VeriSilicon / TIM-VX

VeriSilicon Tensor Interface Module
Other
224 stars 84 forks source link

tflite profiling does not show details when using vx delegate #571

Open DiXcipuli opened 1 year ago

DiXcipuli commented 1 year ago

Using the tflite benchmark executable on my NXP IMX8 board to profile my quantized model (found on this repo under /tensorflow/lite/tools/benchmark/ ) I can get all operation runtime details when running on the CPU, but none is displayed when the param --external_delgate_path=/usr/lib/libvx_delegate.so is given , expected one Vx Delegate Node for the whole model. I would like to have the individual operator details, what is the runtime for each Conv etc.. . I have seen this topic for the GPU, is it the same issue for the vx_delegate ?

Given the documentation from NXP (page 14), I run the benchmark tool with the following arguments: ./benchmark_model --graph=my_model.quant.tflite --enable_op_profiling=true --max_delegated_partitions=1000 --external_delgate_path=/usr/lib/libvx_delegate.so

sunshinemyson commented 1 year ago

@DiXcipuli ,

Sorry for late, We don't support profile mode in tflite yet.