def __setitem__(self, key, value):
sliced = self.__getitem__(key)
# Use fill for scalar values
if type(value) == np.float32 or type(value) == np.float64 or \
type(value) == float:
sliced.fill(value)
elif type(value) == np.int32 or type(value) == np.int64 or \
type(value) == int:
sliced.fill(value)
elif self.tensor.shape == () or np.prod(self.tensor.shape) == 1:
sliced.fill(value)
elif np.sum(self.tensor.strides) == 0:
view = GPUArray((1, ), dtype=self.tensor.dtype)
view.fill(value)
else:
# Convert to correct dtype if necessary
if value.dtype != self.tensor.dtype:
new_value = np.ndarray(self.tensor.shape, dtype=self.tensor.dtype)
new_value[:] = value
value = new_value
# Reshape to satisfy pycuda if necessary
if sliced.shape != value.shape:
sliced = self.tensor.reshape(value.shape)
if self.is_contiguous and self.strides_contiguous(value):
sliced[:] = value
elif type(value) == GPUArray:
self.from_other(value, sliced)
else:
contig_tensor = GPUArray(value.shape, self.tensor.dtype)
contig_tensor[:] = value
self.from_other(contig_tensor, sliced)
def __setitem__(self, key, value):
sliced = self.__getitem__(key)
# Use fill for scalar values
# convert value to numpy
if type(value) == float:
value = np.float64(value)
elif type(value) == int:
value = np.int64(value)
elif isinstance(value, np.ndarray):
# handle 0-d and 1-d conversion to scalar
if value.shape == ():
value = value[()]
elif value.shape == (1, ):
value = value[0]
# flex: added astype to deal with GPUArray dtype int16
# FLEX TODO: assumed same behavior for all cases
if type(value) in (np.int32, np.int64, int, np.uint32, np.float32,
np.float64):
sliced.fill(value.astype(sliced.dtype))
elif self.tensor.shape == () or np.prod(self.tensor.shape) == 1:
sliced.fill(value.astype(sliced.dtype))
elif np.sum(self.tensor.strides) == 0:
view = GPUArray((1, ), dtype=self.tensor.dtype)
view.fill(value.astype(sliced.dtype))
else:
# Convert to correct dtype if necessary
if value.dtype != self.tensor.dtype:
new_value = np.ndarray(value.shape, dtype=self.tensor.dtype)
new_value[:] = value
value = new_value
# Reshape to satisfy pycuda if necessary
if sliced.shape != value.shape:
sliced = self.tensor.reshape(value.shape)
if self.is_contiguous and self.strides_contiguous(value):
if sliced.shape == ():
sliced.reshape((1, ))[:] = value.reshape((1, ))
else:
sliced[:] = value
elif type(value) == GPUArray:
self.from_other(value, sliced)
else:
contig_tensor = GPUArray(value.shape, self.tensor.dtype)
contig_tensor[:] = value
self.from_other(contig_tensor, sliced)