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.

logical_stream_allocator.cc 27 kB

5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
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
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
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
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
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
5 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
5 years ago
4 years ago
4 years ago
4 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
5 years ago
4 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725
  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/build/logical_stream_allocator.h"
  17. #include <queue>
  18. #include "common/ge/ge_util.h"
  19. #include "framework/common/debug/ge_log.h"
  20. #include "framework/common/fmk_error_codes.h"
  21. #include "framework/common/types.h"
  22. #include "graph/debug/ge_attr_define.h"
  23. #include "graph/utils/graph_utils.h"
  24. #include "graph/common/ge_call_wrapper.h"
  25. using std::map;
  26. using std::set;
  27. using std::string;
  28. using std::vector;
  29. using std::queue;
  30. namespace ge {
  31. LogicalStreamPass::LogicalStreamPass(const string &name) : name_(name) {}
  32. const string &LogicalStreamPass::GetName() const {
  33. return name_;
  34. }
  35. bool LogicalStreamPass::IsEngineSkip(const Subgraph &subgraph) const {
  36. return subgraph.engine_conf.skip_assign_stream;
  37. }
  38. bool LogicalStreamPass::IsEngineAttach(const Subgraph &subgraph) const {
  39. return subgraph.engine_conf.attach;
  40. }
  41. bool LogicalStreamPass::IsEngineIndependent(const Subgraph &subgraph) const {
  42. return subgraph.engine_conf.independent;
  43. }
  44. bool LogicalStreamPass::HasStreamLabel(const Subgraph &subgraph) const {
  45. return !subgraph.subgraph_info.GetStreamLabel().empty();
  46. }
  47. bool LogicalStreamPass::HasAssignedStream(const Subgraph &subgraph) const {
  48. return subgraph.stream_id != kInvalidStream;
  49. }
  50. Status AssignByLabelPass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr> &subgraphs, Context &context) {
  51. bool changed = false;
  52. int64_t &next_stream = context.next_stream;
  53. map<string, int64_t> label_streams;
  54. for (const SubgraphPtr &subgraph : subgraphs) {
  55. const string &stream_label = subgraph->subgraph_info.GetStreamLabel();
  56. if (!stream_label.empty()) {
  57. // Subgraphs of the same stream_label are assigned to the same stream,
  58. // and different stream_labels are assigned new streams.
  59. auto iter = label_streams.find(stream_label);
  60. if (iter == label_streams.end()) {
  61. subgraph->stream_id = next_stream;
  62. GELOGI("Assign new stream %ld for label %s.", next_stream, stream_label.c_str());
  63. label_streams.emplace(stream_label, next_stream);
  64. next_stream++;
  65. } else {
  66. subgraph->stream_id = iter->second;
  67. }
  68. changed = true;
  69. }
  70. }
  71. return changed ? SUCCESS : NOT_CHANGED;
  72. }
  73. Status IndependentStreamPass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr> &subgraphs, Context &context) {
  74. bool changed = false;
  75. int64_t &next_stream = context.next_stream;
  76. // <engine, <label, stream>>
  77. map<string, map<string, int64_t>> engine_streams;
  78. for (const SubgraphPtr &subgraph : subgraphs) {
  79. if (!IsEngineIndependent(*subgraph)) {
  80. continue;
  81. }
  82. const string &engine = subgraph->engine_conf.id;
  83. const string &stream_label = subgraph->subgraph_info.GetStreamLabel();
  84. auto &label_streams = engine_streams[engine];
