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.

shape_inference_engine.cc 10 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211
  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 "hybrid/executor/worker/shape_inference_engine.h"
  17. #include "graph/shape_refiner.h"
  18. #include "graph/utils/node_utils.h"
  19. #include "hybrid/node_executor/node_executor.h"
  20. namespace ge {
  21. namespace hybrid {
  22. ShapeInferenceEngine::ShapeInferenceEngine(GraphExecutionContext *execution_context, SubgraphContext *subgraph_context)
  23. : execution_context_(execution_context), subgraph_context_(subgraph_context) {}
  24. Status ShapeInferenceEngine::InferShape(NodeState &node_state) {
  25. // Wait for all input shape become valid
  26. GE_CHK_STATUS_RET_NOLOG(node_state.GetShapeInferenceState().AwaitShapesReady(*execution_context_));
  27. auto &node_item = *node_state.GetNodeItem();
  28. if (node_item.is_output_shape_static) {
  29. return SUCCESS;
  30. }
  31. if (node_item.fused_subgraph != nullptr) {
  32. return InferShapeForSubgraph(node_item, *node_item.fused_subgraph);
  33. }
  34. // Skip shape inference for node of type DEPEND_COMPUTE
  35. if (node_item.shape_inference_type == DEPEND_COMPUTE) {
  36. GELOGD("[%s] Skipping node with unknown shape type DEPEND_COMPUTE", node_item.NodeName().c_str());
  37. return SUCCESS;
  38. }
  39. // Clear shape range in case shape inference func forgot to do it
  40. if (node_item.shape_inference_type == DEPEND_SHAPE_RANGE) {
  41. // in case InferFunc forgot to reset output shape
  42. for (auto &output_desc : node_item.op_desc->GetAllOutputsDescPtr()) {
  43. output_desc->SetShape(GeShape({UNKNOWN_DIM_NUM}));
  44. }
  45. }
  46. // Wait for "const input nodes" if node's shape inference function requires any.
  47. GE_CHK_STATUS_RET_NOLOG(AwaitDependentNodes(node_state));
  48. // Do shape inference
  49. GELOGD("[%s] Start to invoke InferShapeAndType", node_item.NodeName().c_str());
  50. {
  51. std::lock_guard<std::mutex> lk(mu_);
  52. RECORD_SHAPE_INFERENCE_EVENT(execution_context_, node_item.NodeName().c_str(), "[InferShapeAndType] Start");
  53. GE_CHK_STATUS_RET(ShapeRefiner::InferShapeAndType(node_item.node), "Invoke InferShapeAndType failed.");
  54. RECORD_SHAPE_INFERENCE_EVENT(execution_context_, node_item.NodeName().c_str(), "[InferShapeAndType] End");
  55. }
  56. // Check again to make sure shape is valid after shape inference
  57. if (node_item.shape_inference_type != DEPEND_SHAPE_RANGE) {
  58. bool is_unknown_shape = false;
  59. GE_CHK_STATUS_RET(NodeUtils::GetNodeUnknownShapeStatus(*node_item.node, is_unknown_shape),
  60. "Failed to get shape status. node = %s", node_item.NodeName().c_str());
  61. GE_CHK_BOOL_RET_STATUS(!is_unknown_shape, INTERNAL_ERROR, "[%s] Shape is still unknown after shape inference.",
  62. node_item.NodeName().c_str());
  63. }
  64. GELOGD("[%s] [HybridTrace] After shape inference. Node = %s", node_item.NodeName().c_str(),
  65. node_item.DebugString().c_str());
  66. GELOGD("[%s] InferShapeAndType finished successfully.", node_item.NodeName().c_str());
  67. return SUCCESS;
  68. }
  69. Status ShapeInferenceEngine::AwaitDependentNodes(NodeState &node_state) {
  70. auto &node_item = *node_state.GetNodeItem();
  71. for (auto &src_node : node_item.dependents_for_shape_inference) {
  72. GELOGI("[%s] Start to wait for data dependent node: %s", node_item.NodeName().c_str(), src_node->GetName().c_str());
  73. RECORD_SHAPE_INFERENCE_EVENT(execution_context_, node_item.NodeName().c_str(), "[AwaitNodeDone] [%s] Start",
  74. src_node->GetName().c_str());
  75. if (!subgraph_context_->Await(src_node)) {
  76. GELOGE(INTERNAL_ERROR, "[%s] Await node failed.", src_node->GetName().c_str());
  77. return INTERNAL_ERROR;
  78. }
  79. RECORD_SHAPE_INFERENCE_EVENT(execution_context_, node_item.NodeName().c_str(), "[AwaitNodeDone] [%s] End",
  80. src_node->GetName().c_str());
  81. GELOGI("[%s] Done waiting node.", src_node->GetName().c_str());
  82. }
  83. return SUCCESS;
  84. }
  85. Status ShapeInferenceEngine::PropagateOutputShapes(const NodeItem &node_item) {
  86. if (node_item.is_output_shape_static) {
  87. return SUCCESS;
  88. }
  89. // output shape will not be valid until compute is done.
