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.

inner_session.cc 25 kB

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
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
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
5 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
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
4 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
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568
  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 "session/inner_session.h"
  17. #include <map>
  18. #include <memory>
  19. #include <vector>
  20. #include "analyzer/analyzer.h"
  21. #include "adx_datadump_server.h"
  22. #include "common/dump/dump_properties.h"
  23. #include "common/dump/dump_manager.h"
  24. #include "framework/common/util.h"
  25. #include "framework/common/debug/ge_log.h"
  26. #include "graph/ge_context.h"
  27. #include "graph/ge_global_options.h"
  28. #include "graph/ge_local_context.h"
  29. #include "common/local_context.h"
  30. #include "graph/manager/graph_var_manager.h"
  31. #include "graph/utils/tensor_adapter.h"
  32. #include "runtime/mem.h"
  33. #include "ir_build/option_utils.h"
  34. #include "common/profiling/profiling_manager.h"
  35. namespace ge {
  36. namespace {
  37. const int32_t kDumpStatus = 0;
  38. Status CheckReuseMemoryOption(const std::map<string, string> &options) {
  39. auto iter = options.find(OPTION_EXEC_DISABLE_REUSED_MEMORY);
  40. if (iter != options.end()) {
  41. if (iter->second == "0") {
  42. GELOGD("%s=0, reuse memory is open", OPTION_EXEC_DISABLE_REUSED_MEMORY);
  43. } else if (iter->second == "1") {
  44. GELOGD("%s=1, reuse memory is close", OPTION_EXEC_DISABLE_REUSED_MEMORY);
  45. } else {
  46. GELOGE(PARAM_INVALID, "[CheckReuse][MemoryOption]option %s=%s is invalid",
  47. OPTION_EXEC_DISABLE_REUSED_MEMORY, iter->second.c_str());
  48. REPORT_INNER_ERROR("E19999", "CheckReuseMemoryOption failed because option %s=%s is invalid.",
  49. OPTION_EXEC_DISABLE_REUSED_MEMORY, iter->second.c_str());
  50. return FAILED;
  51. }
  52. }
  53. return SUCCESS;
  54. }
  55. }
  56. static std::mutex mutex_; // BuildGraph and RunGraph use
  57. bool InnerSession::is_dump_server_inited_ = false;
  58. InnerSession::InnerSession(uint64_t session_id, const std::map<string, string> &options)
  59. : init_flag_(false), session_id_(session_id), options_(options) {}
  60. Status InnerSession::Initialize() {
  61. if (init_flag_) {
  62. GELOGW("[InnerSession:%lu] session already initialize.", session_id_);
  63. return SUCCESS;
  64. }
  65. // If the global options and the session options are duplicated, the session options is preferred.