  85. auto iter = label_streams.find(stream_label);
  86. if (iter == label_streams.end()) {
  87. subgraph->stream_id = next_stream;
  88. GELOGI("Assign new independent stream %ld for engine %s (label: %s).", next_stream, engine.c_str(),
  89. stream_label.c_str());
  90. label_streams.emplace(stream_label, next_stream);
  91. next_stream++;
  92. } else {
  93. subgraph->stream_id = iter->second;
  94. }
  95. changed = true;
  96. }
  97. return changed ? SUCCESS : NOT_CHANGED;
  98. }
  99. Status AssignByDependencyPass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr> &subgraphs, Context &context) {
  100. bool changed = false;
  101. map<NodePtr, SubgraphPtr> end_subgraph_map;
  102. map<NodePtr, SubgraphPtr> pld_subgraph_map;
  103. InitEndSubgraphMap(subgraphs, end_subgraph_map);
  104. InitPldSubgraphMap(subgraphs, pld_subgraph_map);
  105. for (const SubgraphPtr &subgraph : subgraphs) {
  106. if (HasAssignedStream(*subgraph)) {
  107. continue;
  108. }
  109. SubgraphPtr reusable_subgraph = GetReusableSubgraph(subgraph, end_subgraph_map, pld_subgraph_map);
  110. if (reusable_subgraph == nullptr) {
  111. (void)AssignNewStream(subgraph);
  112. } else {
  113. if (HasAssignedStream(*reusable_subgraph)) {
  114. subgraph->stream_id = reusable_subgraph->stream_id;
  115. } else {
  116. int64_t stream_id = AssignNewStream(reusable_subgraph);
  117. subgraph->stream_id = stream_id;
  118. GELOGI("Reusable subgraph %s has not been assigned a stream, now assign new stream %ld.",
  119. reusable_subgraph->name.c_str(), stream_id);
  120. }
  121. if (reusable_subgraph->reused_subgraph != nullptr) {
  122. reusable_subgraph = reusable_subgraph->reused_subgraph;
  123. }
  124. subgraph->reused_subgraph = reusable_subgraph;
  125. reused_subgraphs_.emplace_back(subgraph, reusable_subgraph);
  126. GELOGI("Subgraph %s of engine %s reuses stream of subgraph %s of engine %s.", subgraph->name.c_str(),
  127. subgraph->engine_conf.id.c_str(), reusable_subgraph->name.c_str(),
  128. reusable_subgraph->engine_conf.id.c_str());
  129. }
  130. changed = true;
  131. }
  132. UpdateAssignedSubgraphs(context);
  133. UpdateReusedSubgraphs();
  134. return changed ? SUCCESS : NOT_CHANGED;
  135. }
  136. void AssignByDependencyPass::InitEndSubgraphMap(const vector<SubgraphPtr> &subgraphs,
  137. map<NodePtr, SubgraphPtr> &end_subgraph_map) {
  138. for (const auto &subgraph : subgraphs) {
  139. const SubGraphInfo &subgraph_info = subgraph->subgraph_info;
  140. for (const auto &item : subgraph_info.GetEnd2PldMap()) {
  141. end_subgraph_map.emplace(item.first, subgraph);
  142. }
  143. }
  144. }
  145. void AssignByDependencyPass::InitPldSubgraphMap(const vector<SubgraphPtr> &subgraphs,
  146. map<NodePtr, SubgraphPtr> &pld_subgraph_map) {
  147. for (const auto &subgraph : subgraphs) {
  148. const SubGraphInfo &subgraph_info = subgraph->subgraph_info;
  149. for (const auto &item : subgraph_info.GetPld2EndMap()) {
  150. pld_subgraph_map.emplace(item.first, subgraph);
  151. }
  152. }
  153. }
  154. bool AssignByDependencyPass::CouldReuse(const SubgraphPtr &subgraph, const SubgraphPtr &pred_subgraph,
  155. const map<NodePtr, SubgraphPtr> &pld_subgraph_map) {
  156. if ((subgraph == nullptr) || (pred_subgraph == nullptr)) {
  157. return false;
  158. }
  159. if (subgraph->engine_conf.scheduler_id != pred_subgraph->engine_conf.scheduler_id) {
  160. return false;
  161. }
  162. if (IsEngineIndependent(*pred_subgraph) || HasStreamLabel(*pred_subgraph)) {
  163. return false;
  164. }
  165. // If the engine of the predecessor subgraph is the same as the other successor subgraphs, the stream is not reused.
