.NET 6 数组拷贝性能对比
本文來(lái)對(duì)比多個(gè)不同的方法進(jìn)行數(shù)組拷貝,和測(cè)試其性能
測(cè)試性能必須采用基準(zhǔn)(標(biāo)準(zhǔn))性能測(cè)試方法,否則測(cè)試結(jié)果不可信。在 dotnet 里面,可以采用?BenchmarkDotNet?進(jìn)行性能測(cè)試。詳細(xì)請(qǐng)看?C# 標(biāo)準(zhǔn)性能測(cè)試
拷貝某個(gè)數(shù)組的從某個(gè)起始點(diǎn)加上某個(gè)長(zhǎng)度的數(shù)據(jù)到另一個(gè)數(shù)組里面,可選方法有很多,本文僅列舉出使用?for?循環(huán)拷貝,和使用 Array.Copy 方法和用 Span 方法進(jìn)行拷貝進(jìn)行對(duì)比
假定有需要被拷貝的數(shù)組是?TestData?其定義如下
static Program(){TestData = new int[1000];for (int i = 0; i < 1000; i++){TestData[i] = i;}}private static readonly int[] TestData;使用?for?循環(huán)拷貝的方法如下
public object CopyByFor(int start, int length){var rawPacketData = TestData;var data = new int[length];for (int localIndex = 0, rawArrayIndex = start; localIndex < data.Length; localIndex++, rawArrayIndex++){data[localIndex] = rawPacketData[rawArrayIndex];}return data;}以上代碼返回 data 作為 object 僅僅只是為了做性能測(cè)試,避免被 dotnet 優(yōu)化掉
另一個(gè)拷貝數(shù)組是采用?Array.Copy?拷貝,邏輯如下
public object CopyByArray(int start, int length){var rawPacketData = TestData;var data = new int[length];Array.Copy(rawPacketData,start,data,0, length);return data;}采用新的 dotnet 提供的 Span 進(jìn)行拷貝,代碼如下
public object CopyBySpan(int start, int length){var rawPacketData = TestData;var rawArrayStartIndex = start;var data = rawPacketData.AsSpan(rawArrayStartIndex, length).ToArray();return data;}接著加上一些性能調(diào)試輔助邏輯
[Benchmark][ArgumentsSource(nameof(ProvideArguments))]public object CopyByFor(int start, int length){var rawPacketData = TestData;var data = new int[length];for (int localIndex = 0, rawArrayIndex = start; localIndex < data.Length; localIndex++, rawArrayIndex++){data[localIndex] = rawPacketData[rawArrayIndex];}return data;}[Benchmark][ArgumentsSource(nameof(ProvideArguments))]public object CopyByArray(int start, int length){var rawPacketData = TestData;var data = new int[length];Array.Copy(rawPacketData,start,data,0, length);return data;}public IEnumerable<object[]> ProvideArguments(){foreach (var start in new[] { 0, 10, 100 }){foreach (var length in new[] { 10, 20, 100 }){yield return new object[] { start, length };}}}在我的設(shè)備上的測(cè)試效果如下
BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19042.1200 (20H2/October2020Update) Intel Core i7-9700K CPU 3.60GHz (Coffee Lake), 1 CPU, 8 logical and 8 physical cores .NET SDK=6.0.100-preview.7.21379.14[Host] : .NET 6.0.0 (6.0.21.37719), X64 RyuJITDefaultJob : .NET 6.0.0 (6.0.21.37719), X64 RyuJIT可以看到,在對(duì)比使用?for?循環(huán)拷貝和使用?Array.Copy?拷貝中,使用?Array.Copy?拷貝的性能更好,在拷貝的數(shù)組長(zhǎng)度越長(zhǎng)的時(shí)候,使用 Array.Copy 拷貝性能優(yōu)勢(shì)就更好
接下來(lái)再加上 Span 的性能比較,如下面代碼
[Benchmark][ArgumentsSource(nameof(ProvideArguments))]public object CopyBySpan(int start, int length){var rawPacketData = TestData;var rawArrayStartIndex = start;var data = rawPacketData.AsSpan(rawArrayStartIndex, length).ToArray();return data;}性能對(duì)比測(cè)試如下
