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
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
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654
  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 "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. Status NetOutputPass::GetRetvalOutputInfo(const ge::NodePtr &node,
  39. std::map<int32_t, RetvalInfo> &retval_node_index_map) {
  40. GE_CHECK_NOTNULL(node);
  41. GE_CHECK_NOTNULL(node->GetOpDesc());
  42. int64_t output_index = 0;
  43. if (!AttrUtils::GetInt(node->GetOpDesc(), RETVAL_ATTR_NAME_INDEX, output_index)) {
  44. GELOGE(PARAM_INVALID, "Get output index failed.");
  45. return PARAM_INVALID;
  46. }
  47. if (retval_node_index_map.count(output_index) > 0) {
  48. GELOGE(PARAM_INVALID, "Retval has duplicate index.");
  49. return PARAM_INVALID;
  50. }
  51. int parent_node_index = -1;
  52. (void)AttrUtils::GetInt(node->GetOpDesc(), ATTR_NAME_PARENT_NODE_INDEX, parent_node_index);
  53. InDataAnchorPtr in_data_anchor = node->GetInDataAnchor(0);
  54. GE_CHECK_NOTNULL(in_data_anchor);
  55. GE_CHECK_NOTNULL(in_data_anchor->GetPeerOutAnchor());
  56. int32_t src_node_index = in_data_anchor->GetPeerOutAnchor()->GetIdx();
  57. NodePtr src_node_ptr = in_data_anchor->GetPeerOutAnchor()->GetOwnerNode();
  58. retval_node_index_map[output_index] = {src_node_ptr, src_node_index, parent_node_index};
  59. // if user targets include retval node,delete it from set and insert its input node instead
  60. // better to GetInNodes here
  61. auto iter = targets_.find(node);
  62. if (iter != targets_.end()) {
  63. targets_.erase(iter);
  64. targets_.insert(src_node_ptr);
  65. GELOGI("node [%s] is in user def targets, do not output result to user!", node->GetName().c_str());
  66. }
  67. is_include_special_node_ = true;
  68. return SUCCESS;
  69. }
  70. Status NetOutputPass::GetOutputNode(const ge::ComputeGraphPtr &graph, std::vector<RetvalInfo> &output_nodes_info) {
  71. std::map<int32_t, RetvalInfo> retval_node_index_map;
  72. for (NodePtr &node : graph->GetDirectNode()) {
  73. Status ret = SUCCESS;
  74. if ((node->GetOpDesc() != nullptr) && (node->GetOpDesc()->HasAttr(RETVAL_ATTR_NAME_INDEX))) {
  75. /// Set the output according to the Retval operator,
  76. /// identify by whether there is an index parameter
  77. ret = GetRetvalOutputInfo(node, retval_node_index_map);
  78. }
  79. if (ret != SUCCESS) {
  80. GELOGE(ret, "GetRetvalOutputInfo failed");
  81. return ret;
  82. }
  83. }
  84. GELOGI("Get retval node size:%zu.", retval_node_index_map.size());
  85. std::vector<RetvalInfo> out_nodes_tmp;
  86. /// The Netoutput output is determined by Retval, and the input order
  87. /// of Netoutput is sorted according to the index value of Retval.
