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 4.7 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146
  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 <atomic>
  21. #include <mutex>
  22. #include <unordered_map>
  23. #include "graph/profiler.h"
  24. #include "external/ge/ge_api_types.h"
  25. #include "toolchain/prof_callback.h"
  26. namespace ge {
  27. namespace profiling {
  28. enum {
  29. kInferShape,
  30. kTiling,
  31. kUpdateShape,
  32. kConstPrepare,
  33. kInitHybridExecuteArgs,
  34. kInitInferShapeContext,
  35. kDestroyInferShapeContext,
  36. kResetSubgraphExecutor,
  37. kCommitInferShapeTask,
  38. kDeviceToHost,
  39. kPrepareTask,
  40. kLaunchTask,
  41. kCommitTilingTask,
  42. kAtomic,
  43. kKernelLaunchPrepare,
  44. kRtKernelLaunch,
  45. kOpExecute,
  46. kAllocMem,
  47. kCopyH2D,
  48. kProfilingIndexEnd
  49. };
  50. constexpr uint64_t kInvalidHashId = 0ULL;
  51. class ProfilingContext {
  52. public:
  53. static bool IsDumpToStdEnabled();
  54. static ProfilingContext &GetInstance();
  55. ProfilingContext();
  56. ~ProfilingContext();
  57. /*
  58. * 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。
  59. * 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。
  60. * 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。
  61. * 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),
  62. * 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
  63. */
  64. bool IsEnabled() const noexcept {
  65. return enabled_ && profiler_ != nullptr;
  66. }
  67. void SetEnable() noexcept {
  68. enabled_ = true;
  69. }
  70. void SetDisable() noexcept {
  71. enabled_ = false;
  72. }
  73. void RecordCurrentThread(int64_t element, int64_t event, EventType et) {
  74. if (IsEnabled()) {
  75. profiler_->RecordCurrentThread(element, event, et);
  76. }
  77. }
  78. const Profiler *GetProfiler() const {
  79. return profiler_.get();
  80. }
  81. void Dump(std::ostream &out_stream) const {
  82. if (IsEnabled()) {
  83. profiler_->Dump(out_stream);
  84. } else {
  85. out_stream << "Profiling not enable, skip to dump" << std::endl;
  86. }
  87. }
  88. void DumpToStdOut() const {
  89. Dump(std::cout);
  90. }
  91. void Reset() {
  92. if (IsEnabled()) {
  93. profiler_->Reset();
  94. }
  95. }
  96. int64_t RegisterString(const std::string &str);
  97. int64_t RegisterStringHash(const uint64_t hash_id, const std::string &str);
  98. void UpdateElementHashId(const MsprofReporterCallback reporter_callback);
  99. static Status QueryHashId(const MsprofReporterCallback reporter_callback, const std::string &src_str,
  100. uint64_t &hash_id);
  101. size_t GetRegisterStringNum() const {
  102. return strings_to_index_.size();
  103. }
  104. private:
  105. void UpdateHashByStr(const std::string &str, const uint64_t hash);
  106. void Init();
  107. private:
  108. bool enabled_;
  109. int64_t str_index_;
  110. std::unordered_map<std::string, int64_t> strings_to_index_;
  111. std::mutex strings_to_index_mutex_;
  112. std::unique_ptr<Profiler> profiler_;
  113. };
  114. class ScopeProfiler {
  115. public:
  116. ScopeProfiler(int64_t element, int64_t event) : element_(element), event_(event) {
  117. ProfilingContext::GetInstance().RecordCurrentThread(element_, event, EventType::kEventStart);
  118. }
  119. ~ScopeProfiler() {
  120. ProfilingContext::GetInstance().RecordCurrentThread(element_, event_, EventType::kEventEnd);
  121. }
  122. private:
  123. int64_t element_;
  124. int64_t event_;
  125. };
  126. } // namespace profiling
  127. } // namespace ge
  128. #define PROFILING_START(element, event) \
  129. profiling::ProfilingContext::GetInstance().RecordCurrentThread(element, event, profiling::EventType::kEventStart)
  130. #define PROFILING_END(element, event) \
  131. profiling::ProfilingContext::GetInstance().RecordCurrentThread(element, event, profiling::EventType::kEventEnd)
  132. #define PROFILING_SCOPE(element, event) profiling::ScopeProfiler profiler(element, event)
  133. #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