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.

variable_accelerate_ctrl.cc 3.4 kB

5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788
  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 "graph/manager/util/variable_accelerate_ctrl.h"
  17. #include "framework/common/debug/ge_log.h"
  18. #include "framework/common/ge_inner_error_codes.h"
  19. #include "framework/common/types.h"
  20. namespace ge {
  21. namespace {
  22. inline bool IsVariable(const std::string &node_type) {
  23. return node_type == VARIABLE || node_type == VARIABLEV2 || node_type == VARHANDLEOP;
  24. }
  25. }
  26. bool VarAccelerateCtrl::IsVarPermitToChangeFormats(const std::string &var_name) {
  27. std::lock_guard<std::mutex> lock(mutex_);
  28. auto iter = var_names_to_change_times_.find(var_name);
  29. if (iter == var_names_to_change_times_.end()) {
  30. return true;
  31. }
  32. return iter->second < kMaxVarChangeTimes_;
  33. }
  34. void VarAccelerateCtrl::SetVarChanged(const std::string &var_name) {
  35. std::lock_guard<std::mutex> lock(mutex_);
  36. auto times = ++var_names_to_change_times_[var_name];
  37. for (auto &graph_id_to_var_names : graph_ids_to_var_names_) {
  38. if (graph_id_to_var_names.second.count(var_name) > 0) {
  39. GELOGI("The format of var %s has been changed, total changed times %d, "
  40. "the graph %u contains which should be re-build before next run",
  41. var_name.c_str(), times, graph_id_to_var_names.first);
  42. /// The graph being compiled right now is also added to the rebuild-list
  43. /// and can be deleted by calling `SetGraphBuildEnd` at the end of compilation.
  44. graph_ids_need_rebuild_.insert(graph_id_to_var_names.first);
  45. }
  46. }
  47. }
  48. void VarAccelerateCtrl::AddGraph(uint32_t graph_id, const ComputeGraphPtr &compute_graph) {
  49. std::lock_guard<std::mutex> lock(mutex_);
  50. if (compute_graph == nullptr) {
  51. GELOGE(PARAM_INVALID, "Failed to add graph %u, the compute graph is null", graph_id);
  52. return;
  53. }
  54. auto &var_names = graph_ids_to_var_names_[graph_id];
  55. for (auto &node : compute_graph->GetAllNodes()) {
  56. auto node_type = node->GetType();
  57. if (IsVariable(node_type)) {
  58. GELOGD("Add graph %u contains variable %s", graph_id, node->GetName().c_str());
  59. var_names.insert(node->GetName());
  60. }
  61. }
  62. GELOGD("Add graph %u, var count %zu", graph_id, var_names.size());
  63. }
  64. void VarAccelerateCtrl::RemoveGraph(uint32_t graph_id) {
  65. std::lock_guard<std::mutex> lock(mutex_);
  66. GELOGD("Remove graph %u", graph_id);
  67. graph_ids_to_var_names_.erase(graph_id);
  68. graph_ids_need_rebuild_.erase(graph_id);
  69. }
  70. bool VarAccelerateCtrl::IsGraphNeedRebuild(uint32_t graph_id) const {
  71. std::lock_guard<std::mutex> lock(mutex_);
  72. return graph_ids_need_rebuild_.count(graph_id) > 0;
  73. }
  74. void VarAccelerateCtrl::SetGraphBuildEnd(uint32_t graph_id) {
  75. std::lock_guard<std::mutex> lock(mutex_);
  76. graph_ids_need_rebuild_.erase(graph_id);
  77. GELOGD("The graph %u has built end, remove it from the rebuild-set", graph_id);
  78. }
  79. } // 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