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
5 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552
  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. map<string, string> new_options;
  67. Status ret = instancePtr_->SetRTSocVersion(options, new_options);
  68. if (ret != SUCCESS) {
  69. GELOGE(ret, "GeLib initial failed.");
  70. return ret;
  71. }
  72. ret = instancePtr_->SetAiCoreNum(new_options);
  73. if (ret != SUCCESS) {
  74. GELOGE(ret, "GeLib initial: SetAiCoreNum failed.");
  75. return ret;
  76. }
  77. instancePtr_->SetDefaultPrecisionMode(new_options);
  78. if (new_options.find("ge.fpCeilingMode") == new_options.end()) {
  79. new_options["ge.fpCeilingMode"] = kGlobalOptionFpCeilingModeDefault;
  80. }
  81. GetMutableGlobalOptions().insert(new_options.begin(), new_options.end());
  82. GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
  83. GE_TIMESTAMP_START(Init);
  84. ret = instancePtr_->InnerInitialize(new_options);
  85. if (ret != SUCCESS) {
  86. GELOGE(ret, "GeLib initial failed.");
  87. instancePtr_ = nullptr;
  88. return ret;
  89. }
  90. GE_TIMESTAMP_EVENT_END(Init, "GELib::Initialize");
  91. return SUCCESS;
  92. }
  93. Status GELib::InnerInitialize(const map<string, string> &options) {
  94. // Multiple initializations are not allowed
  95. if (init_flag_) {
  96. GELOGW("multi initializations");
  97. return SUCCESS;
  98. }
  99. GELOGI("GE System initial.");
  100. GE_TIMESTAMP_START(SystemInitialize);
  101. Status initSystemStatus = SystemInitialize(options);
  102. GE_TIMESTAMP_END(SystemInitialize, "InnerInitialize::SystemInitialize");
  103. if (initSystemStatus != SUCCESS) {
  104. GELOGE(initSystemStatus, "GE system initial failed.");
  105. RollbackInit();
  106. return initSystemStatus;
  107. }
  108. GELOGI("engineManager initial.");
  109. GE_TIMESTAMP_START(EngineInitialize);
  110. Status initEmStatus = engineManager_.Initialize(options);
  111. GE_TIMESTAMP_END(EngineInitialize, "InnerInitialize::EngineInitialize");
  112. if (initEmStatus != SUCCESS) {
  113. GELOGE(initEmStatus, "GE engine manager initial failed.");
  114. RollbackInit();
  115. return initEmStatus;
  116. }
  117. GELOGI("opsManager initial.");
  118. GE_TIMESTAMP_START(OpsManagerInitialize);
  119. Status initOpsStatus = opsManager_.Initialize(options);
  120. GE_TIMESTAMP_END(OpsManagerInitialize, "InnerInitialize::OpsManagerInitialize");
  121. if (initOpsStatus != SUCCESS) {
  122. GELOGE(initOpsStatus, "GE ops manager initial failed.");
  123. RollbackInit();
  124. return initOpsStatus;
  125. }
  126. GELOGI("opsBuilderManager initial.");
  127. GE_TIMESTAMP_START(OpsKernelBuilderManagerInitialize);
  128. Status initOpsBuilderStatus = OpsKernelBuilderManager::Instance().Initialize(options);
  129. GE_TIMESTAMP_END(OpsKernelBuilderManagerInitialize, "InnerInitialize::OpsKernelBuilderManager");
  130. if (initOpsBuilderStatus != SUCCESS) {
  131. GELOGE(initOpsBuilderStatus, "GE ops builder manager initial failed.");
  132. RollbackInit();
  133. return initOpsBuilderStatus;
  134. }
  135. GELOGI("sessionManager initial.");
  136. GE_TIMESTAMP_START(SessionManagerInitialize);
