1 /** 2 * Copyright 2021-2022 Huawei Technologies Co., Ltd 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17 #ifndef MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_TSA_ATOMIC_ADD_TO_FIRST_TENSOR_H_ 18 #define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_TSA_ATOMIC_ADD_TO_FIRST_TENSOR_H_ 19 20 #include <memory> 21 #include <tuple> 22 #include <utility> 23 #include <vector> 24 #include "include/backend/optimizer/optimizer.h" 25 #include "backend/common/graph_kernel/add_atomic_clean.h" 26 #include "include/backend/kernel_graph.h" 27 28 namespace mindspore::graphkernel { 29 /* 30 * output = SubGraph(input_x, indices, update) { 31 * %0 = TensorScatterAdd(%para1, %para2, %para3) 32 * return %0 33 * } 34 * ----------------------------------------------------------------> 35 * // Initialize output with input_x. 36 * output = Reshape(input_x) 37 * fake_out = SubGraph'(output, indices, update) { 38 * %0 = TensorScatterAdd(%para1, %para2, %para3) 39 * %1 = Assign(%para1, %0, umond) // 40 * return %1 41 * } 42 */ 43 class TsaAtomicAddToFirstTensor : public AtomicCleanInserter { 44 public: TsaAtomicAddToFirstTensor()45 TsaAtomicAddToFirstTensor() : AtomicCleanInserter("tensor_scatter_add_atomic_add_to_first_tensor") {} 46 ~TsaAtomicAddToFirstTensor() override = default; 47 48 bool Run(const FuncGraphPtr &func_graph) override; 49 50 private: 51 void ProcessOriginalCNode(const AnfNodePtr &composite_node, 52 const std::vector<std::tuple<InplaceAssignerInfo, AnfNodePtr, size_t>> &outer_nodes) const; 53 void ChangeKernelBuildInfo(const AnfNodePtr &composite_node, 54 const std::vector<std::tuple<InplaceAssignerInfo, AnfNodePtr, size_t>> &outer_infos) const; 55 void ProcessTsa(const KernelGraphPtr &main_graph, const AnfNodePtr &anf_node, 56 const std::vector<InplaceAssignerInfo> &atomic_add_infos, const FuncGraphManagerPtr &mng) const; 57 std::pair<AnfNodePtr, size_t> GetOrCreateNewTsaFirstNode(const KernelGraphPtr &main_graph, 58 const InplaceAssignerInfo &atomic_add_info, 59 const AnfNodePtr &node) const; 60 std::pair<AnfNodePtr, size_t> FindTsaFirstRealInputInGraph(const KernelGraphPtr &, const CNodePtr &tsa_node, 61 const AnfNodePtr &node) const; 62 63 size_t tsa_first_input_index_{0}; // sub-graph parameter index. 64 }; 65 using TsaAtomicAddToFirstTensorPtr = std::shared_ptr<TsaAtomicAddToFirstTensor>; 66 } // namespace mindspore::graphkernel 67 68 #endif // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_TSA_ATOMIC_ADD_TO_FIRST_TENSOR_H_ 69