Hi Bas,<div><br></div><div>First of all, thanks for these numbers. I have previously compared the performance of GP libs [1] and your results confirm what I would expect, except for that last one, <span class="Apple-style-span" style="color: rgb(34, 34, 34); font-family: arial, sans-serif; font-size: 13px; background-color: rgb(255, 255, 255); ">BigSum/fromJSON/generic.</span></div>
<div><span class="Apple-style-span" style="color: rgb(34, 34, 34); font-family: arial, sans-serif; font-size: 13px; background-color: rgb(255, 255, 255); "><br></span></div><div><span class="Apple-style-span" style="color: rgb(34, 34, 34); font-family: arial, sans-serif; font-size: 13px; background-color: rgb(255, 255, 255); ">It's good that you're using INLINE pragmas on the generic function already. What I would also try:</span></div>
<div><span class="Apple-style-span" style="color: rgb(34, 34, 34); font-family: arial, sans-serif; font-size: 13px; background-color: rgb(255, 255, 255); ">- Compile with -O2 and -fno-spec-constr-count (this last one is particularly important)</span></div>
<div><span class="Apple-style-span" style="color: rgb(34, 34, 34); font-family: arial, sans-serif; font-size: 13px; background-color: rgb(255, 255, 255); ">- Add {-# INLINE [1] #-} pragmas to the to/from methods of your Generic instances.</span></div>
<div><span class="Apple-style-span" style="color: rgb(34, 34, 34); font-family: arial, sans-serif; font-size: 13px; background-color: rgb(255, 255, 255); "><br></span></div><div><font class="Apple-style-span" color="#222222" face="arial, sans-serif">To add these INLINE pragmas you will have to give your own instances of Generic (so you can't derive them). I'd suggest you get hold of them with -ddump-deriv, copy-paste and add the pragmas, just for testing purposes. The phase is important: you first want to make sure you inline the generic function definition, and only then the from/to.</font></div>
<div><font class="Apple-style-span" color="#222222" face="arial, sans-serif"><br></font></div><div><font class="Apple-style-span" color="#222222" face="arial, sans-serif">Please keep me posted on the effects of these suggestions. In particular, if the INLINE pragmas on the from/to methods are essential, I'll be happy to add them to the derived instances.</font></div>
<div><br></div><div><br></div><div>Cheers,</div><div>Pedro</div><div><br></div><div>[1] <a href="http://dreixel.net/research/pdf/ogie.pdf">http://dreixel.net/research/pdf/ogie.pdf</a><br><br><div class="gmail_quote">2011/11/3 Bas van Dijk <span dir="ltr"><<a href="mailto:v.dijk.bas@gmail.com">v.dijk.bas@gmail.com</a>></span><br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">Hello,<br>
<br>
I recently added default generic implementations of toJSON and<br>
parseJSON to the aeson package. Now I'm optimizing them. Here are some<br>
benchmark results that compare:<br>
<br>
* th: toJSON and fromJSON generated by template-haskell. Can be<br>
compared to hand-written code. Should be the fastest of all.<br>
<br>
* syb: toJSON and fromJSON from the Data.Aeson.Generic module. Uses<br>
the Data type class.<br>
<br>
* generic: my toJSON and fromJSON using GHC Generics.<br>
<br>
The benchmark itself can be found here:<br>
<a href="https://github.com/basvandijk/aeson/blob/optimizations/benchmarks/AesonCompareAutoInstances.hs" target="_blank">https://github.com/basvandijk/aeson/blob/optimizations/benchmarks/AesonCompareAutoInstances.hs</a><br>
<br>
toJSON<br>
======<br>
<br>
D/toJSON/th 3.631734 us<br>
D/toJSON/syb 32.66679 us<br>
D/toJSON/generic 3.371868 us<br>
<br>
BigRecord/toJSON/th 8.982990 us<br>
BigRecord/toJSON/syb 48.90737 us<br>
BigRecord/toJSON/generic 8.971597 us<br>
<br>
BigProduct/toJSON/th 1.578259 us<br>
BigProduct/toJSON/syb 29.21153 us<br>
BigProduct/toJSON/generic 1.623115 us<br>
<br>
BigSum/toJSON/th 51.81214 ns<br>
BigSum/toJSON/syb 1.256708 us<br>
BigSum/toJSON/generic 71.32851 ns<br>
<br>
<br>
fromJSON<br>
========<br>
<br>
D/fromJSON/th 7.017204 us<br>
D/fromJSON/syb 23.46567 us<br>
D/fromJSON/generic 7.968974 us<br>
<br>
BigRecord/fromJSON/th 8.513789 us<br>
BigRecord/fromJSON/syb 36.64501 us<br>
BigRecord/fromJSON/generic 10.07809 us<br>
<br>
BigProduct/fromJSON/th 2.430677 us<br>
BigProduct/fromJSON/syb 17.97764 us<br>
BigProduct/fromJSON/generic 2.201130 us<br>
<br>
BigSum/fromJSON/th 414.8699 ns<br>
BigSum/fromJSON/syb 4.113170 us<br>
BigSum/fromJSON/generic 13.62614 us !!!<br>
<br>
<br>
As can be seen, in most cases the GHC Generics implementation is much<br>
faster than SYB and just as fast as TH. I'm impressed by how well GHC<br>
optimizes the code!<br>
<br>
Unfortunately the last benchmark, generically parsing a big sum type,<br>
is much slower. The code for parsing sums, which can be found here:<br>
<br>
<a href="https://github.com/basvandijk/aeson/blob/optimizations/Data/Aeson/Types/Internal.hs#L1059" target="_blank">https://github.com/basvandijk/aeson/blob/optimizations/Data/Aeson/Types/Internal.hs#L1059</a><br>
<br>
is basically this:<br>
<br>
<br>
instance (GFromSum a, GFromSum b) => GFromJSON (a :+: b) where<br>
gParseJSON (Object (M.toList -> [keyVal])) = gParseSum keyVal<br>
gParseJSON v = typeMismatch "sum (:+:)" v<br>
{-# INLINE gParseJSON #-}<br>
<br>
<br>
class GFromSum f where<br>
gParseSum :: Pair -> Parser (f a)<br>
<br>
instance (GFromSum a, GFromSum b) => GFromSum (a :+: b) where<br>
gParseSum keyVal = (L1 <$> gParseSum keyVal) <|><br>
(R1 <$> gParseSum keyVal)<br>
{-# INLINE gParseSum #-}<br>
<br>
instance (Constructor c, GFromJSON a, ConsFromJSON a) =><br>
GFromSum (C1 c a) where<br>
gParseSum (key, value)<br>
| key == pack (conName (undefined :: t c a p)) =<br>
gParseJSON value<br>
| otherwise = notFound $ unpack key<br>
{-# INLINE gParseSum #-}<br>
<br>
<br>
notFound :: String -> Parser a<br>
notFound key = fail $ "The key \"" ++ key ++ "\" was not found"<br>
{-# INLINE notFound #-}<br>
<br>
<br>
Any idea how to make it faster?<br>
<br>
Regards,<br>
<span class="HOEnZb"><font color="#888888"><br>
Bas<br>
</font></span></blockquote></div><br></div>