  88. for (auto &it : retval_node_index_map) {
  89. out_nodes_tmp.push_back(it.second);
  90. }
  91. // when user set targets, mean that no output result
  92. for (auto &ele : graph->GetGraphOutNodesInfo()) {
  93. auto iter = targets_.find(ele.first);
  94. if (iter != targets_.end()) {
  95. GELOGI("user set out node [%s] is found in user def targets, out node is prio!", ele.first->GetName().c_str());
  96. targets_.erase(iter);
  97. }
  98. output_nodes_info.push_back({ele.first, ele.second, -1});
  99. }
  100. GELOGI("Output node set by user or leaf node, size:%zu.", output_nodes_info.size());
  101. for (auto &ele : out_nodes_tmp) {
  102. // add member, no need to remove duplicated because we need to keep all edges
  103. output_nodes_info.push_back(ele);
  104. }
  105. GELOGI("Get output node, size:%zu.", output_nodes_info.size());
  106. Status check_ret = CheckOutputNodeInfo(graph, output_nodes_info);
  107. if (check_ret != SUCCESS) {
  108. return check_ret;
  109. }
  110. return SUCCESS;
  111. }
  112. Status NetOutputPass::CheckOutputNodeInfo(const ComputeGraphPtr &graph, const std::vector<RetvalInfo> &outputs) {
  113. for (auto &item : outputs) {
  114. NodePtr node = item.output_node;
  115. if (node == nullptr) {
  116. GELOGE(PARAM_INVALID, "Node in outputs is null.");
  117. return PARAM_INVALID;
  118. } else {
  119. if (graph->FindNode(node->GetName()) == nullptr) {
  120. GELOGE(INTERNAL_ERROR, "Out node (%s) is not in graph.", node->GetName().c_str());
  121. return INTERNAL_ERROR;
  122. }
  123. GE_CHECK_NOTNULL(node->GetOpDesc());
  124. int32_t out_size = node->GetOpDesc()->GetOutputsSize();
  125. int32_t index = item.node_output_index;
  126. if (index < 0 || index >= out_size) {
  127. GELOGE(PARAM_INVALID,
  128. "User declared out node (%s) output index:%d must be smaller "
  129. "than node ouput size:%d and cann't be negative!",
  130. node->GetName().c_str(), index, out_size);
  131. return PARAM_INVALID;
  132. }
  133. }
  134. }
  135. return SUCCESS;
  136. }
  137. Status NetOutputPass::RemoveUnusedNode(const ge::ComputeGraphPtr &graph) {
  138. std::vector<ge::NodePtr> node_to_delete;
  139. // Delete _Retval operator.
  140. for (auto &node : graph->GetDirectNode()) {
  141. GE_IF_BOOL_EXEC(node->GetOpDesc() == nullptr, GELOGW("Node OpDesc is nullptr"); continue);
  142. bool need_be_deleted = node->GetInDataNodes().size() != 0 && node->GetOutDataNodesSize() == 0 &&
  143. (node->GetOpDesc()->HasAttr(RETVAL_ATTR_NAME_INDEX));
  144. if (need_be_deleted) {
  145. node_to_delete.push_back(node);
  146. }
  147. }
  148. for (NodePtr &node : node_to_delete) {
  149. auto iter = targets_.find(node);
  150. if (iter != targets_.end()) {
  151. GELOGI("[Net output pass] node[%s] is in user set targets.so do not remove!", node->GetName().c_str());
  152. continue;
  153. }
  154. if (graph->RemoveNode(node) != GRAPH_SUCCESS) {
  155. GELOGE(INTERNAL_ERROR, "Remove node failed, node name:%s.", node->GetName().c_str());
  156. return INTERNAL_ERROR;
  157. }
  158. }
  159. return SUCCESS;
  160. }
  161. Status NetOutputPass::UpdateNetOutputDesc(const ge::NodePtr &net_output) {
  162. OpDescPtr net_output_desc = net_output->GetOpDesc();
  163. if (net_output_desc == nullptr) {
  164. GELOGE(INTERNAL_ERROR, "Opdesc of net output node is nullptr.");
  165. return INTERNAL_ERROR;
  166. }
  167. if (net_output_desc->GetInputsSize() == 0) {
  168. GELOGE(INTERNAL_ERROR, "Net output node input is empty.");
