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 19 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
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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534
  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 "analyzer/analyzer.h"
  32. #include "ge/ge_api_types.h"
  33. #include "ge_local_engine/engine/host_cpu_engine.h"
  34. #include "graph/common/ge_call_wrapper.h"
  35. #include "graph/ge_context.h"
  36. #include "graph/ge_global_options.h"
  37. #include "graph/load/model_manager/model_manager.h"
  38. #include "graph/manager/graph_mem_allocator.h"
  39. #include "graph/manager/host_mem_manager.h"
  40. #include "graph/manager/graph_var_manager.h"
  41. #include "omm/csa_interact.h"
  42. #include "runtime/kernel.h"
  43. #include "opskernel_manager/ops_kernel_builder_manager.h"
  44. using Json = nlohmann::json;
  45. namespace ge {
  46. namespace {
  47. const int kDecimal = 10;
  48. const int kSocVersionLen = 50;
  49. const int kDefaultDeviceIdForTrain = 0;
  50. const int kDefaultDeviceIdForInfer = -1;
  51. const char *const kGlobalOptionFpCeilingModeDefault = "2";
  52. const int32_t ACL_ERROR_RT_FEATURE_NOT_SUPPORT = 207000;
  53. } // namespace
  54. static std::shared_ptr<GELib> instancePtr_ = nullptr;
  55. // Initial each module of GE, if one failed, release all
  56. Status GELib::Initialize(const map<string, string> &options) {
  57. GELOGI("initial start");
  58. GEEVENT("[GEPERFTRACE] GE Init Start");
  59. // Multiple initializations are not allowed
  60. instancePtr_ = MakeShared<GELib>();
  61. if (instancePtr_ == nullptr) {
  62. GELOGE(GE_CLI_INIT_FAILED, "GeLib initialize failed, malloc shared_ptr failed.");
  63. return GE_CLI_INIT_FAILED;
  64. }
  65. map<string, string> new_options;
  66. Status ret = instancePtr_->SetRTSocVersion(options, new_options);
  67. if (ret != SUCCESS) {
  68. GELOGE(ret, "GeLib initial failed.");
  69. return ret;
  70. }
  71. ret = instancePtr_->SetAiCoreNum(new_options);
  72. if (ret != SUCCESS) {
  73. GELOGE(ret, "GeLib initial: SetAiCoreNum failed.");
  74. return ret;
  75. }
  76. instancePtr_->SetDefaultPrecisionMode(new_options);
  77. if (new_options.find("ge.fpCeilingMode") == new_options.end()) {
  78. new_options["ge.fpCeilingMode"] = kGlobalOptionFpCeilingModeDefault;
  79. }
  80. GetMutableGlobalOptions().insert(new_options.begin(), new_options.end());
  81. GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
  82. GE_TIMESTAMP_START(Init);
  83. ret = instancePtr_->InnerInitialize(new_options);
  84. if (ret != SUCCESS) {
  85. GELOGE(ret, "GeLib initial failed.");
  86. instancePtr_ = nullptr;
  87. return ret;
  88. }
  89. GE_TIMESTAMP_EVENT_END(Init, "GELib::Initialize");
  90. return SUCCESS;
  91. }
  92. Status GELib::InnerInitialize(const map<string, string> &options) {
  93. // Multiple initializations are not allowed
  94. if (init_flag_) {
  95. GELOGW("multi initializations");
  96. return SUCCESS;
  97. }
  98. GELOGI("GE System initial.");
  99. GE_TIMESTAMP_START(SystemInitialize);
  100. Status initSystemStatus = SystemInitialize(options);
  101. GE_TIMESTAMP_END(SystemInitialize, "InnerInitialize::SystemInitialize");
  102. if (initSystemStatus != SUCCESS) {
  103. GELOGE(initSystemStatus, "GE system initial failed.");
  104. RollbackInit();
  105. return initSystemStatus;
  106. }
  107. GELOGI("engineManager initial.");
  108. GE_TIMESTAMP_START(EngineInitialize);
  109. Status initEmStatus = engineManager_.Initialize(options);
  110. GE_TIMESTAMP_END(EngineInitialize, "InnerInitialize::EngineInitialize");
  111. if (initEmStatus != SUCCESS) {
  112. GELOGE(initEmStatus, "GE engine manager initial failed.");
  113. RollbackInit();
  114. return initEmStatus;
  115. }
  116. GELOGI("opsManager initial.");
  117. GE_TIMESTAMP_START(OpsManagerInitialize);
  118. Status initOpsStatus = opsManager_.Initialize(options);
