floordiv_op

→TensorImplPtr

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

source edit

Compute the elementwise floor division $y = \lfloor a / b \rfloor$ .

Floor division is non-differentiable (piecewise constant with jumps at every integer crossing), so no autograd node is registered: the returned tensor never carries a grad_fn and the result effectively detaches its inputs from the graph.

Math

y_i = \left\lfloor \frac{a_i}{b_i} \right\rfloor

Shape

a.shape == b.shape. The result is the same shape; the dtype is always I64.

See Also

div_op : True (real-valued) elementwise division, fully differentiable.

Parameters

aTensorImplPtr

Numerator tensor.

bTensorImplPtr

Denominator tensor. Must have the same shape and dtype as a — no broadcasting and no implicit dtype promotion is performed.

Returns

TensorImplPtr

Tensor of shape a.shape and dtype Dtype::I64 holding $\lfloor a / b \rfloor$ . The cast to I64 is unconditional regardless of the input dtype, matching Python's // operator semantics.

Raises

LucidError

If a.shape != b.shape or a.dtype != b.dtype.

Notes

Behaviour on division by zero is delegated to the backend primitive (backend.floordiv). Because the output is forced to I64, any floating-point NaN / Inf produced by the backend is reinterpreted to a platform-defined integer value.