避免 C# 虚拟调用的开销

Avoiding the overhead of C# virtual calls

我有几个高度优化的数学函数需要 1-2 nanoseconds 才能完成。这些函数每秒被调用数亿次,因此尽管性能已经非常出色,但调用开销仍是一个问题。

为了保持程序的可维护性,提供这些方法的类继承了一个IMathFunction接口,以便其他对象可以直接存储特定的数学函数,并在需要时使用它。

public interface IMathFunction
{
  double Calculate(double input);
  double Derivate(double input);
}

public SomeObject
{
  // Note: There are cases where this is mutable
  private readonly IMathFunction mathFunction_; 

  public double SomeWork(double input, double step)
  {
    var f = mathFunction_.Calculate(input);
    var dv = mathFunction_.Derivate(input);
    return f - (dv * step);
  }
}

由于使用代码的使用方式,与直接调用相比,此接口会产生巨大的开销。 直接调用需要 1-2ns,而虚拟 接口调用需要 8-9ns。显然,接口的存在及其对虚拟调用的后续翻译是这种情况的瓶颈。

如果可能,我希望同时保留可维护性和性能。 有没有一种方法可以在实例化对象时将虚函数解析为直接调用,以便所有后续调用都能够避免开销?我认为这将涉及使用 IL 创建委托, 但我不知道从哪里开始。

所以这有明显的局限性,不应该在任何有接口的地方一直使用,但如果你有一个地方确实需要最大化性能,你可以使用泛型:

public SomeObject<TMathFunction> where TMathFunction: struct, IMathFunction 
{
  private readonly TMathFunction mathFunction_;

  public double SomeWork(double input, double step)
  {
    var f = mathFunction_.Calculate(input);
    var dv = mathFunction_.Derivate(input);
    return f - (dv * step);
  }
}

并且不传递接口,而是将您的实现作为 TMathFunction 传递。这将避免由于接口而导致的 vtable 查找,并且还允许内联。

注意这里使用 struct 很重要,否则泛型将通过接口访问 class。

一些实现:

我做了一个简单的 IMathFunction 实现用于测试:

class SomeImplementationByRef : IMathFunction
{
    public double Calculate(double input)
    {
        return input + input;
    }

    public double Derivate(double input)
    {
        return input * input;
    }
}

...以及结构版本和抽象版本。

所以,这是界面版本的情况。您可以看到它相对低效,因为它执行两个级别的间接寻址:

    return obj.SomeWork(input, step);
sub         esp,40h  
vzeroupper  
vmovaps     xmmword ptr [rsp+30h],xmm6  
vmovaps     xmmword ptr [rsp+20h],xmm7  
mov         rsi,rcx
vmovsd      qword ptr [rsp+60h],xmm2  
vmovaps     xmm6,xmm1
mov         rcx,qword ptr [rsi+8]          ; load mathFunction_ into rcx.
vmovaps     xmm1,xmm6  
mov         r11,7FFED7980020h              ; load vtable address of the IMathFunction.Calculate function.
cmp         dword ptr [rcx],ecx  
call        qword ptr [r11]                ; call IMathFunction.Calculate function which will call the actual Calculate via vtable.
vmovaps     xmm7,xmm0
mov         rcx,qword ptr [rsi+8]          ; load mathFunction_ into rcx.
vmovaps     xmm1,xmm6  
mov         r11,7FFED7980028h              ; load vtable address of the IMathFunction.Derivate function.
cmp         dword ptr [rcx],ecx  
call        qword ptr [r11]                ; call IMathFunction.Derivate function which will call the actual Derivate via vtable.
vmulsd      xmm0,xmm0,mmword ptr [rsp+60h] ; dv * step
vsubsd      xmm7,xmm7,xmm0                 ; f - (dv * step)
vmovaps     xmm0,xmm7  
vmovaps     xmm6,xmmword ptr [rsp+30h]  
vmovaps     xmm7,xmmword ptr [rsp+20h]  
add         rsp,40h  
pop         rsi  
ret  

