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 99 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
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
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
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
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
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
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
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
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
4 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
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248
  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 "graph/common/ge_call_wrapper.h"
  30. #include "graph/common/local_context.h"
  31. #include "graph/common/transop_util.h"
  32. #include "graph/ge_context.h"
  33. #include "graph/shape_refiner.h"
  34. #include "graph/manager/graph_var_manager.h"
  35. #include "graph/manager/util/rt_context_util.h"
  36. #include "graph/optimize/graph_optimize.h"
  37. #include "graph/passes/addn_pass.h"
  38. #include "graph/passes/aicpu_constant_folding_pass.h"
  39. #include "graph/passes/assert_pass.h"
  40. #include "ge/ge_api_types.h"
  41. #include "graph/passes/common_subexpression_elimination_pass.h"
  42. #include "graph/passes/cond_pass.h"
  43. #include "graph/passes/cond_remove_pass.h"
  44. #include "graph/passes/constant_folding_pass.h"
  45. #include "graph/passes/dimension_adjust_pass.h"
  46. #include "graph/passes/dimension_compute_pass.h"
  47. #include "graph/passes/dropout_pass.h"
  48. #include "graph/passes/enter_pass.h"
  49. #include "graph/passes/for_pass.h"
  50. #include "graph/passes/guarantee_const_pass.h"
  51. #include "graph/passes/hccl_memcpy_pass.h"
  52. #include "graph/passes/hccl_group_pass.h"
  53. #include "graph/passes/identity_pass.h"
  54. #include "graph/passes/infershape_pass.h"
  55. #include "graph/passes/merge_pass.h"
  56. #include "graph/passes/net_output_pass.h"
  57. #include "graph/passes/no_use_reshape_remove_pass.h"
  58. #include "graph/passes/parallel_concat_start_op_pass.h"
  59. #include "graph/passes/placeholder_with_default_pass.h"
  60. #include "graph/passes/prevent_gradient_pass.h"
  61. #include "graph/passes/print_op_pass.h"
  62. #include "graph/passes/prune_pass.h"
  63. #include "graph/passes/replace_transshape_pass.h"
  64. #include "graph/passes/replace_with_empty_const_pass.h"
  65. #include "graph/passes/resource_pair_add_control_pass.h"
  66. #include "graph/passes/resource_pair_remove_control_pass.h"
  67. #include "graph/passes/save_pass.h"
  68. #include "graph/passes/shape_operate_op_remove_pass.h"
  69. #include "graph/passes/snapshot_pass.h"
  70. #include "graph/passes/stop_gradient_pass.h"
  71. #include "graph/passes/switch_dead_branch_elimination.h"
  72. #include "graph/passes/unused_const_pass.h"
  73. #include "graph/passes/var_is_initialized_op_pass.h"
  74. #include "graph/passes/variable_prepare_op_pass.h"
  75. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  76. #include "graph/utils/type_utils.h"
  77. #include "inc/pass_manager.h"
  78. #include "init/gelib.h"
  79. #include "multi_batch_copy_graph.h"
  80. #include "graph/passes/data_pass.h"
  81. #include "graph/passes/mark_agnostic_pass.h"
  82. namespace ge {
  83. namespace {
  84. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  85. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  86. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  87. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  88. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  89. // the size of user defined output datatype or format string after split by ":".
  90. const size_t kUserDefinedElementCount = 2;
  91. const int kDataOutIndex = 0;
  92. const int64_t kInvalidDynaimcDimsType = -1;
  93. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  94. GeTensorPtr tensor = MakeShared<GeTensor>();
  95. if (tensor == nullptr) {
  96. REPORT_CALL_ERROR("E19999", "New GeTensor failed");
  97. GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed");
  98. return nullptr;
  99. }
  100. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  101. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  102. auto dst_ge_shape = data_tensor.GetShape();
  103. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  104. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  105. tensor->MutableTensorDesc().SetShape(GeShape());
  106. int32_t dst_shape = 1;
  107. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  108. REPORT_CALL_ERROR("E19999", "Set data to tensor failed");
  109. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  110. return nullptr;
  111. }
  112. } else {
  113. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  114. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  115. if (dst_shape == nullptr) {
  116. REPORT_CALL_ERROR("E19999", "Malloc buffer failed, size:%zu", dim_cnt);
  117. GELOGE(INTERNAL_ERROR, "Create unique ptr failed");
  118. return nullptr;
  119. }
  120. for (int64_t i = 0; i < dim_cnt; ++i) {
  121. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  122. }
  123. GE_IF_BOOL_EXEC(
  124. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  125. REPORT_CALL_ERROR("E19999", "Set data to tensor failed");
  126. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  127. return nullptr;)
  128. }
  129. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  130. return OpDescUtils::CreateConstOp(tensor);
  131. }
  132. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  133. OpDescPtr &op_desc) {
  134. // For format transfer node, the IR definition has src/dst format attrs
  135. if (node_type == TRANSDATA) {
  136. GE_IF_BOOL_EXEC(
  137. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  138. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  139. GE_IF_BOOL_EXEC(
  140. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  141. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  142. }
  143. // For TransposeD node, the IR definition has perm attrs
  144. if (node_type == TRANSPOSED) {
  145. Format src_format = input.GetFormat();
  146. Format dst_format = output.GetFormat();
  147. std::vector<int64_t> perm_arg;
  148. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  149. "Get perm by foramt failed.");
  150. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  151. "SetStr PERMUTE_ATTR_PERM failed")
  152. }
  153. // For cast node, the IR definition has src/dst attrs
  154. if (node_type == CAST) {
  155. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  156. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  157. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  158. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  159. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  160. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  161. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  162. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  163. }
  164. }
  165. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  166. const GeTensorDesc &output, NodePtr &node) {
