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.

ge_generator.cc 30 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
4 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
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
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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775
  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 "generator/ge_generator.h"
  17. #include <atomic>
  18. #include "common/ge/ge_util.h"
  19. #include "common/ge/plugin_manager.h"
  20. #include "common/helper/model_helper.h"
  21. #include "common/helper/om_file_helper.h"
  22. #include "common/util.h"
  23. #include "common/util/error_manager/error_manager.h"
  24. #include "framework/common/debug/ge_log.h"
  25. #include "framework/common/debug/log.h"
  26. #include "ge/ge_api.h"
  27. #include "graph/debug/ge_attr_define.h"
  28. #include "graph/ge_context.h"
  29. #include "graph/manager/graph_manager.h"
  30. #include "graph/manager/util/rt_context_util.h"
  31. #include "graph/opsproto_manager.h"
  32. #include "graph/utils/graph_utils.h"
  33. #include "graph/utils/type_utils.h"
  34. #include "init/gelib.h"
  35. #include "model/ge_model.h"
  36. using std::map;
  37. using std::string;
  38. using std::vector;
  39. namespace {
  40. const char *const kAttrOpType = "op_type";
  41. const char *const kEngineNameDefault = "default";
  42. const char *const kVectorEngine = "VectorEngine";
  43. const char *const kAIcoreEngine = "AIcoreEngine";
  44. const char *const kFileNameSuffix = "online";
  45. std::map<ge::OpEngineType, std::string> engine_type_map{
  46. {ge::ENGINE_SYS, kEngineNameDefault}, {ge::ENGINE_AICORE, kAIcoreEngine}, {ge::ENGINE_VECTOR, kVectorEngine}};
  47. bool ContainsDynamicInpus(const ge::OpDesc &op_desc) {
  48. for (auto &tensor_desc : op_desc.GetAllInputsDescPtr()) {
  49. if (tensor_desc->MutableShape().IsUnknownShape()) {
  50. GELOGI("Contains unknown shape input. set is_dynamic_input to true.");
  51. return true;
  52. }
  53. }
  54. return false;
  55. }
  56. } // namespace
  57. namespace ge {
  58. static Status CheckEngineTypeSupport(const OpDescPtr &op_desc, OpEngineType engine_type) {
  59. GE_CHECK_NOTNULL_EXEC(op_desc, return PARAM_INVALID);
  60. if (engine_type == ENGINE_SYS) {
  61. GELOGI("CheckEngineType: use default engine.");
  62. return SUCCESS;
  63. }
  64. // get op engine name
  65. string op_engine_name;
  66. auto iter = engine_type_map.find(engine_type);
  67. if (iter != engine_type_map.end()) {
  68. op_engine_name = iter->second;
  69. GELOGI("CheckEngineType: engine type: %d", static_cast<int>(engine_type));
  70. } else {
  71. ErrorManager::GetInstance().ATCReportErrMessage("E14001", {"opname", "optype", "value", "reason"},
  72. {op_desc->GetName(), op_desc->GetType(), "engine type",
  73. "it only support kEngineNameDefault/kAIcoreEngine/kVectorEngine"});
  74. GELOGE(FAILED, "CheckEngineType: engine type: %d not support", static_cast<int>(engine_type));
  75. return FAILED;
  76. }
  77. if (op_desc->HasAttr(ATTR_NAME_UNREGST_OPPATH)) {
  78. op_desc->SetOpEngineName(op_engine_name);
  79. op_desc->SetOpKernelLibName(op_engine_name);
  80. return SUCCESS;
  81. }
  82. // set op engine name and opkernelLib. when engine support
  83. std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
  84. if ((instance_ptr == nullptr) || (!instance_ptr->InitFlag())) {
  85. GELOGE(GE_CLI_GE_NOT_INITIALIZED, "CheckEngineType failed.");
  86. return FAILED;
  87. }
  88. OpsKernelManager &ops_kernel_manager = instance_ptr->OpsKernelManagerObj();
  89. std::vector<OpInfo> op_infos = ops_kernel_manager.GetOpsKernelInfo(op_desc->GetType());
  90. if (op_infos.empty()) {
  91. ErrorManager::GetInstance().ATCReportErrMessage("E14001", {"opname", "optype", "value", "reason"},
  92. {op_desc->GetName(), op_desc->GetType(), "optype", "it can not find"});
