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.

reshape_recovery_pass.cc 3.5 kB

4 years ago
4 years ago
4 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788
  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/reshape_recovery_pass.h"
  17. #include "common/ge/ge_util.h"
  18. namespace ge {
  19. namespace {
  20. NodePtr CreateReshape(const ConstGeTensorDescPtr &src, const ConstGeTensorDescPtr &dst, const ComputeGraphPtr &graph) {
  21. static std::atomic_long reshape_num(0);
  22. auto next_num = reshape_num.fetch_add(1);
  23. auto reshape = MakeShared<OpDesc>("Reshape_ReshapeRecoveryPass_" + std::to_string(next_num), RESHAPE);
  24. if (reshape == nullptr) {
  25. return nullptr;
  26. }
  27. auto ret = reshape->AddInputDesc("x", *src);
  28. if (ret != GRAPH_SUCCESS) {
  29. return nullptr;
  30. }
  31. ret = reshape->AddInputDesc("shape", GeTensorDesc(GeShape(), Format(), DT_INT32));
  32. if (ret != GRAPH_SUCCESS) {
  33. return nullptr;
  34. }
  35. ret = reshape->AddOutputDesc("y", *dst);
  36. if (ret != GRAPH_SUCCESS) {
  37. return nullptr;
  38. }
  39. return graph->AddNode(reshape);
  40. }
  41. Status InsertReshapeIfNeed(const NodePtr &node) {
  42. GE_CHECK_NOTNULL(node);
  43. GE_CHECK_NOTNULL(node->GetOpDesc());
  44. for (auto src_anchor : node->GetAllOutDataAnchors()) {
  45. auto src_tensor = node->GetOpDesc()->GetOutputDescPtr(src_anchor->GetIdx());
  46. GE_CHECK_NOTNULL(src_tensor);
  47. for (auto dst_anchor : src_anchor->GetPeerInDataAnchors()) {
  48. auto dst_node = dst_anchor->GetOwnerNode();
  49. GELOGD("Try insert reshape between %s[%d] and %s[%d] to keep the shape continues",
  50. node->GetName().c_str(), src_anchor->GetIdx(), dst_node->GetName().c_str(), dst_anchor->GetIdx());
  51. GE_CHECK_NOTNULL(dst_node);
  52. GE_CHECK_NOTNULL(dst_node->GetOpDesc());
  53. auto dst_tensor = dst_node->GetOpDesc()->GetInputDescPtr(dst_anchor->GetIdx());
  54. GE_CHECK_NOTNULL(dst_tensor);
  55. bool is_need_insert_reshape = src_tensor->GetShape().GetDims() != UNKNOWN_RANK &&
  56. dst_tensor->GetShape().GetDims() != UNKNOWN_RANK &&
  57. src_tensor->GetShape().GetDims() != dst_tensor->GetShape().GetDims();
  58. if (is_need_insert_reshape) {
  59. auto reshape = CreateReshape(src_tensor, dst_tensor, node->GetOwnerComputeGraph());
  60. GE_CHECK_NOTNULL(reshape);
  61. auto ret = GraphUtils::InsertNodeBetweenDataAnchors(src_anchor, dst_anchor, reshape);
  62. if (ret != GRAPH_SUCCESS) {
  63. GELOGE(INTERNAL_ERROR, "Failed to insert reshape between node %s and %s",
  64. node->GetName().c_str(), dst_node->GetName().c_str());
  65. return INTERNAL_ERROR;
  66. }
  67. GELOGI("Insert reshape between %s and %s to keep the shape continues",
  68. node->GetName().c_str(), dst_node->GetName().c_str());
  69. }
  70. }
  71. }
  72. return SUCCESS;
  73. }
  74. } // namespace
  75. Status ReshapeRecoveryPass::Run(ComputeGraphPtr graph) {
  76. for (const auto &node : graph->GetDirectNode()) {
  77. auto ret = InsertReshapeIfNeed(node);
  78. if (ret != SUCCESS) {
  79. return ret;
  80. }
  81. }
  82. return SUCCESS;
  83. }
  84. } // 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