简单SSE循环比非SSE版本慢

Question

我正在尝试将上交所float4添加与标准float4添加进行比较。作为演示，我在使用和不使用SSE的情况下计算相加分量的总和：

#include <iostream>
#include <vector>

struct Point4
{
  Point4()
  {
    data[0] = 0;
    data[1] = 0;
    data[2] = 0;
    data[3] = 0;
  }

  float data[4];
};

void Standard()
{
  Point4 a;
  a.data[0] = 1.0f;
  a.data[1] = 2.0f;
  a.data[2] = 3.0f;
  a.data[3] = 4.0f;

  Point4 b;
  b.data[0] = 1.0f;
  b.data[1] = 6.0f;
  b.data[2] = 3.0f;
  b.data[3] = 5.0f;

  float total = 0.0f;
  for(unsigned int i = 0; i < 1e9; ++i)
  {
    for(unsigned int component = 0; component < 4; ++component)
    {
      total += a.data[component] + b.data[component];
    }
  }

  std::cout << "total: " << total << std::endl;
}

void Vectorized()
{
  typedef float v4sf __attribute__ (( vector_size(4*sizeof(float)) ));

  v4sf a;
  float* aPointer = (float*)&a;
  aPointer[0] = 1.0f; aPointer[1] = 2.0f; aPointer[2] = 3.0f; aPointer[3] = 4.0f;

  v4sf b;
  float* bPointer = (float*)&b;
  bPointer[0] = 1.0f; bPointer[1] = 6.0f; bPointer[2] = 3.0f; bPointer[3] = 5.0f;

  v4sf result;
  float* resultPointer = (float*)&result;
  resultPointer[0] = 0.0f;
  resultPointer[1] = 0.0f;
  resultPointer[2] = 0.0f;
  resultPointer[3] = 0.0f;

  for(unsigned int i = 0; i < 1e9; ++i)
  {
    result += a + b; // Vectorized operation
  }

  // Sum the components of the result (this is done with the "total += " in the Standard() loop
  float total = 0.0f;
  for(unsigned int component = 0; component < 4; ++component)
  {
    total += resultPointer[component];
  }
  std::cout << "total: " << total << std::endl;
}

int main()
{

//  Standard();

  Vectorized();

  return 0;
}

然而，使用标准方法的代码似乎比矢量化方法快(~.2秒) (~.4秒)。是因为for循环对v4sf值求和吗？有没有更好的操作，我可以用来计算这两种技术之间的差异，并仍然比较输出，以确保这两种技术之间没有差异？

Answer 1

那么你的版本比SSE慢的原因是你必须在每次迭代中从SSE寄存器解包到标量寄存器4次，这比你从矢量化加法中获得的开销更大。看一下反汇编，你应该会得到一个更清晰的图像。

我认为你想做的事情如下(使用SSE会更快)：

for(unsigned int i = 0; i < 1e6; ++i)
{
    result += a + b; // Vectorized operation
}

// Sum the components of the result (this is done with the "total += " in the Standard() loop
for(unsigned int component = 0; component < 4; ++component)
{
    total += resultPointer[component];
}

另外，下面的代码可能会更快：

for(unsigned int i = 0; i < 1e6/4; ++i)
{
    result0 += a + b; // Vectorized operation
    result1 += a + b; // Vectorized operation
    result2 += a + b; // Vectorized operation
    result3 += a + b; // Vectorized operation
}

// Sum the components of the result (this is done with the "total += " in the Standard() loop
for(unsigned int component = 0; component < 4; ++component)
{
    total += resultPointer0[component];
    total += resultPointer1[component];
    total += resultPointer2[component];
    total += resultPointer3[component];
}