  93. GELOGE(FAILED, "CheckEngineType: Can not get op info by op type %s", op_desc->GetType().c_str());
  94. return FAILED;
  95. }
  96. string kernel_name;
  97. for (const auto &it : op_infos) {
  98. if (it.engine == op_engine_name) {
  99. kernel_name = it.opKernelLib;
  100. break;
  101. }
  102. }
  103. if (kernel_name.empty()) {
  104. ErrorManager::GetInstance().ATCReportErrMessage("E14001", {"opname", "optype", "value", "reason"},
  105. {op_desc->GetName(), op_desc->GetType(), "engine name" + FmtToStr(op_engine_name), "it can not find"});
  106. GELOGE(FAILED, "CheckEngineType:Can not find ops kernel, engine name: %s.", op_engine_name.c_str());
  107. return FAILED;
  108. }
  109. auto &kernel_map = ops_kernel_manager.GetAllOpsKernelInfoStores();
  110. auto kernel_info_store = kernel_map.find(kernel_name);
  111. if (kernel_info_store != kernel_map.end()) {
  112. std::string unsupported_reason;
  113. if (kernel_info_store->second->CheckSupported(op_desc, unsupported_reason)) {
  114. op_desc->SetOpEngineName(op_engine_name);
  115. op_desc->SetOpKernelLibName(kernel_name);
  116. GELOGI("CheckEngineType:Set OpKernelLibName %s and engine name %s into op_desc %s", kernel_name.c_str(),
  117. op_engine_name.c_str(), op_desc->GetName().c_str());
  118. return SUCCESS;
  119. } else {
  120. ErrorManager::GetInstance().ATCReportErrMessage(
  121. "E13002", {"optype", "opskernel", "reason"}, {op_desc->GetType(), kernel_name, unsupported_reason});
  122. GELOGE(FAILED, "CheckEngineType: check support failed, Op type %s of ops kernel %s is unsupported, reason:%s",
  123. op_desc->GetType().c_str(), kernel_name.c_str(), unsupported_reason.c_str());
  124. return FAILED;
  125. }
  126. } else {
  127. ErrorManager::GetInstance().ATCReportErrMessage(
  128. "E13003", {"opname", "optype"}, {op_desc->GetName(), op_desc->GetType()});
  129. GELOGE(FAILED,
  130. "CheckEngineType:Can not find any supported ops kernel info store by kernel_name %s,"
  131. "op type is %s, op name is %s",
  132. kernel_name.c_str(), op_desc->GetType().c_str(), op_desc->GetName().c_str());
  133. }
  134. return FAILED;
  135. }
  136. static Status AddInputs(const ComputeGraphPtr &graph, const NodePtr &node, GeTensorDesc &tensor, int32_t index,
  137. bool attr) {
  138. GE_CHECK_NOTNULL_EXEC(graph, return PARAM_INVALID);
  139. GE_CHECK_NOTNULL_EXEC(node, return PARAM_INVALID);
  140. auto format = tensor.GetFormat();
  141. auto data_type = tensor.GetDataType();
  142. if (format == FORMAT_RESERVED && data_type == DT_UNDEFINED) {
  143. return SUCCESS;
  144. }
  145. string op_type;
  146. if (!AttrUtils::GetStr(tensor, kAttrOpType, op_type) || op_type.empty()) {
  147. op_type = DATA;
  148. }
  149. string op_name = node->GetName() + "_in_" + std::to_string(index);
  150. OpDescPtr data_op = MakeShared<ge::OpDesc>(op_name, op_type);
  151. if (data_op == nullptr) {
  152. return FAILED;
  153. }
  154. (void)AttrUtils::SetBool(data_op, "_is_single_op", true);
  155. GE_CHK_BOOL_EXEC(data_op->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add input desc fail.");
  156. GE_CHK_BOOL_EXEC(data_op->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add output desc fail.");
  157. if (attr) {
  158. GE_CHK_BOOL_EXEC(AttrUtils::SetInt(data_op, ATTR_NAME_INDEX, index), return FAILED, "Set index fail.");
  159. }
  160. ge::NodePtr arg_node = graph->AddNode(data_op);
  161. GE_CHK_BOOL_EXEC(arg_node != nullptr, return FAILED, "Insert Data node fail.");
  162. GE_CHK_STATUS(GraphUtils::AddEdge(arg_node->GetOutDataAnchor(0), node->GetInDataAnchor(index)),
  163. "Add edge[%s->%s] fail.", data_op->GetName().c_str(), node->GetName().c_str());
  164. return SUCCESS;
  165. }
  166. static Status AddOutputs(const ComputeGraphPtr &graph, const NodePtr &node, const vector<GeTensor> &outputs) {
  167. OpDescPtr op_desc = MakeShared<ge::OpDesc>(graph->GetName() + "_" + NODE_NAME_NET_OUTPUT, NETOUTPUT);
