flopscope.accounting.unwrap_cost

shape:tuple[int, ...]

Shape of the array passed to the analytical cost formula.

:int

Weighted public cost estimate, floored to match runtime accounting.

This helper multiplies the analytical FLOP count by the active weight from flopscope._weights and then applies int(...) so public estimates match budget deductions.

NumPy's unwrap (flops.lib._function_base_impl.unwrap) performs 13 one-FLOP ufunc passes over the N-1 element dd = diff(p) array, then one final add of the cumulative correction into the output. Counting each named ufunc call as one pass:

1. diff — subtract adjacent elements (N-1 elements)

2. dd - low — subtract scalar interval_low from dd

3. mod(..., period)— elementwise modulo

4. + interval_low — add scalar back

5. ddmod == low — elementwise compare (boundary check)

6. dd > 0 — elementwise compare

7. ``& `` — bitwise-and of two bool arrays

   system_message

   <string>:14: (WARNING/2) Inline literal start-string without end-string.

8. copyto/select — conditional write (boundary fix), 3-arg where

9. ddmod - dd — elementwise subtract (ph_correct)

10. abs(dd) — elementwise absolute value

11. < discont — elementwise compare

12. copyto/select — conditional write (small-jump zeroing), 3-arg where

13. cumsum — prefix-sum scan

The p[slice1] + ph_correct.cumsum(axis) expression (final output materialization) involves one add pass but is treated as part of the output-write cost — following the convention that the output buffer fill is attributed to the issuing op. Charging all 13 op-passes against numel(input) rather than N-1 avoids tracking the edge-element correction (one extra element) and gives a clean formula.

Steps 8 and 12 (copyto/select via 3-arg where) were treated as free data movement (selecting by a given mask) from the where-arity branching (Task 4, 2026-06-15) through the select-class rework (Task 6, 2026-07-17), which discounted this formula to 11 * numel for that window. Task 6 also fixed 3-arg where itself to charge 4 * numel(broadcast output) (it derives no selector, but does scan and write every output element), so the discount those two steps relied on no longer holds; this formula reverts to charging all 13 passes, matching the original value from Audit-completion Task 4 (2026-06-12).