Test that we can use Apple Accelerate by default on Mac

[amontoison]

It's documented nowhere but I found this in the tests of LBT:
https://github.com/JuliaLinearAlgebra/libblastrampoline/blob/main/test/runtests.jl#L150-L165

We have CI with a M1 so we can test it.

[dpo]

The last time I checked, LAPACK was incomplete in Accelerate.

[dpo]

@JSOBot runtests

[codecov]

Codecov Report

Base: 78.27% // Head: 81.59% // Increases project coverage by +3.32% 🎉

Coverage data is based on head (c843f00) compared to base (658bc69).
Patch coverage: 33.33% of modified lines in pull request are covered.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #124      +/-   ##
==========================================
+ Coverage   78.27%   81.59%   +3.32%     
==========================================
  Files           5        5              
  Lines         359      364       +5     
==========================================
+ Hits          281      297      +16     
+ Misses         78       67      -11     

Impacted Files	Coverage Δ
src/HSL.jl	58.33% <33.33%> (-27.39%)	⬇️
src/mc21.jl	93.75% <0.00%> (+93.75%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[dpo]

On another note, I don't see the MC21 tests in the output anymore... 🤔

[amontoison]

Is it possible to test the following code on your computer Dominique?
You don't need the modifications in this PR to run it.

# Current version of HSL
using HSL, MatrixMarket, SuiteSparseMatrixCollection
using LinearAlgebra, Printf, BenchmarkTools

ssmc = ssmc_db(verbose=false)
matrix = ssmc_matrices(ssmc, "Boeing", "pwtk")
path = fetch_ssmc(matrix, format="MM")

n = matrix.nrows[1]
A = MatrixMarket.mmread(joinpath(path[1], "$(matrix.name[1]).mtx"))
b = ones(n)

# Solve Ax = b.
LDL = @btime Ma57($A)
@btime ma57_factorize!($LDL)
@btime ma57_solve($LDL, $b)

blas = "/System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/libBLAS.dylib"
lapack = "/System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/libLAPACK.dylib"
LinearAlgebra.BLAS.lbt_forward(blas, clear=true, verbose=true)
LinearAlgebra.BLAS.lbt_forward(lapack, clear=false, verbose=true)

# Solve Ax = b.
LDL = @btime Ma57($A)
@btime ma57_factorize!($LDL)
@btime ma57_solve($LDL, $b)

[amontoison]

On another note, I don't see the MC21 tests in the output anymore... thinking

I changed HSL_MC21_PATH into MC21_PATH with #119.
You could use the new variable HSL_ARCHIVES_PATH now.

[JSOBot]

Testing HSL tests passed: https://gist.github.com/f1c96b39f37c7cc4d81f99f1343b9d46

[dpo]

julia> @btime ma57_factorize($LDL)
ERROR: MethodError: no method matching ma57_factorize(::Ma57{Float64})
Closest candidates are:
  ma57_factorize(::SparseArrays.SparseMatrixCSC{T, Ti}; kwargs...) where {T<:Union{Float32, Float64}, Ti<:Integer} at ~/.julia/packages/HSL/X5ozr/src/hsl_ma57.jl:472

[amontoison]

julia> @btime ma57_factorize($LDL)
ERROR: MethodError: no method matching ma57_factorize(::Ma57{Float64})
Closest candidates are:
  ma57_factorize(::SparseArrays.SparseMatrixCSC{T, Ti}; kwargs...) where {T<:Union{Float32, Float64}, Ti<:Integer} at ~/.julia/packages/HSL/X5ozr/src/hsl_ma57.jl:472

It's ma57_factorize!.

[dpo]

julia> LDL = @btime Ma57($A)
  499.216 ms (40 allocations: 345.10 MiB)

julia> @btime ma57_factorize!($LDL)
  2.054 s (0 allocations: 0 bytes)

julia> @btime ma57_solve($LDL, $b)
  40.659 ms (4 allocations: 3.33 MiB)

After the BLAS change

julia> LDL = @btime Ma57($A)
  503.579 ms (40 allocations: 345.10 MiB)

julia> @btime ma57_factorize!($LDL)
  1.976 s (0 allocations: 0 bytes)

julia> @btime ma57_solve($LDL, $b)
  45.326 ms (4 allocations: 3.33 MiB)

[dpo]

@JSOBot runtests

[amontoison]

julia> LDL = @btime Ma57($A)
  499.216 ms (40 allocations: 345.10 MiB)

julia> @btime ma57_factorize!($LDL)
  2.054 s (0 allocations: 0 bytes)

julia> @btime ma57_solve($LDL, $b)
  40.659 ms (4 allocations: 3.33 MiB)

After the BLAS change

julia> LDL = @btime Ma57($A)
  503.579 ms (40 allocations: 345.10 MiB)

julia> @btime ma57_factorize!($LDL)
  1.976 s (0 allocations: 0 bytes)

julia> @btime ma57_solve($LDL, $b)
  45.326 ms (4 allocations: 3.33 MiB)

OpenBLAS is already efficient 🤔 , we could keep OpenBLAS32 by default and just add in the documentation how to load MKL or Apple Accelerate (#98).

[dpo]

Yes, that sounds good.

[dpo]

@JSOBot runtests

[dpo]

The error with JSOBot is

ERROR: Unsatisfiable requirements detected for package libblastrampoline_jll [8e850b90]:
 libblastrampoline_jll [8e850b90] log:
 ├─libblastrampoline_jll [8e850b90] has no known versions!
 └─restricted to versions * by HSL [34c5aeac] — no versions left
   └─HSL [34c5aeac] log:
     ├─possible versions are: 0.3.2 or uninstalled
     └─HSL [34c5aeac] is fixed to version 0.3.2

It's stuck on Julia 1.6.

[amontoison]

The error with JSOBot is

ERROR: Unsatisfiable requirements detected for package libblastrampoline_jll [8e850b90]:
 libblastrampoline_jll [8e850b90] log:
 ├─libblastrampoline_jll [8e850b90] has no known versions!
 └─restricted to versions * by HSL [34c5aeac] — no versions left
   └─HSL [34c5aeac] log:
     ├─possible versions are: 0.3.2 or uninstalled
     └─HSL [34c5aeac] is fixed to version 0.3.2

It's stuck on Julia 1.6.

LBT_jll is shipped with Julia, just like OpenBLAS_jll. I don't understand why we have this error. 🤔

[dpo]

Could you test locally with 1.6?

[amontoison]

Could you test locally with 1.6?

I just tested and it works fine.

[dpo]

what do we do with this?

[amontoison]

We close it and I will add documentation about LBT (#98)