  168. if (op_desc == nullptr) {
  169. return FAILED;
  170. }
  171. (void)AttrUtils::SetBool(op_desc, "_is_single_op", true);
  172. int32_t count = 0;
  173. for (const auto &out_desc : outputs) {
  174. GeTensorDesc tensor = out_desc.GetTensorDesc();
  175. TensorUtils::SetInputTensor(tensor, true);
  176. GE_CHK_BOOL_EXEC(op_desc->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add input desc fail");
  177. TensorUtils::SetInputTensor(tensor, false);
  178. TensorUtils::SetOutputTensor(tensor, true);
  179. GE_CHK_BOOL_EXEC(op_desc->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add output desc fail");
  180. count++;
  181. }
  182. GE_CHECK_NOTNULL_EXEC(graph, return PARAM_INVALID);
  183. ge::NodePtr out_node = graph->AddNode(op_desc);
  184. GE_CHK_BOOL_EXEC(out_node != nullptr, return FAILED, "Insert Output node fail.");
  185. GE_CHECK_NOTNULL_EXEC(node, return PARAM_INVALID);
  186. for (int32_t i = 0; i < count; ++i) {
  187. GE_CHK_STATUS(GraphUtils::AddEdge(node->GetOutDataAnchor(i), out_node->GetInDataAnchor(i)),
  188. "Add edge[%s->%s] fail.", node->GetName().c_str(), out_node->GetName().c_str());
  189. }
  190. return SUCCESS;
  191. }
  192. static void GetOpsProtoPath(string &opsproto_path) {
  193. GELOGI("Start to get ops proto path schedule.");
  194. const char *path_env = std::getenv("ASCEND_OPP_PATH");
  195. if (path_env != nullptr) {
  196. string path = path_env;
  197. string file_path = RealPath(path.c_str());
  198. if (file_path.empty()) {
  199. GELOGE(FAILED, "File path %s is invalid.", path.c_str());
  200. return;
  201. }
  202. opsproto_path = (path + "/op_proto/custom/" + ":") + (path + "/op_proto/built-in/");
  203. GELOGI("Get opsproto so path from env : %s", path.c_str());
  204. return;
  205. }
  206. string path_base = PluginManager::GetPath();
  207. GELOGI("path_base is %s", path_base.c_str());
  208. path_base = path_base.substr(0, path_base.rfind('/'));
  209. path_base = path_base.substr(0, path_base.rfind('/') + 1);
  210. opsproto_path = (path_base + "ops/op_proto/custom/" + ":") + (path_base + "ops/op_proto/built-in/");
  211. }
  212. class GeGenerator::Impl {
  213. public:
  214. Impl(OmgContext &omg_context) : omg_context_(omg_context) {}
  215. ~Impl() = default;
  216. Status BuildModel(const Graph &graph, const vector<GeTensor> &inputs, GeRootModelPtr &ge_models);
  217. Status SaveModel(const string &file_name_prefix, GeModelPtr &models, ModelBufferData &model);
  218. Status SaveParams(GeModelPtr &ge_model, const string &type, const map<string, GeAttrValue> &attrs,
  219. const vector<GeTensor> &inputs, const vector<GeTensor> &outputs);
  220. Status GenerateInfershapeGraph(const Graph &graph);
  221. OmgContext &omg_context_;
  222. GraphManager graph_manager_;
  223. SaveParam save_param_;
  224. bool is_offline_ = true;
  225. bool is_singleop_unregistered_ = false;
  226. std::string build_mode_;
  227. std::string build_step_;
  228. static std::mutex mutex_;
  229. private:
  230. static std::string Trim(const std::string &str);
  231. bool ParseVersion(const std::string &line, std::string &version);
  232. bool GetVersionFromPath(const std::string &file_path, std::string &version);
  233. bool SetAtcVersionInfo(AttrHolder &obj);
  234. bool SetOppVersionInfo(AttrHolder &obj);
  235. };
  236. Status GeGenerator::Initialize(const map<string, string> &options) {
  237. return Initialize(options, domi::GetContext());
  238. }
  239. Status GeGenerator::Initialize(const map<string, string> &options, OmgContext &omg_context) {
  240. impl_ = ge::MakeShared<Impl>(omg_context);
  241. if (impl_ == nullptr) {
  242. GELOGE(MEMALLOC_FAILED, "Make shared failed");
  243. return MEMALLOC_FAILED;
  244. }
  245. string opsproto_path;
  246. GetOpsProtoPath(opsproto_path);
  247. GELOGI("Get opsproto path is %s", opsproto_path.c_str());
  248. OpsProtoManager *manager = OpsProtoManager::Instance();
