running autodiff twice leads to bad result when result vector is preallocated

[hiemstar]

I went through the documentation and examples and could not find the reason why the MWE below does not work.

I would expect that it is possible to pre-allocate the input and result arrays that are used in autodiff and
run autodiff on different values. Hoever, on the second run the results are not updated.

using Enzyme, Test

struct Gradient
    f::Function
    bx::Vector{Float64}
    y::Vector{Float64}
    by::Vector{Float64}
end

function (G::Gradient)(x::Vector{Float64})
    Enzyme.autodiff(Reverse, G.f, Duplicated(x, G.bx), Duplicated(G.y, G.by));
    return G.bx
end

function f(x::Array{Float64}, y::Array{Float64})
    y[1] = x[1] * x[1] + x[2] * x[1]
    return nothing
end;

bx = [0.0, 0.0]
y  = [0.0]
by = [1.0];

G = Gradient(f, bx, y, by)

x  = [2.0, 2.0];
@test G(x) == [6.0, 2.0] # test passes

x  = [3.0, 2.0];
@test G(x) == [8.0, 3.0] # test fails. Evaluated: [6.0, 2.0] == [8.0, 3.0]

[wsmoses]

You haven't reset dy = 1 and dx = 0 between invocations.

Reverse mode will += the derivative results in, and similarly 0 out the derivative output when it propagates (this is required to get correct behavior inside a loop -- if you were to run the same y[1] = ... code twice in the same function to be differentiated, it would be the same result rather than adding the result twice.

[hiemstar]

Got it. Thanks for the quick help!