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
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
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519
  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 <cstdlib>
  18. #include <mutex>
  19. #include <set>
  20. #include <sstream>
  21. #include <string>
  22. #include <utility>
  23. #include "common/ge/ge_util.h"
  24. #include "common/ge/plugin_manager.h"
  25. #include "common/profiling/profiling_manager.h"
  26. #include "common/properties_manager.h"
  27. #include "framework/common/debug/ge_log.h"
  28. #include "framework/common/debug/log.h"
  29. #include "framework/common/util.h"
  30. #include "framework/omg/ge_init.h"
  31. #include "analyzer/analyzer.h"
  32. #include "external/ge/ge_api_types.h"
  33. #include "ge_local_engine/engine/host_cpu_engine.h"
  34. #include "common/ge_call_wrapper.h"
  35. #include "graph/ge_context.h"
  36. #include "graph/ge_global_options.h"
  37. #include "graph/manager/graph_mem_manager.h"
  38. #include "graph/manager/host_mem_manager.h"
  39. #include "graph/manager/graph_var_manager.h"
  40. #include "runtime/kernel.h"
  41. #include "opskernel_manager/ops_kernel_builder_manager.h"
  42. #include "external/runtime/rt_error_codes.h"
  43. #include "common/profiling/profiling_init.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. } // namespace
  53. static std::shared_ptr<GELib> instancePtr_ = nullptr;
  54. // Initial each module of GE, if one failed, release all
  55. Status GELib::Initialize(const map<string, string> &options) {
  56. GELOGI("initial start");
  57. GEEVENT("[GEPERFTRACE] GE Init Start");
  58. // Multiple initializations are not allowed
  59. instancePtr_ = MakeShared<GELib>();
  60. if (instancePtr_ == nullptr) {
  61. GELOGE(GE_CLI_INIT_FAILED, "[Create][GELib]GeLib initialize failed, malloc shared_ptr failed.");
  62. REPORT_INNER_ERROR("E19999", "GELib Init failed for new GeLib failed.");
  63. return GE_CLI_INIT_FAILED;
  64. }
  65. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kSystemInit);
  66. map<string, string> new_options;
  67. Status ret = instancePtr_->SetRTSocVersion(options, new_options);
  68. if (ret != SUCCESS) {
  69. GELOGE(ret, "[Set][RTSocVersion]GeLib initial: SetRTSocVersion failed.");
  70. REPORT_CALL_ERROR("E19999", "SetRTSocVersion failed.");
  71. return ret;
  72. }
  73. ret = instancePtr_->SetAiCoreNum(new_options);
  74. if (ret != SUCCESS) {
  75. GELOGE(ret, "[Set][AiCoreNum]GeLib initial: SetAiCoreNum failed.");
  76. REPORT_CALL_ERROR("E19999", "SetAiCoreNum failed.");
  77. return ret;
  78. }
  79. instancePtr_->SetDefaultPrecisionMode(new_options);
  80. if (new_options.find("ge.fpCeilingMode") == new_options.end()) {
  81. new_options["ge.fpCeilingMode"] = kGlobalOptionFpCeilingModeDefault;
  82. }
  83. GetMutableGlobalOptions().insert(new_options.begin(), new_options.end());
  84. GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
  85. GE_TIMESTAMP_START(Init);
  86. ret = instancePtr_->InnerInitialize(new_options);
  87. if (ret != SUCCESS) {
  88. GELOGE(ret, "[Init][GeLib]GeLib initial failed.");
  89. REPORT_CALL_ERROR("E19999", "GELib::InnerInitialize failed.");
  90. instancePtr_ = nullptr;
  91. return ret;
  92. }
  93. GE_TIMESTAMP_EVENT_END(Init, "GELib::Initialize");
  94. return SUCCESS;
  95. }
  96. Status GELib::InnerInitialize(const map<string, string> &options) {
  97. // Multiple initializations are not allowed
  98. if (init_flag_) {
  99. GELOGW("multi initializations");
  100. return SUCCESS;
  101. }
  102. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kSystemInit);
