eddys_snippets/avx/dotproduct.c

/* Compute the dot product of two (properly aligned) vectors. */
#include <stdio.h>
#include <immintrin.h>
#include <math.h>

const int N = 83;

double slow_dot_product(const double *a, const double *b) {
    double answer = 0.0;
    for(int ii = 0; ii < N; ++ii)
        answer += a[ii]*b[ii];
    return answer;
}

/* Horizontal add works within 128-bit lanes. Use scalar ops to add
 * across the boundary. */
double reduce_vector1(__m256d input) {
    __m256d temp = _mm256_hadd_pd(input, input);
    return ((double*)&temp)[0] + ((double*)&temp)[2];
}

/* Another way to get around the 128-bit boundary: grab the first 128
 * bits, grab the lower 128 bits and then add them together with a 128
 * bit add instruction. */
double reduce_vector2(__m256d input) {
    __m256d temp = _mm256_hadd_pd(input, input);
    __m128d sum_high = _mm256_extractf128_pd(temp, 1);
    __m128d result = _mm_add_pd(sum_high, _mm256_castpd256_pd128(temp));
    return ((double*)&result)[0];
}

double dot_product(const double *a, const double *b) {
    __m256d sum_vec = _mm256_set_pd(0.0, 0.0, 0.0, 0.0);

    /* Add up partial dot-products in blocks of 256 bits */
    for(int ii = 0; ii < N/4; ++ii) {
        __m256d x = _mm256_load_pd(a+4*ii);
        __m256d y = _mm256_load_pd(b+4*ii);
        __m256d z = _mm256_mul_pd(x,y);
        sum_vec = _mm256_add_pd(sum_vec, z);
    }

    /* Find the partial dot-product for the remaining elements after
    * dealing with all 256-bit blocks. */
    double final = 0.0;
    for(int ii = N-N%4; ii < N; ++ii)
        final += a[ii] * b[ii];

    return reduce_vector2(sum_vec) + final;
}

int main() {
    __attribute__ ((aligned (32))) double a[N], b[N];

    for(int ii = 0; ii < N; ++ii)
        a[ii] = b[ii] = ii/sqrt(N);

    double answer = dot_product(a, b);
    printf("%f\n", answer);
    printf("%f\n", slow_dot_product(a,b));
}