Derivative lost in 3-arg mul

[antoine-levitt]

julia> a = randn(3,3)
b =3×3 Matrix{Float64}:
 ra  0.0237017   1.62282  -1.28889
 -0.186575   -2.00309   0.502572
 -0.413123   -1.99718  -0.174593

julia> b = randn(3,3)
3×3 Matrix{Float64}:
 -0.40664   1.80043     1.2399
  1.22559  -0.828653    1.38103
  1.40994  -0.0784539  -0.127629

julia> ForwardDiff.derivative(x -> sum(x*a*b), 0.0)
0.0

julia> ForwardDiff.derivative(x -> sum(x*(a*b)), 0.0)
-6.900806270529478

[antoine-levitt]

This is present on 1.7beta and master, but not on 1.6

[antoine-levitt]

@mcabbott shot in the dark, is this related to JuliaLang/julia#37898?

[mcabbott]

That's not good! Yes it's caused by the PR, but I don't know if that's where the bug is. This becomes:

julia> ForwardDiff.derivative(x -> sum(LinearAlgebra.mat_mat_scalar(a, b, x)), 0.0)
0.0

julia> ForwardDiff.derivative(x -> sum(LinearAlgebra.mat_mat_scalar(a, b, x)), 2.0)
0.15095195538098594

julia> ForwardDiff.derivative(0.0) do x
         c = similar(a, typeof(@show x))
         mul!(c, a, b, x, false)
         @show sum(c)
       end
x = Dual{ForwardDiff.Tag{var"#27#28", Float64}}(0.0,1.0)
sum(c) = Dual{ForwardDiff.Tag{var"#27#28", Float64}}(0.0,0.0)
0.0

The explicit mul! form gives zero on Julia 1.5 too. It calls generic_matmatmul!(C, 'N', 'N', A, B, MulAddMul(alpha, beta)) which certainly sounds like it should work. But perhaps it's testing alpha==0, in which case this is one more bug which would be fixed by #481 .

[antoine-levitt]

That makes sense. So it's just missing somebody to merge #481?