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.

graph_preprocess.cc 74 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
4 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
4 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
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
4 years ago
4 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
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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863
  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 "graph/preprocess/graph_preprocess.h"
  17. #include <map>
  18. #include <set>
  19. #include <string>
  20. #include <utility>
  21. #include "common/formats/format_transfers/format_transfer_fractal_nz.h"
  22. #include "common/formats/format_transfers/format_transfer_fractal_z.h"
  23. #include "common/formats/format_transfers/format_transfer_nchw_nc1hwc0.h"
  24. #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h"
  25. #include "common/formats/format_transfers/format_transfer_transpose.h"
  26. #include "common/formats/utils/formats_trans_utils.h"
  27. #include "common/helper/model_helper.h"
  28. #include "common/math/math_util.h"
  29. #include "common/op/ge_op_utils.h"
  30. #include "common/util/error_manager/error_manager.h"
  31. #include "common/formats/utils/formats_trans_utils.h"
  32. #include "framework/common/debug/ge_log.h"
  33. #include "graph/common/ge_call_wrapper.h"
  34. #include "graph/common/local_context.h"
  35. #include "graph/common/transop_util.h"
  36. #include "graph/debug/ge_attr_define.h"
  37. #include "graph/ge_context.h"
  38. #include "graph/shape_refiner.h"
  39. #include "graph/manager/graph_var_manager.h"
  40. #include "graph/manager/util/rt_context_util.h"
  41. #include "graph/optimize/graph_optimize.h"
  42. #include "graph/passes/addn_pass.h"
  43. #include "graph/passes/aicpu_constant_folding_pass.h"
  44. #include "graph/passes/assert_pass.h"
  45. #include "graph/passes/assign_pass.h"
  46. #include "graph/passes/base_pass.h"
  47. #include "graph/passes/common_subexpression_elimination_pass.h"
  48. #include "graph/passes/cond_pass.h"
  49. #include "graph/passes/cond_remove_pass.h"
  50. #include "graph/passes/constant_folding_pass.h"
  51. #include "graph/passes/constant_fuse_same_pass.h"
  52. #include "graph/passes/control_trigger_pass.h"
  53. #include "graph/passes/dimension_adjust_pass.h"
  54. #include "graph/passes/dimension_compute_pass.h"
  55. #include "graph/passes/dropout_pass.h"
  56. #include "graph/passes/enter_pass.h"
  57. #include "graph/passes/flow_ctrl_pass.h"
  58. #include "graph/passes/for_pass.h"
  59. #include "graph/passes/get_original_format_pass.h"
  60. #include "graph/passes/guarantee_const_pass.h"
  61. #include "graph/passes/hccl_group_pass.h"
  62. #include "graph/passes/identity_pass.h"
  63. #include "graph/passes/infershape_pass.h"
  64. #include "graph/passes/iterator_op_pass.h"
  65. #include "graph/passes/merge_pass.h"
  66. #include "graph/passes/net_output_pass.h"
  67. #include "graph/passes/next_iteration_pass.h"
  68. #include "graph/passes/no_use_reshape_remove_pass.h"
  69. #include "graph/passes/parallel_concat_start_op_pass.h"
  70. #include "graph/passes/placeholder_with_default_pass.h"
  71. #include "graph/passes/prevent_gradient_pass.h"
  72. #include "graph/passes/print_op_pass.h"
  73. #include "graph/passes/prune_pass.h"
  74. #include "graph/passes/replace_transshape_pass.h"
  75. #include "graph/passes/replace_with_empty_const_pass.h"
  76. #include "graph/passes/resource_pair_add_control_pass.h"
  77. #include "graph/passes/resource_pair_remove_control_pass.h"
  78. #include "graph/passes/save_pass.h"
  79. #include "graph/passes/shape_operate_op_remove_pass.h"
  80. #include "graph/passes/snapshot_pass.h"
  81. #include "graph/passes/stop_gradient_pass.h"
  82. #include "graph/passes/subgraph_pass.h"
  83. #include "graph/passes/switch_data_edges_bypass.h"
  84. #include "graph/passes/switch_dead_branch_elimination.h"
  85. #include "graph/passes/switch_logic_remove_pass.h"
  86. #include "graph/passes/merge_to_stream_merge_pass.h"
  87. #include "graph/passes/switch_to_stream_switch_pass.h"
  88. #include "graph/passes/attach_stream_label_pass.h"
  89. #include "graph/passes/unused_const_pass.h"
  90. #include "graph/passes/unused_op_remove_pass.h"
  91. #include "graph/passes/var_is_initialized_op_pass.h"
  92. #include "graph/passes/variable_prepare_op_pass.h"
  93. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  94. #include "graph/types.h"
  95. #include "graph/utils/tensor_utils.h"
  96. #include "graph/utils/type_utils.h"
  97. #include "inc/pass_manager.h"
  98. #include "init/gelib.h"
  99. #include "multi_batch_copy_graph.h"
  100. #include "runtime/dev.h"
  101. #include "graph/passes/dimension_adjust_pass.h"
  102. #include "graph/passes/link_gen_mask_nodes_pass.h"
  103. #include "graph/passes/permute_pass.h"
  104. #include "graph/passes/reshape_remove_pass.h"
  105. #include "graph/passes/same_transdata_breadth_fusion_pass.h"
  106. #include "graph/passes/transop_breadth_fusion_pass.h"
  107. #include "graph/passes/transop_depth_fusion_pass.h"
  108. #include "graph/passes/transop_nearby_allreduce_fusion_pass.h"
  109. #include "graph/passes/cast_remove_pass.h"
  110. #include "graph/passes/data_pass.h"
  111. #include "graph/passes/transop_without_reshape_fusion_pass.h"
  112. #include "graph/passes/transpose_transdata_pass.h"
  113. #include "graph/passes/variable_op_pass.h"
  114. #include "graph/passes/variable_prepare_op_pass.h"
  115. #include "graph/passes/variable_ref_delete_op_pass.h"
  116. namespace ge {
  117. namespace {
  118. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  119. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  120. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  121. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  122. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  123. // the size of user defined output datatype or format string after split by ":".
