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 21 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
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
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
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581
  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, "[Create][GELib]GeLib initialize failed, malloc shared_ptr failed.");
  64. REPORT_INNER_ERROR("E19999", "GELib Init failed for new GeLib failed.");
  65. return GE_CLI_INIT_FAILED;
  66. }
  67. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kSystemInit);
  68. map<string, string> new_options;
  69. Status ret = instancePtr_->SetRTSocVersion(options, new_options);
  70. if (ret != SUCCESS) {
  71. GELOGE(ret, "[Set][RTSocVersion]GeLib initial: SetRTSocVersion failed.");
  72. REPORT_CALL_ERROR("E19999", "SetRTSocVersion failed.");
  73. return ret;
  74. }
  75. ret = instancePtr_->SetAiCoreNum(new_options);
  76. if (ret != SUCCESS) {
  77. GELOGE(ret, "[Set][AiCoreNum]GeLib initial: SetAiCoreNum failed.");
  78. REPORT_CALL_ERROR("E19999", "SetAiCoreNum failed.");
  79. return ret;
  80. }
  81. instancePtr_->SetDefaultPrecisionMode(new_options);
  82. if (new_options.find("ge.fpCeilingMode") == new_options.end()) {
  83. new_options["ge.fpCeilingMode"] = kGlobalOptionFpCeilingModeDefault;
  84. }
  85. GetMutableGlobalOptions().insert(new_options.begin(), new_options.end());
  86. GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
  87. GE_TIMESTAMP_START(Init);
  88. ret = instancePtr_->InnerInitialize(new_options);
  89. if (ret != SUCCESS) {
  90. GELOGE(ret, "[Init][GeLib]GeLib initial failed.");
  91. REPORT_CALL_ERROR("E19999", "GELib::InnerInitialize failed.");
  92. instancePtr_ = nullptr;
  93. return ret;
  94. }
  95. GE_TIMESTAMP_EVENT_END(Init, "GELib::Initialize");
  96. return SUCCESS;
  97. }
  98. Status GELib::InnerInitialize(const map<string, string> &options) {
  99. // Multiple initializations are not allowed
  100. if (init_flag_) {
  101. GELOGW("multi initializations");
  102. return SUCCESS;
  103. }
  104. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kSystemInit);
  105. GELOGI("GE System initial.");
  106. GE_TIMESTAMP_START(SystemInitialize);
  107. Status initSystemStatus = SystemInitialize(options);
  108. GE_TIMESTAMP_END(SystemInitialize, "InnerInitialize::SystemInitialize");
  109. if (initSystemStatus != SUCCESS) {
  110. GELOGE(initSystemStatus, "[Init][GESystem]GE system initial failed.");
  111. RollbackInit();
  112. return initSystemStatus;
  113. }
  114. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kEngineInit);
  115. GELOGI("engineManager initial.");
  116. GE_TIMESTAMP_START(EngineInitialize);
  117. Status initEmStatus = engineManager_.Initialize(options);
  118. GE_TIMESTAMP_END(EngineInitialize, "InnerInitialize::EngineInitialize");
  119. if (initEmStatus != SUCCESS) {
  120. GELOGE(initEmStatus, "[Init][EngineManager]GE engine manager initial failed.");
  121. REPORT_CALL_ERROR("E19999", "EngineManager initialize failed.");
  122. RollbackInit();
  123. return initEmStatus;
  124. }
  125. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kOpsKernelInit);
  126. GELOGI("opsManager initial.");
  127. GE_TIMESTAMP_START(OpsManagerInitialize);
  128. Status initOpsStatus = opsManager_.Initialize(options);
  129. GE_TIMESTAMP_END(OpsManagerInitialize, "InnerInitialize::OpsManagerInitialize");
  130. if (initOpsStatus != SUCCESS) {
  131. GELOGE(initOpsStatus, "[Init][OpsManager]GE ops manager initial failed.");
  132. REPORT_CALL_ERROR("E19999", "OpsManager initialize failed.");
  133. RollbackInit();
  134. return initOpsStatus;
  135. }
  136. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kOpsKernelBuilderInit);
  137. GELOGI("opsBuilderManager initial.");
  138. GE_TIMESTAMP_START(OpsKernelBuilderManagerInitialize);
  139. Status initOpsBuilderStatus = OpsKernelBuilderManager::Instance().Initialize(options);
