Decrease time for world.gop.reduce

[EricaCMitchell]

Decreases time for world.gop.reduce by shortening the buffer length to the minimum number of elements and only forming the buffers if the child process exists. Significantly, speeds up sum reduction for std::complex type.
@evaleev

[JonathonMisiewicz]

Does std::make_unique not work here? The less new, the better.

[EricaCMitchell]

I'll test timing with make_unique, when I ran tests with it by just switching the original code it resulted in longer test times in MPQC.

[JonathonMisiewicz]

That's annoying. For future reference, if there is a slowdown, there's a good chance this is the cause. Not that we can do anything about that as long as we maintain C++17 support.

[EricaCMitchell]

Haven't seen a difference in the timings between the two in the MPQC test cases. I'll push the std::make_unique implementation.

[evaleev]

I would avoid std::make_unique due to the issues discussed in Jonathon's link ... no reason to pay for value initialization of primitive types.

In fact, since this function only makes sense when T has trivial default ctor and trivial copy ctor (otherwise just copying bytes from MPI will not produce valid T objects) then we should switch to aligned malloc ... This will avoid us having to pay for initialization in the case of T=std::complex.

[EricaCMitchell]

This should now be accomplished with std::malloc. I tried std::aligned_alloc but the alignment constraint using alignof(T) resulted in a nullptr return when child process was present.

[evaleev]

I tried std::aligned_alloc but the alignment constraint using alignof(T) resulted in a nullptr return when child process was present.

possibly due to asking for size that is not a multiple of alignment? See: https://en.cppreference.com/w/cpp/memory/c/aligned_alloc#Notes

[EricaCMitchell]

Double-checked, that doesn't seem to be the issue. From the code below and running h2-sci-x2c-1root-np2.json

std::cout << "Alignment: " << alignof(T) << " \n"
                << "Test: " << (sizeof(T) * std::min(nelem_per_maxmsg, nelem)) % alignof(T) << std::endl;

the output shows that the size is always a multiple of alignment.

[evaleev]

the problem with the current solution is that alignment of T is not guaranteed ... I pushed changes to reduce and concat0 that hopefully address it, but not tested, please test, report

[EricaCMitchell]

Using the most recent changes, the test set of sci-x2c in MPQC is decreased from 27.2s to 13.9s compared to the unaligned malloc.