Valentin Churavy
Valentin Churavy
@wsmoses from our last chat the thing left to do is to "pseudo-zero" the shadow sret?
@gaspardbb does this still reproduce for you?
So in #245 I would evaluate the augmented primal immediatly and then return the gradient thunk to the user. It might generally be more flexible to provide as a base...
> Then there's the section "One must be allowed to take derivatives of integer arguments" in the ChainRules docs (CC @oxinabox), so Julia users are used to diffing in respect...
Can you post the full trace to Gist or something similar?
And give current main a try
``` could not deduce type of integer %value_phi14152 = phi double [ %49, %L174.lr.ph ], [ %value_phi20, %L148 ] julia: /workspace/srcdir/Enzyme/enzyme/Enzyme/TypeAnalysis/TypeAnalysis.cpp:4877: ConcreteType TypeResults::intType(size_t, llvm::Value*, bool, bool): Assertion `0 && "could...
Is this still an issues on current Enzyme?
So the third check is without KA? ``` using ForwardDiff using Enzyme using Test function speelpenning(x) reduce(*, x) end function test_gradient(x::Vector{Float64}, gradient::Vector{Float64}) errg = 0.0 y = speelpenning(x) for (i,...
@michel2323 could you add the one you care about as a test?