c# - Mono SIMD 性能恶化？

Question

基准代码:

using System;
using System.Collections;
using System.Collections.Generic;
using System.Diagnostics;
using System.Linq;
using Mono.Simd;
using MathNet.Numerics.LinearAlgebra.Single;

namespace XXX {
public static class TimeSpanExtensions {
    public static double TotalNanoseconds(this TimeSpan timeSpan) {
        return timeSpan.TotalMilliseconds * 1000000.0;
    }
}

public sealed class SimdBenchmark : Benchmark {
    Vector4f a = new Vector4f(1.0f, 2.0f, 3.0f, 4.0f);
    Vector4f b = new Vector4f(1.0f, 2.0f, 3.0f, 4.0f);
    Vector4f c;

    public override void Do() {
        c = a + b;
    }
}

public sealed class MathNetBenchmark : Benchmark {
    DenseVector a = new DenseVector(new float[]{1.0f,2.0f,3.0f,4.0f});
    DenseVector b = new DenseVector(new float[]{1.0f,2.0f,3.0f,4.0f});
    DenseVector c;

    public override void Do() {
        c = a + b;
    }
}

public sealed class DefaultBenchmark : Benchmark {
    Vector4 a = new Vector4(1.0f, 2.0f, 3.0f, 4.0f);
    Vector4 b = new Vector4(1.0f, 2.0f, 3.0f, 4.0f);
    Vector4 c;

    public override void Do() {
        c = a + b;
    }
}

public sealed class SimpleBenchmark : Benchmark {
    float a = 1.0f;
    float b = 2.0f;
    float c;

    public override void Do() {
        c = a + b;
    }
}

public sealed class DelegateBenchmark : Benchmark {
    private readonly Action _action;

    public DelegateBenchmark(Action action) {
        _action = action;
    }

    public override void Do() {
        _action();
    }
}

public abstract class Benchmark : IEnumerable<TimeSpan> {
    public IEnumerator<TimeSpan> GetEnumerator() {
        Do(); // Warm-up!

        GC.Collect(); // Collect garbage.
        GC.WaitForPendingFinalizers(); // Wait until finalizers finish.

        var stopwatch = new Stopwatch();

        while (true) {
            stopwatch.Reset();
            stopwatch.Start();
            Do();
            stopwatch.Stop();

            yield return stopwatch.Elapsed;
        }
    }

    IEnumerator IEnumerable.GetEnumerator() {
        return GetEnumerator();
    }

    public abstract void Do();
}

public struct Vector4 {
    float x;
    float y;
    float z;
    float w;

    public Vector4(float x, float y, float z, float w) {
        this.x = x;
        this.y = y;
        this.z = z;
        this.w = w;
    }

    public static Vector4 operator +(Vector4 v1, Vector4 v2) {
        return new Vector4(v1.x + v2.x, v1.y + v2.y, v1.z + v2.z, v1.w + v2.w);
    }
}

class MainClass {
    public static void Main(string[] args) {
        var avgNS1 = new SimdBenchmark().Take(1000).Average(timeSpan => timeSpan.TotalNanoseconds());
        var avgNS2 = new SimpleBenchmark().Take(1000).Average(timeSpan => timeSpan.TotalNanoseconds());
        var avgNS3 = new DefaultBenchmark().Take(1000).Average(timeSpan => timeSpan.TotalNanoseconds());
        var avgNS4 = new MathNetBenchmark().Take(1000).Average(timeSpan => timeSpan.TotalNanoseconds());


        Console.WriteLine(avgNS1 + " ns");
        Console.WriteLine(avgNS2 + " ns");
        Console.WriteLine(avgNS3 + " ns");
        Console.WriteLine(avgNS4 + " ns");
    }
}
}

环境设置:

Windows 7/Mono 2.10.8/MonoDevelop 2.8.5

MonoDevelop 设置:

• 工具 > 选项 > .NET 运行时 > Mono 2.10.8(默认)
• 项目 > 选项 > 构建 > 常规 > 目标框架 > Mono/.NET4.0
• 项目 > 选项 > 构建 > 编译器 > 常规选项 > 启用优化
• 项目 > 选项 > 构建 > 编译器 > 常规选项 > 平台目标 > x86
• 项目 > 选项 > 运行 > 常规 > 参数 > -O=simd

结果:

• 94.4 纳秒
• 29.7 纳秒
• 49.9 纳秒
• 231595.2 纳秒

Answer 1

我会首先怀疑您的基准基础设施。

可能有几点:

• 您正在使用 `Stopwatch` 为单个操作计时 - 它没有分辨率
• 您的计时包括虚函数调用
• 您的样本量 ( 1000 ) 太小