Caenorst commented on a change in pull request #11325: Added TensorRT runtime integration URL: https://github.com/apache/incubator-mxnet/pull/11325#discussion_r201521176
########## File path: src/executor/tensorrt_pass.cc ########## @@ -0,0 +1,583 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +/*! + * Copyright (c) 2018 by Contributors + * \file tensorrt_pass.cc + * \brief Replace TRT compatible subgraphs by TRT engines + * \author Clement Fuji Tsang + */ + +#if MXNET_USE_TENSORRT + +#include <mxnet/base.h> +#include <mxnet/operator.h> +#include <mxnet/op_attr_types.h> +#include <nnvm/graph_attr_types.h> +#include <onnx/onnx.pb.h> +#include <NvInfer.h> + +#include "./onnx_to_tensorrt.h" +#include "./exec_pass.h" +#include "../operator/contrib/nnvm_to_onnx-inl.h" + +namespace mxnet { +namespace exec { + +using NodePtr = nnvm::NodePtr; + +/*! + * \brief Custom graph class, which will contain bi-directional nodes + * we need to compute DFS and reverse DFS for graph partitioning + */ +class BidirectionalGraph { + public: + struct Node { + nnvm::Node* nnvmptr; + std::vector<Node*> inputs; + std::vector<Node*> outputs; + }; + std::vector<Node> nodes; + std::unordered_map<nnvm::Node*, uint32_t> nnvm2nid; + std::vector<Node*> outputs; + static const std::unordered_set<std::string> unconditionalTRTop; + + explicit BidirectionalGraph(const Graph &g) { + auto& idx = g.indexed_graph(); + auto num_nodes = idx.num_nodes(); + nodes.reserve(num_nodes); + nnvm2nid.reserve(num_nodes); + outputs.reserve(idx.outputs().size()); + DFSVisit(g.outputs, [this](const nnvm::NodePtr& n) { + BidirectionalGraph::Node new_node; + new_node.nnvmptr = n.get(); + nnvm2nid[n.get()] = static_cast<uint32_t>(nodes.size()); + nodes.emplace_back(std::move(new_node)); + }); + for (const auto& it : nnvm2nid) { + nnvm::Node* nnvmnode = it.first; + uint32_t nid = it.second; + for (auto& n : nnvmnode->inputs) { + uint32_t input_nid = nnvm2nid[n.node.get()]; + nodes[input_nid].outputs.emplace_back(&nodes[nid]); + nodes[nid].inputs.emplace_back(&nodes[input_nid]); + } + } + for (auto& e : g.outputs) { + uint32_t nid = nnvm2nid[e.node.get()]; + outputs.emplace_back(&nodes[nid]); + } + } + + template <typename FVisit> + void DFS(const std::vector<Node*>& heads, bool reverse, FVisit fvisit) { + std::unordered_set<Node*> visited; + std::deque<Node*> stack(heads.begin(), heads.end()); + visited.reserve(heads.size()); + while (!stack.empty()) { + Node* vertex = stack.back(); + stack.pop_back(); + if (visited.count(vertex) == 0) { + visited.insert(vertex); + fvisit(vertex); + std::vector<Node*> nexts = reverse ? vertex->inputs : vertex->outputs; + for (Node* node : nexts) { + if (visited.count(node) == 0) { + stack.emplace_back(node); + } + } + } + } + } + + using t_pairset = std::pair<std::unordered_set<Node*>, std::unordered_set<Node*>>; + using t_pairvec = std::pair<std::vector<Node*>, std::vector<Node*>>; + using t_uncomp_map = std::unordered_map<Node*, std::unordered_set<Node*>>; + + std::unordered_set<Node*> naive_grow_subgraph(Node* head, + std::unordered_set<Node*>* set_unused, + t_uncomp_map* uncomp_map) { + std::unordered_set<Node*> subgraph; + std::unordered_set<Node*> uncomp_set; + std::deque<Node*> stack; + stack.emplace_back(head); + while (!stack.empty()) { + Node* vertex = stack.back(); + stack.pop_back(); + if (set_unused->count(vertex) && !uncomp_set.count(vertex)) { + set_unused->erase(vertex); + subgraph.insert(vertex); + uncomp_set.insert((*uncomp_map)[vertex].begin(), (*uncomp_map)[vertex].end()); + for (Node* input : vertex->inputs) { + if (set_unused->count(input) && !uncomp_set.count(input)) { + stack.emplace_back(input); + } + } + for (Node* output : vertex->outputs) { + if (set_unused->count(output) && !uncomp_set.count(output)) { + stack.emplace_back(output); + } + } + } + } + return subgraph; + } + + std::vector<std::unordered_set<Node*>> get_subsets( + std::unordered_map<std::string, NDArray>* const