Overview

Indexing routines is NumSharp's largest partial gap: 10 of 35 implemented. This tracks the 25 missing functions, grouped into four work clusters. It is dominated by np.diag -- ~115K GitHub hits, one of the most-used NumPy functions overall and absent today (NumSharp has diagonal for extraction but not the diag constructor/dual).

Problem

The gap spans four very different kinds of work -- pure index arithmetic (diagonals/triangles), a slice-expression DSL (r_/c_/s_/ix_/ogrid), iterator objects, and along-axis selection. Splitting by cluster lets the cheap, high-value pieces (esp. diag) ship first.

Verified codebase state

• Implemented (np.*): take, put, compress, place, where, nonzero, diagonal, indices, ravel_multi_index, unravel_index (+ mgrid exists).
• Missing (25 below): verified absent from np.* (including static properties, so c_/r_/s_/ogrid are confirmed missing) and as NDArray methods (reflection + source grep).
• Nuance: the internal multi-operand iterator (public static class NDIter, Backends/Iterators/NDIter.cs) already exists -- so the nditer/ndindex/ndenumerate cluster is mostly a public wrapper, not new machinery.

Proposal

Add under np.*, cluster by cluster. Recommended order: A -> (D, C) -> B, leading with diag.

A. Diagonal & triangle index-math -- pure index arithmetic, no new kernels (do first)

• diag (114,816) -- extract a diagonal or construct a diagonal matrix (the dual of diagonal)
• fill_diagonal (26,032) -- write the main diagonal of an array in place
• triu_indices (8,944) -- upper-triangle indices of an (n,m) array
• triu_indices_from (4,600) -- upper-triangle indices for a given array
• diag_indices (4,340) -- indices to access the main diagonal
• tril_indices (3,800) -- lower-triangle indices of an (n,m) array
• diag_indices_from (3,488) -- main-diagonal indices for a given array
• tril_indices_from (3,140) -- lower-triangle indices for a given array
• mask_indices (2,156) -- indices from a generic triangle-mask function

D. Along-axis selection -- selection kernels (build on take/put/place)

• select (23,040) -- piecewise pick from choice arrays by a condition list
• take_along_axis (10,584) -- gather by matching 1-D index and data slices
• choose (9,416) -- build an array from an index array + a list of choice arrays
• putmask (3,512) -- masked in-place write (broadcasting values)
• put_along_axis (3,088) -- scatter by matching 1-D index and data slices

C. Iterator objects -- mostly surfacing the existing NDIter engine

• nditer (42,416) -- multi-operand iterator object -- internal NDIter engine already exists (Backends/Iterators/NDIter.cs)
• ndindex (20,160) -- N-dimensional index iterator
• ndenumerate (8,288) -- N-dimensional (index, value) iterator
• iterable (5,488) -- predicate: is the object iterable (trivial)
• nested_iters (1,664) -- nested multi-operand iterators
• flatiter (4,800) -- flat-iterator type behind .flat

B. Grid / slice-expression DSL -- needs C# indexer sugar

• r_ (44,016) -- row-wise concatenation / build from slice objects
• c_ (35,600) -- column-wise concatenation / build from slice objects
• ix_ (32,544) -- open-mesh index from multiple sequences
• s_ (19,104) -- slice-literal builder for index tuples
• ogrid (6,888) -- open (non-fleshed) meshgrid (mgrid already exists in NumSharp)

Evidence -- GitHub usage ranking

Combined usage ~ 441,924 hits -- the largest of the three sweep issues, and diag (~114,816) is the single most-used missing function in the entire sweep.

Usage = GitHub REST code-search total_count in each function's canonical call form (np.<fn>(, or np.<fn>[ for the index-expression objects). It is an approximate estimate (noisy plus/minus ~10% on large sets) and counts only the dominant np. alias. Measured 2026-07-19; treat as relative demand, not exact.

Scope / Non-goals

• np.matrix-returning behavior is out of scope (NumSharp has no matrix subclass; np.matlib is deprecated).
• nditer should expose the existing NDIter engine, not a re-implementation; full NumPy op_flags/buffering parity can be staged.
• The DSL objects (r_/c_/s_) require C# indexer ergonomics -- acceptable to expose a close-but-not-identical surface where C# syntax forces it (document divergences).

Related issues

• Part of the NumPy category-coverage sweep; sibling issues: Set routines: implement np.* set operations (isin, intersect1d, union1d, setdiff1d, setxor1d, unique_*) #622 (set) and Sorting & searching: implement partition/argpartition, lexsort, nanargmax/nanargmin, sort_complex #623 (sort). Umbrella / methodology: LinearAlgebra Missing Support #621 (linalg).