• State Resolved
  • Locked Locked
  • Replies 1 reply
  • Subscribers 99 subscribers
  • Views 275 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.

TDA4VM: Slow execution of convolutional node

PR7
Prodigy 160 points

Part Number: TDA4VM

Hello,

I encountered an  issue while executing an ONNX model on the TDA4 board.

The problem is related to a 2D convolution operation that runs very slowly, around 130 ms, while it should take just few microseconds according to the simulator (see attached for simulation output)conv2D.zip  

The operation has the following parameters:

-        Input shape NCHW: 1x128x8x24

-        1x1 kernel shape with stride (2,2)

 

I attached a simple ONNX model which contains only this convolutional node in order to reproduce the issue.

Do you have any clue about that? May it be due to memory limitations of C7x DSP or due to a combination of parameters that is preferable to avoid?

 

Thanks,

Regards

  • 0
    TI__Expert 6335 points

    Hi Pierre,

    This issue will be fixed in the next release of TI-DL, which will be part of PSDKA 6.2. The expected release timeframe is in 2 weeks.

    You will get the following execution cycles:

    # 0 . ..
    Network Cycles 155433
    Layer, Layer Cycles,kernelOnlyCycles, coreLoopCycles,LayerSetupCycles,dmaPipeupCycles, dmaPipeDownCycles, PrefetchCycles,copyKerCoeffCycles,LayerDeinitCycles,LastBlockCycles, paddingTrigger, paddingWait,LayerWithoutPad,LayerHandleCopy, BackupCycles, RestoreCycles,
    1, 98280, 67284, 67689, 9599, 6241, 2704, 21, 18, 714, 67284, 633, 223, 93296, 5565, 0, 0,\

    So around 0.000155 ms.

    regards,

    Victor