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 16 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467
  1. /**
  2. * Copyright 2019-2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #include "init/gelib.h"
  17. #include <dlfcn.h>
  18. #include <cstdlib>
  19. #include <mutex>
  20. #include <set>
  21. #include <sstream>
  22. #include <string>
  23. #include <utility>
  24. #include "common/ge/ge_util.h"
  25. #include "common/ge/plugin_manager.h"
  26. #include "common/profiling/profiling_manager.h"
  27. #include "common/properties_manager.h"
  28. #include "framework/common/debug/ge_log.h"
  29. #include "framework/common/debug/log.h"
  30. #include "framework/common/util.h"
  31. #include "analyzer/analyzer.h"
  32. #include "ge/ge_api_types.h"
  33. #include "ge_local_engine/engine/host_cpu_engine.h"
  34. #include "graph/common/ge_call_wrapper.h"
  35. #include "graph/ge_context.h"
  36. #include "graph/ge_global_options.h"
  37. #include "graph/load/new_model_manager/model_manager.h"
  38. #include "graph/manager/graph_mem_allocator.h"
  39. #include "graph/manager/host_mem_manager.h"
  40. #include "graph/manager/graph_var_manager.h"
  41. #include "omm/csa_interact.h"
  42. #include "runtime/kernel.h"
  43. using Json = nlohmann::json;
  44. namespace ge {
  45. namespace {
  46. const int kDecimal = 10;
  47. const int kSocVersionLen = 50;
  48. const int kDefaultDeviceIdForTrain = 0;
  49. const int kDefaultDeviceIdForInfer = -1;
  50. const uint32_t kAicoreOverflow = (0x1 << 0);
  51. const uint32_t kAtomicOverflow = (0x1 << 1);
  52. const uint32_t kAllOverflow = (kAicoreOverflow | kAtomicOverflow);
  53. } // namespace
  54. static std::shared_ptr<GELib> instancePtr_ = nullptr;
  55. // Initial each module of GE, if one failed, release all
  56. Status GELib::Initialize(const map<string, string> &options) {
  57. GELOGI("initial start");
  58. GEEVENT("[GEPERFTRACE] GE Init Start");
  59. // Multiple initializations are not allowed
  60. instancePtr_ = MakeShared<GELib>();
  61. if (instancePtr_ == nullptr) {
  62. GELOGE(GE_CLI_INIT_FAILED, "GeLib initialize failed, malloc shared_ptr failed.");
  63. return GE_CLI_INIT_FAILED;
  64. }
  65. map<string, string> new_options;
  66. Status ret = instancePtr_->SetRTSocVersion(options, new_options);
  67. if (ret != SUCCESS) {
  68. GELOGE(ret, "GeLib initial failed.");
  69. return ret;
  70. }
  71. GetMutableGlobalOptions().insert(new_options.begin(), new_options.end());
  72. GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
  73. GE_TIMESTAMP_START(Init);
  74. ret = instancePtr_->InnerInitialize(new_options);
  75. if (ret != SUCCESS) {
  76. GELOGE(ret, "GeLib initial failed.");
  77. instancePtr_ = nullptr;
  78. return ret;
  79. }
  80. GE_TIMESTAMP_EVENT_END(Init, "GELib::Initialize");
  81. return SUCCESS;
  82. }
  83. Status GELib::InnerInitialize(const map<string, string> &options) {
  84. // Multiple initializations are not allowed
  85. if (init_flag_) {
  86. GELOGW("multi initializations");
  87. return SUCCESS;
  88. }
  89. GELOGI("GE System initial.");
  90. GE_TIMESTAMP_START(SystemInitialize);
  91. Status initSystemStatus = SystemInitialize(options);
  92. GE_TIMESTAMP_END(SystemInitialize, "InnerInitialize::SystemInitialize");
  93. if (initSystemStatus != SUCCESS) {
  94. GELOGE(initSystemStatus);
  95. RollbackInit();
  96. return initSystemStatus;
  97. }
  98. GELOGI("engineManager initial.");
  99. GE_TIMESTAMP_START(EngineInitialize);
  100. Status initEmStatus = engineManager_.Initialize(options);
  101. GE_TIMESTAMP_END(EngineInitialize, "InnerInitialize::EngineInitialize");
  102. if (initEmStatus != SUCCESS) {
  103. GELOGE(initEmStatus);
  104. RollbackInit();
  105. return initEmStatus;
  106. }
  107. GELOGI("opsManager initial.");
  108. GE_TIMESTAMP_START(OpsManagerInitialize);
  109. Status initOpsStatus = opsManager_.Initialize(options);
  110. GE_TIMESTAMP_END(OpsManagerInitialize, "InnerInitialize::OpsManagerInitialize");
  111. if (initOpsStatus != SUCCESS) {
  112. GELOGE(initOpsStatus);
  113. RollbackInit();
  114. return initOpsStatus;
  115. }
  116. GELOGI("sessionManager initial.");
  117. GE_TIMESTAMP_START(SessionManagerInitialize);
