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.

mark_agnostic_pass.cc 3.4 kB

5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384
  1. /**
  2. * Copyright 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/passes/mark_agnostic_pass.h"
  17. #include "graph/utils/node_utils.h"
  18. #include "graph/utils/tensor_utils.h"
  19. namespace ge {
  20. const size_t kTwoInputNodesSize = 2;
  21. Status MarkAgnosticPass::Run(ComputeGraphPtr graph) {
  22. for (const auto &node : graph->GetDirectNode()) {
  23. auto node_type = NodeUtils::GetNodeType(*node);
  24. if (node_type == SWITCH || node_type == SWITCHN) {
  25. GELOGD("Mark format agnostic and continuous for switch node %s", node->GetName().c_str());
  26. const OpDescPtr op_desc = node->GetOpDesc();
  27. const GeTensorDescPtr op_tensor = op_desc->MutableInputDesc(0);
  28. if (op_tensor == nullptr) {
  29. GELOGD("Op: %s, Index:0,has no input", node->GetName().c_str());
  30. continue;
  31. }
  32. AttrUtils::SetInt(op_tensor, "_format_continuous", 1);
  33. AttrUtils::SetInt(node->GetOpDesc(), "_format_agnostic", 1);
  34. AttrUtils::SetListInt(node->GetOpDesc(), "_format_agnostic_except_input", std::vector<int64_t>({1}));
  35. continue;
  36. }
  37. if (node_type == IDENTITY) {
  38. GELOGD("Mark format agnostic for identity node %s", node->GetName().c_str());
  39. AttrUtils::SetInt(node->GetOpDesc(), "_format_agnostic", 1);
  40. continue;
  41. }
  42. if (node_type == REFMERGE || node_type == REFSWITCH) {
  43. GELOGD("Mark format agnostic for regmerge and refswitch node %s", node->GetName().c_str());
  44. AttrUtils::SetInt(node->GetOpDesc(), "_format_agnostic", 1);
  45. AttrUtils::SetListInt(node->GetOpDesc(), "_format_agnostic_except_input", std::vector<int64_t>({1}));
  46. continue;
  47. }
  48. if (node_type == MERGE) {
  49. GELOGD("Mark format agnostic and continuous for merge node %s", node->GetName().c_str());
  50. const auto &input_nodes = node->GetInAllNodes();
  51. /// Enter-----------+
  52. /// +-> Merge
  53. /// NextIteration---+
  54. if (input_nodes.size() == kTwoInputNodesSize) {
  55. if (input_nodes.at(0)->GetType() == ENTER && input_nodes.at(1)->GetType() == NEXTITERATION) {
  56. continue;
  57. }
  58. }
  59. const OpDescPtr op_desc = node->GetOpDesc();
  60. const GeTensorDescPtr op_tensor = op_desc->MutableOutputDesc(0);
  61. if (op_tensor == nullptr) {
  62. GELOGD("Op: %s, Index:0,has no output", node->GetName().c_str());
  63. continue;
  64. }
  65. AttrUtils::SetInt(op_tensor, "_format_continuous", 1);
  66. // Merge----------->NetOutput only set format_cofntinuous attr
  67. const auto &output_nodes = node->GetOutAllNodes();
  68. if (output_nodes.size() > 0) {
  69. if (output_nodes.at(0)->GetType() == NETOUTPUT) {
  70. continue;
  71. }
  72. }
  73. AttrUtils::SetInt(node->GetOpDesc(), "_format_agnostic", 1);
  74. AttrUtils::SetListInt(node->GetOpDesc(), "_format_agnostic_except_output", std::vector<int64_t>({1}));
  75. continue;
  76. }
  77. }
  78. return SUCCESS;
  79. }
  80. } // 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