  119. GE_TIMESTAMP_END(OpsManagerInitialize, "InnerInitialize::OpsManagerInitialize");
  120. if (initOpsStatus != SUCCESS) {
  121. GELOGE(initOpsStatus, "GE ops manager initial failed.");
  122. RollbackInit();
  123. return initOpsStatus;
  124. }
  125. GELOGI("opsBuilderManager initial.");
  126. GE_TIMESTAMP_START(OpsKernelBuilderManagerInitialize);
  127. Status initOpsBuilderStatus = OpsKernelBuilderManager::Instance().Initialize(options);
  128. GE_TIMESTAMP_END(OpsKernelBuilderManagerInitialize, "InnerInitialize::OpsKernelBuilderManager");
  129. if (initOpsBuilderStatus != SUCCESS) {
  130. GELOGE(initOpsBuilderStatus, "GE ops builder manager initial failed.");
  131. RollbackInit();
  132. return initOpsBuilderStatus;
  133. }
  134. GELOGI("sessionManager initial.");
  135. GE_TIMESTAMP_START(SessionManagerInitialize);
  136. Status initSmStatus = sessionManager_.Initialize(options);
  137. GE_TIMESTAMP_END(SessionManagerInitialize, "InnerInitialize::SessionManagerInitialize");
  138. if (initSmStatus != SUCCESS) {
  139. GELOGE(initSmStatus, "GE session manager initial failed.");
  140. RollbackInit();
  141. return initSmStatus;
  142. }
  143. GELOGI("Start to initialize HostCpuEngine");
  144. GE_TIMESTAMP_START(HostCpuEngineInitialize);
  145. Status initHostCpuEngineStatus = HostCpuEngine::GetInstance().Initialize();
  146. GE_TIMESTAMP_END(HostCpuEngineInitialize, "InnerInitialize::HostCpuEngineInitialize");
  147. if (initHostCpuEngineStatus != SUCCESS) {
  148. GELOGE(initHostCpuEngineStatus, "Failed to initialize HostCpuEngine");
  149. RollbackInit();
  150. return initHostCpuEngineStatus;
  151. }
  152. GELOGI("Start to init Analyzer!");
  153. Status init_analyzer_status = ge::Analyzer::GetInstance()->Initialize();
  154. if (init_analyzer_status != SUCCESS) {
  155. GELOGE(init_analyzer_status, "Failed to initialize HostCpuEngine");
  156. RollbackInit();
  157. return init_analyzer_status;
  158. }
  159. init_flag_ = true;
  160. return SUCCESS;
  161. }
  162. Status GELib::SystemInitialize(const map<string, string> &options) {
  163. Status status = FAILED;
  164. auto iter = options.find(OPTION_GRAPH_RUN_MODE);
  165. if (iter != options.end()) {
  166. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  167. is_train_mode_ = true;
  168. }
  169. }
  170. InitOptions(options);
  171. // In train and infer, profiling is always needed.
  172. InitProfiling(this->options_);
  173. auto model_manager = ModelManager::GetInstance();
  174. GE_CHECK_NOTNULL(model_manager);
  175. GE_IF_BOOL_EXEC(model_manager->EnableExceptionDump(options) != SUCCESS,
  176. GELOGE(FAILED, "Enable exception dump failed");
  177. return FAILED);
  178. // 1.`is_train_mode_` means case: train
  179. // 2.`(!is_train_mode_) && (options_.device_id != kDefaultDeviceIdForInfer)` means case: online infer
  180. // these two case with logical device id
  181. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  182. status = InitSystemWithOptions(this->options_);
  183. } else {
  184. status = InitSystemWithoutOptions();
  185. }
  186. return status;
  187. }
  188. void GELib::InitProfiling(Options &options) {
  189. GELOGI("Init Profiling. session Id: %ld, device id:%d ", options.session_id, options.device_id);
  190. std::lock_guard<std::mutex> lock(status_mutex_);
  191. GetContext().Init();
  192. // Profiling init
  193. if (ProfilingManager::Instance().Init(options) != SUCCESS) {
  194. GELOGW("Profiling init failed.");
  195. }
  196. }
  197. void GELib::SetDefaultPrecisionMode(map<string, string> &new_options) {
  198. auto iter = new_options.find(PRECISION_MODE);
  199. if (iter != new_options.end()) {
  200. GELOGI("Find precision_mode in options, value is %s", iter->second.c_str());
  201. return;
  202. }
  203. iter = new_options.find(OPTION_GRAPH_RUN_MODE);
  204. if (iter != new_options.end()) {
  205. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  206. // only train mode need to be set allow_fp32_to_fp16.