  124. const size_t kUserDefinedElementCount = 2;
  125. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  126. GeTensorPtr tensor = MakeShared<GeTensor>();
  127. if (tensor == nullptr) {
  128. GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed");
  129. return nullptr;
  130. }
  131. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  132. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  133. auto dst_ge_shape = data_tensor.GetShape();
  134. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  135. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  136. tensor->MutableTensorDesc().SetShape(GeShape());
  137. int32_t dst_shape = 1;
  138. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  139. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  140. return nullptr;
  141. }
  142. } else {
  143. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  144. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  145. if (dst_shape == nullptr) {
  146. GELOGE(INTERNAL_ERROR, "Create unique ptr failed");
  147. return nullptr;
  148. }
  149. for (int64_t i = 0; i < dim_cnt; ++i) {
  150. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  151. }
  152. GE_IF_BOOL_EXEC(
  153. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  154. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  155. return nullptr;)
  156. }
  157. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  158. return OpDescUtils::CreateConstOp(tensor);
  159. }
  160. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  161. OpDescPtr &op_desc) {
  162. // For format transfer node, the IR definition has src/dst format attrs
  163. if (node_type == TRANSDATA) {
  164. GE_IF_BOOL_EXEC(
  165. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  166. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  167. GE_IF_BOOL_EXEC(
  168. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  169. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  170. }
  171. // For TransposeD node, the IR definition has perm attrs
  172. if (node_type == TRANSPOSED) {
  173. Format src_format = input.GetFormat();
  174. Format dst_format = output.GetFormat();
  175. std::vector<int64_t> perm_arg;
  176. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  177. "Get perm by foramt failed.");
  178. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  179. "SetStr PERMUTE_ATTR_PERM failed")
  180. }
  181. // For cast node, the IR definition has src/dst attrs
  182. if (node_type == CAST) {
  183. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  184. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  185. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  186. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  187. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  188. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  189. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  190. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  191. }
  192. }
  193. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  194. const GeTensorDesc &output, NodePtr &node) {
  195. if (node == nullptr) {
  196. GELOGE(PARAM_INVALID, "node is null.");
  197. return nullptr;
  198. }
  199. auto graph = node->GetOwnerComputeGraph();
  200. if (graph == nullptr) {
  201. GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str());
  202. return nullptr;
  203. }
  204. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  205. if (index < 0) {
  206. GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str());
  207. return nullptr;
  208. }
  209. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  210. if (op_desc == nullptr) {
  211. GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed");
  212. return nullptr;
  213. }
  214. // for data dump
  215. GE_IF_BOOL_EXEC(
  216. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  217. GELOGW("CreateTransNode: SetListStr failed");)
  218. // Default single input and single output
  219. auto ret = op_desc->AddInputDesc(input);
  220. if (ret != GRAPH_SUCCESS) {
  221. GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str());
  222. return nullptr;
  223. }
  224. ret = op_desc->AddOutputDesc(output);
  225. if (ret != GRAPH_SUCCESS) {
  226. GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str());
  227. return nullptr;
  228. }
  229. AddTransNodeAttr(node_type, input, output, op_desc);
  230. NodePtr shape_node = nullptr;
  231. if (node_type == RESHAPE) {
  232. auto shape_desc = CreateTensorShape(output);
  233. if (shape_desc == nullptr) {
  234. GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input",
  235. node->GetName().c_str());
  236. return nullptr;
  237. }
  238. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  239. if (ret != GRAPH_SUCCESS) {
  240. GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str());
  241. return nullptr;
  242. }
  243. shape_node = graph->AddNode(shape_desc);
  244. if (shape_node == nullptr) {
  245. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str());
  246. return nullptr;
  247. }
  248. }
  249. auto trans_node = graph->AddNode(op_desc);
  250. if (trans_node == nullptr) {
  251. GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str());
  252. return nullptr;
  253. }
  254. if (node_type == RESHAPE) {
  255. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  256. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str());
  257. return nullptr;
  258. }
  259. }
  260. return trans_node;
  261. }
  262. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  263. NodePtr &trans_node) {
  264. GE_CHECK_NOTNULL(node);
  265. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  266. if (trans_node == nullptr) {
  267. return INTERNAL_ERROR;
  268. }
  269. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  270. if (ret != GRAPH_SUCCESS) {
  271. GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s",
  272. node->GetName().c_str(), node->GetType().c_str());
  273. return INTERNAL_ERROR;
  274. }
  275. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  276. if (ret != GRAPH_SUCCESS) {
  277. GELOGE(INTERNAL_ERROR, "Failed to connect node %s to trans node %s", node->GetName().c_str(),
  278. trans_node->GetName().c_str());
  279. return INTERNAL_ERROR;
  280. }
  281. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  282. if (ret != GRAPH_SUCCESS) {
  283. GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.",
  284. node->GetName().c_str(), trans_node->GetName().c_str());
  285. return INTERNAL_ERROR;
  286. }
  287. return SUCCESS;
  288. }
  289. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  290. NodePtr &trans_node) {
  291. GE_CHECK_NOTNULL(node);
  292. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  293. if (trans_node == nullptr) {
  294. return INTERNAL_ERROR;
  295. }
  296. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  297. if (ret != GRAPH_SUCCESS) {
  298. GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s",
  299. node->GetName().c_str(), node->GetType().c_str());
  300. return INTERNAL_ERROR;
  301. }
  302. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  303. if (ret != GRAPH_SUCCESS) {
  304. GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(),
  305. node->GetName().c_str());
  306. return INTERNAL_ERROR;
  307. }
  308. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  309. if (ret != GRAPH_SUCCESS) {
  310. GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.",
  311. node->GetName().c_str(), trans_node->GetName().c_str());
  312. return INTERNAL_ERROR;
  313. }
  314. return SUCCESS;
  315. }
  316. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  317. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  318. GE_CHECK_NOTNULL(var->GetOpDesc());
  319. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  320. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  321. output_desc.SetFormat(tensor_desc.GetFormat());
  322. output_desc.SetDataType(tensor_desc.GetDataType());
  323. output_desc.SetShape(tensor_desc.GetShape());
  324. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  325. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  326. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  327. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  328. GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed");
  329. return INTERNAL_ERROR;);
  330. }
  331. if (var->GetOpDesc()->GetInputsSize() > 0) {
  332. auto desc = var->GetOpDesc()->GetInputDesc(0);
  333. desc.SetFormat(tensor_desc.GetFormat());
  334. desc.SetDataType(tensor_desc.GetDataType());
  335. desc.SetShape(tensor_desc.GetShape());
  336. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  337. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  338. desc.SetOriginShape(tensor_desc.GetOriginShape());
  339. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  340. GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed");
  341. return INTERNAL_ERROR;)
  342. }
  343. return SUCCESS;
  344. }
