4 strided blas tests are the only ones that fail in LinearAlgebra.jl BLAS tests

[ViralBShah]

The BLAS tests are currently disabled since 4 tests fail. Only the LAPACK tests are being run:

strided interface blas: Test Failed at /Users/viral/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:730
  Expression: C == WrappedArray([23 50 + 38im; 35 + 27im 152])
   Evaluated: ComplexF32[23.0f0 + 27.0f0im 50.0f0 + 38.0f0im; 35.0f0 + 27.0f0im 152.0f0 + 36.0f0im] == Complex{Int64}[23 + 0im 50 + 38im; 35 + 27im 152 + 0im]

Stacktrace:
 [1] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Test/src/Test.jl:680 [inlined]
 [2] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:730 [inlined]
 [3] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Test/src/Test.jl:1776 [inlined]
 [4] top-level scope
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:639
strided interface blas: Test Failed at /Users/viral/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:732
  Expression: C == WrappedArray([63 138 + 38im; 35 + 27im 352])
   Evaluated: ComplexF32[63.0f0 + 27.0f0im 138.0f0 + 38.0f0im; 35.0f0 + 27.0f0im 352.0f0 + 36.0f0im] == Complex{Int64}[63 + 0im 138 + 38im; 35 + 27im 352 + 0im]

Stacktrace:
 [1] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Test/src/Test.jl:680 [inlined]
 [2] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:732 [inlined]
 [3] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Test/src/Test.jl:1776 [inlined]
 [4] top-level scope
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:639
strided interface blas: Test Failed at /Users/viral/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:730
  Expression: C == WrappedArray([23 50 + 38im; 35 + 27im 152])
   Evaluated: ComplexF64[23.0 + 27.0im 50.0 + 38.0im; 35.0 + 27.0im 152.0 + 36.0im] == Complex{Int64}[23 + 0im 50 + 38im; 35 + 27im 152 + 0im]

Stacktrace:
 [1] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Test/src/Test.jl:680 [inlined]
 [2] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:730 [inlined]
 [3] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Test/src/Test.jl:1776 [inlined]
 [4] top-level scope
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:639
strided interface blas: Test Failed at /Users/viral/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:732
  Expression: C == WrappedArray([63 138 + 38im; 35 + 27im 352])
   Evaluated: ComplexF64[63.0 + 27.0im 138.0 + 38.0im; 35.0 + 27.0im 352.0 + 36.0im] == Complex{Int64}[63 + 0im 138 + 38im; 35 + 27im 352 + 0im]

Stacktrace:
 [1] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Test/src/Test.jl:680 [inlined]
 [2] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:732 [inlined]
 [3] macro expansion
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Test/src/Test.jl:1776 [inlined]
 [4] top-level scope
   @ ~/.julia/juliaup/julia-1.12.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/test/blas.jl:639
Test Summary:                | Pass  Fail  Total   Time
LinearAlgebra.jl BLAS tests  | 1902     4   1906  31.3s
  vec_pointer_stride         |    7            7   0.2s
  elty = Float32             |  394          394   7.1s
  elty = Float64             |  394          394   5.7s
  elty = ComplexF32          |  430          430   6.5s
  elty = ComplexF64          |  430          430   5.9s
  syr for eltype Float32     |    4            4   0.1s
  syr for eltype Float64     |    4            4   0.1s
  syr for eltype ComplexF32  |    4            4   0.1s
  syr for eltype ComplexF64  |    4            4   0.1s
  her for eltype ComplexF32  |    4            4   0.1s
  her for eltype ComplexF64  |    4            4   0.1s
  strided interface adjtrans |   26           26   0.2s
  strided interface blas     |  146     4    150   2.7s
  get_set_num_threads        |    4            4   0.0s
  test for 0-strides         |   11           11   0.3s
  test for 0-strides         |   11           11   0.3s
  test for 0-strides         |   12           12   0.5s
  test for 0-strides         |   12           12   0.4s
  libblas_name               |    1            1   0.0s
RNG of the outermost testset: Xoshiro(0xfef9c95b5a3f61e8, 0x9db6807c8e2aa3ed, 0x2317a9b6478e87d4, 0x2b6a6b2384eb15fb, 0xb77c89590cf58eb0)
ERROR: LoadError: Some tests did not pass: 1902 passed, 4 failed, 0 errored, 0 broken.
in expression starting at /Users/viral/.julia/dev/AppleAccelerate/test/runtests.jl:398
ERROR: Package AppleAccelerate errored during testing

[ViralBShah]

@imciner2 @giordano It feels like this is a bug in Accelerate.

