Expose NVVM version detection helper as public API

[leofang]

This

cuda-python/cuda_core/tests/test_program.py

Lines 58 to 132 in da7eb1f

	_libnvvm_version = None
	_libnvvm_version_attempted = False
	precheck_nvvm_ir = """target triple = "nvptx64-unknown-cuda"
	target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
	define void @dummy_kernel() {{
	entry:
	ret void
	}}
	!nvvm.annotations = !{{!0}}
	!0 = !{{void ()* @dummy_kernel, !"kernel", i32 1}}
	!nvvmir.version = !{{!1}}
	!1 = !{{i32 {major}, i32 {minor}, i32 {debug_major}, i32 {debug_minor}}}
	""" # noqa: E501
	def _get_libnvvm_version_for_tests():
	"""
	Detect libNVVM version by compiling dummy IR and analyzing the PTX output.
	Workaround for the lack of direct libNVVM version API (nvbugs 5312315).
	The approach:
	- Compile a small dummy NVVM IR to PTX
	- Use PTX version analysis APIs if available to infer libNVVM version
	- Cache the result for future use
	"""
	global _libnvvm_version, _libnvvm_version_attempted
	if _libnvvm_version_attempted:
	return _libnvvm_version
	_libnvvm_version_attempted = True
	try:
	from cuda.core._program import _get_nvvm_module
	nvvm = _get_nvvm_module()
	try:
	from cuda.bindings.utils import get_minimal_required_cuda_ver_from_ptx_ver, get_ptx_ver
	except ImportError:
	_libnvvm_version = None
	return _libnvvm_version
	program = nvvm.create_program()
	try:
	major, minor, debug_major, debug_minor = nvvm.ir_version()
	global precheck_nvvm_ir
	precheck_nvvm_ir = precheck_nvvm_ir.format(
	major=major, minor=minor, debug_major=debug_major, debug_minor=debug_minor
	)
	precheck_ir_bytes = precheck_nvvm_ir.encode("utf-8")
	nvvm.add_module_to_program(program, precheck_ir_bytes, len(precheck_ir_bytes), "precheck.ll")
	options = ["-arch=compute_90"]
	nvvm.verify_program(program, len(options), options)
	nvvm.compile_program(program, len(options), options)
	ptx_size = nvvm.get_compiled_result_size(program)
	ptx_data = bytearray(ptx_size)
	nvvm.get_compiled_result(program, ptx_data)
	ptx_str = ptx_data.decode("utf-8")
	ptx_version = get_ptx_ver(ptx_str)
	cuda_version = get_minimal_required_cuda_ver_from_ptx_ver(ptx_version)
	_libnvvm_version = cuda_version
	return _libnvvm_version
	finally:
	nvvm.destroy_program(program)
	except Exception:
	_libnvvm_version = None
	return _libnvvm_version

has a real need (ex: NVIDIA/numba-cuda#681). We should expose it to cuda.bindings.utils, similar to the PTX version helpers.