  166. for (const auto &end_pld_pair : pred_subgraph->subgraph_info.GetEnd2PldMap()) {
  167. auto iter = pld_subgraph_map.find(end_pld_pair.second);
  168. if (iter != pld_subgraph_map.end()) {
  169. const SubgraphPtr &pred_subgraph_succ = iter->second;
  170. if ((pred_subgraph_succ != subgraph) &&
  171. (pred_subgraph_succ->engine_conf.id == pred_subgraph->engine_conf.id)) {
  172. return false;
  173. }
  174. }
  175. }
  176. if ((subgraph->engine_conf.id == pred_subgraph->engine_conf.id) ||
  177. IsEngineAttach(*subgraph)) {
  178. return true;
  179. }
  180. if ((pred_subgraph->reused_subgraph != nullptr) &&
  181. (pred_subgraph->reused_subgraph->engine_conf.id == subgraph->engine_conf.id)) {
  182. return true;
  183. }
  184. return false;
  185. }
  186. LogicalStreamPass::SubgraphPtr AssignByDependencyPass::GetReusableSubgraph(
  187. const SubgraphPtr &subgraph, const map<NodePtr, SubgraphPtr> &end_subgraph_map,
  188. const map<NodePtr, SubgraphPtr> &pld_subgraph_map) {
  189. const SubGraphInfo &subgraph_info = subgraph->subgraph_info;
  190. for (const auto &pld_2_end : subgraph_info.GetPld2EndMap()) {
  191. const NodePtr &peer_end = pld_2_end.second;
  192. auto iter = end_subgraph_map.find(peer_end);
  193. if (iter != end_subgraph_map.end()) {
  194. const SubgraphPtr &pred_subgraph = iter->second;
  195. if (CouldReuse(subgraph, pred_subgraph, pld_subgraph_map)) {
  196. return pred_subgraph;
  197. }
  198. }
  199. }
  200. return nullptr;
  201. }
  202. int64_t AssignByDependencyPass::AssignNewStream(SubgraphPtr subgraph) {
  203. const string &engine_name = subgraph->engine_conf.id;
  204. int64_t max_parallel_num = subgraph->max_parallel_num;
  205. int64_t stream_id = 0;
  206. auto next_iter = engine_next_streams_.find(engine_name);
  207. if (next_iter != engine_next_streams_.end()) {
  208. stream_id = next_iter->second;
  209. }
  210. if (stream_id >= max_parallel_num) {
  211. stream_id = 0;
  212. }
  213. subgraph->stream_id = stream_id;
  214. engine_next_streams_[engine_name] = stream_id + 1;
  215. assigned_subgraphs_.emplace_back(subgraph);
  216. if ((stream_id + 1) > engine_stream_num_[engine_name]) {
  217. engine_stream_num_[engine_name] = stream_id + 1;
  218. }
  219. GELOGI("Subgraph %s assigns new temp stream %ld (engine: %s).", subgraph->name.c_str(), stream_id,
  220. engine_name.c_str());
  221. return stream_id;
  222. }
  223. void AssignByDependencyPass::UpdateAssignedSubgraphs(Context &context) {
  224. // If the default stream is valid, the first assigned stream will reuse the default stream id
  225. // and other streams use new id. To ensure that the id of the new stream is continuous,
  226. // we first subtract one from next_stream.
  227. int64_t to_be_updated_stream = kInvalidStream;
  228. if (context.default_stream != kInvalidStream) {
  229. context.next_stream--;
  230. to_be_updated_stream = context.next_stream;
  231. }
  232. // Update the starting stream id for each engine.
  233. int64_t &next_stream = context.next_stream;
  234. map<string, int64_t> engine_start_streams;
  235. for (const auto &item : engine_stream_num_) {
  236. int64_t stream_count = item.second;
  237. engine_start_streams[item.first] = next_stream;
  238. next_stream += stream_count;
  239. }
  240. // Update the subgraph streams assigned by engine.
  241. for (auto &subgraph : assigned_subgraphs_) {
  242. subgraph->stream_id += engine_start_streams[subgraph->engine_conf.id];
  243. if (subgraph->stream_id == to_be_updated_stream) {
  244. subgraph->stream_id = context.default_stream;
  245. GELOGI("Subgraph %s of engine %s reuses default stream %ld.", subgraph->name.c_str(),
  246. subgraph->engine_conf.id.c_str(), context.default_stream);
  247. } else {
  248. GELOGI("Stream of subgraph %s has been updated to %ld.", subgraph->name.c_str(), subgraph->stream_id);
  249. }
  250. }
  251. }
  252. void AssignByDependencyPass::UpdateReusedSubgraphs() {
  253. // Update streams for the subgraphs of reusing stream.
