Error with non-integer sizeof when differentiating Base.diff

[gdalle]

Here's an MWE with Enzyme v0.13.19:

julia> import Enzyme

julia> sumdiff(x) = sum(diff(x));

julia> x, dx = float.(1:5), float.(6:10);

julia> Enzyme.hvp(sum, x, dx)  # works
5-element Vector{Float64}:
 0.0
 0.0
 0.0
 0.0
 0.0

julia> Enzyme.hvp(sumdiff, x, dx)  # fails
ERROR: InexactError: Int64(0.125)
Stacktrace:
  [1] Int64
    @ ./float.jl:994 [inlined]
  [2] sizeof
    @ ~/.julia/packages/LLVM/wMjUU/src/datalayout.jl:91 [inlined]
  [3] should_recurse(typ2::Any, arg_t::LLVM.LLVMType, byref::GPUCompiler.ArgumentCC, dl::LLVM.DataLayout)
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/ottqJ/src/absint.jl:208
  [4] abs_typeof(arg::LLVM.Value, partial::Bool, seenphis::Set{LLVM.PHIInst})
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/ottqJ/src/absint.jl:601
  [5] abs_typeof(arg::LLVM.Value, partial::Bool, seenphis::Set{LLVM.PHIInst})
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/ottqJ/src/absint.jl:710
  [6] abs_typeof
    @ ~/.julia/packages/Enzyme/ottqJ/src/absint.jl:283 [inlined]
  [7] codegen(output::Symbol, job::GPUCompiler.CompilerJob{…}; libraries::Bool, deferred_codegen::Bool, optimize::Bool, toplevel::Bool, strip::Bool, validate::Bool, only_entry::Bool, parent_job::Nothing)
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/ottqJ/src/compiler.jl:4255
  [8] codegen
    @ ~/.julia/packages/Enzyme/ottqJ/src/compiler.jl:3218 [inlined]
  [9] _thunk(job::GPUCompiler.CompilerJob{Enzyme.Compiler.EnzymeTarget, Enzyme.Compiler.EnzymeCompilerParams}, postopt::Bool)
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/ottqJ/src/compiler.jl:5265
 [10] cached_compilation(job::GPUCompiler.CompilerJob)
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/ottqJ/src/compiler.jl:5306
 [11] thunkbase(mi::Core.MethodInstance, World::UInt64, FA::Type{…}, A::Type{…}, TT::Type, Mode::Enzyme.API.CDerivativeMode, width::Int64, ModifiedBetween::NTuple{…} where N, ReturnPrimal::Bool, ShadowInit::Bool, ABI::Type, ErrIfFuncWritten::Bool, RuntimeActivity::Bool)
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/ottqJ/src/compiler.jl:5410
 [12] thunk_generator(world::UInt64, source::LineNumberNode, FA::Type, A::Type, TT::Type, Mode::Enzyme.API.CDerivativeMode, Width::Int64, ModifiedBetween::NTuple{…} where N, ReturnPrimal::Bool, ShadowInit::Bool, ABI::Type, ErrIfFuncWritten::Bool, RuntimeActivity::Bool, self::Any, fakeworld::Any, fa::Type, a::Type, tt::Type, mode::Type, width::Type, modifiedbetween::Type, returnprimal::Type, shadowinit::Type, abi::Type, erriffuncwritten::Type, runtimeactivity::Type)
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/ottqJ/src/compiler.jl:5562
 [13] autodiff
    @ ~/.julia/packages/Enzyme/ottqJ/src/Enzyme.jl:640 [inlined]
 [14] autodiff
    @ ~/.julia/packages/Enzyme/ottqJ/src/Enzyme.jl:544 [inlined]
 [15] autodiff
    @ ~/.julia/packages/Enzyme/ottqJ/src/Enzyme.jl:516 [inlined]
 [16] hvp!
    @ ~/.julia/packages/Enzyme/ottqJ/src/sugar.jl:1105 [inlined]
 [17] hvp(f::typeof(sumdiff), x::Vector{Float64}, v::Vector{Float64})
    @ Enzyme ~/.julia/packages/Enzyme/ottqJ/src/sugar.jl:1072
 [18] top-level scope
    @ ~/Work/GitHub/Julia/DifferentiationInterface.jl/DifferentiationInterface/test/playground.jl:7
Some type information was truncated. Use `show(err)` to see complete types.

Note that the error doesn't happen when I reimplement diff manually:

julia> function mydiff(x)
           y = similar(x, length(x)-1)
           for i in eachindex(y)
               y[i] = x[i+1]-x[i]
           end
           return y
       end
mydiff (generic function with 1 method)

julia> summydiff(x) = sum(mydiff(x));

julia> Enzyme.hvp(summydiff, x, dx)  # works
5-element Vector{Float64}:
 0.0
 0.0
 0.0
 0.0
 0.0

[wsmoses]

your test case works for me on main, does it err on the latest Enzyme for you still?

julia> sumdiff(x) = sum(diff(x));

julia> x, dx = float.(1:5), float.(6:10);

julia> Enzyme.hvp(sum, x, dx)  # works
5-element Vector{Float64}:
 0.0
 0.0
 0.0
 0.0
 0.0

julia> Enzyme.hvp(sumdiff, x, dx)  # fails
5-element Vector{Float64}:
 0.0
 0.0
 0.0
 0.0
 0.0

julia> 

[gdalle]

If I'm not mistaken, v0.3.19 is the latest version of Enzyme? It corresponds to commit hash 6606cd9 which is also the last commit on main?

[wsmoses]

oh yeah, and hm odd. Regardless this appears to be a bug in LLVM.jl here: https://github.com/maleadt/LLVM.jl/blob/a1c3b217c1c9e5db4223e4e0af1754aae53d113f/src/datalayout.jl#L91

Maybe file an issue there?

[vchuravy]

@wsmoses not really a bug in LLVM.jl

You need to decide what you want to do for i1 in should_recurse(typ2::Any, arg_t::LLVM.LLVMType, byref::GPUCompiler.ArgumentCC, dl::LLVM.DataLayout)

[wsmoses]

workaround presumably added here #2181