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 27 kB

5 years ago
5 years ago
5 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
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
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
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655
  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 "framework/omg/parser/parser_inner_ctx.h"
  27. #include "graph/debug/ge_attr_define.h"
  28. #include "graph/common/local_context.h"
  29. #include "graph/passes/pass_utils.h"
  30. #include "graph/utils/tensor_utils.h"
  31. #include "graph/utils/type_utils.h"
  32. namespace ge {
  33. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  34. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  35. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  36. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  37. // the size of user defined output datatype or format string after split by ":".
  38. const size_t kUserDefinedElementCount = 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. GELOGE(PARAM_INVALID, "Get output index failed.");
  46. return PARAM_INVALID;
  47. }
  48. if (retval_node_index_map.count(output_index) > 0) {
  49. GELOGE(PARAM_INVALID, "Retval has duplicate index.");
  50. return PARAM_INVALID;
  51. }
  52. int parent_node_index = -1;
  53. (void)AttrUtils::GetInt(node->GetOpDesc(), ATTR_NAME_PARENT_NODE_INDEX, parent_node_index);
  54. InDataAnchorPtr in_data_anchor = node->GetInDataAnchor(0);
  55. GE_CHECK_NOTNULL(in_data_anchor);
  56. GE_CHECK_NOTNULL(in_data_anchor->GetPeerOutAnchor());
  57. int32_t src_node_index = in_data_anchor->GetPeerOutAnchor()->GetIdx();
  58. NodePtr src_node_ptr = in_data_anchor->GetPeerOutAnchor()->GetOwnerNode();
  59. retval_node_index_map[output_index] = {src_node_ptr, src_node_index, parent_node_index};
  60. // if user targets include retval node,delete it from set and insert its input node instead
  61. // better to GetInNodes here
  62. auto iter = targets_.find(node);
  63. if (iter != targets_.end()) {
  64. targets_.erase(iter);
  65. targets_.insert(src_node_ptr);
  66. GELOGI("node [%s] is in user def targets, do not output result to user!", node->GetName().c_str());
  67. }
  68. is_include_special_node_ = true;
  69. return SUCCESS;
  70. }
  71. Status NetOutputPass::GetOutputNode(const ge::ComputeGraphPtr &graph, std::vector<RetvalInfo> &output_nodes_info) {
  72. std::map<int32_t, RetvalInfo> retval_node_index_map;
  73. for (NodePtr &node : graph->GetDirectNode()) {
  74. Status ret = SUCCESS;
  75. if ((node->GetOpDesc() != nullptr) && (node->GetOpDesc()->HasAttr(RETVAL_ATTR_NAME_INDEX))) {
  76. /// Set the output according to the Retval operator,
  77. /// identify by whether there is an index parameter
  78. ret = GetRetvalOutputInfo(node, retval_node_index_map);
  79. }
  80. if (ret != SUCCESS) {
  81. GELOGE(ret, "GetRetvalOutputInfo failed");
  82. return ret;
  83. }
  84. }
  85. GELOGI("Get retval node size:%zu.", retval_node_index_map.size());
  86. std::vector<RetvalInfo> out_nodes_tmp;
  87. /// The Netoutput output is determined by Retval, and the input order
  88. /// of Netoutput is sorted according to the index value of Retval.