  254. for (const auto &item : reused_subgraphs_) {
  255. auto &cur_subgraph = item.first;
  256. auto &reused_graph = item.second;
  257. cur_subgraph->stream_id = reused_graph->stream_id;
  258. GELOGI("Stream of subgraph %s has been updated to %ld.", cur_subgraph->name.c_str(), cur_subgraph->stream_id);
  259. }
  260. }
  261. Status SingleStreamPass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr> &subgraphs, Context &context) {
  262. // context.default_stream can be kInvalidStream only when graph is the root graph.
  263. int64_t new_stream = context.default_stream;
  264. if (new_stream == kInvalidStream) {
  265. new_stream = context.next_stream;
  266. ++context.next_stream;
  267. }
  268. for (const SubgraphPtr &subgraph : subgraphs) {
  269. if (!HasAssignedStream(*subgraph)) {
  270. const string &stream_label = subgraph->subgraph_info.GetStreamLabel();
  271. if (!stream_label.empty()) {
  272. GELOGE(INTERNAL_ERROR, "Stream labels are not supported (subgraph: %s, stream label: %s).",
  273. subgraph->name.c_str(), stream_label.c_str());
  274. return INTERNAL_ERROR;
  275. }
  276. subgraph->stream_id = new_stream;
  277. }
  278. }
  279. return SUCCESS;
  280. }
  281. Status NodeStreamUpdatePass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr> &subgraphs, Context &context) {
  282. // Check if all subgraphs have been assigned a stream.
  283. for (const SubgraphPtr &subgraph : subgraphs) {
  284. const string &engine_name = subgraph->engine_conf.id;
  285. if (!IsEngineSkip(*subgraph) && !HasAssignedStream(*subgraph)) {
  286. GELOGE(INTERNAL_ERROR, "Subgraph %s has not yet been assigned a stream (engine: %s).", subgraph->name.c_str(),
  287. engine_name.c_str());
  288. return INTERNAL_ERROR;
  289. } else {
  290. GELOGI("Subgraph %s is assigned stream %ld (engine: %s).", subgraph->name.c_str(), subgraph->stream_id,
  291. engine_name.c_str());
  292. }
  293. }
  294. // Init the stream id of node.
  295. for (NodePtr &node : graph->GetDirectNode()) {
  296. GE_CHECK_NOTNULL(node->GetOpDesc());
  297. node->GetOpDesc()->SetStreamId(kInvalidStream);
  298. }
  299. // Set the stream id of the subgraph to the node.
  300. for (const SubgraphPtr &subgraph : subgraphs) {
  301. int64_t stream_id = subgraph->stream_id;
  302. const string &engine_name = subgraph->engine_conf.id;
  303. auto compute_graph = subgraph->subgraph_info.GetSubGraph();
  304. for (NodePtr &node : compute_graph->GetDirectNode()) {
  305. GE_CHECK_NOTNULL(node->GetOpDesc());
  306. if (node->GetOpDesc()->HasAttr(ATTR_NAME_RTS_LABEL_NODE)) {
  307. node->GetOpDesc()->SetStreamId(context.default_stream);
  308. GELOGD("Node %s of type %s in subgraph %s is assigned parent stream %ld (engine: %s).", node->GetName().c_str(),
  309. node->GetType().c_str(), subgraph->name.c_str(), context.default_stream, engine_name.c_str());
  310. } else if (IsEngineSkip(*subgraph) && node->GetInNodes().empty()) {
  311. GELOGD("Node %s of type %s in subgraph %s doesn't need to assign a stream (engine: %s).",
  312. node->GetName().c_str(), node->GetType().c_str(), subgraph->name.c_str(), engine_name.c_str());
  313. } else {
  314. node->GetOpDesc()->SetStreamId(stream_id);
  315. GELOGD("Node %s of type %s in subgraph %s is assigned stream %ld (engine: %s).", node->GetName().c_str(),
  316. node->GetType().c_str(), subgraph->name.c_str(), stream_id, engine_name.c_str());
  317. }
  318. }
  319. }
  320. return SUCCESS;
  321. }
