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.

graph_item.cc 4.1 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146
  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 "framework/common/util.h"
  17. #include "hybrid/model/graph_item.h"
  18. namespace ge {
  19. namespace hybrid {
  20. namespace {
  21. constexpr int kInvalidIndex = -1;
  22. } // namespace
  23. GraphItem::~GraphItem() {
  24. GELOGD("[%s] GraphItem destroyed.", name_.c_str());
  25. }
  26. const vector<NodeItem *> &hybrid::GraphItem::GetAllNodes() const {
  27. return node_items_;
  28. }
  29. const vector<NodeItem *> &GraphItem::GetAllNodes(int group) const {
  30. if (group == -1) {
  31. return GetAllNodes();
  32. }
  33. if (group >= static_cast<int>(grouped_node_items_.size())) {
  34. static vector<NodeItem *> empty_nodes;
  35. return empty_nodes;
  36. }
  37. return grouped_node_items_[group];
  38. }
  39. const vector<NodeItem *> &GraphItem::GetRootNodes(int group) const {
  40. if (group == -1) {
  41. return root_items_;
  42. }
  43. if (static_cast<uint32_t>(group) >= grouped_root_items_.size()) {
  44. static vector<NodeItem *> empty_nodes;
  45. return empty_nodes;
  46. }
  47. return grouped_root_items_[group];
  48. }
  49. size_t GraphItem::GetNodeSize(int group) const {
  50. if (group == -1) {
  51. return node_items_.size();
  52. }
  53. return (static_cast<uint32_t>(group) < grouped_node_items_.size()) ? grouped_node_items_[group].size() : 0;
  54. }
  55. const vector<const NodeItem *> &GraphItem::GetInputNodes() const {
  56. return input_nodes_;
  57. }
  58. Status GraphItem::GetOutputDescList(vector<ConstGeTensorDescPtr> &output_desc_list) const {
  59. if (output_node_ == nullptr) {
  60. return SUCCESS;
  61. }
  62. if (is_dynamic_) {
  63. for (auto &tensor_desc : output_node_->GetOpDesc()->GetAllInputsDescPtr()) {
  64. output_desc_list.emplace_back(tensor_desc);
  65. }
  66. } else {
  67. for (auto &tensor_desc : output_node_->GetOpDesc()->GetAllOutputsDescPtr()) {
  68. output_desc_list.emplace_back(tensor_desc);
  69. }
  70. }
  71. return SUCCESS;
  72. }
  73. bool GraphItem::IsDynamic() const {
  74. return is_dynamic_;
  75. }
  76. const vector<int> &GraphItem::GetInputIndexMapping() const {
  77. return input_index_mapping_;
  78. }
  79. int GraphItem::GetParentOutputIndex(size_t index) const {
  80. if (index >= output_index_mapping_.size()) {
  81. return kInvalidIndex;
  82. }
  83. return output_index_mapping_[index];
  84. }
  85. const NodeItem *GraphItem::GetOutputNode() const {
  86. return output_node_;
  87. }
  88. const vector<std::pair<const NodeItem *, int>> &GraphItem::GetOutputEdges() const {
  89. return output_edges_;
  90. }
  91. Status GraphItem::GroupNodes(const std::vector<NodeItem *> &node_items,
  92. std::vector<std::vector<NodeItem *>> &grouped_node_items) const {
  93. int curr_group = 0;
  94. int last_group = INT32_MIN;
  95. std::set<int> seen_groups;
  96. for (auto node : node_items) {
  97. int group = node->group;
  98. if (group != last_group) {
  99. if (seen_groups.find(group) != seen_groups.end()) {
  100. GELOGE(INTERNAL_ERROR,
  101. "[Find][Group]Unordered node group found. node = %s, group = %d", node->NodeName().c_str(), group);
  102. return INTERNAL_ERROR;
  103. } else {
  104. last_group = group;
  105. seen_groups.insert(group);
  106. curr_group = static_cast<int>(grouped_node_items.size());
  107. grouped_node_items.emplace_back(std::vector<NodeItem *>());
  108. }
  109. }
  110. node->group = curr_group;
  111. GELOGD("Adding node [%s] to group %d", node->NodeName().c_str(), node->group);
  112. grouped_node_items.back().emplace_back(node);
  113. }
  114. return SUCCESS;
  115. }
  116. Status GraphItem::GroupNodes() {
  117. GE_CHK_STATUS_RET_NOLOG(GroupNodes(node_items_, grouped_node_items_));
  118. GE_CHK_STATUS_RET_NOLOG(GroupNodes(root_items_, grouped_root_items_));
  119. return SUCCESS;
  120. }
  121. } // namespace hybrid
  122. } // 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