  89. for (auto &it : retval_node_index_map) {
  90. out_nodes_tmp.push_back(it.second);
  91. }
  92. // when user set targets, mean that no output result
  93. for (auto &ele : graph->GetGraphOutNodesInfo()) {
  94. auto iter = targets_.find(ele.first);
  95. if (iter != targets_.end()) {
  96. GELOGI("user set out node [%s] is found in user def targets, out node is prio!", ele.first->GetName().c_str());
  97. targets_.erase(iter);
  98. }
  99. output_nodes_info.push_back({ele.first, ele.second, -1});
  100. }
  101. GELOGI("Output node set by user or leaf node, size:%zu.", output_nodes_info.size());
  102. for (auto &ele : out_nodes_tmp) {
  103. // add member, no need to remove duplicated because we need to keep all edges
  104. output_nodes_info.push_back(ele);
  105. }
  106. GELOGI("Get output node, size:%zu.", output_nodes_info.size());
  107. Status check_ret = CheckOutputNodeInfo(graph, output_nodes_info);
  108. if (check_ret != SUCCESS) {
  109. return check_ret;
  110. }
  111. return SUCCESS;
  112. }
  113. Status NetOutputPass::CheckOutputNodeInfo(const ComputeGraphPtr &graph, const std::vector<RetvalInfo> &outputs) {
  114. for (auto &item : outputs) {
  115. NodePtr node = item.output_node;
  116. if (node == nullptr) {
  117. GELOGE(PARAM_INVALID, "Node in outputs is null.");
  118. return PARAM_INVALID;
  119. } else {
  120. if (graph->FindNode(node->GetName()) == nullptr) {
  121. GELOGE(INTERNAL_ERROR, "Out node (%s) is not in graph.", node->GetName().c_str());
  122. return INTERNAL_ERROR;
  123. }
  124. GE_CHECK_NOTNULL(node->GetOpDesc());
  125. int32_t out_size = node->GetOpDesc()->GetOutputsSize();
  126. int32_t index = item.node_output_index;
  127. if (index < 0 || index >= out_size) {
  128. GELOGE(PARAM_INVALID,
  129. "User declared out node (%s) output index:%d must be smaller "
  130. "than node ouput size:%d and cann't be negative!",
  131. node->GetName().c_str(), index, out_size);
  132. return PARAM_INVALID;
  133. }
  134. }
  135. }
  136. return SUCCESS;
  137. }
  138. Status NetOutputPass::RemoveUnusedNode(const ge::ComputeGraphPtr &graph) {
  139. std::vector<ge::NodePtr> node_to_delete;
  140. // Delete _Retval operator.
  141. for (auto &node : graph->GetDirectNode()) {
  142. GE_IF_BOOL_EXEC(node->GetOpDesc() == nullptr, GELOGW("Node OpDesc is nullptr"); continue);
  143. bool need_be_deleted = node->GetInDataNodes().size() != 0 && node->GetOutDataNodesSize() == 0 &&
  144. (node->GetOpDesc()->HasAttr(RETVAL_ATTR_NAME_INDEX));
  145. if (need_be_deleted) {
  146. node_to_delete.push_back(node);
  147. }
  148. }
  149. for (NodePtr &node : node_to_delete) {
  150. auto iter = targets_.find(node);
  151. if (iter != targets_.end()) {
  152. GELOGI("[Net output pass] node[%s] is in user set targets.so do not remove!", node->GetName().c_str());
  153. continue;
  154. }
  155. if (graph->RemoveNode(node) != GRAPH_SUCCESS) {
  156. GELOGE(INTERNAL_ERROR, "Remove node failed, node name:%s.", node->GetName().c_str());
  157. return INTERNAL_ERROR;
  158. }
  159. }
  160. return SUCCESS;
  161. }
  162. Status NetOutputPass::UpdateNetOutputDesc(const ge::NodePtr &net_output) {
  163. OpDescPtr net_output_desc = net_output->GetOpDesc();
  164. if (net_output_desc == nullptr) {
  165. GELOGE(INTERNAL_ERROR, "Opdesc of net output node is nullptr.");
  166. return INTERNAL_ERROR;
  167. }
  168. if (net_output_desc->GetInputsSize() == 0) {
  169. GELOGE(INTERNAL_ERROR, "Net output node input is empty.");
