| 415 | } |
| 416 | |
| 417 | bool opFuse(std::vector<Schedule::OpCacheInfo>& infos, MNNForwardType type, BackendConfig::PrecisionMode precision) { |
| 418 | std::unordered_map<const Tensor*, Node*> outputTensor; |
| 419 | // build graph |
| 420 | std::vector<std::unique_ptr<Node>> graph; |
| 421 | auto insertEdge = [&outputTensor](const Tensor* inputTensor, Node* succNode) { |
| 422 | if (outputTensor.find(inputTensor) != outputTensor.end()) { |
| 423 | auto preNode = outputTensor[inputTensor]; |
| 424 | succNode->pred.push_back(preNode); |
| 425 | preNode->succ.push_back(succNode); |
| 426 | } |
| 427 | }; |
| 428 | for (int i = 0; i < infos.size(); i++) { |
| 429 | auto& info = infos[i]; |
| 430 | auto& cmdBuffer = info.executeBuffer; |
| 431 | for (int j = 0; j < cmdBuffer.command.size(); j++) { |
| 432 | auto iter = cmdBuffer.command[j]; |
| 433 | /* |
| 434 | if (iter->buffer.get()) { |
| 435 | iter->op = flatbuffers::GetMutableRoot<Op>((void*)iter->buffer); |
| 436 | } |
| 437 | */ |
| 438 | std::unique_ptr<Node> node(new Node); |
| 439 | node->cmd = iter.get(); |
| 440 | node->topoIndex = i; |
| 441 | for (auto input : iter->inputs) { |
| 442 | insertEdge(input, node.get()); |
| 443 | } |
| 444 | for (auto output : iter->outputs) { |
| 445 | outputTensor[output] = node.get(); |
| 446 | } |
| 447 | graph.push_back(std::move(node)); |
| 448 | } |
| 449 | } |
| 450 | |
| 451 | if(type == MNN_FORWARD_OPENCL){ |
| 452 | for(int i = 0; i < graph.size(); ++i){ |
| 453 | mergeConvolutionAndPrelu(graph[i].get(), type); |
| 454 | } |
| 455 | for(auto iter = graph.begin(); iter != graph.end();){ |
| 456 | if(iter->get()->cmd->op == nullptr){ |
| 457 | iter = graph.erase(iter); |
| 458 | }else{ |
| 459 | ++iter; |
| 460 | } |
| 461 | } |
| 462 | } |
| 463 | |
| 464 | std::queue<Node*> postDominateNodeQueue; |
| 465 | // build dominate tree |
| 466 | for (int i = static_cast<int>(graph.size()) - 1; i >= 0; i--) { |
| 467 | auto node = graph[i].get(); |
| 468 | if (!node->succ.empty()) { |
| 469 | auto parent = node->succ[0]; |
| 470 | for (int j = 1; j < node->succ.size(); j++) { |
| 471 | parent = LCA(parent, node->succ[j]); |
| 472 | } |
| 473 | node->domainatePred = parent; |
| 474 | if (parent) { |
no test coverage detected