  137. Status initSmStatus = sessionManager_.Initialize(options);
  138. GE_TIMESTAMP_END(SessionManagerInitialize, "InnerInitialize::SessionManagerInitialize");
  139. if (initSmStatus != SUCCESS) {
  140. GELOGE(initSmStatus, "GE session manager initial failed.");
  141. RollbackInit();
  142. return initSmStatus;
  143. }
  144. GELOGI("Start to initialize HostCpuEngine");
  145. GE_TIMESTAMP_START(HostCpuEngineInitialize);
  146. Status initHostCpuEngineStatus = HostCpuEngine::GetInstance().Initialize();
  147. GE_TIMESTAMP_END(HostCpuEngineInitialize, "InnerInitialize::HostCpuEngineInitialize");
  148. if (initHostCpuEngineStatus != SUCCESS) {
  149. GELOGE(initHostCpuEngineStatus, "Failed to initialize HostCpuEngine");
  150. RollbackInit();
  151. return initHostCpuEngineStatus;
  152. }
  153. GELOGI("Start to init Analyzer!");
  154. Status init_analyzer_status = ge::Analyzer::GetInstance()->Initialize();
  155. if (init_analyzer_status != SUCCESS) {
  156. GELOGE(init_analyzer_status, "Failed to initialize HostCpuEngine");
  157. RollbackInit();
  158. return init_analyzer_status;
  159. }
  160. init_flag_ = true;
  161. return SUCCESS;
  162. }
  163. Status GELib::SystemInitialize(const map<string, string> &options) {
  164. Status status = FAILED;
  165. auto iter = options.find(OPTION_GRAPH_RUN_MODE);
  166. if (iter != options.end()) {
  167. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  168. is_train_mode_ = true;
  169. }
  170. }
  171. InitOptions(options);
  172. // In train and infer, profiling is always needed.
  173. InitProfiling(this->options_);
  174. auto model_manager = ModelManager::GetInstance();
  175. GE_CHECK_NOTNULL(model_manager);
  176. GE_IF_BOOL_EXEC(model_manager->EnableExceptionDump(options) != SUCCESS,
  177. GELOGE(FAILED, "Enable exception dump failed");
  178. return FAILED);
  179. // 1.`is_train_mode_` means case: train
  180. // 2.`(!is_train_mode_) && (options_.device_id != kDefaultDeviceIdForInfer)` means case: online infer
  181. // these two case with logical device id
  182. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  183. status = InitSystemWithOptions(this->options_);
  184. } else {
  185. status = InitSystemWithoutOptions();
  186. }
  187. return status;
  188. }
  189. void GELib::InitProfiling(Options &options) {
  190. GELOGI("Init Profiling. session Id: %ld, device id:%d ", options.session_id, options.device_id);
  191. std::lock_guard<std::mutex> lock(status_mutex_);
  192. GetContext().Init();
  193. // Profiling init
  194. if (ProfilingManager::Instance().Init(options) != SUCCESS) {
  195. GELOGW("Profiling init failed.");
  196. }
  197. }
  198. void GELib::SetDefaultPrecisionMode(map<string, string> &new_options) {
  199. auto iter = new_options.find(PRECISION_MODE);
  200. if (iter != new_options.end()) {
  201. GELOGI("Find precision_mode in options, value is %s", iter->second.c_str());
  202. return;
  203. }
  204. iter = new_options.find(OPTION_GRAPH_RUN_MODE);
  205. if (iter != new_options.end()) {
  206. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  207. // only train mode need to be set allow_fp32_to_fp16.
