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
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292
  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/merge_pass.h"
  57. #include "graph/passes/net_output_pass.h"
  58. #include "graph/passes/no_use_reshape_remove_pass.h"
  59. #include "graph/passes/parallel_concat_start_op_pass.h"
  60. #include "graph/passes/placeholder_with_default_pass.h"
  61. #include "graph/passes/prevent_gradient_pass.h"
  62. #include "graph/passes/print_op_pass.h"
  63. #include "graph/passes/prune_pass.h"
  64. #include "graph/passes/replace_transshape_pass.h"
  65. #include "graph/passes/replace_with_empty_const_pass.h"
  66. #include "graph/passes/resource_pair_add_control_pass.h"
  67. #include "graph/passes/resource_pair_remove_control_pass.h"
  68. #include "graph/passes/save_pass.h"
  69. #include "graph/passes/shape_operate_op_remove_pass.h"
  70. #include "graph/passes/snapshot_pass.h"
  71. #include "graph/passes/stop_gradient_pass.h"
  72. #include "graph/passes/switch_dead_branch_elimination.h"
  73. #include "graph/passes/unused_const_pass.h"
  74. #include "graph/passes/var_is_initialized_op_pass.h"
  75. #include "graph/passes/variable_prepare_op_pass.h"
  76. #include "graph/passes/mark_force_unknown_for_cond_pass.h"
  77. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  78. #include "graph/utils/type_utils.h"
  79. #include "inc/pass_manager.h"
  80. #include "init/gelib.h"
  81. #include "multi_batch_copy_graph.h"
  82. #include "graph/passes/data_pass.h"
  83. #include "graph/passes/mark_agnostic_pass.h"
  84. namespace ge {
  85. namespace {
  86. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  87. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  88. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  89. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  90. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  91. // the size of user defined output datatype or format string after split by ":".
  92. const size_t kUserDefinedElementCount = 2;
  93. const int kDataOutIndex = 0;
  94. const int64_t kInvalidDynaimcDimsType = -1;
  95. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  96. GeTensorPtr tensor = MakeShared<GeTensor>();
  97. if (tensor == nullptr) {
  98. REPORT_CALL_ERROR("E19999", "New GeTensor failed");
  99. GELOGE(INTERNAL_ERROR, "[New][GeTensor] failed");
  100. return nullptr;
  101. }
  102. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  103. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  104. auto dst_ge_shape = data_tensor.GetShape();
  105. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  106. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  107. tensor->MutableTensorDesc().SetShape(GeShape());
  108. int32_t dst_shape = 1;
  109. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  110. REPORT_CALL_ERROR("E19999", "Set data to tensor failed");
  111. GELOGE(INTERNAL_ERROR, "[Set][Data] to tensor failed");
  112. return nullptr;
  113. }
  114. } else {
  115. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  116. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  117. if (dst_shape == nullptr) {
  118. REPORT_CALL_ERROR("E19999", "Malloc buffer failed, size:%zu", dim_cnt);
  119. GELOGE(INTERNAL_ERROR, "[Malloc][Buffer] failed, size:%zu", dim_cnt);
  120. return nullptr;
  121. }
  122. for (int64_t i = 0; i < dim_cnt; ++i) {
  123. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  124. }
  125. GE_IF_BOOL_EXEC(
  126. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  127. REPORT_CALL_ERROR("E19999", "Set data to tensor failed");
  128. GELOGE(INTERNAL_ERROR, "[Set][Data] to tensor failed");
  129. return nullptr;)
  130. }
  131. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  132. return OpDescUtils::CreateConstOp(tensor);
  133. }
  134. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  135. OpDescPtr &op_desc) {
  136. // For format transfer node, the IR definition has src/dst format attrs
  137. if (node_type == TRANSDATA) {
  138. GE_IF_BOOL_EXEC(
  139. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  140. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  141. GE_IF_BOOL_EXEC(
  142. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  143. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  144. }
  145. // For TransposeD node, the IR definition has perm attrs
  146. if (node_type == TRANSPOSED) {
  147. Format src_format = input.GetFormat();
  148. Format dst_format = output.GetFormat();
  149. std::vector<int64_t> perm_arg;
  150. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  151. "Get perm by foramt failed.");
  152. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  153. "SetStr PERMUTE_ATTR_PERM failed")
  154. }
  155. // For cast node, the IR definition has src/dst attrs
  156. if (node_type == CAST) {
  157. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  158. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  159. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  160. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  161. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  162. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  163. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  164. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  165. }
  166. }
  167. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  168. const GeTensorDesc &output, NodePtr &node) {
  169. if (node == nullptr) {
  170. REPORT_INNER_ERROR("E19999", "Param node is nullptr, trans_name:%s, trans_type:%s, check invalid",
  171. name.c_str(), node_type.c_str());
  172. GELOGE(PARAM_INVALID, "[Check][Param] Param node is nullptr, trans_name:%s, trans_type:%s",
  173. name.c_str(), node_type.c_str());
  174. return nullptr;
  175. }
  176. auto graph = node->GetOwnerComputeGraph();
  177. if (graph == nullptr) {
  178. REPORT_INNER_ERROR("E19999", "Owner graph in node is nullptr, trans_name:%s, trans_type:%s, check invalid",
  179. name.c_str(), node_type.c_str());
  180. GELOGE(PARAM_INVALID, "[Get][OwnerGraph] in node is nullptr, trans_name:%s, trans_type:%s",
  181. name.c_str(), node_type.c_str());
  182. return nullptr;
  183. }
  184. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  185. if (index < 0) {
  186. ErrorManager::GetInstance().ATCReportErrMessage(
  187. "E19025", {"situation", "reason"},
  188. {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()});
  189. GELOGE(INTERNAL_ERROR, "[Check][Param] The trans node type %s does not exists", node_type.c_str());
  190. return nullptr;
  191. }
  192. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  193. if (op_desc == nullptr) {
  194. REPORT_CALL_ERROR("E19999", "New OpDesc failed, trans_name:%s, trans_type:%s",
  195. name.c_str(), node_type.c_str());
  196. GELOGE(INTERNAL_ERROR, "[New][OpDesc] failed, trans_name:%s, trans_type:%s",
  197. name.c_str(), node_type.c_str());
  198. return nullptr;
  199. }
  200. // for data dump
  201. GE_IF_BOOL_EXEC(
  202. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  203. GELOGW("CreateTransNode: SetListStr failed");)
  204. // Default single input and single output
  205. auto ret = op_desc->AddInputDesc(input);
  206. if (ret != GRAPH_SUCCESS) {
  207. REPORT_CALL_ERROR("E19999", "Add input desc into op:%s(%s) failed",
  208. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  209. GELOGE(INTERNAL_ERROR, "[Add][InputDesc] into op:%s(%s) failed",
  210. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  211. return nullptr;
  212. }
  213. ret = op_desc->AddOutputDesc(output);
  214. if (ret != GRAPH_SUCCESS) {
  215. REPORT_CALL_ERROR("E19999", "Add output desc into op:%s(%s) failed",
  216. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  217. GELOGE(INTERNAL_ERROR, "[Add][OutputDesc] into op:%s(%s) failed",
  218. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  219. return nullptr;
  220. }
  221. AddTransNodeAttr(node_type, input, output, op_desc);
  222. NodePtr shape_node = nullptr;
  223. if (node_type == RESHAPE) {
  224. auto shape_desc = CreateTensorShape(output);
  225. if (shape_desc == nullptr) {
  226. GELOGE(INTERNAL_ERROR, "[Create][TensorShape] Failed to add shape for reshape %s",
  227. node->GetName().c_str());
  228. return nullptr;
  229. }
  230. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  231. if (ret != GRAPH_SUCCESS) {
  232. REPORT_CALL_ERROR("E19999", "Add input desc into op:%s(%s) failed",
  233. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  234. GELOGE(INTERNAL_ERROR, "[Add][InputDesc] into op:%s(%s) failed",
  235. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  236. return nullptr;
  237. }
  238. shape_node = graph->AddNode(shape_desc);
  239. if (shape_node == nullptr) {
  240. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed",
  241. shape_desc->GetName().c_str(), shape_desc->GetType().c_str(),
  242. graph->GetName().c_str());
  243. GELOGE(INTERNAL_ERROR, "[Add][Node] %s(%s) to graph:%s failed",
  244. shape_desc->GetName().c_str(), shape_desc->GetType().c_str(), graph->GetName().c_str());
  245. return nullptr;
  246. }
  247. }
  248. auto trans_node = graph->AddNode(op_desc);
  249. if (trans_node == nullptr) {
  250. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed",
  251. op_desc->GetName().c_str(), op_desc->GetType().c_str(),
  252. graph->GetName().c_str());
  253. GELOGE(INTERNAL_ERROR, "[Add][Node] %s(%s) to graph:%s failed",
  254. op_desc->GetName().c_str(), op_desc->GetType().c_str(), graph->GetName().c_str());
  255. return nullptr;
  256. }
  257. if (node_type == RESHAPE) {
  258. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  259. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:1) failed",
  260. shape_node->GetName().c_str(), shape_node->GetType().c_str(),
  261. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  262. GELOGE(INTERNAL_ERROR, "[Add][Edge] between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:1) failed",
  263. shape_node->GetName().c_str(), shape_node->GetType().c_str(),
  264. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  265. return nullptr;
  266. }
  267. }
  268. return trans_node;
  269. }
  270. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  271. NodePtr &trans_node) {
  272. GE_CHECK_NOTNULL(node);
  273. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  274. if (trans_node == nullptr) {
  275. return INTERNAL_ERROR;
  276. }
  277. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  278. if (ret != GRAPH_SUCCESS) {
  279. REPORT_CALL_ERROR("E19999", "Replace out anchors of node:%s(%s) by node:%s(%s) failed",
  280. node->GetName().c_str(), node->GetType().c_str(),
  281. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  282. GELOGE(INTERNAL_ERROR, "[Replace][OutAnchors] of node:%s(%s) by node:%s(%s) failed",
  283. node->GetName().c_str(), node->GetType().c_str(),
  284. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  285. return INTERNAL_ERROR;
  286. }
  287. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  288. if (ret != GRAPH_SUCCESS) {
  289. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  290. node->GetName().c_str(), node->GetType().c_str(),
  291. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  292. GELOGE(INTERNAL_ERROR, "[Add][Edge] between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  293. node->GetName().c_str(), node->GetType().c_str(),
  294. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  295. return INTERNAL_ERROR;
  296. }
  297. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  298. if (ret != GRAPH_SUCCESS) {
  299. REPORT_CALL_ERROR("E19999", "Move out control edges from node:%s(%s) to node:%s(%s) failed",
  300. node->GetName().c_str(), node->GetType().c_str(),
  301. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  302. GELOGE(INTERNAL_ERROR, "[MoveOut][ControlEdges] from node:%s(%s) to node:%s(%s) failed",
  303. node->GetName().c_str(), node->GetType().c_str(),
  304. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  305. return INTERNAL_ERROR;
  306. }
  307. return SUCCESS;
  308. }
  309. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  310. NodePtr &trans_node) {
  311. GE_CHECK_NOTNULL(node);
  312. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  313. if (trans_node == nullptr) {
  314. return INTERNAL_ERROR;
  315. }
  316. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  317. if (ret != GRAPH_SUCCESS) {
  318. REPORT_CALL_ERROR("E19999", "Replace out anchors of node:%s(%s) by node:%s(%s) failed",
  319. node->GetName().c_str(), node->GetType().c_str(),
  320. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  321. GELOGE(INTERNAL_ERROR, "[Replace][OutAnchors] of node:%s(%s) by node:%s(%s) failed",
  322. node->GetName().c_str(), node->GetType().c_str(),
  323. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  324. return INTERNAL_ERROR;
  325. }
  326. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  327. if (ret != GRAPH_SUCCESS) {
  328. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  329. trans_node->GetName().c_str(), trans_node->GetType().c_str(),
  330. node->GetName().c_str(), node->GetType().c_str());
  331. GELOGE(INTERNAL_ERROR, "[Add][Edge] between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  332. trans_node->GetName().c_str(), trans_node->GetType().c_str(),
  333. node->GetName().c_str(), node->GetType().c_str());
  334. return INTERNAL_ERROR;
  335. }
