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
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465
  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 "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. // In train and infer, profiling is always needed.
  153. InitOptions(options);
  154. auto model_manager = ModelManager::GetInstance();
  155. GE_CHECK_NOTNULL(model_manager);
  156. GE_IF_BOOL_EXEC(model_manager->EnableExceptionDump(options) != SUCCESS,
  157. GELOGE(FAILED, "Enable exception dump failed");
  158. return FAILED);
  159. // 1.`is_train_mode_` means case: train
  160. // 2.`(!is_train_mode_) && (options_.device_id != kDefaultDeviceIdForInfer)` means case: online infer
  161. // these two case with logical device id
  162. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  163. InitProfiling(this->options_, true);
  164. status = InitSystemWithOptions(this->options_);
  165. } else {
  166. InitProfiling(this->options_);
  167. status = InitSystemWithoutOptions();
  168. }
  169. return status;
  170. }
  171. void GELib::InitProfiling(Options &options, bool convert_2_phy_device_id) {
  172. GELOGI("Init Profiling. session Id: %ld, device id:%d ", options.session_id, options.device_id);
  173. std::lock_guard<std::mutex> lock(status_mutex_);
  174. GetContext().Init();
  175. // Profiling init
  176. if (ProfilingManager::Instance().Init(options, convert_2_phy_device_id) != SUCCESS) {
  177. GELOGW("Profiling init failed.");
  178. }
  179. }
  180. Status GELib::SetRTSocVersion(const map<string, string> &options, map<string, string> &new_options) {
  181. GELOGI("Start to set SOC_VERSION");
  182. new_options.insert(options.begin(), options.end());
  183. auto it = new_options.find(ge::SOC_VERSION);
  184. if (it != new_options.end()) {
  185. GE_CHK_RT_RET(rtSetSocVersion(it->second.c_str()));
  186. GELOGI("Succeeded in setting SOC_VERSION[%s] to runtime.", it->second.c_str());
  187. } else {
  188. GELOGI("SOC_VERSION is not exist in options");
  189. char version[kSocVersionLen] = {0};
  190. rtError_t rt_ret = rtGetSocVersion(version, kSocVersionLen);
  191. GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(rt_ret, "rtGetSocVersion failed"); return FAILED;)
  192. GELOGI("Succeeded in getting SOC_VERSION[%s] from runtime.", version);
  193. new_options.insert(std::make_pair(ge::SOC_VERSION, version));
  194. }
  195. return SUCCESS;
  196. }
  197. void GELib::InitOptions(const map<string, string> &options) {
  198. this->options_.session_id = 0;
  199. auto iter = options.find(OPTION_EXEC_SESSION_ID);
  200. if (iter != options.end()) {
  201. this->options_.session_id = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  202. }
  203. this->options_.device_id = is_train_mode_ ? kDefaultDeviceIdForTrain : kDefaultDeviceIdForInfer;
  204. iter = options.find(OPTION_EXEC_DEVICE_ID);
  205. if (iter != options.end()) {
  206. this->options_.device_id = static_cast<int32_t>(std::strtol(iter->second.c_str(), nullptr, kDecimal));
  207. }
  208. iter = options.find(OPTION_EXEC_JOB_ID);
  209. if (iter != options.end()) {
  210. this->options_.job_id = iter->second.c_str();
  211. }
  212. this->options_.isUseHcom = false;
  213. iter = options.find(OPTION_EXEC_IS_USEHCOM);
  214. if (iter != options.end()) {
  215. std::istringstream(iter->second) >> this->options_.isUseHcom;
  216. }
  217. this->options_.isUseHvd = false;
  218. iter = options.find(OPTION_EXEC_IS_USEHVD);
  219. if (iter != options.end()) {
  220. std::istringstream(iter->second) >> this->options_.isUseHvd;
  221. }
  222. this->options_.deployMode = false;
  223. iter = options.find(OPTION_EXEC_DEPLOY_MODE);
  224. if (iter != options.end()) {
  225. std::istringstream(iter->second) >> this->options_.deployMode;
  226. }
  227. iter = options.find(OPTION_EXEC_POD_NAME);
  228. if (iter != options.end()) {
  229. this->options_.podName = iter->second.c_str();