  66. auto all_options = options_;
  67. all_options.insert(GetMutableGlobalOptions().begin(), GetMutableGlobalOptions().end());
  68. Status ret = CheckReuseMemoryOption(all_options);
  69. if (ret != SUCCESS) {
  70. GELOGE(ret, "[CheckReuse][MemoryOption] failed, [InnerSession:%lu].", session_id_);
  71. REPORT_CALL_ERROR("E19999", "CheckReuseMemoryOption failed, InnerSession=%lu.", session_id_);
  72. return ret;
  73. }
  74. //Check option OP_PRECISION_MODE
  75. auto iter = all_options.find(ge::OP_PRECISION_MODE);
  76. if (iter != all_options.end() && !iter->second.empty() && !ge::CheckInputPathValid(iter->second)) {
  77. REPORT_INPUT_ERROR("E10001", std::vector<std::string>({"parameter", "value", "reason"}),
  78. std::vector<std::string>({ge::OP_PRECISION_MODE, iter->second, "path is not found"}));
  79. GELOGE(PARAM_INVALID, "[Check][OP_PRECISION_MODE] %s not found", iter->second.c_str());
  80. return FAILED;
  81. }
  82. if (iter != all_options.end()) {
  83. GELOGI("Option set successfully, option_key=%s, option_value=%s",
  84. ge::OP_PRECISION_MODE.c_str(), iter->second.c_str());
  85. }
  86. // Check option modify_mixlist
  87. if (ge::CheckModifyMixlistParamValid(all_options) != ge::SUCCESS) {
  88. return FAILED;
  89. }
  90. UpdateThreadContext(std::map<std::string, std::string>{});
  91. // session device id set here
  92. std::string str_session_device_id;
  93. if (GetContext().GetOption("ge.session_device_id", str_session_device_id) == SUCCESS) {
  94. GELOGI("Option session device id has set, value is %s.", str_session_device_id.c_str());
  95. uint32_t session_device_id = 0;
  96. try {
  97. session_device_id = static_cast<uint32_t>(std::stoi(str_session_device_id.c_str()));
  98. // session device id has priority
  99. GetContext().SetCtxDeviceId(session_device_id);
  100. } catch (std::invalid_argument &) {
  101. GELOGW("session device id %s transform to int failed.", str_session_device_id.c_str());
  102. } catch (std::out_of_range &) {
  103. GELOGW("session device id %s transform to int failed.", str_session_device_id.c_str());
  104. }
  105. }
  106. GE_CHK_RT_RET(rtSetDevice(GetContext().DeviceId()));
  107. DumpProperties dump_properties;
  108. GE_CHK_STATUS_RET(dump_properties.InitByOptions(), "Init dump properties failed.");
  109. GE_CHK_STATUS_RET(AddDumpProperties(dump_properties), "[Add][DumpProperties] failed.");
  110. ret = InnerInitialize();
  111. if (ret != SUCCESS) {
  112. GELOGE(ret, "[Init][GraphManager] failed, InnerSession:%lu.", session_id_);
  113. REPORT_CALL_ERROR("E19999", "GraphManager initialize failed, InnerSession:%lu.", session_id_);
  114. GE_CHK_STATUS(RemoveDumpProperties(), "[Remove][DumpProperties] failed.");
  115. return ret;
  116. }
  117. int32_t version = static_cast<int32_t>(SessionVersion::ClOUD_VERSION);
  118. const int DEFAULT_DEVICE_ID = 0;
  119. const int DEFAULT_JOB_ID = 0;
  120. ret = VarManager::Instance(session_id_)->Init(version, session_id_, DEFAULT_DEVICE_ID, DEFAULT_JOB_ID);
  121. if (ret != SUCCESS) {
  122. GELOGE(ret, "[Init][VarManager] failed.");
  123. REPORT_CALL_ERROR("E19999", "VarManager init failed, InnerSession:%lu.", session_id_);
  124. GE_CHK_STATUS(RemoveDumpProperties(), "[Remove][DumpProperties] failed.");
  125. }
