Julia 中 for 循环的性能

Question

我尝试了以下代码，但性能差异很大在代码之间。听说顶层的代码不适合数值计算，但是性能好像也取决于是否顶级变量（此处为 N）出现在 for 循环的范围内。避免此类顶级变量总是更好吗？

N = 10000000

# case1 (slow)
x = 0.0
@time for k = 1:N
    x += float( k )
end

# case2 (slow)
@time let
    y = 0.0
    for j = 1:N
        y += float( j )
    end
end

# case3 (very fast)
@time let
    n::Int64
    n = N
    z = 0.0
    for m = 1:n
        z += float( m )
    end
end

# case 4 (slow)
function func1()
  c = 0.0
  for i = 1:N                                             
    c += float( i )
  end
end

# case 5 (fast)
function func2( n )                                          
  c = 0.0
  for i = 1:n
    c += float( i )
  end
end

# case 6 (fast)
function func3()
  n::Int
  n = N
  c = 0.0
  for i = 1:n                                   
    c += float( i )
  end
end

# case 7 (slow)
function func4()
  n = N        # n = int( N ) is also slow                                  
  c = 0.0
  for i = 1:n
    c += float( i )
  end
end

@time func1()
@time func2( N )
@time func3()
@time func4()

使用 Julia 0.3.7（在 Linux x86_64 上）获得的结果是

elapsed time: 2.595440598 seconds (959985496 bytes allocated, 10.70% gc time)
elapsed time: 2.469471127 seconds (959983688 bytes allocated, 11.49% gc time)
elapsed time: 1.608e-6 seconds (16 bytes allocated)
elapsed time: 2.535243279 seconds (960021976 bytes allocated, 11.21% gc time)
elapsed time: 0.002601149 seconds (75592 bytes allocated)
elapsed time: 0.003471583 seconds (84456 bytes allocated)
elapsed time: 2.480343146 seconds (960020752 bytes allocated, 11.48% gc time)

Answer 1

采用"Is it always better to avoid such top-level variables?"字面量的答案当然是"No, it depends"，但有用的注释是将全局变量声明为常量

const N = 10000000

使案例 2 与案例 3 一样快。

编辑：我应该补充一点，案例 2 的问题是顶层 N 使得范围 1:N 和迭代器变量 j 的类型不稳定，即使累加器变量 y 是本地的。这个问题更灵活的解决方法是

   let
       y = 0.0
       for j = 1:(N::Int)
           y += float( j )
       end
       y
   end

Julia 中 for 循环的性能

Performance of for loops in Julia

performance

for-loop

julia