  169. return INTERNAL_ERROR;
  170. }
  171. std::vector<bool> is_input_const;
  172. for (const auto &in_anchor : net_output->GetAllInDataAnchors()) {
  173. GE_CHECK_NOTNULL(in_anchor);
  174. uint32_t index = static_cast<uint32_t>(in_anchor->GetIdx());
  175. if (index >= net_output_desc->GetAllInputsDesc().size()) {
  176. GELOGE(INTERNAL_ERROR, "Index is invalid, index:%u, size:%zu.", index,
  177. net_output_desc->GetAllInputsDesc().size());
  178. return INTERNAL_ERROR;
  179. }
  180. GE_CHECK_NOTNULL(in_anchor->GetPeerOutAnchor());
  181. is_input_const.push_back(PassUtils::IsConstant(in_anchor->GetPeerOutAnchor()->GetOwnerNode()));
  182. OpDescPtr src_op_desc = in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetOpDesc();
  183. GE_CHECK_NOTNULL(src_op_desc);
  184. uint32_t peer_index = static_cast<uint32_t>(in_anchor->GetPeerOutAnchor()->GetIdx());
  185. ge::GeTensorDesc output_in_desc = src_op_desc->GetOutputDesc(peer_index);
  186. if (net_output_desc->UpdateInputDesc(index, output_in_desc) != GRAPH_SUCCESS) {
  187. GELOGE(INTERNAL_ERROR, "Update input desc failed, index:%u.", index);
  188. return INTERNAL_ERROR;
  189. }
  190. GELOGD("Update desc, format:%s, data type:%s, index:%u.",
  191. TypeUtils::FormatToSerialString(output_in_desc.GetFormat()).c_str(),
  192. TypeUtils::DataTypeToSerialString(output_in_desc.GetDataType()).c_str(), index);
  193. }
  194. net_output_desc->SetIsInputConst(is_input_const);
  195. return SUCCESS;
  196. }
  197. Status NetOutputPass::AddCtrlEdgeForTargets(const ge::NodePtr &net_out_node) {
  198. if (net_out_node == nullptr) {
  199. GELOGE(PARAM_INVALID, "net out node is null.");
  200. return PARAM_INVALID;
  201. }
  202. // Add ctrl edge for targets
  203. for (auto &node : targets_) {
  204. if (node == nullptr) {
  205. continue;
  206. }
  207. // no need to check null because have handled it in run SaveAndRemoveTargets function
  208. graphStatus status = GraphUtils::AddEdge(node->GetOutControlAnchor(), net_out_node->GetInControlAnchor());
  209. if (status != GRAPH_SUCCESS) {
  210. GELOGE(INTERNAL_ERROR, "Add ctrl edge to netoutput node[%s] for target node [%s] failed!",
  211. net_out_node->GetName().c_str(), node->GetName().c_str());
  212. return INTERNAL_ERROR;
  213. }
  214. GELOGD("Add ctrl edge to netoutput node[%s] for target node [%s] success!", net_out_node->GetName().c_str(),
  215. node->GetName().c_str());
  216. }
  217. return SUCCESS;
  218. }
  219. void NetOutputPass::SaveAndRemoveTargets(const ge::ComputeGraphPtr &graph) {
  220. // save user targets node
  221. for (auto &node : graph->GetGraphTargetNodesInfo()) {
  222. if (node == nullptr) {
  223. GELOGW("User pointed targets contains null node.ignore it !");
  224. continue;
  225. }
  226. targets_.insert(node);
  227. }
  228. GELOGI("User pointed targets size is %zu !", targets_.size());
  229. }
  230. Status NetOutputPass::AddEdgesForNetOutput(const ge::ComputeGraphPtr &graph, const ge::NodePtr &net_out_node,
  231. const std::vector<RetvalInfo> &output_nodes_info) {
  232. int32_t net_input_index = 0;
  233. for (auto &item : output_nodes_info) {
  234. NodePtr src_node = item.output_node;
  235. GE_CHECK_NOTNULL(src_node);
  236. graphStatus status = GraphUtils::AddEdge(src_node->GetOutDataAnchor(item.node_output_index),
  237. net_out_node->GetInDataAnchor(net_input_index));
  238. if (status != GRAPH_SUCCESS) {