  140. GE_TIMESTAMP_END(OpsKernelBuilderManagerInitialize, "InnerInitialize::OpsKernelBuilderManager");
  141. if (initOpsBuilderStatus != SUCCESS) {
  142. GELOGE(initOpsBuilderStatus, "[Init][OpsKernelBuilderManager]GE ops builder manager initial failed.");
  143. REPORT_CALL_ERROR("E19999", "OpsBuilderManager initialize failed.");
  144. RollbackInit();
  145. return initOpsBuilderStatus;
  146. }
  147. ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kOther);
  148. GELOGI("sessionManager initial.");
  149. GE_TIMESTAMP_START(SessionManagerInitialize);
  150. Status initSmStatus = sessionManager_.Initialize(options);
  151. GE_TIMESTAMP_END(SessionManagerInitialize, "InnerInitialize::SessionManagerInitialize");
  152. if (initSmStatus != SUCCESS) {
  153. GELOGE(initSmStatus, "[Init][SessionManager] GE session manager initial failed.");
  154. REPORT_CALL_ERROR("E19999", "SessionManager initialize failed.");
  155. RollbackInit();
  156. return initSmStatus;
  157. }
  158. GELOGI("Start to initialize HostCpuEngine");
  159. GE_TIMESTAMP_START(HostCpuEngineInitialize);
  160. Status initHostCpuEngineStatus = HostCpuEngine::GetInstance().Initialize();
  161. GE_TIMESTAMP_END(HostCpuEngineInitialize, "InnerInitialize::HostCpuEngineInitialize");
  162. if (initHostCpuEngineStatus != SUCCESS) {
  163. GELOGE(initHostCpuEngineStatus, "[Init][HostCpuEngine]Failed to initialize HostCpuEngine.");
  164. REPORT_CALL_ERROR("E19999", "HostCpuEngine initialize failed.");
  165. RollbackInit();
  166. return initHostCpuEngineStatus;
  167. }
  168. GELOGI("Start to init Analyzer!");
  169. Status init_analyzer_status = ge::Analyzer::GetInstance()->Initialize();
  170. if (init_analyzer_status != SUCCESS) {
  171. GELOGE(init_analyzer_status, "[Init][Analyzer]Failed to initialize Analyzer.");
  172. REPORT_CALL_ERROR("E19999", "ge::Analyzer initialize failed.");
  173. RollbackInit();
  174. return init_analyzer_status;
  175. }
  176. init_flag_ = true;
  177. return SUCCESS;
  178. }
  179. Status GELib::SystemInitialize(const map<string, string> &options) {
  180. Status status = FAILED;
  181. auto iter = options.find(OPTION_GRAPH_RUN_MODE);
  182. if (iter != options.end()) {
  183. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  184. is_train_mode_ = true;
  185. }
  186. }
  187. InitOptions(options);
  188. // In train and infer, profiling is always needed.
  189. InitProfiling(this->options_);
  190. auto model_manager = ModelManager::GetInstance();
  191. GE_CHECK_NOTNULL(model_manager);
  192. GE_IF_BOOL_EXEC(model_manager->EnableExceptionDump(options) != SUCCESS,
  193. REPORT_CALL_ERROR("E19999", "ModelManager EnableExceptionDump failed.");
  194. GELOGE(FAILED, "[Enable][ExceptionDump] failed.");
  195. return FAILED);
  196. // 1.`is_train_mode_` means case: train
  197. // 2.`(!is_train_mode_) && (options_.device_id != kDefaultDeviceIdForInfer)` means case: online infer
  198. // these two case with logical device id
  199. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  200. status = InitSystemWithOptions(this->options_);
  201. } else {
  202. status = InitSystemWithoutOptions();
  203. }
  204. return status;
  205. }
  206. void GELib::InitProfiling(Options &options) {
  207. GELOGI("Init Profiling. session Id: %ld, device id:%d ", options.session_id, options.device_id);
  208. std::lock_guard<std::mutex> lock(status_mutex_);
  209. GetContext().Init();
  210. // Profiling init
  211. if (ProfilingManager::Instance().Init(options) != SUCCESS) {
  212. GELOGW("Profiling init failed.");
  213. }
  214. }
  215. void GELib::SetDefaultPrecisionMode(map<string, string> &new_options) {
  216. auto iter = new_options.find(PRECISION_MODE);
  217. if (iter != new_options.end()) {
  218. GELOGI("Find precision_mode in options, value is %s", iter->second.c_str());
  219. return;
  220. }
  221. iter = new_options.find(OPTION_GRAPH_RUN_MODE);
  222. if (iter != new_options.end()) {
  223. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  224. // only train mode need to be set allow_fp32_to_fp16.
