How to vectorize this kernel? - c

Question

I made this question in the middle of another one, and it seems that nobody will answer it in the previous topic.
My question is the following. I have been vectorizing one application with success. However, for this particular kernel:
inline int cFunction(Node** A, Node* B){
long dot = 0, i;
for(i=0;i<SIZE;i++)
dot += (*A)->data[i] * B->data[i];
if(abs(2 * dot) <= B->norm)
return 0;
long q = round((double) dot / B->norm);
for(i=0;i<SIZE;i++)
(*A)->data[i] -= q * B->data[i];
(*A)->norm = (*A)->norm + q * q * B->norm - 2 * q * dot;
return 1;
}
I was only able to vectorize the first loop. If I put icpc to run this code, with:
icpc *.c *.h -g -O2 -msse4.2 -vec-report=1
I have the following report:
main.c(558): (col. 9) remark: LOOP WAS VECTORIZED.
main.c(566): (col. 2) remark: LOOP WAS VECTORIZED.
Which tells me that the icpc actually vectorizes the code. Now, if I hand vectorize the first loop I have a good (perfect?) speedup factor for ints. This tells me that the compiler is not doing a good job at vectorizing this (especially because if I use short s the performance is the same). However, for the second loop, I get no whatsoever performance gains if I hand vectorize it like this:
const int q = round((double) dot / B->norm) ;
int32_t * pA = (*A)->data;
int32_t * const pB = B->data;
const __m128i vecQi = _mm_set1_epi32(q);
__m128i vecResi, vecPi, vecCi, vecQCi;
for(i=0;i<SIZE-3;i+=4){
vecPi = _mm_load_si128((__m128i *)&(pA)[i] );
vecCi = _mm_load_si128((__m128i *)&(pB)[i] );
vecQCi = _mm_mullo_epi32(vecQi,vecCi);
vecResi = _mm_sub_epi32(vecPi,vecQCi);
_mm_store_si128((__m128i *) ((pA) + i), vecResi );
}
for(;i<SIZE;i++)
pA[i] -= q * pB[i];
(*A)->norm = (*A)->norm + q * q * B->norm - 2 * q * dot;
Does anyone have a clue of why I am not getting performance gains by vectorizing the second kernel?
Thanks.