  336. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  337. if (ret != GRAPH_SUCCESS) {
  338. REPORT_CALL_ERROR("E19999", "Move in control edges from node:%s(%s) to node:%s(%s) failed",
  339. node->GetName().c_str(), node->GetType().c_str(),
  340. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  341. GELOGE(INTERNAL_ERROR, "[MoveIn][CtrlEdges] from node:%s(%s) to node:%s(%s) failed",
  342. node->GetName().c_str(), node->GetType().c_str(),
  343. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  344. return INTERNAL_ERROR;
  345. }
  346. return SUCCESS;
  347. }
  348. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  349. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  350. GE_CHECK_NOTNULL(var->GetOpDesc());
  351. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  352. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  353. output_desc.SetFormat(tensor_desc.GetFormat());
  354. output_desc.SetDataType(tensor_desc.GetDataType());
  355. output_desc.SetShape(tensor_desc.GetShape());
  356. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  357. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  358. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  359. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  360. REPORT_CALL_ERROR("E19999", "Update output desc of node:%s(%s) failed, index:0,",
  361. var->GetName().c_str(), var->GetType().c_str());
  362. GELOGE(INTERNAL_ERROR, "[Update][OutputDesc] of node:%s(%s) failed, index:0,",
  363. var->GetName().c_str(), var->GetType().c_str());
  364. return INTERNAL_ERROR;);
  365. }
  366. if (var->GetOpDesc()->GetInputsSize() > 0) {
  367. auto desc = var->GetOpDesc()->GetInputDesc(0);
  368. desc.SetFormat(tensor_desc.GetFormat());
  369. desc.SetDataType(tensor_desc.GetDataType());
  370. desc.SetShape(tensor_desc.GetShape());
  371. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  372. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  373. desc.SetOriginShape(tensor_desc.GetOriginShape());
  374. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  375. REPORT_CALL_ERROR("E19999", "Update input desc of node:%s(%s) failed, index:0,",
  376. var->GetName().c_str(), var->GetType().c_str());
  377. GELOGE(INTERNAL_ERROR, "[Update][InputDesc] of node:%s(%s) failed, index:0,",
  378. var->GetName().c_str(), var->GetType().c_str());
  379. return INTERNAL_ERROR;)
  380. }
  381. return SUCCESS;
  382. }
  383. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  384. GE_CHECK_NOTNULL(var);
  385. int index = 0;
  386. NodePtr last_node = var;
  387. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  388. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  389. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  390. if (ret != SUCCESS) {
  391. ErrorManager::GetInstance().ATCReportErrMessage(
  392. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  393. GELOGE(INTERNAL_ERROR, "[Recover][TransNode] for variable %s, index %d, type %s", var->GetName().c_str(),
  394. index, iter->node_type.c_str());
  395. return INTERNAL_ERROR;
  396. }
  397. // set stream_label
  398. OpDescPtr var_desc = var->GetOpDesc();
  399. GE_CHECK_NOTNULL(var_desc);
  400. std::string stream_label;
  401. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  402. if (!stream_label.empty()) {
  403. auto status = SetStreamLabel(last_node, stream_label);
  404. if (status != ge::SUCCESS) {
  405. REPORT_CALL_ERROR("E19999", "Set stream_label:%s to op:%s(%s) failed",
  406. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  407. GELOGE(status, "[Set][StreamLabel] %s to op:%s(%s) failed.",
  408. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  409. return status;
  410. }
  411. }
  412. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  413. REPORT_CALL_ERROR("E19999", "Set Attr:%s to node:%s(%s) failed",
  414. ge::ATTR_INSERTED_BY_GE.c_str(),
  415. last_node->GetName().c_str(), last_node->GetType().c_str());
  416. return INTERNAL_ERROR,
  417. "[Set][Attr] %s to node:%s(%s) failed", ge::ATTR_INSERTED_BY_GE.c_str(),
  418. last_node->GetName().c_str(), last_node->GetType().c_str());
  419. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  420. }
  421. if (road.empty()) {
  422. return SUCCESS;
  423. }
  424. return UpdateVarFormats(var, road.rbegin()->output);
  425. }
  426. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  427. for (auto &var : nodes) {
  428. GE_CHECK_NOTNULL(var);
  429. int index = 0;
  430. NodePtr last_node = var;
  431. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  432. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  433. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  434. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  435. if (ret != SUCCESS) {
  436. ErrorManager::GetInstance().ATCReportErrMessage(
  437. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  438. GELOGE(INTERNAL_ERROR, "[Recover][TransNode] for variable %s failed, index %d, type %s",
  439. var->GetName().c_str(), index, iter->node_type.c_str());
  440. return INTERNAL_ERROR;
  441. }
  442. // set stream_label
  443. OpDescPtr var_desc = var->GetOpDesc();
  444. GE_CHECK_NOTNULL(var_desc);
  445. std::string stream_label;
  446. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  447. if (!stream_label.empty()) {
  448. auto status = SetStreamLabel(last_node, stream_label);
  449. if (status != ge::SUCCESS) {
  450. REPORT_CALL_ERROR("E19999", "Set stream_label:%s to op:%s(%s) failed",
  451. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  452. GELOGE(status, "[Set][StreamLabel] %s to op:%s(%s) failed.",
  453. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  454. return status;
  455. }
  456. }
  457. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  458. REPORT_CALL_ERROR("E19999", "Set Attr:%s of node:%s(%s) failed",
  459. ge::ATTR_INSERTED_BY_GE.c_str(),
  460. last_node->GetName().c_str(), last_node->GetType().c_str());
  461. return INTERNAL_ERROR,
  462. "[Set][Attr] %s of node:%s(%s) failed", ge::ATTR_INSERTED_BY_GE.c_str(),
  463. last_node->GetName().c_str(), last_node->GetType().c_str());
  464. }
  465. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  466. return INTERNAL_ERROR;
  467. }
  468. }
  469. return SUCCESS;
  470. }
  471. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  472. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  473. VarNamesToRefs names_to_refs;
  474. std::string var_name;
  475. if (graph == nullptr) {
  476. REPORT_INNER_ERROR("E19999", "Param graph is nullptr, check invalid");
  477. GELOGE(PARAM_INVALID, "[Check][Param] graph is nullptr.");
  478. return names_to_refs;
  479. }
  480. for (auto &node : graph->GetAllNodes()) {
  481. if (node->GetType() != VARIABLE) {
  482. continue;
  483. }
  484. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  485. (void)names_to_refs[var_name].insert(node);
  486. }
  487. }
  488. return names_to_refs;
  489. }
  490. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  491. std::vector<int64_t> &dst_shape) {
  492. if (src_format == FORMAT_NCHW) {
  493. formats::FormatTransferNchwNc1hwc0 transfer;
  494. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  495. GELOGE(INTERNAL_ERROR, "[Trans][Shape] failed");
  496. return FAILED;
  497. }
  498. } else if (src_format == FORMAT_NHWC) {
  499. formats::FormatTransferNhwcNc1hwc0 transfer;
  500. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  501. GELOGE(INTERNAL_ERROR, "[Trans][Shape] failed");
  502. return FAILED;
  503. }
  504. }
  505. return SUCCESS;
  506. }
  507. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  508. GE_CHECK_NOTNULL(node_ptr);
  509. auto op_desc = node_ptr->GetOpDesc();
  510. GE_CHECK_NOTNULL(op_desc);
  511. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  512. GE_CHECK_NOTNULL(input);
  513. ge::Format old_format = input->GetFormat();
  514. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  515. ge::DataType dt = input->GetDataType();
  516. std::vector<int64_t> dst_shape_dims;
  517. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  518. REPORT_CALL_ERROR("E19999", "Transfer shape to NC1HWC0 failed, op:%s(%s),",
  519. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  520. GELOGE(INTERNAL_ERROR, "[Transfer][Shape] to NC1HWC0 failed, op:%s(%s),",
  521. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  522. return FAILED;
  523. }
  524. input->SetFormat(FORMAT_NC1HWC0);
  525. input->SetShape(ge::GeShape(dst_shape_dims));
  526. auto output = op_desc->MutableOutputDesc(0);
  527. GE_CHECK_NOTNULL(output);
  528. output->SetFormat(FORMAT_NC1HWC0);
  529. output->SetShape(ge::GeShape(dst_shape_dims));
  530. int64_t size = 0;
  531. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  532. if (graph_status != ge::GRAPH_SUCCESS) {
  533. REPORT_CALL_ERROR("E19999", "Get output tensor size failed, op:%s(%s), index:0",
  534. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  535. GELOGE(graph_status, "[Get][TensorSize] In Bytes failed, op:%s(%s), index:0",
  536. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  537. return FAILED;
  538. }
  539. ge::TensorUtils::SetSize(*output, size);
  540. ge::TensorUtils::SetSize(*input, size);
  541. return SUCCESS;
  542. }
  543. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  544. GE_CHECK_NOTNULL(input_output);
  545. ge::Format old_format = input_output->GetFormat();
  546. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  547. ge::DataType dt = input_output->GetDataType();
  548. std::vector<int64_t> dst_shape_dims;
  549. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  550. GELOGE(INTERNAL_ERROR, "[Trans][Shape] to NC1HWC0 failed");
  551. return FAILED;
  552. }
  553. input_output->SetFormat(FORMAT_NC1HWC0);
  554. input_output->SetShape(ge::GeShape(dst_shape_dims));
  555. return SUCCESS;
  556. }
  557. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  558. vector<int64_t> &dst_shape_dims) {
  559. GE_CHECK_NOTNULL(op_desc);
  560. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  561. GE_CHECK_NOTNULL(input);
  562. ge::Format old_format = input->GetFormat();
  563. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  564. input->SetShape(ge::GeShape(dst_shape_dims));
  565. input->SetFormat(storage_format);
  566. auto output = op_desc->MutableOutputDesc(index);
  567. GE_CHECK_NOTNULL(output);
  568. output->SetShape(ge::GeShape(dst_shape_dims));
  569. output->SetFormat(storage_format);
  570. if (!output->MutableShape().IsUnknownShape()) {
  571. int64_t size = 0;
  572. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  573. if (graph_status != ge::GRAPH_SUCCESS) {
  574. REPORT_CALL_ERROR("E19999", "Get output tensor size failed, op:%s(%s), index:%u",
  575. op_desc->GetName().c_str(), op_desc->GetType().c_str(), index);
  576. GELOGE(graph_status, "[Get][TensorSize] In Bytes failed, op:%s(%s), index:%u",
  577. op_desc->GetName().c_str(), op_desc->GetType().c_str(), index);
  578. return FAILED;
  579. }
  580. ge::TensorUtils::SetSize(*input, size);
  581. ge::TensorUtils::SetSize(*output, size);
  582. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  583. "new_shape:%s, new_format:%s, new_size:%lu",
  584. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  585. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  586. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  587. }
  588. return SUCCESS;
  589. }
  590. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &mbatch_node, int32_t &index) {
  591. is_dynamic_batch = false;
  592. std::string related_node_name;
  593. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  594. if (related_node_name.empty()) {
  595. ErrorManager::GetInstance().ATCReportErrMessage(
  596. "E15002", {"opname", "value", "reason"}, {data_node->GetName(), "flag", "but the value is empty"});
  597. GELOGE(INTERNAL_ERROR, "[Check][Param] The data node %s has switchn node flag, but the value is empty",
  598. data_node->GetName().c_str());
  599. return INTERNAL_ERROR;
  600. }
  601. auto out_data_nodes_anchors = data_node->GetOutDataNodesAndAnchors();
  602. for (const auto &out_data_node_anchor : out_data_nodes_anchors) {
  603. if (out_data_node_anchor.first->GetName() == related_node_name) {
  604. mbatch_node = out_data_node_anchor.first;
  605. index = out_data_node_anchor.second->GetIdx();
  606. break;
  607. }
  608. }
  609. if (mbatch_node == nullptr) {
  610. ErrorManager::GetInstance().ATCReportErrMessage(
  611. "E15002", {"opname", "value", "reason"},
  612. {data_node->GetName(), related_node_name, "but can not find it on the graph"});
  613. GELOGE(INTERNAL_ERROR, "[Check][Param] The data node %s has switchn node %s, but can not find it on the graph",
  614. data_node->GetName().c_str(), related_node_name.c_str());
  615. return INTERNAL_ERROR;
  616. }
  617. is_dynamic_batch = true;
  618. }
  619. return SUCCESS;
  620. }
  621. bool CheckOpType(const NodePtr &node, const std::string type) {
  622. if (node->GetType() == type) {
  623. return true;
  624. }
  625. return false;
  626. }
  627. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  628. auto op = node_ptr->GetOpDesc();
  629. GE_CHECK_NOTNULL(op);
  630. auto inputDescsPtr = op->GetAllInputsDescPtr();
  631. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  632. ge::Format format = ge::FORMAT_ND;
  633. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  634. // process, otherwise fe will insert transdata.
