For some reason running the simple serial code
for(i=0;i<1152*1152;i++){
MatrixA3[i] = MatrixA1[i] + z*MatrixA2[i];}
runs faster than or same speed with the vectorized equivalent;
for (int i = 0; i < 1152*1152; i+=4){
load_data1 = _mm256_load_pd(MatrixA1 + i);
load_data2 = _mm256_load_pd(MatrixA2 + i);
_mm256_store_pd(MatrixA3 + i, _mm256_fmadd_pd(load_z,
load_data2,load_data1_dp));
}
On my intel i7-4578U with Intel compiler XE 15.0, the former runs in 1.507millesecs while the later finished in 1.513millisecs with 10000runs.
My experience has been a significant acceleration with avx2 intrinsics but for some reason this line decides to fail me. What am I doing wrong please?
What are you doing wrong? Not trusting your compiler.
This is not a case for manual optimization, any respectable compiler could vectorize that.