search

serokell / universum

:milky_way: Prelude written in @Serokell
MIT License
176 stars 28 forks source link

Add benchmarks (#21) #66

Closed vrom911 closed 7 years ago

vrom911 commented 7 years ago

I made benchmarks on lists where numbers are not greater than 100 ("small") and where I have list consist of numbers from 1 to n repeated twice ("big").

vrom911 commented 7 years ago

@Martoon-00 , could you please review the (...) (super composition) benchmarks? It seems that it's very fast and all is within the limits of error.

vrom911 commented 7 years ago

@Martoon-00 here is the result I received: It still works okay when I'm using functions receiving many args (I used 5) : 1.071 ns and 1.072 ns super and normal composition accordingly. But the interesting thing I discovered is when you don't write down the function type for super composition it works very slow 356.6 ns ((...)/5args/unty benchmark). It seems it inferences very polymorphic type which affects optimizations.. But when types are explicitly specified no difference is noticed.

Martoon-00 commented 7 years ago

Very interesting results! Glad to know that it can work as fast as with writing it in normal way 🙂

Could you please try one more thing: function without type, but with INLINE pragma. My guess is that, ... already has this pragma, so when type of function calling it is known, ... call is rewritten in normal way by GHC. However when type is unknown then ... can not be inlined, and compiler may suggest not to inline unty5args itself as well, thus we get such drawdown

vrom911 commented 7 years ago

@Martoon-00 done! It works as fast as super typed version.

Martoon-00 commented 7 years ago

Great! I suppose these results are worth mentioning in haddoc to '(...)' eventually. Thanks for your thorough work on this!

vrom911 commented 7 years ago

I created benchmarks for concatMap. It turns out that concatMapM in Identity is 3x times slower than concatMap even on not big tests. Is that what was on your mind, @Martoon-00 ?

Martoon-00 commented 7 years ago

Wow, that's worth investigating. Thanks for finding this issue!

chshersh commented 7 years ago

@vrom911 Thanks for your amazing contribution in this issue! I really appreciate your effort for our benchmarks. I ran benchmarks and saw results for nub and other functions (following your discussion with@Martoon-00). Looks great! And I will create new issues by results of this PR.