Rebase to CTK 11.5

Author: vzhurba01

.gitattributes

@@ -0,0 +1 @@
+cuda/_version.py export-subst

.gitignore

@@ -12,6 +12,7 @@ __pycache__/
 .benchmarks/
 *.cpp
 !cuda/_lib/param_packer.cpp
+!cuda/_cuda/loader.cpp
 
 # Distribution / packaging
 .Python
@@ -77,6 +78,7 @@ instance/
 
 # Sphinx documentation
 docs/_build/
+docs_src/_build/
 
 # PyBuilder
 .pybuilder/
@@ -142,4 +144,4 @@ dmypy.json
 .pytype/
 
 # Cython debug symbols
-cython_debug/
+cython_debug/

README.md

@@ -8,9 +8,8 @@ Dependencies of the CUDA-Python bindings and some versions that are known to
 work are as follows:
 
 * Driver: Linux (450.80.02 or later) Windows(456.38 or later)
-* CUDA Toolkit 11.0 to 11.4 - e.g. 11.4.48
+* CUDA Toolkit 11.0 to 11.5
 * Cython - e.g. 0.29.21
-* Versioneer - e.g. 0.20
 
 ### Compilation
 
@@ -36,20 +35,6 @@ to use the module in-place in your current Python environment (e.g. for testing
 of porting other libraries to use the binding).
 
 
-### Build the Docs
-
-```
-conda env create -f docs/environment-docs.yml
-conda activate cuda-python-docs
-```
-Then compile and install `cuda-python` following the steps above.
-
-```
-cd docs
-make html
-open build/html/index.html
-```
-
 ### Build the Docs
 
 ```
@@ -81,7 +66,6 @@ Dependencies of the test execution and some versions that are known to
 work are as follows:
 
 * numpy-1.19.5
-* numba-0.53.1
 * matplotlib-3.3.4
 * scipy-1.6.3
 * pytest-benchmark-3.4.1
@@ -91,30 +75,29 @@ work are as follows:
 You can run the included tests with:
 
 ```
-pytest
+python -m pytest
 ```
+### Benchmark
 
-### Samples
-
-You can run the included tests with:
+You can run benchmark only tests with:
 
 ```
-pytest examples
+python -m pytest --benchmark-only
 ```
 
-### Benchmark
+### Samples
 
-You can run benchmark only tests with:
+You can run the included tests with:
 
