PMADDWD

#include <smmintrin.h>
#include <stdio.h>
#include <stdint.h>



long dp32(const int32_t * const pA, const int32_t * const pB, long SIZE){
    long valor = 0, i=0;

    __m128i vsum, vecPi, vecCi, vecQCi;

    vsum = _mm_set1_epi32(0);

    int sumDot[1];

    for( ; i<SIZE-3 ;i+=4){
            vecPi = _mm_loadu_si128((__m128i *)&(pA)[i] );
            vecCi = _mm_loadu_si128((__m128i *)&(pB)[i] );
            vecQCi = _mm_mullo_epi32(vecPi,vecCi);
            vsum = _mm_add_epi32(vsum,vecQCi);
    }

    vsum = _mm_hadd_epi32(vsum, vsum);
    vsum = _mm_hadd_epi32(vsum, vsum);
    sumDot[0] = (int32_t)_mm_extract_epi32(vsum, 0);

    for( ; i<SIZE; i++)
          valor += pA[i] * pB[i];

    return valor += sumDot[0];
}

long dp16(const int16_t * const pA, const int16_t * const pB, long SIZE){
    long valor = 0, i=0;

    __m128i vsum, vecPi, vecCi, vecQCi;

    vsum = _mm_set1_epi32(0);

    int sumDot[1];

    for( ; i<SIZE-7 ;i+=8){
            vecPi  = _mm_loadu_si128((__m128i *)&(pA)[i] );
            vecCi  = _mm_loadu_si128((__m128i *)&(pB)[i] );
            vecQCi = _mm_madd_epi16(vecPi,vecCi);
            vsum   = _mm_add_epi32(vsum, vecQCi);
    }

    vsum = _mm_hadd_epi32(vsum, vsum);
    vsum = _mm_hadd_epi32(vsum, vsum);
    sumDot[0] = (int32_t)_mm_extract_epi32(vsum, 0);

    for( ; i<SIZE; i++)
          valor += pA[i] * pB[i];

    return valor += sumDot[0];
}

//#define SIZE 9
//#define VECA  0, -1,  2,  -3,  4, -5,   6, -7,  8
//#define VECB -9, 10, 11, -12, 13, 14, -15, 16, 17
#define SIZE 40
#define VECA 169, -341, 487, -50, 116, -66, 170, -217, -236, 72, -296, -469, -178, -1, 61, -26, -107, 593, -558, 125, 111, 111, 184, 141, -228, -490, -357, 535, 206, -85, -29, -216, -490, -265, 12, 250, 306, -291, -46, 180
#define VECB 398, 305, 268, -125, -96, -214, -284, 108, -37, 2, -402, -228, 243, 33, 76, 265, 3, -558, -323, -552, 419, 408, -217, -2, -440, -375, 153, -108, -79, -80, 299, 81, -385, 80, 53, 294, 170, -380, -164, -172
/*            0 -10 +22  +36 +52 -70  -90-112+136 */
/*            -10   +58      -18      -202   +136 */
/*            +48            -220            +136 */
/*            -172                           +136 */
/*            -36                                 */

int main(){
    int32_t arrA32[] = {VECA};
    int16_t arrA16[] = {VECA};
    int32_t arrB32[] = {VECB};
    int16_t arrB16[] = {VECB};

    long v32 = dp32(arrA32, arrB32, SIZE);
    long v16 = dp16(arrA16, arrB16, SIZE);

    if(v32 == v16){
        printf("Success! (32: %ld == %ld : 16)\n", v32, v16);
    }else{
        printf("Failure! (32: %ld != %ld : 16)\n", v32, v16);
    }
    return 0;
}