Crate vecnorm_core

Expand description

Pure-Rust core for vecnorm. Bulk f32 matrix operations:

l2_normalize / l2_normalize_copy — row-wise unit-length scaling. Rows whose norm is below EPS are left at zero rather than dividing by zero.
cosine_similarity — single pair on 1-D vectors. Returns 0 for any pair where either side has zero norm.
top_k_argmax / batch_top_k_argmax — partial-heap top-k that runs in O(n log k). Tied scores are broken by the original index ascending (deterministic).

Enums§

argmax: Single argmax: returns (index, score) of the largest element. Ties broken by ascending index. Errors on empty input.
batch_top_k_argmax: Batch top-k argmax over an (n_rows, n_cols) matrix. With parallel = true distributes rows across rayon’s pool.
cosine_distances: Cosine distance matrix between two (n_a, d) and (n_b, d) matrices. Returns an (n_a, n_b) matrix where out[i, j] is the cosine distance 1 - cos(a_i, b_j). Inputs are not modified; this normalizes copies internally so accuracy is preserved on un-normalized inputs.
cosine_similarity: Cosine similarity between two 1-D vectors. Returns 0 if either side is all-zero.
dot_product: Inner product (dot product) of two 1-D vectors. No normalization. Errors on dim mismatch.
l2_normalize: L2-normalize matrix in place, row by row. Rows with norm below EPS are zeroed out (i.e. left unchanged at all-zero) to avoid NaN.
l2_normalize_copy: L2-normalize a copy. Same semantics as l2_normalize.
top_k_argmax: Top-k argmax over a 1-D score vector. Returns (index, score) pairs in descending order. Ties broken by ascending index.