avxmemcpy.c

#include <string.h>
#include <x86intrin.h>
#include <stddef.h>
#include <stdint.h>
#include <stdio.h>

/* on Zen better use ssememcpy: its about as fast for most cases, and
   for block sizes 5..63 where avxmemcpy uses _mm256_maskload_ps/
   _mm256_maskstore_ps is slow on Zen */

/* the calls to memcpy() here implement unaligned accesses, and are
   compiled by gcc to mov instructions (even if you call this function
   memcpy, too) */
void *avxmemcpy(void *dest, const void *src, size_t n)
{
  static unsigned mask1[] = {~0,~0,~0,~0,~0,~0,~0,~0,
                              0, 0, 0, 0, 0, 0, 0, 0};
  if (n<5) {
    if (n<2) {
      if (n==1) {
        *(char *)dest = *(char *)src;
      }
    } else { /* n in 2..4 */
      short temp1, temp2;
      memcpy(&temp1,src,2);
      memcpy(&temp2,src+(n-2),2);
      memcpy(dest,&temp1,2);
      memcpy(dest+(n-2),&temp2,2);
    }
  } else {
    if (n<64) { /* n in 5..63 */
      size_t masklen = (n>>1)&~3; /* actual mask len is 4 bytes longer, 
                                     that's why this case does not cover n=64 */
      __m256i mask = (__m256i)_mm256_loadu_ps((float const*)(((char *)mask1)+28-masklen));
      __m256 p1 = _mm256_maskload_ps((float const *)src, mask);
      __m256 p2 = _mm256_maskload_ps((float const *)(src+(n-masklen-4)), mask);
      _mm256_maskstore_ps((float *)dest, mask, p1);
      _mm256_maskstore_ps((float *)(dest+(n-masklen-4)), mask, p2);
    } else { /* n>=64 */
      __m256i x1 = _mm256_loadu_si256((__m256i *)src);
      ptrdiff_t off = src-dest;
      void *dlast = dest+n-32;
      void *d = (void *)(((intptr_t)(dest+32))&~31);
      __m256i x3 = _mm256_loadu_si256((__m256i *)(dlast+off));
#ifdef NO_UNROLLING
      for (; d<dlast; d+=32) {
        __m256i x = _mm256_loadu_si256((__m256i *)(d+off));
        _mm256_storeu_si256((__m256i *)d, x);
      }
#else
      __m256i x2 = _mm256_loadu_si256((__m256i *)(dlast+off-32));
      for (; d<dlast-32; d+=64) {
        __m256i x1 = _mm256_loadu_si256((__m256i *)(d+off));
        __m256i x2 = _mm256_loadu_si256((__m256i *)(d+off+32));
        _mm256_storeu_si256((__m256i *)d, x1);
        _mm256_storeu_si256((__m256i *)(d+32), x2);
      }
      _mm256_storeu_si256((__m256i *)(dlast-32), x2);
#endif
      _mm256_storeu_si256((__m256i *)dest, x1);
      _mm256_storeu_si256((__m256i *)dlast, x3);
    }
  }
  return dest;
}