build model with `--dla_core 1` but model looks to run on DLA core 0

[y-okumura-isp]

In my understanding, we can use DLA core 1 by building the model, but we can not specify the core on runtime. https://github.com/jkjung-avt/tensorrt_demos/issues/394 Though I set --dla_core 1 at build time, it looks DLA Core 0 is used.

Here is my environment.

• Board
  • Jetson NX
  • /etc/nv_tegra_release: R32 (release), REVISION: 5.1
  • TensorRT 7.1.3.0
• Version
  • 07f3d4af070014177fe5594ec294234ff628dd0f

I built the tensorrt model as following:

$ cd yolo
$ python3 onnx_to_tensorrt.py -m yolov4-csp-512 --dla_core 1
Using DLA core 1.
(snip)

$ cd ..
$ python3 trt_yolo.py --image dog.jpg -m yolov4-csp-512

And I checked which DLA core is used: We can do this by checking /sys/.../runtime_status according to https://forums.developer.nvidia.com/t/matrix-multiply-on-dla-and-checking-dla-usage/84338/2

$ while true; do 
  echo "===";
  cat /sys/devices/platform/host1x/15880000.nvdla0/power/runtime_status;
  cat /sys/devices/platform/host1x/158c0000.nvdla1/power/runtime_status;
  sleep 0.1;
done

I got the result as below. It looks DLA Core 0 is used.

===
active
suspended

All suggestions are welcome. Thanks.

[jkjung-avt]

Thanks for reporting the issue. Below is the relevant source code, and the relevant documentation is here: IBuilderConfig. I followed NVIDIA's API and set DLA_Core in the TensorRT Builder Config. I'm not sure why it does not work. I think you could re-direct this question to NVIDIA.

https://github.com/jkjung-avt/tensorrt_demos/blob/9dd56b3b8d841dcfc2e5d1868f4bd785a50cbe98/yolo/onnx_to_tensorrt.py#L159-L164

[y-okumura-isp]

Thank you for your comment. It looks the code is correct. I try to post NVIDIA Forum. By the way, can anyone reproduce this situation? How about on boards except for NX?

[xiaochus]

I have the same question and find the solution on my device . I try config->setDLACore(0) and config->setDLACore(1), but the model always run on DLA 0. Then, i try setDLACore at runtime and the model can run on the correct DLA.