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.

net_output_pass.cc 34 kB

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
5 years ago
5 years ago
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
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 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
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 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
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 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
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742
  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/net_output_pass.h"
  17. #include <map>
  18. #include <memory>
  19. #include <string>
  20. #include <utility>
  21. #include <vector>
  22. #include "common/ge/ge_util.h"
  23. #include "framework/common/debug/ge_log.h"
  24. #include "framework/common/ge_inner_error_codes.h"
  25. #include "framework/omg/omg_inner_types.h"
  26. #include "graph/debug/ge_attr_define.h"
  27. #include "graph/common/local_context.h"
  28. #include "graph/passes/pass_utils.h"
  29. #include "graph/utils/tensor_utils.h"
  30. #include "graph/utils/type_utils.h"
  31. namespace ge {
  32. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  33. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  34. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  35. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  36. // the size of user defined output datatype or format string after split by ":".
  37. const size_t kUserDefinedElementCount = 2;
  38. const size_t kNodesCount = 2;
  39. Status NetOutputPass::GetRetvalOutputInfo(const ge::NodePtr &node,
  40. std::map<int32_t, RetvalInfo> &retval_node_index_map) {
  41. GE_CHECK_NOTNULL(node);
  42. GE_CHECK_NOTNULL(node->GetOpDesc());
  43. int64_t output_index = 0;
  44. if (!AttrUtils::GetInt(node->GetOpDesc(), RETVAL_ATTR_NAME_INDEX, output_index)) {
  45. REPORT_CALL_ERROR("E19999", "Get Attr:%s from op:%s(%s) failed", RETVAL_ATTR_NAME_INDEX.c_str(),
  46. node->GetName().c_str(), node->GetType().c_str());
  47. GELOGE(PARAM_INVALID, "Get output index failed.");
  48. return PARAM_INVALID;
  49. }
  50. if (retval_node_index_map.count(output_index) > 0) {
  51. REPORT_INNER_ERROR("E19999", "Attr:%s from op:%s(%s), value:%ld duplicate with other node, check invalid",
  52. RETVAL_ATTR_NAME_INDEX.c_str(), node->GetName().c_str(), node->GetType().c_str(), output_index);
  53. GELOGE(PARAM_INVALID, "Retval has duplicate index.");
  54. return PARAM_INVALID;
  55. }
  56. int parent_node_index = -1;
  57. (void)AttrUtils::GetInt(node->GetOpDesc(), ATTR_NAME_PARENT_NODE_INDEX, parent_node_index);
  58. InDataAnchorPtr in_data_anchor = node->GetInDataAnchor(0);
  59. GE_CHECK_NOTNULL(in_data_anchor);
  60. GE_CHECK_NOTNULL(in_data_anchor->GetPeerOutAnchor());
  61. int32_t src_node_index = in_data_anchor->GetPeerOutAnchor()->GetIdx();
  62. NodePtr src_node_ptr = in_data_anchor->GetPeerOutAnchor()->GetOwnerNode();
  63. retval_node_index_map[output_index] = {src_node_ptr, src_node_index, parent_node_index};
  64. // if user targets include retval node,delete it from set and insert its input node instead
  65. // better to GetInNodes here
  66. auto iter = targets_.find(node);
  67. if (iter != targets_.end()) {
  68. targets_.erase(iter);
  69. targets_.insert(src_node_ptr);
  70. GELOGI("node [%s] is in user def targets, do not output result to user!", node->GetName().c_str());
  71. }
  72. is_include_special_node_ = true;
  73. return SUCCESS;
  74. }
  75. Status NetOutputPass::GetOutputNode(const ge::ComputeGraphPtr &graph, std::vector<RetvalInfo> &output_nodes_info) {
  76. std::map<int32_t, RetvalInfo> retval_node_index_map;
  77. for (NodePtr &node : graph->GetDirectNode()) {
  78. Status ret = SUCCESS;
  79. if ((node->GetOpDesc() != nullptr) && (node->GetOpDesc()->HasAttr(RETVAL_ATTR_NAME_INDEX))) {
  80. /// Set the output according to the Retval operator,
  81. /// identify by whether there is an index parameter
  82. ret = GetRetvalOutputInfo(node, retval_node_index_map);
  83. }
  84. if (ret != SUCCESS) {
  85. GELOGE(ret, "GetRetvalOutputInfo failed");
  86. return ret;
  87. }
  88. }
  89. GELOGI("Get retval node size:%zu.", retval_node_index_map.size());
  90. std::vector<RetvalInfo> out_nodes_tmp;
  91. /// The Netoutput output is determined by Retval, and the input order
  92. /// of Netoutput is sorted according to the index value of Retval.
