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
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
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549
  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 "external/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_manager.h"
  40. #include "graph/manager/host_mem_manager.h"
  41. #include "graph/manager/graph_var_manager.h"
  42. #include "runtime/kernel.h"
  43. #include "opskernel_manager/ops_kernel_builder_manager.h"
  44. #include "external/runtime/rt_error_codes.h"
  45. using Json = nlohmann::json;
  46. namespace ge {
  47. namespace {
  48. const int kDecimal = 10;
  49. const int kSocVersionLen = 50;
  50. const int kDefaultDeviceIdForTrain = 0;
  51. const int kDefaultDeviceIdForInfer = -1;
  52. const char *const kGlobalOptionFpCeilingModeDefault = "2";
  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, "[Create][GELib]GeLib initialize failed, malloc shared_ptr failed.");
  63. REPORT_INNER_ERROR("E19999", "GELib Init failed for new GeLib failed.");
  64. return GE_CLI_INIT_FAILED;
  65. }
  66. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kSystemInit);
  67. map<string, string> new_options;
  68. Status ret = instancePtr_->SetRTSocVersion(options, new_options);
  69. if (ret != SUCCESS) {
  70. GELOGE(ret, "[Set][RTSocVersion]GeLib initial: SetRTSocVersion failed.");
  71. REPORT_CALL_ERROR("E19999", "SetRTSocVersion failed.");
  72. return ret;
  73. }
  74. ret = instancePtr_->SetAiCoreNum(new_options);
  75. if (ret != SUCCESS) {
  76. GELOGE(ret, "[Set][AiCoreNum]GeLib initial: SetAiCoreNum failed.");
  77. REPORT_CALL_ERROR("E19999", "SetAiCoreNum failed.");
  78. return ret;
  79. }
  80. instancePtr_->SetDefaultPrecisionMode(new_options);
  81. if (new_options.find("ge.fpCeilingMode") == new_options.end()) {
  82. new_options["ge.fpCeilingMode"] = kGlobalOptionFpCeilingModeDefault;
  83. }
  84. GetMutableGlobalOptions().insert(new_options.begin(), new_options.end());
  85. GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
  86. GE_TIMESTAMP_START(Init);
  87. ret = instancePtr_->InnerInitialize(new_options);
  88. if (ret != SUCCESS) {
  89. GELOGE(ret, "[Init][GeLib]GeLib initial failed.");
  90. REPORT_CALL_ERROR("E19999", "GELib::InnerInitialize failed.");
  91. instancePtr_ = nullptr;
  92. return ret;
  93. }
  94. GE_TIMESTAMP_EVENT_END(Init, "GELib::Initialize");
  95. return SUCCESS;
  96. }
  97. Status GELib::InnerInitialize(const map<string, string> &options) {
  98. // Multiple initializations are not allowed
  99. if (init_flag_) {
  100. GELOGW("multi initializations");
  101. return SUCCESS;
  102. }
  103. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kSystemInit);
  104. GELOGI("GE System initial.");
  105. GE_TIMESTAMP_START(SystemInitialize);
  106. Status initSystemStatus = SystemInitialize(options);
  107. GE_TIMESTAMP_END(SystemInitialize, "InnerInitialize::SystemInitialize");
  108. if (initSystemStatus != SUCCESS) {
  109. GELOGE(initSystemStatus, "[Init][GESystem]GE system initial failed.");
  110. RollbackInit();
  111. return initSystemStatus;
  112. }
  113. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kEngineInit);
  114. GELOGI("engineManager initial.");
  115. GE_TIMESTAMP_START(EngineInitialize);
  116. Status initEmStatus = engineManager_.Initialize(options);
  117. GE_TIMESTAMP_END(EngineInitialize, "InnerInitialize::EngineInitialize");
  118. if (initEmStatus != SUCCESS) {
  119. GELOGE(initEmStatus, "[Init][EngineManager]GE engine manager initial failed.");
  120. REPORT_CALL_ERROR("E19999", "EngineManager initialize failed.");
  121. RollbackInit();
  122. return initEmStatus;
  123. }
  124. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kOpsKernelInit);
  125. GELOGI("opsManager initial.");
  126. GE_TIMESTAMP_START(OpsManagerInitialize);
  127. Status initOpsStatus = opsManager_.Initialize(options);
  128. GE_TIMESTAMP_END(OpsManagerInitialize, "InnerInitialize::OpsManagerInitialize");
