OpenBLAS 0.3.13 ex-shift patch not working

[pablosanjose]

There must have been some kind of problem with the upgrade to OpenBLAS 0.3.13 in JuliaLang/julia#39216. That PR included a patch called openblas-exshift.patch intended to bring in the fix in OpenMathLib/OpenBLAS#3091 without having to wait for OpenBLAS 0.3.14, but it seems the patch does not work. I don't know why, but the following test, discussed in the original issue #475, still fails in Julia 1.6 and nightly

A1 = ComplexF64[0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0; 3.7796350217469814 -3.3125635598133054 0.0 0.0 0.0 0.0 0.0 0.0 6.418270043493963 -6.625127119626611 0.0 0.0 0.0 0.0 0.0 -1.0; -3.312563559813306 3.779635021746982 0.0 0.0 0.0 0.0 0.0 0.0 -6.625127119626612 6.418270043493964 -1.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 3.7796350217469814 0.0 0.0 -3.3125635598133054 0.0 0.0 0.0 -1.0 6.418270043493963 0.0 0.0 -6.625127119626611 0.0 0.0; 0.0 0.0 0.0 3.779635021746982 -3.312563559813306 0.0 0.0 0.0 0.0 0.0 0.0 6.418270043493964 -6.625127119626612 0.0 -1.0 0.0; 0.0 0.0 0.0 -3.3125635598133054 3.7796350217469814 0.0 0.0 0.0 0.0 0.0 0.0 -6.625127119626611 6.418270043493963 -1.0 0.0 0.0; 0.0 0.0 -3.312563559813306 0.0 0.0 3.779635021746982 0.0 0.0 0.0 0.0 -6.625127119626612 0.0 -1.0 6.418270043493964 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 3.7796350217469814 -3.3125635598133054 0.0 0.0 0.0 -1.0 0.0 0.0 6.418270043493963 -6.625127119626611; 0.0 0.0 0.0 0.0 0.0 0.0 -3.312563559813306 3.779635021746982 -1.0 0.0 0.0 0.0 0.0 0.0 -6.625127119626612 6.418270043493964]
B1 = ComplexF64[1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -3.7796350217469814 3.312563559813306 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 3.3125635598133054 -3.779635021746982 0.0 0.0 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -3.7796350217469814 0.0 0.0 3.312563559813306 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -3.779635021746982 3.3125635598133054 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 3.312563559813306 -3.7796350217469814 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 3.3125635598133054 0.0 0.0 -3.779635021746982 0.0 0.0; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -3.7796350217469814 3.312563559813306; 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 3.3125635598133054 -3.779635021746982]

Then, on Intel + macOS

julia> using LinearAlgebra

julia> eigen(A1, B1)  # or schur(A1, B1)
ERROR: LAPACKException(12)

The failure is platform dependent (other pencils that fail for other platforms are available in #475), and seems similar to JuliaLang/LinearAlgebra.jl#832.

Also, using MKL (Julia 1.5.3 + MKL.jl) never fails.

I'm clueless about the whole Yggdrasil BinaryBuilder machinery, so I cannot fathom why the patch doesn't work (it did work when I compiled locally with it back in the day). However, perhaps the best thing is to wait for OpenBLAS 0.3.14 to be released with the fix and bake that in instead.

[martin-frbg]

Not sure if this belongs in Reference-LAPACK rather than Julia, but anyway...
What I put into OpenBLAS after 0.3.13 was a slightly earlier iteration of JuliaLang/julia#477 (though the only bits imissing should be the "LLSAME" typo fix and the "fix compiler warnings in qz code"). The former is probably harmless (only an EXTERNAL declaration) and the latter was added after you reported success with JuliaLang/julia#477

[martin-frbg]

Checked that current OpenBLAS-develop still works with the (python versions of the) pencils from JuliaLang/julia#475 - on Haswell at least.

[pablosanjose]

My apologies, I opened this issue in the wrong repo. Closing and posting over at Julialang