Improve #789: Remove cyclical dependency between cuda.bindings.{driver|runtime} and c.b.utils

cuda_bindings/cuda/bindings/driver.pyx.in

@@ -14,6 +14,7 @@ from libc.limits cimport CHAR_MIN
 from libcpp.vector cimport vector
 from cpython.buffer cimport PyObject_CheckBuffer, PyObject_GetBuffer, PyBuffer_Release, PyBUF_SIMPLE, PyBUF_ANY_CONTIGUOUS
 from cpython.bytes cimport PyBytes_FromStringAndSize
+from cuda.bindings import utils
 import cuda.bindings.driver
 from libcpp.map cimport map
 
@@ -53948,3 +53949,124 @@ def sizeof(objType):
     if objType == VdpOutputSurface:
         return sizeof(cydriver.VdpOutputSurface){{endif}}
     raise TypeError("Unknown type: " + str(objType))
+
+def _add_native_handle_getters() -> None:
+    _add_cuda_native_handle_getter = utils._add_cuda_native_handle_getter
+    {{if 'CUcontext' in found_types}}
+    def CUcontext_getter(CUcontext x): return <uintptr_t><void*><cydriver.CUcontext>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUcontext, CUcontext_getter)
+    {{endif}}
+    {{if 'CUmodule' in found_types}}
+    def CUmodule_getter(CUmodule x): return <uintptr_t><void*><cydriver.CUmodule>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUmodule, CUmodule_getter)
+    {{endif}}
+    {{if 'CUfunction' in found_types}}
+    def CUfunction_getter(CUfunction x): return <uintptr_t><void*><cydriver.CUfunction>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUfunction, CUfunction_getter)
+    {{endif}}
+    {{if 'CUlibrary' in found_types}}
+    def CUlibrary_getter(CUlibrary x): return <uintptr_t><void*><cydriver.CUlibrary>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUlibrary, CUlibrary_getter)
+    {{endif}}
+    {{if 'CUkernel' in found_types}}
+    def CUkernel_getter(CUkernel x): return <uintptr_t><void*><cydriver.CUkernel>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUkernel, CUkernel_getter)
+    {{endif}}
+    {{if 'CUarray' in found_types}}
+    def CUarray_getter(CUarray x): return <uintptr_t><void*><cydriver.CUarray>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUarray, CUarray_getter)
+    {{endif}}
+    {{if 'CUmipmappedArray' in found_types}}
+    def CUmipmappedArray_getter(CUmipmappedArray x): return <uintptr_t><void*><cydriver.CUmipmappedArray>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUmipmappedArray, CUmipmappedArray_getter)
+    {{endif}}
+    {{if 'CUtexref' in found_types}}
+    def CUtexref_getter(CUtexref x): return <uintptr_t><void*><cydriver.CUtexref>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUtexref, CUtexref_getter)
+    {{endif}}
+    {{if 'CUsurfref' in found_types}}
+    def CUsurfref_getter(CUsurfref x): return <uintptr_t><void*><cydriver.CUsurfref>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUsurfref, CUsurfref_getter)
+    {{endif}}
+    {{if 'CUevent' in found_types}}
+    def CUevent_getter(CUevent x): return <uintptr_t><void*><cydriver.CUevent>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUevent, CUevent_getter)
+    {{endif}}
+    {{if 'CUstream' in found_types}}
+    def CUstream_getter(CUstream x): return <uintptr_t><void*><cydriver.CUstream>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUstream, CUstream_getter)
+    {{endif}}
+    {{if 'CUgraphicsResource' in found_types}}
+    def CUgraphicsResource_getter(CUgraphicsResource x): return <uintptr_t><void*><cydriver.CUgraphicsResource>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUgraphicsResource, CUgraphicsResource_getter)
+    {{endif}}
+    {{if 'CUexternalMemory' in found_types}}
+    def CUexternalMemory_getter(CUexternalMemory x): return <uintptr_t><void*><cydriver.CUexternalMemory>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUexternalMemory, CUexternalMemory_getter)
+    {{endif}}
+    {{if 'CUexternalSemaphore' in found_types}}
+    def CUexternalSemaphore_getter(CUexternalSemaphore x): return <uintptr_t><void*><cydriver.CUexternalSemaphore>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUexternalSemaphore, CUexternalSemaphore_getter)
+    {{endif}}
+    {{if 'CUgraph' in found_types}}
+    def CUgraph_getter(CUgraph x): return <uintptr_t><void*><cydriver.CUgraph>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUgraph, CUgraph_getter)
+    {{endif}}