 ```
-pytest --benchmark-only
+python -m pytest examples
 ```
 
 ## Examples
 
 The included examples are:
 
-- `examples/extra/jit_program.py`: Demonstrates the use of the API to compile and
+- `examples/extra/jit_program_test.py`: Demonstrates the use of the API to compile and
   launch a kernel on the device. Includes device memory allocation /
   deallocation, transfers between host and device, creation and usage of
   streams, and context management.

cuda/_cuda/ccuda.pxd

@@ -41,6 +41,8 @@ cdef CUresult _cuDeviceGetMemPool(CUmemoryPool* pool, CUdevice dev) nogil except
 
 cdef CUresult _cuDeviceGetDefaultMemPool(CUmemoryPool* pool_out, CUdevice dev) nogil except ?CUDA_ERROR_NOT_FOUND
 
+cdef CUresult _cuFlushGPUDirectRDMAWrites(CUflushGPUDirectRDMAWritesTarget target, CUflushGPUDirectRDMAWritesScope scope) nogil except ?CUDA_ERROR_NOT_FOUND
+
 cdef CUresult _cuDeviceGetProperties(CUdevprop* prop, CUdevice dev) nogil except ?CUDA_ERROR_NOT_FOUND
 
 cdef CUresult _cuDeviceComputeCapability(int* major, int* minor, CUdevice dev) nogil except ?CUDA_ERROR_NOT_FOUND
@@ -704,5 +706,3 @@ cdef CUresult _cuGraphicsUnmapResources(unsigned int count, CUgraphicsResource*
 cdef CUresult _cuGetProcAddress(const char* symbol, void** pfn, int cudaVersion, cuuint64_t flags) nogil except ?CUDA_ERROR_NOT_FOUND
 
 cdef CUresult _cuGetExportTable(const void** ppExportTable, const CUuuid* pExportTableId) nogil except ?CUDA_ERROR_NOT_FOUND
-
-cdef CUresult _cuFlushGPUDirectRDMAWrites(CUflushGPUDirectRDMAWritesTarget target, CUflushGPUDirectRDMAWritesScope scope) nogil except ?CUDA_ERROR_NOT_FOUND

cuda/_cuda/ccuda.pyx

@@ -8,8 +8,11 @@
 IF UNAME_SYSNAME == "Windows":
     import win32api
     import struct
+    from pywintypes import error
 ELSE:
     cimport cuda._lib.dlfcn as dlfcn
+import sys
+cimport cuda._cuda.loader as loader
 cdef bint __cuPythonInit = False
 cdef void *__cuGetErrorString = NULL
 cdef void *__cuGetErrorName = NULL
@@ -28,6 +31,7 @@ cdef void *__cuDeviceGetNvSciSyncAttributes = NULL
 cdef void *__cuDeviceSetMemPool = NULL
 cdef void *__cuDeviceGetMemPool = NULL
 cdef void *__cuDeviceGetDefaultMemPool = NULL
+cdef void *__cuFlushGPUDirectRDMAWrites = NULL
 cdef void *__cuDeviceGetProperties = NULL
 cdef void *__cuDeviceComputeCapability = NULL
 cdef void *__cuDevicePrimaryCtxRetain = NULL
@@ -360,29 +364,34 @@ cdef void *__cuGraphicsMapResources = NULL
 cdef void *__cuGraphicsUnmapResources = NULL
 cdef void *__cuGetProcAddress = NULL
 cdef void *__cuGetExportTable = NULL
-cdef void *__cuFlushGPUDirectRDMAWrites = NULL
 
 cdef int cuPythonInit() nogil except -1:
     global __cuPythonInit
     if __cuPythonInit:
         return 0
     __cuPythonInit = True
-    IF UNAME_SYSNAME == "Windows":
-        LOAD_LIBRARY_SEARCH_SYSTEM32 = 0x00000800
-        with gil:
-            if 8 * struct.calcsize("P") == 32:
-                try:
-                    handle = win32api.LoadLibraryEx('nvcuda32.dll', 0, LOAD_LIBRARY_SEARCH_SYSTEM32)
-                except:
-                    handle = win32api.LoadLibraryEx('nvcuda.dll', 0, LOAD_LIBRARY_SEARCH_SYSTEM32)
-            else:
-                handle = win32api.LoadLibraryEx('nvcuda.dll', 0, LOAD_LIBRARY_SEARCH_SYSTEM32)
-    ELSE:
-        handle = dlfcn.dlopen('libcuda.so', dlfcn.RTLD_NOW)
-        if (handle == NULL):
-            with gil:
+    cdef char libPath[260]
+    libPath[0] = 0
