We are using the NPU on a MX93 CPU to perform audio processing. The neural network is recurrent with many state variables. Loading such a vela compiled network with tensorflow+libethosu_delegate.so causes a segmentation violation. Networks with less inputs and outputs work fine.
The problem can be easily reproduced on the i.MX93EVK with the latest OS LF_v6.12.3-1.0.0_images_IMX93EVK. Within python3 execute these commands:
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()
Any ideas?
