GCC 4.8 results posted to source forum:
http://www.jsoftware.com/pipermail/source/2014-January/000527.html
On Wed, Jan 15, 2014 at 8:49 AM, Thomas Costigliola <tcost...@gmail.com>wrote:
> Also, I'm pretty sure I tried with all optimizations turned off.
>
>
> On Wed, Jan 15, 2014 at 8:45 AM, Thomas Costigliola <tcost...@gmail.com>wrote:
>
>> Off the top of my head g000i.ijs is one. There are others, when I get
>> some time I will retry with your suggestion and post the results.
>>
>>
>> On Wed, Jan 15, 2014 at 8:19 AM, bill lam <bbill....@gmail.com> wrote:
>>
>>> I didn't test. Which tests failed with gcc 4.8? I guess it is the loop
>>> vectorize optimization that may break ar.c, you may try disable it with a
>>> CFLAG switch.
>>>
>>> 15.01.2014, в 20:38, Thomas Costigliola <tcost...@gmail.com> написал(а):
>>>
>>> > Hi Bill, did you run the test suite after compiling? I have been
>>> having
>>> > problems with gcc 4.8. I haven't been able to get a passing libj, I
>>> have
>>> > tried a few distros' gcc with different optimizations. It seems to
>>> always
>>> > be overoptimizing some things. Gcc 4.7 doesn't have the issue.
>>> >
>>> > On Jan 14, 2014 10:11 PM, "bill lam" <bbill....@gmail.com> wrote:
>>> >>
>>> >> I get similar result for Fibo that linux needed about 41s and this is
>>> >> strange. I then recompiled the libj.so using gcc 4.8 -O3 switch,
>>> >> and it droped to 14s, a little bit faster than wine's 17s.
>>> >>
>>> >> I guess it depends on compiler and optimization level.
>>> >>
>>> >> Вт, 14 янв 2014, Ben Gorte - CITG писал(а):
>>> >>> With pleasure. Here are two examples. The first is from my work
>>> (image
>>> > analysis). It computes the good old Haralick texture on a 5000x5000
>>> aerial
>>> > image. The adverb Filter uses ;. to activate TextSub at every 25th
>>> pixel
>>> > (1M times).
>>> >>>
>>> >>> ~$ ~/j701/bin/jconsole
>>> >>> load '/home/ben/j701-user/temp/pgm.ijs'
>>> >>>
>>> >>> CM =. 4 : '((# /.)~ y) (<"1 ~.y) } (x,x)$0'"0 2
>>> >>> Norm =: 3 : 'y%+/,y'
>>> >>> Contrast=: 3 : '+/, y * *:(<:#y)%~-/~i.#y'"2
>>> >>> Con =: 3 : 'Contrast Norm (+|:) y'"2 NB. for
>>> > non-normalized non-symmetrical cm
>>> >>>
>>> >>> TextSub =: 4 : 0 NB. texture in a sub-image
>>> >>> co =. |:,.(_1&}."1 ,: 1&}."1) y
>>> >>> cm =. x CM co
>>> >>> Con cm
>>> >>> )
>>> >>>
>>> >>> im =: <.(%16) * pgmread '/home/ben/data/Danbi/redsub.pgm'
>>> >>> NB. Filter applies 12&F to 7x7 subimages of <.im%16, at every 5th
>>> line
>>> > and column.
>>> >>> 6!:2 'conperpix =. <.1000*12 TextSub Filter (5 5,:7 7) <.im%16'
>>> >>> 87.5534
>>> >>>
>>> >>> ~$ wine /media/Windows/Documents\ and\
>>> > Settings/bgorte/j701/bin/jconsole.exe
>>> >>> ....
>>> >>> 6!:2 'conperpix =. <.1000*12 TextSub Filter (5 5,:7 7) <.im%16'
>>> >>> 67.088
>>> >>>
>>> >>> It's a 30% speed difference.
