bug in tensorrt graph analysis and tensorrt engine op
Created by: NHZlX
-
We should use topological sorting to traverse the graph to ensure that the nodes in the subgraph are also topological ordered. https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/analysis/subgraph_splitter.cc#79
-
We should filter the parameters here. https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/tensorrt_engine_op.cc#99
-
There exits loops inside the graph. For example: a = f(a), it will triger bug when setITensor during op converter process. So we should deal with the output name mapping.