  93. for (auto &it : retval_node_index_map) {
  94. out_nodes_tmp.push_back(it.second);
  95. }
  96. // when user set targets, mean that no output result
  97. for (auto &ele : graph->GetGraphOutNodesInfo()) {
  98. auto iter = targets_.find(ele.first);
  99. if (iter != targets_.end()) {
  100. GELOGI("user set out node [%s] is found in user def targets, out node is prio!", ele.first->GetName().c_str());
  101. targets_.erase(iter);
  102. }
  103. auto op_desc = ele.first->GetOpDesc();
  104. GE_CHECK_NOTNULL(op_desc);
  105. if (op_desc->HasAttr(ATTR_ATC_USER_DEFINE_OUTPUT_NODES)) {
  106. is_user_define_ouput_nodes = true;
  107. }
  108. int parent_index = -1;
  109. auto output_desc = op_desc->MutableOutputDesc(ele.second);
  110. if (output_desc == nullptr) {
  111. GELOGE(FAILED, "[Get][OutputDesc]Can not find output tensor desc from node:%s, index %d",
  112. op_desc->GetName().c_str(), ele.second);
  113. return FAILED;
  114. }
  115. (void)ge::AttrUtils::GetInt(output_desc, ge::ATTR_NAME_PARENT_NODE_INDEX, parent_index);
  116. output_nodes_info.push_back({ele.first, ele.second, parent_index});
  117. }
  118. GELOGI("Output node set by user or leaf node, size:%zu.", output_nodes_info.size());
  119. for (auto &ele : out_nodes_tmp) {
  120. // add member, no need to remove duplicated because we need to keep all edges
  121. output_nodes_info.push_back(ele);
  122. }
  123. GELOGI("Get output node, size:%zu.", output_nodes_info.size());
  124. Status check_ret = CheckOutputNodeInfo(graph, output_nodes_info);
  125. if (check_ret != SUCCESS) {
  126. return check_ret;
  127. }
  128. return SUCCESS;
  129. }
  130. Status NetOutputPass::CheckOutputNodeInfo(const ComputeGraphPtr &graph, const std::vector<RetvalInfo> &outputs) {
  131. for (auto &item : outputs) {
  132. NodePtr node = item.output_node;
  133. if (node == nullptr) {
  134. REPORT_INNER_ERROR("E19999", "Param outputs has item which output_node is nullptr, check invalid");
  135. GELOGE(PARAM_INVALID, "Node in outputs is null.");
  136. return PARAM_INVALID;
  137. } else {
  138. if (graph->FindNode(node->GetName()) == nullptr) {
  139. REPORT_INNER_ERROR("E19999", "Find node:%s from graph:%s failed",
  140. node->GetName().c_str(), graph->GetName().c_str());
  141. GELOGE(INTERNAL_ERROR, "Out node (%s) is not in graph.", node->GetName().c_str());
  142. return INTERNAL_ERROR;
  143. }
  144. GE_CHECK_NOTNULL(node->GetOpDesc());
  145. int32_t out_size = node->GetOpDesc()->GetOutputsSize();
  146. int32_t index = item.node_output_index;
  147. if (index < 0 || index >= out_size) {
  148. REPORT_INNER_ERROR("E19999", "Index:%d in param outputs item, < 0 or > output size:%d of node:%s(%s)",
  149. index, out_size, node->GetName().c_str(), node->GetType().c_str());
  150. GELOGE(PARAM_INVALID,
  151. "User declared out node (%s) output index:%d must be smaller "
  152. "than node ouput size:%d and cann't be negative!",
  153. node->GetName().c_str(), index, out_size);
  154. return PARAM_INVALID;
  155. }
  156. }
  157. }
  158. return SUCCESS;
  159. }
  160. Status NetOutputPass::RemoveUnusedNode(const ge::ComputeGraphPtr &graph) {
  161. std::vector<ge::NodePtr> node_to_delete;
  162. // Delete _Retval operator.
  163. for (auto &node : graph->GetDirectNode()) {
  164. GE_IF_BOOL_EXEC(node->GetOpDesc() == nullptr, GELOGW("Node OpDesc is nullptr"); continue);
  165. bool need_be_deleted = node->GetInDataNodes().size() != 0 && node->GetOutDataNodesSize() == 0 &&
  166. (node->GetOpDesc()->HasAttr(RETVAL_ATTR_NAME_INDEX));
  167. if (need_be_deleted) {
  168. node_to_delete.push_back(node);
  169. }
  170. }
  171. for (NodePtr &node : node_to_delete) {
  172. auto iter = targets_.find(node);
  173. if (iter != targets_.end()) {
  174. GELOGI("[Net output pass] node[%s] is in user set targets.so do not remove!", node->GetName().c_str());
  175. continue;
  176. }
  177. if (graph->RemoveNode(node) != GRAPH_SUCCESS) {
  178. REPORT_INNER_ERROR("E19999", "Remove node:%s(%s) from graph:%s failed",
  179. node->GetName().c_str(), node->GetType().c_str(), graph->GetName().c_str());
