add_inplace_op

→TensorImplPtr

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

source edit

In-place elementwise addition: writes $a + b$ back into a.

Equivalent to a = add_op(a, b) but reuses a's storage and bumps its version counter. b may broadcast up to a's shape but the broadcast output shape must equal a->shape().

Math

$a_i \leftarrow a_i + b_i$

See Also

add_op — out-of-place counterpart.

Parameters

aTensorImplPtr

Left operand and in-place target. Must be non-null and must not share storage with a view.

bTensorImplPtr

Right operand. Must be non-null and have a shape broadcastable to a->shape() such that the broadcast result equals a->shape().

Returns

TensorImplPtr

The same pointer as a, now holding the summed storage.

Raises

ShapeMismatch

If the broadcast output of add_op(a, b) does not equal a->shape() (i.e. a would have to grow to fit the result).

RuntimeError

If a shares storage with a view (call .clone() first).