【问题标题】:Why are Swift iterators slower than array building?为什么 Swift 迭代器比数组构建慢?
【发布时间】:2017-04-03 00:22:41
【问题描述】:

这与this question 有点相关,其中假设使用生成器(迭代器)遍历嵌套数组对于遍历元素是最佳的,只要您不需要存储结果,如果您只想展平数组,则最好使用重复的数组连接。

然而,我决定做一些测试,并以惰性和存储形式实现这个函数(将包含 Ints 或 [Int]s 的数组 [Any] 展平),结果存储形式是更快,即使只是用于迭代元素!这意味着不知何故,迭代生成器比在内存中构造一个新数组和然后迭代那个花费更多的时间。

令人难以置信的是,它甚至比同一程序的 python 实现慢了大约 5-70%,后者随着输入的减少而恶化。 Swift 是使用 -O 标志构建的。

这里是三个测试用例 1. 小输入,混合; 2.大输入,[Int]显性,3.大输入,Int显性:

斯威夫特

let array1: [Any] = [Array(1...100), Array(101...105), 106, 
                     Array(107...111), 112, 113, 114, Array(115...125)]
let array2: [Any] = Array(repeating: Array(1...5), count: 2000)
let array3: [Any] = Array(repeating: 31, count: 10000)

Python

A1 = [list(range(1, 101)), list(range(101, 106)), 106, 
      list(range(107, 112)), 112, 113, 114, list(range(115, 126))]
A2 = list(range(1, 6)) * 2000
A3 = [31] * 10000

生成器和数组构建器:

斯威夫特

func chain(_ segments: [Any]) -> AnyIterator<Int>{
    var i = 0
    var j = 0
    return AnyIterator<Int> {
        while i < segments.count {
            switch segments[i] {
                case let e as Int:
                    i += 1
                    return e
                case let E as [Int]:
                    if j < E.count {
                        let val = E[j]
                        j += 1
                        return val
                    }
                    j = 0
                    i += 1

                default:
                    return nil
            }
        }
        return nil
    }
}


func flatten_array(_ segments: [Any]) -> [Int] {
    var result = [Int]()
    for segment in segments {
        switch segment {
            case let segment as Int:
                result.append(segment)
            case let segment as [Int]:
                result.append(contentsOf: segment)
            default:
                break
        }
    }
    return result
}

Python

def chain(L):
    for i in L:
        if type(i) is int:
            yield i
        elif type(i) is list:
            yield from i


def flatten_list(L):
    result = []
    for i in L:
        if type(i) is int:
            result.append(i)
        elif type(i) is list:
            result.extend(i)
    return result

以及基准测试结果(第一个测试用例循环 100000 次,其他测试用例循环 1000 次):

斯威夫特

test case 1 (small mixed input)
    Filling an array                         : 0.068221092224121094 s
    Filling an array, and looping through it : 0.074559926986694336 s
    Looping through a generator              : 1.5902719497680664   s *
    Materializing the generator to an array  : 1.759943962097168    s *

test case 2 (large input, [Int] s)
    Filling an array                         : 0.20634698867797852  s
    Filling an array, and looping through it : 0.21031379699707031  s
    Looping through a generator              : 1.3505551815032959   s *
    Materializing the generator to an array  : 1.4733860492706299   s *

test case 3 (large input, Int s)
    Filling an array                         : 0.27392101287841797  s
    Filling an array, and looping through it : 0.27670192718505859  s
    Looping through a generator              : 0.85304021835327148  s
    Materializing the generator to an array  : 1.0027849674224854   s *

Python

test case 1 (small mixed input)
    Filling an array                         : 0.1622014045715332   s
    Filling an array, and looping through it : 0.4312894344329834   s
    Looping through a generator              : 0.6839139461517334   s
    Materializing the generator to an array  : 0.5300459861755371   s

test case 2 (large input, [int] s)
    Filling an array                         : 1.029205083847046    s
    Filling an array, and looping through it : 1.2195289134979248   s
    Looping through a generator              : 1.0876803398132324   s
    Materializing the generator to an array  : 0.8958714008331299   s

test case 3 (large input, int s)
    Filling an array                         : 1.0181667804718018   s
    Filling an array, and looping through it : 1.244570255279541    s
    Looping through a generator              : 1.1220412254333496   s
    Materializing the generator to an array  : 0.9486079216003418   s

显然,Swift 非常非常擅长构建数组。但是为什么它的生成器这么慢,在某些情况下甚至比 Python 的还要慢呢? (在表中用* 标记。)用非常大的输入(> 100,000,000 个元素,这几乎使 Swift 崩溃)进行测试表明,即使在极限情况下,生成器的速度也会比数组填充慢至少一倍最好的情况是 3.25。

如果这真的是该语言固有的,它有一些有趣的含义。例如,常识(对我作为一个 python 程序员来说)如果我们试图合成一个不可变对象(如字符串),我们应该首先将源提供给生成函数以展开它,然后将输出关闭到joined() 方法,该方法适用于单个浅序列。相反,看起来最有效的策略是通过数组进行序列化;将源展开到中间数组,然后从数组构造输出。

构建一个全新的数组,然后迭代它比在原始数组上进行惰性迭代更快吗?为什么?

