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.

gelib.cc 20 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
4 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
5 years ago
4 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
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
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
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563
  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 "init/gelib.h"
  17. #include <dlfcn.h>
  18. #include <cstdlib>
  19. #include <mutex>
  20. #include <set>
  21. #include <sstream>
  22. #include <string>
  23. #include <utility>
  24. #include "common/ge/ge_util.h"
  25. #include "common/ge/plugin_manager.h"
  26. #include "common/profiling/profiling_manager.h"
  27. #include "common/properties_manager.h"
  28. #include "framework/common/debug/ge_log.h"
  29. #include "framework/common/debug/log.h"
  30. #include "framework/common/util.h"
  31. #include "framework/omg/ge_init.h"
  32. #include "analyzer/analyzer.h"
  33. #include "ge/ge_api_types.h"
  34. #include "ge_local_engine/engine/host_cpu_engine.h"
  35. #include "graph/common/ge_call_wrapper.h"
  36. #include "graph/ge_context.h"
  37. #include "graph/ge_global_options.h"
  38. #include "graph/load/model_manager/model_manager.h"
  39. #include "graph/manager/graph_mem_allocator.h"
  40. #include "graph/manager/host_mem_manager.h"
  41. #include "graph/manager/graph_var_manager.h"
  42. #include "omm/csa_interact.h"
  43. #include "runtime/kernel.h"
  44. #include "opskernel_manager/ops_kernel_builder_manager.h"
  45. #include "external/runtime/rt_error_codes.h"
  46. using Json = nlohmann::json;
  47. namespace ge {
  48. namespace {
  49. const int kDecimal = 10;
  50. const int kSocVersionLen = 50;
  51. const int kDefaultDeviceIdForTrain = 0;
  52. const int kDefaultDeviceIdForInfer = -1;
  53. const char *const kGlobalOptionFpCeilingModeDefault = "2";
  54. } // namespace
  55. static std::shared_ptr<GELib> instancePtr_ = nullptr;
  56. // Initial each module of GE, if one failed, release all
  57. Status GELib::Initialize(const map<string, string> &options) {
  58. GELOGI("initial start");
  59. GEEVENT("[GEPERFTRACE] GE Init Start");
  60. // Multiple initializations are not allowed
  61. instancePtr_ = MakeShared<GELib>();
  62. if (instancePtr_ == nullptr) {
  63. GELOGE(GE_CLI_INIT_FAILED, "GeLib initialize failed, malloc shared_ptr failed.");
  64. return GE_CLI_INIT_FAILED;
  65. }
  66. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kSystemInit);
  67. map<string, string> new_options;
  68. Status ret = instancePtr_->SetRTSocVersion(options, new_options);
  69. if (ret != SUCCESS) {
  70. GELOGE(ret, "GeLib initial failed.");
  71. return ret;
  72. }
  73. ret = instancePtr_->SetAiCoreNum(new_options);
  74. if (ret != SUCCESS) {
  75. GELOGE(ret, "GeLib initial: SetAiCoreNum failed.");
  76. return ret;
  77. }
  78. instancePtr_->SetDefaultPrecisionMode(new_options);
  79. if (new_options.find("ge.fpCeilingMode") == new_options.end()) {
  80. new_options["ge.fpCeilingMode"] = kGlobalOptionFpCeilingModeDefault;
  81. }
  82. GetMutableGlobalOptions().insert(new_options.begin(), new_options.end());
  83. GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
  84. GE_TIMESTAMP_START(Init);
  85. ret = instancePtr_->InnerInitialize(new_options);
  86. if (ret != SUCCESS) {
  87. GELOGE(ret, "GeLib initial failed.");
  88. instancePtr_ = nullptr;
  89. return ret;
  90. }
  91. GE_TIMESTAMP_EVENT_END(Init, "GELib::Initialize");
  92. return SUCCESS;
  93. }
  94. Status GELib::InnerInitialize(const map<string, string> &options) {
  95. // Multiple initializations are not allowed
  96. if (init_flag_) {
  97. GELOGW("multi initializations");
  98. return SUCCESS;
  99. }
  100. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kSystemInit);
  101. GELOGI("GE System initial.");
  102. GE_TIMESTAMP_START(SystemInitialize);
  103. Status initSystemStatus = SystemInitialize(options);
  104. GE_TIMESTAMP_END(SystemInitialize, "InnerInitialize::SystemInitialize");
  105. if (initSystemStatus != SUCCESS) {
  106. GELOGE(initSystemStatus, "GE system initial failed.");
  107. RollbackInit();
  108. return initSystemStatus;
  109. }
  110. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kEngineInit);
  111. GELOGI("engineManager initial.");
  112. GE_TIMESTAMP_START(EngineInitialize);
  113. Status initEmStatus = engineManager_.Initialize(options);
  114. GE_TIMESTAMP_END(EngineInitialize, "InnerInitialize::EngineInitialize");