  103. GELOGI("GE System initial.");
  104. GE_TIMESTAMP_START(SystemInitialize);
  105. Status initSystemStatus = SystemInitialize(options);
  106. GE_TIMESTAMP_END(SystemInitialize, "InnerInitialize::SystemInitialize");
  107. if (initSystemStatus != SUCCESS) {
  108. GELOGE(initSystemStatus, "[Init][GESystem]GE system initial failed.");
  109. RollbackInit();
  110. return initSystemStatus;
  111. }
  112. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kEngineInit);
  113. GELOGI("engineManager initial.");
  114. GE_TIMESTAMP_START(EngineInitialize);
  115. Status initEmStatus = engineManager_.Initialize(options);
  116. GE_TIMESTAMP_END(EngineInitialize, "InnerInitialize::EngineInitialize");
  117. if (initEmStatus != SUCCESS) {
  118. GELOGE(initEmStatus, "[Init][EngineManager]GE engine manager initial failed.");
  119. REPORT_CALL_ERROR("E19999", "EngineManager initialize failed.");
  120. RollbackInit();
  121. return initEmStatus;
  122. }
  123. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kOpsKernelInit);
  124. GELOGI("opsManager initial.");
  125. GE_TIMESTAMP_START(OpsManagerInitialize);
  126. Status initOpsStatus = opsManager_.Initialize(options);
  127. GE_TIMESTAMP_END(OpsManagerInitialize, "InnerInitialize::OpsManagerInitialize");
  128. if (initOpsStatus != SUCCESS) {
  129. GELOGE(initOpsStatus, "[Init][OpsManager]GE ops manager initial failed.");
  130. REPORT_CALL_ERROR("E19999", "OpsManager initialize failed.");
  131. RollbackInit();
  132. return initOpsStatus;
  133. }
  134. ErrorManager::GetInstance().SetStage(error_message::kInitialize, error_message::kOpsKernelBuilderInit);
  135. GELOGI("opsBuilderManager initial.");
  136. GE_TIMESTAMP_START(OpsKernelBuilderManagerInitialize);
  137. Status initOpsBuilderStatus = OpsKernelBuilderManager::Instance().Initialize(options);
  138. GE_TIMESTAMP_END(OpsKernelBuilderManagerInitialize, "InnerInitialize::OpsKernelBuilderManager");
  139. if (initOpsBuilderStatus != SUCCESS) {
  140. GELOGE(initOpsBuilderStatus, "[Init][OpsKernelBuilderManager]GE ops builder manager initial failed.");
  141. REPORT_CALL_ERROR("E19999", "OpsBuilderManager initialize failed.");
  142. RollbackInit();
  143. return initOpsBuilderStatus;
  144. }
  145. GELOGI("Start to initialize HostCpuEngine");
  146. GE_TIMESTAMP_START(HostCpuEngineInitialize);
  147. Status initHostCpuEngineStatus = HostCpuEngine::GetInstance().Initialize();
  148. GE_TIMESTAMP_END(HostCpuEngineInitialize, "InnerInitialize::HostCpuEngineInitialize");
  149. if (initHostCpuEngineStatus != SUCCESS) {
  150. GELOGE(initHostCpuEngineStatus, "[Init][HostCpuEngine]Failed to initialize HostCpuEngine.");
  151. REPORT_CALL_ERROR("E19999", "HostCpuEngine initialize failed.");
  152. RollbackInit();
  153. return initHostCpuEngineStatus;
  154. }
  155. GELOGI("Start to init Analyzer!");
  156. Status init_analyzer_status = ge::Analyzer::GetInstance()->Initialize();
  157. if (init_analyzer_status != SUCCESS) {
  158. GELOGE(init_analyzer_status, "[Init][Analyzer]Failed to initialize Analyzer.");
  159. REPORT_CALL_ERROR("E19999", "ge::Analyzer initialize failed.");
  160. RollbackInit();
  161. return init_analyzer_status;
  162. }
  163. init_flag_ = true;
  164. return SUCCESS;
  165. }
  166. Status GELib::SystemInitialize(const map<string, string> &options) {
  167. Status status = FAILED;
  168. auto iter = options.find(OPTION_GRAPH_RUN_MODE);
  169. if (iter != options.end()) {
  170. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  171. is_train_mode_ = true;
  172. }
  173. }
  174. InitOptions(options);
  175. // In train and infer, profiling is always needed.