  249. map<string, string> option_tmp;
  250. option_tmp.emplace(std::pair<string, string>(string("ge.opsProtoLibPath"), opsproto_path));
  251. (void)manager->Initialize(option_tmp);
  252. Status ret = impl_->graph_manager_.Initialize(options);
  253. if (ret != SUCCESS) {
  254. GELOGE(GE_GENERATOR_GRAPH_MANAGER_INIT_FAILED, "Graph manager initialize failed.");
  255. return GE_GENERATOR_GRAPH_MANAGER_INIT_FAILED;
  256. }
  257. // get ek file
  258. auto iter = options.find(EK_FILE);
  259. if (iter != options.end()) {
  260. impl_->save_param_.ek_file = iter->second;
  261. }
  262. // get cert file
  263. iter = options.find(CERT_FILE);
  264. if (iter != options.end()) {
  265. impl_->save_param_.cert_file = iter->second;
  266. }
  267. // get hw key file
  268. iter = options.find(HW_KEY_FILE);
  269. if (iter != options.end()) {
  270. impl_->save_param_.hw_key_file = iter->second;
  271. }
  272. // get private file
  273. iter = options.find(PRIVATE_KEY_FILE);
  274. if (iter != options.end()) {
  275. impl_->save_param_.pri_key_file = iter->second;
  276. }
  277. // get build mode
  278. iter = options.find(BUILD_MODE);
  279. if (iter != options.end()) {
  280. impl_->build_mode_ = iter->second;
  281. }
  282. // get build step
  283. iter = options.find(BUILD_STEP);
  284. if (iter != options.end()) {
  285. impl_->build_step_ = iter->second;
  286. }
  287. return SUCCESS;
  288. }
  289. Status GeGenerator::Finalize() {
  290. GE_CHECK_NOTNULL_EXEC(impl_, return PARAM_INVALID);
  291. Status ret = impl_->graph_manager_.Finalize();
  292. if (ret != SUCCESS) {
  293. GELOGE(GE_GENERATOR_GRAPH_MANAGER_FINALIZE_FAILED, "Graph manager finalize failed.");
  294. return GE_GENERATOR_GRAPH_MANAGER_FINALIZE_FAILED;
  295. }
  296. return SUCCESS;
  297. }
  298. Status GeGenerator::GenerateOfflineModel(const Graph &graph, const string &file_name_prefix,
  299. const vector<GeTensor> &inputs) {
  300. GELOGI("Start to generate offline model.");
  301. ModelBufferData model;
  302. return GenerateModel(graph, file_name_prefix, inputs, model, true);
  303. }
  304. Status GeGenerator::GenerateOnlineModel(const Graph &graph, const vector<GeTensor> &inputs, ModelBufferData &model) {
  305. return GenerateModel(graph, "online", inputs, model, false);
  306. }
  307. Status GeGenerator::GenerateInfershapeGraph(const Graph &graph) {
  308. GE_CHECK_NOTNULL_EXEC(impl_, return PARAM_INVALID);
  309. Status ret = impl_->GenerateInfershapeGraph(graph);
  310. if (ret != SUCCESS) {
  311. GELOGE(ret, "Dump infershape json failed");
  312. if (impl_->graph_manager_.Finalize() != SUCCESS) {
  313. GELOGE(FAILED, "graph_manager finalize fail.");
  314. }
  315. return ret;
  316. }
  317. GELOGI("Generate infer shape graph success");
  318. return SUCCESS;
  319. }
  320. std::mutex GeGenerator::Impl::mutex_;
  321. // Remove the space and tab before and after the string
  322. std::string GeGenerator::Impl::Trim(const std::string &str) {
  323. if (str.empty()) {
  324. return str;
  325. }
  326. std::string::size_type start = str.find_first_not_of(" \t\r\n");
  327. if (start == std::string::npos) {
  328. return str;
  329. }
  330. std::string::size_type end = str.find_last_not_of(" \t\r\n") + 1;
  331. return str.substr(start, end);
  332. }
  333. // Parsing the command line
  334. bool GeGenerator::Impl::ParseVersion(const std::string &line, std::string &version) {
  335. std::string flag = "Version=";
  336. std::string temp = Trim(line);
  337. if (temp.empty()) {
  338. GELOGW("line is empty.");
  339. return false;
  340. }
  341. std::string::size_type pos = temp.find(flag);
  342. if (pos == std::string::npos) {
  343. GELOGW("Incorrect line [%s], it must include [%s].", line.c_str(), flag.c_str());
  344. return false;
  345. }
  346. if (temp.size() == flag.size()) {
  347. GELOGW("version information is empty. %s", line.c_str());
  348. return false;
