Backport #914 and #894 to 12.9.x branch

.gitignore

@@ -29,10 +29,6 @@ cuda_bindings/cuda/bindings/_bindings/cynvrtc.pxd
 cuda_bindings/cuda/bindings/_bindings/cynvrtc.pyx
 cuda_bindings/cuda/bindings/_internal/nvjitlink.pyx
 cuda_bindings/cuda/bindings/_internal/nvvm.pyx
-cuda_bindings/cuda/bindings/_lib/cyruntime/cyruntime.pxd
-cuda_bindings/cuda/bindings/_lib/cyruntime/cyruntime.pyx
-cuda_bindings/cuda/bindings/_lib/cyruntime/utils.pxd
-cuda_bindings/cuda/bindings/_lib/cyruntime/utils.pyx
 cuda_bindings/cuda/bindings/_lib/utils.pxd
 cuda_bindings/cuda/bindings/_lib/utils.pyx
 cuda_bindings/cuda/bindings/cydriver.pxd
@@ -115,6 +111,7 @@ instance/
 
 # Sphinx documentation
 docs_src/_build/
+*/docs/source/generated/
 
 # PyBuilder
 .pybuilder/

cuda_bindings/cuda/bindings/_bindings/cydriver.pyx.in

@@ -495,10 +495,8 @@ cdef bint __cuPythonInit = False
 ctypedef CUresult (*__cuGetProcAddress_v2_T)(const char*, void**, int, cuuint64_t, CUdriverProcAddressQueryResult*) except?CUDA_ERROR_NOT_FOUND nogil
 cdef __cuGetProcAddress_v2_T _F_cuGetProcAddress_v2 = NULL
 
-cdef int cuPythonInit() except -1 nogil:
+cdef int _cuPythonInit() except -1 nogil:
     global __cuPythonInit
-    if __cuPythonInit:
-        return 0
 
     cdef bint usePTDS
     cdef char libPath[260]
@@ -8946,6 +8944,13 @@ cdef int cuPythonInit() except -1 nogil:
         __cuPythonInit = True
         return 0
 
+# Create a very small function to check whether we are init'ed, so the C
+# compiler can inline it.
+cdef inline int cuPythonInit() except -1 nogil:
+    if __cuPythonInit:
+        return 0
+    return _cuPythonInit()
+
 {{if 'cuGetErrorString' in found_functions}}
 
 cdef CUresult _cuGetErrorString(CUresult error, const char** pStr) except ?CUDA_ERROR_NOT_FOUND nogil:

cuda_bindings/cuda/bindings/_bindings/cynvrtc.pyx.in

@@ -42,10 +42,8 @@ cdef bint __cuPythonInit = False
 {{if 'nvrtcGetPCHHeapSizeRequired' in found_functions}}cdef void *__nvrtcGetPCHHeapSizeRequired = NULL{{endif}}
 {{if 'nvrtcSetFlowCallback' in found_functions}}cdef void *__nvrtcSetFlowCallback = NULL{{endif}}
 
-cdef int cuPythonInit() except -1 nogil:
+cdef int _cuPythonInit() except -1 nogil:
     global __cuPythonInit
-    if __cuPythonInit:
-        return 0
 
     # Load library
     with gil, __symbol_lock:
@@ -349,6 +347,13 @@ cdef int cuPythonInit() except -1 nogil:
         __cuPythonInit = True
         return 0
 
+# Create a very small function to check whether we are init'ed, so the C
+# compiler can inline it.
+cdef inline int cuPythonInit() except -1 nogil:
+    if __cuPythonInit:
+        return 0
+    return _cuPythonInit()
+
 {{if 'nvrtcGetErrorString' in found_functions}}
 
 cdef const char* _nvrtcGetErrorString(nvrtcResult result) except ?NULL nogil:

cuda_bindings/cuda/bindings/_bindings/cyruntime.pxd.in

@@ -4,6 +4,8 @@
 # This code was automatically generated with version 12.9.0. Do not modify it directly.
 include "../cyruntime_types.pxi"
 
+include "../_lib/cyruntime/cyruntime.pxd"
+
 {{if 'cudaDeviceReset' in found_functions}}
 
 cdef cudaError_t _cudaDeviceReset() except ?cudaErrorCallRequiresNewerDriver nogil

cuda_bindings/cuda/bindings/_bindings/cyruntime.pyx.in

@@ -10,16 +10,22 @@ cimport cython
 
 cdef bint __cudaPythonInit = False
 cdef bint __usePTDS = False
-cdef int cudaPythonInit() except -1 nogil:
+cdef int _cudaPythonInit() except -1 nogil:
         global __cudaPythonInit
         global __usePTDS
-        if __cudaPythonInit:
-            return __usePTDS
+
         with gil:
             __usePTDS = os.getenv('CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM', default=False)
         __cudaPythonInit = True
         return __usePTDS
 
