SSE floating point dot product for dummies

Asked Oct 05 '12 at 03:33

Active Oct 08 '12 at 01:58

Viewed 2,056 times

I have read many SO questions about SSE/SIMD (e.g., Getting started with SSE), but I'm still confused by all of it. All I want is a dot product between two double precision floating-point vectors, in C (C99 FWIW). I'm using GCC.

Can someone post a simple and complete example, including how to convert double vectors to the SSE types and back again?

[Edit 2012-10-08]

Here's some SSE2 code I managed to cobble together, critiques?

#include <emmintrin.h>

double dotprod(double *restrict a, double *restrict b, int n)
{
   __m128d aa, bb, cc, ss;
   int i, n1 = n - 1;
   double *s = calloc(2, sizeof(double));
   double s2 = 0;

   ss = _mm_set1_pd(0);

   for(i = 0 ; i < n1 ; i += 2) 
   {
     aa = _mm_load_pd(a + i);
     bb = _mm_load_pd(b + i);
     cc = _mm_mul_pd(aa, bb);
     ss = _mm_add_pd(ss, cc);
   }

   _mm_store_pd(s, ss);
   s2 = s[0] + s[1];

   if(i < n)
      s2 += a[i] * b[i];

   free(s);

   return s2;
}

edited May 23 '17 at 12:13

Community

asked Oct 05 '12 at 03:33

purple51

3

Please specify which versions of SSE you're interested in (e.g. there is even a dot product primitive in most recent). – ndkrempel Oct 05 '12 at 03:34
Both SSE2 and SSE4 would be great. – purple51 Oct 05 '12 at 03:52
1

That's a nasty horizontal thing to want.. it can be done of course (trivially even with `DPPD`), but it could point towards a larger problem with the organization of your data. – harold Oct 05 '12 at 09:29
Well, I'd be grateful for any example code to experiment with and compare. – purple51 Oct 06 '12 at 02:08
@arunmoezhi: note that `_mm_dp_pd` has very high latency (typically 9 cycles). – Paul R Jan 09 '17 at 09:32

SSE floating point dot product for dummies

0 Answers0