>>> >>>
>>> >>> Here's a more drastic example. I guess it spends most of its time
>>> doing
>>> > extended precision things. The Fibo routine is activated recursively
>>> just
>>> > 20 times or so. The result has over 200000 digits.
>>> >>>
>>> >>> ~$ ~/j701/bin/jconsole
>>> >>> Fibo =: 3 : 0"0 NB. Compute large Fibonacci numbers like Fibo
>>> > 1000000
>>> >>> if. y<5 do. NB. returns (F y-1),F y
>>> >>> ((,~<:)y){0 1 1 2 3x
>>> >>> elseif. 0=2|y do.
>>> >>> 'b c'=.Fibo -:y
>>> >>> a=.c-b
>>> >>> d=.c+b
>>> >>> p=.b*a+c
>>> >>> q=.c*b+d
>>> >>> (q-p),q
>>> >>> elseif. do.
>>> >>> 'b c'=.Fibo -:>:y
>>> >>> a=.c-b
>>> >>> d=.c+b
>>> >>> p=.b*a+c
>>> >>> q=.c*b+d
>>> >>> p,q-p
>>> >>> end.
>>> >>> )
>>> >>> 6!:2 'z=.{:Fibo 1000000'
>>> >>> 41.3298
>>> >>>
>>> >>>
>>> >>> ~$ wine /media/Windows/Documents\ and\
>>> > Settings/bgorte/j701/bin/jconsole.exe
>>> >>> ....
>>> >>> 6!:2 'z=.{:Fibo 1000000'
>>> >>> 15.5417
>>> >>>
>>> >>> #":z NB.
>>> >>> 208988
>>> >>>
>>> >>> ________________________________________
>>> >>> From: programming-boun...@forums.jsoftware.com [
>>> > programming-boun...@forums.jsoftware.com] on behalf of Michael Dykman
>>> [
>>> > mdyk...@gmail.com]
>>> >>> Sent: Tuesday, January 14, 2014 17:02
>>> >>> To: J Programming
>>> >>> Subject: Re: [Jprogramming] On benchmarking results from J
>>> programming
>>> > styles
>>> >>>
>>> >>> As the timings that you are reporting are very tiny values, we should
>>> >>> pause for a moment and consider a basic difference between those OSs.
>>> >>> Under windows, the finest-grained application timer available tick
>>> >>> 18/s; on linux that number is 1024/s. I suggest that you try some
>>> >>> longer-running verbs if you expect a fair comparison.
>>> >>>
>>> >>> On Tue, Jan 14, 2014 at 6:11 AM, Ben Gorte - CITG
>>> >>> <b.g.h.go...@tudelft.nl> wrote:
>>> >>>> Speaking about J, performance and Linux, is it true that Windows is
>>> > significantly faster? Or is there something wrong with my installation?
>>> > Also when runnning windows J under wine on my linux PC I get a better
>>> > performance than with native linux J:
>>> >>>>
>>> >>>> NB. Native Linux
>>> >>>> JVERSION
>>> >>>> Engine: j701/2011-01-10/11:25
>>> >>>> Library: 7.01.087
>>> >>>> Platform: Linux 32
>>> >>>> Installer: j701a_linux32.sh
>>> >>>> InstallPath: /home/ben/j701
>>> >>>> time'locs=:nudge"1 locs'
>>> >>>> 1.43086e_5
>>> >>>> time'locs=:pfn"1 locs'
>>> >>>> 7.41384e_6
>>> >>>> time'locs=:(pfn f.)"1 locs'
>>> >>>> 3.77003e_6
>>> >>>> time'locs=:pfns"1 locs'
>>> >>>> 3.7135e_5
>>> >>>>
>>> >>>> NB. wine + Windows J
>>> >>>> JVERSION
>>> >>>> Engine: j701/2011-01-10/11:25
>>> >>>> Library: 7.01.040
>>> >>>> Platform: Win 32
>>> >>>> Installer: j701a_win.exe
>>> >>>> InstallPath: z:/media/windows/documents and settings/bgorte/j701
>>> >>>> time'locs=:nudge"1 locs'
>>> >>>> 1.09025e_5
>>> >>>> time'locs=:pfn"1 locs'
>>> >>>> 5.56416e_6
>>> >>>> time'locs=:(pfn f.)"1 locs'
>>> >>>> 2.88706e_6
>>> >>>> time'locs=:pfns"1 locs'
>>> >>>> 2.77585e_5
>>> >>>>
>>> >>>> Regards,
>>> >>>> Ben
>>> >>>> ________________________________________
>>> >>>> From: programming-boun...@forums.jsoftware.com [
>>> > programming-boun...@forums.jsoftware.com] on behalf of Raul Miller [
>>> > rauldmil...@gmail.com]
>>> >>>> Sent: Monday, January 13, 2014 06:44
>>> >>>> To: Programming forum
>>> >>>> Subject: Re: [Jprogramming] On benchmarking results from J
>>> > programming styles
>>> >>>>
>>> >>>> That sounds about right.
