I suspect that a significant reason for the speed increase is that the verb
ranks of * and + are 0.
* b. 0
+ b. 0
This means that %@+ for example will calculate % for each item of the right
argument, whereas if you use %@:+ then % will wait until all the items have
been processed and then calculate % on the whole array.
In general J performs better if you give it whole arrays to work on at
once, rather than forcing it to operate on the individual items in an
array.
Perhaps this will help to visualise the difference:
7 9 4 <@* 2 9 5
┌──┬──┬──┐
│14│81│20│
└──┴──┴──┘
7 9 4 <@:* 2 9 5
┌────────┐
│14 81 20│
└────────┘
If the verb rank of v is infinite, then is no difference in the processing
of u@v and u@:v so for example:
([ * ]) b. 0 NB. user-defined verbs (in this case a fork) have
infinite rank
_ _ _
7 9 4 <@([ * ]) 2 9 5
┌────────┐
│14 81 20│
└────────┘
7 9 4 <@:([ * ]) 2 9 5
┌────────┐
│14 81 20│
└────────┘
HTH
Ric
On Mon, Feb 16, 2015 at 10:08 AM, Robert Clausecker <[email protected]>
wrote:
>
> Hello!
>
> To see how J compares to other languages out there, I tried to implement
> the spectral-norm [1] language benchmark in J. Part of my code [2] for
> this looks like this:
>
> NB. entry at x,y of infinite matrix A
> A =: >:@:[ %@:+ -:@:(* >:)@:+
> NB. get the first y by y elements of A
> subA =: A"0 _~&i. M.
>
> subA 5500 is called in the benchmark. In the beginning, my definition
> for A looked like this:
>
> A =: >:@[ %@+ + -:@* >:@+
>
> and was about five times slower than the definition above, even though
> it computes the same values. Why is that the case?
>
> Yours,
> Robert Clausecker
>
> [1]:
>
> http://benchmarksgame.alioth.debian.org/u32/performance.php?test=spectralnorm
>
> [2]:
>
> https://gist.github.com/fuzxxl/ea936084dd5038cc4c06/c0db428c351af069332d456a2c5368c5fecb8d11
>
>
> ----------------------------------------------------------------------
> For information about J forums see http://www.jsoftware.com/forums.htm
>
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