  115. if (initEmStatus != SUCCESS) {
  116. GELOGE(initEmStatus, "GE engine manager initial failed.");
  117. RollbackInit();
  118. return initEmStatus;
  119. }
  120. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kOpsKernelInit);
  121. GELOGI("opsManager initial.");
  122. GE_TIMESTAMP_START(OpsManagerInitialize);
  123. Status initOpsStatus = opsManager_.Initialize(options);
  124. GE_TIMESTAMP_END(OpsManagerInitialize, "InnerInitialize::OpsManagerInitialize");
  125. if (initOpsStatus != SUCCESS) {
  126. GELOGE(initOpsStatus, "GE ops manager initial failed.");
  127. RollbackInit();
  128. return initOpsStatus;
  129. }
  130. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kOpsKernelBuilderInit);
  131. GELOGI("opsBuilderManager initial.");
  132. GE_TIMESTAMP_START(OpsKernelBuilderManagerInitialize);
  133. Status initOpsBuilderStatus = OpsKernelBuilderManager::Instance().Initialize(options);
  134. GE_TIMESTAMP_END(OpsKernelBuilderManagerInitialize, "InnerInitialize::OpsKernelBuilderManager");
  135. if (initOpsBuilderStatus != SUCCESS) {
  136. GELOGE(initOpsBuilderStatus, "GE ops builder manager initial failed.");
  137. RollbackInit();
  138. return initOpsBuilderStatus;
  139. }
  140. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kOther);
  141. GELOGI("sessionManager initial.");
  142. GE_TIMESTAMP_START(SessionManagerInitialize);
  143. Status initSmStatus = sessionManager_.Initialize(options);
  144. GE_TIMESTAMP_END(SessionManagerInitialize, "InnerInitialize::SessionManagerInitialize");
  145. if (initSmStatus != SUCCESS) {
  146. GELOGE(initSmStatus, "GE session manager initial failed.");
  147. RollbackInit();
  148. return initSmStatus;
  149. }
  150. GELOGI("Start to initialize HostCpuEngine");
  151. GE_TIMESTAMP_START(HostCpuEngineInitialize);
  152. Status initHostCpuEngineStatus = HostCpuEngine::GetInstance().Initialize();
  153. GE_TIMESTAMP_END(HostCpuEngineInitialize, "InnerInitialize::HostCpuEngineInitialize");
  154. if (initHostCpuEngineStatus != SUCCESS) {
  155. GELOGE(initHostCpuEngineStatus, "Failed to initialize HostCpuEngine");
  156. RollbackInit();
  157. return initHostCpuEngineStatus;
  158. }
  159. GELOGI("Start to init Analyzer!");
  160. Status init_analyzer_status = ge::Analyzer::GetInstance()->Initialize();
  161. if (init_analyzer_status != SUCCESS) {
  162. GELOGE(init_analyzer_status, "Failed to initialize HostCpuEngine");
  163. RollbackInit();
  164. return init_analyzer_status;
  165. }
  166. init_flag_ = true;
  167. return SUCCESS;
  168. }
  169. Status GELib::SystemInitialize(const map<string, string> &options) {
  170. Status status = FAILED;
  171. auto iter = options.find(OPTION_GRAPH_RUN_MODE);
  172. if (iter != options.end()) {
  173. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  174. is_train_mode_ = true;
  175. }
  176. }
  177. InitOptions(options);
  178. // In train and infer, profiling is always needed.
  179. InitProfiling(this->options_);
  180. auto model_manager = ModelManager::GetInstance();
  181. GE_CHECK_NOTNULL(model_manager);
  182. GE_IF_BOOL_EXEC(model_manager->EnableExceptionDump(options) != SUCCESS,
  183. GELOGE(FAILED, "Enable exception dump failed");
  184. return FAILED);
  185. // 1.`is_train_mode_` means case: train
  186. // 2.`(!is_train_mode_) && (options_.device_id != kDefaultDeviceIdForInfer)` means case: online infer
  187. // these two case with logical device id
  188. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  189. status = InitSystemWithOptions(this->options_);
  190. } else {
  191. status = InitSystemWithoutOptions();
  192. }
  193. return status;
  194. }
  195. void GELib::InitProfiling(Options &options) {
  196. GELOGI("Init Profiling. session Id: %ld, device id:%d ", options.session_id, options.device_id);
  197. std::lock_guard<std::mutex> lock(status_mutex_);
  198. GetContext().Init();
  199. // Profiling init
  200. if (ProfilingManager::Instance().Init(options) != SUCCESS) {
  201. GELOGW("Profiling init failed.");
  202. }
  203. }
  204. void GELib::SetDefaultPrecisionMode(map<string, string> &new_options) {
  205. auto iter = new_options.find(PRECISION_MODE);
  206. if (iter != new_options.end()) {
  207. GELOGI("Find precision_mode in options, value is %s", iter->second.c_str());
  208. return;
  209. }
  210. iter = new_options.find(OPTION_GRAPH_RUN_MODE);
  211. if (iter != new_options.end()) {
  212. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  213. // only train mode need to be set allow_fp32_to_fp16.