  322. int64_t UpdateForSkippedEnginePass::GetSingleInoutStream(const NodePtr &node) const {
  323. set<int64_t> stream_ids;
  324. for (const auto &in_node : node->GetInAllNodes()) {
  325. GE_CHECK_NOTNULL_EXEC(in_node->GetOpDesc(), return kInvalidStream);
  326. int64_t stream_id = in_node->GetOpDesc()->GetStreamId();
  327. if (stream_id != kInvalidStream) {
  328. stream_ids.insert(stream_id);
  329. }
  330. }
  331. for (const auto &out_node : node->GetOutAllNodes()) {
  332. GE_CHECK_NOTNULL_EXEC(out_node->GetOpDesc(), return kInvalidStream);
  333. int64_t stream_id = out_node->GetOpDesc()->GetStreamId();
  334. if (stream_id != kInvalidStream) {
  335. stream_ids.insert(stream_id);
  336. }
  337. }
  338. if (stream_ids.size() == 1) {
  339. int64_t stream_id = *(stream_ids.begin());
  340. GELOGI("The stream of all input and output nodes of node %s (type: %s) is %ld.", node->GetName().c_str(),
  341. node->GetType().c_str(), stream_id);
  342. return stream_id;
  343. }
  344. return kInvalidStream;
  345. }
  346. Status UpdateForSkippedEnginePass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr> &subgraphs, Context &context) {
  347. set<OpDescPtr> ops_without_label;
  348. // Check if subgraph is engine skipped and without stream label or not
  349. for (const SubgraphPtr &subgraph : subgraphs) {
  350. if (IsEngineSkip(*subgraph)) {
  351. auto compute_graph = subgraph->subgraph_info.GetSubGraph();
  352. for (NodePtr &node : compute_graph->GetDirectNode()) {
  353. auto op_desc = node->GetOpDesc();
  354. GE_CHECK_NOTNULL(op_desc);
  355. auto stream_id = op_desc->GetStreamId();
  356. if ((stream_id != kInvalidStream) && !HasStreamLabel(*subgraph)) {
  357. ops_without_label.emplace(op_desc);
  358. }
  359. }
  360. }
  361. }
  362. // Try reassign the stream id
  363. for (ge::NodePtr &node : graph->GetDirectNode()) {
  364. auto op_desc = node->GetOpDesc();
  365. GE_CHECK_NOTNULL(op_desc);
  366. int64_t stream_id = op_desc->GetStreamId();
  367. if (ops_without_label.find(op_desc) != ops_without_label.end()) {
  368. if (AreAllPredStreamsInvalid(node) && op_desc->GetSubgraphInstanceNames().empty()) {
  369. op_desc->SetStreamId(kInvalidStream);
  370. GELOGI("Node %s of type %s reassign to stream %ld from stream %ld.", node->GetName().c_str(),
  371. node->GetType().c_str(), kInvalidStream, stream_id);
  372. } else if (!node->GetOutAllNodes().empty()) {
  373. int64_t inout_stream = GetSingleInoutStream(node);
  374. if (inout_stream != kInvalidStream) {
  375. op_desc->SetStreamId(inout_stream);
  376. GELOGI("Node %s of type %s reassign to stream %ld from stream %ld.", node->GetName().c_str(),
  377. node->GetType().c_str(), inout_stream, stream_id);
  378. }
  379. }
  380. }
  381. }
  382. return SUCCESS;
  383. }
  384. bool UpdateForSkippedEnginePass::AreAllPredStreamsInvalid(const NodePtr &node) const {
  385. for (const auto &pre_node : node->GetInAllNodes()) {
  386. auto pre_node_desc = pre_node->GetOpDesc();
  387. if (pre_node_desc != nullptr) {
  388. int64_t stream_id = pre_node_desc->GetStreamId();
  389. if (stream_id != kInvalidStream) {
  390. return false;
  391. }
  392. }
  393. }
  394. return true;
  395. }
  396. Status AllReduceParallelPass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr> &subgraphs, Context &context) {
  397. if (!context.enable_hcom_parallel) {
  398. return NOT_CHANGED;
  399. }
  400. GELOGI("AllReduceParallelPass is enabled.");
  401. GE_DUMP(graph, "BeforeAllReduceParallel");
  402. // All successors of HcomAllReduce.
