Missing rule for LAPACK's dgetrf_64_

[mhauru]

MWE:

using Distributions
using Random
using Enzyme: Enzyme
using LinearAlgebra: I

d = MatrixTDist(
    3.1,
    randn(2, 3),
    Array(I, (2, 2)),
    Array(I, (3, 3)),
)
x = randn(2, 3)
f(x) = logpdf(d, x)
Enzyme.gradient(Enzyme.Forward, Enzyme.Const(f), x, Val(1))

Output:

 fallback?
ERROR:
No forward mode derivative found for dgetrf_64_
 at context:   call void @dgetrf_64_(i8* noundef nonnull %6, i8* noundef nonnull %8, i64 %22, i8* noundef nonnull %10, i64 %26, i64 noundef %bitcast_coercion42) #76 [ "jl_roots"({} addrspace(10)* null, {} addrspace(10)* %19, {} addrspace(10)* null, {} addrspace(10)* %2, {} addrspace(10)* null, {} addrspace(10)* null) ], !dbg !140

Stacktrace:
 [1] #getrf!#1
   @ ~/.julia/juliaup/julia-1.10.5+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/lapack.jl:565


Stacktrace:
  [1] copyto_unaliased!
    @ ./abstractarray.jl:1082
  [2] gemm!
    @ ~/.julia/juliaup/julia-1.10.5+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/blas.jl:1524
  [3] indexed_iterate (repeats 2 times)
    @ ./tuple.jl:92 [inlined]
  [4] log_proc1
    @ ./special/log.jl:165 [inlined]
  [5] _log
    @ ./special/log.jl:295
  [6] LazyString
    @ ./strings/lazy.jl:41 [inlined]
  [7] gemm!
    @ ~/.julia/juliaup/julia-1.10.5+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/blas.jl:1519
  [8] macro expansion
    @ ~/.julia/packages/Enzyme/TiboG/src/compiler.jl:7187 [inlined]
  [9] enzyme_call
    @ ~/.julia/packages/Enzyme/TiboG/src/compiler.jl:6794 [inlined]
 [10] ForwardModeThunk
    @ ~/.julia/packages/Enzyme/TiboG/src/compiler.jl:6674 [inlined]
 [11] runtime_generic_fwd(activity::Type{…}, width::Val{…}, RT::Val{…}, f::typeof(logpdf), df::Nothing, primal_1::MatrixTDist{…}, shadow_1_1::Nothing, primal_2::Matrix{…}, shadow_2_1::Matrix{…})
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/TiboG/src/rules/jitrules.jl:223
 [12] f
    @ ./REPL[17]:1 [inlined]
 [13] fwddiffejulia_f_2311wrap
    @ ./REPL[17]:0
 [14] macro expansion
    @ ~/.julia/packages/Enzyme/TiboG/src/compiler.jl:7187 [inlined]
 [15] enzyme_call
    @ ~/.julia/packages/Enzyme/TiboG/src/compiler.jl:6794 [inlined]
 [16] ForwardModeThunk
    @ ~/.julia/packages/Enzyme/TiboG/src/compiler.jl:6674 [inlined]
 [17] autodiff
    @ ~/.julia/packages/Enzyme/TiboG/src/Enzyme.jl:437 [inlined]
 [18] #112
    @ ~/.julia/packages/Enzyme/TiboG/src/Enzyme.jl:1197 [inlined]
 [19] ntuple
    @ ./ntuple.jl:19 [inlined]
 [20] gradient(::EnzymeCore.ForwardMode{…}, f::EnzymeCore.Const{…}, x::Matrix{…}, ::Val{…}; shadow::NTuple{…})
    @ Enzyme ~/.julia/packages/Enzyme/TiboG/src/Enzyme.jl:1196
 [21] gradient(::EnzymeCore.ForwardMode{…}, f::EnzymeCore.Const{…}, x::Matrix{…}, ::Val{…})
    @ Enzyme ~/.julia/packages/Enzyme/TiboG/src/Enzyme.jl:1195
 [22] top-level scope
    @ REPL[18]:1
Some type information was truncated. Use `show(err)` to see complete types.

[wsmoses]

Can you minimize this?

We should handle all the common ways LU decomposition is used without this rule, so I'm surprised it comes up

[mhauru]

using Random
using Enzyme: Enzyme
using LinearAlgebra: logdet

x = randn(2,2)
Enzyme.autodiff(Enzyme.Forward, Enzyme.Const(logdet), Enzyme.Duplicated(x, zero(x)))

[mhauru]

Or just plain

using Random
using Enzyme: Enzyme
using LinearAlgebra: lu

x = randn(2,2)
f(x) = lu(x, check = false)
Enzyme.autodiff(Enzyme.Forward, Enzyme.Const(f), Enzyme.Duplicated(x, zero(x)))