  126. init_flag_ = true;
  127. return SUCCESS;
  128. }
  129. Status InnerSession::Finalize() {
  130. std::lock_guard<std::mutex> lock(resource_mutex_);
  131. if (!init_flag_) {
  132. GELOGW("[InnerSession:%lu] session does not initialize.", session_id_);
  133. return SUCCESS;
  134. }
  135. UpdateThreadContext(std::map<std::string, std::string>{});
  136. Status ret = InnerFinalize();
  137. if (ret != SUCCESS) {
  138. // Subsequent code execution is required, so no return is required
  139. GELOGE(ret, "[Finalize][GraphManager] failed, InnerSession:%lu.", session_id_);
  140. REPORT_CALL_ERROR("E19999", "GraphManager Finalize failed, InnerSession:%lu.", session_id_);
  141. }
  142. init_flag_ = false;
  143. // release analyzer saved info(Session Level)
  144. Analyzer::GetInstance()->DestroySessionJsonObject(session_id_);
  145. GE_CHK_RT(rtDeviceReset(static_cast<int32_t>(GetContext().DeviceId())));
  146. GE_CHK_STATUS_RET(RemoveDumpProperties(), "[Remove][DumpProperties] failed.");
  147. return ret;
  148. }
  149. Status InnerSession::InnerInitialize() {
  150. Status ret = model_executor_.Initialize(options_, session_id_);
  151. if (ret != SUCCESS) {
  152. GELOGE(ret, "[Init][GraphExecutor] failed, InnerSession:%lu.", session_id_);
  153. REPORT_CALL_ERROR("E19999", "GraphExecutor initialize failed, InnerSession:%lu.", session_id_);
  154. GE_CHK_STATUS(RemoveDumpProperties(), "[Remove][DumpProperties] failed.");
  155. return ret;
  156. }
  157. ret = graph_manager_.Initialize(options_, &model_executor_);
  158. if (ret != SUCCESS) {
  159. GELOGE(ret, "[Init][GraphManager] failed, InnerSession:%lu.", session_id_);
  160. REPORT_CALL_ERROR("E19999", "GraphManager initialize failed, InnerSession:%lu.", session_id_);
  161. GE_CHK_STATUS(RemoveDumpProperties(), "[Remove][DumpProperties] failed.");
  162. return ret;
  163. }
  164. return SUCCESS;
  165. }
  166. Status InnerSession::InnerFinalize() {
  167. Status ret = graph_manager_.Finalize();
  168. if (ret != SUCCESS) {
  169. // Subsequent code execution is required, so no return is required
  170. GELOGE(ret, "[Finalize][GraphManager] failed, InnerSession:%lu.", session_id_);
  171. REPORT_CALL_ERROR("E19999", "GraphManager Finalize failed, InnerSession:%lu.", session_id_);
  172. }
  173. ret = model_executor_.Finalize();
  174. if (ret != SUCCESS) {
  175. // Subsequent code execution is required, so no return is required
  176. GELOGE(ret, "[Finalize][GraphExecutor] failed, InnerSession:%lu.", session_id_);
  177. REPORT_CALL_ERROR("E19999", "GraphExecutor Finalize failed, InnerSession:%lu.", session_id_);
  178. }
  179. return SUCCESS;
  180. }
  181. Status InnerSession::GetVariable(const std::string &name, Tensor &val) {
  182. UpdateThreadContext(std::map<std::string, std::string>{});
  183. return graph_manager_.GetVariable(name, val);
  184. }
  185. Status InnerSession::AddGraph(uint32_t graph_id, const Graph &graph) {
  186. std::map<std::string, std::string> options;
  187. auto device_id = GetContext().DeviceId();
  188. GELOGD("Device id is %u", device_id);
  189. ProfilingManager::Instance().SetGraphIdToDeviceMap(graph_id, device_id);
  190. return AddGraph(graph_id, graph, options);
  191. }
  192. Status InnerSession::AddGraph(uint32_t graph_id, const Graph &graph,
