2 ms·
There are a few critical algorithms where fp error cancellation or simple ifs get optimized out if you disable signed zeros. Typically you would know which thes
by AstralStorm 5y ago
There are a few critical algorithms where fp error cancellation or simple ifs get optimized out if you disable signed zeros. Typically you would know which these are, they tend to appear in statistical machine learning which use sign or expect monotonicity near zero, or filters with coefficients that are near zero (and filter out NaNs explicitly).
- CamperBob2 5y agoWhat would be an example of a filter with coefficients near zero that would be adversely affected by the loss of signed-zero support? You're already in mortal peril if you're working with "coefficients near zero" because of denormals, another bad idea that should have been disabled by default and turned on only in the vanishingly-few applications that benefit from them.