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.

hybrid_profiler.cc 3.3 kB

5 years ago
5 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697
  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. #include "hybrid_profiler.h"
  17. #include <iomanip>
  18. #include <iostream>
  19. #include <cstdarg>
  20. #include "framework/common/debug/ge_log.h"
  21. #include "securec.h"
  22. namespace ge {
  23. namespace hybrid {
  24. namespace {
  25. const int kMaxEvents = 1024 * 500;
  26. const int kEventDescMax = 512;
  27. const int kMaxEventTypes = 8;
  28. const int kIndent = 8;
  29. }
  30. HybridProfiler::HybridProfiler(): counter_(0) {
  31. Reset();
  32. }
  33. void HybridProfiler::RecordEvent(EventType event_type, const char *fmt, ...) {
  34. va_list args;
  35. va_start(args, fmt);
  36. char buf[kEventDescMax];
  37. if (vsnprintf_s(buf, kEventDescMax, kEventDescMax - 1, fmt, args) == -1) {
  38. GELOGE(FAILED, "[Parse][Param:fmt]Format %s failed.", fmt);
  39. REPORT_CALL_ERROR("E19999", "Parse Format %s failed when HybridProfiler %s.", fmt, __FUNCTION__);
  40. va_end(args);
  41. return;
  42. }
  43. va_end(args);
  44. auto index = counter_++;
  45. if (index >= static_cast<int>(events_.size())) {
  46. GELOGE(INTERNAL_ERROR,
  47. "[Check][Range]index out of range. index = %d, max event size = %zu", index, events_.size());
  48. REPORT_INNER_ERROR("E19999", "index out of range when HybridProfiler %s. index = %d, max event size = %zu",
  49. __FUNCTION__, index, events_.size());
  50. return;
  51. }
  52. auto &evt = events_[index];
  53. evt.timestamp = std::chrono::system_clock::now();
  54. evt.desc = std::string(buf);
  55. evt.event_type = event_type;
  56. }
  57. void HybridProfiler::Dump(std::ostream &output_stream) {
  58. if (events_.empty()) {
  59. return;
  60. }
  61. auto start_dump = std::chrono::system_clock::now();
  62. auto first_evt = events_[0];
  63. auto start = first_evt.timestamp;
  64. std::vector<decltype(start)> prev_timestamps;
  65. prev_timestamps.resize(kMaxEventTypes, start);
  66. for (int i = 0; i < counter_; ++i) {
  67. auto &evt = events_[i];
  68. auto elapsed = std::chrono::duration_cast<std::chrono::microseconds>(evt.timestamp - start).count();
  69. auto &prev_ts = prev_timestamps[evt.event_type];
  70. auto cost = std::chrono::duration_cast<std::chrono::microseconds>(evt.timestamp - prev_ts).count();
  71. prev_ts = evt.timestamp;
  72. output_stream << std::setw(kIndent) << elapsed << "\t\t" << cost << "\t\t" << evt.desc << std::endl;
  73. }
  74. auto end_dump = std::chrono::system_clock::now();
  75. auto elapsed_dump = std::chrono::duration_cast<std::chrono::microseconds>(end_dump - start).count();
  76. auto cost_dump = std::chrono::duration_cast<std::chrono::microseconds>(end_dump - start_dump).count();
  77. output_stream << std::setw(kIndent) << elapsed_dump << "\t\t" << cost_dump
  78. << "\t\t" << "[Dump profiling]" << std::endl;
  79. Reset();
  80. }
  81. void HybridProfiler::Reset() {
  82. counter_ = 0;
  83. events_.clear();
  84. events_.resize(kMaxEvents);
  85. }
  86. } // namespace hybrid
  87. } // namespace ge

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