gcl-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Gcl-devel] Bad news on Maxima and proclaim


From: Camm Maguire
Subject: Re: [Gcl-devel] Bad news on Maxima and proclaim
Date: 02 Aug 2004 17:50:29 -0400
User-agent: Gnus/5.09 (Gnus v5.9.0) Emacs/21.2

Greetings, and thank you very much for your report!

This is indeed a bug in gcl, which has now been fixed in CVS head.

I am leaning toward the conclusion that this is important enough to
warrant a 2.6.4.

May I suggest in the meantime that you manually remove TYIPEEK from
the sys-proclaim.lisp file.  If/when we get a 2.6.4 out, we can
restore it.

I broadly confirm your report of a 5-10% speedup with the proclaims on
make check.  This places GCL ahead of cmucl on my machine at least,
especially after gc time is removed.  Please keep in mind our earlier
discussion on the inadequacy of this test as a benchmark.  The
produced code with the proclaims is obviously superior from inspection
(not that it cannot still be further improved :-)).

Here are my make check timings:

GCL w/ proclaims
----------------
real time       :      3.220 secs
run-gbc time    :      1.940 secs
child run time  :      0.670 secs
gbc time        :      0.380 secs

GCL no proclaims
----------------
real time       :      3.510 secs
run-gbc time    :      2.260 secs
child run time  :      0.670 secs
gbc time        :      0.410 secs

CMUCL
-----
; Evaluation took:
;   3.49f0 seconds of real time
;   3.24f0 seconds of user run time
;   0.25f0 seconds of system run time
;   8,380,363,322 CPU cycles
;   [Run times include 0.17f0 seconds GC run time]
;   5815 page faults and
;   92,299,104 bytes consed.
; 

Take care,


James Amundson <address@hidden> writes:

> On Sun, 2004-07-25 at 18:07, James Amundson wrote:
> 
> > Having followed all your directions, I finally see what you were talking
> > about: GCL Maxima runs roughly twice as fast as it used to! I am very
> > impressed. For the first time, GCL is beating CMUCL in my timings. I
> > have committed the necessary changes to the build system.
> 
> Sadly, it appears that these initial reports of spectacular success were
> ill-founded. The mistaken impression that including proper proclaims
> sped up maxima by a factor of 2 came from a deficiency in the test
> suite. Here's what I now know:
> 
> 1) Maxima 5.9.0.9beta2 compiled with GCL 2.6.3 has a bug. The command 
>       trigrat(sin(3*a)/sin(a+%pi/3));
> unexpectedly waits for user input. We have a report that this behavior
> is not seen in GCL 2.7.0. It seems that this must be a bug in 2.6.3.
> 
> 2) The bug goes away if the reference to MAXIMA::TYIPEEK is removed from
> sys-proclaim.lisp.
> 
> 3) The apparent large speed increase with the proclaims was an artifact
> of a flaw in the maxima test suite. When the test suite is run via "make
> check", it quietly stops running further tests when it hits the above
> command and its spurious request for user input. The decrease in time
> was primarily due to the large number of omitted tests!
> 
> I have now tried comparing
> 
>   a) Maxima 5.9.0.9beta2 with MAXIMA::TYIPEEK removed from
> sys-proclaims.lisp
> 
>   b) Maxima 5.9.0.9beta2 with an empty sys-proclaims.lisp.
> 
> After five runs, I conclude that the former is (6 +/- 3)% faster than
> the latter on my machine.
> 
> --Jim
> 
> 
> 
> _______________________________________________
> Gcl-devel mailing list
> address@hidden
> http://lists.gnu.org/mailman/listinfo/gcl-devel
> 
> 
> 

-- 
Camm Maguire                                            address@hidden
==========================================================================
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah




reply via email to

[Prev in Thread] Current Thread [Next in Thread]