  129. if (initOpsStatus != SUCCESS) {
  130. GELOGE(initOpsStatus, "[Init][OpsManager]GE ops manager initial failed.");
  131. REPORT_CALL_ERROR("E19999", "OpsManager initialize failed.");
  132. RollbackInit();
  133. return initOpsStatus;
  134. }
  135. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kOpsKernelBuilderInit);
  136. GELOGI("opsBuilderManager initial.");
  137. GE_TIMESTAMP_START(OpsKernelBuilderManagerInitialize);
  138. Status initOpsBuilderStatus = OpsKernelBuilderManager::Instance().Initialize(options);
  139. GE_TIMESTAMP_END(OpsKernelBuilderManagerInitialize, "InnerInitialize::OpsKernelBuilderManager");
  140. if (initOpsBuilderStatus != SUCCESS) {
  141. GELOGE(initOpsBuilderStatus, "[Init][OpsKernelBuilderManager]GE ops builder manager initial failed.");
  142. REPORT_CALL_ERROR("E19999", "OpsBuilderManager initialize failed.");
  143. RollbackInit();
  144. return initOpsBuilderStatus;
  145. }
  146. GELOGI("Start to initialize HostCpuEngine");
  147. GE_TIMESTAMP_START(HostCpuEngineInitialize);
  148. Status initHostCpuEngineStatus = HostCpuEngine::GetInstance().Initialize();
  149. GE_TIMESTAMP_END(HostCpuEngineInitialize, "InnerInitialize::HostCpuEngineInitialize");
  150. if (initHostCpuEngineStatus != SUCCESS) {
  151. GELOGE(initHostCpuEngineStatus, "[Init][HostCpuEngine]Failed to initialize HostCpuEngine.");
  152. REPORT_CALL_ERROR("E19999", "HostCpuEngine initialize failed.");
  153. RollbackInit();
  154. return initHostCpuEngineStatus;
  155. }
  156. GELOGI("Start to init Analyzer!");
  157. Status init_analyzer_status = ge::Analyzer::GetInstance()->Initialize();
  158. if (init_analyzer_status != SUCCESS) {
  159. GELOGE(init_analyzer_status, "[Init][Analyzer]Failed to initialize Analyzer.");
  160. REPORT_CALL_ERROR("E19999", "ge::Analyzer initialize failed.");
  161. RollbackInit();
  162. return init_analyzer_status;
  163. }
  164. init_flag_ = true;
  165. return SUCCESS;
  166. }
  167. Status GELib::SystemInitialize(const map<string, string> &options) {
  168. Status status = FAILED;
  169. auto iter = options.find(OPTION_GRAPH_RUN_MODE);
  170. if (iter != options.end()) {
  171. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  172. is_train_mode_ = true;
  173. }
  174. }
  175. InitOptions(options);
  176. // In train and infer, profiling is always needed.
  177. InitProfiling(this->options_);
  178. auto model_manager = ModelManager::GetInstance();
  179. GE_CHECK_NOTNULL(model_manager);
  180. GE_IF_BOOL_EXEC(model_manager->EnableExceptionDump(options) != SUCCESS,
  181. REPORT_CALL_ERROR("E19999", "ModelManager EnableExceptionDump failed.");
  182. GELOGE(FAILED, "[Enable][ExceptionDump] failed.");
  183. return FAILED);
  184. // 1.`is_train_mode_` means case: train
  185. // 2.`(!is_train_mode_) && (options_.device_id != kDefaultDeviceIdForInfer)` means case: online infer
  186. // these two case with logical device id
  187. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  188. status = InitSystemWithOptions(this->options_);
  189. } else {
  190. status = InitSystemWithoutOptions();
  191. }
  192. return status;
  193. }
  194. void GELib::InitProfiling(Options &options) {
  195. GELOGI("Init Profiling. session Id: %ld, device id:%d ", options.session_id, options.device_id);
  196. std::lock_guard<std::mutex> lock(status_mutex_);
  197. GetContext().Init();
  198. // Profiling init
  199. if (ProfilingManager::Instance().Init(options) != SUCCESS) {
  200. GELOGW("Profiling init failed.");
  201. }
  202. }
  203. void GELib::SetDefaultPrecisionMode(map<string, string> &new_options) {
  204. auto iter = new_options.find(PRECISION_MODE);
  205. if (iter != new_options.end()) {
  206. GELOGI("Find precision_mode in options, value is %s", iter->second.c_str());
  207. return;
  208. }
  209. iter = new_options.find(OPTION_GRAPH_RUN_MODE);
  210. if (iter != new_options.end()) {
  211. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  212. // only train mode need to be set allow_fp32_to_fp16.