  90. bool shape_is_future =
  91. node_item.shape_inference_type == DEPEND_SHAPE_RANGE || node_item.shape_inference_type == DEPEND_COMPUTE;
  92. GELOGD("[%s] Start to propagate output shapes. shape_type = %d", node_item.NodeName().c_str(),
  93. node_item.shape_inference_type);
  94. RECORD_SHAPE_INFERENCE_EVENT(execution_context_, node_item.NodeName().c_str(), "[PropagateOutputShapes] Start");
  95. // propagate each output
  96. for (int i = 0; i < node_item.num_outputs; ++i) {
  97. auto output_desc = node_item.op_desc->MutableOutputDesc(i);
  98. const auto &shape = output_desc->MutableShape();
  99. const auto &ori_shape = output_desc->GetOriginShape();
  100. auto &output_nodes = node_item.outputs[i];
  101. // propagate output to all sub-inputs
  102. for (auto &dst_input_index_and_node : output_nodes) {
  103. auto &dst_node_item = dst_input_index_and_node.second;
  104. auto dst_node_state = subgraph_context_->GetOrCreateNodeState(dst_node_item);
  105. GE_CHECK_NOTNULL(dst_node_state);
  106. GELOGI("[%s] Update dst node [%s], input index = %d", node_item.NodeName().c_str(),
  107. dst_node_item->NodeName().c_str(), dst_input_index_and_node.first);
  108. // in case type 3 and 4, shape will be valid after computing is done
  109. if (shape_is_future) {
  110. ShapeFuture future(node_item.node, i, subgraph_context_);
  111. dst_node_state->GetShapeInferenceState().UpdateInputShapeFuture(dst_input_index_and_node.first,
  112. std::move(future));
  113. } else {
  114. dst_node_state->GetShapeInferenceState().UpdateInputShape(dst_input_index_and_node.first, ori_shape, shape);
  115. }
  116. }
  117. }
  118. RECORD_SHAPE_INFERENCE_EVENT(execution_context_, node_item.NodeName().c_str(), "[PropagateOutputShapes] End");
  119. GELOGD("[%s] Propagating output shapes finished successfully.", node_item.NodeName().c_str());
  120. return SUCCESS;
  121. }
  122. Status ShapeInferenceEngine::InferShapeForSubgraph(const NodeItem &node_item, const FusedSubgraph &fused_subgraph) {
  123. GELOGD("[%s] Start to infer shape by fused subgraph", node_item.NodeName().c_str());
  124. for (auto &it : fused_subgraph.input_mapping) {
  125. auto parent_tensor_desc = node_item.op_desc->MutableInputDesc(it.first);
  126. GE_CHECK_NOTNULL(parent_tensor_desc);
  127. GELOGD("Start to update shape by input[%u]", it.first);
  128. GELOGD("Update shape to [%s]", parent_tensor_desc->GetShape().ToString().c_str());
  129. GELOGD("Update original shape to [%s]", parent_tensor_desc->GetOriginShape().ToString().c_str());
  130. for (auto &tensor_desc : it.second) {
  131. tensor_desc->SetShape(parent_tensor_desc->GetShape());
  132. tensor_desc->SetOriginShape(parent_tensor_desc->GetOriginShape());
  133. }
  134. }
  135. for (auto &node : fused_subgraph.nodes) {
  136. GELOGD("[%s] Start to invoke InferShapeAndType", node->GetName().c_str());