+# Create a very small function to check whether we are init'ed, so the C
+# compiler can inline it.
+cdef inline int cudaPythonInit() except -1 nogil:
+    if __cudaPythonInit:
+        return __usePTDS
+    return _cudaPythonInit()
+
 {{if 'cudaDeviceReset' in found_functions}}
 
 cdef cudaError_t _cudaDeviceReset() except ?cudaErrorCallRequiresNewerDriver nogil:
@@ -2647,3 +2653,6 @@ cdef cudaError_t _cudaProfilerStop() except ?cudaErrorCallRequiresNewerDriver no
         return ptds._cudaProfilerStop()
     return cudaProfilerStop()
 {{endif}}
+
+
+include "../_lib/cyruntime/cyruntime.pxi"

cuda_bindings/cuda/bindings/_lib/cyruntime/cyruntime.pxd

@@ -0,0 +1,43 @@
+# SPDX-FileCopyrightText: Copyright (c) 2021-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+cimport cuda.bindings.cyruntime as cyruntime
+cimport cuda.bindings._bindings.cydriver as _cydriver
+
+# These graphics API are the reimplemented version of what's supported by CUDA Runtime.
+# Issue https://github.com/NVIDIA/cuda-python/issues/488 will remove them by letting us
+# use call into the static library directly.
+#
+# This is an ABI breaking change which can only happen in a major version bump.
+
+# This file is included from cuda/bindings/_bindings/cyruntime.pxd.in but kept in a
+# separate file to keep it separated from the auto-generated code there.
+
+# Prior to https://github.com/NVIDIA/cuda-python/pull/914, this was two
+# independent modules (c.b._lib.cyruntime.cyruntime and
+# c.b._lib.cyruntime.utils), but was merged into one.
+
+cdef cudaError_t _cudaEGLStreamProducerPresentFrame(cyruntime.cudaEglStreamConnection* conn, cyruntime.cudaEglFrame eglframe, cudaStream_t* pStream) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaEGLStreamProducerReturnFrame(cyruntime.cudaEglStreamConnection* conn, cyruntime.cudaEglFrame* eglframe, cudaStream_t* pStream) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaGraphicsResourceGetMappedEglFrame(cyruntime.cudaEglFrame* eglFrame, cudaGraphicsResource_t resource, unsigned int index, unsigned int mipLevel) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaVDPAUSetVDPAUDevice(int device, cyruntime.VdpDevice vdpDevice, cyruntime.VdpGetProcAddress* vdpGetProcAddress) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaVDPAUGetDevice(int* device, cyruntime.VdpDevice vdpDevice, cyruntime.VdpGetProcAddress* vdpGetProcAddress) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaGraphicsVDPAURegisterVideoSurface(cudaGraphicsResource** resource, cyruntime.VdpVideoSurface vdpSurface, unsigned int flags) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaGraphicsVDPAURegisterOutputSurface(cudaGraphicsResource** resource, cyruntime.VdpOutputSurface vdpSurface, unsigned int flags) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaGLGetDevices(unsigned int* pCudaDeviceCount, int* pCudaDevices, unsigned int cudaDeviceCount, cyruntime.cudaGLDeviceList deviceList) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaGraphicsGLRegisterImage(cudaGraphicsResource** resource, cyruntime.GLuint image, cyruntime.GLenum target, unsigned int flags) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaGraphicsGLRegisterBuffer(cudaGraphicsResource** resource, cyruntime.GLuint buffer, unsigned int flags) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaGraphicsEGLRegisterImage(cudaGraphicsResource_t* pCudaResource, cyruntime.EGLImageKHR image, unsigned int flags) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaEGLStreamConsumerConnect(cyruntime.cudaEglStreamConnection* conn, cyruntime.EGLStreamKHR eglStream) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaEGLStreamConsumerConnectWithFlags(cyruntime.cudaEglStreamConnection* conn, cyruntime.EGLStreamKHR eglStream, unsigned int flags) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaEGLStreamConsumerDisconnect(cyruntime.cudaEglStreamConnection* conn) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaEGLStreamConsumerAcquireFrame(cyruntime.cudaEglStreamConnection* conn, cudaGraphicsResource_t* pCudaResource, cudaStream_t* pStream, unsigned int timeout) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaEGLStreamConsumerReleaseFrame(cyruntime.cudaEglStreamConnection* conn, cudaGraphicsResource_t pCudaResource, cudaStream_t* pStream) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaEGLStreamProducerConnect(cyruntime.cudaEglStreamConnection* conn, cyruntime.EGLStreamKHR eglStream, cyruntime.EGLint width, cyruntime.EGLint height) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaEGLStreamProducerDisconnect(cyruntime.cudaEglStreamConnection* conn) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t _cudaEventCreateFromEGLSync(cudaEvent_t* phEvent, cyruntime.EGLSyncKHR eglSync, unsigned int flags) except ?cudaErrorCallRequiresNewerDriver nogil
+
+# utility functions
+
+cdef cudaError_t getDriverEglFrame(_cydriver.CUeglFrame *cuEglFrame, cyruntime.cudaEglFrame eglFrame) except ?cudaErrorCallRequiresNewerDriver nogil
+cdef cudaError_t getRuntimeEglFrame(cyruntime.cudaEglFrame *eglFrame, _cydriver.CUeglFrame cueglFrame) except ?cudaErrorCallRequiresNewerDriver nogil