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.

built_in_subscriber_definitions.h 2.8 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071
  1. /**
  2. * Copyright (c) Huawei Technologies Co., Ltd. 2022. All rights reserved.
  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_INC_FRAMEWORK_RUNTIME_SUBSCRIBER_BUILT_IN_SUBSCRIBER_DEFINITIONS_H_
  17. #define AIR_CXX_INC_FRAMEWORK_RUNTIME_SUBSCRIBER_BUILT_IN_SUBSCRIBER_DEFINITIONS_H_
  18. #include <type_traits>
  19. #include <vector>
  20. #include "framework/common/ge_visibility.h"
  21. namespace gert {
  22. constexpr size_t kProfilingDataCap = 10UL * 1024UL * 1024UL;
  23. constexpr size_t kInitSize = 10UL * 1024UL;
  24. constexpr size_t kModelStrIdx = 0UL;
  25. constexpr size_t kExecuteStrIdx = 1UL;
  26. constexpr size_t kRegStartIdx = 2UL;
  27. constexpr size_t kDouble = 2UL;
  28. enum class BuiltInSubscriberType { kProfiling, kDumper, kNum };
  29. enum class ProfilingType {
  30. kHost, // 打开Host侧调度的profiling
  31. kDevice,
  32. kGeHost, // 打开GE Host侧调度的profiling
  33. kSingleOpReg, // 单算子需要打开此开关开始register node name和kernel type
  34. kNum,
  35. kAll = kNum
  36. };
  37. static_assert(static_cast<size_t>(ProfilingType::kNum) < sizeof(uint64_t) * 8,
  38. "The max num of profiling type must less than the width of uint64");
  39. enum class DumpType { kDataDump, kExceptionDump, kExceptionSave, kNum, kAll = kNum };
  40. static_assert(static_cast<size_t>(DumpType::kNum) < sizeof(uint64_t) * 8,
  41. "The max num of dumper type must less than the width of uint64");
  42. class VISIBILITY_EXPORT BuiltInSubscriberUtil {
  43. public:
  44. template <typename T,
  45. typename std::enable_if<(std::is_same<T, ProfilingType>::value) || (std::is_same<T, DumpType>::value),
  46. int>::type = 0>
  47. constexpr static uint64_t EnableBit(T et) {
  48. return 1UL << static_cast<size_t>(et);
  49. }
  50. template <typename T,
  51. typename std::enable_if<(std::is_same<T, ProfilingType>::value) || (std::is_same<T, DumpType>::value),
  52. int>::type = 0>
  53. static uint64_t BuildEnableFlags(const std::vector<T> &enable_types) {
  54. uint64_t flag = 0UL;
  55. for (auto et : enable_types) {
  56. if (et == T::kAll) {
  57. return EnableBit(T::kNum) - 1UL;
  58. }
  59. flag |= EnableBit(et);
  60. }
  61. return flag;
  62. }
  63. };
  64. } // namespace gert
  65. #endif // AIR_CXX_INC_FRAMEWORK_RUNTIME_SUBSCRIBER_BUILT_IN_SUBSCRIBER_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