Performance improvement for `NB<T>` about svelto.ecs HOT 5 CLOSED

Here is the benchmark source I've tested

Results:

BenchmarkDotNet=v0.13.2, OS=macOS Monterey 12.3 (21E230) [Darwin 21.4.0]
Apple M1 Pro, 1 CPU, 10 logical and 10 physical cores
.NET SDK=7.0.101
  [Host]     : .NET 6.0.2 (6.0.222.6406), Arm64 RyuJIT AdvSIMD
  DefaultJob : .NET 6.0.2 (6.0.222.6406), Arm64 RyuJIT AdvSIMD

Weirdly, removing using block but not replacing SIZE makes it even slower. I think it is related to static constructor for SIZE while the code got inlined.

from svelto.ecs.

I've added benchmark result for .NET 7 and Mono to the link, and it seems accessing with pointer is much performant for those. I think using T* instead of Unsafe APIs should be considered alongside.

from svelto.ecs.

I am pretty sure (confirmed with sharplab.io) that in release compilation the using will not generate a try catch if it's empty, in any case atm that line can be commented out. However the static field killed me, I have to think a lot about it as I use this everywhere in the code. Why is your benchmarking alocating at all?

from svelto.ecs.

that in release compilation the using will not generate a try catch if it's empty

Main issue would be that using prevents inlining. I've tested with this SharpLab, and if you see the assembly it prevents inlining to CallWithUsing while method WithUsing does not contain try-finally.

Why is your benchmarking alocating at all?

It only happens if I put a very big number for Count for some reason. It happens on irrelevant benchmarks as well. I think it's MemoryDiagnoser issue.

from svelto.ecs.

according my profiing Unsafe.As is as fast as using T*

from svelto.ecs.