matmul_ozaki — DRIVE ozaki op

• Data kinds: matrix × matrix → matrix

• Call: import fullseye as fs; fs.ledger.matmul_ozaki(A, B, n_slices=None, scheme: 'str' = 'ozaki1', tol: 'float' = 1.1102230246251565e-16, on_insufficient: 'str' = 'raise', return_info: 'bool' = False) (to call the implementation directly, import ozakimm; ozakimm.matmul_ozaki(A, B, n_slices=None, scheme: 'str' = 'ozaki1', tol: 'float' = 1.1102230246251565e-16, on_insufficient: 'str' = 'raise', return_info: 'bool' = False); from the registry, opsdrive.get("matmul_ozaki"))

Usage

Builds an FP64-accurate matrix product A @ B from exact low-precision partial products (Ozaki-I / Ozaki-II).

scheme="ozaki1": splits into n_slices slices (n(n+1)/2 products). If n_slices=None, it automatically chooses the smallest number of slices (1–16) for which the guaranteed error upper bound is at most tol · (|A||B|)_ij. If that cannot be reached, on_insufficient="raise" gives ValueError, and "fallback" returns an ordinary float64 product (only in this case is the reproducibility promise dropped; visible from return_info=True in fallback). scheme="ozaki2": the number of moduli must always be given by n_slices (2–20) (n products). There is no automatic selection. With return_info=True, (C, info) (n_slices, n_products, scheme, fallback).

With either scheme, the same input gives the same bits regardless of inner reordering or the number of BLAS threads. On CPU it is 1–2 orders of magnitude slower than DGEMM (see the module docstring for when to use which).

References (sample data, literature)

• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).

• Operator provenance and references — the sources of the research/methods this op family came from.

• The canonical algorithm (author, year) and its uses are named in the family usage guide above.

Runnable examples (verified samples that actually call this op)

• poc_reproducible_icp — py -3.11 examples/poc_reproducible_icp.py

Ops the type connects to (they accept matrix as input)

world_camera · lidar_scan · relative_motion · foe_from_motion · flow_from_depth_motion · ttc_truth · world_materials · mesh_signed_volume

Same category (ozaki)

matmul_reproducible · ozaki_error_bound · cross_covariance_reproducible · kabsch_reproducible · fp64_emulation_probe

*Provenance: ozakimm.py — DRIVE operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*

© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.