+    {{if 'CUgraphNode' in found_types}}
+    def CUgraphNode_getter(CUgraphNode x): return <uintptr_t><void*><cydriver.CUgraphNode>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUgraphNode, CUgraphNode_getter)
+    {{endif}}
+    {{if 'CUgraphExec' in found_types}}
+    def CUgraphExec_getter(CUgraphExec x): return <uintptr_t><void*><cydriver.CUgraphExec>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUgraphExec, CUgraphExec_getter)
+    {{endif}}
+    {{if 'CUmemoryPool' in found_types}}
+    def CUmemoryPool_getter(CUmemoryPool x): return <uintptr_t><void*><cydriver.CUmemoryPool>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUmemoryPool, CUmemoryPool_getter)
+    {{endif}}
+    {{if 'CUuserObject' in found_types}}
+    def CUuserObject_getter(CUuserObject x): return <uintptr_t><void*><cydriver.CUuserObject>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUuserObject, CUuserObject_getter)
+    {{endif}}
+    {{if 'CUgraphDeviceNode' in found_types}}
+    def CUgraphDeviceNode_getter(CUgraphDeviceNode x): return <uintptr_t><void*><cydriver.CUgraphDeviceNode>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUgraphDeviceNode, CUgraphDeviceNode_getter)
+    {{endif}}
+    {{if 'CUasyncCallbackHandle' in found_types}}
+    def CUasyncCallbackHandle_getter(CUasyncCallbackHandle x): return <uintptr_t><void*><cydriver.CUasyncCallbackHandle>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUasyncCallbackHandle, CUasyncCallbackHandle_getter)
+    {{endif}}
+    {{if 'CUgreenCtx' in found_types}}
+    def CUgreenCtx_getter(CUgreenCtx x): return <uintptr_t><void*><cydriver.CUgreenCtx>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUgreenCtx, CUgreenCtx_getter)
+    {{endif}}
+    {{if 'CUlinkState' in found_types}}
+    def CUlinkState_getter(CUlinkState x): return <uintptr_t><void*><cydriver.CUlinkState>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUlinkState, CUlinkState_getter)
+    {{endif}}
+    {{if 'CUdevResourceDesc' in found_types}}
+    def CUdevResourceDesc_getter(CUdevResourceDesc x): return <uintptr_t><void*><cydriver.CUdevResourceDesc>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUdevResourceDesc, CUdevResourceDesc_getter)
+    {{endif}}
+    {{if 'CUlogsCallbackHandle' in found_types}}
+    def CUlogsCallbackHandle_getter(CUlogsCallbackHandle x): return <uintptr_t><void*><cydriver.CUlogsCallbackHandle>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUlogsCallbackHandle, CUlogsCallbackHandle_getter)
+    {{endif}}
+    {{if True}}
+    def CUeglStreamConnection_getter(CUeglStreamConnection x): return <uintptr_t><void*><cydriver.CUeglStreamConnection>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(CUeglStreamConnection, CUeglStreamConnection_getter)
+    {{endif}}
+    {{if True}}
+    def EGLImageKHR_getter(EGLImageKHR x): return <uintptr_t><void*><cydriver.EGLImageKHR>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(EGLImageKHR, EGLImageKHR_getter)
+    {{endif}}
+    {{if True}}
+    def EGLStreamKHR_getter(EGLStreamKHR x): return <uintptr_t><void*><cydriver.EGLStreamKHR>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(EGLStreamKHR, EGLStreamKHR_getter)
+    {{endif}}
+    {{if True}}
+    def EGLSyncKHR_getter(EGLSyncKHR x): return <uintptr_t><void*><cydriver.EGLSyncKHR>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(EGLSyncKHR, EGLSyncKHR_getter)
+    {{endif}}
+_add_native_handle_getters()
+

cuda_bindings/cuda/bindings/nvrtc.pyx.in

@@ -14,6 +14,7 @@ from libc.limits cimport CHAR_MIN
 from libcpp.vector cimport vector
 from cpython.buffer cimport PyObject_CheckBuffer, PyObject_GetBuffer, PyBuffer_Release, PyBUF_SIMPLE, PyBUF_ANY_CONTIGUOUS
 from cpython.bytes cimport PyBytes_FromStringAndSize
+from cuda.bindings import utils
 
 ctypedef unsigned long long signed_char_ptr
 ctypedef unsigned long long unsigned_char_ptr

cuda_bindings/cuda/bindings/runtime.pyx.in

@@ -14,6 +14,7 @@ from libc.limits cimport CHAR_MIN
 from libcpp.vector cimport vector
 from cpython.buffer cimport PyObject_CheckBuffer, PyObject_GetBuffer, PyBuffer_Release, PyBUF_SIMPLE, PyBUF_ANY_CONTIGUOUS
 from cpython.bytes cimport PyBytes_FromStringAndSize
+from cuda.bindings import utils
 import cuda.bindings.driver
 from libcpp.map cimport map
 
