You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

ops_kernel_info_store.h 3.3 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788
  1. /**
  2. * Copyright 2019-2020 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. #ifndef INC_COMMON_OPSKERNEL_OPS_KERNEL_INFO_STORE_H_
  17. #define INC_COMMON_OPSKERNEL_OPS_KERNEL_INFO_STORE_H_
  18. #include <iostream>
  19. #include <map>
  20. #include <string>
  21. #include <vector>
  22. #include "./ge_task_info.h"
  23. #include "./ops_kernel_info_types.h"
  24. #include "cce/aicpu_engine_struct.h"
  25. #include "cce/fwk_adpt_struct.h"
  26. #include "common/ge_inner_error_codes.h"
  27. #include "graph/node.h"
  28. #include "proto/task.pb.h"
  29. using std::map;
  30. using std::string;
  31. using std::to_string;
  32. using std::vector;
  33. namespace ge {
  34. class OpDesc;
  35. class OpsKernelInfoStore {
  36. public:
  37. OpsKernelInfoStore() {}
  38. virtual ~OpsKernelInfoStore() {}
  39. // initialize opsKernelInfoStore
  40. virtual Status Initialize(const map<string, string> &options) = 0; /*lint -e148*/
  41. // close opsKernelInfoStore
  42. virtual Status Finalize() = 0; /*lint -e148*/
  43. virtual Status CreateSession(const std::map<std::string, std::string> &session_options) { return SUCCESS; }
  44. virtual Status DestroySession(const std::map<std::string, std::string> &session_options) { return SUCCESS; }
  45. // get all opsKernelInfo
  46. virtual void GetAllOpsKernelInfo(map<string, OpInfo> &infos) const = 0;
  47. // whether the opsKernelInfoStore is supported based on the operator attribute
  48. virtual bool CheckSupported(const OpDescPtr &opDescPtr, std::string &un_supported_reason) const = 0;
  49. virtual bool CheckAccuracySupported(const OpDescPtr &opDescPtr, std::string &un_supported_reason,
  50. bool realQuery = false) const {
  51. return CheckSupported(opDescPtr, un_supported_reason);
  52. }
  53. // opsFlag opsFlag[0] indicates constant folding is supported or not
  54. virtual void opsFlagCheck(const ge::Node &node, std::string &opsFlag){};
  55. // memory allocation requirement
  56. virtual Status CalcOpRunningParam(Node &node) = 0; /*lint -e148*/
  57. // generate task for op。
  58. virtual Status GenerateTask(const Node &node, RunContext &context,
  59. std::vector<domi::TaskDef> &tasks) = 0; /*lint -e148*/
  60. // only call fe engine interface to compile single op
  61. virtual Status CompileOp(vector<ge::NodePtr> &node_vec) { return SUCCESS; }
  62. virtual Status CompileOpRun(vector<ge::NodePtr> &node_vec) { return SUCCESS; }
  63. // load task for op
  64. virtual Status LoadTask(GETaskInfo &task) { return SUCCESS; }
  65. // only call aicpu interface to generate task struct
  66. virtual Status GenSingleOpRunTask(const NodePtr &node, STR_FWK_OP_KERNEL &task, string &task_info) { return SUCCESS; }
  67. // only call aicpu interface to generate task struct
  68. virtual Status GenMemCopyTask(uint64_t count, STR_FWK_OP_KERNEL &task, string &task_info) { return SUCCESS; }
  69. };
  70. } // namespace ge
  71. #endif // INC_COMMON_OPSKERNEL_OPS_KERNEL_INFO_STORE_H_

图引擎模块(GE)是MindSpore的一个子模块,其代码由C++实现,位于前端模块ME和底层硬件之间,起到承接作用。图引擎模块以ME下发的图作为输入,然后进行一系列的深度图优化操作,最后输出一张可以在底层硬件上高效运行的图。GE针对昇腾AI处理器的硬件结构特点,做了特定的优化工作,以此来充分发挥出昇腾AI处理器的强大算力。在进行模型训练/推理时,GE会被自动调用而用户并不感知。GE主要由GE API和GE Core两部分组成,详细的架构图如下所示