  170. return INTERNAL_ERROR;
  171. }
  172. std::vector<bool> is_input_const;
  173. for (const auto &in_anchor : net_output->GetAllInDataAnchors()) {
  174. GE_CHECK_NOTNULL(in_anchor);
  175. uint32_t index = static_cast<uint32_t>(in_anchor->GetIdx());
  176. if (index >= net_output_desc->GetAllInputsDesc().size()) {
  177. GELOGE(INTERNAL_ERROR, "Index is invalid, index:%u, size:%zu.", index,
  178. net_output_desc->GetAllInputsDesc().size());
  179. return INTERNAL_ERROR;
  180. }
  181. GE_CHECK_NOTNULL(in_anchor->GetPeerOutAnchor());
  182. is_input_const.push_back(PassUtils::IsConstant(in_anchor->GetPeerOutAnchor()->GetOwnerNode()));
  183. OpDescPtr src_op_desc = in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetOpDesc();
  184. GE_CHECK_NOTNULL(src_op_desc);
  185. uint32_t peer_index = static_cast<uint32_t>(in_anchor->GetPeerOutAnchor()->GetIdx());
  186. ge::GeTensorDesc output_in_desc = src_op_desc->GetOutputDesc(peer_index);
  187. if (net_output_desc->UpdateInputDesc(index, output_in_desc) != GRAPH_SUCCESS) {
  188. GELOGE(INTERNAL_ERROR, "Update input desc failed, index:%u.", index);
  189. return INTERNAL_ERROR;
  190. }
  191. GELOGD("Update desc, format:%s, data type:%s, index:%u.",
  192. TypeUtils::FormatToSerialString(output_in_desc.GetFormat()).c_str(),
  193. TypeUtils::DataTypeToSerialString(output_in_desc.GetDataType()).c_str(), index);
  194. }
  195. net_output_desc->SetIsInputConst(is_input_const);
  196. return SUCCESS;
  197. }
  198. Status NetOutputPass::AddCtrlEdgeForTargets(const ge::NodePtr &net_out_node) {
  199. if (net_out_node == nullptr) {
  200. GELOGE(PARAM_INVALID, "net out node is null.");
  201. return PARAM_INVALID;
  202. }
  203. // Add ctrl edge for targets
  204. for (auto &node : targets_) {
  205. if (node == nullptr) {
  206. continue;
  207. }
  208. // no need to check null because have handled it in run SaveAndRemoveTargets function
  209. graphStatus status = GraphUtils::AddEdge(node->GetOutControlAnchor(), net_out_node->GetInControlAnchor());
  210. if (status != GRAPH_SUCCESS) {
  211. GELOGE(INTERNAL_ERROR, "Add ctrl edge to netoutput node[%s] for target node [%s] failed!",
  212. net_out_node->GetName().c_str(), node->GetName().c_str());
  213. return INTERNAL_ERROR;
  214. }
  215. GELOGD("Add ctrl edge to netoutput node[%s] for target node [%s] success!", net_out_node->GetName().c_str(),
  216. node->GetName().c_str());
  217. }
  218. return SUCCESS;
  219. }
  220. void NetOutputPass::SaveAndRemoveTargets(const ge::ComputeGraphPtr &graph) {
  221. // save user targets node
  222. for (auto &node : graph->GetGraphTargetNodesInfo()) {
  223. if (node == nullptr) {
  224. GELOGW("User pointed targets contains null node.ignore it !");
  225. continue;
  226. }
  227. targets_.insert(node);
  228. }
  229. GELOGI("User pointed targets size is %zu !", targets_.size());
  230. }
  231. Status NetOutputPass::AddEdgesForNetOutput(const ge::ComputeGraphPtr &graph, const ge::NodePtr &net_out_node,
  232. const std::vector<RetvalInfo> &output_nodes_info) {
  233. int32_t net_input_index = 0;
  234. for (auto &item : output_nodes_info) {
  235. NodePtr src_node = item.output_node;
  236. GE_CHECK_NOTNULL(src_node);
  237. graphStatus status = GraphUtils::AddEdge(src_node->GetOutDataAnchor(item.node_output_index),
  238. net_out_node->GetInDataAnchor(net_input_index));
  239. if (status != GRAPH_SUCCESS) {
