[vlc-devel] memcpy

[Pierre Ynard]

I ran tests on the old MMX hardware you were talking about, under linux:

        block size 4147200
unaligned: libc  33.22% slower than vlc.
  aligned: libc  51.01% slower than vlc.
        block size 4147200
unaligned: libc  45.66% slower than vlc.
  aligned: libc  53.61% slower than vlc.
        block size 4147200
unaligned: libc  51.62% slower than vlc.
  aligned: libc  52.65% slower than vlc.

        block size 3840
unaligned: libc  43.96% slower than vlc.
  aligned: libc  44.28% slower than vlc.
        block size 3840
unaligned: libc  31.08% slower than vlc.
  aligned: libc  57.30% slower than vlc.
        block size 3840
unaligned: libc  31.91% slower than vlc.
  aligned: libc  42.50% slower than vlc.

        block size 1024
unaligned: libc 301.90% faster than vlc.
  aligned: libc 294.19% faster than vlc.
        block size 1024
unaligned: libc 289.43% faster than vlc.
  aligned: libc 319.45% faster than vlc.
        block size 1024
unaligned: libc 290.36% faster than vlc.
  aligned: libc 301.57% faster than vlc.


Now on 32 bits (Pentium M):

        block size 4147200
unaligned: libc  56.65% slower than vlc.
  aligned: libc  12.16% slower than vlc.
        block size 4147200
unaligned: libc  54.77% slower than vlc.
  aligned: libc  21.04% slower than vlc.
        block size 4147200
unaligned: libc  78.05% slower than vlc.
  aligned: libc  18.13% slower than vlc.

        block size 3840
unaligned: libc  75.20% faster than vlc.
  aligned: libc 539.64% faster than vlc.
        block size 3840
unaligned: libc  75.45% faster than vlc.
  aligned: libc 494.44% faster than vlc.
        block size 3840
unaligned: libc  70.81% faster than vlc.
  aligned: libc 602.57% faster than vlc.

        block size 1024
unaligned: libc  17.15% faster than vlc.
  aligned: libc 182.80% faster than vlc.
        block size 1024
unaligned: libc  20.19% faster than vlc.
  aligned: libc 196.78% faster than vlc.
        block size 1024
unaligned: libc   7.56% faster than vlc.
  aligned: libc 139.34% faster than vlc.

I'm all for removing old cruft but... How come glibc can't do as good?

Tests were ran on Debian unstable with libc6 and libc6-i686 2.13-31

CPUs:
model name      : Pentium MMX
model name      : Intel(R) Pentium(R) M processor 1.70GHz

-- 
Pierre Ynard
"Une âme dans un corps, c'est comme un dessin sur une feuille de papier."