  208. GELOGI("This is train mode, precision_mode need to be set allow_fp32_to_fp16");
  209. new_options.insert(std::make_pair(PRECISION_MODE, "allow_fp32_to_fp16"));
  210. return;
  211. }
  212. }
  213. GELOGI("This is not train mode, precision_mode need to be set force_fp16");
  214. new_options.insert(std::make_pair(PRECISION_MODE, "force_fp16"));
  215. return;
  216. }
  217. Status GELib::SetRTSocVersion(const map<string, string> &options, map<string, string> &new_options) {
  218. GELOGI("Start to set SOC_VERSION");
  219. new_options.insert(options.begin(), options.end());
  220. auto it = new_options.find(ge::SOC_VERSION);
  221. if (it != new_options.end()) {
  222. GE_CHK_RT_RET(rtSetSocVersion(it->second.c_str()));
  223. GELOGI("Succeeded in setting SOC_VERSION[%s] to runtime.", it->second.c_str());
  224. } else {
  225. GELOGI("SOC_VERSION is not exist in options");
  226. char version[kSocVersionLen] = {0};
  227. rtError_t rt_ret = rtGetSocVersion(version, kSocVersionLen);
  228. GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(rt_ret, "rtGetSocVersion failed"); return FAILED;)
  229. GELOGI("Succeeded in getting SOC_VERSION[%s] from runtime.", version);
  230. new_options.insert(std::make_pair(ge::SOC_VERSION, version));
  231. }
  232. return SUCCESS;
  233. }
  234. Status GELib::SetAiCoreNum(map<string, string> &options) {
  235. // Already set or get AICORE_NUM from options in offline mode
  236. if (options.find(AICORE_NUM) != options.end()) {
  237. return SUCCESS;
  238. }
  239. uint32_t aicore_num = 0;
  240. rtError_t ret = rtGetAiCoreCount(&aicore_num);
  241. if (ret == ACL_ERROR_RT_FEATURE_NOT_SUPPORT) { // offline without ATC Input of AiCoreNum
  242. return SUCCESS;
  243. } else if (ret == RT_ERROR_NONE) { // online-mode
  244. options.emplace(std::make_pair(AICORE_NUM, std::to_string(aicore_num)));
  245. return SUCCESS;
  246. }
  247. GELOGE(FAILED, "rtGetAiCoreCount failed.");
  248. return FAILED;
  249. }
  250. void GELib::InitOptions(const map<string, string> &options) {
  251. this->options_.session_id = 0;
  252. auto iter = options.find(OPTION_EXEC_SESSION_ID);
  253. if (iter != options.end()) {
  254. this->options_.session_id = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  255. }
  256. this->options_.device_id = is_train_mode_ ? kDefaultDeviceIdForTrain : kDefaultDeviceIdForInfer;
  257. iter = options.find(OPTION_EXEC_DEVICE_ID);
  258. if (iter != options.end()) {
  259. this->options_.device_id = static_cast<int32_t>(std::strtol(iter->second.c_str(), nullptr, kDecimal));
  260. }
  261. iter = options.find(OPTION_EXEC_JOB_ID);
  262. if (iter != options.end()) {
  263. this->options_.job_id = iter->second.c_str();
  264. }
  265. this->options_.isUseHcom = false;
  266. iter = options.find(OPTION_EXEC_IS_USEHCOM);
  267. if (iter != options.end()) {
  268. std::istringstream(iter->second) >> this->options_.isUseHcom;
  269. }
  270. this->options_.isUseHvd = false;
  271. iter = options.find(OPTION_EXEC_IS_USEHVD);
  272. if (iter != options.end()) {
  273. std::istringstream(iter->second) >> this->options_.isUseHvd;
  274. }
  275. this->options_.deployMode = false;
  276. iter = options.find(OPTION_EXEC_DEPLOY_MODE);
  277. if (iter != options.end()) {
  278. std::istringstream(iter->second) >> this->options_.deployMode;
  279. }
  280. iter = options.find(OPTION_EXEC_POD_NAME);
  281. if (iter != options.end()) {
  282. this->options_.podName = iter->second.c_str();
  283. }
  284. iter = options.find(OPTION_EXEC_PROFILING_MODE);
  285. if (iter != options.end()) {
  286. this->options_.profiling_mode = iter->second.c_str();
  287. }
  288. iter = options.find(OPTION_EXEC_PROFILING_OPTIONS);
  289. if (iter != options.end()) {
