TPAT - TensorRT Plugin Autogen Tool

Introduction

1. Automatically generate high-performance TensorRT plugins for unsupported operators or replacing inefficient kernels.
2. End-to-end command line tool. No requirement for any CUDA programming knowledge. Users only need to provide the ONNX model and assign the node names or types to auto-generate TensorRT plugin.
3. The performance of auto-generated TensorRT plugins in real cases:
   • Performance comparation with hand-written kernels
   • Optimization for TensorRT’s original kernels

Support Matrix

• ONNX Operators supported by TPAT-1.0

Runtime Env : dockerfile

1. Build image

nvidia-docker build .

2. Run container

nvidia-docker run -itd --gpus all -v <TPAT path dir>:/root <Image_ID> /bin/bash

3. Execute conrainer

nvidia-docker exec -it <Container_ID> /bin/bash

4. Modify CUDA_PATH and TRT_PATH in python/trt_plugin/Makefile

CUDA_PATH: local CUDA installation path
TRT_LIB_PATH: local TensorRT installation path

5. Plugin auto generated

cd examples
python test_onehot_dynamic_direct.py

• tpat_onehot.so is stored in python/trt_plugin/lib/

Runtime Env : Build

1. Prerequisites

System Packages

• LLVM >= 9.0.1, (LLVM==9.0.1 recommended)
• GCC >= 7.3.0, (GCC==7.4.0 recommended)
• TensorRT

PyPI packages

• numpy pycuda onnx onnxruntime onnx_graphsurgeon xgboost jinja2 ctypes tornado cloudpickle psutil

  NOTE: these necessary packages are recorded in requirements.txt

Optional packages

• tensorflow-gpu==1.15
• tf2onnx
• torch
• pytest

  NOTE: these optional packages are required by Example and UnitTest

2. Clone the TPAT repository

git clone -b master https://github.com/nvidia/TensorRT TPAT
cd TPAT
git submodule update --init --recursive

3. Build BlazerML-TVM

mkdir build && cp cmake/config.cmake build
#Edit build/config.cmake to customize the compilation options
set(USE_LLVM /usr/local/llvm/bin/llvm-config)
set(USE_CUDA ON)
#gcc compiler is required to support C++14
cd build && cmake .. 
make -j
#TVM Python package
export TVM_HOME=/path/to/tvm
export PYTHONPATH=$TVM_HOME/python:${PYTHONPATH}

4. Plugin Compiler Env

Modify python/trt_plugin/Makefile according to your environment setup.

CUDA_PATH: local CUDA installation path
TRT_LIB_PATH: local TensorRT installation path

Usage

TPAT provides a Python function and command line for usage.

Python function

onnx2plugin(
    input_model_path, 
    output_model_path, 
    node_names=None, 
    node_types=None, 
    plugin_name_dict=None,
    dynamic_bs=False, # if True, this operator support dynamic batchsize
    min_bs=1,
    max_bs=256,
    opt_bs=128
    )

• input_model_path[required] : input onnx model including nodes which require TRT plugin
• output_model_path[required] : output onnx model where the corresponding node types are replaced by plugin names. The output onnx model can be directly converted to TRT with onnx parser and built plugin dynamic library.
• node_names : list of node names for autogen
• node_types : list of node types for autogen
• plugin_name_dict : dict of {plugin_name: node_name} for autogen
• dynamic_bs : if True, TPAT will generate plugin that supported dynamic batch, if False, generated plugin only support fixed shapes but has better performance.
• min_bs: the minium batch size in range of dynamic batch.
• max_bs: the maxium batch size in range of dynamic batch.
• opt_bs: the optimize batch size in range of dynamic batch.

  NOTE: For node_names, node_types, plugin_name_dict, at least one of them should be provided

Command line

# Separate different ops with spaces
python3 Onnx2Plugin.py -i input.onnx -o output.onnx -n op_name1 op_name2 -dynamic=true -min=1 -max=512 -opt=256
python3 Onnx2Plugin.py -i input.onnx -o output.onnx -t op_type1 op_type2 -dynamic=false
python3 Onnx2Plugin.py -i input.onnx -o output.onnx -p '{"op_name1": "plugin_name1", "op_name2": "plugin_name2"}'

• -i[required]: input_model_path
• -o[required]: output_model_path
• -n: node_names
• -t: node_types
• -p: plugin_name_dict
• -dynamic: dynamic_bs
• -min: min_bs
• -max: max_bs
• -opt: opt_bs

Output

1. Assign nodes and plugin names through plugin_name_dict

• trt_plugin/src contains {plugin_name}.cu and {plugin_name}.h
• trt_plugin/lib contains {plugin_name}.so

2. Assign node names or node types

• trt_plugin/src contains tpat_{node_name}.cu and tpat_{node_name}.h
• trt_plugin/lib contains tpat_{node_name}.so

Example && UnitTest

• Example : example_tensorflow.py
• UnitTest : test_tapt.py

Release notes

Changelog

• Support mutiple nodes for autogen
• Support boolean input/outputs
• Able to reuse plugins

Known issues

• Only support dynamic BatchSize
• Opeartors with int8/float16/double inputs/outputs are not supported

TODO

• Support ONNX subgraph for autogen
• Support direction conversion from TensorFlow and PyTorch