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 103 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
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
4 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
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
4 years ago
4 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
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
5 years ago
4 years ago
4 years ago
5 years ago
4 years ago
4 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
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
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 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
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 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
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
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
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 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
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295
  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 "common/formats/format_transfers/format_transfer_fractal_nz.h"
  21. #include "common/formats/format_transfers/format_transfer_nchw_nc1hwc0.h"
  22. #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h"
  23. #include "common/formats/format_transfers/format_transfer_transpose.h"
  24. #include "common/formats/utils/formats_trans_utils.h"
  25. #include "common/util/error_manager/error_manager.h"
  26. #include "common/helper/model_helper.h"
  27. #include "common/math/math_util.h"
  28. #include "common/op/ge_op_utils.h"
  29. #include "ir_build/option_utils.h"
  30. #include "graph/common/ge_call_wrapper.h"
  31. #include "graph/common/local_context.h"
  32. #include "graph/common/transop_util.h"
  33. #include "graph/ge_context.h"
  34. #include "graph/shape_refiner.h"
  35. #include "graph/manager/graph_var_manager.h"
  36. #include "graph/manager/util/rt_context_util.h"
  37. #include "graph/optimize/graph_optimize.h"
  38. #include "graph/passes/addn_pass.h"
  39. #include "graph/passes/aicpu_constant_folding_pass.h"
  40. #include "graph/passes/assert_pass.h"
  41. #include "ge/ge_api_types.h"
  42. #include "graph/passes/common_subexpression_elimination_pass.h"
  43. #include "graph/passes/cond_pass.h"
  44. #include "graph/passes/cond_remove_pass.h"
  45. #include "graph/passes/constant_folding_pass.h"
  46. #include "graph/passes/dimension_adjust_pass.h"
  47. #include "graph/passes/dimension_compute_pass.h"
  48. #include "graph/passes/dropout_pass.h"
  49. #include "graph/passes/enter_pass.h"
  50. #include "graph/passes/for_pass.h"
  51. #include "graph/passes/guarantee_const_pass.h"
  52. #include "graph/passes/hccl_memcpy_pass.h"
  53. #include "graph/passes/hccl_group_pass.h"
  54. #include "graph/passes/identity_pass.h"
  55. #include "graph/passes/infershape_pass.h"
  56. #include "graph/passes/infer_value_range_pass.h"
  57. #include "graph/passes/merge_pass.h"
  58. #include "graph/passes/net_output_pass.h"
  59. #include "graph/passes/no_use_reshape_remove_pass.h"
  60. #include "graph/passes/parallel_concat_start_op_pass.h"
  61. #include "graph/passes/placeholder_with_default_pass.h"
  62. #include "graph/passes/prevent_gradient_pass.h"
  63. #include "graph/passes/print_op_pass.h"
  64. #include "graph/passes/prune_pass.h"
  65. #include "graph/passes/replace_transshape_pass.h"
  66. #include "graph/passes/replace_with_empty_const_pass.h"
  67. #include "graph/passes/resource_pair_add_control_pass.h"
  68. #include "graph/passes/resource_pair_remove_control_pass.h"
  69. #include "graph/passes/save_pass.h"
  70. #include "graph/passes/shape_operate_op_remove_pass.h"
  71. #include "graph/passes/snapshot_pass.h"
  72. #include "graph/passes/stop_gradient_pass.h"
  73. #include "graph/passes/switch_dead_branch_elimination.h"
  74. #include "graph/passes/unused_const_pass.h"
  75. #include "graph/passes/var_is_initialized_op_pass.h"
  76. #include "graph/passes/variable_prepare_op_pass.h"
  77. #include "graph/passes/mark_force_unknown_for_cond_pass.h"
  78. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  79. #include "graph/utils/type_utils.h"
  80. #include "inc/pass_manager.h"
  81. #include "init/gelib.h"
  82. #include "multi_batch_copy_graph.h"
  83. #include "graph/passes/data_pass.h"
  84. #include "graph/passes/mark_agnostic_pass.h"
  85. namespace ge {
  86. namespace {
  87. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  88. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  89. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  90. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  91. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  92. // the size of user defined output datatype or format string after split by ":".
  93. const size_t kUserDefinedElementCount = 2;
  94. const int kDataOutIndex = 0;
  95. const int64_t kInvalidDynaimcDimsType = -1;
  96. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  97. GeTensorPtr tensor = MakeShared<GeTensor>();
  98. if (tensor == nullptr) {
  99. REPORT_CALL_ERROR("E19999", "New GeTensor failed");
  100. GELOGE(INTERNAL_ERROR, "[New][GeTensor] failed");
  101. return nullptr;
  102. }
  103. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  104. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  105. auto dst_ge_shape = data_tensor.GetShape();
  106. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  107. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  108. tensor->MutableTensorDesc().SetShape(GeShape());
  109. int32_t dst_shape = 1;
  110. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  111. REPORT_CALL_ERROR("E19999", "Set data to tensor failed");
  112. GELOGE(INTERNAL_ERROR, "[Set][Data] to tensor failed");
  113. return nullptr;
  114. }
  115. } else {
  116. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  117. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  118. if (dst_shape == nullptr) {
  119. REPORT_CALL_ERROR("E19999", "Malloc buffer failed, size:%zu", dim_cnt);
  120. GELOGE(INTERNAL_ERROR, "[Malloc][Buffer] failed, size:%zu", dim_cnt);
  121. return nullptr;
  122. }
  123. for (int64_t i = 0; i < dim_cnt; ++i) {
  124. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  125. }
  126. GE_IF_BOOL_EXEC(
  127. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  128. REPORT_CALL_ERROR("E19999", "Set data to tensor failed");
  129. GELOGE(INTERNAL_ERROR, "[Set][Data] to tensor failed");
  130. return nullptr;)
  131. }
  132. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  133. return OpDescUtils::CreateConstOp(tensor);
  134. }
  135. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  136. OpDescPtr &op_desc) {
  137. // For format transfer node, the IR definition has src/dst format attrs
  138. if (node_type == TRANSDATA) {
  139. GE_IF_BOOL_EXEC(
  140. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  141. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  142. GE_IF_BOOL_EXEC(
  143. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  144. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  145. }
  146. // For TransposeD node, the IR definition has perm attrs
  147. if (node_type == TRANSPOSED) {
  148. Format src_format = input.GetFormat();
  149. Format dst_format = output.GetFormat();
  150. std::vector<int64_t> perm_arg;
  151. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  152. "Get perm by foramt failed.");
  153. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  154. "SetStr PERMUTE_ATTR_PERM failed")
  155. }
  156. // For cast node, the IR definition has src/dst attrs
  157. if (node_type == CAST) {
  158. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  159. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  160. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  161. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  162. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  163. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  164. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  165. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  166. }
  167. }
  168. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  169. const GeTensorDesc &output, NodePtr &node) {
  170. if (node == nullptr) {
  171. REPORT_INNER_ERROR("E19999", "Param node is nullptr, trans_name:%s, trans_type:%s, check invalid",
  172. name.c_str(), node_type.c_str());
  173. GELOGE(PARAM_INVALID, "[Check][Param] Param node is nullptr, trans_name:%s, trans_type:%s",
  174. name.c_str(), node_type.c_str());
  175. return nullptr;
  176. }
  177. auto graph = node->GetOwnerComputeGraph();
  178. if (graph == nullptr) {
  179. REPORT_INNER_ERROR("E19999", "Owner graph in node is nullptr, trans_name:%s, trans_type:%s, check invalid",
  180. name.c_str(), node_type.c_str());
  181. GELOGE(PARAM_INVALID, "[Get][OwnerGraph] in node is nullptr, trans_name:%s, trans_type:%s",
  182. name.c_str(), node_type.c_str());
  183. return nullptr;
  184. }
  185. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  186. if (index < 0) {
  187. ErrorManager::GetInstance().ATCReportErrMessage(
  188. "E19025", {"situation", "reason"},
  189. {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()});
  190. GELOGE(INTERNAL_ERROR, "[Check][Param] The trans node type %s does not exists", node_type.c_str());
  191. return nullptr;
  192. }
  193. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  194. if (op_desc == nullptr) {
  195. REPORT_CALL_ERROR("E19999", "New OpDesc failed, trans_name:%s, trans_type:%s",
  196. name.c_str(), node_type.c_str());
  197. GELOGE(INTERNAL_ERROR, "[New][OpDesc] failed, trans_name:%s, trans_type:%s",
  198. name.c_str(), node_type.c_str());
  199. return nullptr;
  200. }
  201. // for data dump
  202. GE_IF_BOOL_EXEC(
  203. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  204. GELOGW("CreateTransNode: SetListStr failed");)
  205. // Default single input and single output
  206. auto ret = op_desc->AddInputDesc(input);
  207. if (ret != GRAPH_SUCCESS) {
  208. REPORT_CALL_ERROR("E19999", "Add input desc into op:%s(%s) failed",
  209. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  210. GELOGE(INTERNAL_ERROR, "[Add][InputDesc] into op:%s(%s) failed",
  211. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  212. return nullptr;
  213. }
  214. ret = op_desc->AddOutputDesc(output);
  215. if (ret != GRAPH_SUCCESS) {
  216. REPORT_CALL_ERROR("E19999", "Add output desc into op:%s(%s) failed",
  217. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  218. GELOGE(INTERNAL_ERROR, "[Add][OutputDesc] into op:%s(%s) failed",
  219. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  220. return nullptr;
  221. }
  222. AddTransNodeAttr(node_type, input, output, op_desc);
  223. NodePtr shape_node = nullptr;
  224. if (node_type == RESHAPE) {
  225. auto shape_desc = CreateTensorShape(output);
  226. if (shape_desc == nullptr) {
  227. GELOGE(INTERNAL_ERROR, "[Create][TensorShape] Failed to add shape for reshape %s",
  228. node->GetName().c_str());
  229. return nullptr;
  230. }
  231. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  232. if (ret != GRAPH_SUCCESS) {
  233. REPORT_CALL_ERROR("E19999", "Add input desc into op:%s(%s) failed",
  234. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  235. GELOGE(INTERNAL_ERROR, "[Add][InputDesc] into op:%s(%s) failed",
  236. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  237. return nullptr;
  238. }
  239. shape_node = graph->AddNode(shape_desc);
  240. if (shape_node == nullptr) {
  241. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed",
  242. shape_desc->GetName().c_str(), shape_desc->GetType().c_str(),
  243. graph->GetName().c_str());
  244. GELOGE(INTERNAL_ERROR, "[Add][Node] %s(%s) to graph:%s failed",
  245. shape_desc->GetName().c_str(), shape_desc->GetType().c_str(), graph->GetName().c_str());
  246. return nullptr;
  247. }
  248. }
  249. auto trans_node = graph->AddNode(op_desc);
  250. if (trans_node == nullptr) {
  251. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed",
  252. op_desc->GetName().c_str(), op_desc->GetType().c_str(),
  253. graph->GetName().c_str());
  254. GELOGE(INTERNAL_ERROR, "[Add][Node] %s(%s) to graph:%s failed",
  255. op_desc->GetName().c_str(), op_desc->GetType().c_str(), graph->GetName().c_str());
  256. return nullptr;
  257. }
  258. if (node_type == RESHAPE) {
