• State Resolved
  • Locked Locked
  • Replies 3 replies
  • Answers 1 answer
  • Subscribers 53 subscribers
  • Views 547 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Performance of C6670 cycle approximate simulator...

Mark Brown
Intellectual 360 points

Hi All,

   We have noticed that running a reasonably realistic program using one core of the C6670 quad core cycle approximate simulator is painfully slow! A competitor's simulator running the same level of detail (cycle accurate) runs 68 to 100 times faster for the same program. That is a simulation run that normally takes about eight minutes takes about 24 hours with the C6670 simulator. This is simply unacceptable! We cannot use the functional simulator since we require the EP trace facility. Our runtime requirements are simple. We use a single core with a flat memory with no cache and no peripherals. Is there any way that we can configure the model for this simple mode of operation to improve its performance?

Thanks,

Mark Brown

  • 0
    TI__Intellectual 1420 points

    Hi Mark,

    The C6670 cycle approximate(CX) simulator models accurate SoC including memory system and interconnect latencies. It is recommended to use this variant only if you are running an application for the purpose of accurate profiling, probably for tuning the application performance. If the use case is to run an application for functional verification, you can use 6670 device functional simulator. EP trace is available on both functional and CX simulator configurations. We do not have a CPU + memory single core configuration as part of the release.

     

    Having said that, moving to device functional simulator in itself will give close  more than 100X improvement in speed, compared to CX. Again there will be some speed degradation if EP trace is ON versus it is OFF, on both CX and functional simulator configurations, which is expected. With regard to EP trace data, since it is CPU execute packet related, the accuracy is the same as CX simulator, since we model the execute pipeline accurately for both functional/CX configurations.

    regards,

    Sheshadri

  • 0 in reply to Sheshadri Chakravarthy
    Intellectual 360 points

    Sheshadri,

       Thanks, as long as the cycle accuracy through EP trace is the same for the functional simulator this should work for us. I will give this a try and report back my findings.

    Thanks,

    Mark Brown

  • 0 in reply to Mark Brown
    Intellectual 360 points

    Sheshadri,

       I have verified that the functional-level simulator does in fact support EP trace and runs much faster. The resulting cycle counts are either the same as or within a fraction of a percent of the cycle approximate simulator.

    Thanks,

    Mark