def visit(self, op, input):
self.computation.set_op_rank(op)
# TODO - is treating TensorSizeOp as constants, okay?
# Construct constant list with number of elements = reduction axes size
constant_tensor = [op.reduction_axes.size]
constant_op = Constant(Type.f32, Shape([]), constant_tensor)
self.computation.register_cpp_op(op, constant_op)
def test_reduce():
float_element_type = Type.f32
AddParam1 = Parameter(float_element_type, Shape([]))
AddParam2 = Parameter(float_element_type, Shape([]))
constant_op = Constant(float_element_type, Shape([]), [0.])
reduce_function = Function(NodeVector([Add(AddParam1, AddParam2)]),
[AddParam1, AddParam2], 'add')
A = Parameter(float_element_type, Shape([2, 2, 2]))
parameter_list = [A]
function = Function(
NodeVector([Reduce(A, constant_op, reduce_function, AxisSet({0}))]),
parameter_list, 'test')
backend, cf = make_backend_call_frame(function)
a = backend.make_primary_tensor_view(float_element_type, Shape([2, 2, 2]))
result = backend.make_primary_tensor_view(float_element_type, Shape([2,
2]))
a.write(util.numpy_to_c(np.arange(8, dtype=np.float32).reshape(2, 2, 2)),
0, 32)
result_arr = np.zeros((2, 2), dtype=np.float32)
result.write(util.numpy_to_c(result_arr), 0, 16)
cf.call([result], [a])
result.read(util.numpy_to_c(result_arr), 0, 16)
a_arr = np.arange(8).reshape(2, 2, 2)
result_arr_ref = np.add.reduce(a_arr)
assert np.allclose(result_arr, result_arr_ref)
def visit(self, op, input):
self.computation.set_op_rank(op)
input_axes = list(input.axes.lengths)
constant_op = Constant(Type.f32, Shape(input_axes), [1])
ngraph_cpp_reciprocal_op = constant_op \
/ self.computation.lookup_cpp_op(input)
self.computation.register_cpp_op(op, ngraph_cpp_reciprocal_op)
def make_constant_node(value, dtype=None): # type: (NumericData, NumericType) -> Constant
"""Return an ngraph Constant node with the specified value."""
ndarray = get_ndarray(value)
if dtype:
element_type = get_element_type(dtype)
else:
element_type = get_element_type(ndarray.dtype)
return Constant(element_type, Shape(ndarray.shape), ndarray.flatten().tolist())
def test_constant():
element_type = Type.f32
parameter_list = []
function = Function([Constant(element_type, Shape([3, 3]), list(range(9)))], parameter_list, "test")
runtime = get_runtime()
computation = runtime.computation(function, *parameter_list)
result = computation()[0]
expected = np.arange(9).reshape(3, 3)
assert np.allclose(result, expected)
def make_scalar_constant(elem_type, scalar, shape=None, axis_set=None):
# type: (int, float, List[int], Set[int]) -> float
"""Create a Constant node for scalar value."""
if shape is None:
shape = Shape([])
if axis_set is None:
axis_set = AxisSet(set())
scalar_shape = Shape([]) # type: List[int]
constant_op = Constant(elem_type, scalar_shape, [scalar])
constant_broadcast = Broadcast(constant_op, shape, axis_set)
return constant_broadcast
def visit(self, op):
# Can be visited in the most trivial computation we only a variable is created
if not self.computation.has_cpp_op(op):
if op.is_constant:
# FIXME: make tensors based on data type
constant_op = Constant(Type.f32, Shape(list(op.axes.lengths)),
list(self.flatten(op.const.tolist())))
self.computation.register_cpp_op(op, constant_op)
else:
op_element_type = Parameter(Type.f32,
Shape(list(op.axes.lengths)))
self.computation.register_cpp_op(op, op_element_type)
if not op.is_placeholder:
self.computation.neon_variable_list.append(op)
def visit(self, op):
self.computation.set_op_rank(op)
tensor = op.tensor
if not self.computation.has_cpp_op(op):
if tensor.is_constant:
# FIXME: make tensors based on data type
constant_op = Constant(
Type.f32, Shape(list(tensor.axes.lengths)),
list(self.flatten(tensor.const.tolist())))
self.computation.register_cpp_op(tensor, constant_op)
else:
op_element_type = Parameter(Type.f32,
Shape(list(tensor.axes.lengths)))
self.computation.register_cpp_op(tensor, op_element_type)
if not tensor.is_placeholder:
self.computation.neon_variable_list.append(tensor)
def test_constant():
element_type = Type.f32
parameter_list = []
function = Function(NodeVector([Constant(element_type, Shape([3, 3]), list(range(9)))]),
parameter_list, 'test')
backend = Backend.create(pytest.config.getoption('backend'))
result = backend.create_tensor(element_type, Shape([3, 3]))
result_arr = np.zeros((3, 3), dtype=np.float32)
result.write(util.numpy_to_c(result_arr), 0, 36)
backend.call(backend.compile(function), [result], [])
result.read(util.numpy_to_c(result_arr), 0, 36)
result_arr_ref = np.arange(9).reshape(3, 3)
assert np.allclose(result_arr, result_arr_ref)
def test_constant():
element_type = Type.f32
parameter_list = []
function = Function(NodeVector([Constant(element_type, Shape([3, 3]), list(range(9)))]),
parameter_list, 'test')
backend, cf = make_backend_call_frame(function)
result = backend.make_primary_tensor_view(element_type, Shape([3, 3]))
result_arr = np.zeros((3, 3), dtype=np.float32)
result.write(util.numpy_to_c(result_arr), 0, 36)
cf.call([], [result])
result.read(util.numpy_to_c(result_arr), 0, 36)
result_arr_ref = np.arange(9).reshape(3, 3)
assert np.allclose(result_arr, result_arr_ref)
def visit(self, op, tensor):
self.computation.set_op_rank(op)
variable = tensor.tensor
list_size = 1
for x in list(tensor.axes.lengths):
list_size *= x
constant_list = [op.scalar] * list_size
ngraph_constant_op = Constant(Type.f32,
Shape(list(tensor.axes.lengths)),
constant_list)
if variable not in self.computation.variables_cpp_op:
# treat 'op' as the rhs of assignment for forwarding and lookup purposes
self.computation.variables_cpp_op[variable] = \
(self.computation.scopemark[op.tensor], op)
self.computation.register_cpp_op(op,
ngraph_constant_op,
set_name=False)
else:
raise RuntimeError("Variable updated more than once!")
def test_constant():
element_type = Type.f32
parameter_list = []
function = Function(
[Constant(element_type, Shape([3, 3]), list(range(9)))],
parameter_list, 'test')
backend = Backend.create(test.BACKEND_NAME)
result = backend.create_tensor(element_type, Shape([3, 3]))
result_arr = np.zeros((3, 3), dtype=np.float32)
result.write(util.numpy_to_c(result_arr), 36)
handle = backend.compile(function)
handle.call([result], [])
result.read(util.numpy_to_c(result_arr), 36)
result_arr_ref = np.arange(9).reshape(3, 3)
assert np.allclose(result_arr, result_arr_ref)