  259. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  260. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:1) failed",
  261. shape_node->GetName().c_str(), shape_node->GetType().c_str(),
  262. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  263. GELOGE(INTERNAL_ERROR, "[Add][Edge] between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:1) failed",
  264. shape_node->GetName().c_str(), shape_node->GetType().c_str(),
  265. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  266. return nullptr;
  267. }
  268. }
  269. return trans_node;
  270. }
  271. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  272. NodePtr &trans_node) {
  273. GE_CHECK_NOTNULL(node);
  274. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  275. if (trans_node == nullptr) {
  276. return INTERNAL_ERROR;
  277. }
  278. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  279. if (ret != GRAPH_SUCCESS) {
  280. REPORT_CALL_ERROR("E19999", "Replace out anchors of node:%s(%s) by node:%s(%s) failed",
  281. node->GetName().c_str(), node->GetType().c_str(),
  282. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  283. GELOGE(INTERNAL_ERROR, "[Replace][OutAnchors] of node:%s(%s) by node:%s(%s) failed",
  284. node->GetName().c_str(), node->GetType().c_str(),
  285. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  286. return INTERNAL_ERROR;
  287. }
  288. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  289. if (ret != GRAPH_SUCCESS) {
  290. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  291. node->GetName().c_str(), node->GetType().c_str(),
  292. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  293. GELOGE(INTERNAL_ERROR, "[Add][Edge] between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  294. node->GetName().c_str(), node->GetType().c_str(),
  295. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  296. return INTERNAL_ERROR;
  297. }
  298. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  299. if (ret != GRAPH_SUCCESS) {
  300. REPORT_CALL_ERROR("E19999", "Move out control edges from node:%s(%s) to node:%s(%s) failed",
  301. node->GetName().c_str(), node->GetType().c_str(),
  302. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  303. GELOGE(INTERNAL_ERROR, "[MoveOut][ControlEdges] from node:%s(%s) to node:%s(%s) failed",
  304. node->GetName().c_str(), node->GetType().c_str(),
  305. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  306. return INTERNAL_ERROR;
  307. }
  308. return SUCCESS;
  309. }
  310. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  311. NodePtr &trans_node) {
  312. GE_CHECK_NOTNULL(node);
  313. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  314. if (trans_node == nullptr) {
  315. return INTERNAL_ERROR;
  316. }
  317. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  318. if (ret != GRAPH_SUCCESS) {
  319. REPORT_CALL_ERROR("E19999", "Replace out anchors of node:%s(%s) by node:%s(%s) failed",
  320. node->GetName().c_str(), node->GetType().c_str(),
  321. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  322. GELOGE(INTERNAL_ERROR, "[Replace][OutAnchors] of node:%s(%s) by node:%s(%s) failed",
  323. node->GetName().c_str(), node->GetType().c_str(),
  324. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  325. return INTERNAL_ERROR;
  326. }
  327. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  328. if (ret != GRAPH_SUCCESS) {
  329. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  330. trans_node->GetName().c_str(), trans_node->GetType().c_str(),
  331. node->GetName().c_str(), node->GetType().c_str());
  332. GELOGE(INTERNAL_ERROR, "[Add][Edge] between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  333. trans_node->GetName().c_str(), trans_node->GetType().c_str(),
  334. node->GetName().c_str(), node->GetType().c_str());
  335. return INTERNAL_ERROR;
  336. }
  337. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  338. if (ret != GRAPH_SUCCESS) {
  339. REPORT_CALL_ERROR("E19999", "Move in control edges from node:%s(%s) to node:%s(%s) failed",
  340. node->GetName().c_str(), node->GetType().c_str(),
  341. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  342. GELOGE(INTERNAL_ERROR, "[MoveIn][CtrlEdges] from node:%s(%s) to node:%s(%s) failed",
  343. node->GetName().c_str(), node->GetType().c_str(),
  344. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  345. return INTERNAL_ERROR;
  346. }
  347. return SUCCESS;
  348. }
  349. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  350. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  351. GE_CHECK_NOTNULL(var->GetOpDesc());
  352. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  353. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  354. output_desc.SetFormat(tensor_desc.GetFormat());
  355. output_desc.SetDataType(tensor_desc.GetDataType());
  356. output_desc.SetShape(tensor_desc.GetShape());
  357. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  358. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  359. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  360. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  361. REPORT_CALL_ERROR("E19999", "Update output desc of node:%s(%s) failed, index:0,",
  362. var->GetName().c_str(), var->GetType().c_str());
  363. GELOGE(INTERNAL_ERROR, "[Update][OutputDesc] of node:%s(%s) failed, index:0,",
  364. var->GetName().c_str(), var->GetType().c_str());
  365. return INTERNAL_ERROR;);
  366. }
  367. if (var->GetOpDesc()->GetInputsSize() > 0) {
  368. auto desc = var->GetOpDesc()->GetInputDesc(0);
  369. desc.SetFormat(tensor_desc.GetFormat());
  370. desc.SetDataType(tensor_desc.GetDataType());
  371. desc.SetShape(tensor_desc.GetShape());
  372. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  373. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  374. desc.SetOriginShape(tensor_desc.GetOriginShape());
  375. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  376. REPORT_CALL_ERROR("E19999", "Update input desc of node:%s(%s) failed, index:0,",
  377. var->GetName().c_str(), var->GetType().c_str());
  378. GELOGE(INTERNAL_ERROR, "[Update][InputDesc] of node:%s(%s) failed, index:0,",
  379. var->GetName().c_str(), var->GetType().c_str());
  380. return INTERNAL_ERROR;)
  381. }
  382. return SUCCESS;
  383. }
  384. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  385. GE_CHECK_NOTNULL(var);
  386. int index = 0;
  387. NodePtr last_node = var;
  388. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  389. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  390. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  391. if (ret != SUCCESS) {
  392. ErrorManager::GetInstance().ATCReportErrMessage(
  393. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  394. GELOGE(INTERNAL_ERROR, "[Recover][TransNode] for variable %s, index %d, type %s", var->GetName().c_str(),
  395. index, iter->node_type.c_str());
  396. return INTERNAL_ERROR;
  397. }
  398. // set stream_label
  399. OpDescPtr var_desc = var->GetOpDesc();
  400. GE_CHECK_NOTNULL(var_desc);
  401. std::string stream_label;
  402. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  403. if (!stream_label.empty()) {
  404. auto status = SetStreamLabel(last_node, stream_label);
  405. if (status != ge::SUCCESS) {
  406. REPORT_CALL_ERROR("E19999", "Set stream_label:%s to op:%s(%s) failed",
  407. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  408. GELOGE(status, "[Set][StreamLabel] %s to op:%s(%s) failed.",
  409. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  410. return status;
  411. }
  412. }
  413. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  414. REPORT_CALL_ERROR("E19999", "Set Attr:%s to node:%s(%s) failed",
  415. ge::ATTR_INSERTED_BY_GE.c_str(),
  416. last_node->GetName().c_str(), last_node->GetType().c_str());
  417. return INTERNAL_ERROR,
  418. "[Set][Attr] %s to node:%s(%s) failed", ge::ATTR_INSERTED_BY_GE.c_str(),
  419. last_node->GetName().c_str(), last_node->GetType().c_str());
  420. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  421. }
  422. if (road.empty()) {
  423. return SUCCESS;
  424. }
  425. return UpdateVarFormats(var, road.rbegin()->output);
  426. }
  427. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  428. for (auto &var : nodes) {
  429. GE_CHECK_NOTNULL(var);
  430. int index = 0;
  431. NodePtr last_node = var;
  432. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  433. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  434. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  435. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  436. if (ret != SUCCESS) {
  437. ErrorManager::GetInstance().ATCReportErrMessage(
  438. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  439. GELOGE(INTERNAL_ERROR, "[Recover][TransNode] for variable %s failed, index %d, type %s",
  440. var->GetName().c_str(), index, iter->node_type.c_str());
  441. return INTERNAL_ERROR;
  442. }
  443. // set stream_label
  444. OpDescPtr var_desc = var->GetOpDesc();
  445. GE_CHECK_NOTNULL(var_desc);
  446. std::string stream_label;
  447. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  448. if (!stream_label.empty()) {
  449. auto status = SetStreamLabel(last_node, stream_label);
  450. if (status != ge::SUCCESS) {
  451. REPORT_CALL_ERROR("E19999", "Set stream_label:%s to op:%s(%s) failed",
  452. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  453. GELOGE(status, "[Set][StreamLabel] %s to op:%s(%s) failed.",
  454. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  455. return status;
  456. }
  457. }
  458. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  459. REPORT_CALL_ERROR("E19999", "Set Attr:%s of node:%s(%s) failed",
  460. ge::ATTR_INSERTED_BY_GE.c_str(),
  461. last_node->GetName().c_str(), last_node->GetType().c_str());
  462. return INTERNAL_ERROR,
  463. "[Set][Attr] %s of node:%s(%s) failed", ge::ATTR_INSERTED_BY_GE.c_str(),
  464. last_node->GetName().c_str(), last_node->GetType().c_str());
  465. }
  466. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  467. return INTERNAL_ERROR;
  468. }
  469. }
  470. return SUCCESS;
  471. }
  472. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  473. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  474. VarNamesToRefs names_to_refs;
  475. std::string var_name;
  476. if (graph == nullptr) {
  477. REPORT_INNER_ERROR("E19999", "Param graph is nullptr, check invalid");
  478. GELOGE(PARAM_INVALID, "[Check][Param] graph is nullptr.");
  479. return names_to_refs;
  480. }
  481. for (auto &node : graph->GetAllNodes()) {
  482. if (node->GetType() != VARIABLE) {
  483. continue;
  484. }
  485. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  486. (void)names_to_refs[var_name].insert(node);
  487. }
  488. }
  489. return names_to_refs;
  490. }
  491. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  492. std::vector<int64_t> &dst_shape) {
  493. if (src_format == FORMAT_NCHW) {
  494. formats::FormatTransferNchwNc1hwc0 transfer;
  495. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  496. GELOGE(INTERNAL_ERROR, "[Trans][Shape] failed");
  497. return FAILED;
  498. }
  499. } else if (src_format == FORMAT_NHWC) {
  500. formats::FormatTransferNhwcNc1hwc0 transfer;
  501. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  502. GELOGE(INTERNAL_ERROR, "[Trans][Shape] failed");
  503. return FAILED;
  504. }
  505. }
  506. return SUCCESS;
  507. }
  508. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  509. GE_CHECK_NOTNULL(node_ptr);
  510. auto op_desc = node_ptr->GetOpDesc();
  511. GE_CHECK_NOTNULL(op_desc);
  512. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  513. GE_CHECK_NOTNULL(input);
  514. ge::Format old_format = input->GetFormat();
  515. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  516. ge::DataType dt = input->GetDataType();
  517. std::vector<int64_t> dst_shape_dims;
  518. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  519. REPORT_CALL_ERROR("E19999", "Transfer shape to NC1HWC0 failed, op:%s(%s),",
  520. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  521. GELOGE(INTERNAL_ERROR, "[Transfer][Shape] to NC1HWC0 failed, op:%s(%s),",
  522. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  523. return FAILED;
  524. }
  525. input->SetFormat(FORMAT_NC1HWC0);
  526. input->SetShape(ge::GeShape(dst_shape_dims));
  527. auto output = op_desc->MutableOutputDesc(0);
  528. GE_CHECK_NOTNULL(output);
  529. output->SetFormat(FORMAT_NC1HWC0);
