ttnn.bcast

ttnn.bcast() → None

Perform a binary elementwise operation math_op between tensors input_a and input_b, where values from tensor input_b are broadcast.

Let tensor input_a have shape [W0, Z0, Y0, X0] and tensor input_b shape [W1, Z1, Y1, X1]. dim determines the type of broadcast performed.

For dim=BcastOpDim::W broadcast is performed on dimension X. Y0 and Y1 must be the same and either (W1=1 and Z1=1) or (W0=W1 and Z0=Z1).

For dim=BcastOpDim::H broadcast is performed on dimension Y. X0 and X1 must be the same and either (W1=1 and Z1=1) or (W0=W1 and Z0=Z1).

For dim=BcastOpDim::HW broadcast is performed on dimensions X and Y. Either (W1=1 and Z1=1) or (W0=W1 and Z0=Z1) must hold for input shapes.

Both input tensors must have BFLOAT16 data type.

Output tensor will have BFLOAT16 data type.

Argument	Description	Data type	Valid range	Required
input_a	Input tensor	Tensor	Tensor of shape [W0, Z0, Y0, X0]	Yes
input_b	Input tensor to broadcast	Tensor	Tensor of shape [W1, Z1, Y1, X1]	Yes
math_op	Aggregating math operation	BcastOpMath	ADD, SUB, MUL	Yes
dim	Dimension on which to broadcast	BcastOpDim	W, H, HW	Yes
memory_config	Layout of tensor in TT Accelerator device memory banks	MemoryConfig	Default is interleaved in DRAM	No
output_tensor	Optional preallocated output tensor	Tensor	Default is None	No

:param * input_tensor_a: First Input Tensor for bcast. :param * input_tensor_b: Second Input Tensor for bcast. :param * math_op: Operation to be performed during broadcasting. :param * dim: the dimension to reduce. If None, the bcast of the flattened input is returned

:keyword * memory_config: Memory Config of the output tensor :keyword * output_tensor: Preallocated output tensor