可以看到 Span 的性能比?Array.Copy?拷貝性能更強(qiáng)
在 Span 里面,轉(zhuǎn)換為數(shù)組的邏輯如下
[MethodImpl(MethodImplOptions.AggressiveInlining)]public T[] ToArray(){if (_length == 0)return Array.Empty<t>();var destination = new T[_length];Buffer.Memmove(ref MemoryMarshal.GetArrayDataReference(destination), ref _pointer.Value, (nuint)_length);return destination;}這里使用到的 Buffer 的有黑科技的 Memmove 方法,此方法的實(shí)現(xiàn)如下
[MethodImpl(MethodImplOptions.AggressiveInlining)]internal static void Memmove<t>(ref T destination, ref T source, nuint elementCount){if (!RuntimeHelpers.IsReferenceOrContainsReferences<t>()){// Blittable memmoveMemmove(ref Unsafe.As<t, byte="">(ref destination),ref Unsafe.As<t, byte="">(ref source),elementCount * (nuint)Unsafe.SizeOf<t>());}else{// Non-blittable memmoveBulkMoveWithWriteBarrier(ref Unsafe.As<t, byte="">(ref destination),ref Unsafe.As<t, byte="">(ref source),elementCount * (nuint)Unsafe.SizeOf<t>());}}以上性能測(cè)試使用的是 int 數(shù)組,剛好能進(jìn)入 Memmove 的分支,而不是 BulkMoveWithWriteBarrier 這個(gè)分支。在里層的 Memmove 方法里面用到了很多黑科技,本文只是用來(lái)對(duì)比多個(gè)方法拷貝數(shù)組的性能,黑科技部分就需要大家自己去閱讀 dotnet 的源代碼啦
另外,如果需要做完全的數(shù)組的拷貝,數(shù)組里面存放的是值類型對(duì)象,如 int 類型,那么拷貝整個(gè)數(shù)組還有另一個(gè)可選項(xiàng)是通過(guò)?Clone?方法進(jìn)行拷貝,代碼如下
public object CopyByClone(){var data = (int[]) TestData.Clone();return data;}使用?Clone?的方法的行為是返回?cái)?shù)組的淺表拷貝,也就是說(shuō)數(shù)組里面的元素沒(méi)有做深拷貝,只是拷貝數(shù)組本身而已。對(duì)于值類型來(lái)說(shuō),就沒(méi)有啥問(wèn)題了
稍微更改一下性能測(cè)試,更改的代碼如下
[MemoryDiagnoser]public class Program{static void Main(string[] args){BenchmarkRunner.Run<program>();}static Program(){TestData = new int[1000];for (int i = 0; i < 1000; i++){TestData[i] = i;}}[Benchmark]public object CopyByFor(){var rawPacketData = TestData;var length = TestData.Length;var data = new int[length];for (int localIndex = 0, rawArrayIndex = 0; localIndex < data.Length; localIndex++, rawArrayIndex++){data[localIndex] = rawPacketData[rawArrayIndex];}return data;}[Benchmark]public object CopyByArray(){var length = TestData.Length;var start = 0;var rawPacketData = TestData;var data = new int[length];Array.Copy(rawPacketData,start,data,0, length);return data;}[Benchmark]public object CopyByClone(){var data = (int[]) TestData.Clone();return data;}private static readonly int[] TestData;}通過(guò)下圖可以了解到采用 Clone 方法和采用 Array.Copy 方法的性能差不多,但 Clone 稍微快一點(diǎn)
以上是給 WPF 框架做性能優(yōu)化時(shí)測(cè)試的,詳細(xì)請(qǐng)看
Using?Array.Copy?to make array copy faster in StylusPointCollection by lindexi · Pull Request #5217 · dotnet/wpf
Using the?Clone?method to fast clone the array in StylusPoint by lindexi · Pull Request #5218 · dotnet/wpf
特別感謝ThomasGoulet73大佬教我使用 AsSpan 的方法拷貝數(shù)組
總結(jié)
以上是生活随笔為你收集整理的.NET 6 数组拷贝性能对比的全部?jī)?nèi)容,希望文章能夠幫你解決所遇到的問(wèn)題。
- 上一篇: Resharper 和 Rider 的奇
- 下一篇: Dapr + .NET Core实战(二