  193. const std::map<std::string, std::string> &options) {
  194. std::lock_guard<std::mutex> lock(resource_mutex_);
  195. if (!init_flag_) {
  196. GELOGE(GE_SESS_INIT_FAILED, "[Add][Graph] failed because GraphManager not init, InnerSession:%lu, graph_id:%u.",
  197. session_id_, graph_id);
  198. REPORT_INNER_ERROR("E19999", "AddGraph failed because GraphManager not init, InnerSession:%lu, graph_id:%u.",
  199. session_id_, graph_id);
  200. return GE_SESS_INIT_FAILED;
  201. }
  202. UpdateThreadContext(options);
  203. Status ret = graph_manager_.AddGraph(graph_id, graph, options, domi::GetContext());
  204. if (ret != SUCCESS) {
  205. GELOGE(ret, "[Add][Graph] failed, InnerSession:%lu graphid: %u.", session_id_, graph_id);
  206. REPORT_CALL_ERROR("E19999", "GraphManager AddGraph failed, InnerSession:%lu graphid: %u.", session_id_, graph_id);
  207. return ret;
  208. }
  209. GELOGI("[InnerSession:%lu] add graph success, graph_id=%u.", session_id_, graph_id);
  210. return SUCCESS;
  211. }
  212. Status InnerSession::AddGraphWithCopy(uint32_t graph_id, const Graph &graph,
  213. const std::map<std::string, std::string> &options) {
  214. std::lock_guard<std::mutex> lock(resource_mutex_);
  215. if (!init_flag_) {
  216. GELOGE(GE_SESS_INIT_FAILED, "[Add][Graph] failed because GraphManager not init, InnerSession:%lu, graph_id:%u.",
  217. session_id_, graph_id);
  218. REPORT_INNER_ERROR("E19999",
  219. "AddGraphWithCopy failed because GraphManager not init, InnerSession:%lu, graph_id:%u.",
  220. session_id_, graph_id);
  221. return GE_SESS_INIT_FAILED;
  222. }
  223. UpdateThreadContext(options);
  224. Status ret = graph_manager_.AddGraphWithCopy(graph_id, graph, options, domi::GetContext());
  225. if (ret != SUCCESS) {
  226. GELOGE(ret, "[Add][Graph] failed, InnerSession:%lu graphid: %u.", session_id_, graph_id);
  227. REPORT_CALL_ERROR("E19999",
  228. "GraphManager AddGraphWithCopy failed, InnerSession:%lu graphid: %u.", session_id_, graph_id);
  229. return ret;
  230. }
  231. GELOGI("[InnerSession:%lu] add graph success, graph_id=%u.", session_id_, graph_id);
  232. return SUCCESS;
  233. }
  234. Status InnerSession::RunGraph(uint32_t graph_id, const std::vector<Tensor> &inputs, std::vector<Tensor> &outputs) {
  235. GELOGI("[InnerSession:%lu] run graph on session, graph_id=%u.", session_id_, graph_id);
  236. if (mutex_.try_lock()) {
  237. std::lock_guard<std::mutex> lock(mutex_, std::adopt_lock);
  238. if (!init_flag_) {
  239. GELOGE(GE_SESS_INIT_FAILED, "[Run][Graph]failed because GraphManager not Init, InnerSession:%lu, graph_id:%u.",
  240. session_id_, graph_id);
  241. REPORT_INNER_ERROR("E19999", "RunGraph failed because GraphManager not Init, InnerSession:%lu, graph_id:%u.",
  242. session_id_, graph_id);
  243. return GE_SESS_INIT_FAILED;
  244. }
  245. UpdateThreadContext(graph_id);
  246. vector<GeTensor> geInputs;
  247. for (auto &item : inputs) {
  248. geInputs.push_back(TensorAdapter::AsGeTensor(item));
  249. }
  250. vector<GeTensor> geOutputs;
  251. Status ret = graph_manager_.RunGraph(graph_id, geInputs, geOutputs, session_id_);
  252. domi::GetContext().out_nodes_map.clear();
  253. domi::GetContext().user_out_nodes.clear();
  254. if (ret != SUCCESS) {
