Comments on: Use a Better Algorithm and Beat Haskell Today!

By: Ungraspiness

Ungraspiness — Fri, 30 Nov 2007 05:59:45 +0000

Lisp !

http://jng.imagine27.com/articles/2007/11/29/dude-your-quad-cores-have-been-smoking-way-too-much-haskell

By: Jake

Jake — Fri, 30 Nov 2007 01:38:48 +0000

@dons:

You didn’t do anything wrong, but I had used the C++ interface instead of the C interface for legibility reasons. (there is some overhead involved)

  mpz_class fib(unsigned int num)
  {
    mpz_t a, b;
 
    mpz_init(a);
    mpz_init(b);
 
    mpz_set_ui(a, 1);
    mpz_set_ui(b, 1);
 
    while(num-- > 1)
    {
      mpz_swap(a, b);
      mpz_add(a, a, b);
    }
    cout<<mpz_class(a)<<endl;
  }

Time: 1.85574 seconds

By: magnus

magnus — Fri, 30 Nov 2007 01:27:19 +0000

If I define the equivalent program:

> fibs = 1:1:(zipWith (+) fibs (tail fibs))
>
> main = do let x = fibs !! 200000 in print “done”

and compile it,

% ghc tmp.hs
% time ./a.out
“done”
./a.out 0.00s user 0.00s system 84% cpu 0.003 total

0 seconds!!

anyway if I shove a seq in there to force evaluation I get,
8.73s user 0.04s system 99% cpu 8.773 total

By: Joel Parish

Joel Parish — Fri, 30 Nov 2007 01:06:02 +0000

The Parallel haskell post you referenced specifically stated that they were using the *naive* algorithm in order to test ease of parallelization rather than that specific algorithm. Binet’s Formula, the Matrix Multiplication method will yield faster results than iteration for larger n. If you want an apples to apples comparison to your iterative method consider:
xor ax, ax push ax mov ax, 1 push ax mov cx, 0x24 ; number of loops, in hex


crunch:
pop ax

pop bx

add sp, 4

add ax, bx

push ax

loop crunch

ret
from http://patrickyeon.blogspot.com/2007/11/holy-cannoli-here-comes-assembly.html

By: dons

dons — Fri, 30 Nov 2007 00:29:27 +0000

I’ll see your linear C++, and raise you some linear Haskell:

main = print (fib 200000)

fib :: Int -> Integer
fib n = go n 1 1
where
go 1 a _ = a
go n a b = go (n-1) (a + b) a

On my machine, your code:

$ g++ fib.cpp -O2 -lgmp -lgmpxx -o fib
$ time ./fib
Fibonacci test for Fib(200000)
Time: 1.34055 seconds
./fib 1.32s user 0.02s system 98% cpu 1.360 total

And the Haskell translation:

$ ghc-6.8.1 -O2 -optc-O2 A.hs -o A
$ time ./A
2440914874035154859992549871532124086962886075690…
…622248628614359062
./A 0.48s user 0.01s system 98% cpu 0.504 total

Maybe you need more cores! (Or did I do something wrong?)