Warp <war### [at] tagpovrayorg> wrote:
>   If you compile a generic 386 binary, but otherwise use the optimizations,
> is there any improvement over the official binary?
>
> --
>                                                           - Warp

I'll have to test that one more thoroughly later. I've compiled a few i386
binaries (with -march=i386) on my netbook running Ubuntu, but I'm not entirely
sure the compiler isn't also throwing in SSE, SSE2, SSE3, and SSSE3 instructions
in there.

Post a reply to this message

Ok, here are some more benchmarks, compiled in a 32-bit Ubuntu 10 environment,
then run on 64-bit Debian 6.0 environment on the Athlon II x4. I went into all
of the Makefiles and edited out any references to any other -march and
-msse/-msse2 statements.

gcc 4.4.5, -march=i386 -ffast-math -funroll-loops
  Parse Time:    0 hours  0 minutes  1 seconds (1 seconds)
  Photon Time:   0 hours  0 minutes 24 seconds (24 seconds)
  Render Time:   0 hours 18 minutes 57 seconds (1137 seconds)
  Total Time:    0 hours 19 minutes 22 seconds (1162 seconds)

gcc 4.4.5, -march=i386
  Parse Time:    0 hours  0 minutes  1 seconds (1 seconds)
  Photon Time:   0 hours  0 minutes 26 seconds (26 seconds)
  Render Time:   0 hours 20 minutes 44 seconds (1244 seconds)
  Total Time:    0 hours 21 minutes 11 seconds (1271 seconds)

Povray official binary
  Parse Time:    0 hours  0 minutes  1 seconds (1 seconds)
  Photon Time:   0 hours  0 minutes 28 seconds (28 seconds)
  Render Time:   0 hours 23 minutes  3 seconds (1383 seconds)
  Total Time:    0 hours 23 minutes 32 seconds (1412 seconds)

Post a reply to this message

jhu <nomail@nomail> wrote:
> gcc 4.4.5, -march=i386 -ffast-math -funroll-loops
>   Parse Time:    0 hours  0 minutes  1 seconds (1 seconds)
>   Photon Time:   0 hours  0 minutes 24 seconds (24 seconds)
>   Render Time:   0 hours 18 minutes 57 seconds (1137 seconds)
>   Total Time:    0 hours 19 minutes 22 seconds (1162 seconds)

> gcc 4.4.5, -march=i386
>   Parse Time:    0 hours  0 minutes  1 seconds (1 seconds)
>   Photon Time:   0 hours  0 minutes 26 seconds (26 seconds)
>   Render Time:   0 hours 20 minutes 44 seconds (1244 seconds)
>   Total Time:    0 hours 21 minutes 11 seconds (1271 seconds)

> Povray official binary
>   Parse Time:    0 hours  0 minutes  1 seconds (1 seconds)
>   Photon Time:   0 hours  0 minutes 28 seconds (28 seconds)
>   Render Time:   0 hours 23 minutes  3 seconds (1383 seconds)
>   Total Time:    0 hours 23 minutes 32 seconds (1412 seconds)

  Perhaps the official binary should be updated to a more optimized version.
(Perhaps the reason for the difference is that the official binary has been
compiled with an older version of gcc, and the newer versions optimize
better, even for the 386.)

  It could also be considered if the official binary distribution could
offer several binaries, compiled for different CPUs. For example, there
could be an official binary that works even on a 386, one that requires
at least an SSE-supporting Pentium, and a third that requires a x86-64 CPU.

-- 
                                                          - Warp

Post a reply to this message

Warp <war### [at] tagpovrayorg> wrote:

>   Perhaps the official binary should be updated to a more optimized version.
> (Perhaps the reason for the difference is that the official binary has been
> compiled with an older version of gcc, and the newer versions optimize
> better, even for the 386.)
>
>   It could also be considered if the official binary distribution could
> offer several binaries, compiled for different CPUs. For example, there
> could be an official binary that works even on a 386, one that requires
> at least an SSE-supporting Pentium, and a third that requires a x86-64 CPU.
>
> --
>                                                           - Warp

I was thinking the same thing about having different binaries for different
CPUs. OTOH, compiling 3.6 from source on Unix-like systems isn't really that
difficult. Perhaps we could also make a list of compiler options that produce
the fastest binary.

Post a reply to this message

Thorsten Froehlich <tho### [at] trfde> wrote:

> Did you rebuild boost with icc first? This looks like boost is configured
> for gcc, while the compiler is of course icc.
>
>  Thorsten

Actually, it turns out it's a bug in icc 11.1's gcc compatibility. Intel has
since updated it to remove the bug, but I haven't gotten around to installing it
yet.

Post a reply to this message