@@ -37912,3 +37913,104 @@ def sizeof(objType):
     if objType == cudaEglStreamConnection:
         return sizeof(cyruntime.cudaEglStreamConnection){{endif}}
     raise TypeError("Unknown type: " + str(objType))
+
+def _add_native_handle_getters() -> None:
+    _add_cuda_native_handle_getter = utils._add_cuda_native_handle_getter
+    {{if 'cudaArray_t' in found_types}}
+    def cudaArray_t_getter(cudaArray_t x): return <uintptr_t><void*><cyruntime.cudaArray_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaArray_t, cudaArray_t_getter)
+    {{endif}}
+    {{if 'cudaArray_const_t' in found_types}}
+    def cudaArray_const_t_getter(cudaArray_const_t x): return <uintptr_t><void*><cyruntime.cudaArray_const_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaArray_const_t, cudaArray_const_t_getter)
+    {{endif}}
+    {{if 'cudaMipmappedArray_t' in found_types}}
+    def cudaMipmappedArray_t_getter(cudaMipmappedArray_t x): return <uintptr_t><void*><cyruntime.cudaMipmappedArray_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaMipmappedArray_t, cudaMipmappedArray_t_getter)
+    {{endif}}
+    {{if 'cudaMipmappedArray_const_t' in found_types}}
+    def cudaMipmappedArray_const_t_getter(cudaMipmappedArray_const_t x): return <uintptr_t><void*><cyruntime.cudaMipmappedArray_const_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaMipmappedArray_const_t, cudaMipmappedArray_const_t_getter)
+    {{endif}}
+    {{if 'cudaStream_t' in found_types}}
+    def cudaStream_t_getter(cudaStream_t x): return <uintptr_t><void*><cyruntime.cudaStream_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaStream_t, cudaStream_t_getter)
+    {{endif}}
+    {{if 'cudaEvent_t' in found_types}}
+    def cudaEvent_t_getter(cudaEvent_t x): return <uintptr_t><void*><cyruntime.cudaEvent_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaEvent_t, cudaEvent_t_getter)
+    {{endif}}
+    {{if 'cudaGraphicsResource_t' in found_types}}
+    def cudaGraphicsResource_t_getter(cudaGraphicsResource_t x): return <uintptr_t><void*><cyruntime.cudaGraphicsResource_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaGraphicsResource_t, cudaGraphicsResource_t_getter)
+    {{endif}}
+    {{if 'cudaExternalMemory_t' in found_types}}
+    def cudaExternalMemory_t_getter(cudaExternalMemory_t x): return <uintptr_t><void*><cyruntime.cudaExternalMemory_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaExternalMemory_t, cudaExternalMemory_t_getter)
+    {{endif}}
+    {{if 'cudaExternalSemaphore_t' in found_types}}
+    def cudaExternalSemaphore_t_getter(cudaExternalSemaphore_t x): return <uintptr_t><void*><cyruntime.cudaExternalSemaphore_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaExternalSemaphore_t, cudaExternalSemaphore_t_getter)
+    {{endif}}
+    {{if 'cudaGraph_t' in found_types}}
+    def cudaGraph_t_getter(cudaGraph_t x): return <uintptr_t><void*><cyruntime.cudaGraph_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaGraph_t, cudaGraph_t_getter)
+    {{endif}}
+    {{if 'cudaGraphNode_t' in found_types}}
+    def cudaGraphNode_t_getter(cudaGraphNode_t x): return <uintptr_t><void*><cyruntime.cudaGraphNode_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaGraphNode_t, cudaGraphNode_t_getter)
+    {{endif}}
+    {{if 'cudaUserObject_t' in found_types}}
+    def cudaUserObject_t_getter(cudaUserObject_t x): return <uintptr_t><void*><cyruntime.cudaUserObject_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaUserObject_t, cudaUserObject_t_getter)
+    {{endif}}
+    {{if 'cudaFunction_t' in found_types}}
+    def cudaFunction_t_getter(cudaFunction_t x): return <uintptr_t><void*><cyruntime.cudaFunction_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaFunction_t, cudaFunction_t_getter)
