def setUp(self):
self.old_routine_accelerators = _acc.get_routine_accelerators()
self.old_reduction_accelerators = _acc.get_reduction_accelerators()
if self.backend == 'device':
_acc.set_routine_accelerators(['cub'])
_acc.set_reduction_accelerators([])
elif self.backend == 'block':
_acc.set_routine_accelerators([])
_acc.set_reduction_accelerators(['cub'])
def setUp(self):
cupy._core._optimize_config._clear_all_contexts_cache()
self.old_reductions = _accelerator.get_reduction_accelerators()
_accelerator.set_reduction_accelerators(self.backend)
# avoid shadowed by the cub module
self.old_routines = _accelerator.get_routine_accelerators()
_accelerator.set_routine_accelerators([])
self.x = testing.shaped_arange((3, 4), cupy, dtype=cupy.float32)
def setUp(self):
self.order, self.axis = self.order_and_axis
self.old_routine_accelerators = _acc.get_routine_accelerators()
self.old_reduction_accelerators = _acc.get_reduction_accelerators()
if self.backend == 'device':
if self.axis is not None:
raise unittest.SkipTest('does not support')
_acc.set_routine_accelerators(['cub'])
_acc.set_reduction_accelerators([])
elif self.backend == 'block':
_acc.set_routine_accelerators([])
_acc.set_reduction_accelerators(['cub'])
def with_accelerators(self):
old_accelerators = _accelerator.get_routine_accelerators()
if self.enable_cub:
_accelerator.set_routine_accelerators(['cub'])
else:
_accelerator.set_routine_accelerators([])
yield
_accelerator.set_routine_accelerators(old_accelerators)
def test_can_use_accelerator_set_unset(self):
# ensure we use CUB block reduction and not CUB device reduction
old_routine_accelerators = _accelerator.get_routine_accelerators()
_accelerator.set_routine_accelerators([])
a = cupy.random.random((10, 10))
# this is the only function we can mock; the rest is cdef'd
func_name = ''.join(('cupy._core._cub_reduction.',
'_SimpleCubReductionKernel_get_cached_function'))
func = _cub_reduction._SimpleCubReductionKernel_get_cached_function
with testing.AssertFunctionIsCalled(func_name,
wraps=func,
times_called=2): # two passes
a.sum()
with testing.AssertFunctionIsCalled(func_name,
wraps=func,
times_called=1): # one pass
a.sum(axis=1)
with testing.AssertFunctionIsCalled(func_name,
wraps=func,
times_called=0): # not used
a.sum(axis=0)
_accelerator.set_routine_accelerators(old_routine_accelerators)
def setUp(self):
self.order, self.axis = self.order_and_axis
old_routine_accelerators = _acc.get_routine_accelerators()
old_reduction_accelerators = _acc.get_reduction_accelerators()
if self.backend == 'device':
if self.axis is not None:
pytest.skip('does not support')
_acc.set_routine_accelerators(['cub'])
_acc.set_reduction_accelerators([])
elif self.backend == 'block':
_acc.set_routine_accelerators([])
_acc.set_reduction_accelerators(['cub'])
yield
_acc.set_routine_accelerators(old_routine_accelerators)
_acc.set_reduction_accelerators(old_reduction_accelerators)
def tearDown(self):
_accelerator.set_routine_accelerators(self.old_accelerators)
def setUp(self):
self.old_accelerators = _accelerator.get_routine_accelerators()
_accelerator.set_routine_accelerators(['cub'])
def tearDown(self):
_acc.set_routine_accelerators(self.old_routine_accelerators)
_acc.set_reduction_accelerators(self.old_reduction_accelerators)
def setUp(self):
self.old_accelerators = _acc.get_routine_accelerators()
_acc.set_routine_accelerators([])
# also avoid fallback to CUB via the general reduction kernel
self.old_reduction_accelerators = _acc.get_reduction_accelerators()
_acc.set_reduction_accelerators([])