  240. GELOGE(INTERNAL_ERROR, "AddEdge failed, src name:%s, src index:%d, dst index:%d.", src_node->GetName().c_str(),
  241. item.node_output_index, net_input_index);
  242. return INTERNAL_ERROR;
  243. }
  244. GELOGD("AddEdge to output node, src name:%s, src index:%d, dst index:%d.", src_node->GetName().c_str(),
  245. item.node_output_index, net_input_index);
  246. if (item.parent_node_index >= 0) {
  247. GELOGI("Add parent node index %d for the netoutput input %d on graph %s", item.parent_node_index, net_input_index,
  248. graph->GetName().c_str());
  249. auto input_desc = net_out_node->GetOpDesc()->MutableInputDesc(net_input_index);
  250. if (input_desc == nullptr) {
  251. GELOGE(INTERNAL_ERROR, "Can not find intput tensor desc from NetOutput, index %d", net_input_index);
  252. return INTERNAL_ERROR;
  253. }
  254. if (!AttrUtils::SetInt(input_desc, ATTR_NAME_PARENT_NODE_INDEX, item.parent_node_index)) {
  255. GELOGE(INTERNAL_ERROR, "Failed to add parent index to NetOutput, index %d", net_input_index);
  256. return INTERNAL_ERROR;
  257. }
  258. }
  259. net_input_index++;
  260. }
  261. if (RemoveUnusedNode(graph) != SUCCESS) {
  262. GELOGE(INTERNAL_ERROR, "Remove unused nodes failed.");
  263. return INTERNAL_ERROR;
  264. }
  265. if (AddCtrlEdgeForTargets(net_out_node) != SUCCESS) {
  266. GELOGE(INTERNAL_ERROR, "Add ctrl edge for targets failed.");
  267. return INTERNAL_ERROR;
  268. }
  269. // Add true stream, netoutput is 0
  270. GE_IF_BOOL_EXEC(!ge::AttrUtils::SetInt(net_out_node->GetOpDesc(), ATTR_NAME_TRUE_BRANCH_STREAM, 0),
  271. GELOGE(INTERNAL_ERROR, "set ATTR_NAME_TRUE_BRANCH_STREAM failed");
  272. return INTERNAL_ERROR);
  273. return SUCCESS;
  274. }
  275. bool NetOutputPass::CheckNodeIsInOutputNodes(const ge::ComputeGraphPtr &graph, const ge::NodePtr &node) {
  276. for (auto &ele : graph->GetGraphOutNodesInfo()) {
  277. auto out_node = ele.first;
  278. if (node == out_node) {
  279. return true;
  280. }
  281. }
  282. return false;
  283. }
  284. Status NetOutputPass::UnLinkDataAnchorOfNetoutput(const ge::ComputeGraphPtr &graph, const ge::NodePtr &net_out_node) {
  285. if (net_out_node == nullptr) {
  286. GELOGE(PARAM_INVALID, "net out node is null.");
  287. return PARAM_INVALID;
  288. }
  289. Status ret = SUCCESS;
  290. // unlink all anchor to data anchor of netoutput
  291. for (auto &in_data_anchor : net_out_node->GetAllInDataAnchors()) {
  292. if (in_data_anchor == nullptr) {
  293. continue;
  294. }
  295. auto peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  296. if (peer_out_anchor == nullptr) {
  297. GELOGI("PeerOutAnchor is null!");
  298. continue;
  299. }
  300. auto node = peer_out_anchor->GetOwnerNode();
  301. auto iter = targets_.find(node);
  302. if (iter != targets_.end()) {
  303. if (!CheckNodeIsInOutputNodes(graph, node)) {
  304. ret = in_data_anchor->Unlink(peer_out_anchor);
  305. if (ret != SUCCESS) {
  306. GELOGE(INTERNAL_ERROR, "Unlink peer_out_anchor fail!");
  307. return ret;
  308. }
  309. } else {
  310. targets_.erase(iter);
  311. }
  312. }
  313. }
  314. return ret;
  315. }
  316. Status NetOutputPass::UnLinkControlAnchorOfNetoutput(const ge::ComputeGraphPtr &graph,
  317. const ge::NodePtr &net_out_node) {
  318. if (net_out_node == nullptr) {
  319. GELOGE(PARAM_INVALID, "net out node is null.");
  320. return PARAM_INVALID;
  321. }
  322. Status ret = SUCCESS;
  323. auto in_control_anchor = net_out_node->GetInControlAnchor();
