Finally I found some time to play with final version of Open Watcom 1.0 released
recently. After some tweaking I got working exacutables for win32 console
version of MegaPOV. Already impelented 'working free' MegaPOV specific changes
caused very clean compilation - only a few dedicated lines in config.h. After
that I compiled MegaPOV with 5 free-like compilers I have available for Win32
platform and made quick test to compare speed. Test is not very accurate since
my time is limited but I'm sure it can be interesting for others. All
compilations are made for Pentium Pro (usually heighiest available) with
optimization for speed (libpng, libtiff, libjpeg, zlib libraries always
recompiled with the same settings as core). Everything was done on one machine
which is Pentium II 233 MHz with 128 MB RAM with NT4 Sp6 installed. No other
processes. Everything started from one batch and glowlits.pov from MegaPOV
distribution was used. No binary difference between outputs was obtained. Here
are results:

DigitalMarc C++ 8.33.4 - 339 seconds (windows console)
MinGW 2.0.0-3 gcc 3.2  - 310 seconds (windows console)
Borland C++ 5.5        - 277 seconds (windows gui)
DJGPP 2.3 (gcc 3.2)    - 202 seconds (dos application)
Open Watcom 1.0        - 194 seconds (windows console)

That was quick test with MegaPOV specific scene. I will try 3.5 scene with all
mentioned compilations and official binaries but I need more free time for that.

ABX

Post a reply to this message

ABX wrote:
> 
> [...]
> 
> DigitalMarc C++ 8.33.4 - 339 seconds (windows console)
> MinGW 2.0.0-3 gcc 3.2  - 310 seconds (windows console)
> Borland C++ 5.5        - 277 seconds (windows gui)
> DJGPP 2.3 (gcc 3.2)    - 202 seconds (dos application)
> Open Watcom 1.0        - 194 seconds (windows console)
> 
> That was quick test with MegaPOV specific scene. I will try 3.5 scene with all
> mentioned compilations and official binaries but I need more free time for that.

It looks interesting but i have some doubts whether these numbers are
representative.  My own tests some time ago showed that the MinGW version
is significantly faster than the Borland C++ version.  What were the
optimizations used for GCC?

BTW the support libraries should not matter much when you render without
file output.

Christoph

-- 
POV-Ray tutorials, include files, Sim-POV,
HCR-Edit and more: http://www.tu-bs.de/~y0013390/
Last updated 31 Dec. 2002 _____./\/^>_*_<^\/\.______

Post a reply to this message

On Thu, 27 Feb 2003 16:32:08 +0100, Christoph Hormann <chr### [at] gmxde>
wrote:
> It looks interesting but i have some doubts whether these numbers are
> representative.  My own tests some time ago showed that the MinGW version
> is significantly faster than the Borland C++ version.  What were the
> optimizations used for GCC?

BORLAND:
  CFLAGS=-6 -ff -H- -O2 -Oi -OS -Ov -P -q -V -vi -w
MINGW & DJGPP:
  CFLAGS= -march=pentiumpro -O3 -pedantic
DIGITAL MARS:
  CFLAGS= -o -v0 -cpp -mn -6 -ff -WA -Ae
WATCOM:
  CFLAGS= /oneatx /oh /oi+ /ei /zp8 /6 /fp6 /zq /xs /bt=nt 

I did not played much with settings - mostly used those suggested in docs.
Perhaps I can inspire somebody for further playing. I will be happy to find
propositions of other combinations.

Except Borland, development of all packages is open-like and perhaps some
improvements can be introduced into compilers. For example I would like to test
http://cmeerw.org/prog/dm/rtl.html solution. I have no idea how this can
influence.

> BTW the support libraries should not matter much when you render without
> file output.

Since image io operations are part of typicall using I decided to use it. But of
course for longer renderings its influence is less important.

ABX

Post a reply to this message

"ABX" <abx### [at] abxartpl> wrote:
> Christoph Hormann <chr### [at] gmxde> wrote:
> > It looks interesting but i have some doubts whether these numbers are
> > representative.  My own tests some time ago showed that the MinGW
version
> > is significantly faster than the Borland C++ version.  What were the
> > optimizations used for GCC?
>
> <...>
> MINGW & DJGPP:
>   CFLAGS= -march=pentiumpro -O3 -pedantic
> <...>

From this list, at least one quite important thing is
missing: -fomit-frame-pointer. There are a few other things to tweak, but
the impact of not generating EBP-based stack frames outweighs all the rest.

