SPEC 2000 results on PowerPC

[Dale Johannesen]

On Aug 17, 2004, at 10:16 AM, Mark Mitchell wrote:

> I've now got SPEC 2000 results for PowerPC.
> I'm using a G5 and ran with "-O2 -mcpu=970".  My goal is not to 
> demonstrate how fast this machine is, but rather to compare the 
> current mainline with the last release.
>
> The current mainline is, I would say, somewhat better than 3.4.1.
> Three tests (176.gcc,187.facerec,189.lucas) work with 3.5, but did not 
> run at all with 3.4.1, so that is clearly an improvement.

176.gcc uses nonstandard aliasing, in a way that causes code to be
misscheduled sometimes.  So this one may just be luck.

> Two tests (186.crafty and 200.sixtrack) that did work with 3.4.1 no 
> longer work with 3.5 -- that is clearly a degradation.

186.crafty worked for me the last time I tried, although not with those 
exact options.

>  Most of the results are within 5% of the 3.4.1 results (nine better, 
> six worse).  The outliers are 252.eon (15% improved), 168.wupwise (10% 
> degraded), 177.mesa (7% improved), and 301.apsi (15% improved).
>
> My current plan is to look into fixing the tests that do not work at 
> all, and then see what's going on with 168.wupwise.

The hot code in wupwise is zgemm.f, loops ending 70 and 100, and 
zaxpy.f, loop 30.
Multiply-add insns are very important.