  239. GELOGE(INTERNAL_ERROR, "AddEdge failed, src name:%s, src index:%d, dst index:%d.", src_node->GetName().c_str(),
  240. item.node_output_index, net_input_index);
  241. return INTERNAL_ERROR;
  242. }
  243. GELOGD("AddEdge to output node, src name:%s, src index:%d, dst index:%d.", src_node->GetName().c_str(),
  244. item.node_output_index, net_input_index);
  245. if (item.parent_node_index >= 0) {
  246. GELOGI("Add parent node index %d for the netoutput input %d on graph %s", item.parent_node_index, net_input_index,
  247. graph->GetName().c_str());
  248. auto input_desc = net_out_node->GetOpDesc()->MutableInputDesc(net_input_index);
  249. if (input_desc == nullptr) {
  250. GELOGE(INTERNAL_ERROR, "Can not find intput tensor desc from NetOutput, index %d", net_input_index);
  251. return INTERNAL_ERROR;
  252. }
  253. if (!AttrUtils::SetInt(input_desc, ATTR_NAME_PARENT_NODE_INDEX, item.parent_node_index)) {
  254. GELOGE(INTERNAL_ERROR, "Failed to add parent index to NetOutput, index %d", net_input_index);
  255. return INTERNAL_ERROR;
  256. }
  257. }
  258. net_input_index++;
  259. }
  260. if (RemoveUnusedNode(graph) != SUCCESS) {
  261. GELOGE(INTERNAL_ERROR, "Remove unused nodes failed.");
  262. return INTERNAL_ERROR;
  263. }
  264. if (AddCtrlEdgeForTargets(net_out_node) != SUCCESS) {
  265. GELOGE(INTERNAL_ERROR, "Add ctrl edge for targets failed.");
  266. return INTERNAL_ERROR;
  267. }
  268. // Add true stream, netoutput is 0
  269. GE_IF_BOOL_EXEC(!ge::AttrUtils::SetInt(net_out_node->GetOpDesc(), ATTR_NAME_TRUE_BRANCH_STREAM, 0),
  270. GELOGE(INTERNAL_ERROR, "set ATTR_NAME_TRUE_BRANCH_STREAM failed");
  271. return INTERNAL_ERROR);
  272. return SUCCESS;
  273. }
  274. bool NetOutputPass::CheckNodeIsInOutputNodes(const ge::ComputeGraphPtr &graph, const ge::NodePtr &node) {
  275. for (auto &ele : graph->GetGraphOutNodesInfo()) {
  276. auto out_node = ele.first;
  277. if (node == out_node) {
  278. return true;
  279. }
  280. }
  281. return false;
  282. }
  283. Status NetOutputPass::UnLinkDataAnchorOfNetoutput(const ge::ComputeGraphPtr &graph, const ge::NodePtr &net_out_node) {
  284. if (net_out_node == nullptr) {
  285. GELOGE(PARAM_INVALID, "net out node is null.");
  286. return PARAM_INVALID;
  287. }
  288. Status ret = SUCCESS;
  289. // unlink all anchor to data anchor of netoutput
  290. for (auto &in_data_anchor : net_out_node->GetAllInDataAnchors()) {
  291. if (in_data_anchor == nullptr) {
  292. continue;
  293. }
  294. auto peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  295. if (peer_out_anchor == nullptr) {
  296. GELOGI("PeerOutAnchor is null!");
  297. continue;
  298. }
  299. auto node = peer_out_anchor->GetOwnerNode();
  300. auto iter = targets_.find(node);
  301. if (iter != targets_.end()) {
  302. if (!CheckNodeIsInOutputNodes(graph, node)) {
  303. ret = in_data_anchor->Unlink(peer_out_anchor);
  304. if (ret != SUCCESS) {
  305. GELOGE(INTERNAL_ERROR, "Unlink peer_out_anchor fail!");
  306. return ret;
  307. }
  308. } else {
  309. targets_.erase(iter);
  310. }
  311. }
  312. }
  313. return ret;
  314. }
  315. Status NetOutputPass::UnLinkControlAnchorOfNetoutput(const ge::ComputeGraphPtr &graph,
  316. const ge::NodePtr &net_out_node) {
  317. if (net_out_node == nullptr) {
  318. GELOGE(PARAM_INVALID, "net out node is null.");
  319. return PARAM_INVALID;
  320. }
  321. Status ret = SUCCESS;
  322. auto in_control_anchor = net_out_node->GetInControlAnchor();
