Adam Wiggins <[EMAIL PROTECTED]> writes:
> Oh, they're more or less the same - the difference is that Intel's chips
> have floating point which is comparable to their integer processing,
> whereas AMD's K6's beat the pants off of Intel for integer, but lagged
> way behind on floating point. 3DNow! was basically a quick fix to this
> problem.
You're probably right, but 3dnow is still capable of being faster than
standard instructions. Here is how the different combinations compare
in Mesa. The times are cycles to transform 100 points3_general
vertices. (The K6-2 times are from Holger.)
K6-2 without 3dnow: 11338
K6-2 with 3dnow: 1993
Plain old pentium: 4742
Mesa running q3test spends a little more than 5% of its time in
gl_x86_transform_points3_general on my Pentium (this may have changed
slightly - I haven't profiled it lately), so the 3dnow version is
probably around 2.5% faster if everything else is equal. SSE (I think
this is the current name for KNI) should be similar.
Clipping looks like the biggest bottleneck for q3test on Mesa right
now.
Still, if anyone really wants SSE optimization in Mesa, send me a PIII
and I'll get right on it! (It really shouldn't be very hard to do.
The hardest part is the kernel modifications, but there is already a
patch available.)
Josh
_______________________________________________
Mesa-dev maillist - [EMAIL PROTECTED]
http://lists.mesa3d.org/mailman/listinfo/mesa-dev
