searchsorted_op

→TensorImplPtr

int searchsorted_op(const int & sorted_1d, const int & values, bool right)

source edit

Python wrappers

pylucid.ops.searchsortedfunction

Per-query insertion point into a sorted 1-D reference.

Composite over reshape_op + broadcast_to_op + (less_op | less_equal_op) + astype_op + sum_op. The 1-D reference is broadcast to [1, ..., 1, N] and the queries to S + [1]; both expand to the full S + [N] grid, the comparison is reduced over the trailing axis, and the bool $\to$ F32 $\to$ I64 cast chain produces the standard searchsorted dtype. Not differentiable — integer output.

Math

Let $s$ be the sorted reference and $v$ the query tensor. Then

y_{\ldots, j} = \sum_{i=0}^{N-1} \mathbb{1}\!\left[ s_i \prec v_{\ldots, j} \right]

with $\prec$ being $<$ when right = false (leftmost insertion) and $\le$ when right = true (rightmost insertion).

See Also

bucketize_op — same operation with the argument order flipped.

Parameters

sorted_1dTensorImplPtr

1-D ascending-sorted reference of length

N

valuesTensorImplPtr

Query tensor of arbitrary shape S.

rightbool

If true, returns the rightmost (post-equal) insertion point; otherwise the leftmost.

Returns

TensorImplPtr

int64 tensor of shape S whose entries lie in $[0, N]$ .

Raises

Failure

If either input is null or sorted_1d is not 1-D.

Notes

The reduction kernel only supports float dtypes; the bool comparison is cast to F32 for the sum, then back to I64 to match the canonical searchsorted dtype.