• State Resolved
  • Locked Locked
  • Replies 5 replies
  • Answers 2 answers
  • Subscribers 97 subscribers
  • Views 524 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.

TDA2PXEVM: How can I test the processing time of each layer on TDA

lluo
Intellectual 340 points
Part Number: TDA2PXEVM

Hi, 

I am running my OD net on TDA2P and the output is fine. And it shows the fps of the whole CNN model running on EVE.

Now, I want to analysis and test each layer's running time. 

SO, how can I get each layer's processing time?

  • 0
    TI__Genius 17580 points

    Hi,

    To get each layer's running time, you need to rebuild TIDL with enabling profile in the library which needs source code access of TIDL.

    But, the issue here is that performance is worse because of your are running all the layers on EVE core, as detection output layer is not optimized on EVE core, so you can see that more cycles consumed on EVE.

    Could you please run this last detection output layer on DSP and all the remaining layers on EVE to get better performance. For that please refer to FAQ 21 and 22 in the TIDL user guide for the import and inference steps to partition layers on EVE and DSP.

    Thanks,

    Praveen 

  • 0 in reply to Praveen Eppa1
    Intellectual 340 points

    Hi, Praveen

    I  have followed the FAQ 21 and 22 in user guide to set layerID and run detection output layer on DSP.

    My point is to test different convolution layer to see how does structure of convolution affect the performance.

    For example, testing whether setting group convolution is helpful for reducing running time, testing whether Nx1 + 1XN conv run faster than NXN conv, testing how much does sparse convolution help on different convolution layers. Those tests can help us a lot to run CNN on TIDL more efficiently.

    I am afraid the user guide is not enough for applying.

    So, is that possible to get source code to rebuild TIDL? 

    Or, is there any testing tools for cnn testing?

    Thanks!

  • 0 in reply to lluo
    TI__Genius 17580 points

    Hi,

    Please refer to section 1.2 Performance Summery in the TIDL data sheet where we had given "Convolution Layer Performance" for different combination of kenel sizes and sparse, dense kernels, so you can refer to that table and based on those cycles you can estimate for your network.

    Thanks,

    Praveen

  • 0 in reply to Praveen Eppa1
    Intellectual 340 points

    Hi, Praveen

    Thanks for your advice. 

    I don't have TIDL data sheet. Could you share it to me? I think it can help me a lot with estimating my network.

    Your replay is really helpful.

    Thanks again!

  • 0 in reply to lluo
    TI__Genius 17580 points

    Hi,

    This TIDL data sheet (TIDeepLearningLibrary_DataSheet.pdf) also included along with TIDL user guide in the "REL.TIDL.01.01.03.00\modules\ti_dl\docs" folder.

    Also, attached here for your reference.

    Thanks,

    Praveen2577.TIDeepLearningLibrary_DataSheet.pdf