  345. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  346. GE_CHECK_NOTNULL(var);
  347. int index = 0;
  348. NodePtr last_node = var;
  349. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  350. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  351. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  352. if (ret != SUCCESS) {
  353. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(),
  354. index, iter->node_type.c_str());
  355. return INTERNAL_ERROR;
  356. }
  357. // set stream_label
  358. OpDescPtr var_desc = var->GetOpDesc();
  359. GE_CHECK_NOTNULL(var_desc);
  360. std::string stream_label;
  361. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  362. if (!stream_label.empty()) {
  363. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  364. }
  365. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  366. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  367. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  368. }
  369. if (road.empty()) {
  370. return SUCCESS;
  371. }
  372. return UpdateVarFormats(var, road.rbegin()->output);
  373. }
  374. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  375. for (auto &var : nodes) {
  376. GE_CHECK_NOTNULL(var);
  377. int index = 0;
  378. NodePtr last_node = var;
  379. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  380. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  381. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  382. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  383. if (ret != SUCCESS) {
  384. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s",
  385. var->GetName().c_str(), index, iter->node_type.c_str());
  386. return INTERNAL_ERROR;
  387. }
  388. // set stream_label
  389. OpDescPtr var_desc = var->GetOpDesc();
  390. GE_CHECK_NOTNULL(var_desc);
  391. std::string stream_label;
  392. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  393. if (!stream_label.empty()) {
  394. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  395. }
  396. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  397. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  398. }
  399. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  400. return INTERNAL_ERROR;
  401. }
  402. }
  403. return SUCCESS;
  404. }
  405. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  406. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  407. VarNamesToRefs names_to_refs;
  408. std::string var_name;
  409. if (graph == nullptr) {
  410. GELOGE(PARAM_INVALID, "graph is null.");
  411. return names_to_refs;
  412. }
  413. for (auto &node : graph->GetAllNodes()) {
  414. if (node->GetType() != VARIABLE) {
  415. continue;
  416. }
  417. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  418. (void)names_to_refs[var_name].insert(node);
  419. }
  420. }
  421. return names_to_refs;
  422. }
  423. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  424. std::vector<int64_t> &dst_shape) {
  425. if (src_format == FORMAT_NCHW) {
  426. formats::FormatTransferNchwNc1hwc0 transfer;
  427. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  428. GELOGE(INTERNAL_ERROR, "TransShape failed");
  429. return FAILED;
  430. }
  431. } else if (src_format == FORMAT_NHWC) {
  432. formats::FormatTransferNhwcNc1hwc0 transfer;
  433. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  434. GELOGE(INTERNAL_ERROR, "TransShape failed");
  435. return FAILED;
  436. }
  437. }
  438. return SUCCESS;
  439. }
  440. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  441. GE_CHECK_NOTNULL(node_ptr);
  442. auto op_desc = node_ptr->GetOpDesc();
  443. GE_CHECK_NOTNULL(op_desc);
  444. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  445. GE_CHECK_NOTNULL(input);
  446. ge::Format old_format = input->GetFormat();
  447. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  448. ge::DataType dt = input->GetDataType();
  449. std::vector<int64_t> dst_shape_dims;
  450. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  451. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  452. return FAILED;
  453. }
  454. input->SetFormat(FORMAT_NC1HWC0);
  455. input->SetShape(ge::GeShape(dst_shape_dims));
  456. auto output = op_desc->MutableOutputDesc(0);
  457. GE_CHECK_NOTNULL(output);
  458. output->SetFormat(FORMAT_NC1HWC0);
  459. output->SetShape(ge::GeShape(dst_shape_dims));
  460. int64_t size = 0;
  461. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  462. if (graph_status != ge::GRAPH_SUCCESS) {
  463. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  464. return FAILED;
  465. }
  466. ge::TensorUtils::SetSize(*output, size);
  467. ge::TensorUtils::SetSize(*input, size);
  468. return SUCCESS;
  469. }
  470. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  471. GE_CHECK_NOTNULL(input_output);
  472. ge::Format old_format = input_output->GetFormat();
  473. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  474. ge::DataType dt = input_output->GetDataType();
  475. std::vector<int64_t> dst_shape_dims;
  476. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  477. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  478. return FAILED;
  479. }
  480. input_output->SetFormat(FORMAT_NC1HWC0);
  481. input_output->SetShape(ge::GeShape(dst_shape_dims));
  482. return SUCCESS;
  483. }
  484. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  485. vector<int64_t> &dst_shape_dims) {
  486. GE_CHECK_NOTNULL(op_desc);
  487. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  488. GE_CHECK_NOTNULL(input);
  489. ge::Format old_format = input->GetFormat();
  490. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  491. input->SetShape(ge::GeShape(dst_shape_dims));
  492. input->SetFormat(storage_format);
  493. auto output = op_desc->MutableOutputDesc(index);
  494. GE_CHECK_NOTNULL(output);
  495. output->SetShape(ge::GeShape(dst_shape_dims));
  496. output->SetFormat(storage_format);
  497. if (!output->MutableShape().IsUnknownShape()) {
  498. int64_t size = 0;
  499. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  500. if (graph_status != ge::GRAPH_SUCCESS) {
  501. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  502. return FAILED;
  503. }
  504. ge::TensorUtils::SetSize(*input, size);
  505. ge::TensorUtils::SetSize(*output, size);
  506. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  507. "new_shape:%s, new_format:%s, new_size:%lu",
  508. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  509. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  510. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  511. }
  512. return SUCCESS;
  513. }
  514. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) {
  515. is_dynamic_batch = false;
  516. std::string related_node_name;
  517. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  518. if (related_node_name.empty()) {
  519. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty",
  520. data_node->GetName().c_str());
  521. return INTERNAL_ERROR;
  522. }
  523. for (const NodePtr &next_node : data_node->GetOutNodes()) {
  524. if (next_node->GetName() == related_node_name) {
  525. switchn_node = next_node;
  526. break;
  527. }
  528. }
  529. if (switchn_node == nullptr) {
  530. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph",
  531. data_node->GetName().c_str(), related_node_name.c_str());
  532. return INTERNAL_ERROR;
  533. }
  534. is_dynamic_batch = true;
  535. }
  536. return SUCCESS;
  537. }
  538. bool CheckOpType(const NodePtr &node, const std::string type) {
  539. if (node->GetType() == type) {
  540. return true;
  541. }
  542. return false;
  543. }
  544. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  545. auto op = node_ptr->GetOpDesc();
  546. GE_CHECK_NOTNULL(op);
  547. auto inputDescsPtr = op->GetAllInputsDescPtr();
  548. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  549. ge::Format format = ge::FORMAT_ND;
  550. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  551. // process, otherwise fe will insert transdata.
  552. for (auto &inputDescPtr : inputDescsPtr) {
  553. GE_CHECK_NOTNULL(inputDescPtr);
  554. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  555. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  556. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  557. inputDescPtr->SetFormat(format);
  558. inputDescPtr->SetOriginFormat(format);
  559. }
  560. }
  561. for (auto &outputDescPtr : outputDescsPtr) {
  562. GE_CHECK_NOTNULL(outputDescPtr);
  563. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  564. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  565. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  566. outputDescPtr->SetFormat(format);
  567. outputDescPtr->SetOriginFormat(format);
  568. }
  569. }
  570. return SUCCESS;
  571. }
  572. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  573. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  574. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  575. // data and netoutput, and no longer inserts the transnode.
