[Lang] Add qd.math.fma(...) single-rounding fused multiply-add#478

Open

duburcqa wants to merge 39 commits into

duburcqa/qd_precisefrom

duburcqa/qd_math_fma

Commits on Apr 13, 2026

[Lang] Add qd.precise(...) for per-op IEEE-strict FP
duburcqa
committed
[Lang] qd.precise: cover UnaryOpStmt as well
duburcqa
committed
[Lang] qd.precise: address self-review feedback
duburcqa
committed
[Lang] qd.precise: gate alg_simp folds, cover sqrt, DRY CUDA libdevice
duburcqa
committed
[Lang] qd.precise: scrub non-ASCII from comments
duburcqa
committed
[Lang] qd.precise: replace -- with single - in comments
duburcqa
committed
[Doc] User guide entry for qd.precise
duburcqa
committed
[Lang] qd.precise: factor disable_fast_math helper, add Vector/select tests
duburcqa
committed
[Lang] qd.precise: propagate tag in 2*a rewrite, narrow zero-fold gate, refresh test_api
duburcqa
committed
[Lang] qd.precise: use make_typed to avoid downcast on synthesized 2*a stmt
duburcqa
committed
Cleanup doc.
duburcqa
committed
[Lang] qd.precise: cover walker boundaries (qd.func, bit_cast, alias, idempotent)
duburcqa
committed
[Lang] qd.precise: fix docstring to mention unary FP ops and approximate transcendentals
duburcqa
committed
[Lang] qd.precise: unify precise field comments via canonical reference in ir/expr.h
duburcqa
committed
[Lang] qd.precise: propagate tag through synthesized stmts in alg_simp / demote_operations
duburcqa
committed
[Lang] qd.precise: clear LLVM FMF on intermediate and pre-FPTrunc values
duburcqa
committed
[Lang] qd.precise: SPIR-V inv forwards precise, inline maybe_no_contraction, reformat QD_NOT_IMPLEMENTED block
duburcqa
committed
[Lang] qd.precise: drop bit-ops-on-FP from doc; align __all__ position; 'moral equivalent' -> 'equivalent'
duburcqa
committed
[Lang] qd.precise: clone input subtree instead of mutating in-place; update doc and tests
duburcqa
committed
[Lang] qd.precise: parametrize unary rounding test per op for per-op pass/fail reporting
duburcqa
committed
[Lang] qd.precise: SPIR-V visit(BinaryOpStmt) tags FP transcendental (atan2/pow) results with NoContraction
duburcqa
committed
[Lang] qd.precise: reflow PR-introduced C++ comments to 120 cols
duburcqa
committed
[Lang] qd.precise: propagate tag through cast in 2*a rewrite (and reflow to 120)
duburcqa
committed
[Lang] qd.precise: CUDA emit_extra_unary clears FMF on libdevice call before f16 FPTrunc
duburcqa
committed
[Lang] qd.precise: skip sin/cos unary-rounding on SPIR-V, drop redundant mutation test, fix pow NOTE
duburcqa
committed
[Lang] qd.precise: unary-rounding test restricts to LLVM via arch decorator; reframe as contract check
duburcqa
committed
[Lang] qd.precise: type_check propagates tag through implicit operand-promotion casts
duburcqa
committed
[Lang] qd.precise: document SPIR-V arithmetic/post-hoc two-layer decoration in visit(BinaryOpStmt)
duburcqa
committed
[Lang] qd.precise: scalarize propagates tag onto per-element scalar Binary/Unary stmts
duburcqa
committed
[Lang] qd.precise: SPIR-V decorates FP ops once via post-hoc block; drop duplicate builder-side precise
duburcqa
committed
[Lang] qd.precise: idempotency test also covers AMDGPU (also an LLVM backend)
duburcqa
committed
[Lang] qd.precise: AMDGPU i32 pow clears FMF on __ocml_pow_f64 call before FPToSI
duburcqa
committed
[Lang] qd.precise: exclude cmp_gt/cmp_lt from precise guard (IEEE-false under NaN)
duburcqa
committed
[Lang] qd.precise: iterative worklist in clone_and_tag_precise (O(1) C++ stack depth)
duburcqa
committed
[Lang] qd.precise: precise_fp_add requires FP operand type; integer a+0 folds unconditionally
duburcqa
committed