2

std::inner_product()与手动点积计算相比，我很感兴趣，所以我做了一个测试。

std::inner_product()比手动实现快 4 倍。我觉得这很奇怪，因为确实没有那么多方法可以计算它，当然？！我也看不到在计算时正在使用任何 SSE/AVX 寄存器。

设置：VS2013/MSVC(12?)，Haswell i7 4770 CPU，64位编译，发布模式。

这是 C++ 测试代码：

#include <iostream>
#include <functional>
#include <numeric>
#include <cstdint>

int main() {
   const int arraySize = 1000;
   const int numTests = 500;
   unsigned int x, y = 0;
   unsigned long long* array1 = new unsigned long long[arraySize];
   unsigned long long* array2 = new unsigned long long[arraySize];

   //Initialise arrays
   for (int i = 0; i < arraySize; i++){
      unsigned long long val = __rdtsc();
      array1[i] = val;
      array2[i] = val;
   }

   //std::inner_product test
   unsigned long long timingBegin1 = __rdtscp(&s);
   for (int i = 0; i < numTests; i++){
      volatile unsigned long long result = std::inner_product(array1, array1 + arraySize, array2, static_cast<uint64_t>(0));
   }
   unsigned long long timingEnd1 = __rdtscp(&s);

   f, s = 0;

   //Manual Dot Product test
   unsigned long long timingBegin2 = __rdtscp(&f);
   for (int i = 0; i < numTests; i++){
      volatile unsigned long long result = 0;

      for (int i = 0; i < arraySize; i++){
         result += (array1[i] * array2[i]);
      }
   }
   unsigned long long timeEnd2 = __rdtscp(&f);


   std::cout << "STL:     :  " << static_cast<double>(finish1 - start1) / numTests << " CPU cycles per dot product" << std::endl;
   std::cout << "Manually :  " << static_cast<double>(finish2 - start2) / numTests << " CPU cycles per dot product" << std::endl;
4

1 回答 1

4

你的测试很糟糕，这可能会产生很大的不同。

volatile uint64_t result = 0;

for (int i = 0; i < arraySize; i++){
   result += (array1[i] * array2[i]);

请注意您如何在volatile此处不断使用 -qualified 变量。这会强制编译器将临时结果写入内存。

相反，您的inner_product版本：

volatile uint64_t result = std::inner_product(array1, array1 + arraySize, array2, static_cast<uint64_t>(0));

首先计算内积，允许优化，然后才将结果分配给volatile-qualfied 变量。

于 2014-08-08T10:47:57.743 回答