  403. set<NodePtr> all_reduce_succs;
  404. for (const NodePtr &node : graph->GetDirectNode()) {
  405. if (!IsHcomNode(node->GetType()) ||
  406. (node->GetInDataNodes().size() <= 1)) {
  407. continue;
  408. }
  409. string reduce_stream_label;
  410. GE_CHECK_NOTNULL(node->GetOpDesc());
  411. (void)AttrUtils::GetStr(node->GetOpDesc(), ATTR_NAME_STREAM_LABEL, reduce_stream_label);
  412. set<NodePtr> cur_nodes = {node};
  413. while (!cur_nodes.empty()) {
  414. set<NodePtr> all_out_data_nodes;
  415. for (auto &curr_node : cur_nodes) {
  416. for (const NodePtr &out_node : curr_node->GetOutDataNodes()) {
  417. string out_stream_label;
  418. GE_CHECK_NOTNULL(out_node->GetOpDesc());
  419. (void)AttrUtils::GetStr(out_node->GetOpDesc(), ATTR_NAME_STREAM_LABEL, out_stream_label);
  420. // normally, Allreduce do not have streamLabel. when in horovod scenario Allreduce will have streamLabel
  421. bool isSuccessorParallel =
  422. (out_stream_label == reduce_stream_label) || (!reduce_stream_label.empty() && out_stream_label.empty());
  423. if (isSuccessorParallel) {
  424. all_reduce_succs.emplace(out_node);
  425. all_out_data_nodes.emplace(out_node);
  426. }
  427. }
  428. }
  429. cur_nodes = all_out_data_nodes;
  430. }
  431. }
  432. map<int64_t, int64_t> old_stream_to_new;
  433. for (const NodePtr &node : all_reduce_succs) {
  434. GE_CHECK_NOTNULL(node->GetOpDesc());
  435. auto old_stream = node->GetOpDesc()->GetStreamId();
  436. if (old_stream != kInvalidStream) {
  437. int64_t new_stream = kInvalidStream;
  438. auto iter = old_stream_to_new.find(old_stream);
  439. if (iter != old_stream_to_new.end()) {
  440. new_stream = iter->second;
  441. } else {
  442. new_stream = context.next_stream;
  443. context.next_stream++;
  444. old_stream_to_new.emplace(old_stream, new_stream);
  445. }
  446. if (!IsHcomNode(node->GetType())) {
  447. GELOGI("Stream of node %s has been updated from %ld to %ld.", node->GetName().c_str(), old_stream, new_stream);
  448. node->GetOpDesc()->SetStreamId(new_stream);
  449. }
  450. }
  451. }
  452. return !all_reduce_succs.empty() ? SUCCESS : NOT_CHANGED;
  453. }
  454. bool AllReduceParallelPass::IsHcomNode(const std::string& node_type) {
  455. return (node_type == HCOMALLREDUCE || node_type == HVDCALLBACKALLREDUCE);
  456. }
  457. LogicalStreamAllocator::LogicalStreamAllocator(const map<string, SchedulerConf> &scheduler_confs,
  458. const map<string, int> &max_parallel_num)
  459. : scheduler_confs_(scheduler_confs), max_parallel_num_(max_parallel_num) {}
  460. void LogicalStreamAllocator::EnableSingleStream(bool enable) { context_.enable_single_stream = enable; }
  461. void LogicalStreamAllocator::EnableHcomParallel(bool enable) { context_.enable_hcom_parallel = enable; }
  462. Status LogicalStreamAllocator::Assign(const ComputeGraphPtr &root_graph, const Graph2SubGraphInfoList &subgraph_map,
  463. int64_t &stream_num) {
  464. GE_CHECK_NOTNULL(root_graph);
  465. map<string, EngineConfPtr> engine_confs;
  466. GE_TIMESTAMP_START(InitEngineConfs);
  467. for (const auto &item : scheduler_confs_) {
  468. const SchedulerConf &scheduler = item.second;
  469. for (const auto &engine_pair : scheduler.cal_engines) {