  255. GELOGE(ret, "[Run][Graph]failed, InnerSession:%lu graph_id=%u.", session_id_, graph_id);
  256. REPORT_CALL_ERROR("E19999",
  257. "GraphManager RunGraph failed, InnerSession:%lu graph_id=%u.", session_id_, graph_id);
  258. return ret;
  259. }
  260. outputs.clear();
  261. for (auto &item : geOutputs) {
  262. outputs.push_back(TensorAdapter::AsTensor(item));
  263. }
  264. GELOGI("[InnerSession:%lu] run graph success, graph_id=%u.", session_id_, graph_id);
  265. return SUCCESS;
  266. } else {
  267. GELOGE(GE_SESS_ALREADY_RUNNING, "[Run][Graph]failed, InnerSession:%lu, graph_id=%u.", session_id_, graph_id);
  268. REPORT_INNER_ERROR("E19999",
  269. "RunGraph failed because mutex try_lock false, InnerSession:%lu, graph_id=%u.",
  270. session_id_, graph_id);
  271. return GE_SESS_ALREADY_RUNNING;
  272. }
  273. }
  274. Status InnerSession::RunGraphWithStreamAsync(uint32_t graph_id, rtStream_t stream,
  275. const std::vector<Tensor> &inputs, std::vector<Tensor> &outputs) {
  276. GELOGI("Run graph with stream, session id = %lu, graph id = %u, stream = %p.",
  277. session_id_, graph_id, stream);
  278. if (mutex_.try_lock()) {
  279. std::lock_guard<std::mutex> lock(mutex_, std::adopt_lock);
  280. if (!init_flag_) {
  281. GELOGE(GE_SESS_INIT_FAILED, "[Run][GraphWithStream]failed because GraphManager not Init,"
  282. "session id = %lu, graph id = %u, stream = %p.", session_id_, graph_id, stream);
  283. REPORT_INNER_ERROR("E19999", "RunGraphWithStreamAsync failed because GraphManager not Init,"
  284. "session id = %lu, graph id = %u, stream = %p.", session_id_, graph_id, stream);
  285. return GE_SESS_INIT_FAILED;
  286. }
  287. UpdateThreadContext(graph_id);
  288. vector<GeTensor> ge_inputs;
  289. for (auto &item : inputs) {
  290. ge_inputs.emplace_back(TensorAdapter::AsGeTensor(item));
  291. }
  292. vector<GeTensor> ge_outputs;
  293. for (auto &item : outputs) {
  294. ge_outputs.emplace_back(TensorAdapter::AsGeTensor(item));
  295. }
  296. Status ret = graph_manager_.RunGraphWithStreamAsync(graph_id, stream, session_id_, ge_inputs, ge_outputs);
  297. domi::GetContext().out_nodes_map.clear();
  298. domi::GetContext().user_out_nodes.clear();
  299. if (ret != SUCCESS) {
  300. GELOGE(ret, "[Run][GraphWithStreamAsync]failed,"
  301. "session id = %lu, graph id = %u, stream = %p.", session_id_, graph_id, stream);
  302. REPORT_CALL_ERROR("E19999", "GraphManager RunGrapWithStreamhAsync failed,"
  303. "session id = %lu, graph id = %u, stream = %p.", session_id_, graph_id, stream);
  304. return ret;
  305. }
  306. GELOGI("Run graph with stream success, session id = %lu, graph id = %u, stream = %p.",
  307. session_id_, graph_id, stream);
  308. return SUCCESS;
  309. } else {
  310. GELOGE(GE_SESS_ALREADY_RUNNING, "[Run][GraphWithStreamAsync]failed because mutex try_lock false,"
  311. "session id = %lu, graph id = %u, stream = %p.", session_id_, graph_id, stream);
  312. REPORT_INNER_ERROR("E19999", "[Run][GraphWithStreamAsync]failed failed because mutex try_lock false,"
  313. "session id = %lu, graph id = %u, stream = %p.", session_id_, graph_id, stream);
  314. return GE_SESS_ALREADY_RUNNING;
  315. }
  316. }
  317. Status InnerSession::RemoveGraph(uint32_t graph_id) {
  318. std::lock_guard<std::mutex> lock(resource_mutex_);