  576. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) {
  577. GE_CHECK_NOTNULL(node_ptr);
  578. auto op_desc = node_ptr->GetOpDesc();
  579. GE_CHECK_NOTNULL(op_desc);
  580. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  581. GE_CHECK_NOTNULL(input);
  582. ge::DataType src_dtype = input->GetDataType();
  583. if (src_dtype == dt_set) {
  584. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  585. return SUCCESS;
  586. }
  587. input->SetDataType(dt_set);
  588. int64_t input_shape_size = 0;
  589. int64_t output_shape_size = 0;
  590. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  591. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  592. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  593. GELOGE(GRAPH_FAILED, "GetTensorSize failed!");
  594. return FAILED;
  595. }
  596. ge::TensorUtils::SetSize(*input, input_shape_size);
  597. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  598. GE_CHECK_NOTNULL(output);
  599. output->SetDataType(dt_set);
  600. ge::TensorUtils::SetSize(*output, output_shape_size);
  601. if (is_dynamic_batch) {
  602. GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str());
  603. auto switchn_op_desc = switchn_node->GetOpDesc();
  604. GE_CHECK_NOTNULL(switchn_op_desc);
  605. auto switchn_input = switchn_op_desc->MutableInputDesc(0);
  606. GE_CHECK_NOTNULL(switchn_input);
  607. switchn_input->SetDataType(dt_set);
  608. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  609. const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i);
  610. GE_CHECK_NOTNULL(switchn_output);
  611. switchn_output->SetDataType(dt_set);
  612. }
  613. }
  614. return SUCCESS;
  615. }
  616. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  617. GE_CHECK_NOTNULL(node_ptr);
  618. auto op_desc = node_ptr->GetOpDesc();
  619. GE_CHECK_NOTNULL(op_desc);
  620. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  621. GE_CHECK_NOTNULL(input);
  622. ge::Format old_format = input->GetFormat();
  623. ge::GeShape old_shape = input->GetShape();
  624. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  625. if (!support) {
  626. GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str());
  627. return FAILED;
  628. }
  629. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  630. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  631. return FAILED;
  632. }
  633. if (is_dynamic_batch) {
  634. auto switchn_op_desc = switchn_node->GetOpDesc();
  635. GE_CHECK_NOTNULL(switchn_op_desc);
  636. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  637. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  638. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  639. return FAILED;
  640. }
  641. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  642. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  643. GE_CHECK_NOTNULL(switchn_output);
  644. old_format = switchn_output->GetFormat();
  645. old_shape = switchn_output->GetShape();
  646. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  647. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  648. return FAILED;
  649. }
  650. }
  651. }
  652. return SUCCESS;
  653. }
  654. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  655. auto op_desc = node_ptr->GetOpDesc();
  656. GE_CHECK_NOTNULL(op_desc);
  657. string set_dt_str;
  658. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  659. return SUCCESS;
  660. }
  661. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  662. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  663. bool is_dynamic_batch = false;
  664. NodePtr switchn_node = nullptr;
  665. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) {
  666. GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed");
  667. return FAILED;
  668. }
  669. if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) {
  670. GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed");
  671. return FAILED;
  672. }
  673. // check if need to set format
  674. string set_format;
  675. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  676. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  677. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  678. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) {
  679. GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed");
  680. return FAILED;
  681. }
  682. }
  683. return SUCCESS;
  684. }
  685. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  686. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  687. GE_CHECK_NOTNULL(op_desc);
  688. GE_CHECK_NOTNULL(tensor_desc_ptr);
  689. storage_format = FORMAT_RESERVED;
  690. int64_t format = FORMAT_RESERVED;
  691. dst_shape_dims.clear();
  692. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  693. storage_format = static_cast<Format>(format);
  694. vector<int32_t> storage_shape;
  695. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  696. for (auto dim : storage_shape) {
  697. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  698. }
  699. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  700. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  701. formats::JoinToString(storage_shape).c_str());
  702. } else {
  703. GELOGE(PARAM_INVALID, "Update node by storage format failed, storage_shape not set. "
  704. "node: [%s], storage_format [%s]",
  705. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  706. return FAILED;
  707. }
  708. ge::Format old_format = tensor_desc_ptr->GetFormat();
  709. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  710. if (old_format == storage_format && old_shape == dst_shape_dims) {
  711. GELOGI("Update node by storage format, not changed.");
  712. storage_format = FORMAT_RESERVED;
  713. return SUCCESS;
  714. }
  715. }
  716. return SUCCESS;
  717. }
  718. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  719. NodePtr &node) {
  720. bool is_dynamic = CheckOpType(node, MERGE);
  721. auto src_op_desc = node->GetOpDesc();
  722. GE_CHECK_NOTNULL(src_op_desc);
  723. ge::GeShape src_shape = src_desc.GetShape();
  724. ge::Format src_format = src_desc.GetFormat();
  725. net_output_input_desc->SetDataType(DT_FLOAT16);
  726. if (is_dynamic) {
  727. auto merge_output = src_op_desc->MutableOutputDesc(0);
  728. GE_CHECK_NOTNULL(merge_output);
  729. merge_output->SetDataType(DT_FLOAT16);
  730. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  731. auto merge_input = src_op_desc->MutableInputDesc(i);
  732. GE_CHECK_NOTNULL(merge_input);
  733. merge_input->SetDataType(DT_FLOAT16);
  734. }
  735. }
  736. std::vector<int64_t> dst_shape_dims;
  737. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  738. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  739. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  740. return FAILED;
  741. }
  742. ge::GeShape dst_shape(dst_shape_dims);
  743. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  744. net_output_input_desc->SetShape(dst_shape);
  745. if (is_dynamic) {
  746. auto merge_out = src_op_desc->MutableOutputDesc(0);
  747. GE_CHECK_NOTNULL(merge_out);
  748. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  749. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  750. return FAILED;
  751. }
  752. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  753. auto merge_in = src_op_desc->MutableInputDesc(i);
  754. GE_CHECK_NOTNULL(merge_in);
  755. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  756. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  757. return FAILED;
  758. }
  759. }
  760. }
  761. return SUCCESS;
  762. }
  763. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  764. GE_CHECK_NOTNULL(netout_desc);
  765. vector<string> output_dt_str;
  766. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  767. for (auto dt_str : output_dt_str) {
  768. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  769. if (dt_str_split.size() == kUserDefinedElementCount) {
  770. if (dt_str_split[0] == to_string(index)) {
  771. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  772. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  773. TypeUtils::DataTypeToSerialString(dt).c_str());
  774. return true;
  775. }
  776. }
  777. }
  778. }
  779. return false;
  780. }
  781. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  782. GE_CHECK_NOTNULL(netout_desc);
  783. vector<string> output_format_str;
  784. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  785. for (auto format_str : output_format_str) {
  786. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  787. if (format_str_split.size() == kUserDefinedElementCount) {
  788. if (format_str_split[0] == to_string(index)) {
  789. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  790. return true;
  791. }
  792. }
  793. }
  794. }
  795. return false;
  796. }
  797. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  798. auto op_desc = node->GetOpDesc();
  799. GE_CHECK_NOTNULL(op_desc);
  800. ge::DataType output_data_type = ge::DT_FLOAT;
