The comparison is not really between i.~f and
f{iota}f , because f i. f in J is equally as fast
as i.~f .
The "secret" is that the dyad i. in J has a linear
expected time algorithm for floats even with
non-zero tolerance. I think Dyalog APL has a
quadratic time algorithm. You can test this
by timing
f {iota} f
where f{is}0.01{times}n{rho}1e6 for n in
1e3 * 1 2 4 8 16 32. If the ratio from one time
to the next increases, then you've got a quadratic
(or worst) algorithm. In J:
tt=: 3 : '6!:2 ''f i.f '' [ f=. 0.01 * y [EMAIL PROTECTED] 1e6'
2 %~/\ tt"0 ]1e3 * 1 2 4 8 16 32
1.95527 1.98379 1.99087 2.34902 2.02415
----- Original Message -----
From: [EMAIL PROTECTED]
Date: Friday, June 1, 2007 11:14 am
Subject: Re: [Jgeneral] Re: Challenge to expert J'ers
> RH> I suggest you try the benchmarks in
> RH> http://www.jsoftware.com/jwiki/Essays/Index_in_Nub
> RH> in APL.
>
> I just did ( a simplified version) and there is definitively a
> major
> problem with the dyadic iota of Dyalog APL.
> These results points to either a "bug" in both interpreters or a
> mystery...
> It seems to me that APL implementers of Dyalog would be best to
> require
> your services for dyadic iota for float...
> Here are the results (all tests done on my PC, Dell OptiPlex
> GX270,
> Processor x86 Family 15 Model 2 Stepping 9 GenuineIntel
> ~2593 Mhz
>
> i=:2e5$1e5?2e9
> f=:i%i+0.1
> J J 6.01 APL Dyalog 10.0.2
> Integer .i~i 0.065 i{iota}i 0.109
> Float .i~f 2.826 f{iota}f 1402.875
>
> Can anyone confirm these results and add other APL implementations
> if they
> so have ?
> Roger ! How can we explain that huge difference between .i~f and
> f{iota}f
> ?
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
