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.

merge_pass.cc 8.4 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236
  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/merge_pass.h"
  17. #include <memory>
  18. #include <string>
  19. #include <vector>
  20. #include "framework/common/debug/ge_log.h"
  21. #include "common/ge/ge_util.h"
  22. #include "graph/common/omg_util.h"
  23. #include "graph/debug/ge_attr_define.h"
  24. #include "graph/utils/graph_utils.h"
  25. #include "graph/passes/pass_utils.h"
  26. namespace ge {
  27. const int kValueIndexOutputIndex = 1;
  28. bool IsEmptyTensor(const GeShape &shape) {
  29. const auto &dims = shape.GetDims();
  30. return std::any_of(dims.begin(), dims.end(), [](int64_t dim) { return dim == 0; });
  31. }
  32. Status MergePass::Run(NodePtr &node) {
  33. GELOGD("MergePass running");
  34. if (node == nullptr) {
  35. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  36. return PARAM_INVALID;
  37. }
  38. std::string op_type;
  39. GE_CHK_STATUS_RET(GetOriginalType(node, op_type), "get original type failed");
  40. if (op_type != MERGE) {
  41. return SUCCESS;
  42. }
  43. if (node->GetAllOutDataAnchors().empty()) {
  44. GELOGE(PARAM_INVALID, "[%s] Merge node output anchor is empty", node->GetName().c_str());
  45. return PARAM_INVALID;
  46. }
  47. if (OptimizeEmptyTensorInput(node) != SUCCESS) {
  48. GELOGE(FAILED, "[%s] remove empty_tensor inputs failed.", node->GetName().c_str());
  49. return FAILED;
  50. }
  51. const auto &in_data_nodes = node->GetInDataNodes();
  52. switch (in_data_nodes.size()) {
  53. case 0: {
  54. /// Case A: input_count = 0, the output of merge node is inactive as well
  55. /// In which case the output branch can be removed
  56. /// until another merge node is met
  57. std::vector<NodePtr> del_nodes;
  58. std::vector<NodePtr> end_nodes;
  59. Status ret = PassUtils::RemoveBranch(node, del_nodes, end_nodes);
  60. for (auto &end_node : end_nodes) {
  61. AddRePassNode(end_node);
  62. }
  63. for (const auto &delete_node : del_nodes) {
  64. AddNodeDeleted(delete_node);
  65. }
  66. return ret;
  67. }
  68. case 1: { // Case B: input_count = 1, the merge node can be optimized out
  69. std::vector<int> merge_io_map = {PassUtils::GetUniqueInDataAnchorIndex(node), -1};
  70. if (merge_io_map[0] != -1 && IsNeedChangeIndexToConstant(node)) {
  71. int index = merge_io_map[0];
  72. if (ChangeIndexToConstant(node, index) != SUCCESS) {
  73. GELOGE(FAILED, "[%s] Change value index to be Constant failed.", node->GetName().c_str());
  74. return FAILED;
  75. }
  76. }
  77. auto in_node = in_data_nodes.at(0);
  78. if (IsMergeInputNeedOptimized(in_node)) {
  79. if (IsolateAndDeleteNode(in_node, {0}) != SUCCESS) {
  80. GELOGE(FAILED, "Isolate and delete node %s failed.", in_node->GetName().c_str());
  81. return FAILED;
  82. }
  83. }
  84. return IsolateAndDeleteNode(node, merge_io_map);
  85. }
  86. default: {
  87. // Case C: input_count > 1, the merge node can not be optimized
  88. return SUCCESS;
  89. }
  90. }
  91. }
  92. bool MergePass::IsNeedChangeIndexToConstant(NodePtr &node) const {
  93. /// value_index is the index 1 output of the Merge
  94. /// value_index link to other node, change it to be Constant
  95. GE_IF_BOOL_EXEC(node == nullptr, GELOGW("Node is nullptr"); return false);
  96. auto out_anchor = node->GetOutDataAnchor(kValueIndexOutputIndex);
  97. GE_IF_BOOL_EXEC(out_anchor == nullptr, GELOGW("Out_anchor is nullptr"); return false);
  98. for (const auto &peer_in_anchor : out_anchor->GetPeerInDataAnchors()) {
  99. if (peer_in_anchor != nullptr && peer_in_anchor->GetOwnerNode() != nullptr) {
  100. GELOGI(
  101. "[%s] MergePass, value_index link to other node, "
  102. "change it to be Constant.",
  103. node->GetName().c_str());
  104. return true;
  105. }
  106. }
  107. return false;
  108. }
  109. Status MergePass::ChangeIndexToConstant(NodePtr &node, int &value_index) {
  110. GE_CHECK_NOTNULL(node);
  111. ComputeGraphPtr graph = node->GetOwnerComputeGraph();
  112. if (graph == nullptr) {
  113. GELOGE(FAILED, "[%s] The owner graph must not be null.", node->GetName().c_str());
  114. return FAILED;
  115. }
  116. OpDescPtr constant_op_desc = nullptr;
