search

microsoft / STL

MSVC's implementation of the C++ Standard Library.
Other
10.08k stars 1.49k forks source link

`<random>` Pick a different backing random number engine for `default_random_engine` #4257

Open horenmar opened 10 months ago

horenmar commented 10 months ago

Currently it is alias for mt19937. mt19937 is not a bad generator on its own, but it is not a good default. The reason is that MTs are excessively large, so sizeof(default_random_engine) == 5000 right now. This is much larger than the user would reasonably expect, for comparison, xoshiro fits into 16-32 bytes, PCG usually fits into 8-16, common LCGs will be 8. WELLs can be as large as MTs, but there are also reasonably sized variants at 32 bytes.

The choice of default_random_engine also has knock-on effects beyond just the direct users, e.g. the simple random utilities wanted to use a thread-local instance of default_random_engine, but that means shoving 5k objects into the TLS. As I understand it, there is nothing really blocking that, but it is not great.

fsb4000 commented 10 months ago

Thank you for your suggestion! It seems reasonable, but it would require an ABI-breaking change. We can consider implementing it in the vNext version (not Visual Studio 2022).

vNext note: Resolving this issue will require breaking binary compatibility. We won't be able to accept pull requests for this issue until the vNext branch is available. See https://github.com/microsoft/STL/issues/169 for more information.

fsb4000 commented 10 months ago

изображение

horenmar commented 10 months ago

Yeah I know, I talked with STL somewhere and he said to open an issue so it might happen in vNext.

StephanTLavavej commented 10 months ago

LCGs are garbage but if there were a better option than mt19937[_64] then we should consider it for vNext. There appears to be nothing technically preventing default_random_engine from being a choice outside the Standard set, so it could be an implementation of PCG or some other algorithm.