  225. GELOGI("This is train mode, precision_mode need to be set allow_fp32_to_fp16");
  226. new_options.insert(std::make_pair(PRECISION_MODE, "allow_fp32_to_fp16"));
  227. return;
  228. }
  229. }
  230. GELOGI("This is not train mode, precision_mode need to be set force_fp16");
  231. new_options.insert(std::make_pair(PRECISION_MODE, "force_fp16"));
  232. return;
  233. }
  234. Status GELib::SetRTSocVersion(const map<string, string> &options, map<string, string> &new_options) {
  235. GELOGI("Start to set SOC_VERSION");
  236. new_options.insert(options.begin(), options.end());
  237. auto it = new_options.find(ge::SOC_VERSION);
  238. if (it != new_options.end()) {
  239. GE_CHK_RT_RET(rtSetSocVersion(it->second.c_str()));
  240. GELOGI("Succeeded in setting SOC_VERSION[%s] to runtime.", it->second.c_str());
  241. } else {
  242. GELOGI("SOC_VERSION is not exist in options");
  243. char version[kSocVersionLen] = {0};
  244. rtError_t rt_ret = rtGetSocVersion(version, kSocVersionLen);
  245. GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE,
  246. REPORT_CALL_ERROR("E19999", "rtGetSocVersion failed.");
  247. GELOGE(rt_ret, "[Get][SocVersion]rtGetSocVersion failed");
  248. return FAILED;)
  249. GELOGI("Succeeded in getting SOC_VERSION[%s] from runtime.", version);
  250. new_options.insert(std::make_pair(ge::SOC_VERSION, version));
  251. }
  252. return SUCCESS;
  253. }
  254. Status GELib::SetAiCoreNum(map<string, string> &options) {
  255. // Already set or get AICORE_NUM from options in offline mode
  256. if (options.find(AICORE_NUM) != options.end()) {
  257. return SUCCESS;
  258. }
  259. uint32_t aicore_num = 0;
  260. rtError_t ret = rtGetAiCoreCount(&aicore_num);
  261. if (ret == ACL_ERROR_RT_FEATURE_NOT_SUPPORT) { // offline without ATC Input of AiCoreNum
  262. return SUCCESS;
  263. } else if (ret == RT_ERROR_NONE) { // online-mode
  264. options.emplace(std::make_pair(AICORE_NUM, std::to_string(aicore_num)));
  265. return SUCCESS;
  266. }
  267. GELOGE(FAILED, "[Get][AiCoreCount]rtGetAiCoreCount failed.");
  268. REPORT_CALL_ERROR("E19999", "rtGetAiCoreCount failed.");
  269. return FAILED;
  270. }
  271. void GELib::InitOptions(const map<string, string> &options) {
  272. this->options_.session_id = 0;
  273. auto iter = options.find(OPTION_EXEC_SESSION_ID);
  274. if (iter != options.end()) {
  275. this->options_.session_id = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  276. }
  277. this->options_.device_id = is_train_mode_ ? kDefaultDeviceIdForTrain : kDefaultDeviceIdForInfer;
  278. iter = options.find(OPTION_EXEC_DEVICE_ID);
  279. if (iter != options.end()) {
  280. this->options_.device_id = static_cast<int32_t>(std::strtol(iter->second.c_str(), nullptr, kDecimal));
  281. }
  282. iter = options.find(OPTION_EXEC_JOB_ID);
  283. if (iter != options.end()) {
  284. this->options_.job_id = iter->second.c_str();
  285. }
  286. this->options_.isUseHcom = false;
  287. iter = options.find(OPTION_EXEC_IS_USEHCOM);
  288. if (iter != options.end()) {
  289. std::istringstream(iter->second) >> this->options_.isUseHcom;
  290. }
  291. this->options_.isUseHvd = false;
  292. iter = options.find(OPTION_EXEC_IS_USEHVD);
  293. if (iter != options.end()) {
  294. std::istringstream(iter->second) >> this->options_.isUseHvd;
  295. }
  296. this->options_.deployMode = false;
  297. iter = options.find(OPTION_EXEC_DEPLOY_MODE);
  298. if (iter != options.end()) {
  299. std::istringstream(iter->second) >> this->options_.deployMode;
  300. }
  301. iter = options.find(OPTION_EXEC_POD_NAME);
  302. if (iter != options.end()) {
  303. this->options_.podName = iter->second.c_str();
  304. }
  305. iter = options.find(OPTION_EXEC_PROFILING_MODE);
  306. if (iter != options.end()) {
  307. this->options_.profiling_mode = iter->second.c_str();
  308. }
  309. iter = options.find(OPTION_EXEC_PROFILING_OPTIONS);
  310. if (iter != options.end()) {
  311. this->options_.profiling_options = iter->second.c_str();
  312. }
  313. iter = options.find(OPTION_EXEC_RANK_ID);
  314. if (iter != options.end()) {
  315. this->options_.rankId = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  316. }