  323. if (in_control_anchor == nullptr) {
  324. GELOGE(PARAM_INVALID, "in control anchor is null.");
  325. return PARAM_INVALID;
  326. }
  327. // unlink all data anchor to control anchor of netoutput
  328. for (auto &peer_out_data_anchor : in_control_anchor->GetPeerOutDataAnchors()) {
  329. if (peer_out_data_anchor == nullptr) {
  330. GELOGI("PeerOutControlAnchor is null!");
  331. } else {
  332. auto node = peer_out_data_anchor->GetOwnerNode();
  333. auto iter = targets_.find(node);
  334. if (iter != targets_.end()) {
  335. if (CheckNodeIsInOutputNodes(graph, node) == false) {
  336. ret = in_control_anchor->Unlink(peer_out_data_anchor);
  337. if (ret != SUCCESS) {
  338. GELOGE(INTERNAL_ERROR, "Unlink peer_out_anchor fail!");
  339. return ret;
  340. }
  341. } else {
  342. targets_.erase(iter);
  343. }
  344. }
  345. }
  346. }
  347. /// check all control anchor to control anchor of netoutput and delete it from targets
  348. /// to avoid duplicated add control edge;
  349. for (auto &peer_out_control_anchor : in_control_anchor->GetPeerOutControlAnchors()) {
  350. if (peer_out_control_anchor == nullptr) {
  351. GELOGI("PeerOutControlAnchor is null");
  352. } else {
  353. auto node = peer_out_control_anchor->GetOwnerNode();
  354. auto iter = targets_.find(node);
  355. if (iter != targets_.end()) {
  356. targets_.erase(iter);
  357. }
  358. }
  359. }
  360. return ret;
  361. }
  362. Status NetOutputPass::UnLink(const ge::ComputeGraphPtr &graph, const ge::NodePtr &net_out_node) {
  363. GELOGI("[NetOutputPass] Enter Unlink process.");
  364. Status ret = UnLinkDataAnchorOfNetoutput(graph, net_out_node);
  365. if (ret != SUCCESS) {
  366. GELOGI("[NetOutputPass] UnLinkDataAnchorOfNetoutput process fail.");
  367. return ret;
  368. }
  369. ret = UnLinkControlAnchorOfNetoutput(graph, net_out_node);
  370. if (ret != SUCCESS) {
  371. GELOGI("[NetOutputPass] UnLinkControlAnchorOfNetoutput process fail.");
  372. return ret;
  373. }
  374. return ret;
  375. }
  376. Status NetOutputPass::ProcessWithNetoutput(const ge::ComputeGraphPtr &graph, const ge::NodePtr &output_node) {
  377. if (UpdateNetOutputDesc(output_node) != SUCCESS) {
  378. GELOGE(INTERNAL_ERROR, "Update net output desc failed.");
  379. return INTERNAL_ERROR;
  380. }
  381. if (UnLink(graph, output_node) != SUCCESS) {
  382. GELOGE(INTERNAL_ERROR, "UnLink connection between netoutput node and user set target node");
  383. return INTERNAL_ERROR;
  384. }
  385. if (AddCtrlEdgeForTargets(output_node) != SUCCESS) {
  386. GELOGE(INTERNAL_ERROR, "Add ctrl edge for targets failed.");
  387. return INTERNAL_ERROR;
  388. }
  389. return SUCCESS;
  390. }
  391. Status NetOutputPass::AddCtrlEdgesBetweenLeafAndNetOutput(const ge::ComputeGraphPtr &graph,
  392. const ge::NodePtr &net_out_node) {
  393. GE_CHECK_NOTNULL(net_out_node);
  394. if (!GetLocalOmgContext().user_out_nodes.empty()) {
  395. GELOGI("No need to add ctrl edge to netoutput because user out nodes have been set.");
  396. return SUCCESS;
  397. }
  398. for (const auto &node : graph->GetDirectNode()) {
  399. if (node == nullptr || node->GetOpDesc() == nullptr || node->GetOpDesc()->GetType() == NETOUTPUT) {
  400. continue;
  401. }
  402. if ((node->GetInControlNodes().size() != 0 || node->GetInDataNodes().size() != 0) &&
  403. node->GetOutDataNodesSize() == 0 && node->GetOutControlNodes().size() == 0) {