  324. if (in_control_anchor == nullptr) {
  325. GELOGE(PARAM_INVALID, "in control anchor is null.");
  326. return PARAM_INVALID;
  327. }
  328. // unlink all data anchor to control anchor of netoutput
  329. for (auto &peer_out_data_anchor : in_control_anchor->GetPeerOutDataAnchors()) {
  330. if (peer_out_data_anchor == nullptr) {
  331. GELOGI("PeerOutControlAnchor is null!");
  332. } else {
  333. auto node = peer_out_data_anchor->GetOwnerNode();
  334. auto iter = targets_.find(node);
  335. if (iter != targets_.end()) {
  336. if (CheckNodeIsInOutputNodes(graph, node) == false) {
  337. ret = in_control_anchor->Unlink(peer_out_data_anchor);
  338. if (ret != SUCCESS) {
  339. GELOGE(INTERNAL_ERROR, "Unlink peer_out_anchor fail!");
  340. return ret;
  341. }
  342. } else {
  343. targets_.erase(iter);
  344. }
  345. }
  346. }
  347. }
  348. /// check all control anchor to control anchor of netoutput and delete it from targets
  349. /// to avoid duplicated add control edge;
  350. for (auto &peer_out_control_anchor : in_control_anchor->GetPeerOutControlAnchors()) {
  351. if (peer_out_control_anchor == nullptr) {
  352. GELOGI("PeerOutControlAnchor is null");
  353. } else {
  354. auto node = peer_out_control_anchor->GetOwnerNode();
  355. auto iter = targets_.find(node);
  356. if (iter != targets_.end()) {
  357. targets_.erase(iter);
  358. }
  359. }
  360. }
  361. return ret;
  362. }
  363. Status NetOutputPass::UnLink(const ge::ComputeGraphPtr &graph, const ge::NodePtr &net_out_node) {
  364. GELOGI("[NetOutputPass] Enter Unlink process.");
  365. Status ret = UnLinkDataAnchorOfNetoutput(graph, net_out_node);
  366. if (ret != SUCCESS) {
  367. GELOGI("[NetOutputPass] UnLinkDataAnchorOfNetoutput process fail.");
  368. return ret;
  369. }
  370. ret = UnLinkControlAnchorOfNetoutput(graph, net_out_node);
  371. if (ret != SUCCESS) {
  372. GELOGI("[NetOutputPass] UnLinkControlAnchorOfNetoutput process fail.");
  373. return ret;
  374. }
  375. return ret;
  376. }
  377. Status NetOutputPass::ProcessWithNetoutput(const ge::ComputeGraphPtr &graph, const ge::NodePtr &output_node) {
  378. if (UpdateNetOutputDesc(output_node) != SUCCESS) {
  379. GELOGE(INTERNAL_ERROR, "Update net output desc failed.");
  380. return INTERNAL_ERROR;
  381. }
  382. if (UnLink(graph, output_node) != SUCCESS) {
  383. GELOGE(INTERNAL_ERROR, "UnLink connection between netoutput node and user set target node");
  384. return INTERNAL_ERROR;
  385. }
  386. if (AddCtrlEdgeForTargets(output_node) != SUCCESS) {
  387. GELOGE(INTERNAL_ERROR, "Add ctrl edge for targets failed.");
  388. return INTERNAL_ERROR;
  389. }
  390. return SUCCESS;
  391. }
  392. Status NetOutputPass::AddCtrlEdgesBetweenLeafAndNetOutput(const ge::ComputeGraphPtr &graph,
  393. const ge::NodePtr &net_out_node) {
  394. GE_CHECK_NOTNULL(net_out_node);
  395. if (!GetLocalOmgContext().user_out_nodes.empty() || !GetParserContext().user_out_nodes.empty()) {
  396. GELOGI("No need to add ctrl edge to netoutput because user out nodes have been set.");
  397. return SUCCESS;
  398. }
  399. for (const auto &node : graph->GetDirectNode()) {
  400. if (node == nullptr || node->GetOpDesc() == nullptr || node->GetOpDesc()->GetType() == NETOUTPUT) {
  401. continue;
  402. }
  403. if ((node->GetInControlNodes().size() != 0 || node->GetInDataNodes().size() != 0) &&
  404. node->GetOutDataNodesSize() == 0 && node->GetOutControlNodes().size() == 0) {