这是摘要class。它的效率更高一些,但可以忽略不计:

        return obj.SomeWork(input, step);
 sub         esp,40h  
 vzeroupper  
 vmovaps     xmmword ptr [rsp+30h],xmm6  
 vmovaps     xmmword ptr [rsp+20h],xmm7  
 mov         rsi,rcx  
 vmovsd      qword ptr [rsp+60h],xmm2  
 vmovaps     xmm6,xmm1  
 mov         rcx,qword ptr [rsi+8]           ; load mathFunction_ into rcx.
 vmovaps     xmm1,xmm6  
 mov         rax,qword ptr [rcx]             ; load object type data from mathFunction_.
 mov         rax,qword ptr [rax+40h]         ; load address of vtable into rax.
 call        qword ptr [rax+20h]             ; call Calculate via offset 0x20 of vtable.
 vmovaps     xmm7,xmm0  
 mov         rcx,qword ptr [rsi+8]           ; load mathFunction_ into rcx.
 vmovaps     xmm1,xmm6  
 mov         rax,qword ptr [rcx]             ; load object type data from mathFunction_.
 mov         rax,qword ptr [rax+40h]         ; load address of vtable into rax.
 call        qword ptr [rax+28h]             ; call Derivate via offset 0x28 of vtable.
 vmulsd      xmm0,xmm0,mmword ptr [rsp+60h]  ; dv * step
 vsubsd      xmm7,xmm7,xmm0                  ; f - (dv * step)
 vmovaps     xmm0,xmm7
 vmovaps     xmm6,xmmword ptr [rsp+30h]  
 vmovaps     xmm7,xmmword ptr [rsp+20h]  
 add         rsp,40h  
 pop         rsi  
 ret  

因此接口和抽象 class 都严重依赖分支目标预测来获得可接受的性能。即便如此,您仍然可以看到还有很多事情要做,所以最好的情况仍然相对较慢,而最坏的情况是由于预测错误导致管道停滞。

最后是带有结构的通用版本。您可以看到它的效率大大提高,因为所有内容都已完全内联,因此不涉及分支预测。它还具有删除其中大部分 stack/parameter 管理的不错的副作用,因此代码变得非常紧凑:

    return obj.SomeWork(input, step);
push        rax  
vzeroupper  
movsx       rax,byte ptr [rcx+8]  
vmovaps     xmm0,xmm1  
vaddsd      xmm0,xmm0,xmm1  ; Calculate - got inlined
vmulsd      xmm1,xmm1,xmm1  ; Derivate - got inlined
vmulsd      xmm1,xmm1,xmm2  ; dv * step
vsubsd      xmm0,xmm0,xmm1  ; f - 
add         rsp,8  
ret  

我会将方法分配给委托。这允许您仍然针对接口进行编程,同时避免接口方法解析。

public SomeObject
{
    private readonly Func<double, double> _calculate;
    private readonly Func<double, double> _derivate;

    public SomeObject(IMathFunction mathFunction)
    {
        _calculate = mathFunction.Calculate;
        _derivate = mathFunction.Derivate;
    }

    public double SomeWork(double input, double step)
    {
        var f = _calculate(input);
        var dv = _derivate(input);
        return f - (dv * step);
    }
}

为了回应@CoryNelson 的评论,我进行了测试以了解实际影响。我已经密封了函数 class,但这似乎完全没有区别,因为我的方法不是虚拟的。

大括号中减去空方法时间的测试结果(1 亿次迭代的平均时间,以 ns 为单位):

Empty Work method: 1.48
Interface: 5.69 (4.21)
Delegates: 5.78 (4.30)
Sealed Class: 2.10 (0.62)
Class: 2.12 (0.64)

委托版本时间与接口版本大致相同(具体时间因测试执行而异)。在 class 上工作大约快 6.8 倍(比较时间减去空工作方法时间)!这意味着我与代表合作的建议没有帮助!

