Shootout/Nsieve

From HaskellWiki
Revision as of 18:06, 4 March 2010 by DavidLazar (talk | contribs) (fix table)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

A ShootoutEntry for the nsieve benchmark

diff program output N = 2 with this output file to check your program is correct before contributing.

Each program should count the prime numbers from 2 to M, using the same nave Sieve of Eratosthenes algorithm:

  • create a sequence of M boolean flags
  • for each index number
    • if the flag value at that index is true
      • set all the flag values at multiples of that index false
      • increment the count

Calculate 3 prime counts, for M = 2N 10000, 2N-1 10000, and 2N-2 10000.

The basic benchmark was described in "A High-Level Language Benchmark." BYTE, September 1981, p. 180, Jim Gilbreath.

Of course, there are more efficient implementations of the Sieve of Eratosthenes, and there are more efficient ways to sieve prime numbers, for example "Prime sieves using binary quadratic forms".

For more information see Eric W. Weisstein, "Sieve of Eratosthenes." From MathWorld PrimeCountingFunction--A Wolfram Web Resource.


Benchmarks[edit]

Debian Linux/x86, N=10

Entry Time
Old 1.961
New 0.669

Alternative entry[edit]

Use ST.

{-# OPTIONS -O2 -optc-O -fbang-patterns #-}
--
-- The Computer Language Shootout
-- http://shootout.alioth.debian.org/
--
-- Contributed by Don Stewart
-- nsieve over an ST monad Word8 array
--

import Control.Monad.ST
import Data.Array.ST
import Data.Array.Base
import System
import Control.Monad
import Data.Bits
import Data.Word
import Text.Printf

main = do
    n <- getArgs >>= readIO . head :: IO Int
    mapM_ (\i -> sieve (10000 `shiftL` (n-i))) [0, 1, 2]

sieve n = do
   let r = runST (do a <- newArray (2,n) 0 :: ST s (STUArray s Int Word8)
                     go a n 2 0)
   printf "Primes up to %8d %8d\n" (n::Int) (r::Int) :: IO ()

go !a !m !n !c
    | n == m    = return c
    | otherwise = do
          e <- unsafeRead a n
          if e == 0 then set (n `shiftL` 1)
                       else go a m (n+1) c
  where
    set !j
        | j < m     = unsafeWrite a j 1 >> set (j+n)
        | otherwise = go a m (n+1) (c+1)

Current entry[edit]

Following the spec, uses Word8's to represent bools. Instead of IOUArray Int Word8, which has a slow initialisation phase, we use a ByteString, and call memset to init. Faster than unoptimised C.

The GHC native code gen seems to do a better job on this one.

{-# OPTIONS -O2 -fasm -fbang-patterns #-}
--
-- The Computer Language Shootout
-- http://shootout.alioth.debian.org/
--
-- Contributed by Don Stewart.
-- Uses Word8 values to represent Bools, avoiding a bit-packing Array Bool
--

import System
import Foreign
import Data.ByteString.Internal
import Data.ByteString.Unsafe (unsafeIndex)
import Text.Printf

main = do
    n <- getArgs >>= readIO . head
    mapM_ (sieve . (10000 *) . (2 ^)) ([n, n-1, n-2] :: [Int])

sieve n = do
    a <- create n $ \p -> memset p 0 (fromIntegral n) >> return ()
    r <- go n a 0 2
    printf "Primes up to %8d %8d\n" (n::Int) (r::Int)

go m !a !c !n
    | n == m    = return c
    | true a n  = go m a c (n+1)
    | otherwise = set (n+n)
  where
    set !j | j <= m    = false a j >> set (j+n)
           | otherwise = go m a (c+1) (n+1)

true  !a !n          = unsafeIndex a n == 1
false (PS fp _ _) !n = withForeignPtr fp $ \p -> pokeByteOff p n (1 :: Word8)

IOUArrays[edit]

A bit slower than using ByteStrings, since the initialisation phase is poor.

{-# OPTIONS -fbang-patterns #-}
--
-- The Computer Language Shootout
-- http://shootout.alioth.debian.org/
--
-- Contributed by Don Stewart
-- Uses Word8 values to represent Bools, avoiding a bit-packing Array Bool
--

import System
import Data.Array.IO
import Data.Array.Base
import Text.Printf
import Word

main = do
    n <- getArgs >>= readIO . head
    mapM_ (sieve . (10000 *) . (2 ^)) [n, n-1, n-2]

sieve n = do
    a <- newArray (2,n) 0 :: IO (IOUArray Int Word8) -- avoid bit packing Bool type
    r <- go n a 0 2
    printf "Primes up to %8d %8d\n" (n::Int) (r::Int)

go m !a !c !n
    | n == m    = return c
    | otherwise = do
        e <- unsafeRead a n
        if e == 0
            then let loop !j | j <= m    = unsafeWrite a j 1 >> loop (j+n)
                             | otherwise = go m a (c+1) (n+1)
                 in loop (n+n)
            else go m a c (n+1)

Illegal entry[edit]

Ported to GHC 6.6 Submitted

Uses bit packing, so is instead the nsieve-bits entry.

