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: TDA4VM: How to remove padding on the final output data layer in TIDL

Part Number: TDA4VM

Hi,

I used tidl_j7_01_03_00_11 in psdk7.1.

I was wondering why the final output data layers always have zero paddings in TIDL.

Actually, I want to final output data without padding in TIDL. But It's so hard to remove the padding in importing tool because I don't work around the perfsim.

In the tidlRunPerfSimTool function when a network is importing, some information of tidlNet->TIDLLayers.outData are from perSimInfo->sdataFlowInfo[i].bufInfo.

Could you some give some guide to remove the padding on out-data layers?

Best regards

Yongsig.

  • Hi Yongsig,

        The padding in the output buffer depends on the layer which is producing it. There are certain layers in TIDL which can change padding and there are others which cannot. So if padding cannot be change by the layer then the padding remain same as its input padding. 

    Regards,

    Anshu

  • Hi, Anshu.

    Thank you for your reply.

    The final layer(before data layer) is slice layer. I use slice layer to only bypass data.

    The Input data(tensor) has z-padding in slice layer.

    It seems like the padding remains the same as its input padding in slice layer.

    Could you recommend a layer to bypass data without padding?

    If there is a layer for bypass without padding, let me know how to describe it on prototxt.

    Best regards

    Yongsig.

  • Hi Yongsig,

        

    The final layer(before data layer) is slice layer. I use slice layer to only bypass data

          I didn't follow this, why do you have slice layer before the output? 

    Regards,

    Anshu

  • Hi, Anshu

    I had a problem it didn't make a data layer after our custom layer or conv layer when a network was imported.

    I didn't know the reason. The problem doesn't occur at all networks. I think the problem occurs if the network has a lot of outputs.

    So, I added a slice layer after our custom layer, that problem was solved.

    Best regards

    Yonsig

  • Hi Yonsig,

       First of apologies for delay in response ( we got occupied with SDK 8.0 release). We will look into why the data layer is not coming with custom layer. Can you also help us understand why you do not want padding in the output?

    Regards,

    Anshu

  • If the output has z-padding, we should remove that padding to enhance the time for copying the data in post-processing.

    Best regards

    Yongsig.

  • Hi Yongsig,

        Thanks for explaining the reason. One follow-up question, why do you need to do this copy? Can the post-processing here not directly work on the buffer with padding?


    Regards,

    Anshu