Hi everyone,

libfreevec 1.0.4 has been released a while ago, but it's only today that I've managed to finish with the benchmarks. Check the URL to see how a G4, a G5 and a MPC8610 compare to a Athlon X2 5000! :D

http://www.freevec.org/content/libfreev ... ks_updated

also, you can find a howto of how to LD_PRELOAD libfreevec 1.0.4 on your system!

http://www.freevec.org/content/howto_us ... ld_preload

First of all: thanks, it works great! I have been using libfreevec since 1.0.4 was released. I even created some rpm-s by taking the 1.0.3 spec file from openSUSE 11.0. It is available at ftp://spike.fa.gau.hu/pub/pmppc110/ppc/ ... .1.ppc.rpm I did not make proper benchmarks, but the system feels to be faster, and works stable.

Konstantinos on your page, the G5 has a much faster bus. 1/2 clock speed.

The memory bus != front side bus. The G5's memory bus is at 533Mhz (DDR2). As far as I know, there have been no G5s with a faster memory bus.

ah okay, sorry

Actually the CPC945 supported dual-channel DDR2 so, in effect, it could probably transfer data at 533MHz x2. However the situations where this can happen are limited; it is only a solution to the problem of having RAM which is far, far slower than the CPU speed (imagine a 533MHz memory bus on a 4GHz chip).

It's pretty much only good for streaming huge amounts of data in sequential order (reading less than 128-bits at a time, in non-sequential order spanning different cache lines is not going to work very well), and this is somewhat compounded by the G5 bus (and NetBurst on a Pentium 4 acts the same way) which is somewhat high latency.

The G4 excels at doing semi-random access to RAM, for instance 2-dimensional array lookups. The G5 really does not perform well here, and neither will a high-end dual-core chip from Intel or AMD.

This is probably why there is a marked difference in speeds on certain workloads.

if you need to perform some tests on our fresh new YDL Powerstation (thanks to IBM Linux Technology Center) please feel free to contact me for a shell.


greetz,

Actually, I'd love to run benchmarks on a Power6 if anyone has access to such a beast, I already have G5 benchmarks, but thanks for the offer anyway :)

For now, I'm focused on porting libfreevec to sans-AltiVec CPUs (like the MPC5200B) and afterwards to 64-bit. The rest will be just function additions.

just curious: why is there a separate library for this functions? wouldn't it be better to integrate this stuff into glibc?

Sure. Tell that to the glibc authors. In any case, I'll make a last attempt to do that and if they get over the fact that there IS actually someone that has managed to outperform their own precious functions, then it might just be possible to have these included into glibc. But I wouldn't be too optimistic.

So what then if they don't? Well then it's time for a new libc...

Is there a chance to see implemented your new libfreevec optimizations like this guy did for 4xx?
http://www.cygwin.com/ml/libc-ports/201 ... 00000.html

They are now available in (e)glibc from 2.13.

tia,
nello