  470. EngineConfPtr engine_conf = engine_pair.second;
  471. if (engine_conf != nullptr) {
  472. engine_confs[engine_pair.first] = engine_conf;
  473. }
  474. }
  475. }
  476. GE_TIMESTAMP_END(InitEngineConfs, "GraphBuilder::AssignStreamInitEngineConfs");
  477. Status status = DoAssign(root_graph, subgraph_map, engine_confs);
  478. if (status != SUCCESS) {
  479. GELOGE(status, "Assign streams failed.");
  480. return status;
  481. }
  482. vector<ComputeGraphPtr> subgraphs = root_graph->GetAllSubgraphs();
  483. for (const ComputeGraphPtr &subgraph : subgraphs) {
  484. Status status = DoAssign(subgraph, subgraph_map, engine_confs);
  485. if (status != SUCCESS) {
  486. GELOGE(status, "Assign streams failed.");
  487. return status;
  488. }
  489. }
  490. RefreshContinuousStreams(root_graph);
  491. stream_num = context_.next_stream;
  492. GELOGI("Assigned logical stream num: %ld.", stream_num);
  493. return SUCCESS;
  494. }
  495. Status LogicalStreamAllocator::DoAssign(const ComputeGraphPtr &graph, const Graph2SubGraphInfoList &subgraph_map,
  496. const map<string, EngineConfPtr> &engine_confs) {
  497. GE_CHECK_NOTNULL(graph);
  498. NodePtr parent_node = graph->GetParentNode();
  499. if ((parent_node == nullptr) || (parent_node->GetOpDesc() == nullptr)) {
  500. context_.default_stream = kInvalidStream;
  501. } else {
  502. context_.default_stream = parent_node->GetOpDesc()->GetStreamId();
  503. }
  504. auto iter = subgraph_map.find(graph);
  505. if (iter == subgraph_map.end()) {
  506. GELOGE(FAILED, "Graph %s not found.", graph->GetName().c_str());
  507. return FAILED;
  508. }
  509. const vector<SubGraphInfoPtr> &subgraph_info_list = iter->second;
  510. vector<SubgraphPtr> subgraphs;
  511. GE_TIMESTAMP_START(ConvertSubgraphs);
  512. Status status = ConvertSubgraphs(subgraph_info_list, engine_confs, subgraphs);
  513. GE_TIMESTAMP_END(ConvertSubgraphs, "GraphBuilder::AssignStreamConvertSubgraphs");
  514. if (status != SUCCESS) {
  515. GELOGE(status, "Create subgraphs failed.");
  516. return status;
  517. }
  518. GELOGD("Subgraphs of graph %s", graph->GetName().c_str());
  519. for (const auto &subgraph : subgraphs) {
  520. if (subgraph != nullptr) {
  521. GELOGD("subgraph: %s", subgraph->name.c_str());
  522. }
  523. }
  524. return RunPasses(graph, subgraphs);
  525. }
  526. Status LogicalStreamAllocator::ConvertSubgraphs(const vector<SubGraphInfoPtr> &subgraph_infos,
  527. const map<string, EngineConfPtr> &engine_confs,
  528. vector<SubgraphPtr> &subgraphs) {
  529. for (auto &subgraph_info : subgraph_infos) {
  530. GE_CHECK_NOTNULL(subgraph_info);
  531. string subgraph_name;
  532. ComputeGraphPtr computer_graph = subgraph_info->GetSubGraph();
  533. if (computer_graph != nullptr) {
  534. subgraph_name = computer_graph->GetName();
  535. }
  536. const string &engine_name = subgraph_info->GetEngineName();
  537. auto engine_conf_iter = engine_confs.find(engine_name);
  538. if ((engine_conf_iter == engine_confs.end()) || (engine_conf_iter->second == nullptr)) {
  539. GELOGE(INTERNAL_ERROR, "Engine conf of subgraph %s not found (engine name: %s).", subgraph_name.c_str(),
