Integer dot product using SSE/AVX?

问题:

I am looking at the intel intrinsic guide:

https://software.intel.com/sites/landingpage/IntrinsicsGuide/

and whilst they have _mm_dp_ps and _mm_dp_pd for calculating the dot product for floats and doubles I cannot see anything for calculating the integer dot product.

I have two unsigned int[8] arrays and I would like to:

(a[0] x b[0]) + (a[1] * b[1])....... + (a[num_elements_in_array-1] * b[num_elements_in_array-1])

(in batches of four) and sum the products?

回答1:

Every time someone does this:

temp_1 = _mm_set_epi32(x[j], x[j+1], x[j+2], x[j+3]);

.. a puppy dies.

Use one of these:

temp_1 = _mm_load_si128(x);  // if aligned
temp_1 = _mm_loadu_si128(x); // if not aligned

Cast x as necessary.

There is no integer version of _mm_dp_ps. But you can do what you were about to do: multiply 4 by 4 integers, accumulate the sum of the products.

So something like this (not tested, doesn't compile)

while(j < num_elements_in_array){
    //Load the 4 values from x
    temp_1 = _mm_load_si128(x + j); // add cast
    //Load the 4 values from y
    temp_2 = _mm_load_si128(y + j); // add cast
    j += 4;
    //Multiply x[0] and y[0], x[1] and y[1] etc
    temp_products = _mm_mullo_epi32(temp_1, temp_2);
    //Sum temp_sum
    temp_sum = _mm_add_epi32(temp_sum, temp_products);
}
// take horizontal sum of temp_sum
temp_sum = _mm_add_epi32(temp_sum, _mm_srli_si128(temp_sum, 8));
temp_sum= _mm_add_epi32(temp_sum, _mm_srli_si128(temp_sum, 4));
sum = _mm_cvtsi128_si32(temp_sum);

As discussed in the comments and chat, that reorders the sums in such a way as to minimize the number of horizontal sums required, by doing most sums vertically.