+    {{endif}}
+    {{if 'cudaKernel_t' in found_types}}
+    def cudaKernel_t_getter(cudaKernel_t x): return <uintptr_t><void*><cyruntime.cudaKernel_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaKernel_t, cudaKernel_t_getter)
+    {{endif}}
+    {{if 'cudaLibrary_t' in found_types}}
+    def cudaLibrary_t_getter(cudaLibrary_t x): return <uintptr_t><void*><cyruntime.cudaLibrary_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaLibrary_t, cudaLibrary_t_getter)
+    {{endif}}
+    {{if 'cudaMemPool_t' in found_types}}
+    def cudaMemPool_t_getter(cudaMemPool_t x): return <uintptr_t><void*><cyruntime.cudaMemPool_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaMemPool_t, cudaMemPool_t_getter)
+    {{endif}}
+    {{if 'cudaGraphExec_t' in found_types}}
+    def cudaGraphExec_t_getter(cudaGraphExec_t x): return <uintptr_t><void*><cyruntime.cudaGraphExec_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaGraphExec_t, cudaGraphExec_t_getter)
+    {{endif}}
+    {{if 'cudaGraphDeviceNode_t' in found_types}}
+    def cudaGraphDeviceNode_t_getter(cudaGraphDeviceNode_t x): return <uintptr_t><void*><cyruntime.cudaGraphDeviceNode_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaGraphDeviceNode_t, cudaGraphDeviceNode_t_getter)
+    {{endif}}
+    {{if 'cudaAsyncCallbackHandle_t' in found_types}}
+    def cudaAsyncCallbackHandle_t_getter(cudaAsyncCallbackHandle_t x): return <uintptr_t><void*><cyruntime.cudaAsyncCallbackHandle_t>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaAsyncCallbackHandle_t, cudaAsyncCallbackHandle_t_getter)
+    {{endif}}
+    {{if 'cudaLogsCallbackHandle' in found_types}}
+    def cudaLogsCallbackHandle_getter(cudaLogsCallbackHandle x): return <uintptr_t><void*><cyruntime.cudaLogsCallbackHandle>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaLogsCallbackHandle, cudaLogsCallbackHandle_getter)
+    {{endif}}
+    {{if True}}
+    def EGLImageKHR_getter(EGLImageKHR x): return <uintptr_t><void*><cyruntime.EGLImageKHR>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(EGLImageKHR, EGLImageKHR_getter)
+    {{endif}}
+    {{if True}}
+    def EGLStreamKHR_getter(EGLStreamKHR x): return <uintptr_t><void*><cyruntime.EGLStreamKHR>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(EGLStreamKHR, EGLStreamKHR_getter)
+    {{endif}}
+    {{if True}}
+    def EGLSyncKHR_getter(EGLSyncKHR x): return <uintptr_t><void*><cyruntime.EGLSyncKHR>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(EGLSyncKHR, EGLSyncKHR_getter)
+    {{endif}}
+    {{if True}}
+    def cudaEglStreamConnection_getter(cudaEglStreamConnection x): return <uintptr_t><void*><cyruntime.cudaEglStreamConnection>(x._pvt_ptr[0])
+    _add_cuda_native_handle_getter(cudaEglStreamConnection, cudaEglStreamConnection_getter)
+    {{endif}}
+_add_native_handle_getters()
+

cuda_bindings/cuda/bindings/utils/__init__.py

@@ -1,5 +1,31 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
-from ._get_handle import get_cuda_native_handle
+from typing import Any, Callable
+
 from ._ptx_utils import get_minimal_required_cuda_ver_from_ptx_ver, get_ptx_ver
+
+_handle_getters: dict[type, Callable[[Any], int]] = {}
+
+
+def _add_cuda_native_handle_getter(t: type, getter: Callable[[Any], int]) -> None:
+    _handle_getters[t] = getter
+
+
+def get_cuda_native_handle(obj: Any) -> int:
+    """Returns the address of the provided CUDA Python object as a Python int.
+
+    Parameters
+    ----------
+    obj : Any
+        CUDA Python object
+
+    Returns
+    -------
+    int : The object address.
+    """
+    obj_type = type(obj)
+    try:
+        return _handle_getters[obj_type](obj)
+    except KeyError:
+        raise TypeError("Unknown type: " + str(obj_type)) from None