plugin

This is an implementation of some custom plugins using TensorRT API, CUDA API, and c++. Each sub-module is equipped with gtest samples.

Dependencies

• CUDA 10.0
• TensorRT-7.0.0.11 (for CUDA10.0)
• CUDNN (for CUDA10.0, may not be used)
• libtorch (torch c++ lib of cpu version, gpu version may conflict with the environment)
• gtest (Google C++ testing framework)

Current plugins:

• Resize_v0 : upsample_bilinear2d, inherited from IPluginV2Ext (introduced in TensorRT 5.1)
• Resize_v1 : TODO upsample_bilinear2d, inherited from IPluginV2DynamicExt
• DCN: deformable CNN

Build

see bash

compile.sh

To test cublas apis, compile the cublas_test.cu with this command

nvcc -Xcompiler -fPIC -std=c++11 -lcublas cublas_test.cu -o cublas_test &&
./cublas_test

Precaution:

For debug mode, since the libtorch I used was compiled beforehanded, the lib may confict with your own environment. Please do these operations:

• download and unzip the libtorch lib: libtorch-1.2.0, which was compiled beforehanded with new GCC ABI.
• install the Googletest: https://github.com/google/googletest.git .
• remove the libgtest.a and libgtest_main.a in libtorch to avoid conflicts.

For release mode, as u see in the CMakeLists.txt, we do not need to link the libtorch and gtest.