  635. for (auto &inputDescPtr : inputDescsPtr) {
  636. GE_CHECK_NOTNULL(inputDescPtr);
  637. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  638. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  639. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  640. inputDescPtr->SetFormat(format);
  641. inputDescPtr->SetOriginFormat(format);
  642. }
  643. }
  644. for (auto &outputDescPtr : outputDescsPtr) {
  645. GE_CHECK_NOTNULL(outputDescPtr);
  646. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  647. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  648. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  649. outputDescPtr->SetFormat(format);
  650. outputDescPtr->SetOriginFormat(format);
  651. }
  652. }
  653. return SUCCESS;
  654. }
  655. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  656. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  657. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  658. // data and netoutput, and no longer inserts the transnode.
  659. Status ProcessInputDtDynShape(NodePtr &node_ptr, NodePtr &switchn_node, DataType &dt_set) {
  660. GE_CHECK_NOTNULL(node_ptr);
  661. auto op_desc = node_ptr->GetOpDesc();
  662. GE_CHECK_NOTNULL(op_desc);
  663. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  664. GE_CHECK_NOTNULL(input);
  665. ge::DataType src_dtype = input->GetDataType();
  666. if (src_dtype == dt_set) {
  667. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  668. return SUCCESS;
  669. }
  670. input->SetDataType(dt_set);
  671. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  672. GE_CHECK_NOTNULL(output);
  673. output->SetDataType(dt_set);
  674. GeShape shape = input->GetShape();
  675. if (!shape.IsUnknownShape()) {
  676. int64_t input_shape_size = 0;
  677. int64_t output_shape_size = 0;
  678. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  679. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  680. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  681. REPORT_CALL_ERROR("E19999", "Get input tensor size failed, op:%s(%s), index:0",
  682. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  683. GELOGE(GRAPH_FAILED, "[Process][InputOp] Get tensor size of op [%s] failed!", node_ptr->GetName().c_str());
  684. return FAILED;
  685. }
  686. ge::TensorUtils::SetSize(*input, input_shape_size);
  687. ge::TensorUtils::SetSize(*output, output_shape_size);
  688. GELOGI("[Process][InputDynShape] Set input and output size of node [%s] success.", node_ptr->GetName().c_str());
  689. }
  690. return SUCCESS;
  691. }
  692. Status UpdateInputOutputDataType(NodePtr &mbatch_node, DataType &dt_set, int32_t index) {
  693. auto mbatch_desc = mbatch_node->GetOpDesc();
  694. GE_CHECK_NOTNULL(mbatch_desc);
  695. auto mbatch_input = mbatch_desc->MutableInputDesc(index);
  696. GE_CHECK_NOTNULL(mbatch_input);
  697. mbatch_input->SetDataType(dt_set);
  698. if (mbatch_node->GetType() == SWITCHN) {
  699. for (uint32_t i = 0; i < mbatch_node->GetAllOutDataAnchorsSize(); ++i) {
  700. const GeTensorDescPtr &mbatch_output = mbatch_desc->MutableOutputDesc(i);
  701. GE_CHECK_NOTNULL(mbatch_output);
  702. mbatch_output->SetDataType(dt_set);
  703. }
  704. }
  705. GELOGD("Update input and output data type of node[name: %s, type: %s, input index: %d] to %s.",
  706. mbatch_node->GetName().c_str(), mbatch_node->GetType().c_str(), index,
  707. TypeUtils::DataTypeToSerialString(dt_set).c_str());
  708. return SUCCESS;
  709. }
  710. Status UpdateSubgraphDataOfCase(NodePtr &mbatch_node, DataType &dt_set, int32_t index) {
  711. if (mbatch_node->GetType() != CASE) {
  712. return SUCCESS;
  713. }
  714. auto subgraphs = NodeUtils::GetAllSubgraphs(*mbatch_node);
  715. for (const auto &subgraph : subgraphs) {
  716. GE_CHECK_NOTNULL(subgraph);
  717. for (auto &sub_node : subgraph->GetDirectNode()) {
  718. GE_CHECK_NOTNULL(sub_node);
  719. if (sub_node->GetType() != DATA) {
  720. continue;
  721. }
  722. auto data_desc = sub_node->GetOpDesc();
  723. GE_CHECK_NOTNULL(data_desc);
  724. int32_t parent_node_index = 0;
  725. if (!AttrUtils::GetInt(data_desc, ATTR_NAME_PARENT_NODE_INDEX, parent_node_index) ||
  726. (parent_node_index != index)) {
  727. continue;
  728. }
  729. auto data_input = data_desc->MutableInputDesc(0);
  730. GE_CHECK_NOTNULL(data_input);
  731. data_input->SetDataType(dt_set);
  732. auto data_output = data_desc->MutableOutputDesc(0);
  733. GE_CHECK_NOTNULL(data_output);
  734. data_output->SetDataType(dt_set);
  735. GELOGD("Update input and output data type of node[name: %s, type: %s, parent_node_index: %d] in subgraph %s "
  736. "to %s.", data_desc->GetName().c_str(), data_desc->GetType().c_str(), parent_node_index,
  737. subgraph->GetName().c_str(), TypeUtils::DataTypeToSerialString(dt_set).c_str());
  738. }
  739. }
  740. return SUCCESS;
  741. }
  742. Status ProcessMbatchScene(NodePtr &mbatch_node, DataType &dt_set, int32_t index) {
  743. GELOGI("The node [%s] dtype set fp16.", mbatch_node->GetName().c_str());
  744. if (UpdateInputOutputDataType(mbatch_node, dt_set, index) != SUCCESS) {
  745. GELOGE(FAILED, "[Update][InputOutputDataType] of node[name: %s, type: %s] to %s failed.",
  746. mbatch_node->GetName().c_str(), mbatch_node->GetType().c_str(),
  747. TypeUtils::DataTypeToSerialString(dt_set).c_str());
  748. return FAILED;
  749. }
  750. if (UpdateSubgraphDataOfCase(mbatch_node, dt_set, index) != SUCCESS) {
  751. GELOGE(FAILED, "[Update][SubgraphDataOfCase] node[parent_node_index:%d] in subgraphs of "
  752. "node[name:%s, type:%s] to %s failed.", index, mbatch_node->GetName().c_str(),
  753. mbatch_node->GetType().c_str(), TypeUtils::DataTypeToSerialString(dt_set).c_str());
  754. return FAILED;
  755. }
  756. return SUCCESS;
  757. }
  758. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  759. GE_CHECK_NOTNULL(node_ptr);
  760. auto op_desc = node_ptr->GetOpDesc();
  761. GE_CHECK_NOTNULL(op_desc);
  762. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  763. GE_CHECK_NOTNULL(input);
  764. ge::Format old_format = input->GetFormat();
  765. ge::GeShape old_shape = input->GetShape();
  766. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  767. if (!support) {
  768. ErrorManager::GetInstance().ATCReportErrMessage(
  769. "E19014", {"opname", "value", "reason"},
  770. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  771. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  772. GELOGE(INTERNAL_ERROR, "[Check][Param] The format [%s] is unsupported, op:%s",
  773. TypeUtils::FormatToSerialString(old_format).c_str(), op_desc->GetName().c_str());
  774. return FAILED;
  775. }
  776. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  777. GELOGE(INTERNAL_ERROR, "[Modify][InputFormatAndShape] failed, op:%s(%s)",
  778. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  779. return FAILED;
  780. }
  781. if (is_dynamic_batch) {
  782. auto switchn_op_desc = switchn_node->GetOpDesc();
  783. GE_CHECK_NOTNULL(switchn_op_desc);
  784. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  785. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  786. REPORT_CALL_ERROR("E19999", "Modify format and shape of input:0 in op:%s(%s) failed",
  787. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  788. GELOGE(INTERNAL_ERROR, "[Modify][FormatAndShape] of input:0 in op:%s(%s) failed",
  789. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  790. return FAILED;
  791. }
  792. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  793. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  794. GE_CHECK_NOTNULL(switchn_output);
  795. old_format = switchn_output->GetFormat();
  796. old_shape = switchn_output->GetShape();
  797. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  798. REPORT_CALL_ERROR("E19999", "Modify format and shape of output:%u in op:%s(%s) failed", i,
  799. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  800. GELOGE(INTERNAL_ERROR, "[Modify][FormatAndShape] of output:%u in op:%s(%s) failed", i,
  801. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  802. return FAILED;
  803. }
  804. }
  805. }
  806. return SUCCESS;
  807. }
  808. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  809. auto op_desc = node_ptr->GetOpDesc();
  810. GE_CHECK_NOTNULL(op_desc);
  811. string set_dt_str;
  812. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  813. return SUCCESS;
  814. }
  815. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  816. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  817. bool is_dynamic_batch = false;
  818. NodePtr mbatch_node = nullptr;
  819. int32_t index = 0;
  820. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, mbatch_node, index)) {
  821. GELOGE(INTERNAL_ERROR, "[Call][CheckIfDynamicBatchScene] failed, op:%s", op_desc->GetName().c_str());
  822. return FAILED;
  823. }
  824. if (ProcessInputDtDynShape(node_ptr, mbatch_node, dt_set) != SUCCESS) {
  825. GELOGE(INTERNAL_ERROR, "[Process][InputDtDynShape] ProcessInputFP16 failed, op:%s", op_desc->GetName().c_str());
  826. return FAILED;
  827. }
  828. if (is_dynamic_batch && ProcessMbatchScene(mbatch_node, dt_set, index) != SUCCESS) {
  829. GELOGE(INTERNAL_ERROR, "[Process][MbatchScene] failed");
  830. return FAILED;
  831. }
  832. // check if need to set format
  833. string set_format;
  834. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  835. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  836. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  837. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, mbatch_node) != SUCCESS) {
  838. GELOGE(INTERNAL_ERROR, "[Process][InputNC1HWC0] failed, op:%s", node_ptr->GetName().c_str());
  839. return FAILED;
  840. }
  841. }
  842. return SUCCESS;
  843. }
  844. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  845. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  846. GE_CHECK_NOTNULL(op_desc);
  847. GE_CHECK_NOTNULL(tensor_desc_ptr);
  848. storage_format = FORMAT_RESERVED;
  849. int64_t format = FORMAT_RESERVED;
  850. dst_shape_dims.clear();
  851. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  852. storage_format = static_cast<Format>(format);
  853. vector<int32_t> storage_shape;
  854. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  855. for (auto dim : storage_shape) {
  856. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  857. }
  858. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  859. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  860. formats::JoinToString(storage_shape).c_str());
  861. } else {
  862. ErrorManager::GetInstance().ATCReportErrMessage(
  863. "15003", {"opname", "format"},
  864. {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)});
  865. GELOGE(PARAM_INVALID, "[Check][Param] Update node by storage format failed, storage_shape not set. "
  866. "node:[%s], storage_format [%s]",
  867. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  868. return FAILED;
  869. }
  870. ge::Format old_format = tensor_desc_ptr->GetFormat();
  871. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  872. if (old_format == storage_format && old_shape == dst_shape_dims) {
  873. GELOGI("Update node by storage format, not changed.");
  874. storage_format = FORMAT_RESERVED;
