Personal tools

Wc

From HaskellWiki

(Difference between revisions)
Jump to: navigation, search
(Change Data.ByteString to use B.count which makes it twice as fast.)
m (Use the FFI: drop use of STRICT macro and use bang patterns)
 
Line 143: Line 143:
   
 
<haskell>
 
<haskell>
  +
{-# LANGUAGE BangPatterns #-}
 
import Foreign
 
import Foreign
 
import Foreign.ForeignPtr
 
import Foreign.ForeignPtr
Line 149: Line 150:
 
import System.Environment
 
import System.Environment
 
import qualified Data.ByteString as B
 
import qualified Data.ByteString as B
 
#define STRICT4(f) f a b c d | a `seq` b `seq` c `seq` d `seq` False = undefined
 
   
 
main = do
 
main = do
Line 158: Line 157:
 
where
 
where
 
go :: Ptr Word8 -> Int -> Int -> Int -> IO ()
 
go :: Ptr Word8 -> Int -> Int -> Int -> IO ()
STRICT4(go)
+
go !p !l !n !i
go p l n i
 
 
| n >= l = print i
 
| n >= l = print i
 
| otherwise = do
 
| otherwise = do

Latest revision as of 10:40, 9 April 2013


Some implementations of the 'wc -l' program in Haskell, with an eye to C-like performance. This illustrates the balance to be made between performance and elegance, over several increasingly fast (and more complex) examples.

Contents

[edit] 1 Baseline

The baseline is the C program 'wc'

$ du -hsL /usr/share/dict/words
912K    /usr/share/dict/words

$ time wc -l < /usr/share/dict/words 
98326
wc -l < /usr/share/dict/words  0.00s user 0.00s system 27% cpu 0.015 total

So the best we can probably hope to get is around 0.015s

[edit] 2 Standard [Char]

main :: IO ()
main = print . length . lines =<< getContents
$ ghc -O wc.hs
$ time ./wc < /usr/share/dict/words
98326
./wc < /usr/share/dict/words  0.10s user 0.00s system 94% cpu 0.106 total

Ok. About 10x C, as to be expected with a list representation.

[edit] 3 Faster [Char]

Perhaps writing our loop, rather than the duplication involved in length . lines, will improve things:

main :: IO ()
main = interact (count 0)
    where count i []        = show i
          count i ('\n':xs) = count (i+1) xs
          count i (_:xs)    = count i     xs
$ ghc -O wc.hs
$ time ./wc < /usr/share/dict/words
98326./wc < /usr/share/dict/words  0.06s user 0.00s system 87% cpu 0.073 total

Ok. Not too bad.

[edit] 4 Data.ByteString

Try to improve performance by using the Data.ByteString library. This uses packed byte arrays instead of heap-allocated [Char] to represent strings.

import qualified Data.ByteString.Char8 as B
 
main :: IO ()
main = B.getContents >>= print . B.count '\n'
$ time ./wc < /usr/share/dict/words
98326
./wc < /usr/share/dict/words  0.00s user 0.00s system 25% cpu 0.016 total
 

Much better, it is now becoming competitive with C. This (and the Data.ByteString.Lazy example below) is as fast as we'll get.

[edit] 5 Data.ByteString.Lazy

Or we could use the new lazy bytestring type, a lazy list of strict, L1-cache-sized chunks of bytes. This example due to Chad Scherrer:

import qualified Data.ByteString.Lazy.Char8 as L
 
main :: IO ()
main = L.getContents >>= print . L.count '\n'
$ time ./a < /usr/share/dict/words
98326
./a < /usr/share/dict/british-english  0.00s user 0.00s system 25% cpu 0.016 total

[edit] 6 Line-by-line processing

We can ask the bytestring library to hand us a string line at a time.

import System.IO
import Data.ByteString (hGetLines)
main = hGetLines stdin >>= print . length
$ time ./b < /usr/share/dict/british-english
98326
./b < /usr/share/dict/british-english  0.04s user 0.01s system 94% cpu 0.055 total

Though this is a bit slower, since it needs to hang on to the lines for longer.

[edit] 7 Ptr hacking

ByteStrings give you access to the underlying pointers to bytes in memory, which can be used to optimise some particular code. So when the ByteString api doesn't provide what you want, you can step inside the ForeignPtr and go nuts.

This example also makes use of a cpp macro to force strictness on a function, via a seq guard case.

import Foreign
import Foreign.ForeignPtr
import System.Environment
import qualified Data.ByteString as B
 
#define STRICT4(f) f a b c d | a `seq` b `seq` c `seq` d `seq` False = undefined
 
main = head `fmap` getArgs >>= B.readFile >>= \(B.PS x _ l) ->
    withForeignPtr x $ \p -> go p l 0 0
 
    where go :: Ptr Word8 -> Int -> Int -> Int -> IO ()
          STRICT4(go)
          go p l n i | n >= l    = print i
                     | otherwise = do (w::Word8) <- peek (p `plusPtr` n)
                                      go p l (n+1) $ if w == 0x0a then (i+1) else i
$ ghc -O -package fps -fglasgow-exts -cpp wc.hs
$ time ./wc /usr/share/dict/words                                                                   
98326       
./wc /usr/share/dict/words  0.00s user 0.01s system 67% cpu 0.018 total

Ok, slower than using length . lines. Lets try some other things.

