SIMD: implement _mm256_max_epu64_

[CakeByTheOcean]

Hi.
I want to ask a simple question about SIMD.
I don't get the AVX512 in my CPU but want to have a _mm256_max_epu64.
https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm256_max_epu64&expand=730,4929,3837,624,776,3705,776,3705,1011,931,3705,3616,3582,3591

How can we implement one?

[CakeByTheOcean]

inline __m256i __my_mm256_min_epu64_(__m256i a, __m256i b) {
  uint64_t *val_a = (uint64_t*) &a;
  uint64_t *val_b = (uint64_t*) &b;
  uint64_t e[4];
  for (size_t i = 0; i < 4; ++i) e[i] = (*(val_a + i) < *(val_b + i)) ? *(val_a + i) : *(val_b + i);
  return _mm256_set_epi64x(e[3], e[2], e[1], e[0]);
}

I am the OP and had this trivial one.
Do you have idea to improve it?

[dhayden]

It's unrelated to the original question, but I urge you to use array syntax at line 5:
for (size_t i = 0; i < 4; ++i) e[i] = (val_a[i] < val_b[i]) ? val_a[i] : val_b[i]);
or even better:
for (size_t i = 0; i < 4; ++i) e[i] = std::min(val_a[i], val_b[i]);

[CakeByTheOcean]

Hi dhayden.
Thanks for your reply. Do these ways have different performances?

[dhayden]

Do these ways have different performances?

Probably not. And if they do, I doubt that it's worth optimizing it.

Write code that is clear.
If it's too slow, profile it to find the actual bottlenecks.
Optimize the bottlenecks.

[CakeByTheOcean]

I did this question also in stackoverflow
https://stackoverflow.com/questions/54394350/simd-implement-mm256-max-epu64-and-mm256-min-epu64