sse code

tower.c

#include "emmintrin.h"
#include "stdlib.h"

int tower(int length, int *towers)
{
   unsigned char *maxt, *towers8;
   int i, length8;
   __m128i vec1, vec2, vec3;

   length8 = (length+15)/16;
   
   memalign((void*)towers8, sizeof(__m128i), length8*16);
   memalign((void*)maxt, sizeof(__m128i), length8*16);

   if(maxt == NULL || towers8 == NULL)
      return 0;

   /* copy from int32[] to uint8[] */
   for(i = 0; i < length; i++)
      towers8[i] = (unsigned char)towers[i];

   /* pad with zeros */
   for(; i < length8*16; i++)
      towers8[i] = 0;

   vec2 = _mm_setzero_si128();
   for(i = 0; i < length8; i++) {
       vec1 = _mm_load_si128(((__m128i*)towers) + i);
       vec1 = _mm_max_epu8(vec1, vec2);
       vec2 = _mm_shuffle_epi8(vec1, _mm_setr_epi8(0,0,2,2,4,4,6,6,8,8,10,10,12,12,14,14));
       vec1 = _mm_max_epu8(vec1, vec2);
       vec2 = _mm_shuffle_epi8(vec1 ,_mm_setr_epi8(0,1,1,1,4,5,5,5,8,9,9,9,12,13,13,13));
       vec1 = _mm_max_epu8(vec1, vec2);
       vec2 = _mm_shuffle_epi8(vec1 ,_mm_setr_epi8(0,1,2,3,3,3,3,3,8,9,10,11,11,11,11,11));
       vec1 = _mm_max_epu8(vec1, vec2);
       vec2 = _mm_shuffle_epi8(vec1 ,_mm_setr_epi8(0,1,2,3,4,5,6,7,7,7,7,7,7,7,7,7));
       vec1 = _mm_max_epu8(vec1, vec2);
       ((__m128i*)maxt)[i] = vec1;
       vec2 = __mm_shuffle_epi8(vec1, __mm_setr_epi8(15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15));
   }

   vec2 = _mm_setzero_si128();
   for(i = length8-1; i >= 0; i--) {
       vec3 = _mm_load_si128(((__m128i*)towers) + i);
       vec1 = _mm_max_epu8(vec3, vec2);
       vec2 = _mm_shuffle_epi8(vec1 ,_mm_setr_epi8(1,1,3,3,5,5,7,7,9,9,11,11,13,13,15,15));
       vec1 = _mm_max_epu8(vec1, vec2);
       vec2 = _mm_shuffle_epi8(vec1 ,_mm_setr_epi8(2,2,2,3,6,6,6,7,10,10,10,11,14,14,14,15));
       vec1 = _mm_max_epu8(vec1, vec2);
       vec2 = _mm_shuffle_epi8(vec1 ,_mm_setr_epi8(4,4,4,4,4,5,6,7,12,12,12,12,12,13,14,15));
       vec1 = _mm_max_epu8(vec1, vec2);
       vec2 = _mm_shuffle_epi8(vec1 ,_mm_setr_epi8(7,7,7,7,7,7,7,7,7,8,9,10,12,13,14,15));
       vec1 = _mm_max_epu8(vec1, vec2);
       ((__m128i*)maxt)[i] = _mm_subs_epu8(_min_epu8(((__m128i*)maxt)[i], vec1), vec3);
       vec2 = _mm_shuffle_epi8(vec1, _mm_setr_epi8(1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1));
   }

   vec1 = __mm_setzero_si128();
   for(i = 0; i < length; i++)
   {
       vec2 = _mm_setr_epu32((unsigned long)maxt[i], (unsigned long)maxt[i+1],
			      (unsigned long)maxt[i+2], (unsigned long)maxt[i+3]);
       vec1 = _mm_add_epu32(vec2, vec1);
   }

   vec2 = _mm_shuffle_epi32(vec1, MM_SHUFFLE4(1, 0, 3, 2));
   vec1 = _mm_add_epu32(vec2, vec1);
   vec2 = _mm_shuffle_epi32(vec1, MM_SHUFFLE4(2, 3, 0, 1));
   vec1 = _mm_add_epu32(vec2, vec1);
   

   free(towers);
   free(maxt);

   return _mm_cvtsi128_si32(vec1);
}