BLAS.asum(::ComplexF32) fails on AMD 9950X in Julia 1.10

[mind6]

This was discovered during CUDA testing JuliaGPU/CUDA.jl#2810 (comment). It seems to produce incorrect results for ComplexF32 on AMD 9950X and also on 7950X.

MWE:

using InteractiveUtils
versioninfo()

using Pkg
Pkg.activate(@__DIR__)
Pkg.add(["Random", "LinearAlgebra"])

using LinearAlgebra, Random, Test

@testset "BLAS test" begin
   Random.seed!(1234)
   A = rand(ComplexF32, 10)

   sum1 = BLAS.asum(A)
   sum2 = sum(z -> abs(real(z)) + abs(imag(z)), A)

   @show A, sum1, sum2
   @test sum1 == sum2
end

My output on 1.10:

>julia +1.10 asum_bug.jl
Julia Version 1.10.10
Commit 95f30e51f4 (2025-06-27 09:51 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 32 × AMD Ryzen 9 9950X 16-Core Processor
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, generic)
Threads: 32 default, 0 interactive, 16 GC (on 32 virtual cores)
Environment:
  JULIA_CONDAPKG_BACKEND = Null
  JULIA_DEPOT_PATH = K:\.julia
  JULIA_NUM_THREADS = auto
  JULIA_PYTHONCALL_EXE = K:\Miniconda3\envs\py310\python.exe
  Activating new project at `K:\DevDocuments\BLAS_bug`
   Resolving package versions...
    Updating `K:\DevDocuments\BLAS_bug\Project.toml`
  [37e2e46d] + LinearAlgebra
  [9a3f8284] + Random
    Updating `K:\DevDocuments\BLAS_bug\Manifest.toml`
  [56f22d72] + Artifacts
  [8f399da3] + Libdl
  [37e2e46d] + LinearAlgebra
  [9a3f8284] + Random
  [ea8e919c] + SHA v0.7.0
  [e66e0078] + CompilerSupportLibraries_jll v1.1.1+0
  [4536629a] + OpenBLAS_jll v0.3.23+4
  [8e850b90] + libblastrampoline_jll v5.11.0+0
(A, sum1, sum2) = (ComplexF32[0.32597667f0 + 0.5490511f0im, 0.21858662f0 + 0.8942454f0im, 0.35311162f0 + 0.39425534f0im, 0.9531246f0 + 0.7955469f0im, 0.49424982f0 + 0.748415f0im, 0.57823193f0 + 0.72793496f0im, 0.007447958f0 + 0.19937658f0im, 0.43924308f0 + 0.6825326f0im, 0.9567409f0 + 0.6478553f0im, 0.99666524f0 + 0.749194f0im], 10.34919f0, 11.711785f0)
BLAS test: Test Failed at K:\DevDocuments\BLAS_bug\asum_bug.jl:18
  Expression: sum1 ≈ sum2
   Evaluated: 10.34919f0 ≈ 11.711785f0

The issues seems to be in OpenBLAS_jll v0.3.23. There's no issue using Julia 1.11 OpenBLAS_jll v0.3.27.

>julia +1.11 asum_bug.jl
Julia Version 1.11.6
Commit 9615af0f26 (2025-07-09 12:58 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 32 × AMD Ryzen 9 9950X 16-Core Processor
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, generic)
Threads: 32 default, 0 interactive, 16 GC (on 32 virtual cores)
Environment:
  JULIA_CONDAPKG_BACKEND = Null
  JULIA_DEPOT_PATH = K:\.julia
  JULIA_NUM_THREADS = auto
  JULIA_PYTHONCALL_EXE = K:\Miniconda3\envs\py310\python.exe
  Activating project at `K:\DevDocuments\BLAS_bug`
   Resolving package versions...
    Updating `K:\DevDocuments\BLAS_bug\Project.toml`
  [37e2e46d] ~ LinearAlgebra ⇒ v1.11.0
  [9a3f8284] ~ Random ⇒ v1.11.0
    Updating `K:\DevDocuments\BLAS_bug\Manifest.toml`
  [56f22d72] ~ Artifacts ⇒ v1.11.0
  [8f399da3] ~ Libdl ⇒ v1.11.0
  [37e2e46d] ~ LinearAlgebra ⇒ v1.11.0
  [9a3f8284] ~ Random ⇒ v1.11.0
  [4536629a] ↑ OpenBLAS_jll v0.3.23+4 ⇒ v0.3.27+1
(A, sum1, sum2) = (ComplexF32[0.32597667f0 + 0.5490511f0im, 0.21858662f0 + 0.8942454f0im, 0.35311162f0 + 0.39425534f0im, 0.9531246f0 + 0.7955469f0im, 0.49424982f0 + 0.748415f0im, 0.57823193f0 + 0.72793496f0im, 0.007447958f0 + 0.19937658f0im, 0.43924308f0 + 0.6825326f0im, 0.9567409f0 + 0.6478553f0im, 0.99666524f0 + 0.749194f0im], 11.711786f0, 11.711785f0)
Test Summary: | Pass  Total  Time
BLAS test     |    1      1  0.2s

[ViralBShah]

@imciner2 I think we may need to bring the latest openblas in to solve some of these sorts of issues (even though they may still persist).

[imciner2]

Do we need to go to 0.3.30, or is just bumping to one of our existing builds enough for this? I need to sort out some issues in BinaryBuilder to get a nice 0.3.30 build script still.

[ViralBShah]

My main thought was to use something newer and see if it fixes this. I hadn't paid close attention to realize that we may have something between what we ship now and 0.3.30. Yeah we should just try that first.

[imciner2]

1.10 is currently on OpenBLAS 0.3.23, and we have up to 0.3.29 available (1.11 is 0.3.27, and 1.12 and master are 0.3.29).

OpenBLAS 0.3.29 required dropping libgfortran 3 & 4 on PowerPC unfortunately, so I think it probably is safest to upgrade 1.10 to use OpenBLAS 0.3.27, which I have proposed in JuliaLang/julia#59334.

[imciner2]

I've done a bisect on OpenBLAS on my AMD 7900 (Zen4, same as the 7950X) using the test program above, and it appears this was fixed by OpenMathLib/OpenBLAS@9019bc4, which is first in 0.3.25.

[imciner2]

And there was also a bug in the actual casum kernel, which in this case was the fallback kernel. That is fixed by OpenMathLib/OpenBLAS@f8ad534 (the previous commit I mentioned switches it so the AVX512 kernel will be used instead, which "fixes" this because it uses a different kernel).