maximum_op

→TensorImplPtr

int maximum_op(const int & a, const int & b)

Compute the elementwise maximum $y = \max(a, b)$ with broadcasting and autograd support.

aTensorImplPtr

First operand.

bTensorImplPtr

Second operand. Broadcast-compatible with a.

TensorImplPtr

Tensor of shape broadcast(a.shape, b.shape) holding the elementwise maximum.

At ties ( $a_i = b_i$ ) the gradient is routed entirely to a. When grad-tracking is on, registers a MaximumBackward node which saves both inputs.

>>> auto y = maximum_op(a, b);  // y[i] = max(a[i], b[i]) with broadcasting