TDA4VH-Q1: Batch processing in TDA4VH

srikar varaganti

Part Number: TDA4VH-Q1
Other Parts Discussed in Thread: TDA4VH

Hello,

Currently we are operating on TDA4VH processor and using sdk: 09_00_00_02 in QNX mode.

This query is regarding Batch processing of Object detection model in multi-core mode. Our customized mode inputs single image at a time, we wanted to enhance the speed of operations by enabling multi-core parallel processing feature. I tried generating the artifacts using NumCores =2, inferenceMode=1. But I am confused on how can I implement this parallel processing as part of a single OpenVX graph. Are there any example/sample codes which I can look at that could help me build the application discussed above. The goal is to run the OD app in multiple cores parallelly on different images while still the network inputs only single frame per core.

Thank you.

over 1 year ago

0 Anand Pathak over 1 year ago

TI__Genius 9065 points

Hi,

From what I understand based on question, you are interested in parallelizing inference of a single frame across multiple cores for a given network, is that correct?

In this case, please try using inferenceMode = 2. Would recommend using the latest SDK release for the same. Please refer this link for further documentation : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_multi_c7x.md

Let me know in case of further questions.

Regards,

Anand

0 srikar varaganti over 1 year ago in reply to Anand Pathak

Intellectual 280 points

It's actually the other way, we intend to run multiple frames on different cores on the same neural network, Is there any sample implementation of the openvx graph that I can refer to for developing this application?

0 Pramod Kumar Swami over 1 year ago in reply to srikar varaganti

TI__Genius 14000 points

Hello Srikar,
Since you are trying to tun multiple frames, I assume all the frames are available while calling the inference either from multiple camera sources or via buffering. Please refer the same documentation github.com/.../tidl_fsg_multi_c7x.md and refer TIDL_inferenceModeHighThroughput.

You need to have your ONNX graph indicating batch dimension as 2 (since you are trying NumCores=2) and then you can infer and the network will execute in a way that 1 frame (or batch) on first core and second frame(or batch) on second core

Thanks

Pramod

Processors

Processors forum

TDA4VH-Q1: Batch processing in TDA4VH