  230. }
  231. iter = options.find(OPTION_EXEC_PROFILING_MODE);
  232. if (iter != options.end()) {
  233. this->options_.profiling_mode = iter->second.c_str();
  234. }
  235. iter = options.find(OPTION_EXEC_PROFILING_OPTIONS);
  236. if (iter != options.end()) {
  237. this->options_.profiling_options = iter->second.c_str();
  238. }
  239. iter = options.find(OPTION_EXEC_RANK_ID);
  240. if (iter != options.end()) {
  241. this->options_.rankId = std::strtoll(iter->second.c_str(), nullptr, kDecimal);
  242. }
  243. iter = options.find(OPTION_EXEC_RANK_TABLE_FILE);
  244. if (iter != options.end()) {
  245. this->options_.rankTableFile = iter->second.c_str();
  246. }
  247. this->options_.enable_atomic = true;
  248. iter = options.find(OPTION_EXEC_ATOMIC_FLAG);
  249. GE_IF_BOOL_EXEC(iter != options.end(),
  250. this->options_.enable_atomic = std::strtol(iter->second.c_str(), nullptr, kDecimal));
  251. GELOGI("ge InnerInitialize, the enable_atomic_flag in options_ is %d", this->options_.enable_atomic);
  252. }
  253. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithOptions(Options &options) {
  254. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  255. GELOGI("%s init GELib. session Id:%ld, device id :%d ", mode.c_str(), options.session_id, options.device_id);
  256. GEEVENT("System init with options begin, job id %s", options.job_id.c_str());
  257. std::lock_guard<std::mutex> lock(status_mutex_);
  258. GE_IF_BOOL_EXEC(is_system_inited && !is_shutdown,
  259. GELOGW("System init with options is already inited and not shutdown.");
  260. return SUCCESS);
  261. std::vector<rtMemType_t> mem_type;
  262. mem_type.push_back(RT_MEMORY_HBM);
  263. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  264. if (initMmStatus != SUCCESS) {
  265. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  266. return initMmStatus;
  267. }
  268. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  269. // Update CSA file
  270. CsaInteract::GetInstance().Init(options.device_id, GetContext().TraceId());
  271. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_RUNNING, JOBSUBSTATE_ENV_INIT);
  272. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  273. // set device id
  274. GELOGI("set logical device id:%u", options.device_id);
  275. GetContext().SetCtxDeviceId(static_cast<uint32_t>(options.device_id));
  276. GE_CHK_RT_RET(rtSetDevice(options.device_id));
  277. // In the scenario that the automatic add fusion is set, but there is no cleanaddr operator,
  278. // maybe need to check it
  279. is_system_inited = true;
  280. is_shutdown = false;
  281. GELOGI("%s init GELib success.", mode.c_str());
  282. return SUCCESS;
  283. }
  284. Status GELib::SystemShutdownWithOptions(const Options &options) {
  285. std::string mode = is_train_mode_ ? "Training" : "Online infer";
  286. GELOGI("%s finalize GELib begin.", mode.c_str());
  287. std::lock_guard<std::mutex> lock(status_mutex_);
  288. GE_IF_BOOL_EXEC(is_shutdown || !is_system_inited,
  289. GELOGW("System Shutdown with options is already is_shutdown or system does not inited. "
  290. "is_shutdown:%d is_omm_inited:%d",
  291. is_shutdown, is_system_inited);
  292. return SUCCESS);
  293. GE_CHK_RT(rtDeviceReset(options.device_id));
  294. // Update CSA file
  295. Status ret = CsaInteract::GetInstance().WriteJobState(JOBSTATE_SUCCEED);
  296. GE_LOGE_IF(ret != SUCCESS, "write job state failed, ret:%u", ret);
  297. is_system_inited = false;
  298. is_shutdown = true;
  299. GELOGI("%s finalize GELib success.", mode.c_str());
  300. return SUCCESS;
  301. }
  302. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status GELib::InitSystemWithoutOptions() {
  303. GELOGI("Inference Init GELib begin.");
  304. std::vector<rtMemType_t> mem_type;
  305. mem_type.push_back(RT_MEMORY_HBM);
  306. Status initMmStatus = MemManager::Instance().Initialize(mem_type);
  307. if (initMmStatus != SUCCESS) {
