Some plugin differences between 5.0 RT and 4.0

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

I migrated plugins from TensorRT 4.0.x to TensorRT 5.0 RC. I found something different in plugins that not only need to realize IPluginExt new virtual funcs written in the guide.

  1. the uff parser may something different causing some error. Talking here https://devtalk.nvidia.com/default/topic/1042385/tensorrt/tensorrt-5-0rc-on-xavier-jetpack-4-0ea-vs-tensorrt-4-0ga-on-tx2-jetpack-3-3-/?offset=5#5289402
  2. IPluginFactory create plugins. In 5.0RC: nvuffparser::IPluginFactory::createPlugin needs to new and delete by user. But in nvinfer1::IPluginFactory::createPlugin for deserialization only needs to new plugin , and return the heap pointer. Do not care release here, or else cause a fault. It will done by RT engine?

    In 4.0.x:
    all createPlugin methods need to new and delete by user.

    This may be something strange in 5.0.

  3. In IPluginExt, func 'getOutputDimensions' in 5.0 RC are no longer called for all plugins. Only the first create one, the clone one will not call it. So I must move the init for dims to 'configureWithFormat' in which are still called by every plugin.

Finally I migrated plugins from TensorRT 4.0.x to TensorRT 5.0 RC.
I don’s sure about the reason. Is it right above?

2

Hello,

Thank you for posting your migration experience. I will pass along your input to the development team.