  404. GE_CHK_STATUS_RET(GraphUtils::AddEdge(node->GetOutControlAnchor(), net_out_node->GetInControlAnchor()),
  405. "add edge failed");
  406. GELOGD("Add ctrl edge success. src name :%s, dst name :%s", node->GetName().c_str(),
  407. net_out_node->GetName().c_str());
  408. }
  409. }
  410. return SUCCESS;
  411. }
  412. Status NetOutputPass::CreateNetOutputNode(OpDescPtr &net_output_desc, const ge::ComputeGraphPtr &graph) {
  413. // Only flush subgraph name
  414. string node_name =
  415. (graph->GetParentGraph() != nullptr) ? (graph->GetName() + "_" + NODE_NAME_NET_OUTPUT) : NODE_NAME_NET_OUTPUT;
  416. net_output_desc = MakeShared<OpDesc>(node_name, NETOUTPUT);
  417. if (net_output_desc == nullptr) {
  418. GELOGE(MEMALLOC_FAILED, "Make shared net output op failed.");
  419. return MEMALLOC_FAILED;
  420. }
  421. (void)AttrUtils::SetListStr(net_output_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES,
  422. std::move(std::vector<std::string>()));
  423. return SUCCESS;
  424. }
  425. Status NetOutputPass::Run(ge::ComputeGraphPtr graph) {
  426. if (graph == nullptr) {
  427. GELOGE(GE_GRAPH_PARAM_NULLPTR, "Compute graph is null.");
  428. return GE_GRAPH_PARAM_NULLPTR;
  429. }
  430. GELOGI("NetOutputPass Run.");
  431. NodePtr output_node = graph->FindFirstNodeMatchType(NETOUTPUT);
  432. // save user targets node
  433. SaveAndRemoveTargets(graph);
  434. // If graph already has a netoutput node, doesn't need to create it again.
  435. if (output_node != nullptr) {
  436. (void)AttrUtils::SetListStr(output_node->GetOpDesc(), ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES,
  437. std::move(std::vector<std::string>()));
  438. if (ProcessWithNetoutput(graph, output_node) != SUCCESS) {
  439. GELOGE(INTERNAL_ERROR, "Process with netoutput node failed.");
  440. return INTERNAL_ERROR;
  441. }
  442. } else {
  443. if (AddNetOutputNodeToGraph(graph, output_node) != SUCCESS) {
  444. GELOGE(INTERNAL_ERROR, "Set user define dtype and format for netoutput failed.");
  445. return INTERNAL_ERROR;
  446. }
  447. }
  448. // Add userdef attrs to netoutput node
  449. return SetUserDefDTypeAndFormatFromAtcParams(output_node);
  450. }
  451. Status NetOutputPass::AddNetOutputNodeToGraph(const ge::ComputeGraphPtr &graph, NodePtr &output_node) {
  452. OpDescPtr net_output_desc = nullptr;
  453. if (CreateNetOutputNode(net_output_desc, graph) != SUCCESS) {
  454. GELOGE(INTERNAL_ERROR, "Get net output nodes failed.");
  455. return INTERNAL_ERROR;
  456. }
  457. std::vector<RetvalInfo> output_nodes_info;
  458. if (GetOutputNode(graph, output_nodes_info) != SUCCESS) {
  459. GELOGE(INTERNAL_ERROR, "Get net output nodes failed.");
  460. return INTERNAL_ERROR;
  461. }
  462. GELOGI("[NETOUTPUT PASS] OutNodesInfo size:%zu, Targets Size:%zu, is_include_special_node_:%d",
  463. graph->GetGraphOutNodesInfo().size(), graph->GetGraphTargetNodesInfo().size(), is_include_special_node_);
  464. // If user does not set out nodes and targets and no retval node, return false
  465. if ((graph->GetGraphOutNodesInfo().empty()) && (graph->GetGraphTargetNodesInfo().empty()) &&
  466. !is_include_special_node_) {
  467. GELOGI("[NETOUTPUT PASS] output_nodes and target_nodes and special nodes is empty!It means no need netoutput!");
  468. return SUCCESS;
  469. }
  470. GELOGI("[NETOUTPUT PASS] Output node size:%lu.", output_nodes_info.size());
  471. if (output_nodes_info.empty()) {
  472. // because retval node is contained by output_nodes_info, here means targets is non-empty