  349. }
  350. version = temp.substr(pos + flag.size());
  351. GELOGI("Version=%s", version.c_str());
  352. return true;
  353. }
  354. bool GeGenerator::Impl::GetVersionFromPath(const std::string &file_path, std::string &version) {
  355. // Normalize the path
  356. string resolved_file_path = RealPath(file_path.c_str());
  357. if (resolved_file_path.empty()) {
  358. GELOGW("Invalid input file path [%s], make sure that the file path is correct.", file_path.c_str());
  359. return false;
  360. }
  361. std::ifstream fs(resolved_file_path, std::ifstream::in);
  362. if (!fs.is_open()) {
  363. GELOGW("Open %s failed.", file_path.c_str());
  364. return false;
  365. }
  366. std::string line;
  367. if (getline(fs, line)) {
  368. if (!ParseVersion(line, version)) {
  369. GELOGW("Parse version failed. content is [%s].", line.c_str());
  370. fs.close();
  371. return false;
  372. }
  373. } else {
  374. GELOGW("No version information found in the file path:%s", file_path.c_str());
  375. fs.close();
  376. return false;
  377. }
  378. fs.close(); // close the file
  379. return true;
  380. }
  381. // Set package version information in the model
  382. bool GeGenerator::Impl::SetAtcVersionInfo(AttrHolder &obj) {
  383. std::string path_base = ge::GELib::GetPath();
  384. path_base = path_base.substr(0, path_base.rfind('/'));
  385. path_base = path_base.substr(0, path_base.rfind('/') + 1);
  386. std::string version_path = path_base + "version.info";
  387. GELOGI("version_path is %s", version_path.c_str());
  388. std::string version;
  389. if (!GetVersionFromPath(version_path, version)) {
  390. GELOGW("Get atc version information failed!");
  391. return false;
  392. }
  393. // set version info
  394. if (!ge::AttrUtils::SetStr(obj, ATTR_MODEL_ATC_VERSION, version)) {
  395. GELOGW("Ge model set atc version failed!");
  396. return false;
  397. }
  398. GELOGI("Ge model set atc version information success.");
  399. return true;
  400. }
  401. // Set package version information in the model
  402. bool GeGenerator::Impl::SetOppVersionInfo(AttrHolder &obj) {
  403. const char *path_env = std::getenv("ASCEND_OPP_PATH");
  404. if (path_env == nullptr) {
  405. GELOGW("Get environment variable ASCEND_OPP_PATH failed!");
  406. return false;
  407. }
  408. std::string version_path = path_env;
  409. version_path += "/version.info";
  410. GELOGI("version_path is %s", version_path.c_str());
  411. std::string version;
  412. if (!GetVersionFromPath(version_path, version)) {
  413. GELOGW("Get opp version information failed!");
  414. return false;
  415. }
  416. // set version info
  417. if (!ge::AttrUtils::SetStr(obj, ATTR_MODEL_OPP_VERSION, version)) {
  418. GELOGW("Ge model set opp version failed!");
  419. return false;
  420. }
  421. GELOGI("Ge Model set opp version information success.");
  422. return true;
  423. }
  424. Status GeGenerator::GenerateModel(const Graph &graph, const string &file_name_prefix, const vector<GeTensor> &inputs,
  425. ModelBufferData &model, bool is_offline) {
  426. rtContext_t ctx = nullptr;
  427. auto rt = rtCtxGetCurrent(&ctx);
  428. if (rt != RT_ERROR_NONE) {
  429. GELOGW("Current ctx is null.");
  430. ctx = nullptr;
  431. }
  432. GeRootModelPtr ge_root_model = nullptr;
  433. GE_CHECK_NOTNULL_EXEC(impl_, return PARAM_INVALID);
  434. impl_->is_offline_ = is_offline;
  435. Status ret = impl_->BuildModel(graph, inputs, ge_root_model);
  436. if (ret != SUCCESS) {
  437. GELOGE(ret, "Build model failed.");
  438. if (impl_->graph_manager_.Finalize() != SUCCESS) {
  439. GELOGE(FAILED, "graph_manager finalize fail.");
  440. }
  441. return ret;
  442. }
  443. /// BUILD_MODE_TUNING with BUILD_STEP_BEFORE_UB_MATCH no need save model;
  444. /// BUILD_MODE_TUNING with BUILD_STEP_AFTER_BUILDER no need save model;
  445. /// BUILD_MODE_TUNING with BUILD_STEP_AFTER_BUILDER_SUB no need save model.