  530. output->SetShape(ge::GeShape(dst_shape_dims));
  531. int64_t size = 0;
  532. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  533. if (graph_status != ge::GRAPH_SUCCESS) {
  534. REPORT_CALL_ERROR("E19999", "Get output tensor size failed, op:%s(%s), index:0",
  535. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  536. GELOGE(graph_status, "[Get][TensorSize] In Bytes failed, op:%s(%s), index:0",
  537. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  538. return FAILED;
  539. }
  540. ge::TensorUtils::SetSize(*output, size);
  541. ge::TensorUtils::SetSize(*input, size);
  542. return SUCCESS;
  543. }
  544. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  545. GE_CHECK_NOTNULL(input_output);
  546. ge::Format old_format = input_output->GetFormat();
  547. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  548. ge::DataType dt = input_output->GetDataType();
  549. std::vector<int64_t> dst_shape_dims;
  550. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  551. GELOGE(INTERNAL_ERROR, "[Trans][Shape] to NC1HWC0 failed");
  552. return FAILED;
  553. }
  554. input_output->SetFormat(FORMAT_NC1HWC0);
  555. input_output->SetShape(ge::GeShape(dst_shape_dims));
  556. return SUCCESS;
  557. }
  558. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  559. vector<int64_t> &dst_shape_dims) {
  560. GE_CHECK_NOTNULL(op_desc);
  561. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  562. GE_CHECK_NOTNULL(input);
  563. ge::Format old_format = input->GetFormat();
  564. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  565. input->SetShape(ge::GeShape(dst_shape_dims));
  566. input->SetFormat(storage_format);
  567. auto output = op_desc->MutableOutputDesc(index);
  568. GE_CHECK_NOTNULL(output);
  569. output->SetShape(ge::GeShape(dst_shape_dims));
  570. output->SetFormat(storage_format);
  571. if (!output->MutableShape().IsUnknownShape()) {
  572. int64_t size = 0;
  573. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  574. if (graph_status != ge::GRAPH_SUCCESS) {
  575. REPORT_CALL_ERROR("E19999", "Get output tensor size failed, op:%s(%s), index:%u",
  576. op_desc->GetName().c_str(), op_desc->GetType().c_str(), index);
  577. GELOGE(graph_status, "[Get][TensorSize] In Bytes failed, op:%s(%s), index:%u",
  578. op_desc->GetName().c_str(), op_desc->GetType().c_str(), index);
  579. return FAILED;
  580. }
  581. ge::TensorUtils::SetSize(*input, size);
  582. ge::TensorUtils::SetSize(*output, size);
  583. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  584. "new_shape:%s, new_format:%s, new_size:%lu",
  585. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  586. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  587. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  588. }
  589. return SUCCESS;
  590. }
  591. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &mbatch_node, int32_t &index) {
  592. is_dynamic_batch = false;
  593. std::string related_node_name;
  594. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  595. if (related_node_name.empty()) {
  596. ErrorManager::GetInstance().ATCReportErrMessage(
  597. "E15002", {"opname", "value", "reason"}, {data_node->GetName(), "flag", "but the value is empty"});
  598. GELOGE(INTERNAL_ERROR, "[Check][Param] The data node %s has switchn node flag, but the value is empty",
  599. data_node->GetName().c_str());
  600. return INTERNAL_ERROR;
  601. }
  602. auto out_data_nodes_anchors = data_node->GetOutDataNodesAndAnchors();
  603. for (const auto &out_data_node_anchor : out_data_nodes_anchors) {
  604. if (out_data_node_anchor.first->GetName() == related_node_name) {
  605. mbatch_node = out_data_node_anchor.first;
  606. index = out_data_node_anchor.second->GetIdx();
  607. break;
  608. }
  609. }
  610. if (mbatch_node == nullptr) {
  611. ErrorManager::GetInstance().ATCReportErrMessage(
  612. "E15002", {"opname", "value", "reason"},
  613. {data_node->GetName(), related_node_name, "but can not find it on the graph"});
  614. GELOGE(INTERNAL_ERROR, "[Check][Param] The data node %s has switchn node %s, but can not find it on the graph",
  615. data_node->GetName().c_str(), related_node_name.c_str());
  616. return INTERNAL_ERROR;
  617. }
  618. is_dynamic_batch = true;
  619. }
  620. return SUCCESS;
  621. }
  622. bool CheckOpType(const NodePtr &node, const std::string type) {
  623. if (node->GetType() == type) {
  624. return true;
  625. }
  626. return false;
  627. }
  628. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  629. auto op = node_ptr->GetOpDesc();
  630. GE_CHECK_NOTNULL(op);
  631. auto inputDescsPtr = op->GetAllInputsDescPtr();
  632. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  633. ge::Format format = ge::FORMAT_ND;
  634. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  635. // process, otherwise fe will insert transdata.
  636. for (auto &inputDescPtr : inputDescsPtr) {
  637. GE_CHECK_NOTNULL(inputDescPtr);
  638. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  639. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  640. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  641. inputDescPtr->SetFormat(format);
  642. inputDescPtr->SetOriginFormat(format);
  643. }
  644. }
  645. for (auto &outputDescPtr : outputDescsPtr) {
  646. GE_CHECK_NOTNULL(outputDescPtr);
  647. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  648. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  649. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  650. outputDescPtr->SetFormat(format);
  651. outputDescPtr->SetOriginFormat(format);
  652. }
  653. }
  654. return SUCCESS;
  655. }
  656. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  657. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  658. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  659. // data and netoutput, and no longer inserts the transnode.
  660. Status ProcessInputDtDynShape(NodePtr &node_ptr, NodePtr &switchn_node, DataType &dt_set) {
  661. GE_CHECK_NOTNULL(node_ptr);
  662. auto op_desc = node_ptr->GetOpDesc();
  663. GE_CHECK_NOTNULL(op_desc);
  664. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  665. GE_CHECK_NOTNULL(input);
  666. ge::DataType src_dtype = input->GetDataType();
  667. if (src_dtype == dt_set) {
  668. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  669. return SUCCESS;
  670. }
  671. input->SetDataType(dt_set);
  672. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  673. GE_CHECK_NOTNULL(output);
  674. output->SetDataType(dt_set);
  675. GeShape shape = input->GetShape();
  676. if (!shape.IsUnknownShape()) {
  677. int64_t input_shape_size = 0;
  678. int64_t output_shape_size = 0;
  679. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  680. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  681. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  682. REPORT_CALL_ERROR("E19999", "Get input tensor size failed, op:%s(%s), index:0",
  683. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  684. GELOGE(GRAPH_FAILED, "[Process][InputOp] Get tensor size of op [%s] failed!", node_ptr->GetName().c_str());
  685. return FAILED;
  686. }
  687. ge::TensorUtils::SetSize(*input, input_shape_size);
  688. ge::TensorUtils::SetSize(*output, output_shape_size);
  689. GELOGI("[Process][InputDynShape] Set input and output size of node [%s] success.", node_ptr->GetName().c_str());
  690. }
  691. return SUCCESS;
  692. }
  693. Status UpdateInputOutputDataType(NodePtr &mbatch_node, DataType &dt_set, int32_t index) {
  694. auto mbatch_desc = mbatch_node->GetOpDesc();
  695. GE_CHECK_NOTNULL(mbatch_desc);
  696. auto mbatch_input = mbatch_desc->MutableInputDesc(index);
  697. GE_CHECK_NOTNULL(mbatch_input);
  698. mbatch_input->SetDataType(dt_set);
  699. if (mbatch_node->GetType() == SWITCHN) {
  700. for (uint32_t i = 0; i < mbatch_node->GetAllOutDataAnchorsSize(); ++i) {
  701. const GeTensorDescPtr &mbatch_output = mbatch_desc->MutableOutputDesc(i);
  702. GE_CHECK_NOTNULL(mbatch_output);
  703. mbatch_output->SetDataType(dt_set);
  704. }
  705. }
  706. GELOGD("Update input and output data type of node[name: %s, type: %s, input index: %d] to %s.",
  707. mbatch_node->GetName().c_str(), mbatch_node->GetType().c_str(), index,
  708. TypeUtils::DataTypeToSerialString(dt_set).c_str());
  709. return SUCCESS;
  710. }
  711. Status UpdateSubgraphDataOfCase(NodePtr &mbatch_node, DataType &dt_set, int32_t index) {
  712. if (mbatch_node->GetType() != CASE) {
  713. return SUCCESS;
  714. }
  715. auto subgraphs = NodeUtils::GetAllSubgraphs(*mbatch_node);
  716. for (const auto &subgraph : subgraphs) {
  717. GE_CHECK_NOTNULL(subgraph);
  718. for (auto &sub_node : subgraph->GetDirectNode()) {
  719. GE_CHECK_NOTNULL(sub_node);
  720. if (sub_node->GetType() != DATA) {
  721. continue;
  722. }
  723. auto data_desc = sub_node->GetOpDesc();
  724. GE_CHECK_NOTNULL(data_desc);
  725. int32_t parent_node_index = 0;
  726. if (!AttrUtils::GetInt(data_desc, ATTR_NAME_PARENT_NODE_INDEX, parent_node_index) ||
  727. (parent_node_index != index)) {
  728. continue;
  729. }
  730. auto data_input = data_desc->MutableInputDesc(0);
  731. GE_CHECK_NOTNULL(data_input);
  732. data_input->SetDataType(dt_set);
  733. auto data_output = data_desc->MutableOutputDesc(0);
  734. GE_CHECK_NOTNULL(data_output);
  735. data_output->SetDataType(dt_set);
  736. GELOGD("Update input and output data type of node[name: %s, type: %s, parent_node_index: %d] in subgraph %s "
  737. "to %s.", data_desc->GetName().c_str(), data_desc->GetType().c_str(), parent_node_index,
  738. subgraph->GetName().c_str(), TypeUtils::DataTypeToSerialString(dt_set).c_str());
  739. }
  740. }
  741. return SUCCESS;
  742. }
  743. Status ProcessMbatchScene(NodePtr &mbatch_node, DataType &dt_set, int32_t index) {
  744. GELOGI("The node [%s] dtype set fp16.", mbatch_node->GetName().c_str());
  745. if (UpdateInputOutputDataType(mbatch_node, dt_set, index) != SUCCESS) {
  746. GELOGE(FAILED, "[Update][InputOutputDataType] of node[name: %s, type: %s] to %s failed.",
  747. mbatch_node->GetName().c_str(), mbatch_node->GetType().c_str(),
  748. TypeUtils::DataTypeToSerialString(dt_set).c_str());
  749. return FAILED;
  750. }
  751. if (UpdateSubgraphDataOfCase(mbatch_node, dt_set, index) != SUCCESS) {
  752. GELOGE(FAILED, "[Update][SubgraphDataOfCase] node[parent_node_index:%d] in subgraphs of "
  753. "node[name:%s, type:%s] to %s failed.", index, mbatch_node->GetName().c_str(),
  754. mbatch_node->GetType().c_str(), TypeUtils::DataTypeToSerialString(dt_set).c_str());
  755. return FAILED;
  756. }
  757. return SUCCESS;
  758. }
  759. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  760. GE_CHECK_NOTNULL(node_ptr);
  761. auto op_desc = node_ptr->GetOpDesc();
  762. GE_CHECK_NOTNULL(op_desc);
  763. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  764. GE_CHECK_NOTNULL(input);
  765. ge::Format old_format = input->GetFormat();
  766. ge::GeShape old_shape = input->GetShape();
  767. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  768. if (!support) {
  769. ErrorManager::GetInstance().ATCReportErrMessage(
  770. "E19014", {"opname", "value", "reason"},
  771. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  772. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  773. GELOGE(INTERNAL_ERROR, "[Check][Param] The format [%s] is unsupported, op:%s",
  774. TypeUtils::FormatToSerialString(old_format).c_str(), op_desc->GetName().c_str());
  775. return FAILED;
  776. }
  777. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  778. GELOGE(INTERNAL_ERROR, "[Modify][InputFormatAndShape] failed, op:%s(%s)",
  779. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  780. return FAILED;
  781. }
  782. if (is_dynamic_batch) {
  783. auto switchn_op_desc = switchn_node->GetOpDesc();
  784. GE_CHECK_NOTNULL(switchn_op_desc);
  785. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  786. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  787. REPORT_CALL_ERROR("E19999", "Modify format and shape of input:0 in op:%s(%s) failed",
  788. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  789. GELOGE(INTERNAL_ERROR, "[Modify][FormatAndShape] of input:0 in op:%s(%s) failed",
  790. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  791. return FAILED;
  792. }
  793. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  794. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  795. GE_CHECK_NOTNULL(switchn_output);
