...made just for fun :)
Ten compilations of povray. Two scenes.

Test files:
 benchmark.pov (from tarball)
 http://www.irtc.org/ftp/pub/stills/2003-10-31/forget.zip
 (using default benchmark.ini file for both)

System :
 Fedora Core 5
 Linux shodan 2.6.17.11 #1 Mon Sep 4 19:44:27 CEST 2006 i686 i686 i386 GNU/Linux
 1GB RAM
 povray started with 'nice -n -15'
 
Processor :
 vendor_id       : GenuineIntel
 cpu family      : 15
 model           : 2
 model name      : Intel(R) Pentium(R) 4 CPU 1.80GHz
 stepping        : 4
 cpu MHz         : 2259.249
 cache size      : 512 KB
 cpuid level     : 2
 bogomips        : 4522.53

Compilers:
 0-8: gcc version 4.1.1 20060525 (Red Hat 4.1.1-1)
   9: icpc (ICC) 9.1 20060816
   
Options:
 0: guessed by autoconf
 1: -O3 -march=i386
 2: -O3 -march=i686
 3: -O3 -march=pentium4
 4: -O3 -march=pentium4 -msse2
 5: -O3 -march=pentium4 -msse2 -mfpmath=sse
 6: -O3 -march=pentium4 -msse2 -mfpmath=sse,387
 7: -O3 -march=pentium4 -msse2 -mfpmath=sse,387 -malign-double
 8: -O3 -march=pentium4 -msse2 -mfpmath=sse -malign-double  \
    -minline-all-stringops -ffast-math -fomit-frame-pointer \
    -funroll-loops -fexpensive-optimizations -foptimize-sibling-calls
 9: -O3 -axN -no-prec-div -march=pentium4 -mcpu=pentium4 -msse2 \
    -mtune=pentium4 -funroll-loops -fomit-frame-pointer -static


[benchmark.pov]
Results - povray stats:
 0: 56/484 (photon/render time)
 1: --/---
 2: 57/497
 3: 54/494
 4: 54/493
 5: 55/489
 6: 60/517
 7: 60/511
 8: 54/459
 9: 45/402
 
Results - system stats:
 0: 534.69 (user time)
 1: ---.--
 2: 550.68
 3: 544.73
 4: 542.62
 5: 540.86
 6: 574.55
 7: 567.52
 8: 508.61
 9: 439.43

[forget.pov]
Results - povray stats:
 0: 1546 (render time)
 1: ----
 2: 1576
 3: 1607
 4: 1607
 5: 1575
 6: 1643
 7: 1628
 8: 1489
 9: 1357

Results - system stats:
 0: 1531.61 (user time)
 1: ----.--
 2: 1566.20
 3: 1593.16
 4: 1592.38
 5: 1561.18
 6: 1628.48
 7: 1613.31
 8: 1462.09
 9: 1332.42

========================

Well... I think, next week I'll check changes between this all images :>

Slawek
-- 


 \__ \__ \_______________________________________________________________
 /___/___/ Slawomir Szczyrba                          steev/AT/hot\dot\pl

Post a reply to this message

> ....made just for fun :)

	... or for speed!
	(If only I had the time to write the article summarizing the bench results
for the *hundreds* of binaries I have prepared so far...)

>  0-8: gcc version 4.1.1 20060525 (Red Hat 4.1.1-1)
>    9: icpc (ICC) 9.1 20060816

	FWIW these are basically the compilers that are used to prepare the x86_64
and x86 versions of the Linux 3.7.0.beta.15, respectively  :-)

	- NC

Post a reply to this message

S?awomir Szczyrba <cre### [at] ofthenight> wrote:
>  8: -O3 -march=pentium4 -msse2 -mfpmath=sse -malign-double  \
>     -minline-all-stringops -ffast-math -fomit-frame-pointer \
>     -funroll-loops -fexpensive-optimizations -foptimize-sibling-calls

  According to my own tests some of those options actually slow down
POV-Ray instead of speeding it up. Adding a new optimization option does
not always mean that the binary will be faster.

  My own experiments show that this combination gives the fastest binary,
at least in a pentium4:

-O3 -march=pentium4 -ffast-math -mfpmath=sse -msse2 -funroll-loops
  -ftsp-ordering

-- 
                                                          - Warp

Post a reply to this message

Houston, we got a Warp...

>   My own experiments show that this combination gives the fastest binary,
> at least in a pentium4:
>
> -O3 -march=pentium4 -ffast-math -mfpmath=sse -msse2 -funroll-loops
>   -ftsp-ordering
>

Well, my compiler doesn't know this option :)

[steev]/tmp# gcc -ftsp-ordering hello.c
cc1: error: unrecognized command line option "-ftsp-ordering"

[steev]/tmp# icc -ftsp-ordering hello.c
icc: Command line warning: ignoring unknown option '-ftsp-ordering'


-- 
  ________ 

 \__ \__ \_______________________________________________________________

Post a reply to this message

> Well, my compiler doesn't know this option :)
> 
> [steev]/tmp# gcc -ftsp-ordering hello.c
> cc1: error: unrecognized command line option "-ftsp-ordering"

	That's because it's been removed in the latest gcc series (IIRC 3.4.x inclusive).

	- NC

Post a reply to this message

Nicolas Calimet <pov### [at] freefr> wrote:
> > Well, my compiler doesn't know this option :)
> > 
> > [steev]/tmp# gcc -ftsp-ordering hello.c
> > cc1: error: unrecognized command line option "-ftsp-ordering"

>         That's because it's been removed in the latest gcc series (IIRC 3.4.x
inclusive).

  It doesn't matter anyways. It only gave a very minimal speedup.

-- 
                                                          - Warp

Post a reply to this message

Speed... kills... ;)
Well, '-unroll-loops' kills. Images. :)

'Benchmark' results + images + differences between images...
< http://shodan.selfip.net/~steev/pov/ >


-- 
  ________ 
_/ __/ __/
 \__ \__ \_______________________________________________________________

Post a reply to this message

If I understand your results correctly, the left image is that produced
by the binary compiled with the specified options, the middle image is the
difference-image with a reference (presumably calculated with the official
binary for Linux?) and the right image is the inverted difference.  Correct?

> Well, '-unroll-loops' kills. Images. :)

	From you results it seems that it's actually -ffast-math that changes the
result of the top-left corner (your last test shows that -funroll-loops does
produce a correct image).  But what is really strange is your second result:
a distorted image is also produced with -march=i686 ???  Hard to believe...

	In any case, on an Intel Pentium 4 or higher processor, it is clear again
that one should use ICC to compile POV from sources.

	- NC

Post a reply to this message

S?awomir Szczyrba wrote:
> BOFH excuse 72: S?awomir Szczyrba did it
> 
> Speed... kills... ;)
> Well, '-unroll-loops' kills. Images. :)
> 
> 'Benchmark' results + images + differences between images...

It is perfectly expected that there are slight difference. The huge
difference is precision bug in your math library though. It should not occur
with a standard(s) conforming library.

	Thorsten

Post a reply to this message

S?awomir Szczyrba <cre### [at] ofthenight> wrote:
> 'Benchmark' results + images + differences between images...
> < http://shodan.selfip.net/~steev/pov/ >

  I think the base image is wrong.

  I just tried benchmark.pov in my system, with povray compiled with
-O3 -march=pentium4 -ffast-math -mfpmath=sse -msse2 -funroll-loops
and it rendered correctly.
  There must be something odd with either your version of gcc or your
system libraries or something.

-- 
                                                          - Warp

Post a reply to this message