• State TI Thinks Resolved
  • Locked Locked
  • Replies 5 replies
  • Answers 2 answers
  • Subscribers 98 subscribers
  • Views 782 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.

RTOS/TDA2EXEVM: ssd mdoel in tidl have high latency

lin rimao
Intellectual 370 points
Part Number: TDA2EXEVM
Other Parts Discussed in Thread: TDA2

Tool/software: TI-RTOS

Hi:

Now, we use SSD model in TIDL to do object detection tasks. But the latency in one eve is about 190ms, how I can refine the code to improve the processing speed, in order to decrease the latency.

hope your advise.

  • 0
    TI__Genius 17580 points
    Hi,

    Yes, this 190 ms latency is expected for running SSD model on EVE, I do not think we can reduce it further by simply refining the code. I will confirm.

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    Intellectual 370 points
    Hi Praveen:

    thanks for your reply, I want to know when the model is fixed, what can I do to decrease the latency.

    hope your further answer.
  • 0 in reply to lin rimao
    TI__Mastermind 22145 points

    With a given network configuration, the current latency that you are observing is for one EVE. In TDA2 we have four EVEs. The latency can be reduced by partitioning the computation across all 4xEVEs. For, a dense model is it is recommended to split processing spatially (Feature/Activation in height wise). For sparse model also the same logic can be used. When the feature vector size is small, the overheads will be high for sparse flow. So when feature width /height is small, it is recommended to split the processing channels wise.

    The current VSDK demos are not using this partitioning, this can be done with current TIDL library.

    Thanks and Regards,

    Kumar.D

  • 0 in reply to kumar.desappan
    Intellectual 370 points
    Hi,

    I don't completely understand your idea. For example, we have trained the model with size of 640x320, in your opinion, i can use one eve to calculate the top half image 640x160, with another eve calculating the bottom half image 640x160?

    Is there any example I can refer?


    thanks.
  • 0 in reply to lin rimao
    TI__Genius 17580 points
    Hi,

    Yes, your understanding is correct if feature sizes are large. For small feature sizes, split the channels instead of height and use one eve to process some channels, and another eve process remaining channels.

    We don't have any code to refer for this implementation.

    Thanks,
    Praveen