Fix #789: Remove cyclical import between driver and _lib.utils

.gitignore

@@ -16,6 +16,7 @@ __pycache__/
 cache_driver
 cache_runtime
 cache_nvrtc
+cuda_bindings/cuda/bindings/_lib/utils.pxi
 
 # CUDA Python specific (auto-generated)
 cuda_bindings/cuda/bindings/_bindings/cydriver.pxd

cuda_bindings/cuda/bindings/_lib/param_packer.cpp

@@ -2,45 +2,44 @@
 // SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
 #include <Python.h>
-#include "param_packer.h"
 
 #include <map>
 #include <functional>
 #include <stdexcept>
 #include <string>
 
-PyObject* enum_module = nullptr;
-PyTypeObject* enum_Enum = nullptr;
+static PyObject* enum_module = nullptr;
+static PyTypeObject* enum_Enum = nullptr;
 
-PyObject* ctypes_module = nullptr;
-PyObject* ctypes_addressof = nullptr;
-PyObject* addressof_param_tuple = nullptr;
+static PyObject* ctypes_module = nullptr;
+static PyObject* ctypes_addressof = nullptr;
+static PyObject* addressof_param_tuple = nullptr;
 
-PyTypeObject* ctypes_c_char = nullptr;
-PyTypeObject* ctypes_c_bool = nullptr;
-PyTypeObject* ctypes_c_wchar = nullptr;
-PyTypeObject* ctypes_c_byte = nullptr;
-PyTypeObject* ctypes_c_ubyte = nullptr;
-PyTypeObject* ctypes_c_short = nullptr;
-PyTypeObject* ctypes_c_ushort = nullptr;
-PyTypeObject* ctypes_c_int = nullptr;
-PyTypeObject* ctypes_c_uint = nullptr;
-PyTypeObject* ctypes_c_long = nullptr;
-PyTypeObject* ctypes_c_ulong = nullptr;
-PyTypeObject* ctypes_c_longlong = nullptr;
-PyTypeObject* ctypes_c_ulonglong = nullptr;
-PyTypeObject* ctypes_c_size_t = nullptr;
-PyTypeObject* ctypes_c_float = nullptr;
-PyTypeObject* ctypes_c_double = nullptr;
-PyTypeObject* ctypes_c_void_p = nullptr;
+static PyTypeObject* ctypes_c_char = nullptr;
+static PyTypeObject* ctypes_c_bool = nullptr;
+static PyTypeObject* ctypes_c_wchar = nullptr;
+static PyTypeObject* ctypes_c_byte = nullptr;
+static PyTypeObject* ctypes_c_ubyte = nullptr;
+static PyTypeObject* ctypes_c_short = nullptr;
+static PyTypeObject* ctypes_c_ushort = nullptr;
+static PyTypeObject* ctypes_c_int = nullptr;
+static PyTypeObject* ctypes_c_uint = nullptr;
+static PyTypeObject* ctypes_c_long = nullptr;
+static PyTypeObject* ctypes_c_ulong = nullptr;
+static PyTypeObject* ctypes_c_longlong = nullptr;
+static PyTypeObject* ctypes_c_ulonglong = nullptr;
+static PyTypeObject* ctypes_c_size_t = nullptr;
+static PyTypeObject* ctypes_c_float = nullptr;
+static PyTypeObject* ctypes_c_double = nullptr;
+static PyTypeObject* ctypes_c_void_p = nullptr;
 
-PyTypeObject* ctypes_c_ssize_t = nullptr;
-PyTypeObject* ctypes_c_longdouble = nullptr;
-PyTypeObject* ctypes_c_char_p = nullptr;
-PyTypeObject* ctypes_c_wchar_p = nullptr;
-PyTypeObject* ctypes_c_structure = nullptr;
+static PyTypeObject* ctypes_c_ssize_t = nullptr;
+static PyTypeObject* ctypes_c_longdouble = nullptr;
+static PyTypeObject* ctypes_c_char_p = nullptr;
+static PyTypeObject* ctypes_c_wchar_p = nullptr;
+static PyTypeObject* ctypes_c_structure = nullptr;
 
-void fetch_ctypes()
+static void fetch_ctypes()
 {
     ctypes_module = PyImport_ImportModule("ctypes");
     if (ctypes_module == nullptr)
@@ -72,9 +71,9 @@ void fetch_ctypes()
 
 
 // (target type, source type)
-std::map<std::pair<PyTypeObject*,PyTypeObject*>, std::function<int(void*, PyObject*)>> m_feeders;
+static std::map<std::pair<PyTypeObject*,PyTypeObject*>, std::function<int(void*, PyObject*)>> m_feeders;
 
-void populate_feeders(PyTypeObject* target_t, PyTypeObject* source_t)
+static void populate_feeders(PyTypeObject* target_t, PyTypeObject* source_t)
 {
     if (target_t == ctypes_c_int)
     {
@@ -140,7 +139,7 @@ void populate_feeders(PyTypeObject* target_t, PyTypeObject* source_t)
     }
 }
 