  405. GE_CHK_STATUS_RET(GraphUtils::AddEdge(node->GetOutControlAnchor(), net_out_node->GetInControlAnchor()),
  406. "add edge failed");
  407. GELOGD("Add ctrl edge success. src name :%s, dst name :%s", node->GetName().c_str(),
  408. net_out_node->GetName().c_str());
  409. }
  410. }
  411. return SUCCESS;
  412. }
  413. Status NetOutputPass::CreateNetOutputNode(OpDescPtr &net_output_desc, const ge::ComputeGraphPtr &graph) {
  414. // Only flush subgraph name
  415. string node_name =
  416. (graph->GetParentGraph() != nullptr) ? (graph->GetName() + "_" + NODE_NAME_NET_OUTPUT) : NODE_NAME_NET_OUTPUT;
  417. net_output_desc = MakeShared<OpDesc>(node_name, NETOUTPUT);
  418. if (net_output_desc == nullptr) {
  419. GELOGE(MEMALLOC_FAILED, "Make shared net output op failed.");
  420. return MEMALLOC_FAILED;
  421. }
  422. (void)AttrUtils::SetListStr(net_output_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES,
  423. std::move(std::vector<std::string>()));
  424. return SUCCESS;
  425. }
  426. Status NetOutputPass::Run(ge::ComputeGraphPtr graph) {
  427. if (graph == nullptr) {
  428. GELOGE(GE_GRAPH_PARAM_NULLPTR, "Compute graph is null.");
  429. return GE_GRAPH_PARAM_NULLPTR;
  430. }
  431. GELOGI("NetOutputPass Run.");
  432. NodePtr output_node = graph->FindFirstNodeMatchType(NETOUTPUT);
  433. // save user targets node
  434. SaveAndRemoveTargets(graph);
  435. // If graph already has a netoutput node, doesn't need to create it again.
  436. if (output_node != nullptr) {
  437. (void)AttrUtils::SetListStr(output_node->GetOpDesc(), ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES,
  438. std::move(std::vector<std::string>()));
  439. if (ProcessWithNetoutput(graph, output_node) != SUCCESS) {
  440. GELOGE(INTERNAL_ERROR, "Process with netoutput node failed.");
  441. return INTERNAL_ERROR;
  442. }
  443. } else {
  444. if (AddNetOutputNodeToGraph(graph, output_node) != SUCCESS) {
  445. GELOGE(INTERNAL_ERROR, "Set user define dtype and format for netoutput failed.");
  446. return INTERNAL_ERROR;
  447. }
  448. }
  449. // Add userdef attrs to netoutput node
  450. return SetUserDefDTypeAndFormatFromAtcParams(output_node);
  451. }
  452. Status NetOutputPass::AddNetOutputNodeToGraph(const ge::ComputeGraphPtr &graph, NodePtr &output_node) {
  453. OpDescPtr net_output_desc = nullptr;
  454. if (CreateNetOutputNode(net_output_desc, graph) != SUCCESS) {
  455. GELOGE(INTERNAL_ERROR, "Get net output nodes failed.");
  456. return INTERNAL_ERROR;
  457. }
  458. std::vector<RetvalInfo> output_nodes_info;
  459. if (GetOutputNode(graph, output_nodes_info) != SUCCESS) {
  460. GELOGE(INTERNAL_ERROR, "Get net output nodes failed.");
  461. return INTERNAL_ERROR;
  462. }
  463. GELOGI("[NETOUTPUT PASS] OutNodesInfo size:%zu, Targets Size:%zu, is_include_special_node_:%d",
  464. graph->GetGraphOutNodesInfo().size(), graph->GetGraphTargetNodesInfo().size(), is_include_special_node_);
  465. // If user does not set out nodes and targets and no retval node, return false
  466. if ((graph->GetGraphOutNodesInfo().empty()) && (graph->GetGraphTargetNodesInfo().empty()) &&
  467. !is_include_special_node_) {
  468. GELOGI("[NETOUTPUT PASS] output_nodes and target_nodes and special nodes is empty!It means no need netoutput!");
  469. return SUCCESS;
  470. }
  471. GELOGI("[NETOUTPUT PASS] Output node size:%lu.", output_nodes_info.size());
  472. if (output_nodes_info.empty()) {
  473. // because retval node is contained by output_nodes_info, here means targets is non-empty
