***BOGO*** Re: [Mingw-users] GCC 4.3.0 20080502 (alpha) Released - Please help us test it!

[Stéphane Larouche <stephane.larouche@xxxxxxxxxx>]

Aaron W. LaFramboise <aaron77thyme@...> writes:
> Stéphane Larouche wrote:
> 
> > they are extremely slow. 
> ...
> > The DLLs contain a lot of mathematic functions (mainly trigonometric)
> ...
> > Is this a known problem? Would it be helpful if I tried to identify
 what part of
> > the code are so slow? Do you have any suggestion of how to do it?
> 
> No, this is not a known problem, and in fact, little performance testing 
> has been done so far on MinGW GCC specifically.
> 
> Yes, it would help if you could give us an idea of what part of the code 
> specifically causes this problem.
> 
> It would also be helpful if we knew what optimization flags you are 
> using, and what effect options like -ffast-math had on this speed 
> difference.

I use -O3 optimization flag for my project. -ffast-math is not
really an option since it does not follow IEEE standard. The
additional level of verification needed to avoid problems such
as division by zero would most probably overcome any speed
gain.

I have determined the performance of GCC 3.4.5, 4.2.1
technology preview and 4.3.0 alpha, with various optimization
flags. In all cases, I have used MinGW 5.1.4 typical
installation, except for the version of GCC. The results
include the size of the 2 DLLs and the time it takes to run a
typical operation of OpenFilters, a software of mine:

GCC 3.4.5

(no flag): (321+105)kB, 71s
-O1: (224+89)kB, 51s
-O2: (219+92)kB, 50s
-O3: (219+92)kB, 48s
-ffast-math: (320+105)kB, 69s

GCC 4.2.1-dw2-2

(no flag): (448+214)kB, 92s
-O1: (328+203)kB, 37s
-O2: (325+204)kB, 37s
-O3: (327+205)kB, 36s
-ffast-math: (446+213)kB, 98s

GCC 4.2.1-sjlj-2

(no flag): (383+159)kB, 90s
-O1: (182+150)kB, 36s
-O2: (180+151)kB, 37s
-O3: (181+151)kB, 36s
-ffast-math: (382+159)kB, 98s

GCC 4.3.0 alpha

(no flag): (331+151)kB, 142s
-O1: (272+138)kB, 87s
-O2: (268+138)kB, 84s
-O3: (269+139)kB, 82s
-ffast-math: (326+150)kB, 115s

As you can see, the 4.3.0 alpha version is much slower that the
other ones. All three level of optimization have approximately
the same effect. -ffast-math does not provide a great increase;
in the case of 4.2.1, it even slows down the execution.

In addition to speed and size differences, I have noticed that
the results returned by the software are slightly different in
the different GCC versions. Is it possible that mathematical
functions do not handle floats/doubles correctly?

I will try to find some time during the weekend to generate a
minimal example that reproduces the problem. In the mean time,
those interested can take a look at the code used to generate
those results at
www.polymtl.ca/larfis/English_Version/OF/OF_English.htm.


Sincerely,

Stéphane



-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
_______________________________________________
MinGW-users mailing list
MinGW-users@xxxxxxxxxxxxxxxxxxxxx

You may change your MinGW Account Options or unsubscribe at:
https://lists.sourceforge.net/lists/listinfo/mingw-users