在Julia中优化高度矢量化的代码?

时间:2019-01-15 20:34:29

标签: optimization julia

我想在Julia中优化以下代码,它以高度矢量化的形式编写,而MATLAB等语言则表现出色。 MATLAB中完全相同的代码花费Elapsed time is 0.277608 seconds.,速度提高了2.8倍,所以我认为可以在Julia中完成一些工作。公平地说,我注意到MATLAB默认使用多线程,因此,如果在Julia中也启用了多线程也没问题。谢谢您的帮助。

function fit_xlin(x, y, w)
    n = length(x)
    regularization = 1.0e-5
    xx_0_0 = fill(sum(w.*1)   , n)
    xx_1_0 = fill(sum(w.*x)   , n)
    xx_0_1 = fill(sum(w.*x)   , n)
    xx_1_1 = fill(sum(w.*x.*x), n)
    xy_0   = fill(sum(w.*y)   , n)
    xy_1   = fill(sum(w.*x.*y), n)
    xx_1_0 .+= regularization
    xx_0_1 .+= regularization

    xxk_0_0 = xx_0_0 .- w.*1
    xxk_1_0 = xx_1_0 .- w.*x
    xxk_0_1 = xx_0_1 .- w.*x
    xxk_1_1 = xx_1_1 .- w.*x.*x
    xyk_0   = xy_0   .- w.*y
    xyk_1   = xy_1   .- w.*x.*y

    det = xxk_0_0.*xxk_1_1 .- xxk_0_1.*xxk_1_0
    c0  = (xxk_1_1.*xyk_0  .- xxk_0_1.*xyk_1)./det
    c1  = (-xxk_1_0.*xyk_0 .+ xxk_0_0.*xyk_1)./det

    y_est = c0 .+ c1.*x
end 

using BenchmarkTools
function test_xlin()
    x = rand( 0.0:4.0, 5000_000)
    y = rand( 0.0:4.0, 5000_000)
    w = rand( 0.0:4.0, 5000_000)
    @btime fit_xlin($x, $y, $w)
end 

这次是:

    julia> test_xlin();
      775.292 ms (46 allocations: 877.38 MiB)

1 个答案:

答案 0 :(得分:3)

这是仍在向量化的代码,它不使用多线程。在我的计算机上,它的速度是原始计算机的两倍多,但是仍然可以在此处完成操作。

顺便说一句,我严重怀疑Matlab代码是否经过特别优化,因为正在进行一些非常浪费的操作-不必要的分配,不必要的操作(sum(w.*1)确实很糟糕,将数组乘以1并分配一个此过程中多余的数组非常浪费:)此外,您也不需要在Matlab中分配任何矢量xx_0_0xx_1_0,您可以像在Julia中那样使用广播。

无论如何,这是我的第一次尝试:

function fit_xlin2(x, y, w)
    regularization = 1.0e-5

    sumwx = (w' * x) + regularization
    sumwy = (w' * y)
    sumwxx = sum(a[1]*a[2]^2 for a in zip(w, x))
    sumwxy = sum(prod, zip(w, x, y))

    wx = w .* x
    xxk_0_0 = sum(w) .- w
    xxk_1_0 = sumwx .- wx
    xxk_1_1 = sumwxx .- wx .* x
    xyk_0 = sumwy .- w .* y
    xyk_1 = sumwxy .- wx .* y

    det = xxk_0_0 .* xxk_1_1 .- xxk_1_0 .* xxk_1_0
    c0  = (xxk_1_1 .* xyk_0  .- xxk_1_0 .* xyk_1)./det
    c1  = (-xxk_1_0 .* xyk_0 .+ xxk_0_0 .* xyk_1)./det

    return c0 .+ c1 .* x
end

编辑:您可以通过对主循环进行矢量化处理来获得一定程度的加速。该代码比原始的Julia代码快17倍,仍然是单线程的,并且可读性强:

function fit_xlin_loop(x, y, w)
    if !(size(x) == size(y) == size(w))
        error("Input vectors must have the same size.")
    end

    regularization = 1.0e-5

    sumw = sum(w)
    sumwx = (w' * x) + regularization
    sumwy = (w' * y)
    sumwxx = sum(a[1]*a[2]^2 for a in zip(w, x))
    sumwxy = sum(prod, zip(w, x, y))

    y_est = similar(x)
    @inbounds for i in eachindex(y_est)
        wx = w[i] * x[i]
        xxk_0_0 = sumw - w[i]
        xxk_1_0 = sumwx - wx
        xxk_1_1 = sumwxx - wx * x[i]
        xyk_0 = sumwy - w[i] * y[i]
        xyk_1 = sumwxy - wx * y[i]

        det = xxk_0_0 * xxk_1_1 - xxk_1_0 * xxk_1_0
        c0  = (xxk_1_1 * xyk_0 - xxk_1_0 * xyk_1) / det
        c1  = (-xxk_1_0 * xyk_0 + xxk_0_0 * xyk_1) / det

        y_est[i] = c0 + c1 * x[i]
    end
    return y_est
end