iMX8M Plus: benchmark_model stops when NPU debugging is enabled

shiitaishigaki — Wed, 16 Mar 2022 01:55:48 GMT

Hi all,

I am trying to use TensorflowLite's benchmark_model with NPU debugging (export CNN_PERF=1 VIV_VX_DEBUG_LEVEL=1) enabled as a test for the NPU delegator.

But the program stops in the middle and you cannot exit from the program.

This problem does not occur unless debugging of the NPU is enabled.

The model used in this case is a simple model where the input data is 1 x 224 x 224 x 1 and Conv2D is done 50 times followed by Add 50 times.

i.MX Yocto Project(zeus-5.4.70-2.3.6) is running on EVK.

Running benchmark_model

export CNN_PERF=1 VIV_VX_DEBUG_LEVEL=1
/usr/bin/tensorflow-lite-2.4.0/examples/benchmark_model --graph=model/conv2d_add.tflite --use_nnapi=true

After a short wait.

[galcore]: Stop driver to keep scene.

Why is this problem happening?

I have attached the model in ZIP format.

topic iMX8M Plus: benchmark_model stops when NPU debugging is enabled in eIQ Machine Learning Software

iMX8M Plus: benchmark_model stops when NPU debugging is enabled