  214. GELOGI("This is train mode, precision_mode need to be set allow_fp32_to_fp16");
  215. new_options.insert(std::make_pair(PRECISION_MODE, "allow_fp32_to_fp16"));
  216. return;
  217. }
  218. }
  219. GELOGI("This is not train mode, precision_mode need to be set force_fp16");
  220. new_options.insert(std::make_pair(PRECISION_MODE, "force_fp16"));
  221. return;
  222. }
  223. Status GELib::SetRTSocVersion(const map<string, string> &options, map<string, string> &new_options) {
  224. GELOGI("Start to set SOC_VERSION");
  225. new_options.insert(options.begin(), options.end());
  226. auto it = new_options.find(ge::SOC_VERSION);
  227. if (it != new_options.end()) {
  228. GE_CHK_RT_RET(rtSetSocVersion(it->second.c_str()));
  229. GELOGI("Succeeded in setting SOC_VERSION[%s] to runtime.", it->second.c_str());
  230. } else {
  231. GELOGI("SOC_VERSION is not exist in options");
  232. char version[kSocVersionLen] = {0};
  233. rtError_t rt_ret = rtGetSocVersion(version, kSocVersionLen);
  234. GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(rt_ret, "rtGetSocVersion failed"); return FAILED;)
  235. GELOGI("Succeeded in getting SOC_VERSION[%s] from runtime.", version);
  236. new_options.insert(std::make_pair(ge::SOC_VERSION, version));
  237. }
  238. return SUCCESS;
  239. }
  240. Status GELib::SetAiCoreNum(map<string, string> &options) {
  241. // Already set or get AICORE_NUM from options in offline mode
  242. if (options.find(AICORE_NUM) != options.end()) {
  243. return SUCCESS;
  244. }
  245. uint32_t aicore_num = 0;
  246. rtError_t ret = rtGetAiCoreCount(&aicore_num);
  247. if (ret == ACL_ERROR_RT_FEATURE_NOT_SUPPORT) { // offline without ATC Input of AiCoreNum
  248. return SUCCESS;
  249. } else if (ret == RT_ERROR_NONE) { // online-mode
  250. options.emplace(std::make_pair(AICORE_NUM, std::to_string(aicore_num)));
  251. return SUCCESS;
  252. }
  253. GELOGE(FAILED, "rtGetAiCoreCount failed.");
  254. return FAILED;
  255. }
  256. void GELib::InitOptions(const map<string, string> &options) {
  257. this->options_.session_id = 0;
  258. auto iter = options.find(OPTION_EXEC_SESSION_ID);
  259. if (iter != options.end()) {
  260. this->options_.session_id = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  261. }
  262. this->options_.device_id = is_train_mode_ ? kDefaultDeviceIdForTrain : kDefaultDeviceIdForInfer;
  263. iter = options.find(OPTION_EXEC_DEVICE_ID);
  264. if (iter != options.end()) {
  265. this->options_.device_id = static_cast<int32_t>(std::strtol(iter->second.c_str(), nullptr, kDecimal));
  266. }
  267. iter = options.find(OPTION_EXEC_JOB_ID);
  268. if (iter != options.end()) {
  269. this->options_.job_id = iter->second.c_str();
  270. }
  271. this->options_.isUseHcom = false;
  272. iter = options.find(OPTION_EXEC_IS_USEHCOM);
  273. if (iter != options.end()) {
  274. std::istringstream(iter->second) >> this->options_.isUseHcom;
  275. }
  276. this->options_.isUseHvd = false;
  277. iter = options.find(OPTION_EXEC_IS_USEHVD);
  278. if (iter != options.end()) {
  279. std::istringstream(iter->second) >> this->options_.isUseHvd;
  280. }
  281. this->options_.deployMode = false;
  282. iter = options.find(OPTION_EXEC_DEPLOY_MODE);
  283. if (iter != options.end()) {
  284. std::istringstream(iter->second) >> this->options_.deployMode;
  285. }
  286. iter = options.find(OPTION_EXEC_POD_NAME);
  287. if (iter != options.end()) {
  288. this->options_.podName = iter->second.c_str();
  289. }
  290. iter = options.find(OPTION_EXEC_PROFILING_MODE);
  291. if (iter != options.end()) {
  292. this->options_.profiling_mode = iter->second.c_str();
  293. }
  294. iter = options.find(OPTION_EXEC_PROFILING_OPTIONS);
  295. if (iter != options.end()) {
  296. this->options_.profiling_options = iter->second.c_str();
  297. }
  298. iter = options.find(OPTION_EXEC_RANK_ID);
  299. if (iter != options.end()) {
  300. this->options_.rankId = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  301. }
