Thread-topic: Strange perfromance results on G4 and G5 FFT
Ian,
Thanks for the quick response. The data on different sizes supports your hypothesis :
Size G4 G5
128x128 0.2msec 0.2msec
256x256 1.0msec 0.4msec
512x512 13.4msec 16.6msec
1024x1024 106msec 64msec
So it looks like there is an threshold value of 256x256 beyond which both G4 and G5 overflow the cache and that except for the 512x512 case, the G5 is faster. The exception is presumably because the number of cache misses is greater on the G5 than G4.
Given that our source images are integers, are there altivec libraries for 8 or 16 bit integer fft's ? Any other suggestions for improving performance on larger 2d data sets ?
-----Original Message-----
From: perfoptimization-dev-bounces+chris.bevis=email@hidden [mailto:perfoptimization-dev-bounces+chris.bevis=email@hidden] On Behalf Of Ian Ollmann
Sent: Wednesday, October 13, 2004 12:41 PM
To: Performance optimization list
Subject: Re: Strange perfromance results on G4 and G5 FFT
On Oct 13, 2004, at 12:23 PM, Bevis, Chris wrote:
> I am benchmarking accelerated FFT's ( 512x512 real, using fft2d_zrip
> in vDSP ) and am getting some wierd results :
>
> 1Ghz single G4 ( codegen and instruction scheduling set for G4
> ) 13msec
> 1Ghz single G4 ( codegen and instruction scheduling set for G5
> ) 12msec
> 2Ghz dual G5 ( codegen and instruction scheduling set for G5
> ) 17msec
>
> When I launch two instances simultaneously on the dual G5, THey both
> finish in 24msec.
>
> Does anyone have any idea why a 2x faster dual G5 performs slower than
> a G4 for a single instance and about the same for two instances ?
> This one has me puzzled.
Hi Chris,
A 512x512 array is about 1MB. The G5 has a 512 kB L2 cache and no
L3. A G4 might have up to 2 MB L3 cache. It seems possible you are
falling out of cache on the G5 and not on the G4. Does the trend go for
all sizes or just that size?
Ian
_______________________________________________
Do not post admin requests to the list. They will be ignored.
PerfOptimization-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/perfoptimization-dev/email@hidden
This email sent to email@hidden
_______________________________________________
Do not post admin requests to the list. They will be ignored.
PerfOptimization-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/perfoptimization-dev/email@hidden
This email sent to email@hidden