  473. output_node = graph->AddNode(net_output_desc);
  474. if (output_node == nullptr) {
  475. GELOGE(INTERNAL_ERROR, "Add output node failed.");
  476. return INTERNAL_ERROR;
  477. }
  478. GE_CHK_STATUS_RET(AddCtrlEdgeForTargets(output_node), "add ctrl edge for targets failed");
  479. // Add true stream, netoutput is 0
  480. GE_IF_BOOL_EXEC(!ge::AttrUtils::SetInt(output_node->GetOpDesc(), ATTR_NAME_TRUE_BRANCH_STREAM, 0),
  481. GELOGE(INTERNAL_ERROR, "set ATTR_NAME_TRUE_BRANCH_STREAM failed");
  482. return INTERNAL_ERROR);
  483. return SUCCESS;
  484. }
  485. AddInOutForNetOutputOp(graph, net_output_desc, output_nodes_info);
  486. output_node = graph->AddNode(net_output_desc);
  487. if (output_node == nullptr) {
  488. GELOGE(INTERNAL_ERROR, "Add output node failed.");
  489. return INTERNAL_ERROR;
  490. }
  491. if (AddEdgesForNetOutput(graph, output_node, output_nodes_info) != SUCCESS) {
  492. GELOGE(INTERNAL_ERROR, "Add edges for net output node failed.");
  493. return INTERNAL_ERROR;
  494. }
  495. if (AddCtrlEdgesBetweenLeafAndNetOutput(graph, output_node) != SUCCESS) {
  496. GELOGE(INTERNAL_ERROR, "Add control edges between leaf and netoutput failed.");
  497. return INTERNAL_ERROR;
  498. }
  499. GELOGI("Add NetOutput node success.");
  500. return SUCCESS;
  501. }
  502. void NetOutputPass::AddInOutForNetOutputOp(const ComputeGraphPtr &graph, OpDescPtr &net_output_desc,
  503. vector<RetvalInfo> &output_nodes_info) {
  504. std::vector<bool> is_input_const;
  505. for (auto iter = output_nodes_info.begin(); iter != output_nodes_info.end();) {
  506. NodePtr src_node = iter->output_node;
  507. if (src_node == nullptr) {
  508. continue;
  509. }
  510. int32_t src_index = iter->node_output_index;
  511. // if src_node is in targets_, no need to Add in and out for netoutput
  512. auto it = targets_.find(src_node);
  513. if (it != targets_.end()) {
  514. iter = output_nodes_info.erase(iter);
  515. GELOGD("node [%s] is in processed targets, do not add inout for netoutput!", src_node->GetName().c_str());
  516. continue;
  517. }
  518. /// Get the output attribute of src_node,
  519. /// and set to the input/output of net_out_node.
  520. if (src_node == nullptr || src_node->GetOpDesc() == nullptr || net_output_desc == nullptr) {
  521. GELOGE(INTERNAL_ERROR, "src node or net output desc is null.");
  522. return;
  523. }
  524. ge::GeTensorDesc out_desc = src_node->GetOpDesc()->GetOutputDesc(src_index);
  525. GE_IF_BOOL_EXEC(net_output_desc->AddInputDesc(out_desc) != SUCCESS, GELOGW("add input desc failed"); return );
  526. is_input_const.push_back(PassUtils::IsConstant(src_node));
  527. ++iter;
  528. }
  529. net_output_desc->SetIsInputConst(is_input_const);
  530. }
  531. bool NeedUpdateOutputByOutputTypeParm(std::string &output_type, OpDescPtr &op_desc, uint32_t &src_index,
  532. ge::DataType &dt) {
  533. if (output_type_str_to_datatype.find(output_type) != output_type_str_to_datatype.end()) {
  534. dt = output_type_str_to_datatype[output_type];
  535. return true;
  536. }
  537. vector<string> output_dt_str;
  538. if (ge::AttrUtils::GetListStr(op_desc, "_user_defined_output_data_type", output_dt_str)) {
  539. for (const auto &dt_str : output_dt_str) {
  540. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  541. if (dt_str_split.size() == kUserDefinedElementCount) {
  542. if (dt_str_split[0] == to_string(src_index)) {