  875. return SUCCESS;
  876. }
  877. }
  878. return SUCCESS;
  879. }
  880. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  881. NodePtr &node) {
  882. bool is_dynamic = CheckOpType(node, MERGE);
  883. auto src_op_desc = node->GetOpDesc();
  884. GE_CHECK_NOTNULL(src_op_desc);
  885. ge::GeShape src_shape = src_desc.GetShape();
  886. ge::Format src_format = src_desc.GetFormat();
  887. net_output_input_desc->SetDataType(DT_FLOAT16);
  888. if (is_dynamic) {
  889. auto merge_output = src_op_desc->MutableOutputDesc(0);
  890. GE_CHECK_NOTNULL(merge_output);
  891. merge_output->SetDataType(DT_FLOAT16);
  892. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  893. auto merge_input = src_op_desc->MutableInputDesc(i);
  894. GE_CHECK_NOTNULL(merge_input);
  895. merge_input->SetDataType(DT_FLOAT16);
  896. }
  897. }
  898. std::vector<int64_t> dst_shape_dims;
  899. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  900. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  901. REPORT_CALL_ERROR("E19999", "Transfer output:0 shape of op:%s(%s) to NC1HWC0 format failed, shape:%s, format:%s",
  902. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str(),
  903. src_shape.ToString().c_str(), TypeUtils::FormatToSerialString(src_format).c_str());
  904. GELOGE(INTERNAL_ERROR, "[Trans][Shape] of op:%s(%s) to NC1HWC0 format failed, shape:%s, format:%s",
  905. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str(),
  906. src_shape.ToString().c_str(), TypeUtils::FormatToSerialString(src_format).c_str());
  907. return FAILED;
  908. }
  909. ge::GeShape dst_shape(dst_shape_dims);
  910. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  911. net_output_input_desc->SetShape(dst_shape);
  912. if (is_dynamic) {
  913. auto merge_out = src_op_desc->MutableOutputDesc(0);
  914. GE_CHECK_NOTNULL(merge_out);
  915. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  916. REPORT_CALL_ERROR("E19999", "Modify format and shape of output:0 in op:%s(%s) failed",
  917. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  918. GELOGE(INTERNAL_ERROR, "[Modify][FormatAndShape] of output:0 in op:%s(%s) failed",
  919. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  920. return FAILED;
  921. }
  922. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  923. auto merge_in = src_op_desc->MutableInputDesc(i);
  924. GE_CHECK_NOTNULL(merge_in);
  925. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  926. REPORT_CALL_ERROR("E19999", "Modify format and shape of input:%u in op:%s(%s) failed", i,
  927. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  928. GELOGE(INTERNAL_ERROR, "[Modify][FormatAndShape] of input:%u in op:%s(%s) failed", i,
  929. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  930. return FAILED;
  931. }
  932. }
  933. }
  934. return SUCCESS;
  935. }
  936. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  937. GE_CHECK_NOTNULL(netout_desc);
  938. vector<string> output_dt_str;
  939. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  940. for (auto dt_str : output_dt_str) {
  941. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  942. if (dt_str_split.size() == kUserDefinedElementCount) {
  943. if (dt_str_split[0] == to_string(index)) {
  944. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  945. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  946. TypeUtils::DataTypeToSerialString(dt).c_str());
  947. return true;
  948. }
  949. }
  950. }
  951. }
  952. return false;
  953. }
  954. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  955. GE_CHECK_NOTNULL(netout_desc);
  956. vector<string> output_format_str;
  957. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  958. for (auto format_str : output_format_str) {
  959. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  960. if (format_str_split.size() == kUserDefinedElementCount) {
  961. if (format_str_split[0] == to_string(index)) {
  962. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  963. return true;
  964. }
  965. }
  966. }
  967. }
  968. return false;
  969. }
  970. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  971. auto op_desc = node->GetOpDesc();
  972. GE_CHECK_NOTNULL(op_desc);
  973. ge::DataType output_data_type = ge::DT_FLOAT;
  974. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  975. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  976. auto peer_out = in_anchor->GetPeerOutAnchor();
  977. GE_CHECK_NOTNULL(peer_out);
  978. auto src_node = peer_out->GetOwnerNode();
  979. GE_CHECK_NOTNULL(src_node);
  980. bool is_dynamic = CheckOpType(src_node, MERGE);
  981. OpDescPtr src_op_desc = src_node->GetOpDesc();
  982. GE_CHECK_NOTNULL(src_op_desc);
  983. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  984. GE_CHECK_NOTNULL(net_output_input_desc);
  985. ge::GeShape old_shape = net_output_input_desc->GetShape();
  986. ge::Format old_format = net_output_input_desc->GetFormat();
  987. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  988. // Update datatype
  989. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  990. GELOGI("Enter into process output_type schedule");
  991. net_output_input_desc->SetDataType(output_data_type);
  992. if (is_dynamic) {
  993. auto merge_output = src_op_desc->MutableOutputDesc(0);
  994. GE_CHECK_NOTNULL(merge_output);
  995. merge_output->SetDataType(output_data_type);
  996. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  997. auto merge_input = src_op_desc->MutableInputDesc(i);
  998. GE_CHECK_NOTNULL(merge_input);
  999. merge_input->SetDataType(output_data_type);
  1000. }
  1001. }
  1002. }
  1003. // check if is_output_adjust_hw_layout is set
  1004. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  1005. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  1006. ErrorManager::GetInstance().ATCReportErrMessage(
  1007. "E19014", {"opname", "value", "reason"},
  1008. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  1009. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  1010. GELOGE(INTERNAL_ERROR, "[Check][Param] Format is not one of NCHW, NHWC, NC1HWC0.");
  1011. return FAILED;
  1012. }
  1013. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  1014. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  1015. GELOGE(INTERNAL_ERROR, "[Process][NetOutput] fp16 nc1hwc0 failed.");
  1016. return FAILED;
  1017. }
  1018. }
  1019. }
  1020. return SUCCESS;
  1021. }
  1022. Status GetDynamicInputShapeRange(const std::vector<GeTensor> &user_input, const std::map<string, string> &graph_option,
  1023. vector<vector<std::pair<int64_t, int64_t>>> &range_vec) {
  1024. // check both mode and shape_range option are all enabled
  1025. auto mode_iter = graph_option.find(OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  1026. bool enable_dynamic_execute_mode = (mode_iter != graph_option.end()) && (mode_iter->second == "dynamic_execute");
  1027. if (!enable_dynamic_execute_mode) {
  1028. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  1029. OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  1030. }
  1031. auto iter = graph_option.find(OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1032. bool enable_input_shape_range = (iter != graph_option.end()) && (!iter->second.empty());
  1033. if (!enable_input_shape_range) {
  1034. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  1035. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1036. }
  1037. if (enable_dynamic_execute_mode && enable_input_shape_range) {
  1038. GELOGD("GraphOption: %s value is dynamic_execute, %s value is %s.", OPTION_EXEC_DYNAMIC_EXECUTE_MODE,
  1039. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE, iter->second.c_str());
  1040. } else if (!enable_dynamic_execute_mode && !enable_input_shape_range) {
  1041. return SUCCESS;
  1042. } else {
  1043. REPORT_INNER_ERROR("E19999", "Graph option: %s and %s should be enabled at the same time, check invalid",
  1044. OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1045. GELOGE(PARAM_INVALID, "[Check][Param] Graph option: %s and %s should be enabled at the same time.",
  1046. OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1047. return PARAM_INVALID;
  1048. }
  1049. if (ParseInputShapeRange(iter->second, range_vec) != SUCCESS) {
  1050. GELOGE(PARAM_INVALID, "[Parse][ShapeRange] Parse dynamic input shape range failed.");
  1051. return PARAM_INVALID;
  1052. }
  1053. if (range_vec.size() != user_input.size()) {
  1054. GELOGE(PARAM_INVALID, "[Check][Param] Dynamic input shape range size is %zu, inputs size is %zu. Not match.",
  1055. range_vec.size(), user_input.size());
  1056. return PARAM_INVALID;
  1057. }
  1058. return SUCCESS;
  1059. }
  1060. Status UpdateDynamicInputShapeRange(const ge::GeAttrValue::INT index,
  1061. const vector<vector<std::pair<int64_t, int64_t>>> &range_vec, OpDescPtr &op,
  1062. GeTensorDesc &desc) {
  1063. auto origin_shape = desc.GetShape();
  1064. auto current_shape_range_vec = range_vec.at(index);
  1065. if (current_shape_range_vec.size() != origin_shape.GetDimNum()) {
  1066. REPORT_INNER_ERROR("E19999", "Given shape_range dim num is %zu, current dim:%s num is %zu, not match, "
  1067. "check invalid", current_shape_range_vec.size(), origin_shape.ToString().c_str(),
  1068. origin_shape.GetDimNum());
  1069. GELOGE(PARAM_INVALID, "[Check][Param] Given shape_range dim num is %zu, current dim num is %zu, "
  1070. "not match.Pleace Check.", current_shape_range_vec.size(), origin_shape.GetDimNum());
  1071. return PARAM_INVALID;
  1072. }
  1073. for (size_t i = 0; i < origin_shape.GetDimNum(); ++i) {
  1074. auto curr_dim = origin_shape.GetDim(i);
  1075. auto left_range = current_shape_range_vec.at(i).first;
  1076. auto right_range = current_shape_range_vec.at(i).second;
  1077. if (left_range == right_range) {
  1078. // given shape_range is known dim, check is same as origin or not
  1079. if (curr_dim != left_range) {
  1080. REPORT_INNER_ERROR("E19999", "Given shape range is %ld, current dim shape is %ld, not match, dim_index:%zu, "
  1081. "check invalid", left_range, curr_dim, i);
  1082. GELOGE(PARAM_INVALID, "[Check][Param] Given shape range is %ld, current dim shape is %ld, "
  1083. "not match.Pleace Check.", left_range, curr_dim);
  1084. return PARAM_INVALID;
  1085. }
  1086. origin_shape.SetDim(i, left_range);
  1087. } else {
  1088. // given shape_range is fix range, check input_shape is in this range or not
  1089. if (right_range != UNKNOWN_DIM) {
  1090. if ((curr_dim < left_range) || (curr_dim > right_range)) {
  1091. REPORT_INNER_ERROR("E19999", "Given shape range is [%ld~%ld], current dim shape is %ld, out of range, "
  1092. "dim_index:%zu, check invalid",
  1093. left_range, right_range, curr_dim, i);
  1094. GELOGE(PARAM_INVALID, "[Check][Param] Given shape range is [%ld~%ld], current dim shape is %ld, "
  1095. "out of range.Pleace Check.", left_range, right_range, curr_dim);
  1096. return PARAM_INVALID;
  1097. }
  1098. }
  1099. origin_shape.SetDim(i, UNKNOWN_DIM);
  1100. }
  1101. }
  1102. desc.SetShape(origin_shape);
  1103. desc.SetShapeRange(current_shape_range_vec);
  1104. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1105. GE_CHK_GRAPH_STATUS_RET(graph_ret, "[Update][InputDesc] fail, graph ret: %u", graph_ret);
