Skip to content

PARAMETER_PARITY_TODO.md

Latest commit

 

History

History
61 lines (49 loc) · 3.45 KB

PARAMETER_PARITY_TODO.md

File metadata and controls

61 lines (49 loc) · 3.45 KB
Raw

Parameter Parity TODO — franken_numpy-ksr

Status Legend

  • Not started
  • [~] In progress
  • Done
  • [D] Parity debt/deferred — observable NumPy behavior not yet closed
  • [A] Architecture note — no current behavior gap unless future evidence proves one

GROUP 1: unique(axis=) — Multi-dimensional uniqueness

  • Implement unique_axis(&self, axis: isize) on UFuncArray
  • Handle axis parameter: extract slices along axis, compare them, deduplicate
  • Return unique slices preserving sorted order
  • Add tests: 2D array unique rows, unique columns

GROUP 2: concatenate/stack dtype+casting parameters

  • Add concatenate_with_dtype(arrays, axis, dtype, casting) method
  • Validate casting via fnp_dtype::can_cast for each input → output dtype
  • Cast each input array to target dtype before concatenation
  • Add stack_with_dtype(arrays, axis, dtype, casting) variant
  • Add tests: concatenate with dtype promotion, casting rejection

GROUP 3: sum/prod initial parameter

  • Add reduce_sum_initial(&self, axis, keepdims, initial: f64) method
  • Add reduce_prod_initial(&self, axis, keepdims, initial: f64) method
  • Initial value added to sum / multiplied into product
  • Add tests: sum with initial, prod with initial

GROUP 4: sum/prod/mean/var/std dtype parameter

  • [D] Deferred parity debt — NumPy exposes dtype= as observable API behavior even when f64 accumulation is internally precise
  • [D] Python FFI explicit dtype= keyword conformance is covered for sum, prod, mean, std, and var; remaining follow-up must define Rust-core accumulator/result dtype semantics for integer, unsigned, float, bool, and complex reductions before this gap can be marked closed

GROUP 5: percentile/quantile keepdims parameter

  • Add percentile_keepdims method
  • Add quantile_keepdims method
  • When keepdims=true, insert size-1 dimension at reduced axis
  • Handle axis=None case (all dims become size 1)
  • Add tests: axis keepdims, None axis keepdims

GROUP 6: partition/argpartition kind parameter

  • [A] NumPy currently exposes one accepted algorithm token (introselect); parity work should still verify accepted/invalid token behavior and error text

GROUP 7: unique return_index/inverse/counts with axis

  • Add unique_axis_with_info(axis, return_index, return_inverse, return_counts) on UFuncArray
  • Return NumPy-shaped return_index, return_inverse, and return_counts metadata for unique axis slices
  • Cover large-integer sidecars, signed-zero representatives, and NaN slice distinctness

ALREADY DONE (earlier this session)

  • clip_optional (one-sided clipping with None min/max)
  • reshape_order (F-order reshape)
  • histogram_full (density, range, weights parameters)

KNOWN PARITY DEBT / ARCHITECTURE NOTES

  • [D] out parameter on array functions — Rust internals are immutable, but Python/NumPy-facing APIs still expose in-place output semantics that need explicit parity handling or documented wrapper delegation
  • [D] order on sort/partition — structured/object-array ordering remains observable NumPy behavior and should stay tracked as parity debt
  • [D] overwrite_input on percentile/quantile — immutable internals do not remove the need to match accepted parameter behavior and error surfaces
  • [D] dtype on reductions — f64 accumulation is an implementation detail; NumPy result dtype and accumulator dtype behavior remain parity debt