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

4 years ago
4 years ago
4 years ago
4 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182
  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. Status MarkAgnosticPass::Run(ComputeGraphPtr graph) {
  21. for (const auto &node : graph->GetDirectNode()) {
  22. auto node_type = NodeUtils::GetNodeType(*node);
  23. if (node_type == SWITCH || node_type == SWITCHN) {
  24. GELOGD("Mark format agnostic and continuous for switch node %s", node->GetName().c_str());
  25. const OpDescPtr op_desc = node->GetOpDesc();
  26. const GeTensorDescPtr op_tensor = op_desc->MutableInputDesc(0);
  27. if (op_tensor == nullptr) {
  28. GELOGD("Op: %s, Index:0,has no input", node->GetName().c_str());
  29. continue;
  30. }
  31. AttrUtils::SetInt(op_tensor, "_format_continuous", 1);
  32. AttrUtils::SetInt(node->GetOpDesc(), "_format_agnostic", 1);
  33. AttrUtils::SetListInt(node->GetOpDesc(), "_format_agnostic_except_input", std::vector<int64_t>({1}));
  34. continue;
  35. }
  36. if (node_type == IDENTITY) {
  37. GELOGD("Mark format agnostic for identity node %s", node->GetName().c_str());
  38. AttrUtils::SetInt(node->GetOpDesc(), "_format_agnostic", 1);
  39. continue;
  40. }
  41. if (node_type == REFMERGE || node_type == REFSWITCH) {
  42. GELOGD("Mark format agnostic for regmerge and refswitch node %s", node->GetName().c_str());
  43. AttrUtils::SetInt(node->GetOpDesc(), "_format_agnostic", 1);
  44. AttrUtils::SetListInt(node->GetOpDesc(), "_format_agnostic_except_input", std::vector<int64_t>({1}));
  45. continue;
  46. }
  47. if (node_type == MERGE) {
  48. GELOGD("Mark format agnostic and continuous for merge node %s", node->GetName().c_str());
  49. const auto &input_nodes = node->GetInAllNodes();
  50. /// Enter-----------+
  51. /// +-> Merge
  52. /// NextIteration---+
  53. if (input_nodes.size() == 2) {
  54. if (input_nodes.at(0)->GetType() == ENTER && input_nodes.at(1)->GetType() == NEXTITERATION) {
  55. continue;
  56. }
  57. }
  58. const OpDescPtr op_desc = node->GetOpDesc();
  59. const GeTensorDescPtr op_tensor = op_desc->MutableOutputDesc(0);
  60. if (op_tensor == nullptr) {
  61. GELOGD("Op: %s, Index:0,has no output", node->GetName().c_str());
  62. continue;
  63. }
  64. AttrUtils::SetInt(op_tensor, "_format_continuous", 1);
  65. // Merge----------->NetOutput only set format_cofntinuous attr
  66. const auto &output_nodes = node->GetOutAllNodes();
  67. if (output_nodes.size() > 0) {
  68. if (output_nodes.at(0)->GetType() == NETOUTPUT) {
  69. continue;
  70. }
  71. }
  72. AttrUtils::SetInt(node->GetOpDesc(), "_format_agnostic", 1);
  73. AttrUtils::SetListInt(node->GetOpDesc(), "_format_agnostic_except_output", std::vector<int64_t>({1}));
  74. continue;
  75. }
  76. }
  77. return SUCCESS;
  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