Segfault with function splatting #1942
Comments
This was referenced Oct 8, 2024
|
Here's a pure-Enzyme reproducer: using Enzyme
f(x::T...) where {T} = (1 - x[1])^2 + 100 * (x[2] - x[1]^2)^2
x = [2.0, 3.0]
hvp(splat(f), x, zero(x))Error log: |
gdalle
changed the title
|
Note that with the latest changes to DI's handling of StaticArrays, switching the import DifferentiationInterface as DI
using Enzyme: Enzyme
using StaticArrays
f(x::T...) where {T} = (1 - x[1])^2 + 100 * (x[2] - x[1]^2)^2
f_nosplat(x::AbstractVector) = (1 - x[1])^2 + 100 * (x[2] - x[1]^2)^2
xs = SVector(2.0, 3.0)
backend = AutoEnzyme()
DI.hessian(f_nosplat, backend, xs) # works
DI.hessian(splat(f), backend, xs) # works |
|
Resolved by #1975 |
|
I don't think this is completely solved? With Enzyme v0.13.11 I don't get a segfault but I do get an error: julia> Enzyme.hvp(splat(f), x, zero(x))
ERROR: Enzyme execution failed.
Enzyme: Not yet implemented, mixed activity for jl_new_struct constants=Bool[1, 1, 1, 1, 1, 1, 0, 1, 1] %16 = call noalias nonnull "enzyme_type"="{[-1]:Pointer}" {} addrspace(10)* ({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)*, {} addrspace(10)*, ...) @julia.call({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)* noundef nonnull @jl_f_tuple, {} addrspace(10)* noundef null, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 135706086968176 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 135706049721088 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 135706086968176 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 135703490524368 to {}*) to {} addrspace(10)*), {} addrspace(10)* nonnull %15, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 135703919856144 to {}*) to {} addrspace(10)*), {} addrspace(10)* nonnull %11, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 135705534029648 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 135706276003848 to {}*) to {} addrspace(10)*)) #17, !dbg !58 Tuple{Bool, Any, Any}[(1, Val{false}, GPUCompiler.BITS_REF), (1, Val{1}, GPUCompiler.BITS_REF), (1, Val{false}, GPUCompiler.BITS_REF), (1, Type{@NamedTuple{1, 2, 3}}, GPUCompiler.BITS_REF), (0, nothing, nothing), (1, Type{typeof(f)}, GPUCompiler.BITS_REF), (0, nothing, nothing), (1, typeof(f), GPUCompiler.BITS_REF), (1, Nothing, GPUCompiler.BITS_REF)] LLVM.User[LLVM.ConstantExpr(0x0000000019ed5530), LLVM.ConstantExpr(0x000000001affa930), LLVM.ConstantExpr(0x0000000019ed5530), LLVM.ConstantExpr(0x000000001df82d30), LLVM.CallInst(%15 = call nonnull "enzyme_type"="{[-1]:Pointer}" {} addrspace(10)* ({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)*, {} addrspace(10)*, ...) @julia.call({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)* noundef nonnull @ijl_apply_generic, {} addrspace(10)* noundef addrspacecast ({}* inttoptr (i64 135706048967360 to {}*) to {} addrspace(10)*), {} addrspace(10)* nonnull %14) #16, !dbg !58), LLVM.ConstantExpr(0x000000001b075fb0), LLVM.CallInst(%11 = call nonnull "enzyme_type"="{[-1]:Pointer}" {} addrspace(10)* ({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)*, {} addrspace(10)*, ...) @julia.call({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)* noundef nonnull @jl_f__apply_iterate, {} addrspace(10)* noundef null, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 135706084209456 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 135702402236400 to {}*) to {} addrspace(10)*), {} addrspace(10)* nonnull %10) #15, !dbg !55), LLVM.ConstantExpr(0x0000000014b05f30), LLVM.ConstantExpr(0x0000000012df5ff0)]