  290. this->options_.profiling_options = iter->second.c_str();
  291. }
  292. iter = options.find(OPTION_EXEC_RANK_ID);
  293. if (iter != options.end()) {
  294. this->options_.rankId = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  295. }
  296. iter = options.find(OPTION_EXEC_RANK_TABLE_FILE);
  297. if (iter != options.end()) {
  298. this->options_.rankTableFile = iter->second.c_str();
  299. }
  300. this->options_.enable_atomic = true;
  301. iter = options.find(OPTION_EXEC_ATOMIC_FLAG);
  302. GE_IF_BOOL_EXEC(iter != options.end(),
  303. this->options_.enable_atomic = std::strtol(iter->second.c_str(), nullptr, kDecimal));
  304. GELOGI("ge InnerInitialize, the enable_atomic_flag in options_ is %d", this->options_.enable_atomic);
  305. }
  306. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithOptions(Options &options) {
  307. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  308. GELOGI("%s init GELib. session Id:%ld, device id :%d ", mode.c_str(), options.session_id, options.device_id);
  309. GEEVENT("System init with options begin, job id %s", options.job_id.c_str());
  310. std::lock_guard<std::mutex> lock(status_mutex_);
  311. GE_IF_BOOL_EXEC(is_system_inited && !is_shutdown,
  312. GELOGW("System init with options is already inited and not shutdown.");
  313. return SUCCESS);
  314. std::vector<rtMemType_t> mem_type;
  315. mem_type.push_back(RT_MEMORY_HBM);
  316. mem_type.push_back(RT_MEMORY_P2P_DDR);
  317. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  318. if (initMmStatus != SUCCESS) {
  319. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  320. return initMmStatus;
  321. }
  322. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  323. // Update CSA file
  324. CsaInteract::GetInstance().Init(options.device_id, GetContext().TraceId());
  325. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_RUNNING, JOBSUBSTATE_ENV_INIT);
  326. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  327. // set device id
  328. GELOGI("set logical device id:%u", options.device_id);
  329. GetContext().SetCtxDeviceId(static_cast<uint32_t>(options.device_id));
  330. GE_CHK_RT_RET(rtSetDevice(options.device_id));
  331. // In the scenario that the automatic add fusion is set, but there is no cleanaddr operator,
  332. // maybe need to check it
  333. is_system_inited = true;
  334. is_shutdown = false;
  335. GELOGI("%s init GELib success.", mode.c_str());
  336. return SUCCESS;
  337. }
  338. Status GELib::SystemShutdownWithOptions(const Options &options) {
  339. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  340. GELOGI("%s finalize GELib begin.", mode.c_str());
  341. std::lock_guard<std::mutex> lock(status_mutex_);
  342. GE_IF_BOOL_EXEC(is_shutdown || !is_system_inited,
  343. GELOGW("System Shutdown with options is already is_shutdown or system does not inited. "
  344. "is_shutdown:%d is_omm_inited:%d",
  345. is_shutdown, is_system_inited);
  346. return SUCCESS);
  347. GE_CHK_RT(rtDeviceReset(options.device_id));
  348. // Update CSA file
  349. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_SUCCEED);
  350. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  351. is_system_inited = false;
  352. is_shutdown = true;
  353. GELOGI("%s finalize GELib success.", mode.c_str());
  354. return SUCCESS;
  355. }
  356. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithoutOptions() {
  357. GELOGI("Inference Init GELib begin.");
  358. std::vector<rtMemType_t> mem_type;
  359. mem_type.push_back(RT_MEMORY_HBM);
  360. mem_type.push_back(RT_MEMORY_P2P_DDR);
  361. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  362. if (initMmStatus != SUCCESS) {
  363. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  364. return initMmStatus;