params_map) { + std::vector<std::unordered_set<Node*>> subgraphs; + std::unordered_set<Node*> set_nonTRTnodes; + std::unordered_set<Node*> set_allnodes(nodes.size()); + std::vector<t_pairset> separation_sets; + for (Node& node : nodes) { + if (!IsTRTCompatible(node.nnvmptr)) { + set_nonTRTnodes.insert(&node); + std::unordered_set<Node*> in_graph; + std::unordered_set<Node*> out_graph; + std::vector<Node*> dummy_head; + dummy_head.emplace_back(&node); + DFS(dummy_head, false, [&out_graph](Node* node) { + out_graph.insert(node); + }); + DFS(dummy_head, true, [&in_graph](Node* node) { + in_graph.insert(node); + }); + separation_sets.emplace_back(std::make_pair(in_graph, out_graph)); + } + set_allnodes.emplace(&node); + } + t_uncomp_map uncomp_map; + std::unordered_set<Node*> set_TRTnodes; + set_TRTnodes.insert(set_allnodes.begin(), set_allnodes.end()); + for (Node* n : set_nonTRTnodes) { + set_TRTnodes.erase(n); + } + for (Node* n : set_TRTnodes) { + for (t_pairset p : separation_sets) { + if (p.first.count(n)) { + uncomp_map[n].insert(p.second.begin(), p.second.end()); + } else if (p.second.count(n)) { + uncomp_map[n].insert(p.first.begin(), p.first.end()); + } + } + for (Node* nonTRTn : set_nonTRTnodes) { + uncomp_map[n].erase(nonTRTn); + } + } + std::unordered_set<Node*> set_unused; + set_unused.reserve(set_TRTnodes.size()); + + for (auto& n : set_TRTnodes) { + if (n->nnvmptr->attrs.op != nullptr || params_map->count(n->nnvmptr->attrs.name)) { + set_unused.insert(n); + } + } + std::unordered_set<Node*> visited; + std::deque<Node*> stack(outputs.begin(), outputs.end()); + while (!stack.empty()) { + Node* vertex = stack.front(); + stack.pop_front(); + if (!visited.count(vertex)) { + visited.insert(vertex); + if (set_unused.count(vertex)) { + subgraphs.emplace_back(naive_grow_subgraph(vertex, &set_unused, &uncomp_map)); + } + for (Node* input : vertex->inputs) { + stack.emplace_back(input); + } + } + } + + return subgraphs; + } + + + private: + friend class Graph; + + bool IsTRTCompatible(nnvm::Node* nodeptr) { + if (nodeptr->op() != nullptr) { + const std::string op_name = nodeptr->op()->name; + if (op_name == "Pooling") { + return (nodeptr->attrs.dict.at("pool_type") == "avg" || + nodeptr->attrs.dict.at("pool_type") == "max"); + } else if (unconditionalTRTop.count(op_name)) { + return true; + } else if (op_name == "Activation") { + return nodeptr->attrs.dict.at("act_type") == "relu" || + nodeptr->attrs.dict.at("act_type") == "tanh" || + nodeptr->attrs.dict.at("act_type") == "sigmoid"; + } + return false; + } + return true; + } +}; // class BidirectionalGraph + +/*! + * \brief function which transform std::vector<dmlc::any> back to Attrs (dmlc::any) + */ +const std::unordered_set<std::string> BidirectionalGraph::unconditionalTRTop = { + "Convolution", + "BatchNorm", + "elemwise_add", + "elemwise_sub", + "elemwise_mul", + "rsqrt", + "pad", + "Pad", + "mean", + "FullyConnected", + "Flatten", + "SoftmaxOutput", +}; + + +using NodeEntrySet = std::unordered_set<nnvm::NodeEntry, nnvm::NodeEntryHash, + nnvm::NodeEntryEqual>; + +/*! + * \brief get the output nodes of the subgraph in the main graph + * \return a vector of the output nodes +*/ +std::vector<nnvm::NodeEntry> GetSubgraphOutputs(Graph g, + std::unordered_set<nnvm::Node*> set_subgraph) { + std::vector<nnvm::NodeEntry> outputs; + NodeEntrySet _outputs; + for (auto& e : g.outputs) { + if (set_subgraph.count(e.node.get())) { + _outputs.insert(e); + } + } + DFSVisit(g.outputs, [&set_subgraph, &_outputs](const nnvm::NodePtr &node){ + if (!set_subgraph.count(node.get())) { + for (auto& e : node->inputs) { + if (set_subgraph.count(e.node.get())) { + _outputs.insert(e); + } + } + } + }); + outputs.insert(outputs.begin(), _outputs.begin(), _outputs.end()); Review comment: can you be more specific about A, B, C ? Are they node or node entries (you say that final output is B+C, so are those node entries ?) ? Only Op can / can't be TRT compatible. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
