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.

profiling_definitions.h 6.2 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188
  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. #ifndef AIR_CXX_PROFILING_DEFINITIONS_H
  17. #define AIR_CXX_PROFILING_DEFINITIONS_H
  18. #include <string>
  19. #include <iostream>
  20. #include <mutex>
  21. #include <unordered_map>
  22. #include "graph/profiler.h"
  23. #include "external/ge/ge_api_types.h"
  24. #include "toolchain/prof_callback.h"
  25. namespace ge {
  26. namespace profiling {
  27. enum {
  28. kAclCompileAndExecute,
  29. kAclMatchOpModel,
  30. kAclMatchStaticOpModel,
  31. kAclMatchDynamicOpModel,
  32. kAclExecuteAsync,
  33. kAclLoadSingleOp,
  34. kAclBuildOpModel,
  35. kInferShape,
  36. kTiling,
  37. kUpdateShape,
  38. kConstPrepare,
  39. kInitHybridExecuteArgs,
  40. kInitInferShapeContext,
  41. kDestroyInferShapeContext,
  42. kResetSubgraphExecutor,
  43. kCommitInferShapeTask,
  44. kDeviceToHost,
  45. kPrepareTask,
  46. kLaunchTask,
  47. kCommitTilingTask,
  48. kAtomic,
  49. kKernelLaunchPrepare,
  50. kRtKernelLaunch,
  51. kOpExecute,
  52. kAllocMem,
  53. kCopyH2D,
  54. kPrepareNode,
  55. kWaitForPrepareDone,
  56. kPropgateOutputs,
  57. kOnNodeDoneCallback,
  58. kValidateInputTensor,
  59. kAfterExecuted,
  60. kRtEventSychronize,
  61. kInferShapeWaitDependShape,
  62. kInferShapeWaitInputTensor,
  63. kInferShapeCallInferFunc,
  64. kInferShapePropgate,
  65. // v2 control node
  66. kSelectBranch,
  67. kExecuteSubGraph,
  68. kInitSubGraphExecutor,
  69. // Add new definitions here
  70. kProfilingIndexEnd
  71. };
  72. constexpr uint64_t kInvalidHashId = 0UL;
  73. class ProfilingContext {
  74. public:
  75. static bool IsDumpToStdEnabled();
  76. static ProfilingContext &GetInstance();
  77. ProfilingContext();
  78. ~ProfilingContext();
  79. /*
  80. * 还有一种思路是`IsEnabled`只判断profiler_是否为空指针,不再设置单独的enabled标记位,这样可以少一个标记位。
  81. * 但是这么做就意味着,profiler_实例在未使能profiling时,必须是空指针状态。
  82. * 为了性能考虑,profiling机制在编译和加载时,就会调用`RegisterString`,向profiler_注册字符串,后续执行时,只会使用注册好的index了。
  83. * 因此存在一种场景:编译时并未使能profiling(因为编译时间很长,使能profiling也无法真实反应执行时的耗时状态),
  84. * 因此编译时注册字符串的动作并没有生效。在执行时,动态的打开了profiling,这种场景下,执行时无法拿到注册后字符串
  85. */
  86. bool IsEnabled() const noexcept {
  87. return enabled_ && profiler_ != nullptr;
  88. }
  89. void SetEnable() noexcept {
  90. enabled_ = true;
  91. }
  92. void SetDisable() noexcept {
  93. enabled_ = false;
  94. }
  95. void RecordCurrentThread(const int64_t element, const int64_t event, const EventType et,
  96. const std::chrono::time_point<std::chrono::system_clock> time_point) {
  97. if (IsEnabled()) {
  98. profiler_->RecordCurrentThread(element, event, et, time_point);
  99. }
  100. }
  101. void RecordCurrentThread(const int64_t element, const int64_t event, const EventType et) {
  102. RecordCurrentThread(element, event, et, std::chrono::system_clock::now());
  103. }
  104. const Profiler *GetProfiler() const {
  105. return profiler_.get();
  106. }
  107. void Dump(std::ostream &out_stream) const {
  108. if (IsEnabled()) {
  109. profiler_->Dump(out_stream);
  110. } else {
  111. out_stream << "Profiling not enable, skip to dump" << std::endl;
  112. }
  113. }
  114. void DumpToStdOut() const {
  115. Dump(std::cout);
  116. }
  117. void Reset() {
  118. if (IsEnabled()) {
  119. profiler_->Reset();
  120. }
  121. }
  122. int64_t RegisterString(const std::string &str);
  123. int64_t RegisterStringHash(const uint64_t hash_id, const std::string &str);
  124. void UpdateElementHashId(const MsprofReporterCallback reporter_callback);
  125. static Status QueryHashId(const MsprofReporterCallback reporter_callback, const std::string &src_str,
  126. uint64_t &hash_id);
  127. size_t GetRegisterStringNum() const {
  128. return strings_to_index_.size();
  129. }
  130. void Init();
  131. private:
  132. void UpdateHashByStr(const std::string &str, const uint64_t hash);
  133. private:
  134. bool inited_;
  135. bool enabled_;
  136. int64_t str_index_;
  137. std::unordered_map<std::string, int64_t> strings_to_index_;
  138. std::mutex strings_to_index_mutex_;
  139. std::unique_ptr<Profiler> profiler_;
  140. };
  141. class ScopeProfiler {
  142. public:
  143. ScopeProfiler(const int64_t element, const int64_t event) : element_(element), event_(event) {
  144. if (ProfilingContext::GetInstance().IsEnabled()) {
  145. start_trace_ = std::chrono::system_clock::now();
  146. }
  147. }
  148. ~ScopeProfiler() {
  149. if (ProfilingContext::GetInstance().IsEnabled()) {
  150. ProfilingContext::GetInstance().RecordCurrentThread(element_, event_, EventType::kEventStart, start_trace_);
  151. ProfilingContext::GetInstance().RecordCurrentThread(element_, event_, EventType::kEventEnd);
  152. }
  153. }
  154. void SetElement(const int64_t element) {
  155. element_ = element;
  156. }
  157. private:
  158. std::chrono::time_point<std::chrono::system_clock> start_trace_;
  159. int64_t element_;
  160. int64_t event_;
  161. };
  162. } // namespace profiling
  163. } // namespace ge
  164. #define PROFILING_START(element, event) \
  165. ge::profiling::ProfilingContext::GetInstance().RecordCurrentThread((element), (event), \
  166. ge::profiling::EventType::kEventStart)
  167. #define PROFILING_END(element, event) \
  168. ge::profiling::ProfilingContext::GetInstance().RecordCurrentThread((element), (event), \
  169. ge::profiling::EventType::kEventEnd)
  170. #define PROFILING_SCOPE(element, event) ge::profiling::ScopeProfiler profiler((element), (event))
  171. #define PROFILING_SCOPE_ELEMENT(element) profiler.SetElement((element))
  172. #endif // AIR_CXX_PROFILING_DEFINITIONS_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