  1106. graph_ret = op->UpdateOutputDesc(0, desc);
  1107. GE_CHK_GRAPH_STATUS_RET(graph_ret, "[Update][OutputDesc] fail, graph ret: %u", graph_ret);
  1108. return SUCCESS;
  1109. }
  1110. } // namespace
  1111. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  1112. GraphPrepare::~GraphPrepare() {}
  1113. /**
  1114. * @param graph
  1115. * @return
  1116. */
  1117. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  1118. GE_CHECK_NOTNULL(graph);
  1119. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  1120. for (auto &node : graph->GetAllNodes()) {
  1121. if (node == nullptr) {
  1122. continue;
  1123. }
  1124. if (node->GetType() != VARIABLE) {
  1125. continue;
  1126. }
  1127. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  1128. if (trans_road == nullptr) {
  1129. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  1130. continue;
  1131. }
  1132. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  1133. auto ret = RecoverTransRoadForVar(node, *trans_road);
  1134. if (ret != SUCCESS) {
  1135. GELOGE(INTERNAL_ERROR, "[Recover][TransRoad] for var %s failed", node->GetName().c_str());
  1136. return INTERNAL_ERROR;
  1137. }
  1138. auto iter = var_names_to_refs.find(node->GetName());
  1139. if (iter != var_names_to_refs.end()) {
  1140. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  1141. if (ret != SUCCESS) {
  1142. GELOGE(INTERNAL_ERROR, "[Recover][TransRoad] for var ref %s failed", node->GetName().c_str());
  1143. return INTERNAL_ERROR;
  1144. }
  1145. }
  1146. }
  1147. return SUCCESS;
  1148. }
  1149. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  1150. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  1151. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  1152. if (compute_graph_ != nullptr) {
  1153. compute_graph_->SetSessionID(session_id);
  1154. }
  1155. session_id_ = session_id;
  1156. Status ret = CheckGraph();
  1157. if (ret != SUCCESS) {
  1158. GELOGE(ret, "[Check][Graph] fail, ret:%u", ret);
  1159. return ret;
  1160. }
  1161. (void)compute_graph_->TopologicalSorting();
  1162. ret = CheckRefOp();
  1163. if (ret != SUCCESS) {
  1164. GELOGE(ret, "[Check][RefOp] fail, ret:%u", ret);
  1165. return ret;
  1166. }
  1167. return SUCCESS;
  1168. }
  1169. Status GraphPrepare::CheckGraph() {
  1170. if (compute_graph_ == nullptr) {
  1171. REPORT_INNER_ERROR("E19999", "compute_graph_ is nullptr, check invalid");
  1172. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][Param] compute_graph_ is nullptr");
  1173. return GE_GRAPH_INIT_FAILED;
  1174. }
  1175. auto nodes = compute_graph_->GetAllNodes();
  1176. if (nodes.empty()) {
  1177. REPORT_INNER_ERROR("E19999", "nodes in graph is empty, check invalid");
  1178. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][Param] Invalid graph, no nodes in this graph.");
  1179. return GE_GRAPH_INIT_FAILED;
  1180. }
  1181. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  1182. GE_CHECK_NOTNULL(node);
  1183. if (node->GetOpDesc() == nullptr) {
  1184. REPORT_INNER_ERROR("E19999", "node without opdesc exist in graph, check invalid");
  1185. GELOGE(GE_GRAPH_INIT_FAILED, "[Get][OpDesc] failed, Check Graph node opdesc is NULL");
  1186. return GE_GRAPH_INIT_FAILED;
  1187. }
  1188. }
  1189. return SUCCESS;
  1190. }
  1191. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  1192. const std::set<NodePtr> &ref_nodes) {
  1193. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  1194. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  1195. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  1196. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  1197. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  1198. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH,
  1199. ge::DATA};
  1200. GE_CHECK_NOTNULL(node);
  1201. const auto &op_desc = node->GetOpDesc();
  1202. GE_CHECK_NOTNULL(op_desc);
  1203. const auto input_index = op_desc->GetInputIndexByName(input_name);
  1204. const auto &in_anchor = node->GetInDataAnchor(input_index);
  1205. GE_CHECK_NOTNULL(in_anchor);
  1206. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  1207. GE_CHECK_NOTNULL(peer_out_anchor);
  1208. const auto &input_node = peer_out_anchor->GetOwnerNode();
  1209. GE_CHECK_NOTNULL(input_node);
  1210. const auto &input_op_desc = input_node->GetOpDesc();
  1211. GE_CHECK_NOTNULL(input_op_desc);
  1212. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  1213. if (is_ref) {
  1214. return SUCCESS;
  1215. }
  1216. auto input_type = input_op_desc->GetType();
  1217. if (input_type == ge::FRAMEWORKOP) {
  1218. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  1219. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed",
  1220. ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE.c_str(),
  1221. input_op_desc->GetName().c_str(), input_op_desc->GetType().c_str());
  1222. GELOGE(PARAM_INVALID, "[Get][Attr] %s of op:%s(%s) failed", ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE.c_str(),
  1223. input_op_desc->GetName().c_str(), input_op_desc->GetType().c_str());
  1224. return PARAM_INVALID;
  1225. }
  1226. }
  1227. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  1228. if (!is_acceptable) {
  1229. ErrorManager::GetInstance().ATCReportErrMessage(
  1230. "E15005", {"opname", "optype", "opname1", "optype1"},
  1231. {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()});
  1232. GELOGE(PARAM_INVALID, "[Check][Param] The ref input of ref node %s[%s] must be ref node or variable, "
  1233. "but %s[%s]isn't.", node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  1234. input_op_desc->GetType().c_str());
  1235. return PARAM_INVALID;
  1236. }
  1237. return SUCCESS;
  1238. }
  1239. Status GraphPrepare::CheckRefOp() {
  1240. GE_CHECK_NOTNULL(compute_graph_);
  1241. std::set<NodePtr> ref_nodes;
  1242. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  1243. if (node == nullptr) {
  1244. REPORT_INNER_ERROR("E19999", "nullptr node exist in graph, check invalid");
  1245. GELOGE(PARAM_INVALID, "[Check][Param] param [node] must not be null.");
  1246. return PARAM_INVALID;
  1247. }
  1248. auto op_desc = node->GetOpDesc();
  1249. if (op_desc == nullptr) {
  1250. REPORT_INNER_ERROR("E19999", "node without opdesc exist in graph, check invalid");
  1251. GELOGE(PARAM_INVALID, "[Check][Param] OpDesc of param [node] must not be null.");
  1252. return PARAM_INVALID;
  1253. }
  1254. auto input_name_index = op_desc->GetAllInputName();
  1255. auto outputs = op_desc->GetAllOutputName();
  1256. for (const auto &name_index : input_name_index) {
  1257. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  1258. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  1259. GELOGE(PARAM_INVALID, "[Check][RefInputNode] failed, node:%s.", op_desc->GetName().c_str());
  1260. return PARAM_INVALID;
  1261. }
  1262. (void)ref_nodes.insert(node); // no need to check value
  1263. }
  1264. }
  1265. }
  1266. return SUCCESS;
  1267. };
  1268. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  1269. GE_CHECK_NOTNULL(compute_graph_);
  1270. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  1271. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  1272. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  1273. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  1274. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  1275. return SUCCESS;
  1276. }
  1277. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1278. if (node == nullptr) {
  1279. REPORT_INNER_ERROR("E19999", "Param node is nullptr, check invalid");
  1280. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "[Check][Param] Input node is nullptr");
  1281. return GE_GRAPH_GRAPH_NODE_NULL;
  1282. }
  1283. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1284. if (op_desc_ptr == nullptr) {
  1285. REPORT_INNER_ERROR("E19999", "Param node's op_desc is nullptr, check invalid");
  1286. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "[Get][OpDesc] Input node opdesc is NULL");
  1287. return GE_GRAPH_GRAPH_NODE_NULL;
  1288. }
  1289. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1290. GeShape output_shape = output.GetShape();
  1291. if (output_shape.IsUnknownShape()) {
  1292. GELOGD("[Adjust][DataOpOutput] Shape of op [%s] output is unknown.", node->GetName().c_str());
  1293. return SUCCESS;
  1294. }
  1295. int64_t tensor_size = 0;
  1296. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1297. if (graph_status != GRAPH_SUCCESS) {
  1298. ErrorManager::GetInstance().ATCReportErrMessage(
  1299. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1300. GELOGE(graph_status, "[Call][GetTensorMemorySizeInBytes] failed, op:%s", node->GetName().c_str());
  1301. return FAILED;
  1302. }
  1303. TensorUtils::SetSize(output, tensor_size);
  1304. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1305. if (graph_ret != GRAPH_SUCCESS) {
  1306. REPORT_CALL_ERROR("E19999", "Update output desc of op:%s(%s) failed, index:0",
  1307. op_desc_ptr->GetName().c_str(), op_desc_ptr->GetType().c_str());
  1308. GELOGE(graph_ret, "[Update][OutputDesc] of op:%s(%s) failed, index:0",
  1309. op_desc_ptr->GetName().c_str(), op_desc_ptr->GetType().c_str());
  1310. return graph_ret;
  1311. }
  1312. return SUCCESS;
  1313. }
  1314. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input,
  1315. const std::map<string, string> &graph_option) {
  1316. // Get shape range of input in dynamic_execute mode
  1317. vector<vector<std::pair<int64_t, int64_t>>> dynamic_shape_range_vec;
  1318. auto ret = GetDynamicInputShapeRange(user_input, graph_option, dynamic_shape_range_vec);
  1319. GE_CHK_STATUS_RET(ret, "[Get][DynamicInputShapeRange] failed, Graph option is not right on Dynamic execute mode.");
  1320. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1321. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1322. GE_CHECK_NOTNULL(input_node);
  1323. OpDescPtr op = input_node->GetOpDesc();
  1324. GE_CHECK_NOTNULL(op);
  1325. if (op->GetType() == DATA) {
  1326. GeAttrValue::INT index = 0;
  1327. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1328. GELOGW("Get index from data attr failed");
  1329. continue;
  1330. }
  1331. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1332. std::string situation = "data op index[" + std::to_string(index) + "]";
  1333. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1334. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1335. GELOGE(PARAM_INVALID, "[Check][Param] user_input size = %zu, graph data op index = %ld.",
  1336. user_input.size(), index);
  1337. return FAILED;
  1338. }
  1339. if (IsDynamicDims(input_node)) {
  1340. continue;
  1341. }
  1342. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1343. auto format = desc.GetFormat();
  1344. auto origin_format = desc.GetOriginFormat();
  1345. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1346. auto tune_flag = (options_.build_mode == BUILD_MODE_TUNING) && (options_.build_step == BUILD_STEP_AFTER_BUILDER);
  1347. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op) && (!tune_flag);
  1348. if (need_check_internal_format) {
  1349. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1350. if (is_internal) {
