Broken dtype detection in CHOLMOD in 1.11

[pawel-tarasiuk-quantumz]

Computing qr factorization on SparseMatrixCSC{Float32} causes an error with Julia 1.11:

using SparseArrays, LinearAlgebra
A = sprandn(Float32, 10, 10, 0.5)
qr(Float64.(A))  # works proprely
qr(A)   # !!!

Message:

ERROR: SparseArrays.CHOLMOD.CHOLMODException("dtype=0 not supported")
Stacktrace:
 [1] SparseArrays.CHOLMOD.Dense{Float32}(ptr::Ptr{SparseArrays.LibSuiteSparse.cholmod_dense_struct})
   @ SparseArrays.CHOLMOD ~/.julia/juliaup/julia-1.11.1+0.x64.linux.gnu/share/julia/stdlib/v1.11/SparseArrays/src/solvers/cholmod.jl:276
 [2] qr(A::SparseMatrixCSC{Float32, Int64}; tol::Float32, ordering::Int32)
   @ SparseArrays.SPQR ~/.julia/juliaup/julia-1.11.1+0.x64.linux.gnu/share/julia/stdlib/v1.11/SparseArrays/src/solvers/spqr.jl:209
 [3] qr(A::SparseMatrixCSC{Float32, Int64})
   @ SparseArrays.SPQR ~/.julia/juliaup/julia-1.11.1+0.x64.linux.gnu/share/julia/stdlib/v1.11/SparseArrays/src/solvers/spqr.jl:195
 [4] top-level scope
   @ REPL[4]:1

Tested on clean Julia 1.11 setup with juliaup on Ubuntu 22.04, x86_64. Eqivalent setup with Julia 1.10 works for either element type.

It would seem that

SparseArrays.jl/src/solvers/spqr.jl

Line 200 in 1b4933c

HTau = Ref{Ptr{CHOLMOD.cholmod_dense}}(C_NULL)

does not indicate the proper dtype, as unsafe_load(HTau[]).dtype is 0 when run for either type. This is incidentally equal to CHOLMOD_DOUBLE, which allows the computation for Float64.

[ViralBShah]

Thanks! If you have a PR, much appreciated!

[pawel-tarasiuk-quantumz]

Since dtype mismatch is possible, it might be worth it to support varied results with no assumptions, as proposed in #586.

[pawel-tarasiuk-quantumz]

I believe #586 to be ready. Additional data conversion is required in some cases, but now it is handled by SparseMatrixCSC{Tv, Ti} constructors.

[ViralBShah]

Thanks - merged. I think this started happening because CHOLMOD recently added Float32 support, which we enabled - but we still needed to extend it to QR - which you have now.