没有定义宏的代码重复问题

我正在尝试使用矢量化函数来模仿CUDA/OpenCL工作流,如下所示:

#include <omp.h>
#include <iostream>
#include <string>
#include <functional>
#include<cmath>
template<typename Type, int Simd>
struct KernelData
{
    alignas(32)
    Type data[Simd];

    inline void readFrom(const Type * const __restrict__ ptr) noexcept
    {
        for(int i=0;i<Simd;i++)
        {
            data[i] = ptr[i];
        }
    }

    inline void writeTo(Type * const __restrict__ ptr) const noexcept
    {
        for(int i=0;i<Simd;i++)
        {
            ptr[i] = data[i];
        }
    }

    inline const KernelData<Type,Simd> sqrt() const noexcept
    {
        KernelData<Type,Simd> result;
        for(int i=0;i<Simd;i++)
        {
            result.data[i] = std::sqrt(data[i]);
        }
        return result;
    }
};




template<int mask>
struct KernelDataFactory
{
    KernelDataFactory()
    {

    }


    template<typename Type>
    inline
    KernelData<Type,mask> generate() const
    {
        return KernelData<Type,mask>();
    }
};

template<int SimdWidth, typename... Args>
class Kernel
{
public:
    Kernel(std::function<void(int,int, Args...)> kernelPrm)
    {
        kernel = kernelPrm;
    }


    void run(int n, Args... args)
    {
        const int nLoop = (n/SimdWidth);
        for(int i=0;i<nLoop;i++)
        {
            kernel(i*SimdWidth,SimdWidth, args...);
        }

        if((n/SimdWidth)*SimdWidth != n)
        {
            const int m = n%SimdWidth;
            for(int i=0;i<m;i++)
            {
                kernel(nLoop*SimdWidth+i,1, args...);
            }
        }
    }
private:
    std::function<void(int,int, Args...)> kernel;
};

// cpu cycles from stackoverflow
#include <stdint.h>  // <cstdint> is preferred in C++, but stdint.h works.
#ifdef _MSC_VER
# include <intrin.h>
#else
# include <x86intrin.h>
#endif
inline
uint64_t readTSC() {
    // _mm_lfence();  // optionally wait for earlier insns to retire before reading the clock
    uint64_t tsc = __rdtsc();
    // _mm_lfence();  // optionally block later instructions until rdtsc retires
    return tsc;
}

int main(int argC, char** argV)
{ 
    constexpr int simd = 16;
    constexpr int n = 1003;
    Kernel<simd, float *, float *> kernel([](int simdGroupId, int simdWidth, float * input, float * output){
        const int id = simdGroupId;
        if(simdWidth == simd)
        {
            const KernelDataFactory<simd> factory;
            auto a = factory.generate<float>();

            a.readFrom(input+id);
            const auto b = a.sqrt().sqrt().sqrt().sqrt().sqrt().
            sqrt().sqrt().sqrt().sqrt().sqrt().
            sqrt().sqrt().sqrt().sqrt().sqrt();
            b.writeTo(output+id);
        }
        else
        {
            const KernelDataFactory<1> factory;
            auto a = factory.generate<float>();

            a.readFrom(input+id);
            const auto b = a.sqrt().sqrt().sqrt().sqrt().sqrt().
            sqrt().sqrt().sqrt().sqrt().sqrt().
            sqrt().sqrt().sqrt().sqrt().sqrt();
            b.writeTo(output+id);
        }
    });
  
    alignas(32)
    float i[n],o[n];
    for(int j=0;j<n;j++)
        i[j]=j;
    auto t1 = readTSC();
    for(int k=0;k<10000;k++)
        kernel.run(n,i,o);
    auto t2 = readTSC();
    for(int i=n-10;i<n;i++)
    {

        std::cout<<"i="<<i<<" value="<<o[i]<<std::endl;
    }
    std::cout<<0.0001f*(t2-t1)/(float)(15*n)<<" cycles per sqrt"<<std::endl;
    return 0;
}

但是用户给出的部分必须像这样复制:

Kernel<simd, float *, float *> kernel([](int simdGroupId, int simdWidth, float * input, float * output){
    const int id = simdGroupId;
    if(simdWidth == simd)
    {
        const KernelDataFactory<simd> factory;
        auto a = factory.generate<float>();

        a.readFrom(input+id);
        const auto b = a.sqrt().sqrt().sqrt().sqrt().sqrt().
        sqrt().sqrt().sqrt().sqrt().sqrt().
        sqrt().sqrt().sqrt().sqrt().sqrt();
        b.writeTo(output+id);
    }
    else
    {
        const KernelDataFactory<1> factory;
        auto a = factory.generate<float>();

        a.readFrom(input+id);
        const auto b = a.sqrt().sqrt().sqrt().sqrt().sqrt().
        sqrt().sqrt().sqrt().sqrt().sqrt().
        sqrt().sqrt().sqrt().sqrt().sqrt();
        b.writeTo(output+id);
    }
});

唯一的区别是要生成的两个已知模板的编译时间:

KernelDataFactory<1> and KernelDataFactory<simd>

使用define宏,很容易只复制lambda的函数体。我尝试在不使用任何define宏的情况下完成此操作。有没有一种简单的方法可以让用户只给出以下内容:

auto a = factory.generate<float>();
a.readFrom(input+id);
const auto b = a.sqrt().sqrt().sqrt().sqrt().sqrt().
sqrt().sqrt().sqrt().sqrt().sqrt().
sqrt().sqrt().sqrt().sqrt().sqrt();
b.writeTo(output+id);

并且它被实现自动复制?

目前的实施是:

• 取n
• 把它分成两部分
• 运行矢量化代码,直到 n - (n%simd) 达到点
• 从运行标量代码 n-(n%simd) 到 n

KernelDataFactory模板参数(必须在编译时已知)(1和simd)用于让编译器生成矢量化代码。(在godbolt.org(avx512)上,它以“每平方英尺0.9个周期”的速度运行,在我的系统(avx1)上,速度为每平方英尺3.8个周期。)