[Haskell-cafe] haskell version of fractal benchmark

[Scott Cruzen]

* Andrew Coppin <andrewcoppin at btinternet.com> [070608 02:45]:
>    Bayley, Alistair wrote:
> 
> [[1]mailto:haskell-cafe-bounces at haskell.org] On Behalf Of Andrew Coppin
> 
> Donald Bruce Stewart wrote:
> 
> Some things to remember using Doubles:
> 
>     * {-# OPTIONS -fexcess-precision #-}
>     * -fvia-C
>     * -fbang-patterns
>     * -optc-O2 -optc-mfpmath=sse -optc-msse2
>     * -optc-march=pentium4
> 
> 1. What do all those things do?
> 2. Is the effect actually that large?
> 
> Large? Depends what you mean by large, but adding a few flags to get
> just a 10-20% speedup isn't to be ignored:
> 
>    Sure - if it really is 10-20%. (And not, say, 0.001 - 0.002%.)

A single data point for all of this, I have a program that calculates:

P^1_i = S_i/sum_k S_k
P^m_i = sum_{k!=i} P^1_k*P^m-1_i(S_~k)

Here's timings for the different options:

options                  run time    compile time
none                      46.401       3.136
-O                         5.033       4.906
-O2                        4.967       6.755
-O2 -fexcess-precision     3.710       6.396
all listed options         3.602       6.344

Results with -fexcess-precision are very insignificantly different
(1.0 e-7).