I am trying to optimize some multimedia algorithm on intel atom. I tried to use intrinsics for multiplication but it did not give any gain. I have put the sample code below.
#include
#include
int main()
{
static const float F[4] = { 0.12499890589348f,
0.11994905818636f,
0.10575346285725f,
0.08503254418207f};
float array[4], result=0.0f;
int j,m;
array[0] = F[0]+F[1];
array[1] = F[1]+F[2];
array[2] = F[2]+F[3];
array[3] = F[3]+F[0];
for(m=0; m<100; m++)
for(j=0;j<100000;j++)
{
#if 0
/* C implementation - takes 0.120secs */
result += (F[0]*array[0] + F[1]*array[1] + F[2]*array[2] + F[3]*array[3]);
#else
__m128 reg0, reg1;
float tmp[4];
reg0 = _mm_loadu_ps(F);
reg1 = _mm_loadu_ps(array);
reg0 = _mm_mul_ps(reg0, reg1);
/* using scalar addition - takes 0.124secs*/
_mm_storeu_ps(tmp, reg0);
result += (tmp[0]+tmp[1]+tmp[2]+tmp[3]);
}
}
I converted the filtering into SIMD intrinsics but its not giving any gain. Is there any particular reason for not getting gain here??. I tried writing hand assembly also which was of no benefit !!
