def run_test():
# on windows only nvvm is available to numba
if sys.platform.startswith('win'):
nvvm = NVVM()
print("NVVM version", nvvm.get_version())
return nvvm.get_version() is not None
if not test():
return False
nvvm = NVVM()
print("NVVM version", nvvm.get_version())
# check pkg version matches lib pulled in
gotlib = get_cudalib('cublas')
# check cufft b/c cublas has an incorrect version in 10.1 update 1
gotlib = get_cudalib('cufft')
return bool(get_cudalib('cublas')) and bool(get_cudalib('cufft'))
def run_test():
if not test():
return False
nvvm = NVVM()
print("NVVM version", nvvm.get_version())
# check pkg version matches lib pulled in
gotlib = get_cudalib('cublas')
lookfor = '9.2'
if sys.platform.startswith('win'):
# windows libs have no dot
lookfor = lookfor.replace('.', '')
return lookfor in gotlib
def run_test():
if not test():
return False
nvvm = NVVM()
print("NVVM version", nvvm.get_version())
# check pkg version matches lib pulled in
gotlib = get_cudalib('cublas')
lookfor = os.environ['PKG_VERSION']
if sys.platform.startswith('win'):
# windows libs have no dot
lookfor = lookfor.replace('.', '')
return lookfor in gotlib
def run_test():
# on windows only nvvm is available to numba
if sys.platform.startswith('win'):
nvvm = NVVM()
print("NVVM version", nvvm.get_version())
return nvvm.get_version() is not None
if not test():
return False
nvvm = NVVM()
print("NVVM version", nvvm.get_version())
# check pkg version matches lib pulled in
gotlib = get_cudalib('cublas')
lookfor = os.environ['PKG_VERSION']
return lookfor in gotlib
def run_test():
# on windows only nvvm is available to numba
if sys.platform.startswith("win"):
nvvm = NVVM()
print("NVVM version", nvvm.get_version())
return nvvm.get_version() is not None
if not test():
return False
nvvm = NVVM()
print("NVVM version", nvvm.get_version())
extra_lib_tests = (
"cublas", # check pkg version matches lib pulled in
"cufft", # check cufft b/c cublas has an incorrect version in 10.1 update 1
"cupti", # check this is getting included
)
found_paths = [get_cudalib(lib) for lib in extra_lib_tests]
print(*zip(extra_lib_tests, found_paths), sep="\n")
return all(extra_lib_tests)
def __init__(self,
py_func,
argtypes,
link=None,
debug=False,
lineinfo=False,
inline=False,
fastmath=False,
extensions=None,
max_registers=None,
opt=True,
device=False):
if device:
raise RuntimeError('Cannot compile a device function as a kernel')
super().__init__()
self.py_func = py_func
self.argtypes = argtypes
self.debug = debug
self.lineinfo = lineinfo
self.extensions = extensions or []
nvvm_options = {
'debug': self.debug,
'lineinfo': self.lineinfo,
'fastmath': fastmath,
'opt': 3 if opt else 0
}
cres = compile_cuda(self.py_func,
types.void,
self.argtypes,
debug=self.debug,
lineinfo=self.lineinfo,
inline=inline,
fastmath=fastmath,
nvvm_options=nvvm_options)
tgt_ctx = cres.target_context
code = self.py_func.__code__
filename = code.co_filename
linenum = code.co_firstlineno
lib, kernel = tgt_ctx.prepare_cuda_kernel(cres.library, cres.fndesc,
debug, nvvm_options,
filename, linenum,
max_registers)
if not link:
link = []
# A kernel needs cooperative launch if grid_sync is being used.
self.cooperative = 'cudaCGGetIntrinsicHandle' in lib.get_asm_str()
# We need to link against cudadevrt if grid sync is being used.
if self.cooperative:
link.append(get_cudalib('cudadevrt', static=True))
for filepath in link:
lib.add_linking_file(filepath)
# populate members
self.entry_name = kernel.name
self.signature = cres.signature
self._type_annotation = cres.type_annotation
self._codelibrary = lib
self.call_helper = cres.call_helper
def __init__(self,
py_func,
argtypes,
link=None,
debug=False,
lineinfo=False,
inline=False,
fastmath=False,
extensions=None,
max_registers=None,
opt=True,
device=False):
if device:
raise RuntimeError('Cannot compile a device function as a kernel')
super().__init__()
# _DispatcherBase.nopython_signatures() expects this attribute to be
# present, because it assumes an overload is a CompileResult. In the
# CUDA target, _Kernel instances are stored instead, so we provide this
# attribute here to avoid duplicating nopython_signatures() in the CUDA
# target with slight modifications.
self.objectmode = False
# The finalizer constructed by _DispatcherBase._make_finalizer also
# expects overloads to be a CompileResult. It uses the entry_point to
# remove a CompileResult from a target context. However, since we never
# insert kernels into a target context (there is no need because they
# cannot be called by other functions, only through the dispatcher) it
# suffices to pretend we have an entry point of None.
self.entry_point = None
self.py_func = py_func
self.argtypes = argtypes
self.debug = debug
self.lineinfo = lineinfo
self.extensions = extensions or []
nvvm_options = {
'debug': self.debug,
'lineinfo': self.lineinfo,
'fastmath': fastmath,
'opt': 3 if opt else 0
}
cres = compile_cuda(self.py_func,
types.void,
self.argtypes,
debug=self.debug,
lineinfo=self.lineinfo,
inline=inline,
fastmath=fastmath,
nvvm_options=nvvm_options)
tgt_ctx = cres.target_context
code = self.py_func.__code__
filename = code.co_filename
linenum = code.co_firstlineno
lib, kernel = tgt_ctx.prepare_cuda_kernel(cres.library, cres.fndesc,
debug, nvvm_options,
filename, linenum,
max_registers)
if not link:
link = []
# A kernel needs cooperative launch if grid_sync is being used.
self.cooperative = 'cudaCGGetIntrinsicHandle' in lib.get_asm_str()
# We need to link against cudadevrt if grid sync is being used.
if self.cooperative:
link.append(get_cudalib('cudadevrt', static=True))
for filepath in link:
lib.add_linking_file(filepath)
# populate members
self.entry_name = kernel.name
self.signature = cres.signature
self._type_annotation = cres.type_annotation
self._codelibrary = lib
self.call_helper = cres.call_helper