  319. if (!init_flag_) {
  320. GELOGE(GE_SESS_INIT_FAILED,
  321. "[Remove][Graph] failed because GraphManager not init, InnerSession:%lu, graph_id=%u.",
  322. session_id_, graph_id);
  323. REPORT_INNER_ERROR("E19999",
  324. "RemoveGraph failed, because GraphManager not init, InnerSession:%lu, graph_id=%u.",
  325. session_id_, graph_id);
  326. return GE_SESS_INIT_FAILED;
  327. }
  328. UpdateThreadContext(graph_id);
  329. Status ret = graph_manager_.RemoveGraph(graph_id);
  330. if (ret != SUCCESS) {
  331. GELOGE(ret, "[Remove][Graph] failed, InnerSession:%lu, graph_id=%u.", session_id_, graph_id);
  332. REPORT_CALL_ERROR("E19999",
  333. "GraphManager RemoveGraph failed, InnerSession:%lu, graph_id=%u.", session_id_, graph_id);
  334. return ret;
  335. }
  336. GELOGI("[InnerSession:%lu] remove graph success, graph_id=%u.", session_id_, graph_id);
  337. return SUCCESS;
  338. }
  339. Status InnerSession::RegisterCallBackFunc(
  340. const std::string &key,
  341. const std::function<Status(uint32_t, const std::map<std::string, ge::Tensor> &)> &callback) {
  342. std::lock_guard<std::mutex> lock(resource_mutex_);
  343. if (!init_flag_) {
  344. GELOGE(GE_SESS_INIT_FAILED,
  345. "[Register][CallBackFunc] failed because GraphManager not initialize, InnerSession:%lu.", session_id_);
  346. REPORT_INNER_ERROR("E19999",
  347. "RegisterCallBackFunc failed because GraphManager not init, InnerSession:%lu.", session_id_);
  348. return GE_SESS_INIT_FAILED;
  349. }
  350. UpdateThreadContext(std::map<std::string, std::string>{});
  351. Status ret = graph_manager_.RegisterCallBackFunc(key, callback);
  352. if (ret != SUCCESS) {
  353. GELOGE(ret, "[Register][CallBackFunc] failed, InnerSession:%lu register %s.", session_id_, key.c_str());
  354. REPORT_CALL_ERROR("E19999",
  355. "GraphManager RegisterCallBackFunc failed, InnerSession:%lu register %s.",
  356. session_id_, key.c_str());
  357. return ret;
  358. }
  359. GELOGI("[InnerSession:%lu] register %s callback function success.", session_id_, key.c_str());
  360. return SUCCESS;
  361. }
  362. Status InnerSession::RegisterCallBackFunc(
  363. const std::string &key,
  364. const std::function<Status(uint32_t, const std::map<AscendString, ge::Tensor> &)> &callback) {
  365. std::lock_guard<std::mutex> lock(resource_mutex_);
  366. if (!init_flag_) {
  367. GELOGE(GE_SESS_INIT_FAILED,
  368. "[Register][CallBackFunc]failed because GraphManager not initialize, InnerSession:%lu.", session_id_);
  369. REPORT_INNER_ERROR("E19999",
  370. "RegisterCallBackFunc failed because GraphManager not initialize, InnerSession:%lu.",
  371. session_id_);
  372. return GE_SESS_INIT_FAILED;
  373. }
  374. UpdateThreadContext(std::map<std::string, std::string>{});
  375. Status ret = graph_manager_.RegisterCallBackFunc(key, callback);
  376. if (ret != SUCCESS) {
  377. GELOGE(ret, "[Register][CallBackFunc] failed, InnerSession:%lu register %s.", session_id_, key.c_str());
  378. REPORT_CALL_ERROR("E19999",
  379. "GraphManager RegisterCallBackFunc failed, InnerSession:%lu register %s.",
  380. session_id_, key.c_str());
  381. return ret;
  382. }
  383. GELOGI("[InnerSession:%lu] register %s callback function success.", session_id_, key.c_str());
  384. return SUCCESS;