令我惊讶的是,我预计界面版本的执行时间会更长。由于这种测试不代表 OP 代码的确切上下文,因此其有效性有限。

static class TimingInterfaceVsDelegateCalls
{
    const int N = 100_000_000;
    const double msToNs = 1e6 / N;

    static SquareFunctionSealed _mathFunctionClassSealed;
    static SquareFunction _mathFunctionClass;
    static IMathFunction _mathFunctionInterface;
    static Func<double, double> _calculate;
    static Func<double, double> _derivate;

    static TimingInterfaceVsDelegateCalls()
    {
        _mathFunctionClass = new SquareFunction();
        _mathFunctionClassSealed = new SquareFunctionSealed();
        _mathFunctionInterface = _mathFunctionClassSealed;
        _calculate = _mathFunctionInterface.Calculate;
        _derivate = _mathFunctionInterface.Derivate;
    }

    interface IMathFunction
    {
        double Calculate(double input);
        double Derivate(double input);
    }

    sealed class SquareFunctionSealed : IMathFunction
    {
        public double Calculate(double input)
        {
            return input * input;
        }

        public double Derivate(double input)
        {
            return 2 * input;
        }
    }

    class SquareFunction : IMathFunction
    {
        public double Calculate(double input)
        {
            return input * input;
        }

        public double Derivate(double input)
        {
            return 2 * input;
        }
    }

    public static void Test()
    {
        var stopWatch = new Stopwatch();

        stopWatch.Start();
        for (int i = 0; i < N; i++) {
            double result = SomeWorkEmpty(i);
        }
        stopWatch.Stop();
        double emptyTime = stopWatch.ElapsedMilliseconds * msToNs;
        Console.WriteLine($"Empty Work method: {emptyTime:n2}");

        stopWatch.Restart();
        for (int i = 0; i < N; i++) {
            double result = SomeWorkInterface(i);
        }
        stopWatch.Stop();
        PrintResult("Interface", stopWatch.ElapsedMilliseconds, emptyTime);

        stopWatch.Restart();
        for (int i = 0; i < N; i++) {
            double result = SomeWorkDelegate(i);
        }
        stopWatch.Stop();
        PrintResult("Delegates", stopWatch.ElapsedMilliseconds, emptyTime);

        stopWatch.Restart();
        for (int i = 0; i < N; i++) {
            double result = SomeWorkClassSealed(i);
        }
        stopWatch.Stop();
        PrintResult("Sealed Class", stopWatch.ElapsedMilliseconds, emptyTime);

        stopWatch.Restart();
        for (int i = 0; i < N; i++) {
            double result = SomeWorkClass(i);
        }
        stopWatch.Stop();
        PrintResult("Class", stopWatch.ElapsedMilliseconds, emptyTime);
    }

    private static void PrintResult(string text, long elapsed, double emptyTime)
    {
        Console.WriteLine($"{text}: {elapsed * msToNs:n2} ({elapsed * msToNs - emptyTime:n2})");
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    private static double SomeWorkEmpty(int i)
    {
        return 0.0;
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    private static double SomeWorkInterface(int i)
    {
        double f = _mathFunctionInterface.Calculate(i);
        double dv = _mathFunctionInterface.Derivate(i);
        return f - (dv * 12.34534);
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    private static double SomeWorkDelegate(int i)
    {
        double f = _calculate(i);
        double dv = _derivate(i);
        return f - (dv * 12.34534);
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    private static double SomeWorkClassSealed(int i)
    {
        double f = _mathFunctionClassSealed.Calculate(i);
        double dv = _mathFunctionClassSealed.Derivate(i);
        return f - (dv * 12.34534);
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    private static double SomeWorkClass(int i)
    {
        double f = _mathFunctionClass.Calculate(i);
        double dv = _mathFunctionClass.Derivate(i);
        return f - (dv * 12.34534);
    }
}

[MethodImpl(MethodImplOptions.NoInlining)]的思想是防止编译器在方法被内联时在循环之前计算方法的地址。