(Possibly related javascript question)

编辑

这里是测试代码:

斯威夫特

func time(test_array: [Any], cycles: Int = 1000000) -> (array_iterate: Double, 
                                                        array_store  : Double, 
                                                        generate_iterate: Double, 
                                                        generate_store: Double) {
    func start() -> Double { return Date().timeIntervalSince1970 }
    func lap(_ t0: Double) -> Double {
        return Date().timeIntervalSince1970 - t0
    }
    var t0 = start()

    for _ in 0..<cycles {
        for e in flatten_array(test_array) { e + 1 }
    }
    let ΔE1 = lap(t0)

    t0 = start()
    for _ in 0..<cycles {
        let array: [Int] = flatten_array(test_array)
    }
    let ΔE2 = lap(t0)

    t0 = start()
    for _ in 0..<cycles {
        let G = chain(test_array)
        while let g = G.next() { g + 1 }
    }
    let ΔG1 = lap(t0)

    t0 = start()
    for _ in 0..<cycles {
        let array: [Int] = Array(chain(test_array))
    }
    let ΔG2 = lap(t0)

    return (ΔE1, ΔE2, ΔG1, ΔG2)
}

print(time(test_array: array1, cycles: 100000))
print(time(test_array: array2, cycles: 1000))
print(time(test_array: array3, cycles: 1000))

Python

def time_f(test_array, cycles = 1000000):
    lap = lambda t0: time() - t0
    t0 = time()

    for _ in range(cycles):
        for e in flatten_list(test_array):
            e + 1

    ΔE1 = lap(t0)

    t0 = time()
    for _ in range(cycles):
        array = flatten_list(test_array)

    ΔE2 = lap(t0)

    t0 = time()
    for _ in range(cycles):
        for g in chain(test_array):
            g + 1

    ΔG1 = lap(t0)

    t0 = time()
    for _ in range(cycles):
        array = list(chain(test_array))

    ΔG2 = lap(t0)

    return ΔE1, ΔE2, ΔG1, ΔG2

print(time_f(A1, cycles=100000))
print(time_f(A3, cycles=1000))
print(time_f(A2, cycles=1000))

【问题讨论】:

  • 你的完整测试代码是什么?优化可能很棘手,因为有些行可能会被编译器优化掉
  • @CodeDifferent 添加到问题中

标签: arrays swift performance generator


【解决方案1】:

你问“为什么它的 [Swift] 生成器这么慢,在某些情况下甚至比 Python 还慢?”

我的回答是,我认为它们几乎没有您的结果显示的那么慢。特别是,我将尝试证明循环遍历迭代器应该比为所有测试用例构造数组更快。

在早期的工作中(请参阅http://lemire.me/blog/2016/09/22/swift-versus-java-the-bitset-performance-test/ 的相关博客文章),我发现在处理 bitset 类时,Swift 迭代器的速度大约是 Java 中的等效迭代器的一半。这不是很好,但 Java 在这方面非常有效。与此同时,Go 的表现更差。我向您提出,Swift 迭代器的效率可能不是理想的,但它们可能是原始 C 代码可能的两倍之内。而性能差距可能与 Swift 中的函数内联不足有关。

我看到您使用的是AnyIterator。我建议从IteratorProtocol 派生一个struct,而不是这样做的好处是确保不必进行任何动态调度。这是一个相对有效的可能性:

public struct FastFlattenIterator: IteratorProtocol {
   let segments: [Any]
    var i = 0 // top-level index
    var j = 0 // second-level index
    var jmax = 0 // essentially, this is currentarray.count, but we buffer it
    var currentarray : [Int]! // quick reference to an int array to be flatten

   init(_ segments: [Any]) {
       self.segments = segments
   }

   public mutating func next() -> Int? {
     if j > 0 { // we handle the case where we iterate within an array separately
       let val = currentarray[j]
       j += 1
       if j == jmax {
         j = 0
         i += 1
       }
       return val
     }
     while i < segments.count {
        switch segments[i] {
          case let e as Int: // found an integer value
            i += 1
            return e
          case let E as [Int]: // first encounter with an array
            jmax = E.count
            currentarray = E
            if jmax > 0 {
              j = 1
              return E[0]
            }
            i += 1
          default:
            return nil
        }
     }
     return nil
   }
}

通过这门课,我得到了以下数字。对于每个测试用例,前四种方法取自您的代码示例,而后两种(快速迭代器)是使用新结构构建的。请注意,“循环通过快速迭代器”总是最快的。

test case 1 (small mixed input)
Filling an array                         : 0.0073099999999999997 ms
Filling an array, and looping through it : 0.0069870000000000002 ms
Looping through a generator              : 0.18385799999999999   ms 
Materializing the generator to an array  : 0.18745700000000001   ms 
Looping through a fast iterator          : 0.005372              ms 
Materializing the fast iterator          : 0.015883999999999999  ms

test case 2 (large input, [Int] s)
Filling an array                         : 2.125931            ms
Filling an array, and looping through it : 2.1169820000000001  ms
Looping through a generator              : 15.064767           ms 
Materializing the generator to an array  : 15.45152            ms 
Looping through a fast iterator          : 1.572919            ms
Materializing the fast iterator          : 1.964912            ms 

test case 3 (large input, Int s)
Filling an array                         : 2.9140269999999999  ms
Filling an array, and looping through it : 2.9064290000000002  ms
Looping through a generator              : 9.8297640000000008  ms
Materializing the generator to an array  : 9.8297640000000008  ms 
Looping through a fast iterator          : 1.978038            ms 
Materializing the fast iterator          : 2.2565339999999998  ms 

你可以在 GitHub 上找到我的完整代码示例:https://github.com/lemire/Code-used-on-Daniel-Lemire-s-blog/tree/master/extra/swift/iterators

【讨论】:

  • 这对我来说很有趣,因为我认为创建struct 会比返回闭包引入更多开销。以及显然您可以迭代 并 保存快速迭代器的输出的时间比以“正常方式”构建数组所需的时间更短......
猜你喜欢
  • 2019-07-30
  • 1970-01-01
  • 1970-01-01
  • 2019-04-09
  • 2013-12-12
  • 1970-01-01
  • 2015-10-20
  • 2021-01-07
  • 1970-01-01
相关资源
最近更新 更多