  207. GELOGI("This is train mode, precision_mode need to be set allow_fp32_to_fp16");
  208. new_options.insert(std::make_pair(PRECISION_MODE, "allow_fp32_to_fp16"));
  209. return;
  210. }
  211. }
  212. GELOGI("This is not train mode, precision_mode need to be set force_fp16");
  213. new_options.insert(std::make_pair(PRECISION_MODE, "force_fp16"));
  214. return;
  215. }
  216. Status GELib::SetRTSocVersion(const map<string, string> &options, map<string, string> &new_options) {
  217. GELOGI("Start to set SOC_VERSION");
  218. new_options.insert(options.begin(), options.end());
  219. auto it = new_options.find(ge::SOC_VERSION);
  220. if (it != new_options.end()) {
  221. GE_CHK_RT_RET(rtSetSocVersion(it->second.c_str()));
  222. GELOGI("Succeeded in setting SOC_VERSION[%s] to runtime.", it->second.c_str());
  223. } else {
  224. GELOGI("SOC_VERSION is not exist in options");
  225. char version[kSocVersionLen] = {0};
  226. rtError_t rt_ret = rtGetSocVersion(version, kSocVersionLen);
  227. GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(rt_ret, "rtGetSocVersion failed"); return FAILED;)
  228. GELOGI("Succeeded in getting SOC_VERSION[%s] from runtime.", version);
  229. new_options.insert(std::make_pair(ge::SOC_VERSION, version));
  230. }
  231. return SUCCESS;
  232. }
  233. Status GELib::SetAiCoreNum(map<string, string> &options) {
  234. // Already set or get AICORE_NUM from options in offline mode
  235. if (options.find(AICORE_NUM) != options.end()) {
  236. return SUCCESS;
  237. }
  238. uint32_t aicore_num = 0;
  239. rtError_t ret = rtGetAiCoreCount(&aicore_num);
  240. if (ret == ACL_ERROR_RT_FEATURE_NOT_SUPPORT) { // offline without ATC Input of AiCoreNum
  241. return SUCCESS;
  242. } else if (ret == RT_ERROR_NONE) { // online-mode
  243. options.emplace(std::make_pair(AICORE_NUM, std::to_string(aicore_num)));
  244. return SUCCESS;
  245. }
  246. GELOGE(FAILED, "rtGetAiCoreCount failed.");
  247. return FAILED;
  248. }
  249. void GELib::InitOptions(const map<string, string> &options) {
  250. this->options_.session_id = 0;
  251. auto iter = options.find(OPTION_EXEC_SESSION_ID);
  252. if (iter != options.end()) {
  253. this->options_.session_id = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  254. }
  255. this->options_.device_id = is_train_mode_ ? kDefaultDeviceIdForTrain : kDefaultDeviceIdForInfer;
  256. iter = options.find(OPTION_EXEC_DEVICE_ID);
  257. if (iter != options.end()) {
  258. this->options_.device_id = static_cast<int32_t>(std::strtol(iter->second.c_str(), nullptr, kDecimal));
  259. }
  260. iter = options.find(OPTION_EXEC_JOB_ID);
  261. if (iter != options.end()) {
  262. this->options_.job_id = iter->second.c_str();
  263. }
  264. this->options_.isUseHcom = false;
  265. iter = options.find(OPTION_EXEC_IS_USEHCOM);
  266. if (iter != options.end()) {
  267. std::istringstream(iter->second) >> this->options_.isUseHcom;
  268. }
  269. this->options_.isUseHvd = false;
  270. iter = options.find(OPTION_EXEC_IS_USEHVD);
  271. if (iter != options.end()) {
  272. std::istringstream(iter->second) >> this->options_.isUseHvd;
  273. }
  274. this->options_.deployMode = false;
  275. iter = options.find(OPTION_EXEC_DEPLOY_MODE);
  276. if (iter != options.end()) {
  277. std::istringstream(iter->second) >> this->options_.deployMode;
  278. }
  279. iter = options.find(OPTION_EXEC_POD_NAME);
  280. if (iter != options.end()) {
  281. this->options_.podName = iter->second.c_str();
  282. }
  283. iter = options.find(OPTION_EXEC_PROFILING_MODE);
  284. if (iter != options.end()) {
  285. this->options_.profiling_mode = iter->second.c_str();
  286. }
  287. iter = options.find(OPTION_EXEC_PROFILING_OPTIONS);
  288. if (iter != options.end()) {
  289. this->options_.profiling_options = iter->second.c_str();