  213. GELOGI("This is train mode, precision_mode need to be set allow_fp32_to_fp16");
  214. new_options.insert(std::make_pair(PRECISION_MODE, "allow_fp32_to_fp16"));
  215. return;
  216. }
  217. }
  218. GELOGI("This is not train mode, precision_mode need to be set force_fp16");
  219. new_options.insert(std::make_pair(PRECISION_MODE, "force_fp16"));
  220. return;
  221. }
  222. Status GELib::SetRTSocVersion(const map<string, string> &options, map<string, string> &new_options) {
  223. GELOGI("Start to set SOC_VERSION");
  224. new_options.insert(options.begin(), options.end());
  225. auto it = new_options.find(ge::SOC_VERSION);
  226. if (it != new_options.end()) {
  227. GE_CHK_RT_RET(rtSetSocVersion(it->second.c_str()));
  228. GELOGI("Succeeded in setting SOC_VERSION[%s] to runtime.", it->second.c_str());
  229. } else {
  230. GELOGI("SOC_VERSION is not exist in options");
  231. char version[kSocVersionLen] = {0};
  232. rtError_t rt_ret = rtGetSocVersion(version, kSocVersionLen);
  233. GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE,
  234. REPORT_CALL_ERROR("E19999", "rtGetSocVersion failed.");
  235. GELOGE(rt_ret, "[Get][SocVersion]rtGetSocVersion failed");
  236. return FAILED;)
  237. GELOGI("Succeeded in getting SOC_VERSION[%s] from runtime.", version);
  238. new_options.insert(std::make_pair(ge::SOC_VERSION, version));
  239. }
  240. return SUCCESS;
  241. }
  242. Status GELib::SetAiCoreNum(map<string, string> &options) {
  243. // Already set or get AICORE_NUM from options in offline mode
  244. if (options.find(AICORE_NUM) != options.end()) {
  245. return SUCCESS;
  246. }
  247. uint32_t aicore_num = 0;
  248. rtError_t ret = rtGetAiCoreCount(&aicore_num);
  249. if (ret == ACL_ERROR_RT_FEATURE_NOT_SUPPORT) { // offline without ATC Input of AiCoreNum
  250. return SUCCESS;
  251. } else if (ret == RT_ERROR_NONE) { // online-mode
  252. options.emplace(std::make_pair(AICORE_NUM, std::to_string(aicore_num)));
  253. return SUCCESS;
  254. }
  255. GELOGE(FAILED, "[Get][AiCoreCount]rtGetAiCoreCount failed.");
  256. REPORT_CALL_ERROR("E19999", "rtGetAiCoreCount failed.");
  257. return FAILED;
  258. }
  259. void GELib::InitOptions(const map<string, string> &options) {
  260. this->options_.session_id = 0;
  261. auto iter = options.find(OPTION_EXEC_SESSION_ID);
  262. if (iter != options.end()) {
  263. this->options_.session_id = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  264. }
  265. this->options_.device_id = is_train_mode_ ? kDefaultDeviceIdForTrain : kDefaultDeviceIdForInfer;
  266. iter = options.find(OPTION_EXEC_DEVICE_ID);
  267. if (iter != options.end()) {
  268. this->options_.device_id = static_cast<int32_t>(std::strtol(iter->second.c_str(), nullptr, kDecimal));
  269. }
  270. iter = options.find(OPTION_EXEC_JOB_ID);
  271. if (iter != options.end()) {
  272. this->options_.job_id = iter->second.c_str();
  273. }
  274. this->options_.isUseHcom = false;
  275. iter = options.find(OPTION_EXEC_IS_USEHCOM);
  276. if (iter != options.end()) {
  277. std::istringstream(iter->second) >> this->options_.isUseHcom;
  278. }
  279. this->options_.isUseHvd = false;
  280. iter = options.find(OPTION_EXEC_IS_USEHVD);
  281. if (iter != options.end()) {
  282. std::istringstream(iter->second) >> this->options_.isUseHvd;
  283. }
  284. this->options_.deployMode = false;
  285. iter = options.find(OPTION_EXEC_DEPLOY_MODE);
  286. if (iter != options.end()) {
  287. std::istringstream(iter->second) >> this->options_.deployMode;
  288. }
  289. iter = options.find(OPTION_EXEC_POD_NAME);
  290. if (iter != options.end()) {
  291. this->options_.podName = iter->second.c_str();
  292. }
  293. iter = options.find(OPTION_EXEC_PROFILING_MODE);
  294. if (iter != options.end()) {
  295. this->options_.profiling_mode = iter->second.c_str();
  296. }
  297. iter = options.find(OPTION_EXEC_PROFILING_OPTIONS);
  298. if (iter != options.end()) {
  299. this->options_.profiling_options = iter->second.c_str();
  300. }
  301. iter = options.find(OPTION_EXEC_RANK_ID);
  302. if (iter != options.end()) {
  303. this->options_.rankId = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  304. }
  305. iter = options.find(OPTION_EXEC_RANK_TABLE_FILE);
  306. if (iter != options.end()) {
  307. this->options_.rankTableFile = iter->second.c_str();
  308. }
  309. this->options_.enable_atomic = true;
  310. iter = options.find(OPTION_EXEC_ATOMIC_FLAG);
  311. GE_IF_BOOL_EXEC(iter != options.end(),
  312. this->options_.enable_atomic = std::strtol(iter->second.c_str(), nullptr, kDecimal));
