move init check inside lock

Author: kkraus14

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

@@ -14,7 +14,7 @@ import os
 import sys
 import threading
 cimport cuda.bindings._bindings.loader as loader
-cdef object __symbol_lock = threading.Lock()
+cdef object __symbol_lock = threading.RLock()
 cdef bint __cuPythonInit = False
 {{if 'cuGetErrorString' in found_functions}}cdef void *__cuGetErrorString = NULL{{endif}}
 {{if 'cuGetErrorName' in found_functions}}cdef void *__cuGetErrorName = NULL{{endif}}
@@ -488,13 +488,14 @@ cdef bint __cuPythonInit = False
 
 cdef int cuPythonInit() except -1 nogil:
     global __cuPythonInit
-    if __cuPythonInit:
-        return 0
-    __cuPythonInit = True
-
     cdef bint usePTDS
     cdef char libPath[260]
+
     with gil, __symbol_lock:
+        if __cuPythonInit:
+            return 0
+        __cuPythonInit = True
+
         usePTDS = os.getenv('CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM', default=0)
 
         # Load library

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

@@ -42,10 +42,11 @@ cdef bint __cuPythonInit = False
 
 cdef int cuPythonInit() except -1 nogil:
     global __cuPythonInit
-    if __cuPythonInit:
-        return 0
 
     with gil, __symbol_lock:
+        if __cuPythonInit:
+            return 0
+
         {{if 'Windows' == platform.system()}}
         handle = load_nvidia_dynamic_lib("nvrtc")._handle_uint
 

cuda_bindings/cuda/bindings/_internal/cufile_linux.pyx

@@ -92,12 +92,13 @@ cdef void* load_library(const int driver_ver) except* with gil:
 
 cdef int _check_or_init_cufile() except -1 nogil:
     global __py_cufile_init
-    if __py_cufile_init:
-        return 0
-
     cdef void* handle = NULL
     cdef int err, driver_ver = 0
+
     with gil, __symbol_lock:
+        if __py_cufile_init:
+            return 0
+
         # Load driver to check version
         handle = dlopen('libcuda.so.1', RTLD_NOW | RTLD_GLOBAL)
         if handle == NULL:

cuda_bindings/cuda/bindings/_internal/nvjitlink_linux.pyx

@@ -62,12 +62,13 @@ cdef void* load_library(int driver_ver) except* with gil:
 
 cdef int _check_or_init_nvjitlink() except -1 nogil:
     global __py_nvjitlink_init
-    if __py_nvjitlink_init:
-        return 0
-
     cdef void* handle = NULL
     cdef int err, driver_ver = 0
+
     with gil, __symbol_lock:
+        if __py_nvjitlink_init:
+            return 0
+
         # Load driver to check version
         handle = dlopen('libcuda.so.1', RTLD_NOW | RTLD_GLOBAL)
         if handle == NULL:

cuda_bindings/cuda/bindings/_internal/nvjitlink_windows.pyx

@@ -43,11 +43,12 @@ cdef void* __nvJitLinkVersion = NULL
 
 cdef int _check_or_init_nvjitlink() except -1 nogil:
     global __py_nvjitlink_init
-    if __py_nvjitlink_init:
-        return 0
-
     cdef int err, driver_ver = 0
+
     with gil, __symbol_lock:
+        if __py_nvjitlink_init:
+            return 0
+
         # Load driver to check version
         try:
             handle = win32api.LoadLibraryEx("nvcuda.dll", 0, LOAD_LIBRARY_SEARCH_SYSTEM32)

cuda_bindings/cuda/bindings/_internal/nvvm_linux.pyx

@@ -61,12 +61,13 @@ cdef void* load_library(const int driver_ver) except* with gil:
 
 cdef int _check_or_init_nvvm() except -1 nogil:
     global __py_nvvm_init
-    if __py_nvvm_init:
-        return 0
-
     cdef void* handle = NULL
     cdef int err, driver_ver = 0
+
     with gil, __symbol_lock:
+        if __py_nvvm_init:
+            return 0
+
         # Load driver to check version
         handle = dlopen('libcuda.so.1', RTLD_NOW | RTLD_GLOBAL)
         if handle == NULL:

cuda_bindings/cuda/bindings/_internal/nvvm_windows.pyx

@@ -42,11 +42,12 @@ cdef void* __nvvmGetProgramLog = NULL
 
 cdef int _check_or_init_nvvm() except -1 nogil:
     global __py_nvvm_init
-    if __py_nvvm_init:
-        return 0
-
     cdef int err, driver_ver = 0
+
     with gil, __symbol_lock:
+        if __py_nvvm_init:
+            return 0
+
         # Load driver to check version
         try:
             handle = win32api.LoadLibraryEx("nvcuda.dll", 0, LOAD_LIBRARY_SEARCH_SYSTEM32)