  167. if (node == nullptr) {
  168. REPORT_INNER_ERROR("E19999", "Param node is nullptr, trans_name:%s, trans_type:%s, check invalid",
  169. name.c_str(), node_type.c_str());
  170. GELOGE(PARAM_INVALID, "node is null.");
  171. return nullptr;
  172. }
  173. auto graph = node->GetOwnerComputeGraph();
  174. if (graph == nullptr) {
  175. REPORT_INNER_ERROR("E19999", "Owner graph in node is nullptr, trans_name:%s, trans_type:%s, check invalid",
  176. name.c_str(), node_type.c_str());
  177. GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str());
  178. return nullptr;
  179. }
  180. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  181. if (index < 0) {
  182. ErrorManager::GetInstance().ATCReportErrMessage(
  183. "E19025", {"situation", "reason"},
  184. {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()});
  185. GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str());
  186. return nullptr;
  187. }
  188. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  189. if (op_desc == nullptr) {
  190. REPORT_CALL_ERROR("E19999", "New OpDesc failed, trans_name:%s, trans_type:%s,",
  191. name.c_str(), node_type.c_str());
  192. GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed");
  193. return nullptr;
  194. }
  195. // for data dump
  196. GE_IF_BOOL_EXEC(
  197. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  198. GELOGW("CreateTransNode: SetListStr failed");)
  199. // Default single input and single output
  200. auto ret = op_desc->AddInputDesc(input);
  201. if (ret != GRAPH_SUCCESS) {
  202. REPORT_CALL_ERROR("E19999", "Add input desc into op:%s(%s) failed",
  203. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  204. GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str());
  205. return nullptr;
  206. }
  207. ret = op_desc->AddOutputDesc(output);
  208. if (ret != GRAPH_SUCCESS) {
  209. REPORT_CALL_ERROR("E19999", "Add output desc into op:%s(%s) failed",
  210. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  211. GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str());
  212. return nullptr;
  213. }
  214. AddTransNodeAttr(node_type, input, output, op_desc);
  215. NodePtr shape_node = nullptr;
  216. if (node_type == RESHAPE) {
  217. auto shape_desc = CreateTensorShape(output);
  218. if (shape_desc == nullptr) {
  219. GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input",
  220. node->GetName().c_str());
  221. return nullptr;
  222. }
  223. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  224. if (ret != GRAPH_SUCCESS) {
  225. REPORT_CALL_ERROR("E19999", "Add input desc into op:%s(%s) failed",
  226. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  227. GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str());
  228. return nullptr;
  229. }
  230. shape_node = graph->AddNode(shape_desc);
  231. if (shape_node == nullptr) {
  232. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed",
  233. shape_desc->GetName().c_str(), shape_desc->GetType().c_str(),
  234. graph->GetName().c_str());
  235. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str());
  236. return nullptr;
  237. }
  238. }
  239. auto trans_node = graph->AddNode(op_desc);
  240. if (trans_node == nullptr) {
  241. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed",
  242. op_desc->GetName().c_str(), op_desc->GetType().c_str(),
  243. graph->GetName().c_str());
  244. GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str());
  245. return nullptr;
  246. }
  247. if (node_type == RESHAPE) {
  248. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  249. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  250. shape_node->GetName().c_str(), shape_node->GetType().c_str(),
  251. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  252. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str());
  253. return nullptr;
  254. }
  255. }
  256. return trans_node;
  257. }
  258. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  259. NodePtr &trans_node) {
  260. GE_CHECK_NOTNULL(node);
  261. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  262. if (trans_node == nullptr) {
  263. return INTERNAL_ERROR;
  264. }
  265. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  266. if (ret != GRAPH_SUCCESS) {
  267. REPORT_CALL_ERROR("E19999", "Replace out anchors of node:%s(%s) by node:%s(%s) failed",
  268. node->GetName().c_str(), node->GetType().c_str(),
  269. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  270. GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s",
  271. node->GetName().c_str(), node->GetType().c_str());
  272. return INTERNAL_ERROR;
  273. }
  274. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  275. if (ret != GRAPH_SUCCESS) {
  276. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  277. node->GetName().c_str(), node->GetType().c_str(),
  278. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  279. GELOGE(INTERNAL_ERROR, "Failed to connect node %s to trans node %s", node->GetName().c_str(),
  280. trans_node->GetName().c_str());
  281. return INTERNAL_ERROR;
  282. }
  283. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  284. if (ret != GRAPH_SUCCESS) {
  285. REPORT_CALL_ERROR("E19999", "Move out control edges from node:%s(%s) to node:%s(%s) failed",
  286. node->GetName().c_str(), node->GetType().c_str(),
  287. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  288. GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.",
  289. node->GetName().c_str(), trans_node->GetName().c_str());
  290. return INTERNAL_ERROR;
  291. }
  292. return SUCCESS;
  293. }
  294. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  295. NodePtr &trans_node) {
  296. GE_CHECK_NOTNULL(node);
  297. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  298. if (trans_node == nullptr) {
  299. return INTERNAL_ERROR;
  300. }
  301. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  302. if (ret != GRAPH_SUCCESS) {
  303. REPORT_CALL_ERROR("E19999", "Replace out anchors of node:%s(%s) by node:%s(%s) failed",
  304. node->GetName().c_str(), node->GetType().c_str(),
  305. trans_node->GetName().c_str(), trans_node->GetType().c_str());
  306. GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s",
  307. node->GetName().c_str(), node->GetType().c_str());
  308. return INTERNAL_ERROR;
  309. }
  310. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  311. if (ret != GRAPH_SUCCESS) {
  312. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed",
  313. trans_node->GetName().c_str(), trans_node->GetType().c_str(),
  314. node->GetName().c_str(), node->GetType().c_str());
  315. GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(),
  316. node->GetName().c_str());
  317. return INTERNAL_ERROR;
  318. }
  319. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  320. if (ret != GRAPH_SUCCESS) {
  321. REPORT_CALL_ERROR("E19999", "Move out control edges from node:%s(%s) to 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. GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.",
  325. node->GetName().c_str(), trans_node->GetName().c_str());
  326. return INTERNAL_ERROR;
  327. }
  328. return SUCCESS;
  329. }