  180. GELOGE(INTERNAL_ERROR, "Remove node failed, node name:%s.", node->GetName().c_str());
  181. return INTERNAL_ERROR;
  182. }
  183. }
  184. return SUCCESS;
  185. }
  186. Status NetOutputPass::UpdateNetOutputDesc(const ge::NodePtr &net_output) {
  187. OpDescPtr net_output_desc = net_output->GetOpDesc();
  188. if (net_output_desc == nullptr) {
  189. REPORT_INNER_ERROR("E19999", "OpDesc in Param net_output is nullptr, check invalid");
  190. GELOGE(INTERNAL_ERROR, "Opdesc of net output node is nullptr.");
  191. return INTERNAL_ERROR;
  192. }
  193. if (net_output_desc->GetInputsSize() == 0) {
  194. REPORT_INNER_ERROR("E19999", "Input desc num of node:%s(%s) is 0, check invalid",
  195. net_output_desc->GetName().c_str(), net_output_desc->GetType().c_str());
  196. GELOGE(INTERNAL_ERROR, "Net output node input is empty.");
  197. return INTERNAL_ERROR;
  198. }
  199. std::vector<bool> is_input_const;
  200. for (const auto &in_anchor : net_output->GetAllInDataAnchors()) {
  201. GE_CHECK_NOTNULL(in_anchor);
  202. uint32_t index = static_cast<uint32_t>(in_anchor->GetIdx());
  203. if (index >= net_output_desc->GetAllInputsDesc().size()) {
  204. REPORT_INNER_ERROR("E19999", "Node:%s(%s) has in_anchor index:%u >= its input desc num:%zu, check invalid",
  205. net_output_desc->GetName().c_str(), net_output_desc->GetType().c_str(), index,
  206. net_output_desc->GetAllInputsDesc().size());
  207. GELOGE(INTERNAL_ERROR, "Index is invalid, index:%u, size:%zu.", index,
  208. net_output_desc->GetAllInputsDesc().size());
  209. return INTERNAL_ERROR;
  210. }
  211. GE_CHECK_NOTNULL(in_anchor->GetPeerOutAnchor());
  212. is_input_const.push_back(PassUtils::IsConstant(in_anchor->GetPeerOutAnchor()->GetOwnerNode()));
  213. OpDescPtr src_op_desc = in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetOpDesc();
  214. GE_CHECK_NOTNULL(src_op_desc);
  215. uint32_t peer_index = static_cast<uint32_t>(in_anchor->GetPeerOutAnchor()->GetIdx());
  216. ge::GeTensorDesc output_in_desc = src_op_desc->GetOutputDesc(peer_index);
  217. if (net_output_desc->UpdateInputDesc(index, output_in_desc) != GRAPH_SUCCESS) {
  218. REPORT_CALL_ERROR("E19999", "Update input desc of op:%s(%s) failed, index:%u",
  219. net_output_desc->GetName().c_str(), net_output_desc->GetType().c_str(), index);
  220. GELOGE(INTERNAL_ERROR, "Update input desc failed, index:%u.", index);
  221. return INTERNAL_ERROR;
  222. }
  223. GELOGD("Update desc, format:%s, data type:%s, index:%u.",
  224. TypeUtils::FormatToSerialString(output_in_desc.GetFormat()).c_str(),
  225. TypeUtils::DataTypeToSerialString(output_in_desc.GetDataType()).c_str(), index);
  226. }
  227. net_output_desc->SetIsInputConst(is_input_const);
  228. return SUCCESS;
  229. }
  230. Status NetOutputPass::AddCtrlEdgeForTargets(const ge::NodePtr &net_out_node) {
  231. if (net_out_node == nullptr) {
  232. REPORT_INNER_ERROR("E19999", "Param net_out_node is nullptr, check invalid");
  233. GELOGE(PARAM_INVALID, "net out node is null.");
  234. return PARAM_INVALID;
  235. }
  236. // Add ctrl edge for targets
  237. for (auto &node : targets_) {
  238. if (node == nullptr) {
  239. continue;
  240. }
  241. // no need to check null because have handled it in run SaveAndRemoveTargets function
  242. graphStatus status = GraphUtils::AddEdge(node->GetOutControlAnchor(), net_out_node->GetInControlAnchor());
  243. if (status != GRAPH_SUCCESS) {
  244. REPORT_CALL_ERROR("E19999", "Add control edge between op:%s(%s) and op:%s(%s) failed",
  245. node->GetName().c_str(), node->GetType().c_str(),
  246. net_out_node->GetName().c_str(), net_out_node->GetType().c_str());
  247. GELOGE(INTERNAL_ERROR, "Add ctrl edge to netoutput node[%s] for target node [%s] failed!",
  248. net_out_node->GetName().c_str(), node->GetName().c_str());
  249. return INTERNAL_ERROR;
  250. }
  251. GELOGD("Add ctrl edge to netoutput node[%s] for target node [%s] success!", net_out_node->GetName().c_str(),
  252. node->GetName().c_str());
  253. }
  254. return SUCCESS;
  255. }
  256. void NetOutputPass::SaveAndRemoveTargets(const ge::ComputeGraphPtr &graph) {
  257. // save user targets node
  258. for (auto &node : graph->GetGraphTargetNodesInfo()) {