@jishnub @amontoison Is there a remote possibility that there is a bug in our wrappers or dependency on some behaviour in OpenBLAS that is causing these specific tests to fail with Accelerate? I don't see the issue with MKL, so it feels like it should be an Accelerate bug, but worth digging deeper.

[amontoison]

I highly suspect a bug in Accelerate.

[daviehh]

This is interesting. TLDR: it's Accelerate's NEWLAPACK interface thing, but we are also passing the "wrong" input.

See C code at gist, old interface gives the expected result, while NEWLAPACK interfaces gives another set of results. C output same as julia, so it's not a wrapper issue but at the Accelerate end.
The file test_blas_acc.c in the gist is the same as the data used in the test
https://github.com/JuliaLang/LinearAlgebra.jl/blob/e367ff15e98f9701400f8fb7707405c14ab8645e/test/blas.jl#L724

While it may be considered a 'bug' in Accelerate's NEWLAPACK interface, it may also be sort of an open question whether one should expect the same result to openblas/mkl if the input is 'wrong': the failed tests here are the 32/64-bit hemm, herk, and her2k.

For hemm, A should be hermitian. While A = A' is not necessary since we are only taking the upper/lower triangle, the diagonal elements of hermitian matrices need to be real, but we are passing matrices with complex diagonals. Similar for herk, where C needs to be hermitian.

The file test_blas_acc_herm.c in the gist enforces the correct input values, and it gives the same result as openblas, same in julia, with or without using AppleAccelerate:
copied from the blas.jl test, but fixes the diagonal (C is hermitian after herk so do not need to reset for her2k)

using Test
struct WrappedArray{T,N} <: AbstractArray{T,N}
    A::Array{T,N}
end

Base.size(A::WrappedArray) = size(A.A)
Base.getindex(A::WrappedArray, i::Int) = A.A[i]
Base.getindex(A::WrappedArray{T,N}, I::Vararg{Int,N}) where {T,N} = A.A[I...]
Base.setindex!(A::WrappedArray, v, i::Int) = setindex!(A.A, v, i)
Base.setindex!(A::WrappedArray{T,N}, v, I::Vararg{Int,N}) where {T,N} = setindex!(A.A, v, I...)
Base.cconvert(::Type{Ptr{T}}, A::WrappedArray{T}) where T = Base.cconvert(Ptr{T}, A.A)

Base.strides(A::WrappedArray) = strides(A.A)
Base.elsize(::Type{WrappedArray{T,N}}) where {T,N} = Base.elsize(Array{T,N})

for elty in (ComplexF32, ComplexF64)
    A = WrappedArray(elty[1 2+2im; 3+3im 4])
    B = WrappedArray(elty[1+im 2+2im; 3+3im 4+4im])
    C = WrappedArray(elty[1+im 2+2im; 3+3im 4+4im])
    @test BLAS.hemm!('L', 'U', elty(2), A, B, elty(1), C) isa WrappedArray{elty,2}
    @test C == WrappedArray([3+27im 6+38im; 35+27im 52+36im])
    C[1] = real(C[1])
    C[4] = real(C[4])
    println("="^6)
    println("C before herk!")
    display(C)
    A = WrappedArray(elty[1+1im 2+2im; 3+3im 4+4im]) # A can be non-Hermitain here
    @test BLAS.herk!('U', 'N', real(elty(2)), A, real(elty(1)), C) isa WrappedArray{elty,2}
    println("C after herk!")
    display(C)

    @test BLAS.her2k!('U', 'N', elty(2), A, B, real(elty(1)), C) isa WrappedArray{elty,2}
    println("C after her2k!")
    display(C)
end