TDA4VM: How to configure batch processing using edgeai-tidl-tools

Hi,

Yes, parallel batch processing is only supported for multi core devices. The recommendation is to use a compatible SoC such as AM69A to utilize this feature. 

More on batch processing: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_multi_c7x.md

Changes in 10_01_00_02 release: https://github.com/TexasInstruments/edgeai-tidl-tools/releases/tag/10_01_00_02

Thank you,

Fabiana

Hi Fabiana,

Thank you for your answer. I am aware about the fact that the parallel batch processing is only supported for multi-core devices. However, my question was about whether or not batch processing (not necessarily parallel) is supported for single-core devices. In older PSDK versions, it was supported even for single-core platforms and the model was adjusted to supported during import process. So is it still supported. I assume so, since I was able to import the network. Only issue I was facing is that DataConvert layer seems to not work correctly for multiple batches. Can you confirm this is the issue and how to possibly address it?

Thank you,

Mladen

Hi Mladen,

Yes, batch processing is supported on single DSP devices. Could you provide any logs along with your configuration? 

Thank you,

Fabiana

Hi Fabiana,

Unfortunately, I cannot provide full logs for the actual network because they contain some confidential information which cannot be shared on a public forum. However, in the next few days I'll try to reproduce the issue with some simpler network and share the results.

Regards,

Mladen

Hi Mladen,

Sounds good.

Thank you,

Fabiana

Hi Fabiana,

Here is the zip file with a simple network which I used to reproduce the issue. You can find the network ONNX file and the log file which is generated by the tool during import process. Also, you can find the SVG generated by the tool after conversion and dumped outputs of the DataConvert layer for target (tda4_*) and PC simulation (pc_*). When you look into the files with dumped outputs, you can clearly observe that on PC all batches are correct, but on target, only first batch is correct and others are all 0.

1273.models.zip

The configuration file I pass to the import tool is as provided below.

modelType = 2
outputParamsFile = simple_tidl_io_
outputNetFile = simple_tidl_net.bin
inData = ./testvecs/input/simple/batch_000001.bin
inHeight = 192
inWidth = 192
inNumChannels = 1
numFrames = 1
numBatches = 5
inputNetFile = ../../models/simple_model.onnx
inFileFormat = 1
quantizationStyle = 2
biasCalibrationIterations = 5
calibrationOption = 7
outElementType = 6
rawDataInElementType = 6
addDataConvertToNet = 3
debugTraceLevel = 1
inElementType = 6
perfSimTool = ./ti_cnnperfsim.out
graphVizTool = ./tidl_graphVisualiser.out
perfSimConfig = ./device_config.cfg
tidlStatsTool = ./PC_dsp_test_dl_algo.out

Regards,

Mladen

Hi Mladen,

Thank you for sharing the data. Can you clarify the edgeai-tidl-tools and Processor SDK versions you are using? Can you verify that they are compatible using the version compatibility table: edgeai-tidl-tools/docs/version_compatibility_table.md at master · TexasInstruments/edgeai-tidl-tools · GitHub

It is recommended for you to use the latest versions: edgeai-tidl-tools tag version 10_01_00_02 and Processor SDK version 10.01.00.04.

Thank you,

Fabiana

Hi Fabiana,

We are currently working with PSDK v09.01.00.06 (as this is the version used by our client) and tools from edgeai-tidl-tools v09_01_08_00, but the same behavior is observed with PSDK v10.00.00.05 with tools from edgeai-tidl-tools v10_01_03_00, which was the latest version available when we investigated this issue. As far as I can see, the latest edgeai-tidl-tools version is 10_01_04_00 compatible with PSDK v10.01.00.04, but we have not tried with that one.

Nevertheless, can someone confirm if this is a known issue that is fixed in new versions and how we can fix it (e.g., by applying a patch) for earlier versions (if possible)?

Regards,

Mladen

Hi Mladen,

You can see the issues that are fixed in each release as well as the known issues that are still present here: https://github.com/TexasInstruments/edgeai-tidl-tools/releases

I have gotten confirmation that batch processing should be possible on this SOC, but have yet to replicate the issue internally. Please allow me some time to discuss with the team about this. Thank you for your patience.

Regards,

Fabiana