  308. GELOGE(initMmStatus, "[Initialize] MemoryAllocatorManager initialize failed.");
  309. return initMmStatus;
  310. }
  311. GE_CHK_STATUS_RET(HostMemManager::Instance().Initialize());
  312. static bool is_inited = false;
  313. if (is_inited) {
  314. GELOGW("System init without options is already inited, don't need to init again.");
  315. return SUCCESS;
  316. }
  317. is_inited = true;
  318. GELOGI("Inference init GELib success.");
  319. return SUCCESS;
  320. }
  321. string GELib::GetPath() { return PluginManager::GetPath(); }
  322. // Finalize all modules
  323. Status GELib::Finalize() {
  324. GELOGI("finalization start");
  325. // Finalization is not allowed before initialization
  326. if (!init_flag_) {
  327. GELOGW("not initialize");
  328. return SUCCESS;
  329. }
  330. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  331. GE_CHK_RT_RET(rtSetDevice(options_.device_id));
  332. }
  333. Status final_state = SUCCESS;
  334. Status mid_state;
  335. GELOGI("engineManager finalization.");
  336. mid_state = engineManager_.Finalize();
  337. if (mid_state != SUCCESS) {
  338. GELOGW("engineManager finalize failed");
  339. final_state = mid_state;
  340. }
  341. GELOGI("sessionManager finalization.");
  342. mid_state = sessionManager_.Finalize();
  343. if (mid_state != SUCCESS) {
  344. GELOGW("sessionManager finalize failed");
  345. final_state = mid_state;
  346. }
  347. GELOGI("opsManager finalization.");
  348. mid_state = opsManager_.Finalize();
  349. if (mid_state != SUCCESS) {
  350. GELOGW("opsManager finalize failed");
  351. final_state = mid_state;
  352. }
  353. GELOGI("VarManagerPool finalization.");
  354. VarManagerPool::Instance().Destory();
  355. GELOGI("MemManager finalization.");
  356. MemManager::Instance().Finalize();
  357. GELOGI("HostMemManager finalization.");
  358. HostMemManager::Instance().Finalize();
  359. GELOGI("HostCpuEngine finalization.");
  360. HostCpuEngine::GetInstance().Finalize();
  361. GELOGI("Analyzer finalization");
  362. Analyzer::GetInstance()->Finalize();
  363. // Shut down profiling
  364. ShutDownProfiling();
  365. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  366. GELOGI("System ShutDown.");
  367. mid_state = SystemShutdownWithOptions(this->options_);
  368. if (mid_state != SUCCESS) {
  369. GELOGW("System shutdown with options failed");
  370. final_state = mid_state;
  371. }
  372. }
  373. is_train_mode_ = false;
  374. GetMutableGlobalOptions().erase(ENABLE_SINGLE_STREAM);
  375. if (is_train_mode_ || (options_.device_id != kDefaultDeviceIdForInfer)) {
  376. GE_CHK_RT_RET(rtDeviceReset(options_.device_id));
  377. }
  378. instancePtr_ = nullptr;
  379. init_flag_ = false;
  380. if (final_state != SUCCESS) {
  381. GELOGE(FAILED, "finalization failed.");
  382. return final_state;
  383. }
  384. GELOGI("finalization success.");
  385. return SUCCESS;
  386. }
  387. void GELib::ShutDownProfiling() {
  388. std::lock_guard<std::mutex> lock(status_mutex_);
  389. if (!ProfilingManager::Instance().ProfilingOpTraceOn() && ProfilingManager::Instance().ProfilingOn()) {
  390. ProfilingManager::Instance().StopProfiling();
  391. }
  392. if (ProfilingManager::Instance().ProfilingOn()) {
  393. ProfilingManager::Instance().PluginUnInit(GE_PROFILING_MODULE);
  394. }
  395. }
  396. // Get Singleton Instance
  397. std::shared_ptr<GELib> GELib::GetInstance() { return instancePtr_; }
  398. void GELib::RollbackInit() {
  399. if (engineManager_.init_flag_) {
  400. (void)engineManager_.Finalize();
  401. }
  402. if (opsManager_.init_flag_) {
  403. (void)opsManager_.Finalize();
  404. }
  405. if (sessionManager_.init_flag_) {
  406. (void)sessionManager_.Finalize();
  407. }
  408. MemManager::Instance().Finalize();
  409. HostMemManager::Instance().Finalize();
  410. VarManagerPool::Instance().Destory();
  411. }
  412. } // 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