  317. iter = options.find(OPTION_EXEC_RANK_TABLE_FILE);
  318. if (iter != options.end()) {
  319. this->options_.rankTableFile = iter->second.c_str();
  320. }
  321. this->options_.enable_atomic = true;
  322. iter = options.find(OPTION_EXEC_ATOMIC_FLAG);
  323. GE_IF_BOOL_EXEC(iter != options.end(),
  324. this->options_.enable_atomic = std::strtol(iter->second.c_str(), nullptr, kDecimal));
  325. GELOGI("ge InnerInitialize, the enable_atomic_flag in options_ is %d", this->options_.enable_atomic);
  326. }
  327. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithOptions(Options &options) {
  328. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  329. GELOGI("%s init GELib. session Id:%ld, device id :%d ", mode.c_str(), options.session_id, options.device_id);
  330. GEEVENT("System init with options begin, job id %s", options.job_id.c_str());
  331. std::lock_guard<std::mutex> lock(status_mutex_);
  332. GE_IF_BOOL_EXEC(is_system_inited && !is_shutdown,
  333. GELOGW("System init with options is already inited and not shutdown.");
  334. return SUCCESS);
  335. std::vector<rtMemType_t> mem_type;
  336. mem_type.push_back(RT_MEMORY_HBM);
  337. mem_type.push_back(RT_MEMORY_P2P_DDR);
  338. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  339. if (initMmStatus != SUCCESS) {
  340. GELOGE(initMmStatus, "[Init][MemManager] MemoryAllocatorManager initialize failed.");
  341. REPORT_CALL_ERROR("E19999", "MemManager initialize failed.");
  342. return initMmStatus;
  343. }
  344. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  345. // Update CSA file
  346. CsaInteract::GetInstance().Init(options.device_id, GetContext().TraceId());
  347. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_RUNNING, JOBSUBSTATE_ENV_INIT);
  348. GE_LOGE_IF(ret != SUCCESS, "[Write][JobState] failed, ret:%u ", ret);
  349. // set device id
  350. GELOGI("set logical device id:%u", options.device_id);
  351. GetContext().SetCtxDeviceId(static_cast<uint32_t>(options.device_id));
  352. GE_CHK_RT_RET(rtSetDevice(options.device_id));
  353. // In the scenario that the automatic add fusion is set, but there is no cleanaddr operator,
  354. // maybe need to check it
  355. is_system_inited = true;
  356. is_shutdown = false;
  357. GELOGI("%s init GELib success.", mode.c_str());
  358. return SUCCESS;
  359. }
  360. Status GELib::SystemShutdownWithOptions(const Options &options) {
  361. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  362. GELOGI("%s finalize GELib begin.", mode.c_str());
  363. std::lock_guard<std::mutex> lock(status_mutex_);
  364. GE_IF_BOOL_EXEC(is_shutdown || !is_system_inited,
  365. GELOGW("System Shutdown with options is already is_shutdown or system does not inited. "
  366. "is_shutdown:%d is_omm_inited:%d",
  367. is_shutdown, is_system_inited);
  368. return SUCCESS);
  369. GE_CHK_RT(rtDeviceReset(options.device_id));
  370. // Update CSA file
  371. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_SUCCEED);
  372. GE_LOGE_IF(ret != SUCCESS, "[Write][JobState] failed, ret:%u ", ret);
  373. is_system_inited = false;
  374. is_shutdown = true;
  375. GELOGI("%s finalize GELib success.", mode.c_str());
  376. return SUCCESS;
  377. }
  378. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithoutOptions() {
  379. GELOGI("Inference Init GELib begin.");
  380. std::vector<rtMemType_t> mem_type;
  381. mem_type.push_back(RT_MEMORY_HBM);
  382. mem_type.push_back(RT_MEMORY_P2P_DDR);
  383. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  384. if (initMmStatus != SUCCESS) {
  385. GELOGE(initMmStatus, "[Init][MemoryManager] initialize failed.");
  386. REPORT_CALL_ERROR("E19999", "MemManager initialize failed.");
  387. return initMmStatus;
  388. }
  389. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  390. static bool is_inited = false;
  391. if (is_inited) {
  392. GELOGW("System init without options is already inited, don't need to init again.");
  393. return SUCCESS;
  394. }
  395. is_inited = true;
  396. GELOGI("Inference init GELib success.");
  397. return SUCCESS;
  398. }
  399. string GELib::GetPath() { return PluginManager::GetPath(); }
  400. // Finalize all modules
