iwatake2222 / play_with_tflite

Sample projects for TensorFlow Lite in C++ with delegates such as GPU, EdgeTPU, XNNPACK, NNAPI
Apache License 2.0
357 stars 79 forks source link

Feature request: QNN delegate #62

Open pavloshargan opened 1 week ago

pavloshargan commented 1 week ago

Hi, I'd like to be able to run some tflite models on the smartphones with Qualcomm chip. I'm trying to get close to 1ms inference speed of mediaPipe pose estimation, and I'm getting something close to It by mimicking the approach here:

https://github.com/quic/ai-hub-apps/tree/main/apps/android/SemanticSegmentation

It uses Java and libraries imported from Maven. It doesn't work for my case, because I'm going to run the processing on videos, decoding and processing them directly OpenCV in my c++ code. I want the workflow to be as efficient as possible.

What can I start from in order to add the support for QNN delegate?

Thanks