  330. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  331. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  332. GE_CHECK_NOTNULL(var->GetOpDesc());
  333. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  334. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  335. output_desc.SetFormat(tensor_desc.GetFormat());
  336. output_desc.SetDataType(tensor_desc.GetDataType());
  337. output_desc.SetShape(tensor_desc.GetShape());
  338. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  339. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  340. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  341. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  342. REPORT_CALL_ERROR("E19999", "Update output desc of node:%s(%s) failed, index:0,",
  343. var->GetName().c_str(), var->GetType().c_str());
  344. GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed");
  345. return INTERNAL_ERROR;);
  346. }
  347. if (var->GetOpDesc()->GetInputsSize() > 0) {
  348. auto desc = var->GetOpDesc()->GetInputDesc(0);
  349. desc.SetFormat(tensor_desc.GetFormat());
  350. desc.SetDataType(tensor_desc.GetDataType());
  351. desc.SetShape(tensor_desc.GetShape());
  352. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  353. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  354. desc.SetOriginShape(tensor_desc.GetOriginShape());
  355. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  356. REPORT_CALL_ERROR("E19999", "Update input desc of node:%s(%s) failed, index:0,",
  357. var->GetName().c_str(), var->GetType().c_str());
  358. GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed");
  359. return INTERNAL_ERROR;)
  360. }
  361. return SUCCESS;
  362. }
  363. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  364. GE_CHECK_NOTNULL(var);
  365. int index = 0;
  366. NodePtr last_node = var;
  367. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  368. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  369. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  370. if (ret != SUCCESS) {
  371. ErrorManager::GetInstance().ATCReportErrMessage(
  372. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  373. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(),
  374. index, iter->node_type.c_str());
  375. return INTERNAL_ERROR;
  376. }
  377. // set stream_label
  378. OpDescPtr var_desc = var->GetOpDesc();
  379. GE_CHECK_NOTNULL(var_desc);
  380. std::string stream_label;
  381. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  382. if (!stream_label.empty()) {
  383. auto status = SetStreamLabel(last_node, stream_label);
  384. if (status != ge::SUCCESS) {
  385. REPORT_CALL_ERROR("E19999", "Set stream_label:%s to op:%s(%s) failed",
  386. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  387. GELOGE(status, "Set stream label failed.");
  388. return status;
  389. }
  390. }
  391. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  392. REPORT_CALL_ERROR("E19999", "Set Attr:%s to node:%s(%s) failed",
  393. ge::ATTR_INSERTED_BY_GE.c_str(),
  394. last_node->GetName().c_str(), last_node->GetType().c_str());
  395. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  396. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  397. }
  398. if (road.empty()) {
  399. return SUCCESS;
  400. }
  401. return UpdateVarFormats(var, road.rbegin()->output);
  402. }
  403. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  404. for (auto &var : nodes) {
  405. GE_CHECK_NOTNULL(var);
  406. int index = 0;
  407. NodePtr last_node = var;
  408. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  409. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  410. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  411. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  412. if (ret != SUCCESS) {
  413. ErrorManager::GetInstance().ATCReportErrMessage(
  414. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  415. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s",
  416. var->GetName().c_str(), index, iter->node_type.c_str());
  417. return INTERNAL_ERROR;
  418. }
  419. // set stream_label
  420. OpDescPtr var_desc = var->GetOpDesc();
  421. GE_CHECK_NOTNULL(var_desc);
  422. std::string stream_label;
  423. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  424. if (!stream_label.empty()) {
  425. auto status = SetStreamLabel(last_node, stream_label);
  426. if (status != ge::SUCCESS) {
  427. REPORT_CALL_ERROR("E19999", "Set stream_label:%s to op:%s(%s) failed",
  428. stream_label.c_str(), last_node->GetName().c_str(), last_node->GetType().c_str());
  429. GELOGE(status, "Set stream label failed.");
  430. return status;
  431. }
  432. }
  433. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  434. REPORT_CALL_ERROR("E19999", "Set Attr:%s of node:%s(%s) failed",
  435. ge::ATTR_INSERTED_BY_GE.c_str(),
  436. last_node->GetName().c_str(), last_node->GetType().c_str());
  437. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  438. }
  439. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  440. return INTERNAL_ERROR;
  441. }
  442. }
  443. return SUCCESS;
  444. }
  445. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  446. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  447. VarNamesToRefs names_to_refs;
  448. std::string var_name;
  449. if (graph == nullptr) {
  450. REPORT_INNER_ERROR("E19999", "Param graph is nullptr, check invalid");
  451. GELOGE(PARAM_INVALID, "graph is null.");
  452. return names_to_refs;
  453. }
  454. for (auto &node : graph->GetAllNodes()) {
  455. if (node->GetType() != VARIABLE) {
  456. continue;
  457. }
  458. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  459. (void)names_to_refs[var_name].insert(node);
  460. }
  461. }
  462. return names_to_refs;
  463. }
  464. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  465. std::vector<int64_t> &dst_shape) {
  466. if (src_format == FORMAT_NCHW) {
  467. formats::FormatTransferNchwNc1hwc0 transfer;
  468. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  469. GELOGE(INTERNAL_ERROR, "TransShape failed");
  470. return FAILED;
  471. }
  472. } else if (src_format == FORMAT_NHWC) {
  473. formats::FormatTransferNhwcNc1hwc0 transfer;
  474. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  475. GELOGE(INTERNAL_ERROR, "TransShape failed");
  476. return FAILED;
  477. }
  478. }
  479. return SUCCESS;
  480. }
  481. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  482. GE_CHECK_NOTNULL(node_ptr);
  483. auto op_desc = node_ptr->GetOpDesc();
  484. GE_CHECK_NOTNULL(op_desc);
  485. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  486. GE_CHECK_NOTNULL(input);
  487. ge::Format old_format = input->GetFormat();
  488. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  489. ge::DataType dt = input->GetDataType();
  490. std::vector<int64_t> dst_shape_dims;
  491. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  492. REPORT_CALL_ERROR("E19999", "Transfer shape to NC1HWC0 failed, op:%s(%s),",
  493. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  494. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  495. return FAILED;
  496. }
  497. input->SetFormat(FORMAT_NC1HWC0);
  498. input->SetShape(ge::GeShape(dst_shape_dims));