  290. }
  291. iter = options.find(OPTION_EXEC_RANK_ID);
  292. if (iter != options.end()) {
  293. this->options_.rankId = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  294. }
  295. iter = options.find(OPTION_EXEC_RANK_TABLE_FILE);
  296. if (iter != options.end()) {
  297. this->options_.rankTableFile = iter->second.c_str();
  298. }
  299. this->options_.enable_atomic = true;
  300. iter = options.find(OPTION_EXEC_ATOMIC_FLAG);
  301. GE_IF_BOOL_EXEC(iter != options.end(),
  302. this->options_.enable_atomic = std::strtol(iter->second.c_str(), nullptr, kDecimal));
  303. GELOGI("ge InnerInitialize, the enable_atomic_flag in options_ is %d", this->options_.enable_atomic);
  304. }
  305. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithOptions(Options &options) {
  306. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  307. GELOGI("%s init GELib. session Id:%ld, device id :%d ", mode.c_str(), options.session_id, options.device_id);
  308. GEEVENT("System init with options begin, job id %s", options.job_id.c_str());
  309. std::lock_guard<std::mutex> lock(status_mutex_);
  310. GE_IF_BOOL_EXEC(is_system_inited && !is_shutdown,
  311. GELOGW("System init with options is already inited and not shutdown.");
  312. return SUCCESS);
  313. std::vector<rtMemType_t> mem_type;
  314. mem_type.push_back(RT_MEMORY_HBM);
  315. mem_type.push_back(RT_MEMORY_P2P_DDR);
  316. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  317. if (initMmStatus != SUCCESS) {
  318. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  319. return initMmStatus;
  320. }
  321. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  322. // Update CSA file
  323. CsaInteract::GetInstance().Init(options.device_id, GetContext().TraceId());
  324. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_RUNNING, JOBSUBSTATE_ENV_INIT);
  325. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  326. // set device id
  327. GELOGI("set logical device id:%u", options.device_id);
  328. GetContext().SetCtxDeviceId(static_cast<uint32_t>(options.device_id));
  329. GE_CHK_RT_RET(rtSetDevice(options.device_id));
  330. // In the scenario that the automatic add fusion is set, but there is no cleanaddr operator,
  331. // maybe need to check it
  332. is_system_inited = true;
  333. is_shutdown = false;
  334. GELOGI("%s init GELib success.", mode.c_str());
  335. return SUCCESS;
  336. }
  337. Status GELib::SystemShutdownWithOptions(const Options &options) {
  338. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  339. GELOGI("%s finalize GELib begin.", mode.c_str());
  340. std::lock_guard<std::mutex> lock(status_mutex_);
  341. GE_IF_BOOL_EXEC(is_shutdown || !is_system_inited,
  342. GELOGW("System Shutdown with options is already is_shutdown or system does not inited. "
  343. "is_shutdown:%d is_omm_inited:%d",
  344. is_shutdown, is_system_inited);
  345. return SUCCESS);
  346. GE_CHK_RT(rtDeviceReset(options.device_id));
  347. // Update CSA file
  348. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_SUCCEED);
  349. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  350. is_system_inited = false;
  351. is_shutdown = true;
  352. GELOGI("%s finalize GELib success.", mode.c_str());
  353. return SUCCESS;
  354. }
  355. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithoutOptions() {
  356. GELOGI("Inference Init GELib begin.");
  357. std::vector<rtMemType_t> mem_type;
  358. mem_type.push_back(RT_MEMORY_HBM);
  359. mem_type.push_back(RT_MEMORY_P2P_DDR);
  360. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  361. if (initMmStatus != SUCCESS) {
  362. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  363. return initMmStatus;
  364. }
  365. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  366. static bool is_inited = false;
