Error when enabling DLA for DetectNet (Solved)

Autonomous Machines Jetson & Embedded Systems Jetson AGX Xavier
1

In the following snippet of code:

builder->setMaxBatchSize(maxBatchSize);
builder->setMaxWorkspaceSize(16 << 20);
         
builder->allowGPUFallback(true);
builder->setFp16Mode(true);
builder->setDefaultDeviceType(static_cast<nvinfer1::DeviceType>(2));

// set up the network for paired-fp16 format
if(mEnableFP16)
	builder->setHalf2Mode(true);

nvinfer1::ICudaEngine* engine = builder->buildCudaEngine(*network);

builder->setDefaultDeviceType(static_castnvinfer1::DeviceType(2)); This line is generating the error.

LOG:

[GIE]  TensorRT version 5.0, build 0
[GIE]  attempting to open cache file networks/1280x720/snapshot_iter_8938250.caffemodel.2.tensorcache
[GIE]  cache file not found, profiling network model
[GIE]  platform has FP16 support.
[GIE]  loading networks/1280x720/deploy.prototxt networks/1280x720/snapshot_iter_8938250.caffemodel
[GIE]  retrieved output tensor 'coverage'
[GIE]  retrieved output tensor 'bboxes'
[GIE]  configuring CUDA engine
[GIE]  building CUDA engine
[GIE]  ../builder/cudnnBuilder2.cpp (689) - Misc Error in buildSingleLayer: 1 (Unable to process layer.)
[GIE]  ../builder/cudnnBuilder2.cpp (689) - Misc Error in buildSingleLayer: 1 (Unable to process layer.)
[GIE]  failed to build CUDA engine
failed to load networks/1280x720/snapshot_iter_8938250.caffemodel
detectNet -- failed to initialize.
detectnet-pipeline:   failed to initialize imageNet
nvbuf_utils: dmabuf_fd 1105 mapped entry NOT found
Segmentation fault (core dumped)

How do I fix this issue?

2

It looks like you are using GPU fallback but there is an error when TRT is building the network.

From the console log, it appears that you are using a fork of jetson-inference. I have been adding DLA support to the dev branch here: [url]https://github.com/dusty-nv/jetson-inference/tree/dev[/url]

You may want to try maxBatchSize of 1, since I notice your model is 1280x720. Also you will want to try the trtexec program that comes with TRT samples just to make sure it isn’t a user error.

Note that official DLA support is limited to Alexnet, Googlenet, and Resnet in the early-access JetPack. I will have to try to run this later as well.

3

my model is 1200*1920, also

../builder/cudnnBuilder2.cpp (728) - Misc Error in buildSingleLayer: 1 (Unable to process layer.)
 ../builder/cudnnBuilder2.cpp (728) - Misc Error in buildSingleLayer: 1 (Unable to process layer.)

Is this because my input is too large?

I have tried other model with input size 640*640, it’s ok