  446. if ((impl_->build_mode_ == BUILD_MODE_TUNING) &&
  447. (impl_->build_step_ == BUILD_STEP_BEFORE_UB_MATCH || impl_->build_step_ == BUILD_STEP_AFTER_BUILDER ||
  448. impl_->build_step_ == BUILD_STEP_AFTER_BUILDER_SUB)) {
  449. GELOGI("Build mode:%s with step:%s no need SaveModel.",
  450. impl_->build_mode_.c_str(),
  451. impl_->build_step_.c_str());
  452. return SUCCESS;
  453. }
  454. GE_CHECK_NOTNULL(ge_root_model);
  455. GE_CHECK_NOTNULL(ge_root_model->GetRootGraph());
  456. ModelHelper model_helper;
  457. string model_name = "";
  458. Status name_ret = model_helper.GetModelNameFromMergedGraphName(ge_root_model->GetRootGraph()->GetName(), model_name);
  459. if (name_ret != SUCCESS) {
  460. ErrorManager::GetInstance().ATCReportErrMessage("E10000", {"parameter"}, {"output"});
  461. GELOGE(FAILED, "Get model_name failed. Param --output is invalid");
  462. return PARAM_INVALID;
  463. }
  464. map<string, GeModelPtr> name_to_ge_model = ge_root_model->GetSubgraphInstanceNameToModel();
  465. GeModelPtr &ge_model = name_to_ge_model[ge_root_model->GetRootGraph()->GetName()];
  466. GE_RETURN_WITH_LOG_IF_FALSE(ge_model != nullptr, "ge_model can not be null");
  467. ge_model->SetName(model_name);
  468. ret = impl_->SaveModel(file_name_prefix, ge_model, model);
  469. if (ret != SUCCESS) {
  470. GELOGE(ret, "Save model failed");
  471. if (impl_->graph_manager_.Finalize() != SUCCESS) {
  472. GELOGE(FAILED, "graph_manager finalize fail.");
  473. }
  474. return ret;
  475. }
  476. if (ctx != nullptr) {
  477. (void)rtCtxSetCurrent(ctx);
  478. }
  479. GELOGI("GenerateOfflineModel success.");
  480. return SUCCESS;
  481. }
  482. namespace {
  483. bool IsNeedConnectInputOpForSingleOp(GeTensorDesc &tensor_desc) {
  484. bool is_need = true;
  485. // format and dtype is all reserved, stand for Optional input. When singleop scene
  486. if (tensor_desc.GetFormat() == FORMAT_RESERVED && tensor_desc.GetDataType() == DT_UNDEFINED) {
  487. is_need = false;
  488. }
  489. return is_need;
  490. }
  491. }
  492. Status GeGenerator::CheckForSingleOp(OpDescPtr &op_desc, const vector<GeTensor> &inputs,
  493. const vector<GeTensor> &outputs) {
  494. GE_CHECK_NOTNULL_EXEC(op_desc, return PARAM_INVALID);
  495. if (!inputs.empty() && (inputs.size() != op_desc->GetAllInputsSize())) {
  496. ErrorManager::GetInstance().ATCReportErrMessage("E14001", {"opname", "optype", "value", "reason"},
  497. {op_desc->GetName(), op_desc->GetType(), "inputs size" + FmtToStr(inputs.size()),
  498. "tensor size is " + FmtToStr(op_desc->GetAllInputsSize())});
  499. GELOGE(PARAM_INVALID, "Tensor size: %zu, Inputs size: %zu", inputs.size(), op_desc->GetAllInputsSize());
  500. return PARAM_INVALID;
  501. }
  502. if (!outputs.empty() && (outputs.size() != op_desc->GetOutputsSize())) {
  503. ErrorManager::GetInstance().ATCReportErrMessage("E14001", {"opname", "optype", "value", "reason"},
  504. {op_desc->GetName(), op_desc->GetType(), "outputs size" + FmtToStr(outputs.size()),
  505. "tensor size is " + FmtToStr(op_desc->GetOutputsSize())});
  506. GELOGE(PARAM_INVALID, "Tensor size: %zu, Outputs size: %zu", outputs.size(), op_desc->GetOutputsSize());
  507. return PARAM_INVALID;
  508. }
  509. return SUCCESS;
  510. }
  511. Status GeGenerator::BuildSingleOp(OpDescPtr &op_desc, const vector<GeTensor> &inputs, const vector<GeTensor> &outputs,
  512. const string &model_file_name, OpEngineType engine_type, ModelBufferData &model_buff,
  513. bool is_offline) {
  514. if (CheckForSingleOp(op_desc, inputs, outputs) != SUCCESS) {
  515. GELOGE(PARAM_INVALID, "input param is invalid when build single op!");
  516. return PARAM_INVALID;
  517. }
  518. OmgContext &omg_context = (impl_ == nullptr) ? domi::GetContext() : impl_->omg_context_;
  519. omg_context.is_dynamic_input = ContainsDynamicInpus(*op_desc);
  520. if (op_desc->HasAttr(ATTR_NAME_UNREGST_OPPATH)) {
  521. impl_->is_singleop_unregistered_ = true;
  522. }
  523. // 0. Save original attributes.
  524. OpDescPtr op_desc_tmp = AttrUtils::CloneOpDesc(op_desc);
  525. GE_CHECK_NOTNULL(op_desc_tmp);
  526. // 1. check engine type when compile online
  527. if (model_file_name == kFileNameSuffix) {
  528. Status ret = CheckEngineTypeSupport(op_desc, engine_type);
  529. if (ret != SUCCESS) {
  530. GELOGE(ret, "check engine type failed.");
  531. return ret;
  532. }
  533. }
  534. // 2. Create ComputeGraph.
  535. string name = ge::CurrentTimeInStr() + "_" + model_file_name;
  536. ge::ComputeGraphPtr compute_graph = MakeShared<ComputeGraph>(name);
  537. GE_CHECK_NOTNULL_EXEC(compute_graph, return INTERNAL_ERROR);
  538. // 3. Add Node to ComputeGraph.
