make cuPythonInit reentrant + ensure GIL is released when calling underlying C APIs

Author: leofang

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

@@ -43,12 +43,11 @@ cdef bint __cuPythonInit = False
 
 cdef int cuPythonInit() except -1 nogil:
     global __cuPythonInit
+    if __cuPythonInit:
+        return 0
 
     # Load library
     with gil, __symbol_lock:
-        if __cuPythonInit:
-            return 0
-
         {{if 'Windows' == platform.system()}}
         handle = load_nvidia_dynamic_lib("nvrtc")._handle_uint
 
@@ -237,7 +236,7 @@ cdef int cuPythonInit() except -1 nogil:
         {{endif}}
 
         {{else}}
-        handle = <void*><uintptr_t>load_nvidia_dynamic_lib("nvrtc")._handle_uint
+        handle = <void*><uintptr_t>(load_nvidia_dynamic_lib("nvrtc")._handle_uint)
 
         # Load function
         {{if 'nvrtcGetErrorString' in found_functions}}

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

@@ -79,12 +79,13 @@ 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/cufile_linux.pyx

@@ -61,12 +61,13 @@ 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_linux.pyx

@@ -43,12 +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/nvjitlink_windows.pyx

@@ -60,12 +60,13 @@ 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_linux.pyx

@@ -42,12 +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)