Repository navigation
Decision needed: opt-in fastmath for TypedPython #153
Description
Float reductions do not vectorise because TypedPython builds with -ffp-contract=off -fno-fast-math for bit-identical results; numba fastmath=True and Cython -ffast-math are opt-in there.
Options:
- (a) no fastmath (identity everywhere);
- (b) an explicit per-function opt-in (e.g. @compiled(fastmath=True)), documented as not CPython-identical, never on by default;
- (c) only exact transforms (no reassociation), i.e. no change.
Recommendation: (a) for now. Integer and element-wise loops already vectorise; revisit (b) with measurements once #41 items land. Waiting for the user.
Activity
b-re-w commentedon Oct 25, 2026
Decision (user, via the coordinator): (a) no fastmath. TypedPython keeps -ffp-contract=off -fno-fast-math for bit-identical results everywhere, by default and with no switch today. Float reductions stay sequential; integer and element-wise loops already vectorise. A per-function opt-in (option b) is recorded as a separate proposal, #332, not implemented.
b-re-w commentedon Oct 25, 2026
Closed with decision (a); opt-in proposal tracked in #332.