  796. old_format = switchn_output->GetFormat();
  797. old_shape = switchn_output->GetShape();
  798. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  799. REPORT_CALL_ERROR("E19999", "Modify format and shape of output:%u in op:%s(%s) failed", i,
  800. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  801. GELOGE(INTERNAL_ERROR, "[Modify][FormatAndShape] of output:%u in op:%s(%s) failed", i,
  802. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  803. return FAILED;
  804. }
  805. }
  806. }
  807. return SUCCESS;
  808. }
  809. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  810. auto op_desc = node_ptr->GetOpDesc();
  811. GE_CHECK_NOTNULL(op_desc);
  812. string set_dt_str;
  813. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  814. return SUCCESS;
  815. }
  816. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  817. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  818. bool is_dynamic_batch = false;
  819. NodePtr mbatch_node = nullptr;
  820. int32_t index = 0;
  821. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, mbatch_node, index)) {
  822. GELOGE(INTERNAL_ERROR, "[Call][CheckIfDynamicBatchScene] failed, op:%s", op_desc->GetName().c_str());
  823. return FAILED;
  824. }
  825. if (ProcessInputDtDynShape(node_ptr, mbatch_node, dt_set) != SUCCESS) {
  826. GELOGE(INTERNAL_ERROR, "[Process][InputDtDynShape] ProcessInputFP16 failed, op:%s", op_desc->GetName().c_str());
  827. return FAILED;
  828. }
  829. if (is_dynamic_batch && ProcessMbatchScene(mbatch_node, dt_set, index) != SUCCESS) {
  830. GELOGE(INTERNAL_ERROR, "[Process][MbatchScene] failed");
  831. return FAILED;
  832. }
  833. // check if need to set format
  834. string set_format;
  835. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  836. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  837. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  838. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, mbatch_node) != SUCCESS) {
  839. GELOGE(INTERNAL_ERROR, "[Process][InputNC1HWC0] failed, op:%s", node_ptr->GetName().c_str());
  840. return FAILED;
  841. }
  842. }
  843. return SUCCESS;
  844. }
  845. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  846. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  847. GE_CHECK_NOTNULL(op_desc);
  848. GE_CHECK_NOTNULL(tensor_desc_ptr);
  849. storage_format = FORMAT_RESERVED;
  850. int64_t format = FORMAT_RESERVED;
  851. dst_shape_dims.clear();
  852. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  853. storage_format = static_cast<Format>(format);
  854. vector<int32_t> storage_shape;
  855. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  856. for (auto dim : storage_shape) {
  857. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  858. }
  859. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  860. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  861. formats::JoinToString(storage_shape).c_str());
  862. } else {
  863. ErrorManager::GetInstance().ATCReportErrMessage(
  864. "15003", {"opname", "format"},
  865. {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)});
  866. GELOGE(PARAM_INVALID, "[Check][Param] Update node by storage format failed, storage_shape not set. "
  867. "node:[%s], storage_format [%s]",
  868. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  869. return FAILED;
  870. }
  871. ge::Format old_format = tensor_desc_ptr->GetFormat();
  872. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  873. if (old_format == storage_format && old_shape == dst_shape_dims) {
  874. GELOGI("Update node by storage format, not changed.");
  875. storage_format = FORMAT_RESERVED;
  876. return SUCCESS;
  877. }
  878. }
  879. return SUCCESS;
  880. }
  881. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  882. NodePtr &node) {
  883. bool is_dynamic = CheckOpType(node, MERGE);
  884. auto src_op_desc = node->GetOpDesc();
  885. GE_CHECK_NOTNULL(src_op_desc);
  886. ge::GeShape src_shape = src_desc.GetShape();
  887. ge::Format src_format = src_desc.GetFormat();
  888. net_output_input_desc->SetDataType(DT_FLOAT16);
  889. if (is_dynamic) {
  890. auto merge_output = src_op_desc->MutableOutputDesc(0);
  891. GE_CHECK_NOTNULL(merge_output);
  892. merge_output->SetDataType(DT_FLOAT16);
  893. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  894. auto merge_input = src_op_desc->MutableInputDesc(i);
  895. GE_CHECK_NOTNULL(merge_input);
  896. merge_input->SetDataType(DT_FLOAT16);
  897. }
  898. }
  899. std::vector<int64_t> dst_shape_dims;
  900. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  901. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  902. REPORT_CALL_ERROR("E19999", "Transfer output:0 shape of op:%s(%s) to NC1HWC0 format failed, shape:%s, format:%s",
  903. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str(),
  904. src_shape.ToString().c_str(), TypeUtils::FormatToSerialString(src_format).c_str());
  905. GELOGE(INTERNAL_ERROR, "[Trans][Shape] of op:%s(%s) to NC1HWC0 format failed, shape:%s, format:%s",
  906. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str(),
  907. src_shape.ToString().c_str(), TypeUtils::FormatToSerialString(src_format).c_str());
  908. return FAILED;
  909. }
  910. ge::GeShape dst_shape(dst_shape_dims);
  911. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  912. net_output_input_desc->SetShape(dst_shape);
  913. if (is_dynamic) {
  914. auto merge_out = src_op_desc->MutableOutputDesc(0);
  915. GE_CHECK_NOTNULL(merge_out);
  916. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  917. REPORT_CALL_ERROR("E19999", "Modify format and shape of output:0 in op:%s(%s) failed",
  918. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  919. GELOGE(INTERNAL_ERROR, "[Modify][FormatAndShape] of output:0 in op:%s(%s) failed",
  920. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  921. return FAILED;
  922. }
  923. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  924. auto merge_in = src_op_desc->MutableInputDesc(i);
  925. GE_CHECK_NOTNULL(merge_in);
  926. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  927. REPORT_CALL_ERROR("E19999", "Modify format and shape of input:%u in op:%s(%s) failed", i,
  928. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  929. GELOGE(INTERNAL_ERROR, "[Modify][FormatAndShape] of input:%u in op:%s(%s) failed", i,
  930. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  931. return FAILED;
  932. }
  933. }
  934. }
  935. return SUCCESS;
  936. }
  937. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  938. GE_CHECK_NOTNULL(netout_desc);
  939. vector<string> output_dt_str;
  940. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  941. for (auto dt_str : output_dt_str) {
  942. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  943. if (dt_str_split.size() == kUserDefinedElementCount) {
  944. if (dt_str_split[0] == to_string(index)) {
  945. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  946. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  947. TypeUtils::DataTypeToSerialString(dt).c_str());
  948. return true;
  949. }
  950. }
  951. }
  952. }
  953. return false;
  954. }
  955. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  956. GE_CHECK_NOTNULL(netout_desc);
  957. vector<string> output_format_str;
  958. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  959. for (auto format_str : output_format_str) {
  960. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  961. if (format_str_split.size() == kUserDefinedElementCount) {
  962. if (format_str_split[0] == to_string(index)) {
  963. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  964. return true;
  965. }
  966. }
  967. }
  968. }
  969. return false;
  970. }
  971. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  972. auto op_desc = node->GetOpDesc();
  973. GE_CHECK_NOTNULL(op_desc);
  974. ge::DataType output_data_type = ge::DT_FLOAT;
  975. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  976. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  977. auto peer_out = in_anchor->GetPeerOutAnchor();
  978. GE_CHECK_NOTNULL(peer_out);
  979. auto src_node = peer_out->GetOwnerNode();
  980. GE_CHECK_NOTNULL(src_node);
  981. bool is_dynamic = CheckOpType(src_node, MERGE);
  982. OpDescPtr src_op_desc = src_node->GetOpDesc();
  983. GE_CHECK_NOTNULL(src_op_desc);
  984. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  985. GE_CHECK_NOTNULL(net_output_input_desc);
  986. ge::GeShape old_shape = net_output_input_desc->GetShape();
  987. ge::Format old_format = net_output_input_desc->GetFormat();
  988. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  989. // Update datatype
  990. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  991. GELOGI("Enter into process output_type schedule");
  992. net_output_input_desc->SetDataType(output_data_type);
  993. if (is_dynamic) {
  994. auto merge_output = src_op_desc->MutableOutputDesc(0);
  995. GE_CHECK_NOTNULL(merge_output);
  996. merge_output->SetDataType(output_data_type);
  997. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  998. auto merge_input = src_op_desc->MutableInputDesc(i);
  999. GE_CHECK_NOTNULL(merge_input);
  1000. merge_input->SetDataType(output_data_type);
  1001. }
  1002. }
  1003. }
  1004. // check if is_output_adjust_hw_layout is set
  1005. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  1006. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  1007. ErrorManager::GetInstance().ATCReportErrMessage(
  1008. "E19014", {"opname", "value", "reason"},
  1009. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  1010. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  1011. GELOGE(INTERNAL_ERROR, "[Check][Param] Format is not one of NCHW, NHWC, NC1HWC0.");
  1012. return FAILED;
  1013. }
  1014. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  1015. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  1016. GELOGE(INTERNAL_ERROR, "[Process][NetOutput] fp16 nc1hwc0 failed.");
  1017. return FAILED;
  1018. }
  1019. }
  1020. }
  1021. return SUCCESS;
  1022. }
  1023. Status GetDynamicInputShapeRange(const std::vector<GeTensor> &user_input, const std::map<string, string> &graph_option,
  1024. vector<vector<std::pair<int64_t, int64_t>>> &range_vec) {
  1025. // check both mode and shape_range option are all enabled
  1026. auto mode_iter = graph_option.find(OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  1027. bool enable_dynamic_execute_mode = (mode_iter != graph_option.end()) && (mode_iter->second == "dynamic_execute");
  1028. if (!enable_dynamic_execute_mode) {
  1029. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  1030. OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  1031. }
  1032. auto iter = graph_option.find(OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1033. bool enable_input_shape_range = (iter != graph_option.end()) && (!iter->second.empty());
  1034. if (!enable_input_shape_range) {
  1035. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  1036. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1037. }
  1038. if (enable_dynamic_execute_mode && enable_input_shape_range) {
  1039. GELOGD("GraphOption: %s value is dynamic_execute, %s value is %s.", OPTION_EXEC_DYNAMIC_EXECUTE_MODE,
  1040. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE, iter->second.c_str());
  1041. } else if (!enable_dynamic_execute_mode && !enable_input_shape_range) {
  1042. return SUCCESS;
  1043. } else {
  1044. REPORT_INNER_ERROR("E19999", "Graph option: %s and %s should be enabled at the same time, check invalid",
  1045. OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1046. GELOGE(PARAM_INVALID, "[Check][Param] Graph option: %s and %s should be enabled at the same time.",
  1047. OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1048. return PARAM_INVALID;
  1049. }
  1050. if (ParseInputShapeRange(iter->second, range_vec) != SUCCESS) {
  1051. GELOGE(PARAM_INVALID, "[Parse][ShapeRange] Parse dynamic input shape range failed.");
  1052. return PARAM_INVALID;
  1053. }
  1054. if (range_vec.size() != user_input.size()) {
  1055. GELOGE(PARAM_INVALID, "[Check][Param] Dynamic input shape range size is %zu, inputs size is %zu. Not match.",
  1056. range_vec.size(), user_input.size());
  1057. return PARAM_INVALID;
  1058. }
  1059. return SUCCESS;
  1060. }
  1061. Status UpdateDynamicInputShapeRange(const ge::GeAttrValue::INT index,