  499. auto output = op_desc->MutableOutputDesc(0);
  500. GE_CHECK_NOTNULL(output);
  501. output->SetFormat(FORMAT_NC1HWC0);
  502. output->SetShape(ge::GeShape(dst_shape_dims));
  503. int64_t size = 0;
  504. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  505. if (graph_status != ge::GRAPH_SUCCESS) {
  506. REPORT_CALL_ERROR("E19999", "Get output tensor size failed, op:%s(%s), index:0",
  507. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  508. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  509. return FAILED;
  510. }
  511. ge::TensorUtils::SetSize(*output, size);
  512. ge::TensorUtils::SetSize(*input, size);
  513. return SUCCESS;
  514. }
  515. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  516. GE_CHECK_NOTNULL(input_output);
  517. ge::Format old_format = input_output->GetFormat();
  518. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  519. ge::DataType dt = input_output->GetDataType();
  520. std::vector<int64_t> dst_shape_dims;
  521. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  522. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  523. return FAILED;
  524. }
  525. input_output->SetFormat(FORMAT_NC1HWC0);
  526. input_output->SetShape(ge::GeShape(dst_shape_dims));
  527. return SUCCESS;
  528. }
  529. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  530. vector<int64_t> &dst_shape_dims) {
  531. GE_CHECK_NOTNULL(op_desc);
  532. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  533. GE_CHECK_NOTNULL(input);
  534. ge::Format old_format = input->GetFormat();
  535. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  536. input->SetShape(ge::GeShape(dst_shape_dims));
  537. input->SetFormat(storage_format);
  538. auto output = op_desc->MutableOutputDesc(index);
  539. GE_CHECK_NOTNULL(output);
  540. output->SetShape(ge::GeShape(dst_shape_dims));
  541. output->SetFormat(storage_format);
  542. if (!output->MutableShape().IsUnknownShape()) {
  543. int64_t size = 0;
  544. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  545. if (graph_status != ge::GRAPH_SUCCESS) {
  546. REPORT_CALL_ERROR("E19999", "Get output tensor size failed, op:%s(%s), index:%u",
  547. op_desc->GetName().c_str(), op_desc->GetType().c_str(), index);
  548. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  549. return FAILED;
  550. }
  551. ge::TensorUtils::SetSize(*input, size);
  552. ge::TensorUtils::SetSize(*output, size);
  553. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  554. "new_shape:%s, new_format:%s, new_size:%lu",
  555. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  556. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  557. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  558. }
  559. return SUCCESS;
  560. }
  561. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) {
  562. is_dynamic_batch = false;
  563. std::string related_node_name;
  564. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  565. if (related_node_name.empty()) {
  566. ErrorManager::GetInstance().ATCReportErrMessage(
  567. "E15002", {"opname", "value", "reason"}, {data_node->GetName(), "flag", "but the value is empty"});
  568. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty",
  569. data_node->GetName().c_str());
  570. return INTERNAL_ERROR;
  571. }
  572. for (const NodePtr &next_node : data_node->GetOutNodes()) {
  573. if (next_node->GetName() == related_node_name) {
  574. switchn_node = next_node;
  575. break;
  576. }
  577. }
  578. if (switchn_node == nullptr) {
  579. ErrorManager::GetInstance().ATCReportErrMessage(
  580. "E15002", {"opname", "value", "reason"},
  581. {data_node->GetName(), related_node_name, "but can not find it on the graph"});
  582. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph",
  583. data_node->GetName().c_str(), related_node_name.c_str());
  584. return INTERNAL_ERROR;
  585. }
  586. is_dynamic_batch = true;
  587. }
  588. return SUCCESS;
  589. }
  590. bool CheckOpType(const NodePtr &node, const std::string type) {
  591. if (node->GetType() == type) {
  592. return true;
  593. }
  594. return false;
  595. }
  596. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  597. auto op = node_ptr->GetOpDesc();
  598. GE_CHECK_NOTNULL(op);
  599. auto inputDescsPtr = op->GetAllInputsDescPtr();
  600. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  601. ge::Format format = ge::FORMAT_ND;
  602. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  603. // process, otherwise fe will insert transdata.
  604. for (auto &inputDescPtr : inputDescsPtr) {
  605. GE_CHECK_NOTNULL(inputDescPtr);
  606. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  607. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  608. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  609. inputDescPtr->SetFormat(format);
  610. inputDescPtr->SetOriginFormat(format);
  611. }
  612. }
  613. for (auto &outputDescPtr : outputDescsPtr) {
  614. GE_CHECK_NOTNULL(outputDescPtr);
  615. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  616. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  617. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  618. outputDescPtr->SetFormat(format);
  619. outputDescPtr->SetOriginFormat(format);
  620. }
  621. }
  622. return SUCCESS;
  623. }
  624. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  625. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  626. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  627. // data and netoutput, and no longer inserts the transnode.
  628. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) {
  629. GE_CHECK_NOTNULL(node_ptr);
  630. auto op_desc = node_ptr->GetOpDesc();
  631. GE_CHECK_NOTNULL(op_desc);
  632. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  633. GE_CHECK_NOTNULL(input);
  634. ge::DataType src_dtype = input->GetDataType();
  635. if (src_dtype == dt_set) {
  636. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  637. return SUCCESS;
  638. }
  639. input->SetDataType(dt_set);
  640. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  641. GE_CHECK_NOTNULL(output);
  642. output->SetDataType(dt_set);
  643. GeShape shape = input->GetShape();
  644. if (!shape.IsUnknownShape()) {
  645. int64_t input_shape_size = 0;
  646. int64_t output_shape_size = 0;
  647. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  648. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  649. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  650. REPORT_CALL_ERROR("E19999", "Get input tensor size failed, op:%s(%s), index:0",
  651. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  652. GELOGE(GRAPH_FAILED, "[Process][InputOp] Get tensor size of op [%s] failed!", node_ptr->GetName().c_str());
  653. return FAILED;
  654. }
  655. ge::TensorUtils::SetSize(*input, input_shape_size);
  656. ge::TensorUtils::SetSize(*output, output_shape_size);
  657. GELOGI("[Process][InputDynShape] Set input and output size of node [%s] success.", node_ptr->GetName().c_str());
  658. }
  659. if (is_dynamic_batch) {
  660. GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str());
  661. auto switchn_op_desc = switchn_node->GetOpDesc();
  662. GE_CHECK_NOTNULL(switchn_op_desc);
