Repository navigation
fma and muladd ignore NaN-safe mode #860
Description
With NaN-safe mode enabled (nansafe_mode = true), fma and muladd propagate NaN partials for dual numbers with zero partials, unlike the equivalent x * y + z:
julia> using ForwardDiff: Dual # nansafe_mode = true
julia> x, y, z = Dual(Inf, 0.0), Dual(2.0, 0.0), Dual(0.0, 0.0);
julia> x * y + z
Dual{Nothing}(Inf,0.0)
julia> muladd(x, y, z)
Dual{Nothing}(Inf,NaN)
julia> fma(x, y, z)
Dual{Nothing}(Inf,NaN)
julia> x, y, z = Dual(1.0, 0.0), Inf, Dual(0.0, 1.0);
julia> x * y + z
Dual{Nothing}(Inf,1.0)
julia> muladd(x, y, z)
Dual{Nothing}(Inf,NaN)
julia> fma(x, y, z)
Dual{Nothing}(Inf,NaN)The methods for three dual arguments, and for dual-real-dual arguments, compute the partials with fma/muladd directly instead of going through the NaN-safe multiplication. Related: #857, #858.
ForwardDiff master (a3c0f4f), Julia 1.13.1, macOS aarch64.
Activity