Backport #865 to 12.9.x

.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/_bindings/cydriver.pyx.in

@@ -12,9 +12,10 @@ cimport cuda.bindings._lib.dlfcn as dlfcn
 from libc.stdint cimport intptr_t
 import os
 import sys
-import threading
 cimport cuda.bindings._bindings.loader as loader
-cdef object __symbol_lock = threading.RLock()
+import threading
+
+cdef object __symbol_lock = threading.Lock()
 cdef bint __cuPythonInit = False
 {{if 'cuGetErrorString' in found_functions}}cdef void *__cuGetErrorString = NULL{{endif}}
 {{if 'cuGetErrorName' in found_functions}}cdef void *__cuGetErrorName = NULL{{endif}}
@@ -504,7 +505,7 @@ cdef int cuPythonInit() except -1 nogil:
 
     with gil, __symbol_lock:
         usePTDS = os.getenv('CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM', default=0)
-    
+
         # Load library
         libPath[0] = 0
         status = loader.getCUDALibraryPath(libPath, sys.maxsize > 2**32)
@@ -516,7 +517,7 @@ cdef int cuPythonInit() except -1 nogil:
             {{else}}
             path = 'libcuda.so.1'
             {{endif}}
-    
+
         {{if 'Windows' == platform.system()}}
         LOAD_LIBRARY_SEARCH_SYSTEM32 = 0x00000800
         try:
@@ -528,7 +529,7 @@ cdef int cuPythonInit() except -1 nogil:
         if (handle == NULL):
             raise RuntimeError('Failed to dlopen ' + path)
         {{endif}}
-    
+
         # Get latest __cuGetProcAddress_v2
         global __cuGetProcAddress_v2
         {{if 'Windows' == platform.system()}}
@@ -539,7 +540,7 @@ cdef int cuPythonInit() except -1 nogil:
         {{else}}
         __cuGetProcAddress_v2 = dlfcn.dlsym(handle, 'cuGetProcAddress_v2')
         {{endif}}
-    
+
         # Load using cuGetProcAddress if available
         if __cuGetProcAddress_v2 != NULL:
             _F_cuGetProcAddress_v2 = <__cuGetProcAddress_v2_T>__cuGetProcAddress_v2
@@ -2782,10 +2783,9 @@ cdef int cuPythonInit() except -1 nogil:
             global __cuGraphicsVDPAURegisterOutputSurface
             _F_cuGetProcAddress_v2('cuGraphicsVDPAURegisterOutputSurface', &__cuGraphicsVDPAURegisterOutputSurface, 3010, CU_GET_PROC_ADDRESS_DEFAULT, NULL)
             {{endif}}
-
             __cuPythonInit = True
             return 0
-    
+
         {{if 'Windows' == platform.system()}}
         # Load using win32GetAddr
         if usePTDS:
@@ -8943,7 +8943,6 @@ cdef int cuPythonInit() except -1 nogil:
         __cuGraphicsVDPAURegisterOutputSurface = dlfcn.dlsym(handle, 'cuGraphicsVDPAURegisterOutputSurface')
         {{endif}}
         {{endif}}
-
         __cuPythonInit = True
         return 0
 

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

@@ -12,6 +12,7 @@ from libc.stdint cimport uintptr_t
 from cuda.pathfinder import load_nvidia_dynamic_lib
 from libc.stdint cimport intptr_t
 import threading
+
 cdef object __symbol_lock = threading.Lock()
 cdef bint __cuPythonInit = False
 {{if 'nvrtcGetErrorString' in found_functions}}cdef void *__nvrtcGetErrorString = NULL{{endif}}
@@ -343,8 +344,8 @@ cdef int cuPythonInit() except -1 nogil:
         global __nvrtcSetFlowCallback
         __nvrtcSetFlowCallback = dlfcn.dlsym(handle, 'nvrtcSetFlowCallback')
         {{endif}}
-        {{endif}}
 
+        {{endif}}
         __cuPythonInit = True
         return 0