  663. auto switchn_input = switchn_op_desc->MutableInputDesc(0);
  664. GE_CHECK_NOTNULL(switchn_input);
  665. switchn_input->SetDataType(dt_set);
  666. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  667. const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i);
  668. GE_CHECK_NOTNULL(switchn_output);
  669. switchn_output->SetDataType(dt_set);
  670. }
  671. }
  672. return SUCCESS;
  673. }
  674. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  675. GE_CHECK_NOTNULL(node_ptr);
  676. auto op_desc = node_ptr->GetOpDesc();
  677. GE_CHECK_NOTNULL(op_desc);
  678. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  679. GE_CHECK_NOTNULL(input);
  680. ge::Format old_format = input->GetFormat();
  681. ge::GeShape old_shape = input->GetShape();
  682. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  683. if (!support) {
  684. ErrorManager::GetInstance().ATCReportErrMessage(
  685. "E19014", {"opname", "value", "reason"},
  686. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  687. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  688. GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str());
  689. return FAILED;
  690. }
  691. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  692. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  693. return FAILED;
  694. }
  695. if (is_dynamic_batch) {
  696. auto switchn_op_desc = switchn_node->GetOpDesc();
  697. GE_CHECK_NOTNULL(switchn_op_desc);
  698. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  699. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  700. REPORT_CALL_ERROR("E19999", "Modify format and shape of input:0 in op:%s(%s) failed",
  701. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  702. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  703. return FAILED;
  704. }
  705. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  706. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  707. GE_CHECK_NOTNULL(switchn_output);
  708. old_format = switchn_output->GetFormat();
  709. old_shape = switchn_output->GetShape();
  710. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  711. REPORT_CALL_ERROR("E19999", "Modify format and shape of output:%u in op:%s(%s) failed", i,
  712. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str());
  713. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  714. return FAILED;
  715. }
  716. }
  717. }
  718. return SUCCESS;
  719. }
  720. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  721. auto op_desc = node_ptr->GetOpDesc();
  722. GE_CHECK_NOTNULL(op_desc);
  723. string set_dt_str;
  724. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  725. return SUCCESS;
  726. }
  727. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  728. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  729. bool is_dynamic_batch = false;
  730. NodePtr switchn_node = nullptr;
  731. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) {
  732. GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed");
  733. return FAILED;
  734. }
  735. if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) {
  736. GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed");
  737. return FAILED;
  738. }
  739. // check if need to set format
  740. string set_format;
  741. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  742. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  743. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  744. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) {
  745. GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed");
  746. return FAILED;
  747. }
  748. }
  749. return SUCCESS;
  750. }
  751. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  752. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  753. GE_CHECK_NOTNULL(op_desc);
  754. GE_CHECK_NOTNULL(tensor_desc_ptr);
  755. storage_format = FORMAT_RESERVED;
  756. int64_t format = FORMAT_RESERVED;
  757. dst_shape_dims.clear();
  758. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  759. storage_format = static_cast<Format>(format);
  760. vector<int32_t> storage_shape;
  761. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  762. for (auto dim : storage_shape) {
  763. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  764. }
  765. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  766. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  767. formats::JoinToString(storage_shape).c_str());
  768. } else {
  769. ErrorManager::GetInstance().ATCReportErrMessage(
  770. "15003", {"opname", "format"},
  771. {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)});
  772. GELOGE(PARAM_INVALID, "Update node by storage format failed, storage_shape not set. "
  773. "node: [%s], storage_format [%s]",
  774. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  775. return FAILED;
  776. }
  777. ge::Format old_format = tensor_desc_ptr->GetFormat();
  778. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  779. if (old_format == storage_format && old_shape == dst_shape_dims) {
  780. GELOGI("Update node by storage format, not changed.");
  781. storage_format = FORMAT_RESERVED;
  782. return SUCCESS;
  783. }
  784. }
  785. return SUCCESS;
  786. }
  787. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  788. NodePtr &node) {
  789. bool is_dynamic = CheckOpType(node, MERGE);
  790. auto src_op_desc = node->GetOpDesc();
  791. GE_CHECK_NOTNULL(src_op_desc);
  792. ge::GeShape src_shape = src_desc.GetShape();
  793. ge::Format src_format = src_desc.GetFormat();
  794. net_output_input_desc->SetDataType(DT_FLOAT16);
  795. if (is_dynamic) {
  796. auto merge_output = src_op_desc->MutableOutputDesc(0);
  797. GE_CHECK_NOTNULL(merge_output);
  798. merge_output->SetDataType(DT_FLOAT16);
  799. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  800. auto merge_input = src_op_desc->MutableInputDesc(i);
  801. GE_CHECK_NOTNULL(merge_input);
  802. merge_input->SetDataType(DT_FLOAT16);
  803. }
  804. }
  805. std::vector<int64_t> dst_shape_dims;
  806. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  807. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  808. REPORT_CALL_ERROR("E19999", "Transfer output:0 shape of op:%s(%s) to NC1HWC0 format failed, shape:%s, format:%s",
  809. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str(),
  810. src_shape.ToString().c_str(), TypeUtils::FormatToSerialString(src_format).c_str());
  811. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  812. return FAILED;
  813. }
  814. ge::GeShape dst_shape(dst_shape_dims);
  815. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  816. net_output_input_desc->SetShape(dst_shape);
  817. if (is_dynamic) {
  818. auto merge_out = src_op_desc->MutableOutputDesc(0);
  819. GE_CHECK_NOTNULL(merge_out);
  820. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  821. REPORT_CALL_ERROR("E19999", "Modify format and shape of output:0 in op:%s(%s) failed",
  822. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  823. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  824. return FAILED;
  825. }
  826. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  827. auto merge_in = src_op_desc->MutableInputDesc(i);
