1 /** 2 * Copyright 2021 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_FL_SERVER_KERNEL_AGGREGATION_KERNEL_H_ 18 #define MINDSPORE_CCSRC_FL_SERVER_KERNEL_AGGREGATION_KERNEL_H_ 19 20 #include <memory> 21 #include <string> 22 #include <vector> 23 #include "backend/kernel_compiler/cpu/cpu_kernel.h" 24 #include "fl/server/common.h" 25 #include "fl/server/memory_register.h" 26 #include "fl/server/kernel/params_info.h" 27 28 namespace mindspore { 29 namespace fl { 30 namespace server { 31 namespace kernel { 32 // AggregationKernel is the kernel for weight, grad or other kinds of parameters' aggregation. 33 // For example, dense gradients accumulation, federated average, etc. 34 // Normally the aggregation process in AggregationKernel is like a finite-state machine: 35 // Initial->Aggregating->Aggregation done->Initial. 36 class AggregationKernel : public CPUKernel { 37 public: AggregationKernel()38 AggregationKernel() : name_(""), done_(false), done_count_(0), accum_count_(0) {} 39 virtual ~AggregationKernel() = default; 40 41 // InitKernel and Launch methods are inherited from pure virtual function of CPUKernel so it must have implementation. InitKernel(const CNodePtr & kernel_node)42 virtual void InitKernel(const CNodePtr &kernel_node) {} Launch(const std::vector<AddressPtr> & inputs,const std::vector<AddressPtr> & workspace,const std::vector<AddressPtr> & outputs)43 virtual bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace, 44 const std::vector<AddressPtr> &outputs) { 45 return true; 46 } 47 48 // Server kernel's memory allocation method, which is different from the workflow in 49 // Session(GPUSession/CPUSession/AscendSession). 50 // virtual void AssignMemory(const CNodePtr &kernel_node, std::shared_ptr<MemoryRegister> memory_register) = 0; 51 52 // Set the cumulative count this aggregation kernel needs before aggregation is done. set_done_count(size_t count)53 void set_done_count(size_t count) { done_count_ = count; } 54 55 // So we use Reset to set the finite-state machine state to Initial after considering this round of aggregation is 56 // done. 57 virtual void Reset() = 0; 58 59 virtual bool IsAggregationDone() = 0; 60 61 // Some kernels should know the inputs/workspace/outputs addresses at initializing phase. For example, FedAvgKernel. SetParameterAddress(const std::vector<AddressPtr> & inputs,const std::vector<AddressPtr> & workspace,const std::vector<AddressPtr> & outputs)62 virtual void SetParameterAddress(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace, 63 const std::vector<AddressPtr> &outputs) { 64 return; 65 } 66 67 // Reinitialize aggregation kernel after scaling operations are done. ReInitForScaling()68 virtual bool ReInitForScaling() { return true; } 69 ReInitForUpdatingHyperParams(size_t)70 virtual bool ReInitForUpdatingHyperParams(size_t) { return true; } 71 72 // Setter and getter of kernels parameters information. set_params_info(const ParamsInfo & params_info)73 void set_params_info(const ParamsInfo ¶ms_info) { params_info_ = params_info; } input_names()74 const std::vector<std::string> &input_names() { return params_info_.inputs_names(); } workspace_names()75 const std::vector<std::string> &workspace_names() { return params_info_.workspace_names(); } output_names()76 const std::vector<std::string> &output_names() { return params_info_.outputs_names(); } 77 78 // Returns information about whether some inputs should reuse kernel node inputs memory. reuse_kernel_node_inputs_info()79 const ReuseKernelNodeInfo &reuse_kernel_node_inputs_info() { return reuse_kernel_node_inputs_info_; } 80 81 protected: 82 virtual void GenerateReuseKernelNodeInfo() = 0; 83 // Aggregation kernel's name which is set by kernel register function. 84 std::string name_; 85 86 // The aggregation is considered done after done_count_ times of accumulation. 87 bool done_; 88 89 // Cumulative count this aggregation kernel needs before aggregation is done. 90 size_t done_count_; 91 92 // Current cumulative count. 93 size_t accum_count_; 94 95 // Parameters information used for kernel register, memory assignment, etc. 96 ParamsInfo params_info_; 97 98 // Information about server kernel reusing kernel node inputs memory from the front end. 99 // Key refers to the server kernel's input index. Value refers to the kernel node's input index. 100 ReuseKernelNodeInfo reuse_kernel_node_inputs_info_; 101 }; 102 } // namespace kernel 103 } // namespace server 104 } // namespace fl 105 } // namespace mindspore 106 #endif // MINDSPORE_CCSRC_FL_SERVER_KERNEL_AGGREGATION_KERNEL_H_ 107