  801. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  802. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  803. auto peer_out = in_anchor->GetPeerOutAnchor();
  804. GE_CHECK_NOTNULL(peer_out);
  805. auto src_node = peer_out->GetOwnerNode();
  806. GE_CHECK_NOTNULL(src_node);
  807. bool is_dynamic = CheckOpType(src_node, MERGE);
  808. OpDescPtr src_op_desc = src_node->GetOpDesc();
  809. GE_CHECK_NOTNULL(src_op_desc);
  810. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  811. GE_CHECK_NOTNULL(net_output_input_desc);
  812. ge::GeShape old_shape = net_output_input_desc->GetShape();
  813. ge::Format old_format = net_output_input_desc->GetFormat();
  814. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  815. // Update datatype
  816. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  817. GELOGI("Enter into process output_type schedule");
  818. net_output_input_desc->SetDataType(output_data_type);
  819. if (is_dynamic) {
  820. auto merge_output = src_op_desc->MutableOutputDesc(0);
  821. GE_CHECK_NOTNULL(merge_output);
  822. merge_output->SetDataType(output_data_type);
  823. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  824. auto merge_input = src_op_desc->MutableInputDesc(i);
  825. GE_CHECK_NOTNULL(merge_input);
  826. merge_input->SetDataType(output_data_type);
  827. }
  828. }
  829. }
  830. // check if is_output_adjust_hw_layout is set
  831. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  832. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  833. GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0.");
  834. return FAILED;
  835. }
  836. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  837. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  838. GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0.");
  839. return FAILED;
  840. }
  841. }
  842. }
  843. return SUCCESS;
  844. }
  845. } // namespace
  846. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  847. GraphPrepare::~GraphPrepare() {}
  848. /**
  849. * @param graph
  850. * @return
  851. */
  852. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  853. GE_CHECK_NOTNULL(graph);
  854. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  855. for (auto &node : graph->GetAllNodes()) {
  856. if (node == nullptr) {
  857. continue;
  858. }
  859. if (node->GetType() != VARIABLE) {
  860. continue;
  861. }
  862. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  863. if (trans_road == nullptr) {
  864. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  865. continue;
  866. }
  867. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  868. auto ret = RecoverTransRoadForVar(node, *trans_road);
  869. if (ret != SUCCESS) {
  870. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str());
  871. return INTERNAL_ERROR;
  872. }
  873. auto iter = var_names_to_refs.find(node->GetName());
  874. if (iter != var_names_to_refs.end()) {
  875. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  876. if (ret != SUCCESS) {
  877. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str());
  878. return INTERNAL_ERROR;
  879. }
  880. }
  881. }
  882. return SUCCESS;
  883. }
  884. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  885. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  886. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  887. if (compute_graph_ != nullptr) {
  888. compute_graph_->SetSessionID(session_id);
  889. }
  890. session_id_ = session_id;
  891. Status ret = CheckGraph();
  892. if (ret != SUCCESS) {
  893. GELOGE(ret, "RunGraph graph check fail, ret:%u", ret);
  894. return ret;
  895. }
  896. (void)compute_graph_->TopologicalSorting();
  897. ret = CheckRefOp();
  898. if (ret != SUCCESS) {
  899. GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret);
  900. return ret;
  901. }
  902. return SUCCESS;
  903. }
  904. Status GraphPrepare::CheckGraph() {
  905. if (compute_graph_ == nullptr) {
  906. GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR");
  907. return GE_GRAPH_INIT_FAILED;
  908. }
  909. auto nodes = compute_graph_->GetAllNodes();
  910. if (nodes.empty()) {
  911. GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph.");
  912. return GE_GRAPH_INIT_FAILED;
  913. }
  914. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  915. GE_CHECK_NOTNULL(node);
  916. if (node->GetOpDesc() == nullptr) {
  917. GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL");
  918. return GE_GRAPH_INIT_FAILED;
  919. }
  920. }
  921. return SUCCESS;
  922. }
  923. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  924. const std::set<NodePtr> &ref_nodes) {
  925. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  926. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  927. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  928. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  929. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  930. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH};
  931. GE_CHECK_NOTNULL(node);
  932. const auto &op_desc = node->GetOpDesc();
  933. GE_CHECK_NOTNULL(op_desc);
  934. const auto input_index = op_desc->GetInputIndexByName(input_name);
  935. const auto &in_anchor = node->GetInDataAnchor(input_index);
  936. GE_CHECK_NOTNULL(in_anchor);
  937. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  938. GE_CHECK_NOTNULL(peer_out_anchor);
  939. const auto &input_node = peer_out_anchor->GetOwnerNode();
  940. GE_CHECK_NOTNULL(input_node);
  941. const auto &input_op_desc = input_node->GetOpDesc();
  942. GE_CHECK_NOTNULL(input_op_desc);
  943. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  944. if (is_ref) {
  945. return SUCCESS;
  946. }
  947. auto input_type = input_op_desc->GetType();
  948. if (input_type == ge::FRAMEWORKOP) {
  949. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  950. GELOGE(PARAM_INVALID, "Get original type failed.");
  951. return PARAM_INVALID;
  952. }
  953. }
  954. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  955. if (!is_acceptable) {
  956. GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.",
  957. node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  958. input_op_desc->GetType().c_str());
  959. return PARAM_INVALID;
  960. }
  961. return SUCCESS;
  962. }
  963. Status GraphPrepare::CheckRefOp() {
  964. GE_CHECK_NOTNULL(compute_graph_);
  965. std::set<NodePtr> ref_nodes;
  966. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  967. if (node == nullptr) {
  968. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  969. return PARAM_INVALID;
  970. }
  971. auto op_desc = node->GetOpDesc();
  972. if (op_desc == nullptr) {
  973. GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null.");
  974. return PARAM_INVALID;
  975. }
  976. auto input_name_index = op_desc->GetAllInputName();
  977. auto outputs = op_desc->GetAllOutputName();
  978. for (const auto &name_index : input_name_index) {
  979. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  980. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  981. GELOGE(PARAM_INVALID, "CheckRefInputNode failed.");
  982. return PARAM_INVALID;
  983. }
  984. (void)ref_nodes.insert(node); // no need to check value
  985. }
  986. }
  987. }
  988. return SUCCESS;
  989. };
  990. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  991. GE_CHECK_NOTNULL(compute_graph_);
  992. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  993. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  994. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  995. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  996. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  997. return SUCCESS;
  998. }
  999. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1000. if (node == nullptr) {
  1001. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL");
  1002. return GE_GRAPH_GRAPH_NODE_NULL;
  1003. }
  1004. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1005. if (op_desc_ptr == nullptr) {
  1006. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL");
  1007. return GE_GRAPH_GRAPH_NODE_NULL;
  1008. }
  1009. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1010. int64_t tensor_size = 0;
  1011. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1012. if (graph_status != GRAPH_SUCCESS) {
  1013. ErrorManager::GetInstance().ATCReportErrMessage(
  1014. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1015. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  1016. return FAILED;
  1017. }
  1018. TensorUtils::SetSize(output, tensor_size);
  1019. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1020. if (graph_ret != GRAPH_SUCCESS) {
  1021. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1022. return graph_ret;
  1023. }
  1024. return SUCCESS;
  1025. }
  1026. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input) {
  1027. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1028. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1029. GE_CHECK_NOTNULL(input_node);
  1030. OpDescPtr op = input_node->GetOpDesc();
  1031. GE_CHECK_NOTNULL(op);
  1032. if (op->GetType() == DATA) {
  1033. GeAttrValue::INT index = 0;
