@lru_cache on functions with no arguments

[tom at tomforb.es]

On Thursday, 3 August 2017 16:37:22 UTC+1, Ian  wrote:
> On Thu, Aug 3, 2017 at 8:35 AM, Paul  Moore ... at gmail.com> wrote:
> > On Tuesday, 1 August 2017 15:54:42 UTC+1, t... at tomforb.es  wrote:
> >> > _sentinel = object()
> >> > _val = _sentinel
> >> > def val():
> >> >     if _val is _sentinel:
> >> >         # Calculate _val
> >> >     return _val
> >> >
> >> > seems entirely sufficient for this case. Write a custom decorator if you use the idiom often enough to make it worth the effort.
> >>
> >> I did some timings with this as part of my timings above and found it to be significantly slower than lru_cache with the C extension. I had to add `nonlocal` to get `_val` to resolve, which I think kills performance a bit.
> >>
> >> I agree with the premise though, it might be worth exploring.
> >
> > It's worth pointing out that there's nothing *wrong* with using lru_cache with maxsize=None. You're going to find it hard to get a pure-Python equivalent that's faster (after all, even maintaining a single variable is still a dict lookup, which is all the cache does when LRU functionality is disabled).
> 
> The single variable is only a dict lookup if it's a global. Locals and
> closures are faster.
> 
> def simple_cache(function):
>     sentinel = object()
>     cached = sentinel
> 
>     @functools.wraps(function)
>     def wrapper(*args, **kwargs):
>         nonlocal cached
>         if args or kwargs:
>             return function(*args, **kwargs)  # No caching with args
>         if cached is sentinel:
>             cached = function()
>         return cached
>     return wrapper
> 
> *Zero* dict lookups at call-time. If that's not (marginally) faster
> than lru_cache with maxsize=None I'll eat my socks.

I hope you washed them!

def test_func():
   return 1

%timeit test_simple_cache()
The slowest run took 11.06 times longer than the fastest. This could mean that an intermediate result is being cached.
10000000 loops, best of 3: 183 ns per loop

%timeit test_lru_cache()
The slowest run took 42.92 times longer than the fastest. This could mean that an intermediate result is being cached.
10000000 loops, best of 3: 46.9 ns per loop

%timeit test_func()
10000000 loops, best of 3: 55.3 ns per loop