  118. Status initSmStatus = sessionManager_.Initialize(options);
  119. GE_TIMESTAMP_END(SessionManagerInitialize, "InnerInitialize::SessionManagerInitialize");
  120. if (initSmStatus != SUCCESS) {
  121. GELOGE(initSmStatus);
  122. RollbackInit();
  123. return initSmStatus;
  124. }
  125. GELOGI("Start to initialize HostCpuEngine");
  126. GE_TIMESTAMP_START(HostCpuEngineInitialize);
  127. Status initHostCpuEngineStatus = HostCpuEngine::GetInstance().Initialize();
  128. GE_TIMESTAMP_END(HostCpuEngineInitialize, "InnerInitialize::HostCpuEngineInitialize");
  129. if (initHostCpuEngineStatus != SUCCESS) {
  130. GELOGE(initHostCpuEngineStatus, "Failed to initialize HostCpuEngine");
  131. RollbackInit();
  132. return initHostCpuEngineStatus;
  133. }
  134. GELOGI("Start to init Analyzer!");
  135. Status init_analyzer_status = ge::Analyzer::GetInstance()->Initialize();
  136. if (init_analyzer_status != SUCCESS) {
  137. GELOGE(init_analyzer_status, "Failed to initialize HostCpuEngine");
  138. RollbackInit();
  139. return init_analyzer_status;
  140. }
  141. init_flag_ = true;
  142. return SUCCESS;
  143. }
  144. Status GELib::SystemInitialize(const map<string, string> &options) {
  145. Status status = FAILED;
  146. auto iter = options.find(OPTION_GRAPH_RUN_MODE);
  147. if (iter != options.end()) {
  148. if (GraphRunMode(std::strtol(iter->second.c_str(), nullptr, kDecimal)) >= TRAIN) {
  149. is_train_mode_ = true;
  150. }
  151. }
  152. InitOptions(options);
  153. // In train and infer, profiling is always needed.
  154. InitProfiling(this->options_);
  155. auto model_manager = ModelManager::GetInstance();
  156. GE_CHECK_NOTNULL(model_manager);
  157. GE_IF_BOOL_EXEC(model_manager->EnableExceptionDump(options) != SUCCESS,
  158. GELOGE(FAILED, "Enable exception dump failed");
  159. return FAILED);
  160. // 1.`is_train_mode_` means case: train
  161. // 2.`(!is_train_mode_) && (options_.device_id != kDefaultDeviceIdForInfer)` means case: online infer
  162. // these two case with logical device id
  163. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  164. status = InitSystemWithOptions(this->options_);
  165. } else {
  166. status = InitSystemWithoutOptions();
  167. }
  168. return status;
  169. }
  170. void GELib::InitProfiling(Options &options) {
  171. GELOGI("Init Profiling. session Id: %ld, device id:%d ", options.session_id, options.device_id);
  172. std::lock_guard<std::mutex> lock(status_mutex_);
  173. GetContext().Init();
  174. // Profiling init
  175. if (ProfilingManager::Instance().Init(options) != SUCCESS) {
  176. GELOGW("Profiling init failed.");
  177. }
  178. }
  179. Status GELib::SetRTSocVersion(const map<string, string> &options, map<string, string> &new_options) {
  180. GELOGI("Start to set SOC_VERSION");
  181. new_options.insert(options.begin(), options.end());
  182. auto it = new_options.find(ge::SOC_VERSION);
  183. if (it != new_options.end()) {
  184. GE_CHK_RT_RET(rtSetSocVersion(it->second.c_str()));
  185. GELOGI("Succeeded in setting SOC_VERSION[%s] to runtime.", it->second.c_str());
  186. } else {
  187. GELOGI("SOC_VERSION is not exist in options");
  188. char version[kSocVersionLen] = {0};
  189. rtError_t rt_ret = rtGetSocVersion(version, kSocVersionLen);
  190. GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(rt_ret, "rtGetSocVersion failed"); return FAILED;)
  191. GELOGI("Succeeded in getting SOC_VERSION[%s] from runtime.", version);
  192. new_options.insert(std::make_pair(ge::SOC_VERSION, version));
  193. }
  194. return SUCCESS;
  195. }
  196. void GELib::InitOptions(const map<string, string> &options) {
  197. this->options_.session_id = 0;
  198. auto iter = options.find(OPTION_EXEC_SESSION_ID);
  199. if (iter != options.end()) {
  200. this->options_.session_id = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  201. }
  202. this->options_.device_id = is_train_mode_ ? kDefaultDeviceIdForTrain : kDefaultDeviceIdForInfer;
  203. iter = options.find(OPTION_EXEC_DEVICE_ID);
  204. if (iter != options.end()) {
  205. this->options_.device_id = static_cast<int32_t>(std::strtol(iter->second.c_str(), nullptr, kDecimal));
  206. }
  207. iter = options.find(OPTION_EXEC_JOB_ID);
  208. if (iter != options.end()) {