Post a reply to this message

Vadim Sytnikov wrote:

> "ABX" <abx### [at] abxartpl> wrote:
>> Christoph Hormann <chr### [at] gmxde> wrote:
>> > It looks interesting but i have some doubts whether these numbers are
>> > representative.  My own tests some time ago showed that the MinGW
> version
>> > is significantly faster than the Borland C++ version.  What were the
>> > optimizations used for GCC?
>>
>> <...>
>> MINGW & DJGPP:
>>   CFLAGS= -march=pentiumpro -O3 -pedantic
>> <...>
> 
> From this list, at least one quite important thing is
> missing: -fomit-frame-pointer. There are a few other things to tweak, but
> the impact of not generating EBP-based stack frames outweighs all the
> rest.

I'm missing the -ffast-math flag which should have quite some impact as 
well.

-- 
POV-Ray Objects Collection: http://objects.povworld.org

Post a reply to this message

On Fri, 28 Feb 2003 14:47:21 +0100, Micha Riser <mri### [at] gmxnet> wrote:
> > From this list, at least one quite important thing is
> > missing: -fomit-frame-pointer. There are a few other things to tweak, but
> > the impact of not generating EBP-based stack frames outweighs all the
> > rest.
>
> I'm missing the -ffast-math flag which should have quite some impact as 
> well.

Ok. I added -fommit-frame-pointer, droped file output and tried again with:

megapov.exe +Iglowlits.pov +GAoutput.txt +W640 +H480 -A -J0.0 -FN -D

  dmc:     329
  mingw:   299
  borland: 264
  djgpp:   204
  watcom:  185

then I added -ffast-math to mingw and djgpp:

  dmc:     329
  borland: 264
  mingw:   256
  watcom:  185
  djgpp:   169

what are next suggestions ?

ABX

Post a reply to this message

ABX wrote:
> 
> [...]
> 
> what are next suggestions ?

- use a slower scene for more accurate results (benchmark.pov for example)
- try the official WinPOV and the MSVC megapov for comparison

Christoph

-- 
POV-Ray tutorials, include files, Sim-POV,
HCR-Edit and more: http://www.tu-bs.de/~y0013390/
Last updated 28 Feb. 2003 _____./\/^>_*_<^\/\.______

Post a reply to this message

ABX wrote:

> 
> what are next suggestions ?

* Memory aligning: -malign-double
* Loop unrolling: -funroll-loops (might get worse though), 
-funroll-all-loops (likely will get worse)
* Use of profiling information.

Flags are for gcc, other compilers might have similar flags for it or 
already enabled by default.

- Micha

-- 
POV-Ray Objects Collection: http://objects.povworld.org

Post a reply to this message

In article <18qu5v0b9uf1amnnpum810vr8upmr5qgh1@4ax.com>,
 ABX <abx### [at] abxartpl> wrote:

> megapov.exe +Iglowlits.pov +GAoutput.txt +W640 +H480 -A -J0.0 -FN -D

glowlits.pov is not a benchmark scene, and is nothing like typical 
scenes. I would suggest using benchmark.pov.

-- 
Christopher James Huff <cja### [at] earthlinknet>
http://home.earthlink.net/~cjameshuff/
POV-Ray TAG: chr### [at] tagpovrayorg
http://tag.povray.org/

Post a reply to this message

On Fri, 28 Feb 2003 16:06:35 -0500, Christopher James Huff
<cja### [at] earthlinknet> wrote:
> > megapov.exe +Iglowlits.pov +GAoutput.txt +W640 +H480 -A -J0.0 -FN -D
>
> glowlits.pov is not a benchmark scene, and is nothing like typical 
> scenes. I would suggest using benchmark.pov.

I was not concerned much on choosing the right sphere because I considered it as
begining for further tests with various conditions. The longer test with
benchmark.pov I leaved for weekend. Now I have results:

app.       | compiler    | GUI | P-II | PIII |
-----------+-------------+-----+------+------+
povray 3.5 | intel       | win | 2740 |  815 |
megapov    | djgpp(gcc)  | dos | 2865 |  852 |
megapov    | mingw(gcc)  | con | 2909 |  872 |
megapov    | open watcom | con | 4094 | 1234 |
megapov    | borland 5.5 | win | 4407 | 1332 |
megapov    | dmc         | con | 4907 | 1454 |

where P-II is P-II 233 MHz with

call:
  application.exe benchmark.ini +Oapp +GAoutput_app.txt -J0.0 +FN

benchmark.ini:
  -w96 -h96 +a0.3 +am1 +v -d -f -x
  +Ibenchmark.pov

P-II:
  Pentium 2 233 MHz, 128 MB, NT4 Sp 6

P-III:
  Pentium 3 730 MHz, 128 MB, Win2000

Seems time for checking profiliers.

While MinGW is in fact faster than Borland compilation it seems strange that DOS
application (DJGPP) is faster than console application. It is signifficant
difference but it can means a day for a long renderings. AFAIK they use were
build with the same GCC version and I applied the same config.h for both.

And BTW: I noticed all images produced in this test are significantly different.
I have to track differences....

ABX

Post a reply to this message