Archive 17/01/2023.

Yet another vector benchmark

1vanK

Urho3D::Vector vs Urho3D::PODVector vs std::vector

https://github.com/1vanK/Urho3DContainersBenchmark

Result: https://github.com/1vanK/Urho3DContainersBenchmark/blob/master/vector_result.txt Lower values are better.

Some conclusions:

  • Debug version of std::vector is very slow, but release version is fast (perhabs some bounds checks)
  • Urho3D::Vector the slowest (release version)
  • It seems std::vector have some optimizations for POD types, but Urho3D::PODVector is faster
  • Mingw is faster than Visual Studio
  • 64 bit is faster than 32 bit
vmost

Interesting that std::vector::iterator is slower than operator[]. I thought they were equivalent (naively).

Nvm was looking at debug. It is only slower for std::vector<int> in release test for MinGW 32-bit (with a flipped speed relationship in VStudio…).

Eugene

I must point out issues with this benchmarking (which may or may not substantially affect the result, I don’t know):

  • In Push tests you may be measuring performance of memory allocator (which is probably not what you want)

  • I don’t 100% trust hand-made benchmarking systems. Benchmarking libraries exist out there for a reason. Sometimes it matters, sometimes it doesn’t :man_shrugging:

  • Resulting numbers are not really helpful without error margin. I have significant fluctuations (5-10%) between runs when I run these benchmarks locally. Which means that vector_result.txt is meaningless (or even harmful/misleading) on its own.

Having said that, I have run your benchmarks locally and I think these conclusions are correct.

1vanK

In any case, the difference between implementations is minimal, so we can replace Vector and PODVector with std::vector.

rku

GCC/Clang on linux/macos + older msvc version tests would be good. msvc only recently made debug performance of containers not stupid.

This would make rbfx maintenance easier so regardless, please do it :slight_smile:

SirNate0

More importantly, this would make integration with many 3rd party libraries (for geometry processing, for example) much easier, and allow the use of the STL algorithms, and make a lot of the Stack Overflow answers applicable with the vectors without changing all of the function names. I am in favor of the change.

vmost

I am also in favor (fwiw). I have been using the STL in my own project as much as possible already :D.

SirNate0

If anyone is interested in it actually happening, I’ve started work on switching over to std::vector on this branch. There seem to be some issues with the AngelScript bindings at present, and I’m not sure when I’ll get to them as this is a lower-priority project for me.

For this, I’ve opted to change the other container’s method names to also match the standard library (so they use size() instead of Size(), for example).

glebedev

Is it possible to include eastl to the test?

1vanK

Added (txts with results in root of repo)