  176. // 1.`is_train_mode_` means case: train
  177. // 2.`(!is_train_mode_) && (options_.device_id != kDefaultDeviceIdForInfer)` means case: online infer
  178. // these two case with logical device id
  179. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  180. status = InitSystemWithOptions(this->options_);
  181. } else {
  182. status = InitSystemWithoutOptions();
  183. }
  184. return status;
  185. }
  186. void GELib::SetDefaultPrecisionMode(map<string, string> &new_options) {
  187. auto iter = new_options.find(PRECISION_MODE);
  188. if (iter != new_options.end()) {
  189. GELOGI("Find precision_mode in options, value is %s", iter->second.c_str());
  190. return;
  191. }
  192. iter = new_options.find(OPTION_GRAPH_RUN_MODE);
  193. if (iter != new_options.end()) {
  194. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  195. // only train mode need to be set allow_fp32_to_fp16.
  196. GELOGI("This is train mode, precision_mode need to be set allow_fp32_to_fp16");
  197. new_options.insert(std::make_pair(PRECISION_MODE, "allow_fp32_to_fp16"));
  198. return;
  199. }
  200. }
  201. GELOGI("This is not train mode, precision_mode need to be set force_fp16");
  202. new_options.insert(std::make_pair(PRECISION_MODE, "force_fp16"));
  203. return;
  204. }
  205. Status GELib::SetRTSocVersion(const map<string, string> &options, map<string, string> &new_options) {
  206. GELOGI("Start to set SOC_VERSION");
  207. new_options.insert(options.begin(), options.end());
  208. auto it = new_options.find(ge::SOC_VERSION);
  209. if (it != new_options.end()) {
  210. GE_CHK_RT_RET(rtSetSocVersion(it->second.c_str()));
  211. GELOGI("Succeeded in setting SOC_VERSION[%s] to runtime.", it->second.c_str());
  212. } else {
  213. GELOGI("SOC_VERSION is not exist in options");
  214. char version[kSocVersionLen] = {0};
  215. rtError_t rt_ret = rtGetSocVersion(version, kSocVersionLen);
  216. GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE,
  217. REPORT_CALL_ERROR("E19999", "rtGetSocVersion failed.");
  218. GELOGE(rt_ret, "[Get][SocVersion]rtGetSocVersion failed");
  219. return FAILED;)
  220. GELOGI("Succeeded in getting SOC_VERSION[%s] from runtime.", version);
  221. new_options.insert(std::make_pair(ge::SOC_VERSION, version));
  222. }
  223. return SUCCESS;
  224. }
  225. Status GELib::SetAiCoreNum(map<string, string> &options) {
  226. // Already set or get AICORE_NUM from options in offline mode
  227. if (options.find(AICORE_NUM) != options.end()) {
  228. return SUCCESS;
  229. }
  230. uint32_t aicore_num = 0;
  231. rtError_t ret = rtGetAiCoreCount(&aicore_num);
  232. if (ret == ACL_ERROR_RT_FEATURE_NOT_SUPPORT) { // offline without ATC Input of AiCoreNum
  233. return SUCCESS;
  234. } else if (ret == RT_ERROR_NONE) { // online-mode
  235. options.emplace(std::make_pair(AICORE_NUM, std::to_string(aicore_num)));
  236. return SUCCESS;
  237. }
  238. GELOGE(FAILED, "[Get][AiCoreCount]rtGetAiCoreCount failed.");
  239. REPORT_CALL_ERROR("E19999", "rtGetAiCoreCount failed.");
  240. return FAILED;
  241. }
  242. void GELib::InitOptions(const map<string, string> &options) {
  243. this->options_.session_id = 0;
  244. auto iter = options.find(OPTION_EXEC_SESSION_ID);
  245. if (iter != options.end()) {
  246. this->options_.session_id = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  247. }
  248. this->options_.device_id = is_train_mode_ ? kDefaultDeviceIdForTrain : kDefaultDeviceIdForInfer;
  249. iter = options.find(OPTION_EXEC_DEVICE_ID);
  250. if (iter != options.end()) {
  251. this->options_.device_id = static_cast<int32_t>(std::strtol(iter->second.c_str(), nullptr, kDecimal));
  252. }
  253. iter = options.find(OPTION_EXEC_JOB_ID);
  254. if (iter != options.end()) {
  255. this->options_.job_id = iter->second.c_str();
  256. }
  257. this->options_.isUseHcom = false;
  258. iter = options.find(OPTION_EXEC_IS_USEHCOM);
  259. if (iter != options.end()) {
  260. std::istringstream(iter->second) >> this->options_.isUseHcom;
  261. }
  262. this->options_.isUseHvd = false;
  263. iter = options.find(OPTION_EXEC_IS_USEHVD);
