def _autotune(self, args, setup):
if setup is False:
return args
elif setup is True:
level = configuration['autotuning'].level or 'basic'
args, summary = autotune(self, args, level,
configuration['autotuning'].mode)
elif isinstance(setup, str):
args, summary = autotune(self, args, setup,
configuration['autotuning'].mode)
elif isinstance(setup, tuple) and len(setup) == 2:
level, mode = setup
if level is False:
return args
else:
args, summary = autotune(self, args, level, mode)
else:
raise ValueError(
"Expected bool, str, or 2-tuple, got `%s` instead" %
type(setup))
# Record the tuned values
self._state.setdefault('autotuning', []).append(summary)
return args
def _autotune(self, arguments):
"""
Use auto-tuning on this Operator to determine empirically the
best block sizes when loop blocking is in use.
"""
if self.dle_flags.get('blocking', False):
return autotune(self, arguments, self.dle_arguments)
else:
return arguments
def _autotune(self, args):
"""
Use auto-tuning on this Operator to determine empirically the
best block sizes when loop blocking is in use.
"""
if self.dle_flags.get('blocking', False):
# AT assumes and ordered dict, so let's feed it one
args = OrderedDict([(p.name, args[p.name])
for p in self.parameters])
return autotune(self, args, self.dle_args)
else:
return args
def _autotune(self, args, setup):
if setup is False:
return args
elif setup is True:
level = configuration['autotuning'].level or 'basic'
args, summary = autotune(self, args, level, configuration['autotuning'].mode)
elif isinstance(setup, str):
args, summary = autotune(self, args, setup, configuration['autotuning'].mode)
elif isinstance(setup, tuple) and len(setup) == 2:
level, mode = setup
if level is False:
return args
else:
args, summary = autotune(self, args, level, mode)
else:
raise ValueError("Expected bool, str, or 2-tuple, got `%s` instead"
% type(setup))
# Record the tuned values
self._state.setdefault('autotuning', []).append(summary)
return args
def _autotune(self, args):
if self._dle_flags.get('blocking', False):
return autotune(self, args, self.parameters, self._dle_args)
else:
return args