  1062. const vector<vector<std::pair<int64_t, int64_t>>> &range_vec, OpDescPtr &op,
  1063. GeTensorDesc &desc) {
  1064. auto origin_shape = desc.GetShape();
  1065. auto current_shape_range_vec = range_vec.at(index);
  1066. if (current_shape_range_vec.size() != origin_shape.GetDimNum()) {
  1067. REPORT_INNER_ERROR("E19999", "Given shape_range dim num is %zu, current dim:%s num is %zu, not match, "
  1068. "check invalid", current_shape_range_vec.size(), origin_shape.ToString().c_str(),
  1069. origin_shape.GetDimNum());
  1070. GELOGE(PARAM_INVALID, "[Check][Param] Given shape_range dim num is %zu, current dim num is %zu, "
  1071. "not match.Pleace Check.", current_shape_range_vec.size(), origin_shape.GetDimNum());
  1072. return PARAM_INVALID;
  1073. }
  1074. for (size_t i = 0; i < origin_shape.GetDimNum(); ++i) {
  1075. auto curr_dim = origin_shape.GetDim(i);
  1076. auto left_range = current_shape_range_vec.at(i).first;
  1077. auto right_range = current_shape_range_vec.at(i).second;
  1078. if (left_range == right_range) {
  1079. // given shape_range is known dim, check is same as origin or not
  1080. if (curr_dim != left_range) {
  1081. REPORT_INNER_ERROR("E19999", "Given shape range is %ld, current dim shape is %ld, not match, dim_index:%zu, "
  1082. "check invalid", left_range, curr_dim, i);
  1083. GELOGE(PARAM_INVALID, "[Check][Param] Given shape range is %ld, current dim shape is %ld, "
  1084. "not match.Pleace Check.", left_range, curr_dim);
  1085. return PARAM_INVALID;
  1086. }
  1087. origin_shape.SetDim(i, left_range);
  1088. } else {
  1089. // given shape_range is fix range, check input_shape is in this range or not
  1090. if (right_range != UNKNOWN_DIM) {
  1091. if ((curr_dim < left_range) || (curr_dim > right_range)) {
  1092. REPORT_INNER_ERROR("E19999", "Given shape range is [%ld~%ld], current dim shape is %ld, out of range, "
  1093. "dim_index:%zu, check invalid",
  1094. left_range, right_range, curr_dim, i);
  1095. GELOGE(PARAM_INVALID, "[Check][Param] Given shape range is [%ld~%ld], current dim shape is %ld, "
  1096. "out of range.Pleace Check.", left_range, right_range, curr_dim);
  1097. return PARAM_INVALID;
  1098. }
  1099. }
  1100. origin_shape.SetDim(i, UNKNOWN_DIM);
  1101. }
  1102. }
  1103. desc.SetShape(origin_shape);
  1104. desc.SetShapeRange(current_shape_range_vec);
  1105. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1106. GE_CHK_GRAPH_STATUS_RET(graph_ret, "[Update][InputDesc] fail, graph ret: %u", graph_ret);
  1107. graph_ret = op->UpdateOutputDesc(0, desc);
  1108. GE_CHK_GRAPH_STATUS_RET(graph_ret, "[Update][OutputDesc] fail, graph ret: %u", graph_ret);
  1109. return SUCCESS;
  1110. }
  1111. } // namespace
  1112. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  1113. GraphPrepare::~GraphPrepare() {}
  1114. /**
  1115. * @param graph
  1116. * @return
  1117. */
  1118. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  1119. GE_CHECK_NOTNULL(graph);
  1120. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  1121. for (auto &node : graph->GetAllNodes()) {
  1122. if (node == nullptr) {
  1123. continue;
  1124. }
  1125. if (node->GetType() != VARIABLE) {
  1126. continue;
  1127. }
  1128. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  1129. if (trans_road == nullptr) {
  1130. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  1131. continue;
  1132. }
  1133. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  1134. auto ret = RecoverTransRoadForVar(node, *trans_road);
  1135. if (ret != SUCCESS) {
  1136. GELOGE(INTERNAL_ERROR, "[Recover][TransRoad] for var %s failed", node->GetName().c_str());
  1137. return INTERNAL_ERROR;
  1138. }
  1139. auto iter = var_names_to_refs.find(node->GetName());
  1140. if (iter != var_names_to_refs.end()) {
  1141. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  1142. if (ret != SUCCESS) {
  1143. GELOGE(INTERNAL_ERROR, "[Recover][TransRoad] for var ref %s failed", node->GetName().c_str());
  1144. return INTERNAL_ERROR;
  1145. }
  1146. }
  1147. }
  1148. return SUCCESS;
  1149. }
  1150. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  1151. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  1152. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  1153. if (compute_graph_ != nullptr) {
  1154. compute_graph_->SetSessionID(session_id);
  1155. }
  1156. session_id_ = session_id;
  1157. Status ret = CheckGraph();
  1158. if (ret != SUCCESS) {
  1159. GELOGE(ret, "[Check][Graph] fail, ret:%u", ret);
  1160. return ret;
  1161. }
  1162. (void)compute_graph_->TopologicalSorting();
  1163. ret = CheckRefOp();
  1164. if (ret != SUCCESS) {
  1165. GELOGE(ret, "[Check][RefOp] fail, ret:%u", ret);
  1166. return ret;
  1167. }
  1168. return SUCCESS;
  1169. }
  1170. Status GraphPrepare::CheckGraph() {
  1171. if (compute_graph_ == nullptr) {
  1172. REPORT_INNER_ERROR("E19999", "compute_graph_ is nullptr, check invalid");
  1173. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][Param] compute_graph_ is nullptr");
  1174. return GE_GRAPH_INIT_FAILED;
  1175. }
  1176. auto nodes = compute_graph_->GetAllNodes();
  1177. if (nodes.empty()) {
  1178. REPORT_INNER_ERROR("E19999", "nodes in graph is empty, check invalid");
  1179. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][Param] Invalid graph, no nodes in this graph.");
  1180. return GE_GRAPH_INIT_FAILED;
  1181. }
  1182. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  1183. GE_CHECK_NOTNULL(node);
  1184. if (node->GetOpDesc() == nullptr) {
  1185. REPORT_INNER_ERROR("E19999", "node without opdesc exist in graph, check invalid");
  1186. GELOGE(GE_GRAPH_INIT_FAILED, "[Get][OpDesc] failed, Check Graph node opdesc is NULL");
  1187. return GE_GRAPH_INIT_FAILED;
  1188. }
  1189. }
  1190. return SUCCESS;
  1191. }
  1192. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  1193. const std::set<NodePtr> &ref_nodes) {
  1194. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  1195. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  1196. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  1197. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  1198. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  1199. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH,
  1200. ge::DATA};
  1201. GE_CHECK_NOTNULL(node);
  1202. const auto &op_desc = node->GetOpDesc();
  1203. GE_CHECK_NOTNULL(op_desc);
  1204. const auto input_index = op_desc->GetInputIndexByName(input_name);
  1205. const auto &in_anchor = node->GetInDataAnchor(input_index);
  1206. GE_CHECK_NOTNULL(in_anchor);
  1207. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  1208. GE_CHECK_NOTNULL(peer_out_anchor);
  1209. const auto &input_node = peer_out_anchor->GetOwnerNode();
  1210. GE_CHECK_NOTNULL(input_node);
  1211. const auto &input_op_desc = input_node->GetOpDesc();
  1212. GE_CHECK_NOTNULL(input_op_desc);
  1213. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  1214. if (is_ref) {
  1215. return SUCCESS;
  1216. }
  1217. auto input_type = input_op_desc->GetType();
  1218. if (input_type == ge::FRAMEWORKOP) {
  1219. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  1220. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed",
  1221. ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE.c_str(),
  1222. input_op_desc->GetName().c_str(), input_op_desc->GetType().c_str());
  1223. GELOGE(PARAM_INVALID, "[Get][Attr] %s of op:%s(%s) failed", ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE.c_str(),
  1224. input_op_desc->GetName().c_str(), input_op_desc->GetType().c_str());
  1225. return PARAM_INVALID;
  1226. }
  1227. }
  1228. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  1229. if (!is_acceptable) {
  1230. ErrorManager::GetInstance().ATCReportErrMessage(
  1231. "E15005", {"opname", "optype", "opname1", "optype1"},
  1232. {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()});
  1233. GELOGE(PARAM_INVALID, "[Check][Param] The ref input of ref node %s[%s] must be ref node or variable, "
  1234. "but %s[%s]isn't.", node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  1235. input_op_desc->GetType().c_str());
  1236. return PARAM_INVALID;
  1237. }
  1238. return SUCCESS;
  1239. }
  1240. Status GraphPrepare::CheckRefOp() {
  1241. GE_CHECK_NOTNULL(compute_graph_);
  1242. std::set<NodePtr> ref_nodes;
  1243. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  1244. if (node == nullptr) {
  1245. REPORT_INNER_ERROR("E19999", "nullptr node exist in graph, check invalid");
  1246. GELOGE(PARAM_INVALID, "[Check][Param] param [node] must not be null.");
  1247. return PARAM_INVALID;
  1248. }
  1249. auto op_desc = node->GetOpDesc();
  1250. if (op_desc == nullptr) {
  1251. REPORT_INNER_ERROR("E19999", "node without opdesc exist in graph, check invalid");
  1252. GELOGE(PARAM_INVALID, "[Check][Param] OpDesc of param [node] must not be null.");
  1253. return PARAM_INVALID;
  1254. }
  1255. auto input_name_index = op_desc->GetAllInputName();
  1256. auto outputs = op_desc->GetAllOutputName();
  1257. for (const auto &name_index : input_name_index) {
  1258. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  1259. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  1260. GELOGE(PARAM_INVALID, "[Check][RefInputNode] failed, node:%s.", op_desc->GetName().c_str());
  1261. return PARAM_INVALID;
  1262. }
  1263. (void)ref_nodes.insert(node); // no need to check value
  1264. }
  1265. }
  1266. }
  1267. return SUCCESS;
  1268. };
  1269. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  1270. GE_CHECK_NOTNULL(compute_graph_);
  1271. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  1272. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  1273. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  1274. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  1275. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  1276. return SUCCESS;
  1277. }
  1278. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1279. if (node == nullptr) {
  1280. REPORT_INNER_ERROR("E19999", "Param node is nullptr, check invalid");
  1281. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "[Check][Param] Input node is nullptr");
  1282. return GE_GRAPH_GRAPH_NODE_NULL;
  1283. }
  1284. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1285. if (op_desc_ptr == nullptr) {
  1286. REPORT_INNER_ERROR("E19999", "Param node's op_desc is nullptr, check invalid");
  1287. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "[Get][OpDesc] Input node opdesc is NULL");
  1288. return GE_GRAPH_GRAPH_NODE_NULL;
  1289. }
  1290. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1291. GeShape output_shape = output.GetShape();
  1292. if (output_shape.IsUnknownShape()) {
  1293. GELOGD("[Adjust][DataOpOutput] Shape of op [%s] output is unknown.", node->GetName().c_str());
  1294. return SUCCESS;
  1295. }
  1296. int64_t tensor_size = 0;
  1297. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1298. if (graph_status != GRAPH_SUCCESS) {
  1299. ErrorManager::GetInstance().ATCReportErrMessage(
  1300. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1301. GELOGE(graph_status, "[Call][GetTensorMemorySizeInBytes] failed, op:%s", node->GetName().c_str());
  1302. return FAILED;
  1303. }
  1304. TensorUtils::SetSize(output, tensor_size);
  1305. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1306. if (graph_ret != GRAPH_SUCCESS) {
  1307. REPORT_CALL_ERROR("E19999", "Update output desc of op:%s(%s) failed, index:0",
  1308. op_desc_ptr->GetName().c_str(), op_desc_ptr->GetType().c_str());
  1309. GELOGE(graph_ret, "[Update][OutputDesc] of op:%s(%s) failed, index:0",
  1310. op_desc_ptr->GetName().c_str(), op_desc_ptr->GetType().c_str());
  1311. return graph_ret;
  1312. }
  1313. return SUCCESS;
  1314. }
  1315. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input,
  1316. const std::map<string, string> &graph_option) {
  1317. // Get shape range of input in dynamic_execute mode
  1318. vector<vector<std::pair<int64_t, int64_t>>> dynamic_shape_range_vec;
  1319. auto ret = GetDynamicInputShapeRange(user_input, graph_option, dynamic_shape_range_vec);
  1320. GE_CHK_STATUS_RET(ret, "[Get][DynamicInputShapeRange] failed, Graph option is not right on Dynamic execute mode.");
  1321. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1322. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1323. GE_CHECK_NOTNULL(input_node);
  1324. OpDescPtr op = input_node->GetOpDesc();
  1325. GE_CHECK_NOTNULL(op);
