Issue building yolo plugin with TRT 6.5

[lazatsoc]

Hi,

Thanks for your useful contribution.

I am trying to run Yolo (Demo 4,5) on x86_64 machine without success. I use a GTX 1080Ti (CC=61), TensorRT version 6.5 and CUDA 10.2. According to prerequisites, Yolo should run with this configuration.

When trying to build the yolo-layer plugin, I get kind of the same error with https://github.com/jkjung-avt/tensorrt_demos/issues/576, https://github.com/jkjung-avt/tensorrt_demos/issues/222#issuecomment-732463357. The output of make is shown here: https://pastebin.com/QeJnYUPh.

I have given the correct CC in Makefile:

compute=61

I have ensured that the libraries and the headers of TRT can be found, changing the paths in the Makefile like this:

TENSORRT_INCS=-I"/usr/include/x86_64_linux_gnu" TENSORRT_LIBS=-I"/usr/lib/x86_64_linux_gnu.

The errors seem like an incompatibility with the headers of TensorRT 6.5.

Is it true that Yolo requires 6.x+ and not 7.x+? Is there a workaround for these errors?

[jkjung-avt]

Errors like this one could be fixed easily by modifying function declarations in the header and C++ source files.

yolo_layer.h(91): error: exception specification for virtual function "nvinfer1::YoloLayerPlugin::detachFromContext" is incompatible with that of overridden function "nvinfer1::IPluginV2Ext::detachFromContext"

I'm not able to find TensorRT 6.5 header files from NVIDIA web site (https://developer.nvidia.com/nvidia-tensorrt-6x-download) so I don't know what exactly went wrong for you.

Just as a quick check, could you remove the following lines of code.

https://github.com/jkjung-avt/tensorrt_demos/blob/a061e44a82e1ca097f57e5a32f20daf5bebe7ade/plugins/yolo_layer.h#L13-L17

And replace them with 1 single line of: (force NOEXCEPT to be empty)

#define NOEXCEPT

Then try to make the plugin again.

[lazatsoc]

Thanks for your quick reply.

I managed to build libyolo_layer.so. The solution was to replace all conditions with NV_TENSORRT_MAJOR and to keep the configuration for TRT 8. I don't know why this is correct for TRT 6.5 that I use. This version is found on DRIVE AGX and nowhere else, so maybe it is a special case.

Specifically, I did the following changes: 1) I replaced https://github.com/jkjung-avt/tensorrt_demos/blob/a061e44a82e1ca097f57e5a32f20daf5bebe7ade/plugins/yolo_layer.h#L13-L17 with https://github.com/jkjung-avt/tensorrt_demos/blob/a061e44a82e1ca097f57e5a32f20daf5bebe7ade/plugins/yolo_layer.h#L14 2) I replaced https://github.com/jkjung-avt/tensorrt_demos/blob/a061e44a82e1ca097f57e5a32f20daf5bebe7ade/plugins/yolo_layer.h#L93-L97 with https://github.com/jkjung-avt/tensorrt_demos/blob/a061e44a82e1ca097f57e5a32f20daf5bebe7ade/plugins/yolo_layer.h#L94 3) I replaced https://github.com/jkjung-avt/tensorrt_demos/blob/a061e44a82e1ca097f57e5a32f20daf5bebe7ade/plugins/yolo_layer.cu#L260-L264 with https://github.com/jkjung-avt/tensorrt_demos/blob/a061e44a82e1ca097f57e5a32f20daf5bebe7ade/plugins/yolo_layer.cu#L261

[jkjung-avt]

OK. Thanks for sharing.