  1351. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1352. {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" +
  1353. TypeUtils::FormatToSerialString(origin_format) + "]", "it is not support"});
  1354. GELOGE(PARAM_INVALID, "[Check][Param] Input format %s or origin_format %s is not support.",
  1355. TypeUtils::FormatToSerialString(format).c_str(),
  1356. TypeUtils::FormatToSerialString(origin_format).c_str());
  1357. return FAILED;
  1358. }
  1359. }
  1360. auto data_type = desc.GetDataType();
  1361. uint32_t length = 1;
  1362. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1363. if (!type_ret) {
  1364. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1365. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1366. GELOGE(PARAM_INVALID, "[Check][Param] Input datatype %s is not support.",
  1367. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1368. return FAILED;
  1369. }
  1370. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1371. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1372. int64_t shape_size = desc_shape * length;
  1373. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1374. int64_t size = 0;
  1375. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1376. REPORT_CALL_ERROR("E19999", "Get size of user input tensor failed, index:%ld", index);
  1377. GELOGE(INTERNAL_ERROR, "[Get][Size] of user input tensor failed, index:%ld", index);
  1378. return FAILED);
  1379. bool size_check = (size != 0 && shape_size != size);
  1380. if (size_check) {
  1381. std::string situation = "input data size[" + std::to_string(size) +
  1382. "] and shape_size[" + std::to_string(size) + "]";
  1383. std::string reason = "because size != 0 and shape_size != size";
  1384. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1385. GELOGE(PARAM_INVALID, "[Check][Param] input data size = %ld, shape_size = %ld.", size, shape_size);
  1386. return FAILED;
  1387. }
  1388. ge::TensorUtils::SetSize(desc, shape_size);
  1389. if (!tune_flag) {
  1390. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1391. if (graph_ret != GRAPH_SUCCESS) {
  1392. REPORT_CALL_ERROR("E19999", "Update input desc of op:%s(%s) failed, index:0",
  1393. op->GetName().c_str(), op->GetType().c_str());
  1394. GELOGE(graph_ret, "[Update][InputDesc] of op:%s(%s) failed, index:0",
  1395. op->GetName().c_str(), op->GetType().c_str());
  1396. return graph_ret;
  1397. }
  1398. // Size will be recalculated in the build stage
  1399. ge::TensorUtils::SetSize(desc, 0);
  1400. graph_ret = op->UpdateOutputDesc(0, desc);
  1401. if (graph_ret != GRAPH_SUCCESS) {
  1402. REPORT_CALL_ERROR("E19999", "Update output desc of op:%s(%s) failed, index:0",
  1403. op->GetName().c_str(), op->GetType().c_str());
  1404. GELOGE(graph_ret, "[Update][OutputDesc] of op:%s(%s) failed, index:0",
  1405. op->GetName().c_str(), op->GetType().c_str());
  1406. return graph_ret;
  1407. }
  1408. } else {
  1409. GELOGI("data %s skip update info in tune mode", op->GetName().c_str());
  1410. }
  1411. if (!dynamic_shape_range_vec.empty()) {
  1412. ret = UpdateDynamicInputShapeRange(index, dynamic_shape_range_vec, op, desc);
  1413. GE_CHK_STATUS_RET(ret, "[Update][DynamicInputShapeRange] on %s failed.", op->GetName().c_str());
  1414. continue;
  1415. }
  1416. if (!options_.train_graph_flag) {
  1417. Status ret = AdjustDataOpOutput(input_node);
  1418. GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(ret, "[Adjust][DataOpOutput] fail, ret:%u", ret); return ret);
  1419. }
  1420. }
  1421. }
  1422. return SUCCESS;
  1423. }
  1424. Status GraphPrepare::TryDoAipp() {
  1425. // infer and with aipp configure file, then call aipp insert
  1426. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1427. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1428. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1429. if (ret != SUCCESS) {
  1430. GELOGE(INTERNAL_ERROR, "[Init][InsertNewOpUtil] failed.");
  1431. return INTERNAL_ERROR;
  1432. }
  1433. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1434. if (ret != SUCCESS) {
  1435. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "[Parse][ConfigFile] %s failed",
  1436. options_.insert_op_file.c_str());
  1437. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1438. }
  1439. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1440. if (ret != SUCCESS) {
  1441. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "[Insert][AippOps] failed, ret:%u", ret);
  1442. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1443. }
  1444. }
  1445. return SUCCESS;
  1446. }
  1447. Status GraphPrepare::FormatAndShapeProcess() {
  1448. Status ret = ResourcePairProcess("add");
  1449. if (ret != SUCCESS) {
  1450. GELOGE(ret, "ResourcePairProcess failed");
  1451. return ret;
  1452. }
  1453. GE_TIMESTAMP_START(InferOriginFormat1);
  1454. ret = compute_graph_->InferOriginFormat();
  1455. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1456. GE_DUMP(compute_graph_, "after_first_inferformat");
  1457. if (ret != SUCCESS) {
  1458. GELOGE(ret, "[Call][InferOriginFormat] Prepare Graph first inferformat failed");
  1459. return ret;
  1460. }
  1461. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1462. ret = InferShapeForPreprocess();
  1463. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1464. GE_DUMP(compute_graph_, "after_infershape");
  1465. if (ret != SUCCESS) {
  1466. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "[Call][InferShapeForPreprocess] Prepare Graph infershape failed");
  1467. return GE_GRAPH_INFERSHAPE_FAILED;
  1468. }
  1469. GE_TIMESTAMP_START(InferOriginFormat2);
  1470. ret = compute_graph_->InferOriginFormat();
  1471. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1472. if (ret != SUCCESS) {
  1473. GELOGE(ret, "[Call][InferOriginFormat] Prepare Graph inferformat failed");
  1474. return ret;
  1475. }
  1476. ret = ResourcePairProcess("remove");
  1477. if (ret != SUCCESS) {
  1478. return ret;
  1479. }
  1480. return ret;
  1481. }
  1482. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1483. PassManager control_pass;
  1484. // Graph pass tmp logic for resource infershape
  1485. if (options_.train_graph_flag) {
  1486. try {
  1487. if (action == "add") {
  1488. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1489. } else {
  1490. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1491. new ResourcePairRemoveControlPass);
  1492. }
  1493. } catch (std::bad_alloc &e) {
  1494. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add ResourcePair Pass");
  1495. GELOGE(INTERNAL_ERROR, "[Add][Pass] failed, bad memory allocation occur, action:%s.", action.c_str());
  1496. return INTERNAL_ERROR;
  1497. }
  1498. }
  1499. Status ret = control_pass.Run(compute_graph_);
  1500. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1501. GELOGE(ret, "[Run][ResourcePairControlPass] failed, action:%s, ret:%u.", action.c_str(), ret);
  1502. return ret;
  1503. }
  1504. return SUCCESS;
  1505. }
  1506. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1507. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1508. GE_CHECK_NOTNULL(node_ptr);
  1509. if (node_ptr->GetType() == DATA) {
  1510. uint32_t index = 0;
  1511. auto op_desc = node_ptr->GetOpDesc();
  1512. GE_CHECK_NOTNULL(op_desc);
  1513. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1514. Format storage_format = FORMAT_RESERVED;
  1515. vector<int64_t> dst_shape_dims;
  1516. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1517. GELOGE(INTERNAL_ERROR, "[Get][StorageFormatAndShape] for input failed, op:%s, index:0",
  1518. op_desc->GetName().c_str());
  1519. return FAILED;
  1520. }
  1521. if (storage_format == FORMAT_RESERVED) {
  1522. continue;
  1523. }
  1524. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1525. GELOGE(INTERNAL_ERROR, "[Modify][DataNetOutputFormatAndShape] for input failed, op:%s, index:0",
  1526. op_desc->GetName().c_str());
  1527. return FAILED;
  1528. }
  1529. }
  1530. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1531. auto op_desc = node_ptr->GetOpDesc();
  1532. GE_CHECK_NOTNULL(op_desc);
  1533. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1534. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1535. Format storage_format = FORMAT_RESERVED;
  1536. vector<int64_t> dst_shape_dims;
  1537. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1538. GELOGE(INTERNAL_ERROR, "[Get][StorageFormatAndShape] from output failed, op:%s, index:%u",
  1539. op_desc->GetName().c_str(), index);
  1540. return FAILED;
  1541. }
  1542. if (storage_format == FORMAT_RESERVED) {
  1543. continue;
  1544. }
  1545. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1546. GELOGE(INTERNAL_ERROR, "[Modify][DataNetOutputFormatAndShape] for output failed, op:%s, index:%u",
  1547. op_desc->GetName().c_str(), index);
  1548. return FAILED;
  1549. }
  1550. }
  1551. }
  1552. }
  1553. return SUCCESS;
  1554. }
  1555. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1556. if (options_.save_original_model == "true") {
  1557. ModelHelper model_helper;
  1558. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1559. options_.original_model_file);
  1560. if (ret != SUCCESS) {
  1561. // If save original model fail, process continue
  1562. GELOGW("SaveOriginalGraphToOmModel fail");
  1563. }
  1564. }
  1565. return SUCCESS;
  1566. }
  1567. #define PP_RUN_AND_DUMP(name, func, ...) \
  1568. do { \
  1569. GE_RUN(Prepare, func, __VA_ARGS__); \
  1570. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1571. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1572. } while (0)
  1573. #define PP_RUN(name, func, ...) \
  1574. do { \
  1575. GE_RUN(Prepare, func, __VA_ARGS__); \
  1576. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1577. } while (0)
  1578. Status GraphPrepare::PrepareDynShape(const GraphNodePtr &graph_node, const std::vector<GeTensor> &user_input,
  1579. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1580. GE_CHECK_NOTNULL(graph_node->GetGraph());
  1581. GE_CHECK_NOTNULL(compute_graph);
  1582. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1583. const Graph &const_graph = *graph_node->GetGraph();
  1584. PP_RUN("Init", Init, const_graph, session_id);
  1585. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1586. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input, graph_node->GetOptions());
  1587. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1588. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1589. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1590. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1591. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1592. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1593. PP_RUN_AND_DUMP("CtrlFlowPreProcess", CtrlFlowPreProcess);
  1594. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1595. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1596. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1597. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1598. return SUCCESS;
  1599. }
  1600. Status GraphPrepare::CtrlFlowPreProcess() {
  1601. PassManager graph_pass;
  1602. // After InferShape Mark v1 control flow for unknown shape.