  828. GE_CHECK_NOTNULL(merge_in);
  829. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  830. REPORT_CALL_ERROR("E19999", "Modify format and shape of input:%u in op:%s(%s) failed", i,
  831. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str());
  832. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  833. return FAILED;
  834. }
  835. }
  836. }
  837. return SUCCESS;
  838. }
  839. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  840. GE_CHECK_NOTNULL(netout_desc);
  841. vector<string> output_dt_str;
  842. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  843. for (auto dt_str : output_dt_str) {
  844. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  845. if (dt_str_split.size() == kUserDefinedElementCount) {
  846. if (dt_str_split[0] == to_string(index)) {
  847. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  848. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  849. TypeUtils::DataTypeToSerialString(dt).c_str());
  850. return true;
  851. }
  852. }
  853. }
  854. }
  855. return false;
  856. }
  857. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  858. GE_CHECK_NOTNULL(netout_desc);
  859. vector<string> output_format_str;
  860. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  861. for (auto format_str : output_format_str) {
  862. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  863. if (format_str_split.size() == kUserDefinedElementCount) {
  864. if (format_str_split[0] == to_string(index)) {
  865. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  866. return true;
  867. }
  868. }
  869. }
  870. }
  871. return false;
  872. }
  873. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  874. auto op_desc = node->GetOpDesc();
  875. GE_CHECK_NOTNULL(op_desc);
  876. ge::DataType output_data_type = ge::DT_FLOAT;
  877. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  878. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  879. auto peer_out = in_anchor->GetPeerOutAnchor();
  880. GE_CHECK_NOTNULL(peer_out);
  881. auto src_node = peer_out->GetOwnerNode();
  882. GE_CHECK_NOTNULL(src_node);
  883. bool is_dynamic = CheckOpType(src_node, MERGE);
  884. OpDescPtr src_op_desc = src_node->GetOpDesc();
  885. GE_CHECK_NOTNULL(src_op_desc);
  886. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  887. GE_CHECK_NOTNULL(net_output_input_desc);
  888. ge::GeShape old_shape = net_output_input_desc->GetShape();
  889. ge::Format old_format = net_output_input_desc->GetFormat();
  890. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  891. // Update datatype
  892. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  893. GELOGI("Enter into process output_type schedule");
  894. net_output_input_desc->SetDataType(output_data_type);
  895. if (is_dynamic) {
  896. auto merge_output = src_op_desc->MutableOutputDesc(0);
  897. GE_CHECK_NOTNULL(merge_output);
  898. merge_output->SetDataType(output_data_type);
  899. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  900. auto merge_input = src_op_desc->MutableInputDesc(i);
  901. GE_CHECK_NOTNULL(merge_input);
  902. merge_input->SetDataType(output_data_type);
  903. }
  904. }
  905. }
  906. // check if is_output_adjust_hw_layout is set
  907. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  908. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  909. ErrorManager::GetInstance().ATCReportErrMessage(
  910. "E19014", {"opname", "value", "reason"},
  911. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  912. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  913. GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0.");
  914. return FAILED;
  915. }
  916. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  917. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  918. GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0.");
  919. return FAILED;
  920. }
  921. }
  922. }
  923. return SUCCESS;
  924. }
  925. long StringToLongNoThrow(const string &str) {
  926. try {
  927. return std::stol(str);
  928. } catch (const std::invalid_argument) {
  929. GELOGE(PARAM_INVALID,
  930. "Parse shape range of input failed when transfer from string to int64. Given %s, while correct example:"
  931. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  932. str.c_str());
  933. return PARAM_INVALID;
  934. } catch (const std::out_of_range) {
  935. GELOGE(PARAM_INVALID,
  936. "Parse shape range of input failed when transfer from string to int64. Given %s, while correct example:"
  937. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  938. str.c_str());
  939. return PARAM_INVALID;
  940. }
  941. }
  942. /**
  943. * Parser shape_range from string to vector
  944. * shape_range from option normally is "[1~20,3,3~6,-1],[1~20,3,3~6,-1]"
  945. * @param shape_range
  946. */
  947. Status ParseDynamicInputShapeRange(const std::string &shape_range,
  948. std::vector<std::vector<std::pair<int64_t, int64_t>>> &range) {
  949. if (shape_range.size() < 2) {
  950. REPORT_INNER_ERROR("E19999", "shape_range.size:%zu < 2, check invalid", shape_range.size());
  951. GELOGE(PARAM_INVALID, "Shape range %s is invalid.", shape_range.c_str());
  952. return PARAM_INVALID;
  953. }
  954. // different shape_range of single input are split by ']'
  955. vector<string> shape_range_set = ge::StringUtils::Split(shape_range, ']');
  956. if (shape_range_set.empty()) {
  957. REPORT_INNER_ERROR("E19999", "Shape range %s is not valid. Correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  958. shape_range.c_str());
  959. GELOGE(PARAM_INVALID, "Shape range %s is not valid. Correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  960. shape_range.c_str());
  961. return PARAM_INVALID;
  962. }
  963. for (auto &shape_range_str : shape_range_set) {
  964. if (shape_range_str.size() < 3) {
  965. // shape_range_str should be "[2~3,1"
  966. // or ",[2~3,1". because we should trim '[' or ',['
  967. // so shape_range_str.size() < 3 is invalid
  968. continue;
  969. }
  970. // trim start bytes, after that, single input should be "1~20,3,3~6,-1"
  971. if (ge::StringUtils::StartWith(shape_range_str, "[")) {
  972. shape_range_str = shape_range_str.substr(1, shape_range_str.size());
  973. }
  974. if (ge::StringUtils::StartWith(shape_range_str, ",")) {
  975. shape_range_str = shape_range_str.substr(2, shape_range_str.size());
  976. }
  977. // parse shape_range of single input. eg. "1~20,3,3~6,-1"
  978. std::vector<std::pair<int64_t, int64_t>> range_of_single_input;
  979. vector<string> dim_range_set = ge::StringUtils::Split(shape_range_str, ',');
  980. for (const auto &range_pair_str : dim_range_set) {
  981. vector<string> range_pair_set = ge::StringUtils::Split(range_pair_str, '~');
  982. pair<int64_t, int64_t> range_pair;
  983. if (range_pair_set.size() == 1) {
  984. // fix dim
  985. auto range_value = StringToLongNoThrow(range_pair_set.at(0).c_str());
  986. if (range_value < 0) {
  987. range_pair = std::make_pair(1, range_value);
  988. } else {
  989. range_pair = std::make_pair(range_value, range_value);
  990. }
  991. } else if (range_pair_set.size() == 2) {
  992. // unknown dim, should get range.