  264. if (iter != options.end()) {
  265. std::istringstream(iter->second) >> this->options_.isUseHvd;
  266. }
  267. this->options_.deployMode = false;
  268. iter = options.find(OPTION_EXEC_DEPLOY_MODE);
  269. if (iter != options.end()) {
  270. std::istringstream(iter->second) >> this->options_.deployMode;
  271. }
  272. iter = options.find(OPTION_EXEC_POD_NAME);
  273. if (iter != options.end()) {
  274. this->options_.podName = iter->second.c_str();
  275. }
  276. iter = options.find(OPTION_EXEC_PROFILING_MODE);
  277. if (iter != options.end()) {
  278. this->options_.profiling_mode = iter->second.c_str();
  279. }
  280. iter = options.find(OPTION_EXEC_PROFILING_OPTIONS);
  281. if (iter != options.end()) {
  282. this->options_.profiling_options = iter->second.c_str();
  283. }
  284. iter = options.find(OPTION_EXEC_RANK_ID);
  285. if (iter != options.end()) {
  286. this->options_.rankId = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  287. }
  288. iter = options.find(OPTION_EXEC_RANK_TABLE_FILE);
  289. if (iter != options.end()) {
  290. this->options_.rankTableFile = iter->second.c_str();
  291. }
  292. this->options_.enable_atomic = true;
  293. iter = options.find(OPTION_EXEC_ATOMIC_FLAG);
  294. GE_IF_BOOL_EXEC(iter != options.end(),
  295. this->options_.enable_atomic = std::strtol(iter->second.c_str(), nullptr, kDecimal));
  296. GELOGI("ge InnerInitialize, the enable_atomic_flag in options_ is %d", this->options_.enable_atomic);
  297. }
  298. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithOptions(Options &options) {
  299. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  300. GELOGI("%s init GELib. session Id:%ld, device id :%d ", mode.c_str(), options.session_id, options.device_id);
  301. GEEVENT("System init with options begin, job id %s", options.job_id.c_str());
  302. std::lock_guard<std::mutex> lock(status_mutex_);
  303. GE_IF_BOOL_EXEC(is_system_inited && !is_shutdown,
  304. GELOGW("System init with options is already inited and not shutdown.");
  305. return SUCCESS);
  306. std::vector<rtMemType_t> mem_type;
  307. mem_type.push_back(RT_MEMORY_HBM);
  308. mem_type.push_back(RT_MEMORY_P2P_DDR);
  309. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  310. if (initMmStatus != SUCCESS) {
  311. GELOGE(initMmStatus, "[Init][MemManager] MemoryAllocatorManager initialize failed.");
  312. REPORT_CALL_ERROR("E19999", "MemManager initialize failed.");
  313. return initMmStatus;
  314. }
  315. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  316. // set device id
  317. GELOGI("set logical device id:%u", options.device_id);
  318. GetContext().SetCtxDeviceId(static_cast<uint32_t>(options.device_id));
  319. GE_CHK_RT_RET(rtSetDevice(options.device_id));
  320. // In the scenario that the automatic add fusion is set, but there is no cleanaddr operator,
  321. // maybe need to check it
  322. is_system_inited = true;
  323. is_shutdown = false;
  324. GELOGI("%s init GELib success.", mode.c_str());
  325. return SUCCESS;
  326. }
  327. Status GELib::SystemShutdownWithOptions(const Options &options) {
  328. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  329. GELOGI("%s finalize GELib begin.", mode.c_str());
  330. std::lock_guard<std::mutex> lock(status_mutex_);
  331. GE_IF_BOOL_EXEC(is_shutdown || !is_system_inited,
  332. GELOGW("System Shutdown with options is already is_shutdown or system does not inited. "
  333. "is_shutdown:%d is_omm_inited:%d",
  334. is_shutdown, is_system_inited);
  335. return SUCCESS);
  336. GE_CHK_RT(rtDeviceReset(options.device_id));
  337. is_system_inited = false;
  338. is_shutdown = true;
  339. GELOGI("%s finalize GELib success.", mode.c_str());
  340. return SUCCESS;
  341. }
  342. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithoutOptions() {
  343. GELOGI("Inference Init GELib begin.");
  344. std::vector<rtMemType_t> mem_type;
  345. mem_type.push_back(RT_MEMORY_HBM);
  346. mem_type.push_back(RT_MEMORY_P2P_DDR);
  347. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  348. if (initMmStatus != SUCCESS) {
  349. GELOGE(initMmStatus, "[Init][MemoryManager] initialize failed.");
