I am having a problem while building the TensorRT engine on Xavier AGX. I used GoogleNet and assign the first 58 layers to the GPU while the rest (82) to NVDLA. The error and warning messages are inserted below:
[TensorRT] WARNING: Layer prob can’t run on DLA with given parameters, falling back to GPU.
[TensorRT] WARNING: Detected invalid timing cache, setup a local cache instead
[TensorRT] ERROR: 2: [nvdlaUtils.cpp::getInputDesc::176] Error Code 2: Internal Error (Assertion idx < num failed.Index is out of range of valid number of input tensors.)
I started to face this problem after upgrading from TensorRT 7.1.3 to TensorRT 8.0.2. I followed the same building process with TensorRT 7.1.3 and everything was okay.
The same problem occurs with the first 34 layers to GPU and the rest to DLA.
I believe my scripts is fine since I can build for different points such as 52 GPU and the rest to DLA. Also, it was working without no problem on TensorRT 7.1.3 on every possible transition point from GPU to DLA.
Setup:
I downloaded the GoogleNet prototxt file from here: dusty-nv/jetson-inference.
I have exactly the same setup with JetPack 4.6