  993. auto range_left = StringToLongNoThrow(range_pair_set.at(0).c_str());
  994. auto range_right = StringToLongNoThrow(range_pair_set.at(1).c_str());
  995. if (range_left < 0 || range_right < 0) {
  996. REPORT_INNER_ERROR("E19999", "Shape range of input is invalid. Given range pair [%ld,%ld], "
  997. "while correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"", range_left, range_right);
  998. GELOGE(PARAM_INVALID,
  999. "Shape range of input is invalid. Given range pair [%ld,%ld], while correct example: "
  1000. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  1001. range_left, range_right);
  1002. return PARAM_INVALID;
  1003. }
  1004. range_pair = std::make_pair(range_left, range_right);
  1005. } else {
  1006. REPORT_INNER_ERROR("E19999", "Shape range of input is invalid. Given %s, "
  1007. "while correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"", shape_range.c_str());
  1008. GELOGE(PARAM_INVALID,
  1009. "Shape range of input is invalid. Given %s, while correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  1010. shape_range.c_str());
  1011. return PARAM_INVALID;
  1012. }
  1013. range_of_single_input.emplace_back(range_pair);
  1014. }
  1015. range.emplace_back(range_of_single_input);
  1016. }
  1017. return SUCCESS;
  1018. }
  1019. Status GetDynamicInputShapeRange(const std::vector<GeTensor> &user_input, const std::map<string, string> &graph_option,
  1020. vector<vector<std::pair<int64_t, int64_t>>> &range_vec) {
  1021. // check both mode and shape_range option are all enabled
  1022. auto mode_iter = graph_option.find(OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  1023. bool enable_dynamic_execute_mode = (mode_iter != graph_option.end()) && (mode_iter->second == "dynamic_execute");
  1024. if (!enable_dynamic_execute_mode) {
  1025. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  1026. OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  1027. }
  1028. auto iter = graph_option.find(OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1029. bool enable_input_shape_range = (iter != graph_option.end()) && (!iter->second.empty());
  1030. if (!enable_input_shape_range) {
  1031. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  1032. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1033. }
  1034. if (enable_dynamic_execute_mode && enable_input_shape_range) {
  1035. GELOGD("GraphOption: %s value is dynamic_execute, %s value is %s.", OPTION_EXEC_DYNAMIC_EXECUTE_MODE,
  1036. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE, iter->second.c_str());
  1037. } else if (!enable_dynamic_execute_mode && !enable_input_shape_range) {
  1038. return SUCCESS;
  1039. } else {
  1040. REPORT_INNER_ERROR("E19999", "Graph option: %s and %s should be enabled at the same time, check invalid",
  1041. OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1042. GELOGE(PARAM_INVALID, "Graph option: %s and %s should be enabled at the same time.",
  1043. OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1044. return PARAM_INVALID;
  1045. }
  1046. auto ret = ParseDynamicInputShapeRange(iter->second, range_vec);
  1047. GE_CHK_STATUS_RET(ret, "Parse dynamic input shape range failed.");
  1048. if (range_vec.size() != user_input.size()) {
  1049. GELOGE(PARAM_INVALID, "Dynamic input shape range size is %zu, inputs size is %zu. Not match.", range_vec.size(),
  1050. user_input.size());
  1051. return PARAM_INVALID;
  1052. }
  1053. return SUCCESS;
  1054. }
  1055. Status UpdateDynamicInputShapeRange(const ge::GeAttrValue::INT index,
  1056. const vector<vector<std::pair<int64_t, int64_t>>> &range_vec, OpDescPtr &op,
  1057. GeTensorDesc &desc) {
  1058. auto origin_shape = desc.GetShape();
  1059. auto current_shape_range_vec = range_vec.at(index);
  1060. if (current_shape_range_vec.size() != origin_shape.GetDimNum()) {
  1061. REPORT_INNER_ERROR("E19999", "Given shape_range dim num is %zu, current dim:%s num is %zu, not match, "
  1062. "check invalid", current_shape_range_vec.size(), origin_shape.ToString().c_str(),
  1063. origin_shape.GetDimNum());
  1064. GELOGE(PARAM_INVALID, "Given shape_range dim num is %zu, current dim num is %zu, not match.Pleace Check.",
  1065. current_shape_range_vec.size(), origin_shape.GetDimNum());
  1066. return PARAM_INVALID;
  1067. }
  1068. for (size_t i = 0; i < origin_shape.GetDimNum(); ++i) {
  1069. auto curr_dim = origin_shape.GetDim(i);
  1070. auto left_range = current_shape_range_vec.at(i).first;
  1071. auto right_range = current_shape_range_vec.at(i).second;
  1072. if (left_range == right_range) {
  1073. // given shape_range is known dim, check is same as origin or not
  1074. if (curr_dim != left_range) {
  1075. REPORT_INNER_ERROR("E19999", "Given shape range is %ld, current dim shape is %ld, not match, dim_index:%zu, "
  1076. "check invalid", left_range, curr_dim, i);
  1077. GELOGE(PARAM_INVALID, "Given shape range is %ld, current dim shape is %ld, not match.Pleace Check.",
  1078. left_range, curr_dim);
  1079. return PARAM_INVALID;
  1080. }
  1081. origin_shape.SetDim(i, left_range);
  1082. } else {
  1083. // given shape_range is fix range, check input_shape is in this range or not
  1084. if (right_range != UNKNOWN_DIM) {
  1085. if ((curr_dim < left_range) || (curr_dim > right_range)) {
  1086. REPORT_INNER_ERROR("E19999", "Given shape range is [%ld~%ld], current dim shape is %ld, out of range, "
  1087. "dim_index:%zu, check invalid",
  1088. left_range, right_range, curr_dim, i);
  1089. GELOGE(PARAM_INVALID, "Given shape range is [%ld~%ld], current dim shape is %ld, out of range.Pleace Check.",
  1090. left_range, right_range, curr_dim);
  1091. return PARAM_INVALID;
  1092. }
  1093. }
  1094. origin_shape.SetDim(i, UNKNOWN_DIM);
  1095. }
  1096. }
  1097. desc.SetShape(origin_shape);
  1098. desc.SetShapeRange(current_shape_range_vec);
  1099. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1100. GE_CHK_GRAPH_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  1101. graph_ret = op->UpdateOutputDesc(0, desc);
  1102. GE_CHK_GRAPH_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  1103. return SUCCESS;
  1104. }
  1105. } // namespace
  1106. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  1107. GraphPrepare::~GraphPrepare() {}
  1108. /**
  1109. * @param graph
  1110. * @return
  1111. */
  1112. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  1113. GE_CHECK_NOTNULL(graph);
  1114. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  1115. for (auto &node : graph->GetAllNodes()) {
  1116. if (node == nullptr) {
  1117. continue;
  1118. }
  1119. if (node->GetType() != VARIABLE) {
  1120. continue;
  1121. }
  1122. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  1123. if (trans_road == nullptr) {
  1124. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  1125. continue;
  1126. }
  1127. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  1128. auto ret = RecoverTransRoadForVar(node, *trans_road);
  1129. if (ret != SUCCESS) {
  1130. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str());
  1131. return INTERNAL_ERROR;
  1132. }
  1133. auto iter = var_names_to_refs.find(node->GetName());
  1134. if (iter != var_names_to_refs.end()) {
  1135. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  1136. if (ret != SUCCESS) {
  1137. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str());
  1138. return INTERNAL_ERROR;
  1139. }
  1140. }
  1141. }
  1142. return SUCCESS;
  1143. }
  1144. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  1145. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  1146. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  1147. if (compute_graph_ != nullptr) {
  1148. compute_graph_->SetSessionID(session_id);
  1149. }
  1150. session_id_ = session_id;
