CODEWITHSUNDEEP
c++gccsse
David Doria
David Doria

Reputation: 10273

Simple SSE loop slower than non-SSE version

I am trying to compare SSE float[4] addition to standard float[4] addition. As a demo I compute the sum of the summed components, with and without SSE:

#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;
}

However, the code seems to be faster (~.2 seconds) with the standard method than with the vectorized (~.4 seconds) method. Is it because of the for loop to sum the v4sf values? Is there a better operation I can use to time the difference between these two techniques and still compare the output to make sure there were no differences between the two?

Upvotes: 1

Views: 418

Answers (1)

ronag
ronag

Reputation: 51283

Then reason your version is slower as SSE is that you have to unpack from an SSE register to a scalar register 4 times every iteration, which has more of an overhead than what you gain from the vectorized addition. Look at the disassembly and you should get a clearer picture.

I think what you want to do is the following (which is faster with SSE):

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];
}

Also the following might be even faster:

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];
}

Upvotes: 1

Related Questions