TDA4VH-Q1: SDK11.00

hongyao.jin

Part Number: TDA4VH-Q1

Tool/software:

Hello,

We are using model import tool to gererate board model files(SimpleBev_OD_io.bin and SimpleBev_OD_net.bin).

For the same onnx model (including the gridsample layer, as shown in following figure), when we export an 8-bit model, both the emulation output and the board-end output results are normal. However, when we export a 16-bit model, the board-end output result is incorrect. As shown in the following figure:

Could you help check where could possibly be wrong? (SDK version 11.00)

All relevant data and images are placed in a compressed package, including the tidl_config file, the board-end model, the onxx model, the output results of the board-end model and the comparison results with the onnx output, the model emulation output results and the comparison results with the onnx output.

THANKS

0630 128x128 8bit extracted.zip 0630 128x128 16bit extracted.zip

2 months ago

0 Chris Tsongas 2 months ago

TI__Genius 14420 points

HI,

I will assign this to Christina, but I think their may be an issue with GridSample.

Thanks,

Chris

0 hongyao.jin 2 months ago in reply to Chris Tsongas

Expert 1220 points

Thanks for the confirmation.

+1 Christina Kuruvilla 2 months ago in reply to hongyao.jin

TI__Expert 4680 points

Hi Hongyao,

There is a new prerelease 11_00_08_00 that was specifically been released for fixes for the GridSample issues. Could you try it out PC emulation on that and let me know if you have any issues?

github.com/.../releases

Warm regards,

Christina

0 Christina Kuruvilla 2 months ago in reply to Christina Kuruvilla

TI__Expert 4680 points

Hi Hongyao,

I was trying to run your import config and am running into an error related to reading the input traces (.y files). How did you read the traces as inputs? InFileFormat = 2 reads a text file list of compressed images.

Also, the files you gave me has two different model names for the model used. Are these models different models?

Warm regards,

Christina

0 hongyao.jin 2 months ago in reply to Christina Kuruvilla

Expert 1220 points

The input_data.txt is in the compressed package, the content is as shown below:

../../model_import/simplebev_od/Model/input/Original_File/C7x_1_tidl_trace_0200_0001_0001_00001_00064_00024x00018.y ../../model_import/simplebev_od/Model/input/Original_File/C7x_1_tidl_trace_0018_0001_0001_00001_00001_00002x32768.y

They are all the same model, just with different names.

0630 128x128 8bit extracted.zip: This compressed package contains the correct results. We imported it with 8 bits.

0630 128x128 16bit extracted.zip: This compressed package contains incorrect results.We imported it with 16 bits.

+1 hongyao.jin 2 months ago in reply to Christina Kuruvilla

Expert 1220 points

We tried c7x_mma_tidl in 11.00.08, and this version of the model import tool solved this problem.

Processors

Processors forum

TDA4VH-Q1: SDK11.00