为什么这个 F# 代码这么慢? [英] Why is this F# code so slow?
问题描述
在 C# 和 F# 中的 Levenshtein 实现.对于大约 1500 个字符的两个字符串,C# 版本的速度提高了 10 倍.C#:69 毫秒,F# 867 毫秒.为什么?据我所知,他们做的事情完全一样?无论是发布版本还是调试版本都没有关系.
A Levenshtein implementation in C# and F#. The C# version is 10 times faster for two strings of about 1500 chars. C#: 69 ms, F# 867 ms. Why? As far as I can tell, they do the exact same thing? Doesn't matter if it is a Release or a Debug build.
如果有人来这里专门寻找编辑距离"实现,那么它就坏了.工作代码在这里.
If anyone comes here looking specifically for the Edit Distance implementation, it is broken. Working code is here.
C#:
private static int min3(int a, int b, int c)
{
return Math.Min(Math.Min(a, b), c);
}
public static int EditDistance(string m, string n)
{
var d1 = new int[n.Length];
for (int x = 0; x < d1.Length; x++) d1[x] = x;
var d0 = new int[n.Length];
for(int i = 1; i < m.Length; i++)
{
d0[0] = i;
var ui = m[i];
for (int j = 1; j < n.Length; j++ )
{
d0[j] = 1 + min3(d1[j], d0[j - 1], d1[j - 1] + (ui == n[j] ? -1 : 0));
}
Array.Copy(d0, d1, d1.Length);
}
return d0[n.Length - 1];
}
F#:
let min3(a, b, c) = min a (min b c)
let levenshtein (m:string) (n:string) =
let d1 = Array.init n.Length id
let d0 = Array.create n.Length 0
for i=1 to m.Length-1 do
d0.[0] <- i
let ui = m.[i]
for j=1 to n.Length-1 do
d0.[j] <- 1 + min3(d1.[j], d0.[j-1], d1.[j-1] + if ui = n.[j] then -1 else 0)
Array.blit d0 0 d1 0 n.Length
d0.[n.Length-1]
推荐答案
问题在于 min3
函数被编译为使用泛型比较的泛型函数(我认为这仅使用了 IComparable
,但它实际上更复杂 - 它会使用 F# 类型的结构比较,而且逻辑相当复杂.
The problem is that the min3
function is compiled as a generic function that uses generic comparison (I thought this uses just IComparable
, but it is actually more complicated - it would use structural comparison for F# types and it's fairly complex logic).
> let min3(a, b, c) = min a (min b c);;
val min3 : 'a * 'a * 'a -> 'a when 'a : comparison
在 C# 版本中,该函数不是通用的(它只需要 int
).您可以通过添加类型注释来改进 F# 版本(以获得与 C# 中相同的内容):
In the C# version, the function is not generic (it just takes int
). You can improve the F# version by adding type annotations (to get the same thing as in C#):
let min3(a:int, b, c) = min a (min b c)
...或者通过将 min3
设为 inline
(在这种情况下,它会在使用时专门用于 int
):>
...or by making min3
as inline
(in which case, it will be specialized to int
when used):
let inline min3(a, b, c) = min a (min b c);;
对于长度为 300 的随机字符串 str
,我得到以下数字:
For a random string str
of length 300, I get the following numbers:
> levenshtein str ("foo" + str);;
Real: 00:00:03.938, CPU: 00:00:03.900, GC gen0: 275, gen1: 1, gen2: 0
val it : int = 3
> levenshtein_inlined str ("foo" + str);;
Real: 00:00:00.068, CPU: 00:00:00.078, GC gen0: 0, gen1: 0, gen2: 0
val it : int = 3
这篇关于为什么这个 F# 代码这么慢?的文章就介绍到这了,希望我们推荐的答案对大家有所帮助,也希望大家多多支持IT屋!