BLAS.axpy! と forループ愚直計算はどっちが速いのか実験

blas2.jl

function perf(a=1.0, N=1000; n=100)
    srand(98765)
    X = rand(N, N)
    Y = rand(N, N)

    println("N=$N")
    print("  BLAS: ")
    gc_enable(false)
    gc()
    @time for i in 1:n
        BLAS.axpy!(a, X, Y)
    end

    print("  for:  ")
    gc()
    @time begin
        for j in 1:n
            for i in eachindex(X)
                @inbounds Y[i] = a*X[i] + Y[i]
            end
        end
    end
    gc_enable(true)
end

versioninfo()

for N in [100, 1000, 3000, 5000, 10000]
   perf(5.0, N, n=1000)
end

実行結果：

% julia-master blas2.jl
Julia Version 0.5.0-dev+1256
Commit 5393150* (2015-11-12 04:42 UTC)
Platform Info:
  System: Darwin (x86_64-apple-darwin14.4.0)
  CPU: Intel(R) Core(TM) i5-4258U CPU @ 2.40GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.8.0svn
N=100
  BLAS:   0.005899 seconds
  for:    0.004070 seconds
N=1000
  BLAS:   1.553393 seconds
  for:    1.450331 seconds
N=3000
  BLAS:  12.617937 seconds
  for:   13.150426 seconds
N=5000
  BLAS:  34.130196 seconds
  for:   36.054273 seconds
N=10000
  BLAS: 137.215003 seconds
  for:  169.501480 seconds