  1326. if (op->GetType() == DATA) {
  1327. GeAttrValue::INT index = 0;
  1328. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1329. GELOGW("Get index from data attr failed");
  1330. continue;
  1331. }
  1332. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1333. std::string situation = "data op index[" + std::to_string(index) + "]";
  1334. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1335. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1336. GELOGE(PARAM_INVALID, "[Check][Param] user_input size = %zu, graph data op index = %ld.",
  1337. user_input.size(), index);
  1338. return FAILED;
  1339. }
  1340. if (IsDynamicDims(input_node)) {
  1341. continue;
  1342. }
  1343. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1344. auto format = desc.GetFormat();
  1345. auto origin_format = desc.GetOriginFormat();
  1346. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1347. auto tune_flag = (options_.build_mode == BUILD_MODE_TUNING) && (options_.build_step == BUILD_STEP_AFTER_BUILDER);
  1348. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op) && (!tune_flag);
  1349. if (need_check_internal_format) {
  1350. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1351. if (is_internal) {
  1352. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1353. {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" +
  1354. TypeUtils::FormatToSerialString(origin_format) + "]", "it is not support"});
  1355. GELOGE(PARAM_INVALID, "[Check][Param] Input format %s or origin_format %s is not support.",
  1356. TypeUtils::FormatToSerialString(format).c_str(),
  1357. TypeUtils::FormatToSerialString(origin_format).c_str());
  1358. return FAILED;
  1359. }
  1360. }
  1361. auto data_type = desc.GetDataType();
  1362. uint32_t length = 1;
  1363. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1364. if (!type_ret) {
  1365. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1366. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1367. GELOGE(PARAM_INVALID, "[Check][Param] Input datatype %s is not support.",
  1368. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1369. return FAILED;
  1370. }
  1371. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1372. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1373. int64_t shape_size = desc_shape * length;
  1374. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1375. int64_t size = 0;
  1376. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1377. REPORT_CALL_ERROR("E19999", "Get size of user input tensor failed, index:%ld", index);
  1378. GELOGE(INTERNAL_ERROR, "[Get][Size] of user input tensor failed, index:%ld", index);
  1379. return FAILED);
  1380. bool size_check = (size != 0 && shape_size != size);
  1381. if (size_check) {
  1382. std::string situation = "input data size[" + std::to_string(size) +
  1383. "] and shape_size[" + std::to_string(size) + "]";
  1384. std::string reason = "because size != 0 and shape_size != size";
  1385. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1386. GELOGE(PARAM_INVALID, "[Check][Param] input data size = %ld, shape_size = %ld.", size, shape_size);
  1387. return FAILED;
  1388. }
  1389. ge::TensorUtils::SetSize(desc, shape_size);
  1390. if (!tune_flag) {
  1391. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1392. if (graph_ret != GRAPH_SUCCESS) {
  1393. REPORT_CALL_ERROR("E19999", "Update input desc of op:%s(%s) failed, index:0",
  1394. op->GetName().c_str(), op->GetType().c_str());
  1395. GELOGE(graph_ret, "[Update][InputDesc] of op:%s(%s) failed, index:0",
  1396. op->GetName().c_str(), op->GetType().c_str());
  1397. return graph_ret;
  1398. }
  1399. // Size will be recalculated in the build stage
  1400. ge::TensorUtils::SetSize(desc, 0);
  1401. graph_ret = op->UpdateOutputDesc(0, desc);
  1402. if (graph_ret != GRAPH_SUCCESS) {
  1403. REPORT_CALL_ERROR("E19999", "Update output desc of op:%s(%s) failed, index:0",
  1404. op->GetName().c_str(), op->GetType().c_str());
  1405. GELOGE(graph_ret, "[Update][OutputDesc] of op:%s(%s) failed, index:0",
  1406. op->GetName().c_str(), op->GetType().c_str());
  1407. return graph_ret;
  1408. }
  1409. } else {
  1410. GELOGI("data %s skip update info in tune mode", op->GetName().c_str());
  1411. }
  1412. if (!dynamic_shape_range_vec.empty()) {
  1413. ret = UpdateDynamicInputShapeRange(index, dynamic_shape_range_vec, op, desc);
  1414. GE_CHK_STATUS_RET(ret, "[Update][DynamicInputShapeRange] on %s failed.", op->GetName().c_str());
  1415. continue;
  1416. }
  1417. if (!options_.train_graph_flag) {
  1418. Status ret = AdjustDataOpOutput(input_node);
  1419. GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(ret, "[Adjust][DataOpOutput] fail, ret:%u", ret); return ret);
  1420. }
  1421. }
  1422. }
  1423. return SUCCESS;
  1424. }
  1425. Status GraphPrepare::TryDoAipp() {
  1426. // infer and with aipp configure file, then call aipp insert
  1427. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1428. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1429. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1430. if (ret != SUCCESS) {
  1431. GELOGE(INTERNAL_ERROR, "[Init][InsertNewOpUtil] failed.");
  1432. return INTERNAL_ERROR;
  1433. }
  1434. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1435. if (ret != SUCCESS) {
  1436. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "[Parse][ConfigFile] %s failed",
  1437. options_.insert_op_file.c_str());
  1438. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1439. }
  1440. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1441. if (ret != SUCCESS) {
  1442. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "[Insert][AippOps] failed, ret:%u", ret);
  1443. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1444. }
  1445. }
  1446. return SUCCESS;
  1447. }
  1448. Status GraphPrepare::FormatAndShapeProcess() {
  1449. Status ret = ResourcePairProcess("add");
  1450. if (ret != SUCCESS) {
  1451. GELOGE(ret, "ResourcePairProcess failed");
  1452. return ret;
  1453. }
  1454. GE_TIMESTAMP_START(InferOriginFormat1);
  1455. ret = compute_graph_->InferOriginFormat();
  1456. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1457. GE_DUMP(compute_graph_, "after_first_inferformat");
  1458. if (ret != SUCCESS) {
  1459. GELOGE(ret, "[Call][InferOriginFormat] Prepare Graph first inferformat failed");
  1460. return ret;
  1461. }
  1462. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1463. ret = InferShapeForPreprocess();
  1464. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1465. GE_DUMP(compute_graph_, "after_infershape");
  1466. if (ret != SUCCESS) {
  1467. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "[Call][InferShapeForPreprocess] Prepare Graph infershape failed");
  1468. return GE_GRAPH_INFERSHAPE_FAILED;
  1469. }
  1470. GE_TIMESTAMP_START(InferOriginFormat2);
  1471. ret = compute_graph_->InferOriginFormat();
  1472. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1473. if (ret != SUCCESS) {
  1474. GELOGE(ret, "[Call][InferOriginFormat] Prepare Graph inferformat failed");
  1475. return ret;
  1476. }
  1477. ret = ResourcePairProcess("remove");
  1478. if (ret != SUCCESS) {
  1479. return ret;
  1480. }
  1481. return ret;
  1482. }
  1483. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1484. PassManager control_pass;
  1485. // Graph pass tmp logic for resource infershape
  1486. if (options_.train_graph_flag) {
  1487. try {
  1488. if (action == "add") {
  1489. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1490. } else {
  1491. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1492. new ResourcePairRemoveControlPass);
  1493. }
  1494. } catch (std::bad_alloc &e) {
  1495. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add ResourcePair Pass");
  1496. GELOGE(INTERNAL_ERROR, "[Add][Pass] failed, bad memory allocation occur, action:%s.", action.c_str());
  1497. return INTERNAL_ERROR;
  1498. }
  1499. }
  1500. Status ret = control_pass.Run(compute_graph_);
  1501. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1502. GELOGE(ret, "[Run][ResourcePairControlPass] failed, action:%s, ret:%u.", action.c_str(), ret);
  1503. return ret;
  1504. }
  1505. return SUCCESS;
  1506. }
  1507. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1508. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1509. GE_CHECK_NOTNULL(node_ptr);
  1510. if (node_ptr->GetType() == DATA) {
  1511. uint32_t index = 0;
  1512. auto op_desc = node_ptr->GetOpDesc();
  1513. GE_CHECK_NOTNULL(op_desc);
  1514. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1515. Format storage_format = FORMAT_RESERVED;
  1516. vector<int64_t> dst_shape_dims;
  1517. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1518. GELOGE(INTERNAL_ERROR, "[Get][StorageFormatAndShape] for input failed, op:%s, index:0",
  1519. op_desc->GetName().c_str());
  1520. return FAILED;
  1521. }
  1522. if (storage_format == FORMAT_RESERVED) {
  1523. continue;
  1524. }
  1525. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1526. GELOGE(INTERNAL_ERROR, "[Modify][DataNetOutputFormatAndShape] for input failed, op:%s, index:0",
  1527. op_desc->GetName().c_str());
  1528. return FAILED;
  1529. }
  1530. }
  1531. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1532. auto op_desc = node_ptr->GetOpDesc();
  1533. GE_CHECK_NOTNULL(op_desc);
  1534. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1535. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1536. Format storage_format = FORMAT_RESERVED;
  1537. vector<int64_t> dst_shape_dims;
  1538. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1539. GELOGE(INTERNAL_ERROR, "[Get][StorageFormatAndShape] from output failed, op:%s, index:%u",
  1540. op_desc->GetName().c_str(), index);
  1541. return FAILED;
  1542. }
  1543. if (storage_format == FORMAT_RESERVED) {
  1544. continue;
  1545. }
  1546. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1547. GELOGE(INTERNAL_ERROR, "[Modify][DataNetOutputFormatAndShape] for output failed, op:%s, index:%u",
  1548. op_desc->GetName().c_str(), index);
  1549. return FAILED;
  1550. }
  1551. }
  1552. }
  1553. }
  1554. return SUCCESS;
  1555. }
  1556. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1557. if (options_.save_original_model == "true") {
  1558. ModelHelper model_helper;
  1559. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1560. options_.original_model_file);
  1561. if (ret != SUCCESS) {
  1562. // If save original model fail, process continue
  1563. GELOGW("SaveOriginalGraphToOmModel fail");
  1564. }
  1565. }
  1566. return SUCCESS;
  1567. }
  1568. #define PP_RUN_AND_DUMP(name, func, ...) \
  1569. do { \
  1570. GE_RUN(Prepare, func, __VA_ARGS__); \
  1571. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1572. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1573. } while (0)
  1574. #define PP_RUN(name, func, ...) \
  1575. do { \
  1576. GE_RUN(Prepare, func, __VA_ARGS__); \
  1577. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1578. } while (0)
  1579. Status GraphPrepare::PrepareDynShape(const GraphNodePtr &graph_node, const std::vector<GeTensor> &user_input,
  1580. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1581. GE_CHECK_NOTNULL(graph_node->GetGraph());
  1582. GE_CHECK_NOTNULL(compute_graph);
  1583. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1584. const Graph &const_graph = *graph_node->GetGraph();
  1585. PP_RUN("Init", Init, const_graph, session_id);
  1586. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1587. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input, graph_node->GetOptions());
  1588. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1589. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1590. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1591. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1592. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1593. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1594. PP_RUN_AND_DUMP("CtrlFlowPreProcess", CtrlFlowPreProcess);
  1595. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1596. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1597. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1598. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1599. return SUCCESS;
  1600. }
  1601. Status GraphPrepare::CtrlFlowPreProcess() {
  1602. PassManager graph_pass;
  1603. // After InferShape Mark v1 control flow for unknown shape.
