Archives Discussions

[ryta1203]

I have a question regarding the reported ALU:Fetch ratios.

For the black scholes model sample (and the other samples as well), there is no branching or control flow so the calculated ALU:Fetch ratio should be quite simple to obtain.

However, the SKA gives 440 ALU and 13 Fetch for an ALU:Fetch of 110.00.

Profiler gives 440 ALU and 13 Fetch for an ALU:Fetch of 33.85.

While the ALU and Fetch numbers are consistent, somehow the ALU:Fetch ratio numbers are NOT.

The given numbers are for 5870.

Any explanation??

It seems that if you take 440/13 = 33.85 but then why does the SKA report 110?

[ryta1203]

I'm going to assume that the SKA still takes into account the 4:1 ratio while the profiler does not, this seems to be the case for all the samples I have looked at.

[bpurnomo]

Yes.  The ALU:Fetch ratio's calculation is diffferent between SKA and ATI Stream Profiler.  For the profiler, the ALU:Fetch ratio is exactly the number of ALU instructions executed divided by Fetch instructions executed.  For SKA, the equation is a bit more complicated and we took into account the 4:1 ratio in the hardware.

[ryta1203]

Originally posted by: bpurnomo Yes.  The ALU:Fetch ratio's calculation is diffferent between SKA and ATI Stream Profiler.  For the profiler, the ALU:Fetch ratio is exactly the number of ALU instructions executed divided by Fetch instructions executed.  For SKA, the equation is a bit more complicated and we took into account the 4:1 ratio in the hardware.

So the profiler is much more accurate? Or the more "complicated" formula takes "other" things into account?