  259. if (node == nullptr) {
  260. GELOGW("User pointed targets contains null node.ignore it !");
  261. continue;
  262. }
  263. targets_.insert(node);
  264. }
  265. GELOGI("User pointed targets size is %zu !", targets_.size());
  266. }
  267. Status NetOutputPass::AddEdgesForNetOutput(const ge::ComputeGraphPtr &graph, const ge::NodePtr &net_out_node,
  268. const std::vector<RetvalInfo> &output_nodes_info) {
  269. int32_t net_input_index = 0;
  270. for (auto &item : output_nodes_info) {
  271. NodePtr src_node = item.output_node;
  272. GE_CHECK_NOTNULL(src_node);
  273. graphStatus status = GraphUtils::AddEdge(src_node->GetOutDataAnchor(item.node_output_index),
  274. net_out_node->GetInDataAnchor(net_input_index));
  275. if (status != GRAPH_SUCCESS) {
  276. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(index:%u) and op:%s(%s)(index:%d) failed",
  277. src_node->GetName().c_str(), src_node->GetType().c_str(), item.node_output_index,
  278. net_out_node->GetName().c_str(), net_out_node->GetType().c_str(), net_input_index);
  279. GELOGE(INTERNAL_ERROR, "AddEdge failed, src name:%s, src index:%d, dst index:%d.", src_node->GetName().c_str(),
  280. item.node_output_index, net_input_index);
  281. return INTERNAL_ERROR;
  282. }
  283. GELOGD("AddEdge to output node, src name:%s, src index:%d, dst index:%d.", src_node->GetName().c_str(),
  284. item.node_output_index, net_input_index);
  285. if (item.parent_node_index >= 0) {
  286. GELOGI("Add parent node index %d for the netoutput input %d on graph %s", item.parent_node_index, net_input_index,
  287. graph->GetName().c_str());
  288. auto input_desc = net_out_node->GetOpDesc()->MutableInputDesc(net_input_index);
  289. if (input_desc == nullptr) {
  290. REPORT_CALL_ERROR("E19999", "Node:%s(%s) has no input desc index is %d, check invalid",
  291. net_out_node->GetName().c_str(), net_out_node->GetType().c_str(), net_input_index);
  292. GELOGE(INTERNAL_ERROR, "Can not find intput tensor desc from NetOutput, index %d", net_input_index);
  293. return INTERNAL_ERROR;
  294. }
  295. if (!AttrUtils::SetInt(input_desc, ATTR_NAME_PARENT_NODE_INDEX, item.parent_node_index)) {
  296. REPORT_CALL_ERROR("E19999", "Set Attr:%s to input:%d tensor of op:%s(%s) failed",
  297. ATTR_NAME_PARENT_NODE_INDEX.c_str(), net_input_index,
  298. net_out_node->GetName().c_str(), net_out_node->GetType().c_str());
  299. GELOGE(INTERNAL_ERROR, "Failed to add parent index to NetOutput, index %d", net_input_index);
  300. return INTERNAL_ERROR;
  301. }
  302. }
  303. net_input_index++;
  304. }
  305. if (RemoveUnusedNode(graph) != SUCCESS) {
  306. GELOGE(INTERNAL_ERROR, "Remove unused nodes failed.");
  307. return INTERNAL_ERROR;
  308. }
  309. if (AddCtrlEdgeForTargets(net_out_node) != SUCCESS) {
  310. GELOGE(INTERNAL_ERROR, "Add ctrl edge for targets failed.");
  311. return INTERNAL_ERROR;
  312. }
  313. // Add true stream, netoutput is 0
  314. GE_IF_BOOL_EXEC(!ge::AttrUtils::SetInt(net_out_node->GetOpDesc(), ATTR_NAME_TRUE_BRANCH_STREAM, 0),
  315. REPORT_CALL_ERROR("E19999", "Set Attr:%s to op:%s(%s) failed", ATTR_NAME_TRUE_BRANCH_STREAM.c_str(),
  316. net_out_node->GetName().c_str(), net_out_node->GetType().c_str());
  317. GELOGE(INTERNAL_ERROR, "set ATTR_NAME_TRUE_BRANCH_STREAM failed");
  318. return INTERNAL_ERROR);
  319. return SUCCESS;
  320. }
  321. bool NetOutputPass::CheckNodeIsInOutputNodes(const ge::ComputeGraphPtr &graph, const ge::NodePtr &node) {
  322. for (auto &ele : graph->GetGraphOutNodesInfo()) {
  323. auto out_node = ele.first;
  324. if (node == out_node) {
  325. return true;
  326. }
  327. }
  328. return false;
  329. }
  330. Status NetOutputPass::UnLinkDataAnchorOfNetoutput(const ge::ComputeGraphPtr &graph, const ge::NodePtr &net_out_node) {
  331. if (net_out_node == nullptr) {
  332. REPORT_INNER_ERROR("E19999", "Param net_out_node is nullptr, check invalid");
  333. GELOGE(PARAM_INVALID, "net out node is null.");
  334. return PARAM_INVALID;
  335. }
  336. Status ret = SUCCESS;
  337. // unlink all anchor to data anchor of netoutput
  338. for (auto &in_data_anchor : net_out_node->GetAllInDataAnchors()) {