  1604. auto mark_force_unknown_pass = new (std::nothrow) MarkForceUnknownForCondPass;
  1605. GE_CHK_STATUS_RET(graph_pass.AddPass("PreRun::MarkForceUnknownForCondPass", mark_force_unknown_pass));
  1606. GE_CHK_STATUS_RET(graph_pass.Run(compute_graph_));
  1607. return SUCCESS;
  1608. }
  1609. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1610. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1611. return SUCCESS;
  1612. }
  1613. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1614. auto compute_graph = compute_graph_;
  1615. PassManager pass_manager;
  1616. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1617. new (std::nothrow) VariablePrepareOpPass))
  1618. GE_TIMESTAMP_START(pass_manager);
  1619. auto ret = pass_manager.Run(compute_graph);
  1620. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1621. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1622. GELOGE(ret, "[Run][Passes] for running format refiner failed, ret:%u.", ret);
  1623. return ret;
  1624. }
  1625. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1626. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1627. return SUCCESS;
  1628. }
  1629. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1630. if (compute_graph == nullptr) {
  1631. REPORT_INNER_ERROR("E19999", "Param compute_graph is nullptr, check invalid");
  1632. GELOGE(GE_GRAPH_NULL_INPUT, "[Check][Param] Input Graph is NULL");
  1633. return GE_GRAPH_NULL_INPUT;
  1634. }
  1635. GEPass ge_passes(compute_graph);
  1636. NamesToPass hccl_group;
  1637. HcclGroupPass hccl_group_pass;
  1638. GELOGD("Add hccl group pass success");
  1639. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1640. auto ret = ge_passes.Run(hccl_group);
  1641. if (ret != SUCCESS) {
  1642. GELOGE(ret, "[Run][HcclGroupPass] pass for preprocess failed, ret:%u.", ret);
  1643. return ret;
  1644. }
  1645. ret = compute_graph->TopologicalSorting();
  1646. if (ret != SUCCESS) {
  1647. REPORT_CALL_ERROR("E19999", "Topological sorting failed");
  1648. GELOGE(ret, "[Call][TopologicalSorting] Graph topological sort failed, ret:%u.", ret);
  1649. return ret;
  1650. }
  1651. return SUCCESS;
  1652. }
  1653. #undef PP_RUN_AND_DUMP
  1654. #undef PP_RUN
  1655. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1656. if (graph == nullptr) {
  1657. REPORT_INNER_ERROR("E19999", "Param graph is nullptr, check invalid");
  1658. GELOGE(GE_GRAPH_NULL_INPUT, "[Check][Param] Input Graph is NULL");
  1659. return GE_GRAPH_NULL_INPUT;
  1660. }
  1661. const Graph &const_graph = *graph;
  1662. Status ret = Init(const_graph, 0);
  1663. if (ret != SUCCESS) {
  1664. GELOGE(ret, "[Init][GraphPrepare] fail, ret:%u", ret);
  1665. return ret;
  1666. }
  1667. GE_DUMP(compute_graph_, "after_parser");
  1668. GELOGI("Start infershape for dump json process.");
  1669. ret = compute_graph_->InferOriginFormat();
  1670. GE_DUMP(compute_graph_, "after_inferformat");
  1671. if (ret != SUCCESS) {
  1672. REPORT_CALL_ERROR("E19999", "Infer OriginFormat failed");
  1673. GELOGE(ret, "[Infer][OriginFormat] failed");
  1674. return ret;
  1675. }
  1676. InferShapePass infer_shape_pass;
  1677. NamesToPass names_to_passes;
  1678. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1679. GEPass ge_passes(compute_graph_);
  1680. ret = ge_passes.Run(names_to_passes);
  1681. GE_DUMP(compute_graph_, "after_infershape");
  1682. if (ret != SUCCESS) {
  1683. GELOGE(ret, "[Run][GePasses] infershape for preprocess failed, ret:%u.", ret);
  1684. return ret;
  1685. }
  1686. ShapeRefiner::ClearContextMap();
  1687. return SUCCESS;
  1688. }
  1689. Status GraphPrepare::CheckConstOp() {
  1690. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1691. GE_CHECK_NOTNULL(node_ptr);
  1692. if (node_ptr->GetType() == CONSTANT) {
  1693. Status ret = VerifyConstOp(node_ptr);
  1694. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1695. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1696. auto op_desc = node_ptr->GetOpDesc();
  1697. if (op_desc == nullptr) {
  1698. REPORT_INNER_ERROR("E19999", "op_desc is nullptr, check invalid");
  1699. GELOGE(PARAM_INVALID, "[Get][OpDesc] of node failed, op_desc is nullptr, node type:FRAMEWORKOP.");
  1700. return PARAM_INVALID;
  1701. }
  1702. std::string original_type;
  1703. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1704. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1705. GELOGI("original type is %s", original_type.c_str());
  1706. if (original_type == CONSTANT) {
  1707. Status ret = VerifyConstOp(node_ptr);
  1708. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1709. }
  1710. }
  1711. }
  1712. return SUCCESS;
  1713. }
  1714. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1715. GE_CHECK_NOTNULL(node);
  1716. auto op_desc = node->GetOpDesc();
  1717. GE_CHECK_NOTNULL(op_desc);
  1718. ConstGeTensorPtr ge_tensor_ptr;
  1719. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1720. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1721. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  1722. GELOGE(PARAM_INVALID, "[Get][Attr] %s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1723. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  1724. return PARAM_INVALID;
  1725. }
  1726. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1727. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1728. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1729. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1730. auto data_type = ge_tensor_desc.GetDataType();
  1731. if (data_type == DT_STRING) {
  1732. return SUCCESS;
  1733. }
  1734. uint32_t length = 1;
  1735. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1736. if (!type_ret) {
  1737. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1738. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1739. GELOGE(PARAM_INVALID, "[Check][Param] Input datatype %s is not support.",
  1740. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1741. return FAILED;
  1742. }
  1743. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1744. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1745. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1746. if (shape_size == 0) {
  1747. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1748. // shape = [], means it's a sclar tensor.
  1749. GE_CHK_BOOL_EXEC(data_size / length == 1,
  1750. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid scalar tensor."});
  1751. return PARAM_INVALID, "[Check][Param] Const is invalid scalar tensor.");
  1752. } else {
  1753. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1754. GE_CHK_BOOL_EXEC(data_size == 0,
  1755. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid vector scalar."});
  1756. return PARAM_INVALID, "[Check][Param] Const is invalid vector scalar.");
  1757. }
  1758. } else {
  1759. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0,
  1760. ErrorManager::GetInstance().ATCReportErrMessage(
  1761. "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"});
  1762. return PARAM_INVALID, "[Check][Param] Const input data size is not equal with tensor desc shape");
  1763. }
  1764. return SUCCESS;
  1765. }
  1766. bool GraphPrepare::IsDynamicDims(const NodePtr &input_node) {
  1767. auto data_shape = NodeUtils::GetOutputDesc(*input_node, kDataOutIndex).GetShape();
  1768. const auto &dims = data_shape.GetDims();
  1769. bool all_is_positive = false;
  1770. if (std::all_of(dims.begin(), dims.end(), [](int64_t val) { return val >= 0; })) {
  1771. all_is_positive = true;
  1772. }
  1773. if (!all_is_positive && !options_.input_shape.empty() && !options_.dynamic_dims.empty() &&
  1774. options_.dynamic_node_type != kInvalidDynaimcDimsType) {
  1775. GELOGI("No need to check and update desc info, the dims of %s is %s.", input_node->GetName().c_str(),
  1776. formats::JoinToString(dims).c_str());
  1777. return true;
  1778. }
  1779. return false;
  1780. }
  1781. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1782. if (GetLocalOmgContext().is_dynamic_input) {
  1783. return SUCCESS;
  1784. }
  1785. unsigned int node_num = 0;
  1786. unsigned int data_num = 0;
  1787. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1788. GE_CHECK_NOTNULL(input_node);
  1789. OpDescPtr op = input_node->GetOpDesc();
  1790. GE_CHECK_NOTNULL(op);
  1791. node_num++;
  1792. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1793. data_num++;
  1794. GeAttrValue::INT index = 0;
  1795. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1796. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1797. op->GetName().c_str(), op->GetType().c_str());
  1798. GELOGE(GE_GRAPH_INIT_FAILED, "[Get][Attr] %s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1799. op->GetName().c_str(), op->GetType().c_str());
  1800. return GE_GRAPH_INIT_FAILED;
  1801. }
  1802. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1803. std::string situation = "data op index[" + std::to_string(index) + "]";
  1804. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1805. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1806. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][Param] user_input size:%zu must larger than data op index:%ld.",
  1807. user_input.size(), index);
  1808. return GE_GRAPH_INIT_FAILED;
  1809. }
  1810. if (IsDynamicDims(input_node)) {
  1811. continue;
  1812. }
  1813. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1814. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1815. int64_t dim = desc.GetShape().GetDim(i);
  1816. if (dim < UNKNOWN_DIM_NUM) {
  1817. std::string situation = "data dim[" + std::to_string(i) + "][" + std::to_string(dim) + "]" ;
  1818. std::string reason = "it need >= -2";
  1819. REPORT_INPUT_ERROR(
  1820. "E19025", std::vector<std::string>({"situation", "reason"}), std::vector<std::string>({situation, reason}));
  1821. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][InputDim]data dim %zu is not supported, need >= -2, real:%ld.", i, dim);
  1822. return GE_GRAPH_INIT_FAILED;
  1823. }
  1824. }
  1825. }
  1826. }
  1827. if (node_num <= data_num) {
  1828. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1829. }
  1830. return SUCCESS;
  1831. }
  1832. Status GraphPrepare::InferShapeForPreprocess() {
  1833. GELOGI("Start infershape for preprocess.");
  1834. GEPass ge_passes(compute_graph_);
  1835. NamesToPass names_to_passes;
  1836. AssertPass assert_pass;
  1837. if (!options_.train_graph_flag) {
  1838. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1839. }
  1840. SwitchDeadBranchElimination switch_dead_branch_elimination;
  1841. names_to_passes.emplace_back("SwitchDeadBranchElimination", &switch_dead_branch_elimination);
  1842. MergePass merge_pass;
  1843. names_to_passes.emplace_back("MergePass", &merge_pass);
  1844. InferShapePass infer_shape_pass;
  1845. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1846. InferValueRangePass infer_value_pass;
  1847. names_to_passes.emplace_back("InferValuePass", &infer_value_pass);
  1848. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1849. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1850. DimensionComputePass dimension_compute_pass;
  1851. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1852. ConstantFoldingPass constant_folding_pass;
  1853. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1854. int32_t dev_count = 0;