  209. this->options_.job_id = iter->second.c_str();
  210. }
  211. this->options_.isUseHcom = false;
  212. iter = options.find(OPTION_EXEC_IS_USEHCOM);
  213. if (iter != options.end()) {
  214. std::istringstream(iter->second) >> this->options_.isUseHcom;
  215. }
  216. this->options_.isUseHvd = false;
  217. iter = options.find(OPTION_EXEC_IS_USEHVD);
  218. if (iter != options.end()) {
  219. std::istringstream(iter->second) >> this->options_.isUseHvd;
  220. }
  221. this->options_.deployMode = false;
  222. iter = options.find(OPTION_EXEC_DEPLOY_MODE);
  223. if (iter != options.end()) {
  224. std::istringstream(iter->second) >> this->options_.deployMode;
  225. }
  226. iter = options.find(OPTION_EXEC_POD_NAME);
  227. if (iter != options.end()) {
  228. this->options_.podName = iter->second.c_str();
  229. }
  230. iter = options.find(OPTION_EXEC_PROFILING_MODE);
  231. if (iter != options.end()) {
  232. this->options_.profiling_mode = iter->second.c_str();
  233. }
  234. iter = options.find(OPTION_EXEC_PROFILING_OPTIONS);
  235. if (iter != options.end()) {
  236. this->options_.profiling_options = iter->second.c_str();
  237. }
  238. iter = options.find(OPTION_EXEC_RANK_ID);
  239. if (iter != options.end()) {
  240. this->options_.rankId = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  241. }
  242. iter = options.find(OPTION_EXEC_RANK_TABLE_FILE);
  243. if (iter != options.end()) {
  244. this->options_.rankTableFile = iter->second.c_str();
  245. }
  246. this->options_.enable_atomic = true;
  247. iter = options.find(OPTION_EXEC_ATOMIC_FLAG);
  248. GE_IF_BOOL_EXEC(iter != options.end(),
  249. this->options_.enable_atomic = std::strtol(iter->second.c_str(), nullptr, kDecimal));
  250. GELOGI("ge InnerInitialize, the enable_atomic_flag in options_ is %d", this->options_.enable_atomic);
  251. }
  252. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithOptions(Options &options) {
  253. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  254. GELOGI("%s init GELib. session Id:%ld, device id :%d ", mode.c_str(), options.session_id, options.device_id);
  255. GEEVENT("System init with options begin, job id %s", options.job_id.c_str());
  256. std::lock_guard<std::mutex> lock(status_mutex_);
  257. GE_IF_BOOL_EXEC(is_system_inited && !is_shutdown,
  258. GELOGW("System init with options is already inited and not shutdown.");
  259. return SUCCESS);
  260. std::vector<rtMemType_t> mem_type;
  261. mem_type.push_back(RT_MEMORY_HBM);
  262. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  263. if (initMmStatus != SUCCESS) {
  264. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  265. return initMmStatus;
  266. }
  267. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  268. // Update CSA file
  269. CsaInteract::GetInstance().Init(options.device_id, GetContext().TraceId());
  270. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_RUNNING, JOBSUBSTATE_ENV_INIT);
  271. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  272. // set device id
  273. GELOGI("set logical device id:%u", options.device_id);
  274. GetContext().SetCtxDeviceId(static_cast<uint32_t>(options.device_id));
  275. GE_CHK_RT_RET(rtSetDevice(options.device_id));
  276. // In the scenario that the automatic add fusion is set, but there is no cleanaddr operator,
  277. // maybe need to check it
  278. is_system_inited = true;
  279. is_shutdown = false;
  280. GELOGI("%s init GELib success.", mode.c_str());
  281. return SUCCESS;
  282. }
  283. Status GELib::SystemShutdownWithOptions(const Options &options) {
  284. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  285. GELOGI("%s finalize GELib begin.", mode.c_str());
  286. std::lock_guard<std::mutex> lock(status_mutex_);
  287. GE_IF_BOOL_EXEC(is_shutdown || !is_system_inited,
  288. GELOGW("System Shutdown with options is already is_shutdown or system does not inited. "
  289. "is_shutdown:%d is_omm_inited:%d",
  290. is_shutdown, is_system_inited);
  291. return SUCCESS);
  292. GE_CHK_RT(rtDeviceReset(options.device_id));
  293. // Update CSA file
  294. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_SUCCEED);
  295. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  296. is_system_inited = false;
  297. is_shutdown = true;