  539. NodePtr op_node = compute_graph->AddNode(op_desc);
  540. GE_CHECK_NOTNULL_EXEC(op_node, return INTERNAL_ERROR);
  541. // 4. Create InputData node.
  542. int32_t arg_index = 0;
  543. if (inputs.empty()) {
  544. for (const auto &input_desc : op_desc->GetAllInputsDescPtr()) {
  545. GE_CHECK_NOTNULL_EXEC(input_desc, return INTERNAL_ERROR);
  546. if (!IsNeedConnectInputOpForSingleOp(*input_desc)) {
  547. continue;
  548. }
  549. GE_CHK_STATUS_RET_NOLOG(AddInputs(compute_graph, op_node, *input_desc, arg_index, false));
  550. arg_index++;
  551. }
  552. } else {
  553. for (const auto &in_desc : inputs) {
  554. GeTensorDesc input_desc = in_desc.GetTensorDesc();
  555. GE_CHK_STATUS_RET_NOLOG(AddInputs(compute_graph, op_node, input_desc, arg_index, true));
  556. arg_index++;
  557. }
  558. }
  559. // 5. Create Output node.
  560. if (!outputs.empty()) {
  561. GE_CHK_STATUS_RET_NOLOG(AddOutputs(compute_graph, op_node, outputs));
  562. }
  563. // dump ComputeGraph.
  564. compute_graph->Dump();
  565. Graph graph = ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph);
  566. GELOGI("ATC parser success in single op build.");
  567. GeRootModelPtr ge_root_model = nullptr;
  568. GE_CHECK_NOTNULL_EXEC(impl_, return PARAM_INVALID);
  569. impl_->is_offline_ = is_offline;
  570. GE_CHK_STATUS_RET_NOLOG(impl_->BuildModel(graph, inputs, ge_root_model));
  571. map<string, GeAttrValue> op_attrs = op_desc_tmp->GetAllAttrs();
  572. GE_CHECK_NOTNULL(ge_root_model);
  573. GE_CHECK_NOTNULL(ge_root_model->GetRootGraph());
  574. map<string, GeModelPtr> name_to_ge_model = ge_root_model->GetSubgraphInstanceNameToModel();
  575. if (name_to_ge_model.empty()) {
  576. GELOGE(PARAM_INVALID, "GetSubgraphInstanceNameToModel is empty.");
  577. return PARAM_INVALID;
  578. }
  579. GeModelPtr &ge_model = name_to_ge_model.begin()->second;
  580. GELOGD("The opType in op_desc_tmp is [%s]", op_desc_tmp->GetType().c_str());
  581. GE_CHK_STATUS_RET_NOLOG(impl_->SaveParams(ge_model, op_desc_tmp->GetType(), op_attrs, inputs, outputs));
  582. GE_CHK_STATUS_RET_NOLOG(impl_->SaveModel(model_file_name, ge_model, model_buff));
  583. return SUCCESS;
  584. }
  585. /**
  586. * @ingroup ge
  587. * @brief Compiling a single operator into an offline model
  588. * @param [in] OpDescPtr &op_desc: Operator description info that needs to be compiled into an offline model file
  589. * @param [in] vector<GeTensor> &inputs: Operator input data description information.
  590. * @param [in] vector<GeTensor> &outputs: Operator output data description information.
  591. * @param [in] const string &model_file_name: Offline model filename.
  592. * @return SUCCESS handle successfully / others handle failed
  593. */
  594. Status GeGenerator::BuildSingleOpModel(OpDescPtr &op_desc, const vector<GeTensor> &inputs,
  595. const vector<GeTensor> &outputs, const string &model_file_name) {
  596. GELOGI("Start to build single op offline model.");
  597. ModelBufferData model_buff;
  598. OpEngineType engine_type = ENGINE_SYS;
  599. return BuildSingleOp(op_desc, inputs, outputs, model_file_name, engine_type, model_buff, true);
  600. }
  601. /**
  602. * @ingroup ge
  603. * @brief Compiling a single operator into online buffer
  604. * @param [in] OpDescPtr &op_desc: Operator description info that needs to be compiled into an offline model file
  605. * @param [in] vector<GeTensor> &inputs: Operator input data description information.
  606. * @param [in] vector<GeTensor> &outputs: Operator output data description information.
  607. * @param [in] engine_type: specific engine.
  608. * @param [out] ModelBufferData &Model_buff: Model_buff: model buffer of the op.