  1603. auto mark_force_unknown_pass = new (std::nothrow) MarkForceUnknownForCondPass;
  1604. GE_CHK_STATUS_RET(graph_pass.AddPass("PreRun::MarkForceUnknownForCondPass", mark_force_unknown_pass));
  1605. GE_CHK_STATUS_RET(graph_pass.Run(compute_graph_));
  1606. return SUCCESS;
  1607. }
  1608. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1609. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1610. return SUCCESS;
  1611. }
  1612. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1613. auto compute_graph = compute_graph_;
  1614. PassManager pass_manager;
  1615. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1616. new (std::nothrow) VariablePrepareOpPass))
  1617. GE_TIMESTAMP_START(pass_manager);
  1618. auto ret = pass_manager.Run(compute_graph);
  1619. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1620. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1621. GELOGE(ret, "[Run][Passes] for running format refiner failed, ret:%u.", ret);
  1622. return ret;
  1623. }
  1624. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1625. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1626. return SUCCESS;
  1627. }
  1628. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1629. if (compute_graph == nullptr) {
  1630. REPORT_INNER_ERROR("E19999", "Param compute_graph is nullptr, check invalid");
  1631. GELOGE(GE_GRAPH_NULL_INPUT, "[Check][Param] Input Graph is NULL");
  1632. return GE_GRAPH_NULL_INPUT;
  1633. }
  1634. GEPass ge_passes(compute_graph);
  1635. NamesToPass hccl_group;
  1636. HcclGroupPass hccl_group_pass;
  1637. GELOGD("Add hccl group pass success");
  1638. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1639. auto ret = ge_passes.Run(hccl_group);
  1640. if (ret != SUCCESS) {
  1641. GELOGE(ret, "[Run][HcclGroupPass] pass for preprocess failed, ret:%u.", ret);
  1642. return ret;
  1643. }
  1644. ret = compute_graph->TopologicalSorting();
  1645. if (ret != SUCCESS) {
  1646. REPORT_CALL_ERROR("E19999", "Topological sorting failed");
  1647. GELOGE(ret, "[Call][TopologicalSorting] Graph topological sort failed, ret:%u.", ret);
  1648. return ret;
  1649. }
  1650. return SUCCESS;
  1651. }
  1652. #undef PP_RUN_AND_DUMP
  1653. #undef PP_RUN
  1654. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1655. if (graph == nullptr) {
  1656. REPORT_INNER_ERROR("E19999", "Param graph is nullptr, check invalid");
  1657. GELOGE(GE_GRAPH_NULL_INPUT, "[Check][Param] Input Graph is NULL");
  1658. return GE_GRAPH_NULL_INPUT;
  1659. }
  1660. const Graph &const_graph = *graph;
  1661. Status ret = Init(const_graph, 0);
  1662. if (ret != SUCCESS) {
  1663. GELOGE(ret, "[Init][GraphPrepare] fail, ret:%u", ret);
  1664. return ret;
  1665. }
  1666. GE_DUMP(compute_graph_, "after_parser");
  1667. GELOGI("Start infershape for dump json process.");
  1668. ret = compute_graph_->InferOriginFormat();
  1669. GE_DUMP(compute_graph_, "after_inferformat");
  1670. if (ret != SUCCESS) {
  1671. REPORT_CALL_ERROR("E19999", "Infer OriginFormat failed");
  1672. GELOGE(ret, "[Infer][OriginFormat] failed");
  1673. return ret;
  1674. }
  1675. InferShapePass infer_shape_pass;
  1676. NamesToPass names_to_passes;
  1677. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1678. GEPass ge_passes(compute_graph_);
  1679. ret = ge_passes.Run(names_to_passes);
  1680. GE_DUMP(compute_graph_, "after_infershape");
  1681. if (ret != SUCCESS) {
  1682. GELOGE(ret, "[Run][GePasses] infershape for preprocess failed, ret:%u.", ret);
  1683. return ret;
  1684. }
  1685. ShapeRefiner::ClearContextMap();
  1686. return SUCCESS;
  1687. }
  1688. Status GraphPrepare::CheckConstOp() {
  1689. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1690. GE_CHECK_NOTNULL(node_ptr);
  1691. if (node_ptr->GetType() == CONSTANT) {
  1692. Status ret = VerifyConstOp(node_ptr);
  1693. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1694. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1695. auto op_desc = node_ptr->GetOpDesc();
  1696. if (op_desc == nullptr) {
  1697. REPORT_INNER_ERROR("E19999", "op_desc is nullptr, check invalid");
  1698. GELOGE(PARAM_INVALID, "[Get][OpDesc] of node failed, op_desc is nullptr, node type:FRAMEWORKOP.");
  1699. return PARAM_INVALID;
  1700. }
  1701. std::string original_type;
  1702. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1703. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1704. GELOGI("original type is %s", original_type.c_str());
  1705. if (original_type == CONSTANT) {
  1706. Status ret = VerifyConstOp(node_ptr);
  1707. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1708. }
  1709. }
  1710. }
  1711. return SUCCESS;
  1712. }
  1713. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1714. GE_CHECK_NOTNULL(node);
  1715. auto op_desc = node->GetOpDesc();
  1716. GE_CHECK_NOTNULL(op_desc);
  1717. ConstGeTensorPtr ge_tensor_ptr;
  1718. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1719. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1720. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  1721. GELOGE(PARAM_INVALID, "[Get][Attr] %s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1722. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  1723. return PARAM_INVALID;
  1724. }
  1725. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1726. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1727. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1728. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1729. auto data_type = ge_tensor_desc.GetDataType();
  1730. if (data_type == DT_STRING) {
  1731. return SUCCESS;
  1732. }
  1733. uint32_t length = 1;
  1734. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1735. if (!type_ret) {
  1736. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1737. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1738. GELOGE(PARAM_INVALID, "[Check][Param] Input datatype %s is not support.",
  1739. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1740. return FAILED;
  1741. }
  1742. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1743. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1744. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1745. if (shape_size == 0) {
  1746. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1747. // shape = [], means it's a sclar tensor.
  1748. GE_CHK_BOOL_EXEC(data_size / length == 1,
  1749. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid scalar tensor."});
  1750. return PARAM_INVALID, "[Check][Param] Const is invalid scalar tensor.");
  1751. } else {
  1752. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1753. GE_CHK_BOOL_EXEC(data_size == 0,
  1754. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid vector scalar."});
  1755. return PARAM_INVALID, "[Check][Param] Const is invalid vector scalar.");
  1756. }
  1757. } else {
  1758. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0,
  1759. ErrorManager::GetInstance().ATCReportErrMessage(
  1760. "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"});
  1761. return PARAM_INVALID, "[Check][Param] Const input data size is not equal with tensor desc shape");
  1762. }
  1763. return SUCCESS;
  1764. }
  1765. bool GraphPrepare::IsDynamicDims(const NodePtr &input_node) {
  1766. auto data_shape = NodeUtils::GetOutputDesc(*input_node, kDataOutIndex).GetShape();
  1767. const auto &dims = data_shape.GetDims();
  1768. bool all_is_positive = false;
  1769. if (std::all_of(dims.begin(), dims.end(), [](int64_t val) { return val >= 0; })) {
  1770. all_is_positive = true;
  1771. }
  1772. if (!all_is_positive && !options_.input_shape.empty() && !options_.dynamic_dims.empty() &&
  1773. options_.dynamic_node_type != kInvalidDynaimcDimsType) {
  1774. GELOGI("No need to check and update desc info, the dims of %s is %s.", input_node->GetName().c_str(),
  1775. formats::JoinToString(dims).c_str());
  1776. return true;
  1777. }
  1778. return false;
  1779. }
  1780. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1781. if (GetLocalOmgContext().is_dynamic_input) {
  1782. return SUCCESS;
  1783. }
  1784. unsigned int node_num = 0;
  1785. unsigned int data_num = 0;
  1786. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1787. GE_CHECK_NOTNULL(input_node);
  1788. OpDescPtr op = input_node->GetOpDesc();
  1789. GE_CHECK_NOTNULL(op);
  1790. node_num++;
  1791. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1792. data_num++;
  1793. GeAttrValue::INT index = 0;
  1794. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1795. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1796. op->GetName().c_str(), op->GetType().c_str());
  1797. GELOGE(GE_GRAPH_INIT_FAILED, "[Get][Attr] %s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1798. op->GetName().c_str(), op->GetType().c_str());
  1799. return GE_GRAPH_INIT_FAILED;
  1800. }
  1801. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1802. std::string situation = "data op index[" + std::to_string(index) + "]";
  1803. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1804. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1805. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][Param] user_input size:%zu must larger than data op index:%ld.",
  1806. user_input.size(), index);
  1807. return GE_GRAPH_INIT_FAILED;
  1808. }
  1809. if (IsDynamicDims(input_node)) {
  1810. continue;
  1811. }
  1812. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1813. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1814. int64_t dim = desc.GetShape().GetDim(i);
  1815. if (dim < UNKNOWN_DIM_NUM) {
  1816. std::string situation = "data dim[" + std::to_string(i) + "][" + std::to_string(dim) + "]" ;
  1817. std::string reason = "it need >= -2";
  1818. REPORT_INPUT_ERROR(
  1819. "E19025", std::vector<std::string>({"situation", "reason"}), std::vector<std::string>({situation, reason}));
  1820. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][InputDim]data dim %zu is not supported, need >= -2, real:%ld.", i, dim);
  1821. return GE_GRAPH_INIT_FAILED;
  1822. }
  1823. }
  1824. }
  1825. }
  1826. if (node_num <= data_num) {
  1827. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1828. }
  1829. return SUCCESS;
  1830. }
  1831. Status GraphPrepare::InferShapeForPreprocess() {
  1832. GELOGI("Start infershape for preprocess.");
  1833. GEPass ge_passes(compute_graph_);
  1834. NamesToPass names_to_passes;
  1835. AssertPass assert_pass;
  1836. if (!options_.train_graph_flag) {
  1837. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1838. }
  1839. SwitchDeadBranchElimination switch_dead_branch_elimination;
  1840. names_to_passes.emplace_back("SwitchDeadBranchElimination", &switch_dead_branch_elimination);
  1841. MergePass merge_pass;
  1842. names_to_passes.emplace_back("MergePass", &merge_pass);
  1843. InferShapePass infer_shape_pass;
  1844. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1845. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1846. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1847. DimensionComputePass dimension_compute_pass;
  1848. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1849. ConstantFoldingPass constant_folding_pass;
  1850. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1851. int32_t dev_count = 0;