  1034. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1035. GELOGW("Get index from data attr failed");
  1036. continue;
  1037. }
  1038. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1039. GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index);
  1040. return FAILED;
  1041. }
  1042. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1043. auto format = desc.GetFormat();
  1044. auto origin_format = desc.GetOriginFormat();
  1045. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1046. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op);
  1047. if (need_check_internal_format) {
  1048. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1049. if (is_internal) {
  1050. GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.",
  1051. TypeUtils::FormatToSerialString(format).c_str(),
  1052. TypeUtils::FormatToSerialString(origin_format).c_str());
  1053. return FAILED;
  1054. }
  1055. }
  1056. auto data_type = desc.GetDataType();
  1057. uint32_t length = 1;
  1058. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1059. if (!type_ret) {
  1060. GELOGE(PARAM_INVALID, "Input datatype %s is not support.",
  1061. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1062. return FAILED;
  1063. }
  1064. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1065. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1066. int64_t shape_size = desc_shape * length;
  1067. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1068. int64_t size = 0;
  1069. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1070. GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed");
  1071. return FAILED);
  1072. bool size_check = (size != 0 && shape_size != size);
  1073. if (size_check) {
  1074. GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size);
  1075. return FAILED;
  1076. }
  1077. ge::TensorUtils::SetSize(desc, shape_size);
  1078. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1079. if (graph_ret != GRAPH_SUCCESS) {
  1080. GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
  1081. return graph_ret;
  1082. }
  1083. // Size will be recalculated in the build stage
  1084. ge::TensorUtils::SetSize(desc, 0);
  1085. graph_ret = op->UpdateOutputDesc(0, desc);
  1086. if (graph_ret != GRAPH_SUCCESS) {
  1087. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1088. return graph_ret;
  1089. }
  1090. if (!options_.train_graph_flag) {
  1091. Status ret = AdjustDataOpOutput(input_node);
  1092. if (ret != SUCCESS) {
  1093. GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret);
  1094. return ret;
  1095. }
  1096. }
  1097. }
  1098. }
  1099. return SUCCESS;
  1100. }
  1101. Status GraphPrepare::TryDoAipp() {
  1102. // infer and with aipp configure file, then call aipp insert
  1103. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1104. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1105. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1106. if (ret != SUCCESS) {
  1107. GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed.");
  1108. return INTERNAL_ERROR;
  1109. }
  1110. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1111. if (ret != SUCCESS) {
  1112. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed",
  1113. options_.insert_op_file.c_str());
  1114. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1115. }
  1116. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1117. if (ret != SUCCESS) {
  1118. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret);
  1119. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1120. }
  1121. }
  1122. return SUCCESS;
  1123. }
  1124. Status GraphPrepare::FormatAndShapeProcess() {
  1125. Status ret = ResourcePairProcess("add");
  1126. if (ret != SUCCESS) {
  1127. GELOGE(ret, "ResourcePairProcess failed");
  1128. return ret;
  1129. }
  1130. GE_TIMESTAMP_START(InferOriginFormat1);
  1131. ret = compute_graph_->InferOriginFormat();
  1132. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1133. GE_DUMP(compute_graph_, "after_first_inferformat");
  1134. if (ret != SUCCESS) {
  1135. GELOGE(ret, "Prepare Graph first inferformat failed");
  1136. return ret;
  1137. }
  1138. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1139. ret = InferShapeForPreprocess();
  1140. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1141. GE_DUMP(compute_graph_, "after_infershape");
  1142. if (ret != SUCCESS) {
  1143. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed");
  1144. return GE_GRAPH_INFERSHAPE_FAILED;
  1145. }
  1146. GE_TIMESTAMP_START(InferOriginFormat2);
  1147. ret = compute_graph_->InferOriginFormat();
  1148. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1149. if (ret != SUCCESS) {
  1150. GELOGE(ret, "Prepare Graph inferformat failed");
  1151. return ret;
  1152. }
  1153. ret = ResourcePairProcess("remove");
  1154. if (ret != SUCCESS) {
  1155. return ret;
  1156. }
  1157. return ret;
  1158. }
  1159. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1160. PassManager control_pass;
  1161. // Graph pass tmp logic for resource infershape
  1162. if (options_.train_graph_flag) {
  1163. try {
  1164. if (action == "add") {
  1165. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1166. } else {
  1167. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1168. new ResourcePairRemoveControlPass);
  1169. }
  1170. } catch (std::bad_alloc &e) {
  1171. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str());
  1172. return INTERNAL_ERROR;
  1173. }
  1174. }
  1175. Status ret = control_pass.Run(compute_graph_);
  1176. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1177. GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret);
  1178. return ret;
  1179. }
  1180. return SUCCESS;
  1181. }
  1182. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1183. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1184. GE_CHECK_NOTNULL(node_ptr);
  1185. if (node_ptr->GetType() == DATA) {
  1186. uint32_t index = 0;
  1187. auto op_desc = node_ptr->GetOpDesc();
  1188. GE_CHECK_NOTNULL(op_desc);
  1189. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1190. Format storage_format = FORMAT_RESERVED;
  1191. vector<int64_t> dst_shape_dims;
  1192. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1193. GELOGE(INTERNAL_ERROR, "Get storage format for input failed");
  1194. return FAILED;
  1195. }
  1196. if (storage_format == FORMAT_RESERVED) {
  1197. continue;
  1198. }
  1199. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1200. GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed");
  1201. return FAILED;
  1202. }
  1203. }
  1204. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1205. auto op_desc = node_ptr->GetOpDesc();
  1206. GE_CHECK_NOTNULL(op_desc);
  1207. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1208. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1209. Format storage_format = FORMAT_RESERVED;
  1210. vector<int64_t> dst_shape_dims;
  1211. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1212. GELOGE(INTERNAL_ERROR, "Get storage format from output failed");
  1213. return FAILED;
  1214. }
  1215. if (storage_format == FORMAT_RESERVED) {
  1216. continue;
  1217. }
  1218. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1219. GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed");
  1220. return FAILED;
  1221. }
  1222. }
  1223. }
  1224. }
  1225. return SUCCESS;
  1226. }
  1227. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1228. if (options_.save_original_model == "true") {
  1229. ModelHelper model_helper;
  1230. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1231. options_.original_model_file);
  1232. if (ret != SUCCESS) {
  1233. // If save original model fail, process continue
  1234. GELOGW("SaveOriginalGraphToOmModel fail");
  1235. }
  1236. }
  1237. return SUCCESS;
  1238. }
  1239. #define PP_RUN_AND_DUMP(name, func, ...) \
  1240. do { \
  1241. GE_RUN(Prepare, func, __VA_ARGS__); \
  1242. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1243. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1244. } while (0)
  1245. #define PP_RUN(name, func, ...) \
  1246. do { \
  1247. GE_RUN(Prepare, func, __VA_ARGS__); \
  1248. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1249. } while (0)
  1250. Status GraphPrepare::PrepareDynShape(ConstGraphPtr graph, const std::vector<GeTensor> &user_input,
  1251. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1252. GE_CHECK_NOTNULL(graph);
  1253. GE_CHECK_NOTNULL(compute_graph);
  1254. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1255. const Graph &const_graph = *graph;
  1256. PP_RUN("Init", Init, const_graph, session_id);
  1257. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1258. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input);