  313. GELOGI("ge InnerInitialize, the enable_atomic_flag in options_ is %d", this->options_.enable_atomic);
  314. }
  315. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithOptions(Options &options) {
  316. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  317. GELOGI("%s init GELib. session Id:%ld, device id :%d ", mode.c_str(), options.session_id, options.device_id);
  318. GEEVENT("System init with options begin, job id %s", options.job_id.c_str());
  319. std::lock_guard<std::mutex> lock(status_mutex_);
  320. GE_IF_BOOL_EXEC(is_system_inited && !is_shutdown,
  321. GELOGW("System init with options is already inited and not shutdown.");
  322. return SUCCESS);
  323. std::vector<rtMemType_t> mem_type;
  324. mem_type.push_back(RT_MEMORY_HBM);
  325. mem_type.push_back(RT_MEMORY_P2P_DDR);
  326. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  327. if (initMmStatus != SUCCESS) {
  328. GELOGE(initMmStatus, "[Init][MemManager] MemoryAllocatorManager initialize failed.");
  329. REPORT_CALL_ERROR("E19999", "MemManager initialize failed.");
  330. return initMmStatus;
  331. }
  332. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  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. is_system_inited = false;
  355. is_shutdown = true;
  356. GELOGI("%s finalize GELib success.", mode.c_str());
  357. return SUCCESS;
  358. }
  359. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithoutOptions() {
  360. GELOGI("Inference Init GELib begin.");
  361. std::vector<rtMemType_t> mem_type;
  362. mem_type.push_back(RT_MEMORY_HBM);
  363. mem_type.push_back(RT_MEMORY_P2P_DDR);
  364. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  365. if (initMmStatus != SUCCESS) {
  366. GELOGE(initMmStatus, "[Init][MemoryManager] initialize failed.");
  367. REPORT_CALL_ERROR("E19999", "MemManager initialize failed.");
  368. return initMmStatus;
  369. }
  370. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  371. static bool is_inited = false;
  372. if (is_inited) {
  373. GELOGW("System init without options is already inited, don't need to init again.");
  374. return SUCCESS;
  375. }
  376. is_inited = true;
  377. GELOGI("Inference init GELib success.");
  378. return SUCCESS;
  379. }
  380. string GELib::GetPath() { return PluginManager::GetPath(); }
  381. // Finalize all modules
  382. Status GELib::Finalize() {
  383. ErrorManager::GetInstance().SetStage(error_message::kFinalize, error_message::kFinalize);
  384. GELOGI("finalization start");
  385. // Finalization is not allowed before initialization
  386. if (!init_flag_) {
  387. GELOGW("not initialize");
  388. return SUCCESS;
  389. }
  390. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  391. GE_CHK_RT_RET(rtSetDevice(options_.device_id));
  392. }
  393. Status final_state = SUCCESS;
  394. Status mid_state;
  395. GELOGI("engineManager finalization.");
  396. mid_state = engineManager_.Finalize();
  397. if (mid_state != SUCCESS) {
  398. GELOGW("engineManager 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, "[Check][State]finalization failed.");
  442. REPORT_INNER_ERROR("E19999", "GELib::Finalize failed.");
  443. return final_state;
  444. }
  445. GELOGI("finalization success.");
  446. return SUCCESS;
  447. }
  448. void GELib::ShutDownProfiling() {
  449. std::lock_guard<std::mutex> lock(status_mutex_);
  450. if (ProfilingManager::Instance().ProfilingOn()) {
  451. ProfilingManager::Instance().StopProfiling();
  452. ProfilingManager::Instance().PluginUnInit();
  453. }
  454. }
  455. // Get Singleton Instance
  456. std::shared_ptr<GELib> GELib::GetInstance() { return instancePtr_; }
  457. void GELib::RollbackInit() {
  458. if (engineManager_.init_flag_) {
  459. (void)engineManager_.Finalize();
  460. }
  461. if (opsManager_.init_flag_) {
  462. (void)opsManager_.Finalize();
  463. }
  464. MemManager::Instance().Finalize();
  465. HostMemManager::Instance().Finalize();
  466. VarManagerPool::Instance().Destory();
  467. }
  468. Status GEInit::Initialize(const map<string, string> &options) {
  469. Status ret = SUCCESS;
  470. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  471. if (instance_ptr == nullptr || !instance_ptr->InitFlag()) {
  472. ret = GELib::Initialize(options);
  473. }
  474. return ret;
  475. }
  476. Status GEInit::Finalize() {
  477. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  478. if (instance_ptr != nullptr) {
  479. return instance_ptr->Finalize();
  480. }
  481. return SUCCESS;
  482. }
  483. string GEInit::GetPath() {
  484. return GELib::GetPath();
  485. }
  486. } // 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