/*
compile with:
gcc -c -Wall -O2 -march=skylake-avx512 sum64avx512.c
g++ -s -Wall -O2 -march=skylake-avx512 aligned_add_main.o sum64avx512.o -o tst_64b
*/
#include <stdint.h>
#include <string.h>
#include <intrin.h>

uint64_t sum64(char* buf, int incr)
{
  __m512i acc0=_mm512_setzero_si512();
  __m512i acc1=_mm512_setzero_si512();
  __m512i acc2=_mm512_setzero_si512();
  __m512i acc3=_mm512_setzero_si512();
  for (int i = 0; i < 4000; ++i) {
    char* p = buf;
    for (int k = 0; k < 125; ++k) {
      __m512i x0 = _mm512_loadu_si512((__m512i*)&p[0*sizeof(__m512i)]);
      __m512i x1 = _mm512_loadu_si512((__m512i*)&p[1*sizeof(__m512i)]);
      __m512i x2 = _mm512_loadu_si512((__m512i*)&p[2*sizeof(__m512i)]);
      __m512i x3 = _mm512_loadu_si512((__m512i*)&p[3*sizeof(__m512i)]);
      acc0 = _mm512_add_epi64(acc0, x0);
      acc1 = _mm512_add_epi64(acc1, x1);
      acc2 = _mm512_add_epi64(acc2, x2);
      acc3 = _mm512_add_epi64(acc3, x3);
      p += sizeof(__m512i)*4;
    }
    buf += incr;
  }
  acc0 = _mm512_add_epi64(acc0, acc1);
  acc2 = _mm512_add_epi64(acc2, acc3);
  acc0 = _mm512_add_epi64(acc0, acc2);
  __m256i acc = _mm256_add_epi64(_mm512_castsi512_si256(acc0), _mm512_extracti64x4_epi64(acc0, 1));
  return
    _mm256_extract_epi64(acc, 0) +
    _mm256_extract_epi64(acc, 1) +
    _mm256_extract_epi64(acc, 2) +
    _mm256_extract_epi64(acc, 3) ;
}