[edit] 8 Use the FFI

Try and step around the inefficent need to inspect each character in Haskell, by using memchr(3), the C function to find each newline for us.

{-# LANGUAGE BangPatterns #-}
import Foreign
import Foreign.ForeignPtr
import Foreign.C.Types
 
import System.Environment
import qualified Data.ByteString as B
 
main = do
    f <- head `fmap` getArgs
    B.readFile f >>= \(B.PS x _ l) -> withForeignPtr x $ \p -> go p l 0 0
 
    where
        go :: Ptr Word8 -> Int -> Int -> Int -> IO ()
        go !p !l !n !i
           | n >= l    = print i
           | otherwise = do
                let p' = p `plusPtr` n
                    q  = memchr p' 0x0a (fromIntegral (l-n))
                if q == nullPtr
                    then print i
                    else do let k = q `minusPtr` p'
                            go p l (n+k+1) (i+1)
 
foreign import ccall unsafe "string.h memchr" memchr
    :: Ptr Word8 -> CInt -> CSize -> Ptr Word8
$ time ./wc /usr/share/dict/words
98326                            
./wc /usr/share/dict/words  0.00s user 0.00s system 47% cpu 0.017 total

Slowly inching forwards.

[edit] 9 Read the Core

While we're here, we can check whether the strictness on the 'go' function makes any difference, by reading the GHC Core:

$ ghc -O -package fps -cpp -ffi wc.hs -ddump-simpl | less

Search for the 'go' function:

Main.$wgo :: GHC.Prim.Addr#
            -> GHC.Prim.Int#
            -> GHC.Prim.Int#
            -> GHC.Prim.Int#
            -> GHC.IOBase.IO ()

And without the strictness:

Main.$wgo :: GHC.Ptr.Ptr GHC.Word.Word8
            -> GHC.Prim.Int#
            -> GHC.Prim.Int#
            -> GHC.Base.Int
            -> GHC.IOBase.IO ()

So GHC is helpfully unboxing the Ptr Word8 into a raw machine Addr#.

[edit] 10 Avoid some code

The guard that checks the length is unneeded, since memchr takes a length argument anyway. It also calculates the next pointer for us, so avoid recalculating it. (Note that this is equivalent to using the 'count' function, which has the same implementation).

import Foreign
import Foreign.ForeignPtr
import Foreign.C.Types
 
import System.Environment
import qualified Data.ByteString as B
 
#define STRICT3(f) f a b c | a `seq` b `seq` c `seq` False = undefined
 
main = do
    f <- head `fmap` getArgs
    B.readFile f >>= \(B.PS x s l) -> withForeignPtr x $ \p -> 
        go (p `plusPtr` s) (fromIntegral l) 0
    where
        go :: Ptr Word8 -> CSize -> Int -> IO ()
        STRICT3(go)
        go p l i = do
            let q  = memchr p 0x0a l
            if q == nullPtr
                then print i
                else do let k = fromIntegral $ q `minusPtr` p
                        go (q `plusPtr` 1) (l-k) (i+1)
 
foreign import ccall unsafe "string.h memchr" memchr
    :: Ptr Word8 -> CInt -> CSize -> Ptr Word8

Checking the Core, 'go' is now:

Main.$wgo :: GHC.Prim.Addr#
             -> GHC.Prim.Word#
             -> GHC.Prim.Int#
             -> GHC.IOBase.IO ()

The code is certainly a bit simpler, at least.

$ ghc -O -package fps -cpp -ffi wc.hs
$ time ./wc /usr/share/dict/words
98326
./wc /usr/share/dict/words  0.00s user 0.01s system 70% cpu 0.017 total

But we can't seem to squeeze any more out, at least on data this size.

[edit] 11 Going via C

We reach a point where I can't think of any more tricks, so we can always code up a little C and call into that, for this tight loop. Sometimes we just have to do this, and that's what the ffi is for, after all.

-- wc.hs
 
import Foreign
import System.Environment
import qualified Data.ByteString as B
 
main = do
    f <- head `fmap` getArgs
    B.readFile f >>= \(B.PS x _ l) -> withForeignPtr x $ \p -> print (c_wc p l)
 
foreign import ccall unsafe "wc.h wc" c_wc :: Ptr Word8 -> Int -> Int
 
-- wc_c.c
int wc(char *p, int len) {
    int c;
    for (c = 0; len--; ++p)
        if (*p == '\n')
            ++c;
    return c;
}
 
-- wc.h
int wc(char *p, int len);
$ gcc -O3 -c wc_c.c
$ ghc -O -package fps wc.hs -o wc -fglasgow-exts wc_c.o
$ time ./wc /usr/share/dict/words
98326
./wc /usr/share/dict/words  0.00s user 0.00s system 25% cpu 0.016 total

And we are done. Note that the tight C loop didn't give us anything in the end over the naive ByteString code, which is a very satisfying result.