stillwater-sc / universal / 25007204123

84%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

feat(lns): Mitchell-style LookupAddSub algorithm (Phase B of #777) (#785)

* feat(lns): Mitchell-style LookupAddSub algorithm (Phase B of #777)

Add the second algorithm in the configurable lns add/sub framework:
a Mitchell-style precomputed-table policy with linear interpolation.

For ML inference accelerators and embedded DSP, this is the SRAM-friendly
hardware default -- one indexed memory read + one mul-add per operation,
no transcendentals at runtime. This was the original Mitchell 1962 LNS
formulation and remains the gold standard for hardware LNS.

LookupAddSub<Lns, IndexBits>
----------------------------
- Compile-time table generation via sw::math::constexpr_math::log2/exp2
  (Epic #763) -- no offline tooling needed, table follows the
  instantiation, a different (Lns, IndexBits) gets its own table.
- IndexBits parameter (log2 of table entries). Default
  min(Lns::rbits + 2, 10) caps at 1024 entries so a 64-bit lns doesn't
  try to allocate millions of entries; users override for higher
  accuracy or smaller SRAM.
- d-range chosen as Lns::rbits + 2: past that, the correction is below
  lns ULP and rounds to 0, so storing those entries gains nothing.
- sb_sub special-cases the |d| < step cell: linear interp on
  log2(1 - 2^d) with unbounded slope at d=0 is unsafe, so the
  cancellation-regime cell falls back to direct cm::log2/cm::exp2
  evaluation. This costs one transcendental pair but only in that
  narrow band, and preserves accuracy where it matters most.
- log_add dispatcher (special values, sign routing) is structurally
  identical to DirectEvaluationAddSub; duplicated for self-containment
  rather than refactored, since the dispatcher is small and the
  algorithm files are meant to be read independently.

Regression extensions (log_add_algorithms.cpp)
----------------------------------------------
- VerifyAddAlgorithmsAgree / VerifySubAlgorithmsAgree generalized to
  take the two policies as template parameters. Phase A's 1-arg form
  ... (continued)

Coverage Stats

57 of 57 new or added lines in 1 file covered. (100.0%)

21 existing lines in 3 files now uncovered.

45544 of 54021 relevant lines covered (84.31%)

6353101.3 hits per line