We are using the MX93 NPU to perform audio processing. The neural network is recurrent with many input and output states. Loading such a vela compiled network with tensorflow lite+libethosu_delegate.so causes a segmentation violation.
The issue can be reproduced on an i.MX93EVK with the latest firmware LF_v6.12.3-1.0.0_images_IMX93EVK. Executed the commands in python3:
import tflite_runtime.interpreter as tflite
ext_delegate = [tflite.load_delegate("/usr/lib/libethosu_delegate.so")]
interpreter=tflite.Interpreter(model_path='vela-segv.tflite', experimental_delegates=ext_delegate)
interpreter.allocate_tensors()
Networks with less variables work just fine.
Any ideas?
