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 3.1 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114
  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 "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<const NodeItem *> &GraphItem::GetInputNodes() const {
  40. return input_nodes_;
  41. }
  42. Status GraphItem::GetOutputDescList(vector<ConstGeTensorDescPtr> &output_desc_list) const {
  43. if (output_node_ == nullptr) {
  44. return SUCCESS;
  45. }
  46. if (is_dynamic_) {
  47. for (auto &tensor_desc : output_node_->GetOpDesc()->GetAllInputsDescPtr()) {
  48. output_desc_list.emplace_back(tensor_desc);
  49. }
  50. } else {
  51. for (auto &tensor_desc : output_node_->GetOpDesc()->GetAllOutputsDescPtr()) {
  52. output_desc_list.emplace_back(tensor_desc);
  53. }
  54. }
  55. return SUCCESS;
  56. }
  57. bool GraphItem::IsDynamic() const {
  58. return is_dynamic_;
  59. }
  60. const vector<int> &GraphItem::GetInputIndexMapping() const {
  61. return input_index_mapping_;
  62. }
  63. int GraphItem::GetParentOutputIndex(size_t index) const {
  64. if (index >= output_index_mapping_.size()) {
  65. return kInvalidIndex;
  66. }
  67. return output_index_mapping_[index];
  68. }
  69. const NodeItem *GraphItem::GetOutputNode() const {
  70. return output_node_;
  71. }
  72. const vector<std::pair<const NodeItem *, int>> &GraphItem::GetOutputEdges() const {
  73. return output_edges_;
  74. }
  75. Status GraphItem::GroupNodes() {
  76. int last_group = INT32_MIN;
  77. std::set<int> seen_groups;
  78. for (auto node : node_items_) {
  79. int group = node->group;
  80. if (group != last_group) {
  81. if (seen_groups.find(group) != seen_groups.end()) {
  82. GELOGE(INTERNAL_ERROR, "Unordered node group found. node = %s, group = %d", node->NodeName().c_str(), group);
  83. return INTERNAL_ERROR;
  84. } else {
  85. last_group = group;
  86. seen_groups.insert(group);
  87. grouped_node_items_.emplace_back(std::vector<NodeItem *>());
  88. }
  89. }
  90. GELOGD("Adding node [%s] to group %d", node->NodeName().c_str(), group);
  91. grouped_node_items_.back().emplace_back(node);
  92. }
  93. return SUCCESS;
  94. }
  95. } // namespace hybrid
  96. } // 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