{-# OPTIONS -fbang-patterns #-}
--
-- The Computer Language Shootout
-- http://shootout.alioth.debian.org/
--
-- Contributed by Don Stewart
-- Nsieve over a Bool array
--

import Data.Array.IO
import Data.Array.Base
import System
import Text.Printf

main = do
    n <- getArgs >>= readIO . head :: IO Int
    mapM_ (sieve . (10000 *) . (2 ^)) [n, n-1, n-2]

sieve n = do
    a <- newArray (2,n) True :: IO (IOUArray Int Bool) -- an array of Bool
    r <- go a n 2 0
    printf "Primes up to %8d %8d\n" (n::Int) (r::Int) :: IO ()

go !a !m !n !c
    | n == m    = return c
    | otherwise = do
            e <- unsafeRead a n
            if e
                then let loop !j
                            | j <= m    = unsafeWrite a j False >> loop (j+n)
                            | otherwise = go a m (n+1) (c+1)
                     in loop (n+n)
                else go a m (n+1) c

Old entry[edit]

Studying the Core shows the the mapM was preventing things from being unboxed properly. Big speedup. The extra `seqs` also help the unboxing. Should be around the fastest entry now.

Is anyone else bothered by the usage of unsafeWrite? I agree that excellent speed is being gained, but this speed is gained at the cost of functional-style coding. I'm a novice Haskeller, but using unsafeWrite seems like an imperative hack to gain speed. -- AlsonKemp

Response: unsafeWrite is just the same as a normal array update, without an extraneous bounds check (since it's already performed in the outer loop). We're just avoiding redundant computation (as we do in all entries for this problem). SPJ actually recommends unsafe* ops in arrays, see the "loop performance bug" thread in the haskell mail archives from 2005.

In summary, its not a hack -- the name just makes clear that you have to do your own bounds check (Haskell is good at letting you know when there are extra proof considerations for the programmer to take). Since this program runs faster than GCC, then I think we should not be worried about avoiding a redundant bounds check :) -- DonStewart

--
-- $Id: nsieve-ghc.code,v 1.27 2006/01/08 22:44:56 igouy-guest Exp $
-- Written by Einar Karttunen, optimised further by Don Stewart
--

import Data.Array.IO; import Data.Array.Base; import Data.Bits; import System; import Text.Printf

main = (\n -> mapM_ (sieve.(10000 *).shiftL 1) [n,n-1,n-2]) . read . head =<< getArgs

sieve m = do c <- newArray (2,m) True >>= \a -> loop a m 2 0
             printf "Primes up to %8d %8d\n" (m::Int) (c::Int) :: IO ()

loop arr m n c | arr `seq` m `seq` n `seq` c `seq` False = undefined
loop arr m n c = if n == m then return c else do
    el <- unsafeRead arr n
    if el then let loop' j | j > m     = loop arr m (n+1) (c + 1)
                           | otherwise = unsafeWrite arr j False >> loop' (j+n)
               in loop' (n+n)
          else loop (arr :: IOUArray Int Bool) m (n+1) c

Entry 2[edit]

Shortest entry in any language

{-# OPTIONS -O2 -optc-O3 #-}

-- $Id: nsieve-ghc.code,v 1.27 2006/01/08 22:44:56 igouy-guest Exp $
-- Written by Einar Karttunen, shortened by Don Stewart

import Data.Array.IO; import Data.Array.Base; import Data.Bits; import System; import Text.Printf

loop arr m n c = if n == m then return c else do
    el <- unsafeRead arr n
    if el then do mapM_ (flip (unsafeWrite arr) False) (tail [n,n+n..m])
                  loop arr m (n+1) $! c + 1
          else loop (arr :: IOUArray Int Bool) m (n+1) c

sieve m = do c <- newArray (2,m) True >>= \a -> loop a m 2 0
             printf "Primes up to %8d %8d\n" (m::Int) (c::Int) :: IO ()

main = (\n -> mapM_ (sieve.(10000 *).shiftL 1) [n,n-1,n-2]) . read . head =<< getArgs

Original entry[edit]

{-# OPTIONS -O2 -optc-O3 #-}

-- $Id: nsieve-ghc.code,v 1.27 2006/01/08 22:44:56 igouy-guest Exp $
-- written by Einar Karttunen

import Data.Array.IO
import Data.Array.Base
import Data.Bits (shiftL)
import System (getArgs)

loop :: IOUArray Int Bool -> Int -> Int -> Int -> IO Int
loop arr m n c | n == m    = return c
               | otherwise = do el <- unsafeRead arr n
                                if el then do mapM_ (\i -> unsafeWrite arr i False) (tail [n,n+n..m])
                                              loop arr m (n+1) $! c+1
                                      else do loop arr m (n+1) c

fmt width i = let is = show i in (replicate (width - length is) ' ') ++ is

sieve n = do let m = (1 `shiftL` n) * 10000
             arr <- newArray (2,m) True
             c   <- loop arr m 2 0
             putStrLn ("Primes up to "++fmt 8 m++" "++fmt 8 c)

main = do n <- getArgs >>= readIO.head
          sieve n >> sieve (n-1) >> sieve (n-2)