  367. if (is_inited) {
  368. GELOGW("System init without options is already inited, don't need to init again.");
  369. return SUCCESS;
  370. }
  371. is_inited = true;
  372. GELOGI("Inference init GELib success.");
  373. return SUCCESS;
  374. }
  375. string GELib::GetPath() { return PluginManager::GetPath(); }
  376. // Finalize all modules
  377. Status GELib::Finalize() {
  378. GELOGI("finalization start");
  379. // Finalization is not allowed before initialization
  380. if (!init_flag_) {
  381. GELOGW("not initialize");
  382. return SUCCESS;
  383. }
  384. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  385. GE_CHK_RT_RET(rtSetDevice(options_.device_id));
  386. }
  387. Status final_state = SUCCESS;
  388. Status mid_state;
  389. GELOGI("engineManager finalization.");
  390. mid_state = engineManager_.Finalize();
  391. if (mid_state != SUCCESS) {
  392. GELOGW("engineManager finalize failed");
  393. final_state = mid_state;
  394. }
  395. GELOGI("sessionManager finalization.");
  396. mid_state = sessionManager_.Finalize();
  397. if (mid_state != SUCCESS) {
  398. GELOGW("sessionManager finalize failed");
  399. final_state = mid_state;
  400. }
  401. GELOGI("opsBuilderManager finalization.");
  402. mid_state = OpsKernelBuilderManager::Instance().Finalize();
  403. if (mid_state != SUCCESS) {
  404. GELOGW("opsBuilderManager finalize failed");
  405. final_state = mid_state;
  406. }
  407. GELOGI("opsManager finalization.");
  408. mid_state = opsManager_.Finalize();
  409. if (mid_state != SUCCESS) {
  410. GELOGW("opsManager finalize failed");
  411. final_state = mid_state;
  412. }
  413. GELOGI("VarManagerPool finalization.");
  414. VarManagerPool::Instance().Destory();
  415. GELOGI("MemManager finalization.");
  416. MemManager::Instance().Finalize();
  417. GELOGI("HostMemManager finalization.");
  418. HostMemManager::Instance().Finalize();
  419. GELOGI("HostCpuEngine finalization.");
  420. HostCpuEngine::GetInstance().Finalize();
  421. GELOGI("Analyzer finalization");
  422. Analyzer::GetInstance()->Finalize();
  423. // Shut down profiling
  424. ShutDownProfiling();
  425. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  426. GELOGI("System ShutDown.");
  427. mid_state = SystemShutdownWithOptions(this->options_);
  428. if (mid_state != SUCCESS) {
  429. GELOGW("System shutdown with options failed");
  430. final_state = mid_state;
  431. }
  432. }
  433. is_train_mode_ = false;
  434. GetMutableGlobalOptions().erase(ENABLE_SINGLE_STREAM);
  435. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  436. GE_CHK_RT_RET(rtDeviceReset(options_.device_id));
  437. }
  438. instancePtr_ = nullptr;
  439. init_flag_ = false;
  440. if (final_state != SUCCESS) {
  441. GELOGE(FAILED, "finalization failed.");
  442. return final_state;
  443. }
  444. GELOGI("finalization success.");
  445. return SUCCESS;
  446. }
  447. void GELib::ShutDownProfiling() {
  448. std::lock_guard<std::mutex> lock(status_mutex_);
  449. if (ProfilingManager::Instance().ProfilingOn()) {
  450. ProfilingManager::Instance().StopProfiling();
  451. ProfilingManager::Instance().PluginUnInit();
  452. }
  453. }
  454. // Get Singleton Instance
  455. std::shared_ptr<GELib> GELib::GetInstance() { return instancePtr_; }
  456. void GELib::RollbackInit() {
  457. if (engineManager_.init_flag_) {
  458. (void)engineManager_.Finalize();
  459. }
  460. if (opsManager_.init_flag_) {
  461. (void)opsManager_.Finalize();
  462. }
  463. if (sessionManager_.init_flag_) {
  464. (void)sessionManager_.Finalize();
  465. }
  466. MemManager::Instance().Finalize();
  467. HostMemManager::Instance().Finalize();
  468. VarManagerPool::Instance().Destory();
  469. }
  470. } // 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