  365. }
  366. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  367. static bool is_inited = false;
  368. if (is_inited) {
  369. GELOGW("System init without options is already inited, don't need to init again.");
  370. return SUCCESS;
  371. }
  372. is_inited = true;
  373. GELOGI("Inference init GELib success.");
  374. return SUCCESS;
  375. }
  376. string GELib::GetPath() { return PluginManager::GetPath(); }
  377. // Finalize all modules
  378. Status GELib::Finalize() {
  379. GELOGI("finalization start");
  380. // Finalization is not allowed before initialization
  381. if (!init_flag_) {
  382. GELOGW("not initialize");
  383. return SUCCESS;
  384. }
  385. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  386. GE_CHK_RT_RET(rtSetDevice(options_.device_id));
  387. }
  388. Status final_state = SUCCESS;
  389. Status mid_state;
  390. GELOGI("engineManager finalization.");
  391. mid_state = engineManager_.Finalize();
  392. if (mid_state != SUCCESS) {
  393. GELOGW("engineManager finalize failed");
  394. final_state = mid_state;
  395. }
  396. GELOGI("sessionManager finalization.");
  397. mid_state = sessionManager_.Finalize();
  398. if (mid_state != SUCCESS) {
  399. GELOGW("sessionManager finalize failed");
  400. final_state = mid_state;
  401. }
  402. GELOGI("opsBuilderManager finalization.");
  403. mid_state = OpsKernelBuilderManager::Instance().Finalize();
  404. if (mid_state != SUCCESS) {
  405. GELOGW("opsBuilderManager finalize failed");
  406. final_state = mid_state;
  407. }
  408. GELOGI("opsManager finalization.");
  409. mid_state = opsManager_.Finalize();
  410. if (mid_state != SUCCESS) {
  411. GELOGW("opsManager finalize failed");
  412. final_state = mid_state;
  413. }
  414. GELOGI("VarManagerPool finalization.");
  415. VarManagerPool::Instance().Destory();
  416. GELOGI("MemManager finalization.");
  417. MemManager::Instance().Finalize();
  418. GELOGI("HostMemManager finalization.");
  419. HostMemManager::Instance().Finalize();
  420. GELOGI("HostCpuEngine finalization.");
  421. HostCpuEngine::GetInstance().Finalize();
  422. GELOGI("Analyzer finalization");
  423. Analyzer::GetInstance()->Finalize();
  424. // Shut down profiling
  425. ShutDownProfiling();
  426. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  427. GELOGI("System ShutDown.");
  428. mid_state = SystemShutdownWithOptions(this->options_);
  429. if (mid_state != SUCCESS) {
  430. GELOGW("System shutdown with options failed");
  431. final_state = mid_state;
  432. }
  433. }
  434. is_train_mode_ = false;
  435. GetMutableGlobalOptions().erase(ENABLE_SINGLE_STREAM);
  436. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  437. GE_CHK_RT_RET(rtDeviceReset(options_.device_id));
  438. }
  439. instancePtr_ = nullptr;
  440. init_flag_ = false;
  441. if (final_state != SUCCESS) {
  442. GELOGE(FAILED, "finalization 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. if (sessionManager_.init_flag_) {
  465. (void)sessionManager_.Finalize();
  466. }
  467. MemManager::Instance().Finalize();
  468. HostMemManager::Instance().Finalize();
  469. VarManagerPool::Instance().Destory();
  470. }
  471. Status GEInit::Initialize(const map<string, string> &options) {
  472. Status ret = SUCCESS;
  473. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  474. if (instance_ptr == nullptr || !instance_ptr->InitFlag()) {
  475. ret = GELib::Initialize(options);
  476. }
  477. return ret;
  478. }
  479. Status GEInit::Finalize() {
  480. return GELib::GetInstance()->Finalize();
  481. }
  482. string GEInit::GetPath() {
  483. return GELib::GetPath();
  484. }
  485. } // 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.