wegamekinglc / Derivatives-Algorithms-Lib / 28288380668

81%

master: 81%

Build Type

Pull #152

github

Committed by wegamekinglc

Commit Message

Fix AAD tape Clear regression: clear adjointsMulti_ unconditionally

Copilot on PR #152 correctly flagged that the ForEachBlock refactor changed
Clear's semantics: the original native-branch Clear cleared adjointsMulti_
unconditionally, but ForEachBlock guarded it on Tape_::multi_ (a runtime flag
toggled by SetNumResultsForAAD). A tape that ran multi (populating
adjointsMulti_), then toggled to non-multi, would retain stale adjoints on
Clear instead of wiping them.

- Split ForEachBlock (guarded, for Mark/Rewind/RewindToMark) from
  ForEachBlockAll (unconditional, for Clear) so each function preserves its
  original adjointsMulti_ clearing semantics. Mark/Rewind/RewindToMark were
  already guarded in the original and stay guarded.
- Add a one-line comment at ForEachBlockAll explaining why adjointsMulti_ is
  cleared unconditionally (so the next reader/Copilot sees it is intentional).
- Add regression test TestClearEmptiesAdjointsMultiAfterMultiToNonMultiToggle
  (native-gated) that toggles multi->non-multi and asserts adjointsMulti_ is
  emptied by Clear. Verified it fails against the buggy ForEachBlock-only form
  (8 stale adjoints retained) and passes with the fix.
- Add // namespace Dal closing-brace comments to fdm1dmesher.hpp and
  concentrating1dmesher.cpp (style item, files were touched by this PR).

Native: 753/753. XAD: 752/752 (regression test compiled out).

Pull Request Pull Request #152: chore: dedup math structural (AAD iterators, tape loop, Sobol seek, mesher epilogue)

Coverage Stats

24 of 26 new or added lines in 4 files covered. (92.31%)

122 existing lines in 8 files now uncovered.

6293 of 7812 relevant lines covered (80.56%)

3223299.49 hits per line