Sum 4 integer from a 128 bit __m128 Intel Intrinsic

Question

I have a __m128 intrinsic element of 128 bits. It contains 32 bit integers. Is there an easy way to sum all four of these integers? I am concerned with speed and cache optimization, so I'm trying to avoid storing the 128 in an array and then grabbing the elements from the array.

Thanks

Answer 1

A pair of phaddd instructions might be the best solution. Here is an example:

#define __USE_MINGW_ANSI_STDIO 1
#include <stdio.h>
#include <intrin.h>

//---------------------------------------------------------------------------

__m128i add_32X4 (__m128i value)
    {
    __m128i result;

    result = _mm_hadd_epi32 (value, value);
    result = _mm_hadd_epi32 (result, result);
    return result;
    }

//---------------------------------------------------------------------------

int main (void)
    {
    __m128i input = _mm_set_epi32 (1, 10, 100, 1000);
    __m128i result = add_32X4 (input);
    printf ("%d\n", _mm_cvtsi128_si32 (result));
    return 0;
    }

//---------------------------------------------------------------------------

program output: 1111

sample code generation (gcc 4.8.1):

<add_32X4>:
 30:    66 0f 6f 01             movdqa xmm0,XMMWORD PTR [rcx]
 34:    66 0f 38 02 c0          phaddd xmm0,xmm0
 39:    66 0f 38 02 c0          phaddd xmm0,xmm0
 3e:    c3                      ret