One theory I have is that , is optimized when ,~ is not.

in tobase64, the second line does an optional append with , .  When the length 
of the string is a multiple of 3, it appends 0#'=' (nothing).  For length 100, 
it appends 2#'='

   20 timespacex 'tobase64"1 ] 1000 100 $ a' 
0.0250075 429824 
   20 timespacex 'tob64"1 ] 1000 100 $ a' 
0.0196931 413824 
   20 timespacex 'tobase64"1 ] 1000 99 $ a' 
0.0268403 429824 
   20 timespacex 'tob64"1 ] 1000 99 $ a' 
0.0194453 413824

for the explicit version, it is strange that reading 100 chars per line and 
spitting out 2 extra takes less time than 99 chars per line with no extra 
output.

the tacit version appends through ,~ , and shows the expected improvement of 
reading and appending less.



----- Original Message -----
From: Joe Bogner <[email protected]>
To: [email protected]
Cc: 
Sent: Wednesday, August 6, 2014 2:04:15 PM
Subject: Re: [Jprogramming] unexpected explicit vs tacit benchmark

My timings suggest that they are all about the same:

100 timespacex 'tob64 a'
0.0317417 1.78275e7
   100 timespacex 'tobase64 a'
0.0316219 2.85809e7
   100 timespacex 'tb64 a'
0.0313129 2.04608e7

Considering the explicit only has 2 lines, I'm not surprised the
timings are about the same. I thought the biggest performance penalty
from explicit is the parsing for each invocation. The parsing in this
case is likely negligible since the explicit function isn't being
called in a loop.

On Wed, Aug 6, 2014 at 1:52 PM, Raul Miller <[email protected]> wrote:
> You might try timings on sub-expressions to find where the speed
> differences occur?
>
> Thanks,
>
> --
> Raul
>
>
> On Wed, Aug 6, 2014 at 1:27 PM, 'Pascal Jasmin' via Programming
> <[email protected]> wrote:
>> BASE64=: (a.{~ ,(a.i.'Aa') +/i.26),'0123456789+/'
>>
>> tobase64=: 3 : 0
>> res=. BASE64 {~ #. _6 [\ , (8#2) #: a. i. y
>> res, (0 2 1 i. 3 | # y) # '='
>> )
>>
>> tob64 =: ('=' #~  0 2 1 i. 3 | # )  ,~ BASE64 {~  [: #. _6  ]\    (8#2) 
>> ,@:#: a.&i.
>> tb64 =: 3 : '(''='' #~  0 2 1 i. 3 | # y)  ,~ BASE64 {~  #. _6  ]\    (8#2) 
>> ,@:#: a.&i. y'
>>
>> frombase64=: 3 : 0
>> pad=. _2 >. (y i. '=') - #y
>> pad }. a. {~ #. _8 [\ , (6#2) #: BASE64 i. y
>> )
>> fb64 =:a. {~ [: #.  _8 [\  (6#2) ,@:#: BASE64&i. }.~ _2 >. # -~ i.&'='
>>
>> tobase64 is original code from addon.  It is explicit with a temporary 
>> variable.  tb64 is also explicit but without the temp var.  tob64 is a tacit 
>> version.
>>
>>  a =. 100000 $ a.
>>
>>    timespacex 'tob64 a'
>>
>> 0.0312112 1.78275e7
>>    timespacex 'tobase64 a'
>> 0.0277369 2.85809e7
>>    timespacex 'tb64 a'
>> 0.0289411 2.04608e7
>>
>> all are impressively fast, but the 2 explicit functions are faster, and 
>> somehow the one with a temp variable is fastest.
>>
>> the tacit version does use less space, as well as the one line tacit entry.  
>> Its also faster for array item application
>>
>>
>>    timespacex 'tob64"1 ] 100 1000 $ a'
>> 0.0154246 539264
>>    timespacex 'tobase64"1 ] 100 1000 $ a'
>> 0.016376 628736
>>
>> Its also surprising though that applying the function to 100 smaller items 
>> is faster than the whole.
>>
>>
>> Any insights on why this is happening?
>> ----------------------------------------------------------------------
>> For information about J forums see http://www.jsoftware.com/forums.htm



> ----------------------------------------------------------------------
> For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to