  298. GELOGI("%s finalize GELib success.", mode.c_str());
  299. return SUCCESS;
  300. }
  301. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithoutOptions() {
  302. GELOGI("Inference Init GELib begin.");
  303. std::vector<rtMemType_t> mem_type;
  304. mem_type.push_back(RT_MEMORY_HBM);
  305. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  306. if (initMmStatus != SUCCESS) {
  307. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  308. return initMmStatus;
  309. }
  310. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  311. static bool is_inited = false;
  312. if (is_inited) {
  313. GELOGW("System init without options is already inited, don't need to init again.");
  314. return SUCCESS;
  315. }
  316. is_inited = true;
  317. GELOGI("Inference init GELib success.");
  318. return SUCCESS;
  319. }
  320. string GELib::GetPath() { return PluginManager::GetPath(); }
  321. // Finalize all modules
  322. Status GELib::Finalize() {
  323. GELOGI("finalization start");
  324. // Finalization is not allowed before initialization
  325. if (!init_flag_) {
  326. GELOGW("not initialize");
  327. return SUCCESS;
  328. }
  329. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  330. GE_CHK_RT_RET(rtSetDevice(options_.device_id));
  331. }
  332. Status final_state = SUCCESS;
  333. Status mid_state;
  334. GELOGI("engineManager finalization.");
  335. mid_state = engineManager_.Finalize();
  336. if (mid_state != SUCCESS) {
  337. GELOGW("engineManager finalize failed");
  338. final_state = mid_state;
  339. }
  340. GELOGI("sessionManager finalization.");
  341. mid_state = sessionManager_.Finalize();
  342. if (mid_state != SUCCESS) {
  343. GELOGW("sessionManager finalize failed");
  344. final_state = mid_state;
  345. }
  346. GELOGI("opsManager finalization.");
  347. mid_state = opsManager_.Finalize();
  348. if (mid_state != SUCCESS) {
  349. GELOGW("opsManager finalize failed");
  350. final_state = mid_state;
  351. }
  352. GELOGI("VarManagerPool finalization.");
  353. VarManagerPool::Instance().Destory();
  354. GELOGI("MemManager finalization.");
  355. MemManager::Instance().Finalize();
  356. GELOGI("HostMemManager finalization.");
  357. HostMemManager::Instance().Finalize();
  358. GELOGI("HostCpuEngine finalization.");
  359. HostCpuEngine::GetInstance().Finalize();
  360. GELOGI("Analyzer finalization");
  361. Analyzer::GetInstance()->Finalize();
  362. // Shut down profiling
  363. ShutDownProfiling();
  364. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  365. GELOGI("System ShutDown.");
  366. mid_state = SystemShutdownWithOptions(this->options_);
  367. if (mid_state != SUCCESS) {
  368. GELOGW("System shutdown with options failed");
  369. final_state = mid_state;
  370. }
  371. }
  372. is_train_mode_ = false;
  373. GetMutableGlobalOptions().erase(ENABLE_SINGLE_STREAM);
  374. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  375. GE_CHK_RT_RET(rtDeviceReset(options_.device_id));
  376. }
  377. instancePtr_ = nullptr;
  378. init_flag_ = false;
  379. if (final_state != SUCCESS) {
  380. GELOGE(FAILED, "finalization failed.");
  381. return final_state;
  382. }
  383. GELOGI("finalization success.");
  384. return SUCCESS;
  385. }
  386. void GELib::ShutDownProfiling() {
  387. std::lock_guard<std::mutex> lock(status_mutex_);
  388. if (!ProfilingManager::Instance().ProfilingOpTraceOn() && ProfilingManager::Instance().ProfilingOn()) {
  389. ProfilingManager::Instance().StopProfiling();
  390. }
  391. if (ProfilingManager::Instance().ProfilingOn()) {
  392. ProfilingManager::Instance().PluginUnInit(GE_PROFILING_MODULE);
  393. }
  394. }
  395. // Get Singleton Instance
  396. std::shared_ptr<GELib> GELib::GetInstance() { return instancePtr_; }
  397. void GELib::RollbackInit() {
  398. if (engineManager_.init_flag_) {
  399. (void)engineManager_.Finalize();
  400. }
  401. if (opsManager_.init_flag_) {
  402. (void)opsManager_.Finalize();
  403. }
  404. if (sessionManager_.init_flag_) {
  405. (void)sessionManager_.Finalize();
  406. }
  407. MemManager::Instance().Finalize();
  408. HostMemManager::Instance().Finalize();
  409. VarManagerPool::Instance().Destory();
  410. }
  411. } // 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