  339. if (in_data_anchor == nullptr) {
  340. continue;
  341. }
  342. auto peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  343. if (peer_out_anchor == nullptr) {
  344. GELOGI("PeerOutAnchor is null!");
  345. continue;
  346. }
  347. auto node = peer_out_anchor->GetOwnerNode();
  348. auto iter = targets_.find(node);
  349. if (iter != targets_.end()) {
  350. if (!CheckNodeIsInOutputNodes(graph, node)) {
  351. ret = in_data_anchor->Unlink(peer_out_anchor);
  352. if (ret != SUCCESS) {
  353. REPORT_CALL_ERROR("E19999",
  354. "Op:%s(%s) out index:%d unlink from op:%s(%s) in index:%d failed",
  355. net_out_node->GetName().c_str(), net_out_node->GetType().c_str(), in_data_anchor->GetIdx(),
  356. node->GetName().c_str(), node->GetType().c_str(), peer_out_anchor->GetIdx());
  357. GELOGE(INTERNAL_ERROR, "Unlink peer_out_anchor fail!");
  358. return ret;
  359. }
  360. } else {
  361. targets_.erase(iter);
  362. }
  363. }
  364. }
  365. return ret;
  366. }
  367. Status NetOutputPass::UnLinkControlAnchorOfNetoutput(const ge::ComputeGraphPtr &graph,
  368. const ge::NodePtr &net_out_node) {
  369. if (net_out_node == nullptr) {
  370. REPORT_INNER_ERROR("E19999", "Param net_out_node is nullptr, check invalid");
  371. GELOGE(PARAM_INVALID, "net out node is null.");
  372. return PARAM_INVALID;
  373. }
  374. Status ret = SUCCESS;
  375. auto in_control_anchor = net_out_node->GetInControlAnchor();
  376. if (in_control_anchor == nullptr) {
  377. REPORT_INNER_ERROR("E19999", "Param net_out_node's in control anchor is nullptr, check invalid");
  378. GELOGE(PARAM_INVALID, "in control anchor is null.");
  379. return PARAM_INVALID;
  380. }
  381. // unlink all data anchor to control anchor of netoutput
  382. for (auto &peer_out_data_anchor : in_control_anchor->GetPeerOutDataAnchors()) {
  383. if (peer_out_data_anchor == nullptr) {
  384. GELOGI("PeerOutControlAnchor is null!");
  385. } else {
  386. auto node = peer_out_data_anchor->GetOwnerNode();
  387. auto iter = targets_.find(node);
  388. if (iter != targets_.end()) {
  389. if (CheckNodeIsInOutputNodes(graph, node) == false) {
  390. ret = in_control_anchor->Unlink(peer_out_data_anchor);
  391. if (ret != SUCCESS) {
  392. REPORT_CALL_ERROR("E19999", "Op:%s(%s) unlink control edge from op:%s(%s) failed",
  393. net_out_node->GetName().c_str(), net_out_node->GetType().c_str(),
  394. node->GetName().c_str(), node->GetType().c_str());
  395. GELOGE(INTERNAL_ERROR, "Unlink peer_out_anchor fail!");
  396. return ret;
  397. }
  398. } else {
  399. targets_.erase(iter);
  400. }
  401. }
  402. }
  403. }
  404. /// check all control anchor to control anchor of netoutput and delete it from targets
  405. /// to avoid duplicated add control edge;
  406. for (auto &peer_out_control_anchor : in_control_anchor->GetPeerOutControlAnchors()) {
  407. if (peer_out_control_anchor == nullptr) {
  408. GELOGI("PeerOutControlAnchor is null");
  409. } else {
  410. auto node = peer_out_control_anchor->GetOwnerNode();
  411. auto iter = targets_.find(node);
  412. if (iter != targets_.end()) {
  413. targets_.erase(iter);
  414. }
  415. }
  416. }
  417. return ret;
  418. }
  419. Status NetOutputPass::UnLink(const ge::ComputeGraphPtr &graph, const ge::NodePtr &net_out_node) {
  420. GELOGI("[NetOutputPass] Enter Unlink process.");
  421. Status ret = UnLinkDataAnchorOfNetoutput(graph, net_out_node);
  422. if (ret != SUCCESS) {
  423. GELOGI("[NetOutputPass] UnLinkDataAnchorOfNetoutput process fail.");
  424. return ret;
  425. }
  426. ret = UnLinkControlAnchorOfNetoutput(graph, net_out_node);
  427. if (ret != SUCCESS) {
  428. GELOGI("[NetOutputPass] UnLinkControlAnchorOfNetoutput process fail.");
  429. return ret;
  430. }
  431. return ret;
  432. }
  433. Status NetOutputPass::ProcessWithNetoutput(const ge::ComputeGraphPtr &graph, const ge::NodePtr &output_node) {
  434. if (UpdateNetOutputDesc(output_node) != SUCCESS) {
  435. GELOGE(INTERNAL_ERROR, "Update net output desc failed.");
  436. return INTERNAL_ERROR;
  437. }
  438. if (UnLink(graph, output_node) != SUCCESS) {