  1852. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1853. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1854. rtError_t rt_err = RT_ERROR_NONE;
  1855. if (aicpu_constant_folding_on != nullptr) {
  1856. rt_err = rtGetDeviceCount(&dev_count);
  1857. if (rt_err == RT_ERROR_NONE) {
  1858. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1859. if (result != SUCCESS) {
  1860. GELOGE(result, "[Set][RtContext] failed, mode = RT_CTX_NORMAL_MODE.");
  1861. return result;
  1862. }
  1863. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1864. }
  1865. }
  1866. Status ret = ge_passes.Run(names_to_passes);
  1867. if (aicpu_constant_folding_on != nullptr) {
  1868. if (rt_err == RT_ERROR_NONE) {
  1869. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1870. if (result != SUCCESS) {
  1871. GELOGE(result, "[Set][RtContext] failed, mode = RT_CTX_GEN_MODE.");
  1872. return result;
  1873. }
  1874. }
  1875. }
  1876. ShapeRefiner::ClearContextMap();
  1877. if (ret != SUCCESS) {
  1878. GELOGE(ret, "[Run][GePasses] infershape for preprocess failed, ret:%u.", ret);
  1879. return ret;
  1880. }
  1881. return SUCCESS;
  1882. }
  1883. Status GraphPrepare::PrepareOptimize() {
  1884. GELOGI("Start optimize for preprocess.");
  1885. // check rw type
  1886. GraphOptimize graph_optimize;
  1887. bool has_conflict = false;
  1888. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1889. if (has_conflict) {
  1890. GELOGE(GRAPH_PARAM_INVALID, "[Check][RWConflict] There has rw conflict.Stop optimize.");
  1891. return FAILED;
  1892. }
  1893. PassManager original_graph_passes;
  1894. // Graph pass
  1895. try {
  1896. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1897. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1898. (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass);
  1899. } catch (std::bad_alloc &e) {
  1900. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass");
  1901. GELOGE(INTERNAL_ERROR, "[Add][Pass] failed, bad memory allocation occurs.");
  1902. return INTERNAL_ERROR;
  1903. }
  1904. GE_TIMESTAMP_START(original_graph_passes);
  1905. Status ret = original_graph_passes.Run(compute_graph_);
  1906. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1907. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1908. GELOGE(ret, "[Run][GraphPasses] optimize for preprocess failed, ret:%u.", ret);
  1909. return ret;
  1910. }
  1911. // New pass
  1912. GEPass ge_passes(compute_graph_);
  1913. NamesToPass names_to_passes;
  1914. EnterPass enter_pass;
  1915. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1916. CondPass cond_pass;
  1917. names_to_passes.emplace_back("CondPass", &cond_pass);
  1918. PrintOpPass print_pass;
  1919. if (options_.enable_print_op_pass) {
  1920. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1921. }
  1922. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1923. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1924. DropOutPass dropout_pass;
  1925. AssertPass assert_pass;
  1926. UnusedConstPass unused_const_pass;
  1927. StopGradientPass stop_gradient_pass;
  1928. PreventGradientPass prevent_gradient_pass;
  1929. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1930. GuaranteeConstPass guarantee_const_pass;
  1931. VarIsInitializedOpPass var_is_initialized_pass;
  1932. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1933. IdentityPass identity_pass(false);
  1934. SnapshotPass snapshot_pass;
  1935. if (!options_.train_graph_flag) {
  1936. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1937. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1938. }
  1939. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1940. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1941. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1942. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1943. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1944. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1945. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1946. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1947. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1948. GE_TIMESTAMP_START(names_to_passes);
  1949. ret = ge_passes.Run(names_to_passes);
  1950. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1951. if (ret != SUCCESS) {
  1952. GELOGE(ret, "[Run][GePasses] optimize for preprocess failed, ret:%u.", ret);
  1953. return ret;
  1954. }
  1955. PassManager graph_pass;
  1956. try {
  1957. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1958. // can't move to optimize1/2 directly, may cause more identity insert, cause CI fail
  1959. (void)graph_pass.AddPass("PrepareOptimize::HcclMemcpyPass", new HcclMemcpyPass);
  1960. } catch (std::bad_alloc &e) {
  1961. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass");
  1962. GELOGE(INTERNAL_ERROR, "[Add][Pass] failed, bad memory allocation occurs.");
  1963. return INTERNAL_ERROR;
  1964. }
  1965. GE_TIMESTAMP_START(graph_passes);
  1966. ret = graph_pass.Run(compute_graph_);
  1967. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1968. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1969. GELOGE(ret, "[Run][GraphPasses] optimize for preprocess failed, ret:%u.", ret);
  1970. return ret;
  1971. }
  1972. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1973. TypeConversionOfConstant();
  1974. ret = compute_graph_->TopologicalSorting();
  1975. if (ret != SUCCESS) {
  1976. REPORT_CALL_ERROR("E19999", "Topological sorting failed");
  1977. GELOGE(ret, "[Call][TopologicalSorting] Graph topological sort failed, ret:%u.", ret);
  1978. return ret;
  1979. }
  1980. GELOGI("End optimize for preprocess.");
  1981. return SUCCESS;
  1982. }
  1983. void GraphPrepare::TypeConversionOfConstant() {
  1984. bool is_acl_compile = false;
  1985. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1986. // This can ensure that n is not a null pointer
  1987. // No Conversion when called by aclOpCompile
  1988. (void)AttrUtils::GetBool(n->GetOpDesc(), ATTR_SINGLE_OP_SCENE, is_acl_compile);
  1989. if (is_acl_compile) {
  1990. return;
  1991. }
  1992. }
  1993. if (options_.train_graph_flag) {
  1994. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1995. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1996. // This can ensure that n is not a null pointer
  1997. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1998. n->GetOpDesc()->SetType(CONSTANTOP);
  1999. }
  2000. }
  2001. } else {
  2002. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  2003. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  2004. // This can ensure that n is not a null pointer
  2005. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  2006. n->GetOpDesc()->SetType(CONSTANT);
  2007. }
  2008. }
  2009. }
  2010. }
  2011. Status GraphPrepare::GraphEquivalentTransformation() {
  2012. NamesToPass names_to_pass;
  2013. ForPass for_pass;
  2014. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  2015. return GEPass(compute_graph_).Run(names_to_pass);
  2016. }
  2017. Status GraphPrepare::ProcessBeforeInfershape() {
  2018. NamesToPass names_to_passes;
  2019. CondRemovePass condition_remove_pass;
  2020. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  2021. GE_TIMESTAMP_START(ProcessCondRemove);
  2022. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  2023. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  2024. if (ret != SUCCESS) {
  2025. GELOGE(ret, "[Run][GEPass] optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  2026. return ret;
  2027. }
  2028. return SUCCESS;
  2029. }
  2030. Status GraphPrepare::ProcessNetOutput() {
  2031. PassManager graph_passes_before_infershape;
  2032. try {
  2033. if (options_.train_graph_flag) {
  2034. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  2035. }
  2036. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  2037. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  2038. new (std::nothrow) DataPass); // Add NetOutput first.
  2039. } catch (std::bad_alloc) {
  2040. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass");
  2041. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  2042. return INTERNAL_ERROR;
  2043. }
  2044. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  2045. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  2046. GELOGE(ret, "[Run][GraphPasses] before Infershape failed, ret:%d.", ret);
  2047. return ret;
  2048. }
  2049. return SUCCESS;
  2050. }
  2051. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input,
  2052. const std::map<string, string> &graph_option) {
  2053. compute_graph_->SetInputSize(user_input.size());
  2054. if (user_input.empty()) {
  2055. return SUCCESS;
  2056. }
  2057. auto ret = CheckUserInput(user_input);
  2058. if (ret != SUCCESS) {
  2059. GELOGE(ret, "[Check][UserInput] failed, ret:%u", ret);
  2060. return ret;
  2061. }
  2062. ret = UpdateInput(user_input, graph_option);
  2063. if (ret != SUCCESS) {
  2064. GELOGE(ret, "[Update][Input] fail, ret:%u", ret);
  2065. return ret;
  2066. }
  2067. if (user_input.size() != 0) {
  2068. ret = CheckConstOp();
  2069. if (ret != SUCCESS) {
  2070. GELOGE(ret, "[Check][ConstOp] fail, ret:%u", ret);
  2071. return ret;
  2072. }
  2073. } else {
  2074. ret = compute_graph_->TopologicalSorting();
  2075. if (ret != SUCCESS) {
  2076. REPORT_CALL_ERROR("E19999", "Topological sorting failed");
  2077. GELOGE(ret, "[Call][TopologicalSorting] failed.");
  2078. return FAILED;
  2079. }
  2080. }
  2081. return SUCCESS;
  2082. }
  2083. Status GraphPrepare::UpdateInputOutputByOptions() {
  2084. auto ret = UpdateDataNetOutputByStorageFormat();
  2085. if (ret != SUCCESS) {
  2086. GELOGE(ret, "[Update][DataNetOutputByStorageFormat] failed.");
  2087. return ret;
  2088. }
  2089. if (options_.train_graph_flag) {
  2090. GELOGI("This is train mode, no need to do this schedule.");
  2091. return SUCCESS;
  2092. }
  2093. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  2094. GE_CHECK_NOTNULL(node_ptr);
  2095. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  2096. GELOGE(INTERNAL_ERROR, "[Set][NdFormat] for node:%s failed", node_ptr->GetName().c_str());
  2097. return FAILED;
  2098. }
  2099. if (node_ptr->GetType() == DATA) {
  2100. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  2101. GELOGE(INTERNAL_ERROR, "[Call][ProcessDataNodeDynShape] for node:%s failed", node_ptr->GetName().c_str());
  2102. return FAILED;
  2103. }
  2104. }
  2105. if (node_ptr->GetType() == ge::NETOUTPUT) {
  2106. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  2107. GELOGE(INTERNAL_ERROR, "[Call][ProcessNetoutputNodeDynShape] for node:%s failed", node_ptr->GetName().c_str());
  2108. return FAILED;
  2109. }
  2110. }
  2111. }
  2112. return SUCCESS;
  2113. }
  2114. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  2115. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  2116. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  2117. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  2118. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  2119. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  2120. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  2121. if (dst_node->GetType() == TRANSDATA) {
  2122. return true;
  2123. }
  2124. }
  2125. }
  2126. return false;
  2127. }
  2128. } // 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