I recently wrote an article an SoA
https://maikklein.github.io/post/soa-d/
But now I wanted to actually benchmark SoA vs AoS and it is so
much harder than I thought.
In DMD SoA basically always beats AoS by a huge chuck. SoA is
always at least twice as fast compared to AoS.
But with LDC it is just really hard to get reliable results.
I have created a example without any dependencies
http://dpaste.dzfl.pl/877a925e0a33
dub run -b release --compiler=dmd
benchmarking complete access
AoS: 419 ms, 938 μs, and 2 hnsecs
SoA: 11 μs and 8 hnsecs
benchmarking partial access
AoS: 521 ms, 381 μs, and 3 hnsecs
SoA: 5 μs
dub run -b release --compiler=ldc
benchmarking complete access
AoS: 1 μs and 6 hnsecs
SoA: 1 hnsec
benchmarking partial access
AoS: 1 hnsec
SoA: 1 hnsec
The problem I have is that LDC always seems to optimize the
functions too much. At least one function executes always in "1
hnsec".
When I do manage do create an testcase where AoS and SoA have a
reasonable time, then they are equally fast, no matter what I do.
What I wanted to see:
If I iterate over a collection and I always access all members, I
would assume that AoS vs SoA should be equally fast.
If I iterate over a collection where the elements have a big size
and I only access some members I would assume SoA to be much
faster.
Any help tips are greatly appreciated.