  609. * @return SUCCESS handle successfully / others handle failed
  610. */
  611. Status GeGenerator::BuildSingleOpModel(OpDescPtr &op_desc, const vector<GeTensor> &inputs,
  612. const vector<GeTensor> &outputs, OpEngineType engine_type,
  613. ModelBufferData &model_buff) {
  614. GELOGI("Start to build single op online");
  615. return BuildSingleOp(op_desc, inputs, outputs, kFileNameSuffix, engine_type, model_buff, false);
  616. }
  617. Status GeGenerator::Impl::SaveParams(GeModelPtr &ge_model, const string &type, const map<string, GeAttrValue> &attrs,
  618. const vector<GeTensor> &inputs, const vector<GeTensor> &outputs) {
  619. GE_CHECK_NOTNULL_EXEC(ge_model, return PARAM_INVALID);
  620. GE_CHK_BOOL_EXEC_NOLOG(graph_manager_.SaveParams(*ge_model, type, attrs, inputs, outputs) == SUCCESS,
  621. (void)graph_manager_.Finalize();
  622. return FAILED);
  623. return SUCCESS;
  624. }
  625. Status GeGenerator::Impl::SaveModel(const string &file_name_prefix, GeModelPtr &model, ModelBufferData &model_buff) {
  626. // set atc version
  627. if (!SetAtcVersionInfo(*(model.get()))) {
  628. GELOGW("SetPackageVersionInfo of atc failed!");
  629. }
  630. // set opp version
  631. if (!SetOppVersionInfo(*(model.get()))) {
  632. GELOGW("SetPackageVersionInfo of ops failed!");
  633. }
  634. ModelHelper model_helper;
  635. model_helper.SetSaveMode(is_offline_);
  636. Status ret = model_helper.SaveToOmModel(model, save_param_, file_name_prefix, model_buff);
  637. if (ret != SUCCESS) {
  638. GELOGE(ret, "Save to om model failed");
  639. return ret;
  640. }
  641. return SUCCESS;
  642. }
  643. Status GeGenerator::Impl::BuildModel(const Graph &graph, const vector<GeTensor> &inputs,
  644. GeRootModelPtr &ge_root_model) {
  645. static std::atomic<GraphId> atomic_graph_id(0);
  646. auto graph_id = atomic_graph_id.fetch_add(1);
  647. const std::map<std::string, std::string> options;
  648. Status ret = graph_manager_.AddGraph(graph_id, graph, options, omg_context_);
  649. if (ret != SUCCESS) {
  650. GELOGE(GE_GENERATOR_GRAPH_MANAGER_ADD_GRAPH_FAILED, "GraphManager add graph fail, graph id: %u", graph_id);
  651. (void)graph_manager_.Finalize();
  652. return GE_GENERATOR_GRAPH_MANAGER_ADD_GRAPH_FAILED;
  653. }
  654. GELOGI("Model inputs size is %zu", inputs.size());
  655. graph_manager_.SetOptionsRunGraphFlag(false);
  656. static std::atomic<uint64_t> atomic_session_id(0);
  657. auto session_id = atomic_session_id.fetch_add(1);
  658. if (is_singleop_unregistered_) {
  659. ret = graph_manager_.BuildGraphForUnregisteredOp(graph_id, inputs, ge_root_model, session_id);
  660. } else {
  661. ret = graph_manager_.BuildGraph(graph_id, inputs, ge_root_model, session_id);
  662. }
  663. if (ret != SUCCESS) {
  664. GELOGE(GE_GENERATOR_GRAPH_MANAGER_BUILD_GRAPH_FAILED, "GraphManager build graph fail, graph id: %u", graph_id);
  665. VarManagerPool::Instance().RemoveVarManager(session_id);
  666. return GE_GENERATOR_GRAPH_MANAGER_BUILD_GRAPH_FAILED;
  667. }
  668. VarManagerPool::Instance().RemoveVarManager(session_id);
  669. return SUCCESS;
  670. }
  671. Status GeGenerator::Impl::GenerateInfershapeGraph(const Graph &graph) {
  672. static std::atomic<GraphId> atomic_graph_id(0);
  673. auto graph_id = atomic_graph_id.fetch_add(1);
  674. const std::map<std::string, std::string> options;
  675. Status ret = graph_manager_.AddGraph(graph_id, graph, options, omg_context_);
  676. if (ret != SUCCESS) {
  677. GELOGE(GE_GENERATOR_GRAPH_MANAGER_ADD_GRAPH_FAILED, "GraphManager add graph failed, graph id: %u", graph_id);
  678. (void)graph_manager_.Finalize();
  679. return GE_GENERATOR_GRAPH_MANAGER_ADD_GRAPH_FAILED;
  680. }
  681. ret = graph_manager_.GenerateInfershapeGraph(graph_id);
  682. if (ret != SUCCESS) {
  683. GELOGE(GE_GENERATOR_GRAPH_MANAGER_BUILD_GRAPH_FAILED, "GraphManager generate graph failed");
  684. return GE_GENERATOR_GRAPH_MANAGER_BUILD_GRAPH_FAILED;
  685. }
  686. return SUCCESS;
  687. }
  688. } // 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