-int feed(void* ptr, PyObject* value, PyObject* type)
+static int feed(void* ptr, PyObject* value, PyObject* type)
 {
     PyTypeObject* pto = (PyTypeObject*)type;
     if (ctypes_c_int == nullptr)

cuda_bindings/cuda/bindings/_lib/param_packer.pxd

@@ -1,5 +1,7 @@
 # SPDX-FileCopyrightText: Copyright (c) 2021-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
-cdef extern from "param_packer.h":
+# Include "param_packer.cpp" so its contents get compiled into every
+# Cython extension module that depends on param_packer.pxd.
+cdef extern from "param_packer.cpp":
     int feed(void* ptr, object o, object ct)

cuda_bindings/cuda/bindings/_lib/utils.pxd.in

@@ -6,21 +6,21 @@ cimport cuda.bindings.cydriver as cydriver
 cimport cuda.bindings.cyruntime as cyruntime
 from libcpp.vector cimport vector
 
-cdef class HelperKernelParams:
+cdef class _HelperKernelParams:
     cdef Py_buffer _pybuffer
     cdef bint _pyobj_acquired
     cdef void** _ckernelParams
     cdef char* _ckernelParamsData
     cdef int _length
     cdef bint _malloc_list_created
 
-cdef class HelperInputVoidPtr:
+cdef class _HelperInputVoidPtr:
     cdef Py_buffer _pybuffer
     cdef void* _cptr
     cdef bint _pyobj_acquired
 {{if 'CUmemPool_attribute_enum' in found_types}}
 
-cdef class HelperCUmemPool_attribute:
+cdef class _HelperCUmemPool_attribute:
     cdef void* _cptr
     cdef cydriver.CUmemPool_attribute_enum _attr
     cdef bint _is_getter
@@ -31,7 +31,7 @@ cdef class HelperCUmemPool_attribute:
 {{endif}}
 {{if 'CUmem_range_attribute_enum' in found_types}}
 
-cdef class HelperCUmem_range_attribute:
+cdef class _HelperCUmem_range_attribute:
     cdef void* _cptr
     cdef cydriver.CUmem_range_attribute_enum _attr
     cdef size_t _data_size
@@ -42,7 +42,7 @@ cdef class HelperCUmem_range_attribute:
 {{endif}}
 {{if 'CUpointer_attribute_enum' in found_types}}
 
-cdef class HelperCUpointer_attribute:
+cdef class _HelperCUpointer_attribute:
     cdef void* _cptr
     cdef cydriver.CUpointer_attribute_enum _attr
     cdef bint _is_getter
@@ -60,7 +60,7 @@ cdef class HelperCUpointer_attribute:
 {{endif}}
 {{if 'CUgraphMem_attribute_enum' in found_types}}
 
-cdef class HelperCUgraphMem_attribute:
+cdef class _HelperCUgraphMem_attribute:
     cdef void* _cptr
     cdef cydriver.CUgraphMem_attribute_enum _attr
     cdef bint _is_getter
@@ -70,7 +70,7 @@ cdef class HelperCUgraphMem_attribute:
 {{endif}}
 {{if 'CUjit_option_enum' in found_types}}
 
-cdef class HelperCUjit_option:
+cdef class _HelperCUjit_option:
     cdef void* _cptr
     cdef cydriver.CUjit_option_enum _attr
 
@@ -83,11 +83,11 @@ cdef class HelperCUjit_option:
     cdef int _int
     cdef cydriver.CUjit_cacheMode_enum _cacheMode
     cdef vector[char*] _charstarstar # list of names
-    cdef InputVoidPtrPtrHelper _voidstarstar # list of addresses
+    cdef _InputVoidPtrPtrHelper _voidstarstar # list of addresses
 {{endif}}
 {{if 'cudaJitOption' in found_types}}
 
-cdef class HelperCudaJitOption:
+cdef class _HelperCudaJitOption:
     cdef void* _cptr
     cdef cyruntime.cudaJitOption _attr
 
@@ -101,7 +101,7 @@ cdef class HelperCudaJitOption:
 {{endif}}
 {{if 'CUlibraryOption_enum' in found_types}}
 
-cdef class HelperCUlibraryOption:
+cdef class _HelperCUlibraryOption:
     cdef void* _cptr
     cdef cydriver.CUlibraryOption_enum _attr
 
@@ -110,7 +110,7 @@ cdef class HelperCUlibraryOption:
 {{endif}}
 {{if 'cudaLibraryOption' in found_types}}
 
-cdef class HelperCudaLibraryOption:
+cdef class _HelperCudaLibraryOption:
     cdef void* _cptr
     cdef cyruntime.cudaLibraryOption _attr
 
@@ -119,7 +119,7 @@ cdef class HelperCudaLibraryOption:
 {{endif}}
 {{if 'CUmemAllocationHandleType_enum' in found_types}}
 
-cdef class HelperCUmemAllocationHandleType:
+cdef class _HelperCUmemAllocationHandleType:
     cdef void* _cptr
     cdef cydriver.CUmemAllocationHandleType_enum _type
 
@@ -132,12 +132,12 @@ cdef class HelperCUmemAllocationHandleType:
     {{endif}}
 {{endif}}
 
-cdef class InputVoidPtrPtrHelper:
+cdef class _InputVoidPtrPtrHelper:
     cdef void** _cptr
 
 {{if 'CUcoredumpSettings_enum' in found_types}}
 
-cdef class HelperCUcoredumpSettings:
+cdef class _HelperCUcoredumpSettings:
     cdef void* _cptr
     cdef cydriver.CUcoredumpSettings_enum _attrib
     cdef bint _is_getter