comparison for load/store vs loadu/storeu

[mperikov]

The following instructions were compared:

1. _mm256_loadu_si256 vs _mm256_load_si256
2. _mm256_storeu_si256 vs _mm256_store_si256

No obvious differences in execution time were observed.

Also need to compare analogs from AVX512.

[Malkovsky]

Missing core functionality

[Malkovsky]

We discussed that we want to compare:
-- Aligned store/load
-- Aligned storeu/loadu
-- Unaligned storeu/loadu within a single 64-byte block (for <=256-bit registers)
-- Unaligned storeu/loadu crossing a 64-byte block border
-- (optional) test that unaligned store/load crashes

[Malkovsky]

Also I think that benchmarks are better be organized so that store/load are performed to a conditionally random address to an array of different sizes so that we also see the impact of the cache misses. Specifically we definitely expect some degradation for store/load on an address that crosses 64-byte block border.

[Malkovsky]

I.e. for example unaligned storeu/loadu within a 64-byte block should be something like

  alignas(64) uint8_t data[64 * n];

  for (auto _ : state) {
    const __m256i* ptr = reinterpret_cast<const __m256i*>(data + 1 + 32 * (rng() % (n - 1)));
    benchmark::DoNotOptimize(_mm256_loadu_si256(ptr));
  }

Note that rng() call and % might be heavy in this context

[Malkovsky]

it is probably better to make n=2^k+1 and perform rng() & ((1 << k) - 1).