>>> >>>>
>>> >>>> The big caution I would place on interpreting these results is:
>>> "This
>>> >>>> won't necessarily apply for games implemented in J for Linux, where
>>> I
>>> >>>> intend to rely on the SDL and byte-per-pixel graphics layouts.
>>> >>>> Nonetheless, I retain the logic here, since it's representative of a
>>> >>>> real-world design decision which directly influences performance on
>>> >>>> the slower Kestrel architecture."
>>> >>>>
>>> >>>> If J is to perform well when applied in suboptimal fashion we'll
>>> need
>>> >>>> some way of representing the code which strips out a lot of the
>>> >>>> functionality (type checks, size checks, rank handling, maybe even
>>> >>>> overflow handling?), at least for the time-critical routines. (As
>>> much
>>> >>>> as possible, hoisting redundant operations out of primitives used in
>>> >>>> bottleneck loops.)
>>> >>>>
>>> >>>> Thanks,
>>> >>>>
>>> >>>> --
>>> >>>> Raul
>>> >>>>
>>> >>>>
>>> >>>> On Sun, Jan 12, 2014 at 11:59 PM, William Tanksley, Jr
>>> >>>> <wtanksle...@gmail.com> wrote:
>>> >>>>> A friend of mine wrote the following paper describing his attempt
>>> to
>>> >>>>> characterize the differences between a few different styles of
>>> >>>>> implementing the same code in J a few different ways -- explicit,
>>> >>>>> implicit, and a few variations. He also baselined against a Forth
>>> >>>>> implementation.
>>> >>>>>
>>> >>>>> I found his writeup very interesting. What do you think?
>>> >>>>>
>>> >>>>> http://sam-falvo.github.io/2014/01/05/subroutine-performance-in-j/
>>> >>>>>
>>> >>>>> -Wm
>>> >>>>>
>>> > ----------------------------------------------------------------------
>>> >>>>> 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
>>> >>>
>>> >>>
>>> >>>
>>> >>> --
>>> >>> - michael dykman
>>> >>> - mdyk...@gmail.com
>>> >>>
>>> >>> May the Source be with you.
>>> >>>
>>> ----------------------------------------------------------------------
>>> >>> For information about J forums see
>>> http://www.jsoftware.com/forums.htm
>>> >>>
>>> ----------------------------------------------------------------------
>>> >>> For information about J forums see
>>> http://www.jsoftware.com/forums.htm
>>> >>
>>> >> --
>>> >> regards,
>>> >> ====================================================
>>> >> GPG key 1024D/4434BAB3 2008-08-24
>>> >> gpg --keyserver subkeys.pgp.net --recv-keys 4434BAB3
>>> >> gpg --keyserver subkeys.pgp.net --armor --export 4434BAB3
>>> >> ----------------------------------------------------------------------
>>> >> 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