Stacktrace:
[1] runtime_iterate_augfwd
@ ~/.julia/packages/Enzyme/vgArw/src/rules/jitrules.jl:77 [inlined]
[2] runtime_iterate_augfwd
@ ~/.julia/packages/Enzyme/vgArw/src/rules/jitrules.jl:0 [inlined]
[3] fwddiffejulia_runtime_iterate_augfwd_29831_inner_1wrap
@ ~/.julia/packages/Enzyme/vgArw/src/rules/jitrules.jl:0
[4] macro expansion
@ ~/.julia/packages/Enzyme/vgArw/src/compiler.jl:8136 [inlined]
[5] enzyme_call
@ ~/.julia/packages/Enzyme/vgArw/src/compiler.jl:7702 [inlined]
[6] ForwardModeThunk
@ ~/.julia/packages/Enzyme/vgArw/src/compiler.jl:7491 [inlined]
[7] runtime_generic_fwd(activity::Type{…}, runtimeActivity::Val{…}, width::Val{…}, RT::Val{…}, f::typeof(Enzyme.Compiler.runtime_iterate_augfwd), df::Nothing, primal_1::Type{…}, shadow_1_1::Nothing, primal_2::Val{…}, shadow_2_1::Nothing, primal_3::Val{…}, shadow_3_1::Nothing, primal_4::Val{…}, shadow_4_1::Nothing, primal_5::Val{…}, shadow_5_1::Nothing, primal_6::typeof(f), shadow_6_1::Nothing, primal_7::Nothing, shadow_7_1::Nothing, primal_8::Vector{…}, shadow_8_1::Vector{…}, primal_9::Vector{…}, shadow_9_1::Vector{…})
@ Enzyme.Compiler ~/.julia/packages/Enzyme/vgArw/src/rules/jitrules.jl:305
[8] Splat
@ ./operators.jl:1271 [inlined]
[9] diffejulia_Splat_30897wrap
@ ./operators.jl:0 [inlined]
[10] macro expansion
@ ~/.julia/packages/Enzyme/vgArw/src/compiler.jl:8136 [inlined]
[11] enzyme_call
@ ~/.julia/packages/Enzyme/vgArw/src/compiler.jl:7702 [inlined]
[12] CombinedAdjointThunk
@ ~/.julia/packages/Enzyme/vgArw/src/compiler.jl:7475 [inlined]
[13] autodiff_deferred
@ ~/.julia/packages/Enzyme/vgArw/src/Enzyme.jl:781 [inlined]
[14] autodiff
@ ~/.julia/packages/Enzyme/vgArw/src/Enzyme.jl:512 [inlined]
[15] gradient!
@ ~/.julia/packages/Enzyme/vgArw/src/Enzyme.jl:1786 [inlined]
[16] fwddiffejulia_gradient__30894wrap
@ ~/.julia/packages/Enzyme/vgArw/src/Enzyme.jl:0
[17] macro expansion
@ ~/.julia/packages/Enzyme/vgArw/src/compiler.jl:8136 [inlined]
[18] enzyme_call
@ ~/.julia/packages/Enzyme/vgArw/src/compiler.jl:7702 [inlined]
[19] ForwardModeThunk
@ ~/.julia/packages/Enzyme/vgArw/src/compiler.jl:7491 [inlined]
[20] autodiff
@ ~/.julia/packages/Enzyme/vgArw/src/Enzyme.jl:647 [inlined]
[21] autodiff
@ ~/.julia/packages/Enzyme/vgArw/src/Enzyme.jl:537 [inlined]
[22] autodiff
@ ~/.julia/packages/Enzyme/vgArw/src/Enzyme.jl:504 [inlined]
[23] hvp!
@ ~/.julia/packages/Enzyme/vgArw/src/Enzyme.jl:2497 [inlined]
[24] hvp(f::Base.Splat{typeof(f)}, x::Vector{Float64}, v::Vector{Float64})
@ Enzyme ~/.julia/packages/Enzyme/vgArw/src/Enzyme.jl:2464
[25] top-level scope
@ REPL[50]:1
Some type information was truncated. Use `show(err)` to see complete types. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
EDIT: pure Enyzme MWE is below
Here's an MWE for the bug I uncovered in the JuMP docs PR (jump-dev/JuMP.jl#3836). It seems to be due to splatting, which is only necessary because JuMP doesn't use vector arguments:
And here's the error log (on 1.10):
The text was updated successfully, but these errors were encountered: