dgemm_bench.jl

if !isdir(Pkg.dir("Benchmarks"))
    Pkg.clone("https://github.com/johnmyleswhite/Benchmarks.jl")
end
using Benchmarks
versioninfo()
# atlas lib from github.com/matthew-brett/np-wheel-builder
# commit 35038507f0d5860955f0bb4602982867d888e3c9
# github.com/flame/blis commit 0b126de1342c11c65623bcb38e258e21e9244e3d
# built in cygwin64 via
# mkdir buildref01 && cd buildref01 && ../configure reference
# make -j8 CC=x86_64-w64-mingw32-gcc CPICFLAGS="" BLIS_ENABLE_DYNAMIC_BUILD=yes
# mkdir ../buildsandy02 && cd ../buildsandy02 && ../configure sandybridge
# make -j8 CC=x86_64-w64-mingw32-gcc CPICFLAGS="" BLIS_ENABLE_DYNAMIC_BUILD=yes

toppath = "C:/cygwin64/home/Tony/github" # CHANGEME
ENV["PATH"] *= ";C:/cygwin64/usr/x86_64-w64-mingw32/sys-root/mingw/bin" # libgomp etc
for (func, BlasInt, lib) in
        ((:dgemm_atlas, Int32, "$toppath/np-wheel-builder/atlas-builds/atlas-3.10.1-sse2-64/lib/numpy-atlas.dll"),
         (:dgemm_blisref, Int64, "$toppath/blis/buildref01/lib/reference/libblis.so"),
         (:dgemm_blissandy, Int64, "$toppath/blis/buildsandy02/lib/sandybridge/libblis.so"),
         (:dgemm_openblas, Int64, "libopenblas64_"))
    if lib == "libopenblas64_"
        # Julia's ILP64 symbol renaming convention
        dgemm = Libdl.dlsym(Libdl.dlopen(lib), :dgemm_64_)
    else
        dgemm = Libdl.dlsym(Libdl.dlopen(lib), :dgemm_)
    end
    @eval function $func(A, B)
        transa = transb = 'N'
        (M, K) = size(A)
        (K2, N) = size(B)
        @assert K == K2
        @assert eltype(A) == eltype(B) == Float64
        C = Array(eltype(A), M, N)
        alpha = convert(eltype(A), 1.0)
        beta = convert(eltype(A), 0.0)
        lda = M
        ldb = K
        ldc = M
        ccall($dgemm, Void, (Ptr{Cchar}, Ptr{Cchar},
            Ptr{$BlasInt}, Ptr{$BlasInt}, Ptr{$BlasInt},
            Ptr{eltype(A)}, Ptr{eltype(A)}, Ptr{$BlasInt},
            Ptr{eltype(A)}, Ptr{$BlasInt}, Ptr{eltype(A)},
            Ptr{eltype(A)}, Ptr{$BlasInt}),
            &transa, &transb, &M, &N, &K,
            &alpha, A, &lda, B, &ldb, &beta, C, &ldc)
        return C
    end
end
A = rand(2000, 2000)
B = rand(2000, 2000)
Catlas = dgemm_atlas(A, B)
Cblisref = dgemm_blisref(A, B)
Cblissandy = dgemm_blissandy(A, B)
Copenblas = dgemm_openblas(A, B)
@show @benchmark dgemm_atlas(A, B)
@show @benchmark dgemm_blisref(A, B)
@show @benchmark dgemm_blissandy(A, B)
@show @benchmark dgemm_openblas(A, B)
@show maximum(abs(Catlas - Cblisref))
@show maximum(abs(Catlas - Cblissandy))
@show maximum(abs(Catlas - Copenblas))

output.log

Julia Version 0.4.3
Commit a2f713d (2016-01-12 21:37 UTC)
Platform Info:
  System: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.3
@benchmark(dgemm_atlas(A,B)) = ================ Benchmark Results ========================
     Time per evaluation: 2.76 s [2.64 s, 2.89 s]
Proportion of time in GC: 0.00% [0.00%, 0.00%]
        Memory allocated: 30.52 mb
   Number of allocations: 2 allocations
       Number of samples: 2
   Number of evaluations: 2
 Time spent benchmarking: 8.43 s

@benchmark(dgemm_blisref(A,B)) = ================ Benchmark Results ========================
Warning: function may not have been precompiled
     Time per evaluation: 5.24 s
Proportion of time in GC: 0.00%
        Memory allocated: 30.52 mb
   Number of allocations: 2 allocations
       Number of samples: 1
   Number of evaluations: 1
 Time spent benchmarking: 5.28 s

@benchmark(dgemm_blissandy(A,B)) = ================ Benchmark Results ========================
     Time per evaluation: 658.65 ms [643.63 ms, 673.66 ms]
Proportion of time in GC: 0.28% [0.00%, 0.57%]
        Memory allocated: 30.52 mb
   Number of allocations: 2 allocations
       Number of samples: 14
   Number of evaluations: 14
 Time spent benchmarking: 9.98 s

@benchmark(dgemm_openblas(A,B)) = ================ Benchmark Results ========================
     Time per evaluation: 119.70 ms [114.81 ms, 124.59 ms]
Proportion of time in GC: 1.93% [1.54%, 2.33%]
        Memory allocated: 30.52 mb
   Number of allocations: 2 allocations
       Number of samples: 84
   Number of evaluations: 84
 Time spent benchmarking: 10.27 s

maximum(abs(Catlas - Cblisref)) = 3.0127011996228248e-12
maximum(abs(Catlas - Cblissandy)) = 3.0127011996228248e-12
maximum(abs(Catlas - Copenblas)) = 3.0127011996228248e-12