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.

task_context.h 4.1 kB

4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121
  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 GE_HYBRID_KERNEL_TASK_CONTEXT_H_
  17. #define GE_HYBRID_KERNEL_TASK_CONTEXT_H_
  18. #include <map>
  19. #include <mutex>
  20. #include <vector>
  21. #include "common/properties_manager.h"
  22. #include "external/ge/ge_api_error_codes.h"
  23. #include "hybrid/common/tensor_value.h"
  24. #include "hybrid/common/npu_memory_allocator.h"
  25. #include "hybrid/executor/rt_callback_manager.h"
  26. #include "hybrid/model/node_item.h"
  27. namespace ge {
  28. namespace hybrid {
  29. class GraphExecutionContext;
  30. class SubgraphContext;
  31. class TaskContext {
  32. public:
  33. static std::unique_ptr<TaskContext> Create(const NodeItem &node_item,
  34. GraphExecutionContext *execution_context,
  35. SubgraphContext *subgraph_context);
  36. ~TaskContext();
  37. int NumInputs() const;
  38. int NumOutputs() const;
  39. size_t NumWorkspaces() const;
  40. const NodeItem &GetNodeItem() const;
  41. const char *GetNodeName() const;
  42. TensorValue *MutableInput(int index);
  43. ConstGeTensorDescPtr GetInputDesc(int index) const;
  44. ConstGeTensorDescPtr GetOutputDesc(int index) const;
  45. GeTensorDescPtr MutableInputDesc(int index) const;
  46. GeTensorDescPtr MutableOutputDesc(int index) const;
  47. void ReleaseInput(int index);
  48. const TensorValue *GetInput(int index) const;
  49. const TensorValue *GetOutput(int index) const;
  50. TensorValue *MutableOutput(int index);
  51. TensorValue *GetVariable(const std::string &name);
  52. rtStream_t GetStream();
  53. int64_t GetSessionId() const;
  54. uint64_t GetIterationNumber() const;
  55. void NodeDone();
  56. void OnError(Status error);
  57. Status SetOutput(int index, const TensorValue &tensor);
  58. Status AllocateOutput(int index,
  59. const GeTensorDesc &tensor_desc,
  60. TensorValue **tensor,
  61. AllocationAttr *attr = nullptr);
  62. Status AllocateOutputs(AllocationAttr *attr = nullptr);
  63. Status AllocateWorkspaces();
  64. Status AllocateWorkspace(size_t size, void **buffer, void *ori_addr = nullptr);
  65. bool IsTraceEnabled() const;
  66. bool IsDumpEnabled() const;
  67. const DumpProperties& GetDumpProperties() const;
  68. const GraphExecutionContext *GetExecutionContext() {
  69. return execution_context_;
  70. }
  71. Status AllocateTensor(size_t size, TensorValue &tensor, AllocationAttr *attr = nullptr);
  72. void *MutableWorkspace(int index);
  73. const void *GetVarBaseAddr();
  74. Status RegisterCallback(const std::function<void()> &callback_fun) const;
  75. Status TryExecuteCallback(const std::function<void()> &callback_fun) const;
  76. Status PropagateOutputs();
  77. Status GetStatus() const;
  78. void SetStatus(Status status);
  79. bool IsForceInferShape() const;
  80. void SetForceInferShape(bool force_infer_shape);
  81. void *handle_ = nullptr;
  82. private:
  83. TaskContext(GraphExecutionContext *execution_context,
  84. const NodeItem *node_item,
  85. SubgraphContext *subgraph_context);
  86. static string TensorDesc2String(const GeTensorDesc &desc);
  87. Status AllocateTensor(const GeTensorDesc &tensor_desc, TensorValue &tensor, AllocationAttr *attr);
  88. const NodeItem *node_item_ = nullptr;
  89. bool force_infer_shape_ = false;
  90. GraphExecutionContext *execution_context_;
  91. SubgraphContext *subgraph_context_;
  92. TensorValue *inputs_start_ = nullptr;
  93. TensorValue *outputs_start_ = nullptr;
  94. Status status_ = SUCCESS;
  95. std::vector<void *> workspaces_;
  96. uint64_t iteration_ = 0;
  97. };
  98. } // namespace hybrid
  99. } // namespace ge
  100. #endif // GE_HYBRID_KERNEL_TASK_CONTEXT_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