  543. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  544. return true;
  545. }
  546. } else {
  547. GELOGW("The size of [%s] is not 2 after split.", dt_str.c_str());
  548. continue;
  549. }
  550. }
  551. }
  552. return false;
  553. }
  554. bool NeedUpdateOutputFp16Nc1hwc0(OpDescPtr &op_desc, uint32_t &src_index) {
  555. vector<string> output_dt_str;
  556. if (ge::AttrUtils::GetListStr(op_desc, "_user_defined_output_fp16_5hd", output_dt_str)) {
  557. for (const auto &dt_str : output_dt_str) {
  558. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  559. if (dt_str_split.size() == kUserDefinedElementCount) {
  560. if (dt_str_split[0] == to_string(src_index)) {
  561. return true;
  562. }
  563. } else {
  564. GELOGW("The size of [%s] is not 2 after split.", dt_str.c_str());
  565. continue;
  566. }
  567. }
  568. }
  569. return false;
  570. }
  571. Status NetOutputPass::SetUserDefDTypeAndFormatFromAtcParams(const NodePtr &output_node) {
  572. if (output_node == nullptr) {
  573. GELOGI("[NETOUTPUT PASS] The graph no need netoutput node!");
  574. return SUCCESS;
  575. }
  576. auto output_type = GetLocalOmgContext().output_type;
  577. auto op_desc = output_node->GetOpDesc();
  578. GE_CHECK_NOTNULL(op_desc);
  579. std::vector<std::string> userdef_dtypes;
  580. std::vector<std::string> userdef_formats;
  581. ge::DataType output_data_type = ge::DT_FLOAT;
  582. for (const auto &in_anchor : output_node->GetAllInDataAnchors()) {
  583. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  584. auto peer_out = in_anchor->GetPeerOutAnchor();
  585. if (peer_out == nullptr) {
  586. // If user set target, peer_out anchor will be unlinked.
  587. continue;
  588. }
  589. auto src_index = static_cast<uint32_t>(peer_out->GetIdx());
  590. auto src_node = peer_out->GetOwnerNode();
  591. GE_CHECK_NOTNULL(src_node);
  592. OpDescPtr src_op_desc = src_node->GetOpDesc();
  593. GE_CHECK_NOTNULL(src_op_desc);
  594. // Update datatype
  595. if (NeedUpdateOutputByOutputTypeParm(output_type, src_op_desc, src_index, output_data_type)) {
  596. GELOGD("Add user-define datatype:%s to netoutput node.",
  597. TypeUtils::DataTypeToSerialString(output_data_type).c_str());
  598. userdef_dtypes.push_back(
  599. std::to_string(index).append(":").append(TypeUtils::DataTypeToSerialString(output_data_type)));
  600. continue;
  601. }
  602. // Output_node is not set,check if is_output_adjust_hw_layout is set
  603. bool set_fp16_nc1hwc0 = NeedUpdateOutputFp16Nc1hwc0(src_op_desc, src_index);
  604. if (set_fp16_nc1hwc0) {
  605. // Set DT_FLOAT16 & FORMAT_NC1HWC0
  606. userdef_dtypes.push_back(std::to_string(index).append(":").append(TypeUtils::DataTypeToSerialString(DT_FLOAT16)));
  607. userdef_formats.push_back(
  608. std::to_string(index).append(":").append(TypeUtils::FormatToSerialString(FORMAT_NC1HWC0)));
  609. }
  610. }
  611. if (!userdef_dtypes.empty() && !ge::AttrUtils::SetListStr(op_desc, ATTR_ATC_USER_DEFINE_DATATYPE, userdef_dtypes)) {
  612. GELOGE(INTERNAL_ERROR, "Set user_define_dtype attr list for netoutput failed.");
  613. return INTERNAL_ERROR;
  614. }
  615. if (!userdef_formats.empty() && !ge::AttrUtils::SetListStr(op_desc, ATTR_ATC_USER_DEFINE_FORMAT, userdef_formats)) {
  616. GELOGE(INTERNAL_ERROR, "Set user_define_format attr list for netoutput failed.");
  617. return INTERNAL_ERROR;
  618. }
  619. return SUCCESS;
  620. }
  621. } // 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