Hi Zacharias, Here are the answers for your questions. 1) The formula is correct because the Idle counters count the time when the processor is in the Idle loop. These cycles are not a part of the benchmark and they do not contribute to anything useful, we don't include them in the UIPC calculation. 2) Can you give us more details about your simulation configuration? Which simulator and benchmark do you use? You can also send some of the debug outputs so that we can see how long you run and how much difference you have. 3) Yes (you listed 'User' counter twice, I assume you wanted to say 'Idle' for the second one). 4) I am not aware of a counter for the executed instructions. However, you might use branch predictor information to find the executed but not committed instructions.
Regards, Onur On Sep 6, 2012, at 2:02 PM, [email protected] wrote: > Hi, > > 1) In flexus 4.1 distibution the postprocess.sh calculates UIPC with this > formula User:Commits:Busy / (User:AccountedCycles + System:AccountedCycles). > Shouldn't it be user:commits:busy / (User:AccountedCycles + > System:AccountedCycles + Idle::AccountedCycles + Trap::AccountedCycles ) ? > > 2) I guess the sum of User:AccountedCycles, System:AccountedCycles, > Idle::AccountedCycles and Trap::AccountedCycles represents the total > execution cycles and it should be equal to flexpoints * > measured_region_cycles * cores. In all simulations I ran these numbers where > not equal (but close). > And I'm sure that all flex points ran correctly. Why is this happening? > > 3) Is the total number of instructions commited equal to User:Commits:Busy + > System:Commits:Busy + Trap:Commits:Busy + User:Commits:Busy ? > > 4) Does flexus have separate statistics for instructions committed and > executed? > > thanks -Zacharias Hadjilambrou