  117. if (CreateConstByValue(node, value_index, constant_op_desc) != SUCCESS) {
  118. return FAILED;
  119. }
  120. NodePtr const_node = graph->AddNode(constant_op_desc);
  121. if (const_node == nullptr) {
  122. return FAILED;
  123. }
  124. // Change peer in anchors from value_index to new Constant node
  125. if (GraphUtils::ReplaceNodeAnchors(const_node, node, {}, {1}) != GRAPH_SUCCESS) {
  126. GELOGE(FAILED, "[%s] ReplaceNodeAnchors failed.", node->GetName().c_str());
  127. return FAILED;
  128. }
  129. auto out_control_anchor = node->GetOutControlAnchor();
  130. GE_CHECK_NOTNULL(out_control_anchor);
  131. // Add control anchor between Merge and Constant
  132. if (out_control_anchor->LinkTo(const_node->GetInControlAnchor()) != GRAPH_SUCCESS) {
  133. return FAILED;
  134. }
  135. return SUCCESS;
  136. }
  137. Status MergePass::CreateConstByValue(NodePtr &node, int value_index, OpDescPtr &op_desc) {
  138. std::string constant_name = node->GetName() + "_value_index";
  139. // 1. create Constant OpDesc
  140. op_desc = MakeShared<OpDesc>(constant_name, CONSTANT);
  141. if (op_desc == nullptr) {
  142. GELOGE(FAILED, "[%s] Make shared of Constant op desc failed.", constant_name.c_str());
  143. return FAILED;
  144. }
  145. // 2. get OpDesc of output number one of Merge(value_index)
  146. OpDescPtr original_op_desc = node->GetOpDesc();
  147. if (original_op_desc == nullptr) {
  148. GELOGE(FAILED, "[%s] Op desc must not be null.", constant_name.c_str());
  149. return FAILED;
  150. }
  151. GeTensorDesc original_out_tensor_desc = original_op_desc->GetOutputDesc(1);
  152. original_out_tensor_desc.SetDataType(DT_INT32);
  153. // 3. create attr value of Constant, is a tensor
  154. GeTensorPtr const_tensor_ptr =
  155. MakeShared<GeTensor>(original_out_tensor_desc, reinterpret_cast<uint8_t *>(&value_index), sizeof(int));
  156. if (const_tensor_ptr == nullptr) {
  157. GELOGE(FAILED, "[%s] Make shared of Constant tensor failed.", constant_name.c_str());
  158. return FAILED;
  159. }
  160. GE_IF_BOOL_EXEC(!AttrUtils::SetTensor(op_desc, ATTR_NAME_WEIGHTS, const_tensor_ptr),
  161. GELOGE(FAILED, "get ATTR_NAME_WEIGHTS failed"); return FAILED);
  162. // 4. set Constant output desc
  163. GE_CHK_STATUS_RET(op_desc->AddOutputDesc(original_out_tensor_desc), "add out put desc failed");
  164. return SUCCESS;
  165. }
  166. bool MergePass::IsMergeInputNeedOptimized(NodePtr &node) const {
  167. if (node == nullptr) {
  168. return false;
  169. }
  170. // node is not inserted by MergeInputMemcpyPass
  171. if ((node->GetType() != MEMCPYASYNC) && (node->GetType() != MEMCPYADDRASYNC)) {
  172. return false;
  173. }
  174. if (node->GetInDataNodes().size() != 1) {
  175. return false;
  176. }
  177. auto in_node = node->GetInDataNodes().at(0);
  178. if (in_node == nullptr) {
  179. return false;
  180. }
  181. // in_node may be global_step var
  182. if ((in_node->GetType() == VARIABLE) || (in_node->GetType() == VARIABLEV2)) {
  183. return false;
  184. }
  185. return true;
  186. }
  187. Status MergePass::OptimizeEmptyTensorInput(const NodePtr &node) {
  188. for (const auto &in_data_anchor : node->GetAllInDataAnchors()) {
  189. const auto &peer_data_anchor = in_data_anchor->GetPeerOutAnchor();
  190. if (peer_data_anchor == nullptr) {
  191. continue;
  192. }
  193. if ((peer_data_anchor->GetOwnerNode() == nullptr) ||
  194. (peer_data_anchor->GetOwnerNode()->GetOpDesc() == nullptr)) {
  195. continue;
  196. }
  197. const auto &op_desc = peer_data_anchor->GetOwnerNode()->GetOpDesc();
  198. if (IsEmptyTensor(op_desc->GetOutputDesc(peer_data_anchor->GetIdx()).GetShape())) {
  199. if (GraphUtils::RemoveEdge(peer_data_anchor, in_data_anchor) != GRAPH_SUCCESS) {
  200. GELOGE(FAILED, "Remove data edge %s:%d->%s:%d failed.",
  201. op_desc->GetName().c_str(), peer_data_anchor->GetIdx(),
  202. node->GetName().c_str(), in_data_anchor->GetIdx());
  203. return FAILED;
  204. }
  205. GELOGD("Remove data edge %s:%d->%s:%d",
  206. op_desc->GetName().c_str(), peer_data_anchor->GetIdx(),
  207. node->GetName().c_str(), in_data_anchor->GetIdx());
  208. }
  209. }
  210. return SUCCESS;
  211. }
  212. } // 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