  401. Status GELib::Finalize() {
  402. ErrorManager::GetInstance().SetStage(ErrorMessage::kFinalize, ErrorMessage::kFinalize);
  403. GELOGI("finalization start");
  404. // Finalization is not allowed before initialization
  405. if (!init_flag_) {
  406. GELOGW("not initialize");
  407. return SUCCESS;
  408. }
  409. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  410. GE_CHK_RT_RET(rtSetDevice(options_.device_id));
  411. }
  412. Status final_state = SUCCESS;
  413. Status mid_state;
  414. GELOGI("engineManager finalization.");
  415. mid_state = engineManager_.Finalize();
  416. if (mid_state != SUCCESS) {
  417. GELOGW("engineManager finalize failed");
  418. final_state = mid_state;
  419. }
  420. GELOGI("sessionManager finalization.");
  421. mid_state = sessionManager_.Finalize();
  422. if (mid_state != SUCCESS) {
  423. GELOGW("sessionManager finalize failed");
  424. final_state = mid_state;
  425. }
  426. GELOGI("opsBuilderManager finalization.");
  427. mid_state = OpsKernelBuilderManager::Instance().Finalize();
  428. if (mid_state != SUCCESS) {
  429. GELOGW("opsBuilderManager finalize failed");
  430. final_state = mid_state;
  431. }
  432. GELOGI("opsManager finalization.");
  433. mid_state = opsManager_.Finalize();
  434. if (mid_state != SUCCESS) {
  435. GELOGW("opsManager finalize failed");
  436. final_state = mid_state;
  437. }
  438. GELOGI("VarManagerPool finalization.");
  439. VarManagerPool::Instance().Destory();
  440. GELOGI("MemManager finalization.");
  441. MemManager::Instance().Finalize();
  442. GELOGI("HostMemManager finalization.");
  443. HostMemManager::Instance().Finalize();
  444. GELOGI("HostCpuEngine finalization.");
  445. HostCpuEngine::GetInstance().Finalize();
  446. GELOGI("Analyzer finalization");
  447. Analyzer::GetInstance()->Finalize();
  448. // Shut down profiling
  449. ShutDownProfiling();
  450. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  451. GELOGI("System ShutDown.");
  452. mid_state = SystemShutdownWithOptions(this->options_);
  453. if (mid_state != SUCCESS) {
  454. GELOGW("System shutdown with options failed");
  455. final_state = mid_state;
  456. }
  457. }
  458. is_train_mode_ = false;
  459. GetMutableGlobalOptions().erase(ENABLE_SINGLE_STREAM);
  460. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  461. GE_CHK_RT_RET(rtDeviceReset(options_.device_id));
  462. }
  463. instancePtr_ = nullptr;
  464. init_flag_ = false;
  465. if (final_state != SUCCESS) {
  466. GELOGE(FAILED, "[Check][State]finalization failed.");
  467. REPORT_INNER_ERROR("E19999", "GELib::Finalize failed.");
  468. return final_state;
  469. }
  470. GELOGI("finalization success.");
  471. return SUCCESS;
  472. }
  473. void GELib::ShutDownProfiling() {
  474. std::lock_guard<std::mutex> lock(status_mutex_);
  475. if (ProfilingManager::Instance().ProfilingOn()) {
  476. ProfilingManager::Instance().StopProfiling();
  477. ProfilingManager::Instance().PluginUnInit();
  478. }
  479. }
  480. // Get Singleton Instance
  481. std::shared_ptr<GELib> GELib::GetInstance() { return instancePtr_; }
  482. void GELib::RollbackInit() {
  483. if (engineManager_.init_flag_) {
  484. (void)engineManager_.Finalize();
  485. }
  486. if (opsManager_.init_flag_) {
  487. (void)opsManager_.Finalize();
  488. }
  489. if (sessionManager_.init_flag_) {
  490. (void)sessionManager_.Finalize();
  491. }
  492. MemManager::Instance().Finalize();
  493. HostMemManager::Instance().Finalize();
  494. VarManagerPool::Instance().Destory();
  495. }
  496. Status GEInit::Initialize(const map<string, string> &options) {
  497. Status ret = SUCCESS;
  498. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  499. if (instance_ptr == nullptr || !instance_ptr->InitFlag()) {
  500. ret = GELib::Initialize(options);
  501. }
  502. return ret;
  503. }
  504. Status GEInit::Finalize() {
  505. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  506. if (instance_ptr != nullptr) {
  507. return instance_ptr->Finalize();
  508. }
  509. return SUCCESS;
  510. }
  511. string GEInit::GetPath() {
  512. return GELib::GetPath();
  513. }
  514. } // 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