  1259. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1260. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1261. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1262. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1263. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1264. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1265. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1266. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1267. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1268. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1269. return SUCCESS;
  1270. }
  1271. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1272. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1273. return SUCCESS;
  1274. }
  1275. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1276. auto compute_graph = compute_graph_;
  1277. PassManager pass_manager;
  1278. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1279. new (std::nothrow) VariablePrepareOpPass))
  1280. GE_TIMESTAMP_START(pass_manager);
  1281. auto ret = pass_manager.Run(compute_graph);
  1282. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1283. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1284. GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret);
  1285. return ret;
  1286. }
  1287. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1288. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1289. return SUCCESS;
  1290. }
  1291. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1292. if (compute_graph == nullptr) {
  1293. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1294. return GE_GRAPH_NULL_INPUT;
  1295. }
  1296. GEPass ge_passes(compute_graph);
  1297. NamesToPass hccl_group;
  1298. HcclGroupPass hccl_group_pass;
  1299. GELOGD("Add hccl group pass success");
  1300. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1301. auto ret = ge_passes.Run(hccl_group);
  1302. if (ret != SUCCESS) {
  1303. GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret);
  1304. return ret;
  1305. }
  1306. ret = compute_graph->TopologicalSorting();
  1307. if (ret != SUCCESS) {
  1308. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1309. return ret;
  1310. }
  1311. return SUCCESS;
  1312. }
  1313. #undef PP_RUN_AND_DUMP
  1314. #undef PP_RUN
  1315. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1316. if (graph == nullptr) {
  1317. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1318. return GE_GRAPH_NULL_INPUT;
  1319. }
  1320. const Graph &const_graph = *graph;
  1321. Status ret = Init(const_graph, 0);
  1322. if (ret != SUCCESS) {
  1323. GELOGE(ret, "Init graph_prepare fail, ret:%u", ret);
  1324. return ret;
  1325. }
  1326. GE_DUMP(compute_graph_, "after_parser");
  1327. GELOGI("Start infershape for dump json process.");
  1328. ret = compute_graph_->InferOriginFormat();
  1329. GE_DUMP(compute_graph_, "after_inferformat");
  1330. if (ret != SUCCESS) {
  1331. GELOGE(ret, "Prepare Graph inferformat failed");
  1332. return ret;
  1333. }
  1334. InferShapePass infer_shape_pass;
  1335. NamesToPass names_to_passes;
  1336. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1337. GEPass ge_passes(compute_graph_);
  1338. ret = ge_passes.Run(names_to_passes);
  1339. GE_DUMP(compute_graph_, "after_infershape");
  1340. if (ret != SUCCESS) {
  1341. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1342. return ret;
  1343. }
  1344. ShapeRefiner::ClearContextMap();
  1345. return SUCCESS;
  1346. }
  1347. Status GraphPrepare::CheckConstOp() {
  1348. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1349. GE_CHECK_NOTNULL(node_ptr);
  1350. if (node_ptr->GetType() == CONSTANT) {
  1351. Status ret = VerifyConstOp(node_ptr);
  1352. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1353. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1354. auto op_desc = node_ptr->GetOpDesc();
  1355. if (op_desc == nullptr) {
  1356. GELOGE(PARAM_INVALID, "Get op desc failed");
  1357. return PARAM_INVALID;
  1358. }
  1359. std::string original_type;
  1360. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1361. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1362. GELOGI("original type is %s", original_type.c_str());
  1363. if (original_type == CONSTANT) {
  1364. Status ret = VerifyConstOp(node_ptr);
  1365. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1366. }
  1367. }
  1368. }
  1369. return SUCCESS;
  1370. }
  1371. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1372. GE_CHECK_NOTNULL(node);
  1373. auto op_desc = node->GetOpDesc();
  1374. GE_CHECK_NOTNULL(op_desc);
  1375. ConstGeTensorPtr ge_tensor_ptr;
  1376. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1377. GELOGE(PARAM_INVALID, "Get value from const attr failed");
  1378. return PARAM_INVALID;
  1379. }
  1380. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1381. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1382. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1383. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1384. auto data_type = ge_tensor_desc.GetDataType();
  1385. uint32_t length = 1;
  1386. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1387. if (!type_ret) {
  1388. GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str());
  1389. return FAILED;
  1390. }
  1391. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1392. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1393. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1394. if (shape_size == 0) {
  1395. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1396. // shape = [], means it's a sclar tensor.
  1397. GE_CHK_BOOL_EXEC(data_size / length == 1, return PARAM_INVALID, "Const is invalid scalar tensor.");
  1398. } else {
  1399. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1400. GE_CHK_BOOL_EXEC(data_size == 0, return PARAM_INVALID, "Const is invalid vector scalar.");
  1401. }
  1402. } else {
  1403. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0, return PARAM_INVALID,
  1404. "Const input data size is not equal with tensor desc shape");
  1405. }
  1406. return SUCCESS;
  1407. }
  1408. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1409. if (GetLocalOmgContext().is_dynamic_input) {
  1410. return SUCCESS;
  1411. }
  1412. unsigned int node_num = 0;
  1413. unsigned int data_num = 0;
  1414. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1415. GE_CHECK_NOTNULL(input_node);
  1416. OpDescPtr op = input_node->GetOpDesc();
  1417. GE_CHECK_NOTNULL(op);
  1418. node_num++;
  1419. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1420. data_num++;
  1421. GeAttrValue::INT index = 0;
  1422. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1423. GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed");
  1424. return GE_GRAPH_INIT_FAILED;
  1425. }
  1426. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1427. GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index);
  1428. return GE_GRAPH_INIT_FAILED;
  1429. }
  1430. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1431. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1432. if (desc.GetShape().GetDim(i) < 0) {
  1433. GELOGE(GE_GRAPH_INIT_FAILED, "data dim %zu is not supported, need >= 0, real:%ld.", i,
  1434. desc.GetShape().GetDim(i));
  1435. return GE_GRAPH_INIT_FAILED;
  1436. }
  1437. }
  1438. }
  1439. }
  1440. if (node_num <= data_num) {
  1441. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1442. }
  1443. return SUCCESS;
  1444. }
  1445. Status GraphPrepare::InferShapeForPreprocess() {
  1446. GELOGI("Start infershape for preprocess.");
  1447. GEPass ge_passes(compute_graph_);
  1448. NamesToPass names_to_passes;
  1449. AssertPass assert_pass;
  1450. if (!options_.train_graph_flag) {
  1451. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1452. }
  1453. InferShapePass infer_shape_pass;
  1454. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1455. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1456. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1457. DimensionComputePass dimension_compute_pass;
  1458. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1459. ConstantFoldingPass constant_folding_pass;
  1460. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1461. int32_t dev_count = 0;
  1462. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1463. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1464. rtError_t rt_err = RT_ERROR_NONE;
  1465. if (aicpu_constant_folding_on != nullptr) {
  1466. rt_err = rtGetDeviceCount(&dev_count);
  1467. if (rt_err == RT_ERROR_NONE) {
  1468. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1469. if (result != SUCCESS) {
  1470. GELOGE(result, "Set rt context failed.");
  1471. return result;
  1472. }
  1473. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1474. }
  1475. }
  1476. Status ret = ge_passes.Run(names_to_passes);
  1477. if (aicpu_constant_folding_on != nullptr) {
  1478. if (rt_err == RT_ERROR_NONE) {
  1479. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1480. if (result != SUCCESS) {
  1481. GELOGE(result, "Set rt context failed.");
  1482. return result;
  1483. }
  1484. }
  1485. }
  1486. ShapeRefiner::ClearContextMap();
  1487. if (ret != SUCCESS) {
  1488. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1489. return ret;
