Hi!

   I have had the chance to compile and run POV3.6 on various platforms
  using Gnu C/C++ Compiler and Intel C/C++ Compiler.

  I have run tests on these Linux-Machines (further specs see below):
  * INTEL Itanium, 0.8 GHz,
  * AMD Athlon TB, 0.8 GHz,
  * AMD Opteron 244, 1.8 GHz,
  * INTEL Pentium IV Xeon, 2.4 GHz

  Of course, it's not particularly fair to compare a 1.8GHz Opteron machine
  with a 0.8 GHz Athlon --keep that in mind, please!


Benchmark for var. archs and compilers:

using POV-Rays official benchmark.pov script
with suggested options -w384 -h384 +a0.3 +v -d -f -x

numbers = processor time [seconds]
LOWER NUMBERS = BETTER

                ICC8.0   GCC3.x.x
---------------------------------
ITANIUM (64)   10880     8715
---------------------------------
OPTERON (64)     n.a.    1580
---------------------------------
XEON    (32)    2458     3757
---------------------------------
ATHLON  (32)     n.a.    5012
---------------------------------
(64=64bit arch, 32=32bit arch)

* AMD Opteron 244, 1.8 GHz, 4 GByte RAM
   (dual processor boards in cluster)
   RedHat Linux
* INTEL Pentium IV Xeon, 2.4 GHz, 2 GByte RAM
   (dual processor boards in cluster)
   RedHat Linux
* INTEL Itanium, 0.8 GHz, 2 GByte Ram
   (dual processor board)
   Debian Linux
* AMD Athlon TB, 0.8 GHz, 256 kB Ram
   (single processor board)
   Debian Linux



  Obviously, you won't make use of an Itanium machine
because of it's speed ;-) I was wondering why the Intel Compiler
produced a code that performed ~20% slower than the one produced
by GCC, so I randomly chose some of POVs example-scene files for
render with both codes, and here is what I got:


SCRIPT               ICC8.0     GCC3.x
--------------------------------------
box.pov                5.87       5.80
mesh2.pov             45.14      40.06
glassthing.pov       299.09     320.25
parallel_lights.pov   19.37      21.69
circular.pov         143.74     175.14
shadows.pov           67.86      89.82
fog_ft.pov            10.30       9.21
atten2.pov            30.70      34.06
caustic2.pov          22.60      19.50
skysph2.pov           32.85      21.26
radiosity2.pov       187.19     207.37
cornell.pov           71.10      81.8
shear.pov              0.03       0.03
perspective.pov        0.02       0.03
panoramic.pov          0.03       0.03
focalblur.pov          0.03       0.03
abyss.pov            486.54     536.31
gaussianblob.pov     213.71     230.85
--------------------------------------
MEAN                  91.00      99.70


  So here is what I expected. Overall-performance codes produced by
the Intel-Compiler is ~10% better than GCC-Codes. It seems, that calculation
of simple objects with no further options like fog etc perform better with
GCC-POV Code. However, scripts like 'circular.pov', 'abyss.pov' or
'shadows.pov' show drastical improovement with ICC-POV-Code.

  I run the same scripts on the Xeon Machine and noticed an overall
improovement of approx. 12% using ICC (ICC: 28.9s, GCC: 33.0s).

  Comments and suggestions welcome...

Sincerely

S. Tayefeh

Post a reply to this message