  439. GELOGE(INTERNAL_ERROR, "UnLink connection between netoutput node and user set target node");
  440. return INTERNAL_ERROR;
  441. }
  442. if (AddCtrlEdgeForTargets(output_node) != SUCCESS) {
  443. GELOGE(INTERNAL_ERROR, "Add ctrl edge for targets failed.");
  444. return INTERNAL_ERROR;
  445. }
  446. return SUCCESS;
  447. }
  448. Status NetOutputPass::AddCtrlEdgesBetweenLeafAndNetOutput(const ge::ComputeGraphPtr &graph,
  449. const ge::NodePtr &net_out_node) {
  450. GE_CHECK_NOTNULL(net_out_node);
  451. if (!GetLocalOmgContext().user_out_nodes.empty() || is_user_define_ouput_nodes) {
  452. GELOGI("No need to add ctrl edge to netoutput because user out nodes have been set.");
  453. return SUCCESS;
  454. }
  455. bool graph_has_only_one_node_except_netoutput = (graph->GetDirectNodesSize() == kNodesCount);
  456. for (const auto &node : graph->GetDirectNode()) {
  457. if (node == nullptr || node->GetOpDesc() == nullptr || node->GetOpDesc()->GetType() == NETOUTPUT) {
  458. continue;
  459. }
  460. if ((node->GetInControlNodes().size() != 0 || node->GetInDataNodes().size() != 0 ||
  461. graph_has_only_one_node_except_netoutput) &&
  462. node->GetOutDataNodesSize() == 0 && node->GetOutControlNodes().size() == 0) {
  463. GE_CHK_GRAPH_STATUS_RET(GraphUtils::AddEdge(node->GetOutControlAnchor(), net_out_node->GetInControlAnchor()),
  464. "add edge failed");
  465. GELOGD("Add ctrl edge success. src name :%s, dst name :%s", node->GetName().c_str(),
  466. net_out_node->GetName().c_str());
  467. }
  468. }
  469. return SUCCESS;
  470. }
  471. Status NetOutputPass::CreateNetOutputNode(OpDescPtr &net_output_desc, const ge::ComputeGraphPtr &graph) {
  472. // Only flush subgraph name
  473. string node_name =
  474. (graph->GetParentGraph() != nullptr) ? (graph->GetName() + "_" + NODE_NAME_NET_OUTPUT) : NODE_NAME_NET_OUTPUT;
  475. net_output_desc = MakeShared<OpDesc>(node_name, NETOUTPUT);
  476. if (net_output_desc == nullptr) {
  477. REPORT_CALL_ERROR("E19999", "New OpDesc failed");
  478. GELOGE(MEMALLOC_FAILED, "Make shared net output op failed.");
  479. return MEMALLOC_FAILED;
  480. }
  481. (void)AttrUtils::SetListStr(net_output_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES,
  482. std::move(std::vector<std::string>()));
  483. return SUCCESS;
  484. }
  485. Status NetOutputPass::Run(ge::ComputeGraphPtr graph) {
  486. if (graph == nullptr) {
  487. REPORT_INNER_ERROR("E19999", "Param graph is nullptr, check invalid");
  488. GELOGE(GE_GRAPH_PARAM_NULLPTR, "Compute graph is null.");
  489. return GE_GRAPH_PARAM_NULLPTR;
  490. }
  491. GELOGI("[NETOUTPUT PASS] Run.graph is [%s]", graph->GetName().c_str());
  492. NodePtr output_node = graph->FindFirstNodeMatchType(NETOUTPUT);
  493. // save user targets node
  494. SaveAndRemoveTargets(graph);
  495. // If graph already has a netoutput node, doesn't need to create it again.
  496. if (output_node != nullptr) {
  497. (void)AttrUtils::SetListStr(output_node->GetOpDesc(), ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES,
  498. std::move(std::vector<std::string>()));
  499. if (ProcessWithNetoutput(graph, output_node) != SUCCESS) {
  500. GELOGE(INTERNAL_ERROR, "Process with netoutput node failed.");
  501. return INTERNAL_ERROR;
  502. }
  503. } else {
  504. if (AddNetOutputNodeToGraph(graph, output_node) != SUCCESS) {
  505. GELOGE(INTERNAL_ERROR, "Set user define dtype and format for netoutput failed.");
  506. return INTERNAL_ERROR;
  507. }
  508. }
  509. // Add userdef attrs to netoutput node
  510. return SetUserDefDTypeAndFormatFromAtcParams(output_node);
  511. }
  512. Status NetOutputPass::AddNetOutputNodeToGraph(const ge::ComputeGraphPtr &graph, NodePtr &output_node) {
  513. OpDescPtr net_output_desc = nullptr;
  514. if (CreateNetOutputNode(net_output_desc, graph) != SUCCESS) {
  515. GELOGE(INTERNAL_ERROR, "Get net output nodes failed.");
  516. return INTERNAL_ERROR;
  517. }
  518. std::vector<RetvalInfo> output_nodes_info;
  519. if (GetOutputNode(graph, output_nodes_info) != SUCCESS) {
  520. GELOGE(INTERNAL_ERROR, "Get net output nodes failed.");
  521. return INTERNAL_ERROR;
  522. }