  302. iter = options.find(OPTION_EXEC_RANK_TABLE_FILE);
  303. if (iter != options.end()) {
  304. this->options_.rankTableFile = iter->second.c_str();
  305. }
  306. this->options_.enable_atomic = true;
  307. iter = options.find(OPTION_EXEC_ATOMIC_FLAG);
  308. GE_IF_BOOL_EXEC(iter != options.end(),
  309. this->options_.enable_atomic = std::strtol(iter->second.c_str(), nullptr, kDecimal));
  310. GELOGI("ge InnerInitialize, the enable_atomic_flag in options_ is %d", this->options_.enable_atomic);
  311. }
  312. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithOptions(Options &options) {
  313. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  314. GELOGI("%s init GELib. session Id:%ld, device id :%d ", mode.c_str(), options.session_id, options.device_id);
  315. GEEVENT("System init with options begin, job id %s", options.job_id.c_str());
  316. std::lock_guard<std::mutex> lock(status_mutex_);
  317. GE_IF_BOOL_EXEC(is_system_inited && !is_shutdown,
  318. GELOGW("System init with options is already inited and not shutdown.");
  319. return SUCCESS);
  320. std::vector<rtMemType_t> mem_type;
  321. mem_type.push_back(RT_MEMORY_HBM);
  322. mem_type.push_back(RT_MEMORY_P2P_DDR);
  323. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  324. if (initMmStatus != SUCCESS) {
  325. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  326. return initMmStatus;
  327. }
  328. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  329. // Update CSA file
  330. CsaInteract::GetInstance().Init(options.device_id, GetContext().TraceId());
  331. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_RUNNING, JOBSUBSTATE_ENV_INIT);
  332. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  333. // set device id
  334. GELOGI("set logical device id:%u", options.device_id);
  335. GetContext().SetCtxDeviceId(static_cast<uint32_t>(options.device_id));
  336. GE_CHK_RT_RET(rtSetDevice(options.device_id));
  337. // In the scenario that the automatic add fusion is set, but there is no cleanaddr operator,
  338. // maybe need to check it
  339. is_system_inited = true;
  340. is_shutdown = false;
  341. GELOGI("%s init GELib success.", mode.c_str());
  342. return SUCCESS;
  343. }
  344. Status GELib::SystemShutdownWithOptions(const Options &options) {
  345. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  346. GELOGI("%s finalize GELib begin.", mode.c_str());
  347. std::lock_guard<std::mutex> lock(status_mutex_);
  348. GE_IF_BOOL_EXEC(is_shutdown || !is_system_inited,
  349. GELOGW("System Shutdown with options is already is_shutdown or system does not inited. "
  350. "is_shutdown:%d is_omm_inited:%d",
  351. is_shutdown, is_system_inited);
  352. return SUCCESS);
  353. GE_CHK_RT(rtDeviceReset(options.device_id));
  354. // Update CSA file
  355. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_SUCCEED);
  356. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  357. is_system_inited = false;
  358. is_shutdown = true;
  359. GELOGI("%s finalize GELib success.", mode.c_str());
  360. return SUCCESS;
  361. }
  362. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithoutOptions() {
  363. GELOGI("Inference Init GELib begin.");
  364. std::vector<rtMemType_t> mem_type;
  365. mem_type.push_back(RT_MEMORY_HBM);
  366. mem_type.push_back(RT_MEMORY_P2P_DDR);
  367. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  368. if (initMmStatus != SUCCESS) {
  369. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  370. return initMmStatus;
  371. }
  372. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  373. static bool is_inited = false;
  374. if (is_inited) {
  375. GELOGW("System init without options is already inited, don't need to init again.");
  376. return SUCCESS;
  377. }
  378. is_inited = true;
  379. GELOGI("Inference init GELib success.");
  380. return SUCCESS;
  381. }
  382. string GELib::GetPath() { return PluginManager::GetPath(); }
  383. // Finalize all modules
  384. Status GELib::Finalize() {
  385. ErrorManager::GetInstance().SetStage(ErrorMessage::kFinalize, ErrorMessage::kFinalize);