  350. REPORT_CALL_ERROR("E19999", "MemManager initialize failed.");
  351. return initMmStatus;
  352. }
  353. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  354. static bool is_inited = false;
  355. if (is_inited) {
  356. GELOGW("System init without options is already inited, don't need to init again.");
  357. return SUCCESS;
  358. }
  359. is_inited = true;
  360. GELOGI("Inference init GELib success.");
  361. return SUCCESS;
  362. }
  363. string GELib::GetPath() { return PluginManager::GetPath(); }
  364. // Finalize all modules
  365. Status GELib::Finalize() {
  366. ErrorManager::GetInstance().SetStage(error_message::kFinalize, error_message::kFinalize);
  367. GELOGI("finalization start");
  368. // Finalization is not allowed before initialization
  369. if (!init_flag_) {
  370. GELOGW("not initialize");
  371. return SUCCESS;
  372. }
  373. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  374. GE_CHK_RT_RET(rtSetDevice(options_.device_id));
  375. }
  376. Status final_state = SUCCESS;
  377. Status mid_state;
  378. GELOGI("engineManager finalization.");
  379. mid_state = engineManager_.Finalize();
  380. if (mid_state != SUCCESS) {
  381. GELOGW("engineManager finalize failed");
  382. final_state = mid_state;
  383. }
  384. GELOGI("opsBuilderManager finalization.");
  385. mid_state = OpsKernelBuilderManager::Instance().Finalize();
  386. if (mid_state != SUCCESS) {
  387. GELOGW("opsBuilderManager finalize failed");
  388. final_state = mid_state;
  389. }
  390. GELOGI("opsManager finalization.");
  391. mid_state = opsManager_.Finalize();
  392. if (mid_state != SUCCESS) {
  393. GELOGW("opsManager finalize failed");
  394. final_state = mid_state;
  395. }
  396. GELOGI("VarManagerPool finalization.");
  397. VarManagerPool::Instance().Destory();
  398. GELOGI("MemManager finalization.");
  399. MemManager::Instance().Finalize();
  400. GELOGI("HostMemManager finalization.");
  401. HostMemManager::Instance().Finalize();
  402. GELOGI("HostCpuEngine finalization.");
  403. HostCpuEngine::GetInstance().Finalize();
  404. GELOGI("Analyzer finalization");
  405. Analyzer::GetInstance()->Finalize();
  406. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  407. GELOGI("System ShutDown.");
  408. mid_state = SystemShutdownWithOptions(this->options_);
  409. if (mid_state != SUCCESS) {
  410. GELOGW("System shutdown with options failed");
  411. final_state = mid_state;
  412. }
  413. }
  414. is_train_mode_ = false;
  415. GetMutableGlobalOptions().erase(ENABLE_SINGLE_STREAM);
  416. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  417. GE_CHK_RT_RET(rtDeviceReset(options_.device_id));
  418. }
  419. instancePtr_ = nullptr;
  420. init_flag_ = false;
  421. if (final_state != SUCCESS) {
  422. GELOGE(FAILED, "[Check][State]finalization failed.");
  423. REPORT_INNER_ERROR("E19999", "GELib::Finalize failed.");
  424. return final_state;
  425. }
  426. GELOGI("finalization success.");
  427. return SUCCESS;
  428. }
  429. // Get Singleton Instance
  430. std::shared_ptr<GELib> GELib::GetInstance() { return instancePtr_; }
  431. void GELib::RollbackInit() {
  432. if (engineManager_.init_flag_) {
  433. (void)engineManager_.Finalize();
  434. }
  435. if (opsManager_.init_flag_) {
  436. (void)opsManager_.Finalize();
  437. }
  438. MemManager::Instance().Finalize();
  439. HostMemManager::Instance().Finalize();
  440. VarManagerPool::Instance().Destory();
  441. }
  442. Status GEInit::Initialize(const map<string, string> &options) {
  443. Status ret = SUCCESS;
  444. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  445. if (instance_ptr == nullptr || !instance_ptr->InitFlag()) {
  446. ret = GELib::Initialize(options);
  447. }
  448. return ret;
  449. }
  450. Status GEInit::Finalize() {
  451. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  452. if (instance_ptr != nullptr) {
  453. return instance_ptr->Finalize();
  454. }
  455. return SUCCESS;
  456. }
  457. string GEInit::GetPath() {
  458. return GELib::GetPath();
  459. }
  460. } // 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