  137. GE_CHK_STATUS_RET(ShapeRefiner::InferShapeAndType(node));
  138. GELOGD("[%s] Done invoking InferShapeAndType", node->GetName().c_str());
  139. GE_CHK_STATUS_RET(UpdatePeerNodeShape(*node), "[%s] Failed to update shapes of peer node.",
  140. node->GetName().c_str());
  141. }
  142. for (auto &it : fused_subgraph.output_mapping) {
  143. uint32_t parent_output_idx = it.first;
  144. const auto &op_desc = it.second;
  145. GELOGD("Update parent output[%d] by [%s]", parent_output_idx, op_desc->GetName().c_str());
  146. auto input_desc = op_desc->MutableInputDesc(0);
  147. GE_CHECK_NOTNULL(input_desc);
  148. auto parent_output_tensor_desc = node_item.op_desc->MutableOutputDesc(parent_output_idx);
  149. GE_CHECK_NOTNULL(parent_output_tensor_desc);
  150. GELOGD("Update shape to [%s]", input_desc->GetShape().ToString().c_str());
  151. GELOGD("Update original shape to [%s]", input_desc->GetOriginShape().ToString().c_str());
  152. parent_output_tensor_desc->SetOriginShape(input_desc->GetOriginShape());
  153. parent_output_tensor_desc->SetShape(input_desc->GetShape());
  154. }
  155. GELOGD("[%s] Done shape inference by subgraph successfully.", node_item.NodeName().c_str());
  156. return SUCCESS;
  157. }
  158. Status ShapeInferenceEngine::UpdatePeerNodeShape(const Node &node) {
  159. auto op_desc = node.GetOpDesc();
  160. for (const auto &out_anchor : node.GetAllOutDataAnchors()) {
  161. auto output_tensor = op_desc->MutableOutputDesc(out_anchor->GetIdx());
  162. for (const auto &peer_anchor : out_anchor->GetPeerInDataAnchors()) {
  163. auto peer_node = peer_anchor->GetOwnerNode();
  164. GE_CHECK_NOTNULL(peer_node);
  165. auto peer_op_desc = peer_node->GetOpDesc();
  166. GE_CHECK_NOTNULL(peer_op_desc);
  167. auto peer_input_desc = peer_op_desc->MutableInputDesc(peer_anchor->GetIdx());
  168. if (peer_input_desc == nullptr) {
  169. GELOGE(GRAPH_FAILED, "peer_input_desc is nullptr");
  170. continue;
  171. }
  172. GELOGI("Peer input op desc name is %s, need to flush: shape size is %zu, datatype is %d, original datatype is %d",
  173. peer_anchor->GetOwnerNode()->GetOpDesc()->GetName().c_str(), output_tensor->GetShape().GetDimNum(),
  174. output_tensor->GetDataType(), output_tensor->GetOriginDataType());
  175. peer_input_desc->SetOriginShape(output_tensor->GetOriginShape());
  176. peer_input_desc->SetShape(output_tensor->GetShape());
  177. GELOGI("Peer input op desc name is %s, shape size is %zu, datatype is %d, original datatype is %d",
  178. peer_anchor->GetOwnerNode()->GetOpDesc()->GetName().c_str(), peer_input_desc->GetShape().GetDimNum(),
  179. peer_input_desc->GetDataType(), peer_input_desc->GetOriginDataType());
  180. }
  181. }
  182. return SUCCESS;
  183. }
  184. } // namespace hybrid
  185. } // 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