  386. GELOGI("finalization start");
  387. // Finalization is not allowed before initialization
  388. if (!init_flag_) {
  389. GELOGW("not initialize");
  390. return SUCCESS;
  391. }
  392. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  393. GE_CHK_RT_RET(rtSetDevice(options_.device_id));
  394. }
  395. Status final_state = SUCCESS;
  396. Status mid_state;
  397. GELOGI("engineManager finalization.");
  398. mid_state = engineManager_.Finalize();
  399. if (mid_state != SUCCESS) {
  400. GELOGW("engineManager finalize failed");
  401. final_state = mid_state;
  402. }
  403. GELOGI("sessionManager finalization.");
  404. mid_state = sessionManager_.Finalize();
  405. if (mid_state != SUCCESS) {
  406. GELOGW("sessionManager finalize failed");
  407. final_state = mid_state;
  408. }
  409. GELOGI("opsBuilderManager finalization.");
  410. mid_state = OpsKernelBuilderManager::Instance().Finalize();
  411. if (mid_state != SUCCESS) {
  412. GELOGW("opsBuilderManager finalize failed");
  413. final_state = mid_state;
  414. }
  415. GELOGI("opsManager finalization.");
  416. mid_state = opsManager_.Finalize();
  417. if (mid_state != SUCCESS) {
  418. GELOGW("opsManager finalize failed");
  419. final_state = mid_state;
  420. }
  421. GELOGI("VarManagerPool finalization.");
  422. VarManagerPool::Instance().Destory();
  423. GELOGI("MemManager finalization.");
  424. MemManager::Instance().Finalize();
  425. GELOGI("HostMemManager finalization.");
  426. HostMemManager::Instance().Finalize();
  427. GELOGI("HostCpuEngine finalization.");
  428. HostCpuEngine::GetInstance().Finalize();
  429. GELOGI("Analyzer finalization");
  430. Analyzer::GetInstance()->Finalize();
  431. // Shut down profiling
  432. ShutDownProfiling();
  433. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  434. GELOGI("System ShutDown.");
  435. mid_state = SystemShutdownWithOptions(this->options_);
  436. if (mid_state != SUCCESS) {
  437. GELOGW("System shutdown with options failed");
  438. final_state = mid_state;
  439. }
  440. }
  441. is_train_mode_ = false;
  442. GetMutableGlobalOptions().erase(ENABLE_SINGLE_STREAM);
  443. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  444. GE_CHK_RT_RET(rtDeviceReset(options_.device_id));
  445. }
  446. instancePtr_ = nullptr;
  447. init_flag_ = false;
  448. if (final_state != SUCCESS) {
  449. GELOGE(FAILED, "finalization failed.");
  450. return final_state;
  451. }
  452. GELOGI("finalization success.");
  453. return SUCCESS;
  454. }
  455. void GELib::ShutDownProfiling() {
  456. std::lock_guard<std::mutex> lock(status_mutex_);
  457. if (ProfilingManager::Instance().ProfilingOn()) {
  458. ProfilingManager::Instance().StopProfiling();
  459. ProfilingManager::Instance().PluginUnInit();
  460. }
  461. }
  462. // Get Singleton Instance
  463. std::shared_ptr<GELib> GELib::GetInstance() { return instancePtr_; }
  464. void GELib::RollbackInit() {
  465. if (engineManager_.init_flag_) {
  466. (void)engineManager_.Finalize();
  467. }
  468. if (opsManager_.init_flag_) {
  469. (void)opsManager_.Finalize();
  470. }
  471. if (sessionManager_.init_flag_) {
  472. (void)sessionManager_.Finalize();
  473. }
  474. MemManager::Instance().Finalize();
  475. HostMemManager::Instance().Finalize();
  476. VarManagerPool::Instance().Destory();
  477. }
  478. Status GEInit::Initialize(const map<string, string> &options) {
  479. Status ret = SUCCESS;
  480. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  481. if (instance_ptr == nullptr || !instance_ptr->InitFlag()) {
  482. ret = GELib::Initialize(options);
  483. }
  484. return ret;
  485. }
  486. Status GEInit::Finalize() {
  487. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  488. if (instance_ptr != nullptr) {
  489. return instance_ptr->Finalize();
  490. }
  491. return SUCCESS;
  492. }
  493. string GEInit::GetPath() {
  494. return GELib::GetPath();
  495. }
  496. } // 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