  1855. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1856. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1857. rtError_t rt_err = RT_ERROR_NONE;
  1858. if (aicpu_constant_folding_on != nullptr) {
  1859. rt_err = rtGetDeviceCount(&dev_count);
  1860. if (rt_err == RT_ERROR_NONE) {
  1861. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1862. if (result != SUCCESS) {
  1863. GELOGE(result, "[Set][RtContext] failed, mode = RT_CTX_NORMAL_MODE.");
  1864. return result;
  1865. }
  1866. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1867. }
  1868. }
  1869. Status ret = ge_passes.Run(names_to_passes);
  1870. if (aicpu_constant_folding_on != nullptr) {
  1871. if (rt_err == RT_ERROR_NONE) {
  1872. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1873. if (result != SUCCESS) {
  1874. GELOGE(result, "[Set][RtContext] failed, mode = RT_CTX_GEN_MODE.");
  1875. return result;
  1876. }
  1877. }
  1878. }
  1879. ShapeRefiner::ClearContextMap();
  1880. if (ret != SUCCESS) {
  1881. GELOGE(ret, "[Run][GePasses] infershape for preprocess failed, ret:%u.", ret);
  1882. return ret;
  1883. }
  1884. return SUCCESS;
  1885. }
  1886. Status GraphPrepare::PrepareOptimize() {
  1887. GELOGI("Start optimize for preprocess.");
  1888. // check rw type
  1889. GraphOptimize graph_optimize;
  1890. bool has_conflict = false;
  1891. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1892. if (has_conflict) {
  1893. GELOGE(GRAPH_PARAM_INVALID, "[Check][RWConflict] There has rw conflict.Stop optimize.");
  1894. return FAILED;
  1895. }
  1896. PassManager original_graph_passes;
  1897. // Graph pass
  1898. try {
  1899. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1900. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1901. (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass);
  1902. } catch (std::bad_alloc &e) {
  1903. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass");
  1904. GELOGE(INTERNAL_ERROR, "[Add][Pass] failed, bad memory allocation occurs.");
  1905. return INTERNAL_ERROR;
  1906. }
  1907. GE_TIMESTAMP_START(original_graph_passes);
  1908. Status ret = original_graph_passes.Run(compute_graph_);
  1909. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1910. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1911. GELOGE(ret, "[Run][GraphPasses] optimize for preprocess failed, ret:%u.", ret);
  1912. return ret;
  1913. }
  1914. // New pass
  1915. GEPass ge_passes(compute_graph_);
  1916. NamesToPass names_to_passes;
  1917. EnterPass enter_pass;
  1918. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1919. CondPass cond_pass;
  1920. names_to_passes.emplace_back("CondPass", &cond_pass);
  1921. PrintOpPass print_pass;
  1922. if (options_.enable_print_op_pass) {
  1923. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1924. }
  1925. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1926. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1927. DropOutPass dropout_pass;
  1928. AssertPass assert_pass;
  1929. UnusedConstPass unused_const_pass;
  1930. StopGradientPass stop_gradient_pass;
  1931. PreventGradientPass prevent_gradient_pass;
  1932. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1933. GuaranteeConstPass guarantee_const_pass;
  1934. VarIsInitializedOpPass var_is_initialized_pass;
  1935. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1936. IdentityPass identity_pass(false);
  1937. SnapshotPass snapshot_pass;
  1938. if (!options_.train_graph_flag) {
  1939. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1940. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1941. }
  1942. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1943. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1944. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1945. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1946. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1947. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1948. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1949. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1950. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1951. GE_TIMESTAMP_START(names_to_passes);
  1952. ret = ge_passes.Run(names_to_passes);
  1953. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1954. if (ret != SUCCESS) {
  1955. GELOGE(ret, "[Run][GePasses] optimize for preprocess failed, ret:%u.", ret);
  1956. return ret;
  1957. }
  1958. PassManager graph_pass;
  1959. try {
  1960. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1961. // can't move to optimize1/2 directly, may cause more identity insert, cause CI fail
  1962. (void)graph_pass.AddPass("PrepareOptimize::HcclMemcpyPass", new HcclMemcpyPass);
  1963. } catch (std::bad_alloc &e) {
  1964. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass");
  1965. GELOGE(INTERNAL_ERROR, "[Add][Pass] failed, bad memory allocation occurs.");
  1966. return INTERNAL_ERROR;
  1967. }
  1968. GE_TIMESTAMP_START(graph_passes);
  1969. ret = graph_pass.Run(compute_graph_);
  1970. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1971. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1972. GELOGE(ret, "[Run][GraphPasses] optimize for preprocess failed, ret:%u.", ret);
  1973. return ret;
  1974. }
  1975. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1976. TypeConversionOfConstant();
  1977. ret = compute_graph_->TopologicalSorting();
  1978. if (ret != SUCCESS) {
  1979. REPORT_CALL_ERROR("E19999", "Topological sorting failed");
  1980. GELOGE(ret, "[Call][TopologicalSorting] Graph topological sort failed, ret:%u.", ret);
  1981. return ret;
  1982. }
  1983. GELOGI("End optimize for preprocess.");
  1984. return SUCCESS;
  1985. }
  1986. void GraphPrepare::TypeConversionOfConstant() {
  1987. bool is_acl_compile = false;
  1988. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1989. // This can ensure that n is not a null pointer
  1990. // No Conversion when called by aclOpCompile
  1991. (void)AttrUtils::GetBool(n->GetOpDesc(), ATTR_SINGLE_OP_SCENE, is_acl_compile);
  1992. if (is_acl_compile) {
  1993. return;
  1994. }
  1995. }
  1996. if (options_.train_graph_flag) {
  1997. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1998. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1999. // This can ensure that n is not a null pointer
  2000. if (n->GetOpDesc()->GetType() == CONSTANT) {
  2001. n->GetOpDesc()->SetType(CONSTANTOP);
  2002. }
  2003. }
  2004. } else {
  2005. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  2006. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  2007. // This can ensure that n is not a null pointer
  2008. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  2009. n->GetOpDesc()->SetType(CONSTANT);
  2010. }
  2011. }
  2012. }
  2013. }
  2014. Status GraphPrepare::GraphEquivalentTransformation() {
  2015. NamesToPass names_to_pass;
  2016. ForPass for_pass;
  2017. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  2018. return GEPass(compute_graph_).Run(names_to_pass);
  2019. }
  2020. Status GraphPrepare::ProcessBeforeInfershape() {
  2021. NamesToPass names_to_passes;
  2022. CondRemovePass condition_remove_pass;
  2023. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  2024. GE_TIMESTAMP_START(ProcessCondRemove);
  2025. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  2026. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  2027. if (ret != SUCCESS) {
  2028. GELOGE(ret, "[Run][GEPass] optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  2029. return ret;
  2030. }
  2031. return SUCCESS;
  2032. }
  2033. Status GraphPrepare::ProcessNetOutput() {
  2034. PassManager graph_passes_before_infershape;
  2035. try {
  2036. if (options_.train_graph_flag) {
  2037. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  2038. }
  2039. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  2040. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  2041. new (std::nothrow) DataPass); // Add NetOutput first.
  2042. } catch (std::bad_alloc) {
  2043. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass");
  2044. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  2045. return INTERNAL_ERROR;
  2046. }
  2047. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  2048. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  2049. GELOGE(ret, "[Run][GraphPasses] before Infershape failed, ret:%d.", ret);
  2050. return ret;
  2051. }
  2052. return SUCCESS;
  2053. }
  2054. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input,
  2055. const std::map<string, string> &graph_option) {
  2056. compute_graph_->SetInputSize(user_input.size());
  2057. if (user_input.empty()) {
  2058. return SUCCESS;
  2059. }
  2060. auto ret = CheckUserInput(user_input);
  2061. if (ret != SUCCESS) {
  2062. GELOGE(ret, "[Check][UserInput] failed, ret:%u", ret);
  2063. return ret;
  2064. }
  2065. ret = UpdateInput(user_input, graph_option);
  2066. if (ret != SUCCESS) {
  2067. GELOGE(ret, "[Update][Input] fail, ret:%u", ret);
  2068. return ret;
  2069. }
  2070. if (user_input.size() != 0) {
  2071. ret = CheckConstOp();
  2072. if (ret != SUCCESS) {
  2073. GELOGE(ret, "[Check][ConstOp] fail, ret:%u", ret);
  2074. return ret;
  2075. }
  2076. } else {
  2077. ret = compute_graph_->TopologicalSorting();
  2078. if (ret != SUCCESS) {
  2079. REPORT_CALL_ERROR("E19999", "Topological sorting failed");
  2080. GELOGE(ret, "[Call][TopologicalSorting] failed.");
  2081. return FAILED;
  2082. }
  2083. }
  2084. return SUCCESS;
  2085. }
  2086. Status GraphPrepare::UpdateInputOutputByOptions() {
  2087. auto ret = UpdateDataNetOutputByStorageFormat();
  2088. if (ret != SUCCESS) {
  2089. GELOGE(ret, "[Update][DataNetOutputByStorageFormat] failed.");
  2090. return ret;
  2091. }
  2092. if (options_.train_graph_flag) {
  2093. GELOGI("This is train mode, no need to do this schedule.");
  2094. return SUCCESS;
  2095. }
  2096. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  2097. GE_CHECK_NOTNULL(node_ptr);
  2098. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  2099. GELOGE(INTERNAL_ERROR, "[Set][NdFormat] for node:%s failed", node_ptr->GetName().c_str());
  2100. return FAILED;
  2101. }
  2102. if (node_ptr->GetType() == DATA) {
  2103. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  2104. GELOGE(INTERNAL_ERROR, "[Call][ProcessDataNodeDynShape] for node:%s failed", node_ptr->GetName().c_str());
  2105. return FAILED;
  2106. }
  2107. }
  2108. if (node_ptr->GetType() == ge::NETOUTPUT) {
  2109. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  2110. GELOGE(INTERNAL_ERROR, "[Call][ProcessNetoutputNodeDynShape] for node:%s failed", node_ptr->GetName().c_str());
  2111. return FAILED;
  2112. }
  2113. }
  2114. }
  2115. return SUCCESS;
  2116. }
  2117. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  2118. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  2119. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  2120. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  2121. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  2122. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  2123. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  2124. if (dst_node->GetType() == TRANSDATA) {
  2125. return true;
  2126. }
  2127. }
  2128. }
  2129. return false;
  2130. }
  2131. } // 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