Merge #1175

1175: xlogy broadcast adjoint r=MikeInnes a=MikeInnes

This is helpful for performance, since it avoids having to differentiate `xlogy` itself inside of a map.

Co-authored-by: Mike J Innes <mike.j.innes@gmail.com>

src/layers/stateless.jl

@@ -288,3 +288,8 @@ CuArrays.@cufunc function xlogy(x, y)
   result = x * log(y)
   ifelse(iszero(x), zero(result), result)
 end
+
+@adjoint function broadcasted(::typeof(xlogy), x::Zygote.Numeric, y::Zygote.Numeric)
+  res = xlogy.(x, y)
+  res, Δ -> (nothing, Zygote.unbroadcast(x, xlogy.(Δ, y)), Zygote.unbroadcast(y, Δ .* x ./ y))
+end