  385. }
  386. Status InnerSession::BuildGraph(uint32_t graph_id, const std::vector<InputTensorInfo> &inputs) {
  387. UpdateThreadContext(graph_id);
  388. GELOGI("[InnerSession:%lu] build graph on session, graph_id=%u.", session_id_, graph_id);
  389. std::vector<ge::GeTensor> ge_inputs;
  390. for (auto const &input : inputs) {
  391. std::vector<int64_t> input_dims;
  392. std::transform(input.dims.begin(), input.dims.end(), std::back_inserter(input_dims),
  393. [](int64_t x) -> int64_t { return x; });
  394. GeShape input_shape(input_dims);
  395. GeTensorDesc input_tensor_desc;
  396. input_tensor_desc.SetShape(input_shape);
  397. input_tensor_desc.SetDataType(static_cast<ge::DataType>(input.data_type));
  398. ge_inputs.emplace_back(input_tensor_desc);
  399. }
  400. GeRootModelPtr ge_root_model = nullptr;
  401. Status ret = graph_manager_.BuildGraph(graph_id, ge_inputs, ge_root_model, session_id_, true);
  402. if (ret != SUCCESS) {
  403. GELOGE(ret, "[Build][Graph] failed, InnerSession:%lu graph_id=%u.", session_id_, graph_id);
  404. REPORT_CALL_ERROR("E19999",
  405. "GraphManager BuildGraph failed, InnerSession:%lu graph_id=%u.", session_id_, graph_id);
  406. return ret;
  407. }
  408. GELOGI("[InnerSession:%lu] build graph success, graph_id=%u.", session_id_, graph_id);
  409. return ret;
  410. }
  411. Status InnerSession::BuildGraph(uint32_t graph_id, const std::vector<ge::Tensor> &inputs) {
  412. UpdateThreadContext(graph_id);
  413. GELOGI("[InnerSession:%lu] build graph on session, graph_id=%u.", session_id_, graph_id);
  414. std::vector<ge::GeTensor> ge_inputs;
  415. for (const auto &input : inputs) {
  416. ge_inputs.emplace_back(TensorAdapter::AsGeTensor(input));
  417. }
  418. GeRootModelPtr ge_root_model = nullptr;
  419. Status ret = graph_manager_.BuildGraph(graph_id, ge_inputs, ge_root_model, session_id_, true);
  420. if (ret != SUCCESS) {
  421. GELOGE(ret, "[Build][Graph] failed, InnerSession:%lu graph_id=%u.", session_id_, graph_id);
  422. REPORT_CALL_ERROR("E19999",
  423. "GraphManager BuildGraph failed, InnerSession:%lu graph_id=%u.", session_id_, graph_id);
  424. return ret;
  425. }
  426. GELOGI("[InnerSession:%lu] build graph success, graph_id=%u.", session_id_, graph_id);
  427. return ret;
  428. }
  429. Status InnerSession::RunGraphAsync(uint32_t graph_id, const std::vector<ge::Tensor> &inputs,
  430. RunAsyncCallback callback) {
  431. UpdateThreadContext(graph_id);
  432. GELOGI("[InnerSession:%lu] run graph on session, graph_id=%u.", session_id_, graph_id);
  433. Status ret = graph_manager_.RunGraphAsync(graph_id, inputs, session_id_, callback);
  434. if (ret != SUCCESS) {
  435. GELOGE(ret, "[Run][GraphAsync]failed, InnerSession:%lu graph_id=%u.", session_id_, graph_id);
  436. REPORT_CALL_ERROR("E19999",
  437. "GraphManager RunGraphAsync failed, InnerSession:%lu graph_id=%u.", session_id_, graph_id);
  438. return ret;
  439. }
  440. GELOGI("[InnerSession:%lu] run graph success, graph_id=%u.", session_id_, graph_id);
  441. return ret;
  442. }
  443. const GraphManager &InnerSession::getGraphManagerObj() const { return graph_manager_; }
  444. void InnerSession::UpdateThreadContext(const std::map<std::string, std::string> &options) {
  445. GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