  474. output_node = graph->AddNode(net_output_desc);
  475. if (output_node == nullptr) {
  476. GELOGE(INTERNAL_ERROR, "Add output node failed.");
  477. return INTERNAL_ERROR;
  478. }
  479. GE_CHK_STATUS_RET(AddCtrlEdgeForTargets(output_node), "add ctrl edge for targets failed");
  480. // Add true stream, netoutput is 0
  481. GE_IF_BOOL_EXEC(!ge::AttrUtils::SetInt(output_node->GetOpDesc(), ATTR_NAME_TRUE_BRANCH_STREAM, 0),
  482. GELOGE(INTERNAL_ERROR, "set ATTR_NAME_TRUE_BRANCH_STREAM failed");
  483. return INTERNAL_ERROR);
  484. return SUCCESS;
  485. }
  486. AddInOutForNetOutputOp(graph, net_output_desc, output_nodes_info);
  487. output_node = graph->AddNode(net_output_desc);
  488. if (output_node == nullptr) {
  489. GELOGE(INTERNAL_ERROR, "Add output node failed.");
  490. return INTERNAL_ERROR;
  491. }
  492. if (AddEdgesForNetOutput(graph, output_node, output_nodes_info) != SUCCESS) {
  493. GELOGE(INTERNAL_ERROR, "Add edges for net output node failed.");
  494. return INTERNAL_ERROR;
  495. }
  496. if (AddCtrlEdgesBetweenLeafAndNetOutput(graph, output_node) != SUCCESS) {
  497. GELOGE(INTERNAL_ERROR, "Add control edges between leaf and netoutput failed.");
  498. return INTERNAL_ERROR;
  499. }
  500. GELOGI("Add NetOutput node success.");
  501. return SUCCESS;
  502. }
  503. void NetOutputPass::AddInOutForNetOutputOp(const ComputeGraphPtr &graph, OpDescPtr &net_output_desc,
  504. vector<RetvalInfo> &output_nodes_info) {
  505. std::vector<bool> is_input_const;
  506. for (auto iter = output_nodes_info.begin(); iter != output_nodes_info.end();) {
  507. NodePtr src_node = iter->output_node;
  508. if (src_node == nullptr) {
  509. continue;
  510. }
  511. int32_t src_index = iter->node_output_index;
  512. // if src_node is in targets_, no need to Add in and out for netoutput
  513. auto it = targets_.find(src_node);
  514. if (it != targets_.end()) {
  515. iter = output_nodes_info.erase(iter);
  516. GELOGD("node [%s] is in processed targets, do not add inout for netoutput!", src_node->GetName().c_str());
  517. continue;
  518. }
  519. /// Get the output attribute of src_node,
  520. /// and set to the input/output of net_out_node.
  521. if (src_node == nullptr || src_node->GetOpDesc() == nullptr || net_output_desc == nullptr) {
  522. GELOGE(INTERNAL_ERROR, "src node or net output desc is null.");
  523. return;
  524. }
  525. ge::GeTensorDesc out_desc = src_node->GetOpDesc()->GetOutputDesc(src_index);
  526. GE_IF_BOOL_EXEC(net_output_desc->AddInputDesc(out_desc) != SUCCESS, GELOGW("add input desc failed"); return );
  527. is_input_const.push_back(PassUtils::IsConstant(src_node));
  528. ++iter;
  529. }
  530. net_output_desc->SetIsInputConst(is_input_const);
  531. }
  532. bool NeedUpdateOutputByOutputTypeParm(std::string &output_type, OpDescPtr &op_desc, uint32_t &src_index,
  533. ge::DataType &dt) {
  534. if (output_type_str_to_datatype.find(output_type) != output_type_str_to_datatype.end()) {
  535. dt = output_type_str_to_datatype[output_type];
  536. return true;
  537. }
  538. vector<string> output_dt_str;
  539. if (ge::AttrUtils::GetListStr(op_desc, "_user_defined_output_data_type", output_dt_str)) {
  540. for (const auto &dt_str : output_dt_str) {
  541. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  542. if (dt_str_split.size() == kUserDefinedElementCount) {
  543. if (dt_str_split[0] == to_string(src_index)) {
