Fix typing issues in dppy_lowerer

[diptorupd]

Several places in dppy_lowerer.py we are using long to mean int64. As long is not guaranteed to be 64 bits this assumption can cause subtle issues on windows. Instead, we should be using long long, size_t, or unsigned long long that are guaranteed to be at least 64 bit wide.

The problem is this function: https://github.com/IntelPython/numba-dppy/blob/0105e830a42829ea7206cf4d87d9f9d9246253f2/numba_dppy/dppy_host_fn_call_gen.py#L115

dppy_lowerer.py defines kernel arguments based on the enum values defined in :
https://github.com/IntelPython/dpctl/blob/8db2b2c389dab0c2e6792c7a10b8d5fbec9125df/dpctl-capi/include/dpctl_sycl_enum_types.h#L71. As can be seen from the enum, "7" points to long, "9" points to long long and "11" points to size_t.

• Please add documentation to the resolve_and_return_dpctl_type function explaining how it is looking up the dpctl defined enum values and what the hard-coded integer values mean.
• Change this line to use DPCTL_LONG_LONG (9) https://github.com/IntelPython/numba-dppy/blob/0105e830a42829ea7206cf4d87d9f9d9246253f2/numba_dppy/dppy_host_fn_call_gen.py#L123
• We are passing all literal integer kernel arguments as 64 bit values and not as int. Therefore, this line should also be changed to use DPCTL_LONG_LONG (9)
  https://github.com/IntelPython/numba-dppy/blob/0105e830a42829ea7206cf4d87d9f9d9246253f2/numba_dppy/dppy_host_fn_call_gen.py#L117
• Similarly, uint32 should also get passed as DPCTL_UNSIGNED_LONG_LONG(10) or as DPCTL_SIZE_T (11)
  https://github.com/IntelPython/numba-dppy/blob/0105e830a42829ea7206cf4d87d9f9d9246253f2/numba_dppy/dppy_host_fn_call_gen.py#L120
• It should be DPCTL_SIZE_T (11) here:
  https://github.com/IntelPython/numba-dppy/blob/0105e830a42829ea7206cf4d87d9f9d9246253f2/numba_dppy/dppy_host_fn_call_gen.py#L125
• The name of the variable is confusing as long is not guaranteed to be 8 bytes. It should be renamed to int64_t
  https://github.com/IntelPython/numba-dppy/blob/0105e830a42829ea7206cf4d87d9f9d9246253f2/numba_dppy/dppy_host_fn_call_gen.py#L55

I have not tested #46 on Windows, but I found at least a semi-reproducer on Linux.

from numba import njit
import dpctl
import numba_dppy
import numpy as np

@njit
def func(b):
    a = np.ones((64), dtype=np.float64)
    np.sin(a, b)

numba_dppy.compiler.DEBUG = 1
expected = np.ones((64), dtype=np.float64)
got_cpu = np.ones((64), dtype=np.float64)

with dpctl.device_context("opencl:cpu"):
    func(got_cpu)

numba_dppy.compiler.DEBUG = 0
func(expected)

print(got_cpu)
print(expected)

Without the changes I get Native API failed. Native API returns: -51 (CL_INVALID_ARG_SIZE) -51 (CL_INVALID_ARG_SIZE) due to int32 being passed as 32 bit integer. With the above changes, I do not get the error.

We should reevaluate if 32 bit integer values should be upcasted to 64 bit always, but that is a separate discussion.

PS: I still get OMP: Info #275: omp_set_nested routine deprecated, please use omp_set_max_active_levels instead. not sure why. But, that is unrelated to this ticket.

[PokhodenkoSA]

Fixed in #46.