Hi,
The accuracy of the simulator from Hexagon 2.0 SDK is according to the simulator user guide:
The simulator is not cycle-accurate. With cache modeling enabled it is cycle-approximate, with current performance within 2% of cycle-accuracy (andprojected performance of less than 1% difference).
My question is, what type of errors is these 2%, random or systematic?
I reran a program multiple times with the simulator and the statistics from the simulator were identical between the runs.
A follow-up question is, without cache modeling, what is the accuracy of the simulator?
Thanks in advance
Best Regards
Martin Persson
Hi martin,-
I would think (!) the simulator would be quite deterministic - provided the test data also are the same. And why don't you upgrade the SDK to version 3.1 ?
Thinking about it - if the test data are of random nature then the cache may show quite large variations.
A related topic - is it possible to switch cache modeling on / off by any option?
BR,
Henrik
Hi, and thanks for your quick reply.
Yes, I would agree on that the simulator seems to be deterministic, which should indicate that there is some systematic error which results in the +-2% stated in the manual.
I use the simulator to simulate performance of code for the Hexagon v4 DSP, and I don't see any reason to upgrade. Is there any benefit with the simulator included in SDK 3.1?
For the simulator provided in SDK 2.0, you can use the --timing argument to simulate caches and processor stalls,
from the simulator user guide:
Best Regards
Martin