  544. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  545. return true;
  546. }
  547. } else {
  548. GELOGW("The size of [%s] is not 2 after split.", dt_str.c_str());
  549. continue;
  550. }
  551. }
  552. }
  553. return false;
  554. }
  555. bool NeedUpdateOutputFp16Nc1hwc0(OpDescPtr &op_desc, uint32_t &src_index) {
  556. vector<string> output_dt_str;
  557. if (ge::AttrUtils::GetListStr(op_desc, "_user_defined_output_fp16_5hd", output_dt_str)) {
  558. for (const auto &dt_str : output_dt_str) {
  559. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  560. if (dt_str_split.size() == kUserDefinedElementCount) {
  561. if (dt_str_split[0] == to_string(src_index)) {
  562. return true;
  563. }
  564. } else {
  565. GELOGW("The size of [%s] is not 2 after split.", dt_str.c_str());
  566. continue;
  567. }
  568. }
  569. }
  570. return false;
  571. }
  572. Status NetOutputPass::SetUserDefDTypeAndFormatFromAtcParams(const NodePtr &output_node) {
  573. if (output_node == nullptr) {
  574. GELOGI("[NETOUTPUT PASS] The graph no need netoutput node!");
  575. return SUCCESS;
  576. }
  577. auto output_type = GetLocalOmgContext().output_type;
  578. auto op_desc = output_node->GetOpDesc();
  579. GE_CHECK_NOTNULL(op_desc);
  580. std::vector<std::string> userdef_dtypes;
  581. std::vector<std::string> userdef_formats;
  582. ge::DataType output_data_type = ge::DT_FLOAT;
  583. for (const auto &in_anchor : output_node->GetAllInDataAnchors()) {
  584. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  585. auto peer_out = in_anchor->GetPeerOutAnchor();
  586. if (peer_out == nullptr) {
  587. // If user set target, peer_out anchor will be unlinked.
  588. continue;
  589. }
  590. auto src_index = static_cast<uint32_t>(peer_out->GetIdx());
  591. auto src_node = peer_out->GetOwnerNode();
  592. GE_CHECK_NOTNULL(src_node);
  593. OpDescPtr src_op_desc = src_node->GetOpDesc();
  594. GE_CHECK_NOTNULL(src_op_desc);
  595. // Update datatype
  596. if (NeedUpdateOutputByOutputTypeParm(output_type, src_op_desc, src_index, output_data_type)) {
  597. GELOGD("Add user-define datatype:%s to netoutput node.",
  598. TypeUtils::DataTypeToSerialString(output_data_type).c_str());
  599. userdef_dtypes.push_back(
  600. std::to_string(index).append(":").append(TypeUtils::DataTypeToSerialString(output_data_type)));
  601. continue;
  602. }
  603. // Output_node is not set,check if is_output_adjust_hw_layout is set
  604. bool set_fp16_nc1hwc0 = NeedUpdateOutputFp16Nc1hwc0(src_op_desc, src_index);
  605. if (set_fp16_nc1hwc0) {
  606. // Set DT_FLOAT16 & FORMAT_NC1HWC0
  607. userdef_dtypes.push_back(std::to_string(index).append(":").append(TypeUtils::DataTypeToSerialString(DT_FLOAT16)));
  608. userdef_formats.push_back(
  609. std::to_string(index).append(":").append(TypeUtils::FormatToSerialString(FORMAT_NC1HWC0)));
  610. }
  611. }
  612. if (!userdef_dtypes.empty() && !ge::AttrUtils::SetListStr(op_desc, ATTR_ATC_USER_DEFINE_DATATYPE, userdef_dtypes)) {
  613. GELOGE(INTERNAL_ERROR, "Set user_define_dtype attr list for netoutput failed.");
  614. return INTERNAL_ERROR;
  615. }
  616. if (!userdef_formats.empty() && !ge::AttrUtils::SetListStr(op_desc, ATTR_ATC_USER_DEFINE_FORMAT, userdef_formats)) {
  617. GELOGE(INTERNAL_ERROR, "Set user_define_format attr list for netoutput failed.");
  618. return INTERNAL_ERROR;
  619. }
  620. return SUCCESS;
  621. }
  622. } // 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