Foo_ (foo.delete@this.nomail.com) on January 6, 2020 10:06 am wrote:
> Chester (lamchester.delete@this.gmail.com) on January 6, 2020 9:19 am wrote:
> >
> > So I'm wondering why spinlocks were used in the first place (not necessarily by the author - tons
> > of games use it apparently). My thought is frequency scaling.
>
> My guess is micro-optimization obsession. Having worked previously with video games developers
> (admiteddly long ago), my experience is that some of them think they are smarter than experts
> - smarter than compiler authors, smarter than standard libary developers, etc. They will insist
> to express possible micro-optimizations explicitly, and their C++ looks like assembler code.
> Often those people were (former or current) demomakers. So I'm not surprised that some video
> game developers would sprinkle their code with hand-written spinlock implementations.
>
> /rant end :-)

I think over-emphasizing micro-optimization is a common failure when doing anything that is perceived as requiring optimization. Instead of taking a step back and gaining orders of magnitude by some changes in algorithms, it's much easier to notice something small, like slow hashmap, bad allocator in vector or expensive locking, and try to optimize that.

Although not a micro-optimization, one of my usual complaints over the years has been people's eagerness to jump to multi-threading code, while it will only offer 2-3 times more performance when it's most needed (e.g. Nintendo Wii U or Switch, dual-core laptop). Some types of optimizations are just more tempting.

That said, at least during PS3 time frame, the standard libs did some really stupid things in some pretty trivial cases. Even when they aren't stupid, it is often useful to know what sort of implementation of vector, map or sorting algorithm one is using, especially in multi-platform development. That's the reason why we have in-house implementations for vectors, various maps, lists and so on.

We don't implement mutexes, semaphores or atomics ourselves, but wrap platform specific structures and functions instead (we can currently use C++11 and std::stuff is still too random in details on various platforms). In some places we have used clever locking algorithms. The number of problems in those parts of the code have convinced me that we shouldn't, but maybe the performance benefits are great.

-JLarja