  540. engine_name.c_str());
  541. return INTERNAL_ERROR;
  542. }
  543. SubgraphPtr subgraph = MakeShared<Subgraph>(*subgraph_info, *engine_conf_iter->second);
  544. GE_CHECK_NOTNULL(subgraph);
  545. subgraph->name = subgraph_name;
  546. auto parallel_iter = max_parallel_num_.find(engine_name);
  547. if (parallel_iter != max_parallel_num_.end()) {
  548. subgraph->max_parallel_num = parallel_iter->second;
  549. }
  550. subgraphs.emplace_back(subgraph);
  551. }
  552. return SUCCESS;
  553. }
  554. Status LogicalStreamAllocator::RunPasses(const ComputeGraphPtr &graph, const vector<SubgraphPtr> &subgraphs) {
  555. vector<LogicalStreamPassPtr> passes;
  556. if (context_.enable_single_stream) {
  557. passes.emplace_back(MakeShared<SingleStreamPass>());
  558. passes.emplace_back(MakeShared<NodeStreamUpdatePass>());
  559. passes.emplace_back(MakeShared<UpdateForSkippedEnginePass>());
  560. } else {
  561. passes.emplace_back(MakeShared<AssignByLabelPass>());
  562. passes.emplace_back(MakeShared<IndependentStreamPass>());
  563. passes.emplace_back(MakeShared<AssignByDependencyPass>());
  564. passes.emplace_back(MakeShared<NodeStreamUpdatePass>());
  565. passes.emplace_back(MakeShared<AllReduceParallelPass>());
  566. passes.emplace_back(MakeShared<UpdateForSkippedEnginePass>());
  567. }
  568. for (auto &pass : passes) {
  569. GE_CHECK_NOTNULL(pass);
  570. Status status = pass->Run(graph, subgraphs, context_);
  571. if (status == SUCCESS) {
  572. GELOGD("Stream pass %s return SUCCESS.", pass->GetName().c_str());
  573. } else if (status == NOT_CHANGED) {
  574. GELOGD("Stream pass %s return NOT_CHANGED.", pass->GetName().c_str());
  575. } else {
  576. GELOGE(status, "Stream pass %s failed.", pass->GetName().c_str());
  577. return status;
  578. }
  579. }
  580. return SUCCESS;
  581. }
  582. void LogicalStreamAllocator::RefreshContinuousStreams(const ComputeGraphPtr &graph) {
  583. int64_t stream_num = context_.next_stream;
  584. vector<bool> stream_has_node(stream_num);
  585. for (const NodePtr &node : graph->GetNodes(graph->GetGraphUnknownFlag())) {
  586. if (node != nullptr) {
  587. auto op_desc = node->GetOpDesc();
  588. if (op_desc != nullptr) {
  589. int64_t stream_id = op_desc->GetStreamId();
  590. if ((stream_id != kInvalidStream) && (stream_id < stream_num)) {
  591. stream_has_node[stream_id] = true;
  592. }
  593. }
  594. }
  595. }
  596. context_.next_stream = 0;
  597. vector<int64_t> old_to_new_streams(stream_num, kInvalidStream);
  598. for (size_t old_stream = 0; old_stream < stream_has_node.size(); old_stream++) {
  599. if (stream_has_node[old_stream]) {
  600. old_to_new_streams[old_stream] = context_.next_stream;
  601. context_.next_stream++;
  602. }
  603. }
  604. for (const NodePtr &node : graph->GetNodes(graph->GetGraphUnknownFlag())) {
  605. auto op_desc = node->GetOpDesc();
  606. if (op_desc != nullptr) {
  607. int64_t stream_id = op_desc->GetStreamId();
  608. if ((stream_id != kInvalidStream) && (stream_id < stream_num)) {
  609. op_desc->SetStreamId(old_to_new_streams[stream_id]);
  610. }
  611. }
  612. }
  613. }
  614. } // 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