  446. GetThreadLocalContext().SetSessionOption(options_);
  447. GetThreadLocalContext().SetGraphOption(options);
  448. GetContext().SetSessionId(session_id_);
  449. SetRtSocVersion();
  450. }
  451. void InnerSession::UpdateThreadContext(uint32_t graph_id) {
  452. auto options = graph_manager_.GetGraphOptions(graph_id);
  453. if (options == nullptr) {
  454. GELOGW("graph level options is null.");
  455. UpdateThreadContext(std::map<std::string, std::string>{});
  456. } else {
  457. UpdateThreadContext(*options);
  458. }
  459. }
  460. bool InnerSession::IsGraphNeedRebuild(uint32_t graph_id) {
  461. UpdateThreadContext(graph_id);
  462. return graph_manager_.IsGraphNeedRebuild(graph_id);
  463. }
  464. Status InnerSession::GetAllVariables(std::map<std::string, GeTensorDesc> &all_variables) {
  465. return VarManager::Instance(session_id_)->GetAllVariables(all_variables);
  466. }
  467. Status InnerSession::GenCheckPointGraph(const std::map<std::string, GeTensorDesc> &all_variables, Graph &graph) {
  468. return graph_manager_.GenCheckPointGraph(all_variables, graph);
  469. }
  470. Status InnerSession::SaveVariables(const Graph &graph, const std::vector<std::string> &var_names,
  471. const std::vector<Tensor> &outputs, std::vector<Tensor> &var_values) {
  472. return graph_manager_.SaveVariables(graph, var_names, outputs, var_values);
  473. }
  474. Status InnerSession::AddDumpProperties(const DumpProperties &dump_properties) {
  475. if (!is_dump_server_inited_) {
  476. if (dump_properties.IsDumpOpen() || dump_properties.IsOpDebugOpen()) {
  477. GE_IF_BOOL_EXEC(AdxDataDumpServerInit() != kDumpStatus,
  478. GELOGE(PARAM_INVALID, "[Init][AdxDataDumpServer] failed, session_id:%lu.", session_id_);
  479. return PARAM_INVALID)
  480. GELOGI("Init adx data dump server success");
  481. is_dump_server_inited_ = true;
  482. }
  483. }
  484. DumpManager::GetInstance().AddDumpProperties(session_id_, dump_properties);
  485. return SUCCESS;
  486. }
  487. Status InnerSession::RemoveDumpProperties() {
  488. DumpManager::GetInstance().RemoveDumpProperties(session_id_);
  489. if (is_dump_server_inited_ && DumpManager::GetInstance().GetDumpPropertiesMap().empty()) {
  490. GE_IF_BOOL_EXEC(AdxDataDumpServerUnInit() != kDumpStatus,
  491. GELOGE(PARAM_INVALID, "[UnInit][AdxDataDumpServer] failed, session_id:%lu.", session_id_);
  492. REPORT_INNER_ERROR("E19999", "RemoveDumpProperties failed because AdxDataDumpServerUnInit failed,"
  493. "session_id:%lu", session_id_);
  494. return PARAM_INVALID)
  495. GELOGI("UnInit adx data dump server success");
  496. is_dump_server_inited_ = false;
  497. }
  498. return SUCCESS;
  499. }
  500. void InnerSession::SetRtSocVersion() {
  501. const auto &global_options = GetMutableGlobalOptions();
  502. auto it = global_options.find(ge::SOC_VERSION);
  503. if (it != global_options.end()) {
  504. const char *soc_version = it->second.c_str();
  505. rtError_t rt_ret = rtSetSocVersion(soc_version);
  506. if (rt_ret != RT_ERROR_NONE) {
  507. GELOGW("Set soc version %s failed. ret:0x%X", soc_version, rt_ret);
  508. }
  509. GELOGI("Set soc version %s success.", soc_version);
  510. }
  511. }
  512. } // 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