You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Summary:
ConstantInputSpectrumStress used a fixed 1e-3 absolute bound for theoretically-zero non-DC components. A deterministic 20x23 constant-255 transform in explicit float produces 0.005859375 residual (4.9952e-8 of its 117300 DC magnitude), proving the fixed bound rejects ordinary single-precision roundoff.
The test now bounds leakage by max(1, abs(expectedDC)) * NumericT<T>::eps(): a precision-derived 1e-6 relative budget for float and 1e-12 for double. This is 100x stricter than the abandoned D111055971 proposal of 1e-4, preserves the only-DC invariant, and strengthens low-magnitude and double validation. Existing DC real and imaginary checks are unchanged. An explicit-float 20x23 regression covers the affected precision on double-default hosts.
___
Differential Revision: D115898011
fbshipit-source-id: 48c7a05b78114e237626a9bfe34d8bd4b3d4094c
0 commit comments