  523. GELOGI("[NETOUTPUT PASS] OutNodesInfo size:%zu, Targets Size:%zu, is_include_special_node_:%d",
  524. graph->GetGraphOutNodesInfo().size(), graph->GetGraphTargetNodesInfo().size(), is_include_special_node_);
  525. // If user does not set out nodes and targets and no retval node, also add netoutput node
  526. if ((graph->GetGraphOutNodesInfo().empty()) && (graph->GetGraphTargetNodesInfo().empty()) &&
  527. !is_include_special_node_) {
  528. GELOGI("[NETOUTPUT PASS] Both output, target and special nodes are empty! add net output node");
  529. output_node = graph->AddNode(net_output_desc);
  530. GE_CHK_STATUS_RET(AddCtrlEdgesBetweenLeafAndNetOutput(graph, output_node),
  531. "add ctrl edge between leaf and netoutput failed");
  532. if (!ge::AttrUtils::SetInt(output_node->GetOpDesc(), ATTR_NAME_TRUE_BRANCH_STREAM, 0)) {
  533. REPORT_CALL_ERROR("E19999", "Set Attr:%s to op:%s(%s) failed", ATTR_NAME_TRUE_BRANCH_STREAM.c_str(),
  534. output_node->GetName().c_str(), output_node->GetType().c_str());
  535. GELOGE(INTERNAL_ERROR, "set ATTR_NAME_TRUE_BRANCH_STREAM failed");
  536. return INTERNAL_ERROR;
  537. }
  538. GELOGI("[NETOUTPUT PASS] Add net output node succeed");
  539. return SUCCESS;
  540. }
  541. GELOGI("[NETOUTPUT PASS] Output node size:%lu.", output_nodes_info.size());
  542. if (output_nodes_info.empty()) {
  543. // because retval node is contained by output_nodes_info, here means targets is non-empty
  544. output_node = graph->AddNode(net_output_desc);
  545. if (output_node == nullptr) {
  546. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed",
  547. net_output_desc->GetName().c_str(), net_output_desc->GetType().c_str(),
  548. graph->GetName().c_str());
  549. GELOGE(INTERNAL_ERROR, "Add output node failed.");
  550. return INTERNAL_ERROR;
  551. }
  552. GE_CHK_STATUS_RET(AddCtrlEdgeForTargets(output_node), "add ctrl edge for targets failed");
  553. // Add true stream, netoutput is 0
  554. GE_IF_BOOL_EXEC(!ge::AttrUtils::SetInt(output_node->GetOpDesc(), ATTR_NAME_TRUE_BRANCH_STREAM, 0),
  555. REPORT_CALL_ERROR("E19999", "Set Attr:%s to op:%s(%s) failed", ATTR_NAME_TRUE_BRANCH_STREAM.c_str(),
  556. output_node->GetName().c_str(), output_node->GetType().c_str());
  557. GELOGE(INTERNAL_ERROR, "set ATTR_NAME_TRUE_BRANCH_STREAM failed");
  558. return INTERNAL_ERROR);
  559. return SUCCESS;
  560. }
  561. AddInOutForNetOutputOp(graph, net_output_desc, output_nodes_info);
  562. output_node = graph->AddNode(net_output_desc);
  563. if (output_node == nullptr) {
  564. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed",
  565. net_output_desc->GetName().c_str(), net_output_desc->GetType().c_str(),
  566. graph->GetName().c_str());
  567. GELOGE(INTERNAL_ERROR, "Add output node failed.");
  568. return INTERNAL_ERROR;
  569. }
  570. if (AddEdgesForNetOutput(graph, output_node, output_nodes_info) != SUCCESS) {
  571. GELOGE(INTERNAL_ERROR, "Add edges for net output node failed.");
  572. return INTERNAL_ERROR;
  573. }
  574. if (AddCtrlEdgesBetweenLeafAndNetOutput(graph, output_node) != SUCCESS) {
  575. GELOGE(INTERNAL_ERROR, "Add control edges between leaf and netoutput failed.");
  576. return INTERNAL_ERROR;
  577. }
  578. GELOGI("Add NetOutput node success.");
  579. return SUCCESS;
  580. }
  581. void NetOutputPass::AddInOutForNetOutputOp(const ComputeGraphPtr &graph, OpDescPtr &net_output_desc,
  582. vector<RetvalInfo> &output_nodes_info) {
  583. std::vector<bool> is_input_const;
  584. for (auto iter = output_nodes_info.begin(); iter != output_nodes_info.end();) {
  585. NodePtr src_node = iter->output_node;
  586. if (src_node == nullptr) {
  587. continue;
  588. }
  589. int32_t src_index = iter->node_output_index;
  590. // if src_node is in targets_, no need to Add in and out for netoutput
  591. auto it = targets_.find(src_node);
  592. if (it != targets_.end()) {
  593. iter = output_nodes_info.erase(iter);
  594. GELOGD("node [%s] is in processed targets, do not add inout for netoutput!", src_node->GetName().c_str());
  595. continue;
  596. }
  597. /// Get the output attribute of src_node,
  598. /// and set to the input/output of net_out_node.
