AoC 2020 - Day 15: Rambunctious Recitation
It took my quite a while to parse the instructions. I know that in these cases I
make quite a few mistakes, so I wanted to be extra cautious.
In the end, the code was quite simple. In a different language, I would have
used recursion, but this would have hit python’s limit before reaching the
result, so I went with an iterative approach.
Then I went to part 2. It looked like my code was able to do it without any
special modification so I tried it.
And then I stopped after about 20 seconds, thinking that it was wrong, and that
it needed some special optimization. So I started to trace the output, trying to
find cycles, but with no luck.
Eventually, I put a progress bar – to discover that yes, it was slow, but it
was going to give a solution, if only I were to wait a bit more.
After that, I tried to optimize the code.
The initial approach was taking a bit more than a minute, but it was maintaining
a list with all the occurrences, while we needed only the last two.
So I swapped that with a tuple, and processing time went considerably down.
I was also able to shave a bit more time, down to about 26 seconds, using some
of the usual tricks, like caching dictionary lookups.
One of the most interesting finding was that, according to profiler, most of the
time now is spent in subtracting two integers (PyNumber_Subtract
). Now, I
remember watching a presentation, about the costs of simple operations like
additions, and that the support for seamless switching to infinite precision
increased that cost even more; but I find it a bit disappointing that we cannot
“hint” that this is a small number, and that elementary operations could be done
in a much more efficient way. Maybe the assumption is that in such cases you
would directly use numpy?
Another round of optimization, this time with subsequent algorithm refactoring,
brought to another nice speed increase.
The idea was to get rid of the tuple maintaining the two last times the number
was spoken. Doing so, allowed getting rid of quite a few checks, for cases that
were not happening any longer. Also, getting rid of the tuple allowed to reduce
the number of object allocated, giving a good improvement as well.
Now, the time is down to approximately 10s, and again, the “hot path” is either
on the increment operator (when using while
loop) or in the for declaration.