresponding to SYCL kernel argument types

[diptorupd]

Currently to submit (SyclQueue.submit) a kernel all scalar arguments for that kernel should be valid ctypes objects (e.g. ctypes.c_int). Inside SyclQueue._populate_args we have a if-elif statements to first cast the address of the ctypes object to void* and also generate a DPPLKernelArgType enum value for the argument. The two things need to be done to call DPPLQueue_SubmitRange and DPPLQueue_SubmitNdRange

A better design will be if we have the equivalent of ctypes.c_xxx classes defined in dpctl, e.g. dpctl.c_xxx, and users cast the Python objects directly to the required dpctl type. The DPPLKernelArgType enum should then be a an attribute of the dpctl type that can be directly extracted without needed an extra switch. Moreover, users will want us to support other data types than just ctypes, e.g., numpy.dtypes.

The design is also needed if we want to support reduced precision inside dpctl.

• Design the dpctl.c_xxx data type classes.
• Change SyclQueue.submit to use dpctl_.c_xxx type instead of ctypes.c_xxx.

Related #72, #85

[diptorupd]

@oleksandr-pavlyk need your help with designing this, especially how we can extend to support reduced precision types.

[diptorupd]

@oleksandr-pavlyk adding you to help us design this feature.

[oleksandr-pavlyk]

Does OpenCL or Level-Zero allow for arguments to be structs passed by values?

[PokhodenkoSA]

Data types we have:

1. Numba - numba.core.types
2. Ctypes - ctypes
3. DPCTLKernelArgType - enum
4. SYCL types which partially are ctypes

Numba -> Ctypes
DPPY. numba_dppy/compiler.py:DPPYKernel._unpack_argument - kernelargs (ctypes with values)
Numba. numba/core/typing/ctypes_utils.py - Numba <-> Ctypes conversion in both ways

Ctypes -> DPCTLKernelArgType
dpCtl.Cyhton.

1. sycl_queue.submit -> _populate_args - Split to 2 arrays: values pointers and DPCTLKernelArgType types
2. sycl_queue.submit -> DPCTLQueue_SubmitRange

DPCTLKernelArgType -> SYCL types
dpCtl.backend. DPCTLQueue_SubmitRange -> set_kernel_arg - Dispatch for call handler cgh.set_arg(..., SYCL_TYPE)

[diptorupd]

The requested feature is no longer required in numba-dpex. Currently, in numba_dpex.experimental the Python interface of dpctl is no longer used to submit kernels. Instead we directly use libsyclinterface.