  599. if (src_node == nullptr || src_node->GetOpDesc() == nullptr || net_output_desc == nullptr) {
  600. REPORT_INNER_ERROR("E19999", "Param output_nodes_info has RetvalInfo item, which src_node is invalid; "
  601. "or Param net_output_desc is nullptr, check invalid");
  602. GELOGE(INTERNAL_ERROR, "src node or net output desc is null.");
  603. return;
  604. }
  605. ge::GeTensorDesc out_desc = src_node->GetOpDesc()->GetOutputDesc(src_index);
  606. out_desc.SetFormat(FORMAT_ND);
  607. out_desc.SetOriginFormat(FORMAT_ND);
  608. GE_IF_BOOL_EXEC(net_output_desc->AddInputDesc(out_desc) != SUCCESS, GELOGW("add input desc failed"); return );
  609. is_input_const.push_back(PassUtils::IsConstant(src_node));
  610. ++iter;
  611. }
  612. net_output_desc->SetIsInputConst(is_input_const);
  613. }
  614. bool NeedUpdateOutputByOutputTypeParm(std::string &output_type, OpDescPtr &op_desc, uint32_t &src_index,
  615. ge::DataType &dt) {
  616. if (output_type_str_to_datatype.find(output_type) != output_type_str_to_datatype.end()) {
  617. dt = output_type_str_to_datatype[output_type];
  618. return true;
  619. }
  620. vector<string> output_dt_str;
  621. if (ge::AttrUtils::GetListStr(op_desc, "_user_defined_output_data_type", output_dt_str)) {
  622. for (const auto &dt_str : output_dt_str) {
  623. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  624. if (dt_str_split.size() == kUserDefinedElementCount) {
  625. if (dt_str_split[0] == to_string(src_index)) {
  626. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  627. return true;
  628. }
  629. } else {
  630. GELOGW("The size of [%s] is not 2 after split.", dt_str.c_str());
  631. continue;
  632. }
  633. }
  634. }
  635. return false;
  636. }
  637. bool NeedUpdateOutputFp16Nc1hwc0(OpDescPtr &op_desc, uint32_t &src_index) {
  638. vector<string> output_dt_str;
  639. if (ge::AttrUtils::GetListStr(op_desc, "_user_defined_output_fp16_5hd", output_dt_str)) {
  640. for (const auto &dt_str : output_dt_str) {
  641. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  642. if (dt_str_split.size() == kUserDefinedElementCount) {
  643. if (dt_str_split[0] == to_string(src_index)) {
  644. return true;
  645. }
  646. } else {
  647. GELOGW("The size of [%s] is not 2 after split.", dt_str.c_str());
  648. continue;
  649. }
  650. }
  651. }
  652. return false;
  653. }
  654. Status NetOutputPass::SetUserDefDTypeAndFormatFromAtcParams(const NodePtr &output_node) {
  655. if (output_node == nullptr) {
  656. GELOGI("[NETOUTPUT PASS] The graph no need netoutput node!");
  657. return SUCCESS;
  658. }
  659. auto output_type = GetLocalOmgContext().output_type;
  660. auto op_desc = output_node->GetOpDesc();
  661. GE_CHECK_NOTNULL(op_desc);
  662. std::vector<std::string> userdef_dtypes;
  663. std::vector<std::string> userdef_formats;
  664. ge::DataType output_data_type = ge::DT_FLOAT;
  665. for (const auto &in_anchor : output_node->GetAllInDataAnchors()) {
  666. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  667. auto peer_out = in_anchor->GetPeerOutAnchor();
  668. if (peer_out == nullptr) {
  669. // If user set target, peer_out anchor will be unlinked.
  670. continue;
  671. }
  672. auto src_index = static_cast<uint32_t>(peer_out->GetIdx());
  673. auto src_node = peer_out->GetOwnerNode();
  674. GE_CHECK_NOTNULL(src_node);
  675. OpDescPtr src_op_desc = src_node->GetOpDesc();
  676. GE_CHECK_NOTNULL(src_op_desc);
  677. // Update datatype
  678. if (NeedUpdateOutputByOutputTypeParm(output_type, src_op_desc, src_index, output_data_type)) {
  679. GELOGD("Add user-define datatype:%s to netoutput node.",
  680. TypeUtils::DataTypeToSerialString(output_data_type).c_str());
  681. userdef_dtypes.push_back(
  682. std::to_string(index).append(":").append(TypeUtils::DataTypeToSerialString(output_data_type)));
  683. continue;
  684. }
  685. // Output_node is not set,check if is_output_adjust_hw_layout is set
  686. bool set_fp16_nc1hwc0 = NeedUpdateOutputFp16Nc1hwc0(src_op_desc, src_index);
  687. if (set_fp16_nc1hwc0) {
  688. // Set DT_FLOAT16 & FORMAT_NC1HWC0
  689. userdef_dtypes.push_back(std::to_string(index).append(":").append(TypeUtils::DataTypeToSerialString(DT_FLOAT16)));
  690. userdef_formats.push_back(
  691. std::to_string(index).append(":").append(TypeUtils::FormatToSerialString(FORMAT_NC1HWC0)));
  692. }
  693. }
  694. if (!userdef_dtypes.empty() && !ge::AttrUtils::SetListStr(op_desc, ATTR_ATC_USER_DEFINE_DATATYPE, userdef_dtypes)) {
  695. REPORT_INNER_ERROR("E19999", "User define datatype is empty or Set Attr:%s to op:%s(%s) failed",
  696. ATTR_ATC_USER_DEFINE_DATATYPE.c_str(), op_desc->GetName().c_str(), op_desc->GetType().c_str());
  697. GELOGE(INTERNAL_ERROR, "Set user_define_dtype attr list for netoutput failed.");
  698. return INTERNAL_ERROR;
  699. }
  700. if (!userdef_formats.empty() && !ge::AttrUtils::SetListStr(op_desc, ATTR_ATC_USER_DEFINE_FORMAT, userdef_formats)) {
  701. REPORT_INNER_ERROR("E19999", "User define format is empty or Set Attr:%s to op:%s(%s) failed",
  702. ATTR_ATC_USER_DEFINE_FORMAT.c_str(), op_desc->GetName().c_str(), op_desc->GetType().c_str());
  703. GELOGE(INTERNAL_ERROR, "Set user_define_format attr list for netoutput failed.");
  704. return INTERNAL_ERROR;
  705. }
  706. return SUCCESS;
  707. }
  708. } // 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