  1490. }
  1491. return SUCCESS;
  1492. }
  1493. Status GraphPrepare::PrepareOptimize() {
  1494. GELOGI("Start optimize for preprocess.");
  1495. // check rw type
  1496. GraphOptimize graph_optimize;
  1497. bool has_conflict = false;
  1498. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1499. if (has_conflict) {
  1500. GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize.");
  1501. return FAILED;
  1502. }
  1503. PassManager original_graph_passes;
  1504. // Graph pass
  1505. try {
  1506. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1507. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1508. } catch (std::bad_alloc &e) {
  1509. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1510. return INTERNAL_ERROR;
  1511. }
  1512. GE_TIMESTAMP_START(original_graph_passes);
  1513. Status ret = original_graph_passes.Run(compute_graph_);
  1514. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1515. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1516. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1517. return ret;
  1518. }
  1519. // New pass
  1520. GEPass ge_passes(compute_graph_);
  1521. NamesToPass names_to_passes;
  1522. EnterPass enter_pass;
  1523. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1524. CondPass cond_pass;
  1525. names_to_passes.emplace_back("CondPass", &cond_pass);
  1526. PrintOpPass print_pass;
  1527. if (options_.enable_print_op_pass) {
  1528. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1529. }
  1530. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1531. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1532. DropOutPass dropout_pass;
  1533. AssertPass assert_pass;
  1534. UnusedConstPass unused_const_pass;
  1535. StopGradientPass stop_gradient_pass;
  1536. PreventGradientPass prevent_gradient_pass;
  1537. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1538. GuaranteeConstPass guarantee_const_pass;
  1539. VarIsInitializedOpPass var_is_initialized_pass;
  1540. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1541. IdentityPass identity_pass(false);
  1542. AssignPass assign_pass;
  1543. SnapshotPass snapshot_pass;
  1544. if (!options_.train_graph_flag) {
  1545. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1546. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1547. }
  1548. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1549. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1550. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1551. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1552. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1553. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1554. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1555. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1556. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1557. if (GetContext().GetHostExecFlag()) {
  1558. names_to_passes.emplace_back("AssignPass", &assign_pass);
  1559. }
  1560. GE_TIMESTAMP_START(names_to_passes);
  1561. ret = ge_passes.Run(names_to_passes);
  1562. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1563. if (ret != SUCCESS) {
  1564. GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret);
  1565. return ret;
  1566. }
  1567. PassManager graph_pass;
  1568. try {
  1569. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1570. } catch (std::bad_alloc &e) {
  1571. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1572. return INTERNAL_ERROR;
  1573. }
  1574. GE_TIMESTAMP_START(graph_passes);
  1575. ret = graph_pass.Run(compute_graph_);
  1576. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1577. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1578. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1579. return ret;
  1580. }
  1581. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1582. TypeConversionOfConstant();
  1583. ret = compute_graph_->TopologicalSorting();
  1584. if (ret != SUCCESS) {
  1585. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1586. return ret;
  1587. }
  1588. GELOGI("End optimize for preprocess.");
  1589. return SUCCESS;
  1590. }
  1591. void GraphPrepare::TypeConversionOfConstant() {
  1592. if (options_.train_graph_flag) {
  1593. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1594. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1595. // This can ensure that n is not a null pointer
  1596. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1597. n->GetOpDesc()->SetType(CONSTANTOP);
  1598. }
  1599. }
  1600. } else {
  1601. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  1602. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1603. // This can ensure that n is not a null pointer
  1604. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  1605. n->GetOpDesc()->SetType(CONSTANT);
  1606. }
  1607. }
  1608. }
  1609. }
  1610. Status GraphPrepare::GraphEquivalentTransformation() {
  1611. NamesToPass names_to_pass;
  1612. ForPass for_pass;
  1613. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  1614. return GEPass(compute_graph_).Run(names_to_pass);
  1615. }
  1616. Status GraphPrepare::ProcessBeforeInfershape() {
  1617. NamesToPass names_to_passes;
  1618. CondRemovePass condition_remove_pass;
  1619. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  1620. GE_TIMESTAMP_START(ProcessCondRemove);
  1621. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  1622. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  1623. if (ret != SUCCESS) {
  1624. GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  1625. return ret;
  1626. }
  1627. return SUCCESS;
  1628. }
  1629. Status GraphPrepare::ProcessNetOutput() {
  1630. PassManager graph_passes_before_infershape;
  1631. try {
  1632. if (options_.train_graph_flag) {
  1633. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  1634. }
  1635. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  1636. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  1637. new (std::nothrow) DataPass); // Add NetOutput first.
  1638. } catch (std::bad_alloc) {
  1639. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1640. return INTERNAL_ERROR;
  1641. }
  1642. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  1643. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  1644. GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret);
  1645. return ret;
  1646. }
  1647. return SUCCESS;
  1648. }
  1649. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input) {
  1650. compute_graph_->SetInputSize(user_input.size());
  1651. if (user_input.empty()) {
  1652. return SUCCESS;
  1653. }
  1654. auto ret = CheckUserInput(user_input);
  1655. if (ret != SUCCESS) {
  1656. GELOGE(ret, "Check user input failed.");
  1657. return ret;
  1658. }
  1659. ret = UpdateInput(user_input);
  1660. if (ret != SUCCESS) {
  1661. GELOGE(ret, "UpdateInput fail, ret:%u", ret);
  1662. return ret;
  1663. }
  1664. if (user_input.size() != 0) {
  1665. ret = CheckConstOp();
  1666. if (ret != SUCCESS) {
  1667. GELOGE(ret, "CheckConstOp fail, ret:%u", ret);
  1668. return ret;
  1669. }
  1670. } else {
  1671. ret = compute_graph_->TopologicalSorting();
  1672. if (ret != SUCCESS) {
  1673. GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting");
  1674. return FAILED;
  1675. }
  1676. }
  1677. return SUCCESS;
  1678. }
  1679. Status GraphPrepare::UpdateInputOutputByOptions() {
  1680. auto ret = UpdateDataNetOutputByStorageFormat();
  1681. if (ret != SUCCESS) {
  1682. GELOGE(ret, "Update format acoording to storage format failed.");
  1683. return ret;
  1684. }
  1685. if (options_.train_graph_flag) {
  1686. GELOGI("This is train mode, no need to do this schedule.");
  1687. return SUCCESS;
  1688. }
  1689. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  1690. GE_CHECK_NOTNULL(node_ptr);
  1691. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  1692. GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str());
  1693. return FAILED;
  1694. }
  1695. if (node_ptr->GetType() == DATA) {
  1696. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  1697. GELOGE(INTERNAL_ERROR, "Process data node failed");
  1698. return FAILED;
  1699. }
  1700. }
  1701. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1702. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  1703. GELOGE(INTERNAL_ERROR, "Process netoutput node failed");
  1704. return FAILED;
  1705. }
  1706. }
  1707. }
  1708. return SUCCESS;
  1709. }
  1710. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  1711. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  1712. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  1713. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  1714. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  1715. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  1716. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  1717. if (dst_node->GetType() == TRANSDATA) {
  1718. return true;
  1719. }
  1720. }
  1721. }
  1722. return false;
  1723. }
  1724. } // 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