+    with gil:
+        status = loader.getCUDALibraryPath(libPath, sys.maxsize > 2**32)
+        if status == 0 and len(libPath) != 0:
+            path = libPath.decode('utf-8')
+        else:
+            IF UNAME_SYSNAME == "Windows":
+                path = 'nvcuda.dll'
+            ELSE:
+                path = 'libcuda.so'
+
+        IF UNAME_SYSNAME == "Windows":
+            LOAD_LIBRARY_SEARCH_SYSTEM32 = 0x00000800
+            try:
+                handle = win32api.LoadLibraryEx(path, 0, LOAD_LIBRARY_SEARCH_SYSTEM32)
+            except error as e:
+                raise RuntimeError('Failed to LoadLibraryEx ' + path)
+        ELSE:
+            handle = dlfcn.dlopen(bytes(path, encoding='utf-8'), dlfcn.RTLD_NOW)
+            if (handle == NULL):
                 raise RuntimeError('Failed to dlopen libcuda.so')
-
     # All Globals
     global __cuGetErrorString
     global __cuGetErrorName
@@ -401,6 +410,7 @@ cdef int cuPythonInit() nogil except -1:
     global __cuDeviceSetMemPool
     global __cuDeviceGetMemPool
     global __cuDeviceGetDefaultMemPool
+    global __cuFlushGPUDirectRDMAWrites
     global __cuDeviceGetProperties
     global __cuDeviceComputeCapability
     global __cuDevicePrimaryCtxRetain
@@ -733,7 +743,6 @@ cdef int cuPythonInit() nogil except -1:
     global __cuGraphicsUnmapResources
     global __cuGetProcAddress
     global __cuGetExportTable
-    global __cuFlushGPUDirectRDMAWrites
     # Get latest __cuGetProcAddress
     IF UNAME_SYSNAME == "Windows":
         with gil:
@@ -763,6 +772,7 @@ cdef int cuPythonInit() nogil except -1:
         _cuGetProcAddress('cuDeviceSetMemPool', &__cuDeviceSetMemPool, 11020, 0)
         _cuGetProcAddress('cuDeviceGetMemPool', &__cuDeviceGetMemPool, 11020, 0)
         _cuGetProcAddress('cuDeviceGetDefaultMemPool', &__cuDeviceGetDefaultMemPool, 11020, 0)
+        _cuGetProcAddress('cuFlushGPUDirectRDMAWrites', &__cuFlushGPUDirectRDMAWrites, 11030, 0)
         _cuGetProcAddress('cuDeviceGetProperties', &__cuDeviceGetProperties, 2000, 0)
         _cuGetProcAddress('cuDeviceComputeCapability', &__cuDeviceComputeCapability, 2000, 0)
         _cuGetProcAddress('cuDevicePrimaryCtxRetain', &__cuDevicePrimaryCtxRetain, 7000, 0)
@@ -1095,7 +1105,6 @@ cdef int cuPythonInit() nogil except -1:
         _cuGetProcAddress('cuGraphicsUnmapResources', &__cuGraphicsUnmapResources, 3000, 0)
         _cuGetProcAddress('cuGetProcAddress', &__cuGetProcAddress, 11030, 0)
         _cuGetProcAddress('cuGetExportTable', &__cuGetExportTable, 3000, 0)
-        _cuGetProcAddress('cuFlushGPUDirectRDMAWrites', &__cuFlushGPUDirectRDMAWrites, 11030, 0)
         return 0
     # dlsym calls
     IF UNAME_SYSNAME == "Windows":
@@ -1168,6 +1177,10 @@ cdef int cuPythonInit() nogil except -1:
                 __cuDeviceGetDefaultMemPool = <void*><unsigned long long>win32api.GetProcAddress(handle, 'cuDeviceGetDefaultMemPool')
             except:
                 pass
+            try:
+                __cuFlushGPUDirectRDMAWrites = <void*><unsigned long long>win32api.GetProcAddress(handle, 'cuFlushGPUDirectRDMAWrites')
+            except:
+                pass
             try:
                 __cuDeviceGetProperties = <void*><unsigned long long>win32api.GetProcAddress(handle, 'cuDeviceGetProperties')
             except:
@@ -2496,10 +2509,6 @@ cdef int cuPythonInit() nogil except -1:
                 __cuGetExportTable = <void*><unsigned long long>win32api.GetProcAddress(handle, 'cuGetExportTable')
             except:
                 pass
-            try:
-                __cuFlushGPUDirectRDMAWrites = <void*><unsigned long long>win32api.GetProcAddress(handle, 'cuFlushGPUDirectRDMAWrites')
-            except:
-                pass
     ELSE:
         __cuGetErrorString = dlfcn.dlsym(handle, 'cuGetErrorString')
         __cuGetErrorName = dlfcn.dlsym(handle, 'cuGetErrorName')
@@ -2518,6 +2527,7 @@ cdef int cuPythonInit() nogil except -1:
         __cuDeviceSetMemPool = dlfcn.dlsym(handle, 'cuDeviceSetMemPool')
         __cuDeviceGetMemPool = dlfcn.dlsym(handle, 'cuDeviceGetMemPool')
         __cuDeviceGetDefaultMemPool = dlfcn.dlsym(handle, 'cuDeviceGetDefaultMemPool')
+        __cuFlushGPUDirectRDMAWrites = dlfcn.dlsym(handle, 'cuFlushGPUDirectRDMAWrites')
         __cuDeviceGetProperties = dlfcn.dlsym(handle, 'cuDeviceGetProperties')
         __cuDeviceComputeCapability = dlfcn.dlsym(handle, 'cuDeviceComputeCapability')
         __cuDevicePrimaryCtxRetain = dlfcn.dlsym(handle, 'cuDevicePrimaryCtxRetain')
@@ -2850,7 +2860,6 @@ cdef int cuPythonInit() nogil except -1:
         __cuGraphicsUnmapResources = dlfcn.dlsym(handle, 'cuGraphicsUnmapResources')
         __cuGetProcAddress = dlfcn.dlsym(handle, 'cuGetProcAddress')
         __cuGetExportTable = dlfcn.dlsym(handle, 'cuGetExportTable')
-        __cuFlushGPUDirectRDMAWrites = dlfcn.dlsym(handle, 'cuFlushGPUDirectRDMAWrites')
 
 cdef CUresult _cuGetErrorString(CUresult error, const char** pStr) nogil except ?CUDA_ERROR_NOT_FOUND:
     global __cuGetErrorString
@@ -3005,6 +3014,15 @@ cdef CUresult _cuDeviceGetDefaultMemPool(CUmemoryPool* pool_out, CUdevice dev) n
     err = (<CUresult (*)(CUmemoryPool*, CUdevice) nogil> __cuDeviceGetDefaultMemPool)(pool_out, dev)
     return err
 
+cdef CUresult _cuFlushGPUDirectRDMAWrites(CUflushGPUDirectRDMAWritesTarget target, CUflushGPUDirectRDMAWritesScope scope) nogil except ?CUDA_ERROR_NOT_FOUND:
+    global __cuFlushGPUDirectRDMAWrites
+    cuPythonInit()
+    if __cuFlushGPUDirectRDMAWrites == NULL:
+        with gil:
+            raise RuntimeError('Function "cuFlushGPUDirectRDMAWrites" not found')
+    err = (<CUresult (*)(CUflushGPUDirectRDMAWritesTarget, CUflushGPUDirectRDMAWritesScope) nogil> __cuFlushGPUDirectRDMAWrites)(target, scope)
+    return err
+
 cdef CUresult _cuDeviceGetProperties(CUdevprop* prop, CUdevice dev) nogil except ?CUDA_ERROR_NOT_FOUND:
     global __cuDeviceGetProperties
     cuPythonInit()
@@ -5992,12 +6010,3 @@ cdef CUresult _cuGetExportTable(const void** ppExportTable, const CUuuid* pExpor
             raise RuntimeError('Function "cuGetExportTable" not found')
     err = (<CUresult (*)(const void**, const CUuuid*) nogil> __cuGetExportTable)(ppExportTable, pExportTableId)
     return err
-
-cdef CUresult _cuFlushGPUDirectRDMAWrites(CUflushGPUDirectRDMAWritesTarget target, CUflushGPUDirectRDMAWritesScope scope) nogil except ?CUDA_ERROR_NOT_FOUND:
-    global __cuFlushGPUDirectRDMAWrites
-    cuPythonInit()
-    if __cuFlushGPUDirectRDMAWrites == NULL:
-        with gil:
-            raise RuntimeError('Function "cuFlushGPUDirectRDMAWrites" not found')
-    err = (<CUresult (*)(CUflushGPUDirectRDMAWritesTarget, CUflushGPUDirectRDMAWritesScope) nogil> __cuFlushGPUDirectRDMAWrites)(target, scope)
-    return err