TDA4VM: TIDL_Doubts

Kaviraj S

Intellectual 320 points

Part Number: TDA4VM

Hi,

I have a few doubts about the TIDL.

Is there any way I can benchmark or test individual tidl layers, for example, If I want to test the conv or pooling layer alone is there any unit test available? If yes how to compile and execute it
Apart from using the inference file provided in the edgeai-tidl-tools is there any way that I can write a stand-alone application or architecture any ML algorithm on the combination of different layers that runs on c7x alone using mma

Thanks,

Kaviraj

over 1 year ago

0 Pratik Kedar over 1 year ago

TI__Mastermind 24041 points

Kaviraj S said:
Is there any way I can benchmark or test individual tidl layers, for example, If I want to test the conv or pooling layer alone is there any unit test available? If yes how to compile and execute it

You can create onnx file which only contains specific layer/layers that you want to compile - test/infer, by removing other layers (With best of my knowledge, I dont recall this layer fission/fusion based support is available as part of TIDL offering.

Kaviraj S said:
Apart from using the inference file provided in the edgeai-tidl-tools is there any way that I can write a stand-alone application or architecture any ML algorithm on the combination of different layers that runs on c7x alone using mma

As part of OSRT, offering we have provided examples, users are more than welcome to write there own script by referring the provided scripts, example : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/examples/osrt_python/ort/onnxrt_ep.py

Now coming to running all the layers using c7x-mma, you can check what all layers are supported by tidl that can be offloaded for acceleration here : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/supported_ops_rts_versions.md and take the call

0 Kaviraj S over 1 year ago

Intellectual 320 points

Hi,

I also would like to know how we can find/how to specify a block of code in my application to run in MMA along with c7x. Will it take automatically or is there any specific compiler flags or defines to be enabled in the code?

Can you provide any examples available in the RTOS SDK?

Thanks,

Kaviraj

0 Pratik Kedar over 1 year ago in reply to Kaviraj S

TI__Mastermind 24041 points

Kaviraj S said:
I also would like to know how we can find/how to specify a block of code in my application to run in MMA along with c7x. Will it take automatically or is there any specific compiler flags or defines to be enabled in the code?

Did you mean, you want to add custom operator to model that runs on C7x-MMA ?

0 Kaviraj S over 1 year ago in reply to Pratik Kedar

Intellectual 320 points

Yes, I want to build my ML application with custom operators and run it from c7x with MMA. To activate/to run specific blocks of code in MMA is there a way to check it?

Thanks,

Kaviraj

0 Pratik Kedar over 1 year ago in reply to Kaviraj S

TI__Mastermind 24041 points

If I comprehend your exact ask,

You wanted to understand how to write C7x/MMA specific code for certain functionality ? (General Purpose)

You can understand more on C7x programming here :

https://www.ti.com/lit/ug/spruig6j/spruig6j.pdf?ts=1704777698135&ref_url=https%253A%252F%252Fwww.ti.com%252Fsitesearch%252Fen-us%252Fdocs%252Funiversalsearch.tsp%253FlangPref%253Den-US%2526searchTerm%253DC7x%2526nr%253D16

https://www.ti.com/document-viewer/lit/html/spruig8?keyMatch=C7X

Please go through the content, for understanding C7x programming in general. Also you can look at TRM of specific SoC you are working on for more details.

To utilize the MMA functionality you can look into MMA Lib documentation here : https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/09_01_00_06/exports/docs/mmalib_09_01_00_06/docs/user_guide/index.html

AFAIK, explicit MMA programming guide is not available currently, you can refer above for reference.

Processors

Processors forum

TDA4VM: TIDL_Doubts