  1151. Status ret = CheckGraph();
  1152. if (ret != SUCCESS) {
  1153. GELOGE(ret, "RunGraph graph check fail, ret:%u", ret);
  1154. return ret;
  1155. }
  1156. (void)compute_graph_->TopologicalSorting();
  1157. ret = CheckRefOp();
  1158. if (ret != SUCCESS) {
  1159. GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret);
  1160. return ret;
  1161. }
  1162. return SUCCESS;
  1163. }
  1164. Status GraphPrepare::CheckGraph() {
  1165. if (compute_graph_ == nullptr) {
  1166. REPORT_INNER_ERROR("E19999", "compute_graph_ is nullptr, check invalid");
  1167. GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR");
  1168. return GE_GRAPH_INIT_FAILED;
  1169. }
  1170. auto nodes = compute_graph_->GetAllNodes();
  1171. if (nodes.empty()) {
  1172. REPORT_INNER_ERROR("E19999", "nodes in graph is empty, check invalid");
  1173. GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph.");
  1174. return GE_GRAPH_INIT_FAILED;
  1175. }
  1176. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  1177. GE_CHECK_NOTNULL(node);
  1178. if (node->GetOpDesc() == nullptr) {
  1179. REPORT_INNER_ERROR("E19999", "node without opdesc exist in graph, check invalid");
  1180. GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL");
  1181. return GE_GRAPH_INIT_FAILED;
  1182. }
  1183. }
  1184. return SUCCESS;
  1185. }
  1186. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  1187. const std::set<NodePtr> &ref_nodes) {
  1188. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  1189. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  1190. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  1191. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  1192. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  1193. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH};
  1194. GE_CHECK_NOTNULL(node);
  1195. const auto &op_desc = node->GetOpDesc();
  1196. GE_CHECK_NOTNULL(op_desc);
  1197. const auto input_index = op_desc->GetInputIndexByName(input_name);
  1198. const auto &in_anchor = node->GetInDataAnchor(input_index);
  1199. GE_CHECK_NOTNULL(in_anchor);
  1200. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  1201. GE_CHECK_NOTNULL(peer_out_anchor);
  1202. const auto &input_node = peer_out_anchor->GetOwnerNode();
  1203. GE_CHECK_NOTNULL(input_node);
  1204. const auto &input_op_desc = input_node->GetOpDesc();
  1205. GE_CHECK_NOTNULL(input_op_desc);
  1206. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  1207. if (is_ref) {
  1208. return SUCCESS;
  1209. }
  1210. auto input_type = input_op_desc->GetType();
  1211. if (input_type == ge::FRAMEWORKOP) {
  1212. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  1213. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed",
  1214. ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE.c_str(),
  1215. input_op_desc->GetName().c_str(), input_op_desc->GetType().c_str());
  1216. GELOGE(PARAM_INVALID, "Get original type failed.");
  1217. return PARAM_INVALID;
  1218. }
  1219. }
  1220. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  1221. if (!is_acceptable) {
  1222. ErrorManager::GetInstance().ATCReportErrMessage(
  1223. "E15005", {"opname", "optype", "opname1", "optype1"},
  1224. {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()});
  1225. GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.",
  1226. node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  1227. input_op_desc->GetType().c_str());
  1228. return PARAM_INVALID;
  1229. }
  1230. return SUCCESS;
  1231. }
  1232. Status GraphPrepare::CheckRefOp() {
  1233. GE_CHECK_NOTNULL(compute_graph_);
  1234. std::set<NodePtr> ref_nodes;
  1235. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  1236. if (node == nullptr) {
  1237. REPORT_INNER_ERROR("E19999", "nullptr node exist in graph, check invalid");
  1238. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  1239. return PARAM_INVALID;
  1240. }
  1241. auto op_desc = node->GetOpDesc();
  1242. if (op_desc == nullptr) {
  1243. REPORT_INNER_ERROR("E19999", "node without opdesc exist in graph, check invalid");
  1244. GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null.");
  1245. return PARAM_INVALID;
  1246. }
  1247. auto input_name_index = op_desc->GetAllInputName();
  1248. auto outputs = op_desc->GetAllOutputName();
  1249. for (const auto &name_index : input_name_index) {
  1250. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  1251. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  1252. GELOGE(PARAM_INVALID, "CheckRefInputNode failed.");
  1253. return PARAM_INVALID;
  1254. }
  1255. (void)ref_nodes.insert(node); // no need to check value
  1256. }
  1257. }
  1258. }
  1259. return SUCCESS;
  1260. };
  1261. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  1262. GE_CHECK_NOTNULL(compute_graph_);
  1263. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  1264. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  1265. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  1266. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  1267. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  1268. return SUCCESS;
  1269. }
  1270. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1271. if (node == nullptr) {
  1272. REPORT_INNER_ERROR("E19999", "Param node is nullptr, check invalid");
  1273. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL");
  1274. return GE_GRAPH_GRAPH_NODE_NULL;
  1275. }
  1276. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1277. if (op_desc_ptr == nullptr) {
  1278. REPORT_INNER_ERROR("E19999", "Param node's op_desc is nullptr, check invalid");
  1279. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL");
  1280. return GE_GRAPH_GRAPH_NODE_NULL;
  1281. }
  1282. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1283. GeShape output_shape = output.GetShape();
  1284. if (output_shape.IsUnknownShape()) {
  1285. GELOGD("[Adjust][DataOpOutput] Shape of op [%s] output is unknown.", node->GetName().c_str());
  1286. return SUCCESS;
  1287. }
  1288. int64_t tensor_size = 0;
  1289. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1290. if (graph_status != GRAPH_SUCCESS) {
  1291. ErrorManager::GetInstance().ATCReportErrMessage(
  1292. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1293. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  1294. return FAILED;
  1295. }
  1296. TensorUtils::SetSize(output, tensor_size);
  1297. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1298. if (graph_ret != GRAPH_SUCCESS) {
  1299. REPORT_CALL_ERROR("E19999", "Update output desc of op:%s(%s) failed, index:0",
  1300. op_desc_ptr->GetName().c_str(), op_desc_ptr->GetType().c_str());
  1301. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1302. return graph_ret;
  1303. }
  1304. return SUCCESS;
  1305. }
  1306. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input,
  1307. const std::map<string, string> &graph_option) {
  1308. // Get shape range of input in dynamic_execute mode
  1309. vector<vector<std::pair<int64_t, int64_t>>> dynamic_shape_range_vec;
  1310. auto ret = GetDynamicInputShapeRange(user_input, graph_option, dynamic_shape_range_vec);
  1311. GE_CHK_STATUS_RET(ret, "Graph option is not right on Dynamic execute mode.");
  1312. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1313. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1314. GE_CHECK_NOTNULL(input_node);
  1315. OpDescPtr op = input_node->GetOpDesc();
  1316. GE_CHECK_NOTNULL(op);
  1317. if (op->GetType() == DATA) {
  1318. GeAttrValue::INT index = 0;
  1319. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1320. GELOGW("Get index from data attr failed");
  1321. continue;
  1322. }
  1323. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1324. std::string situation = "data op index[" + std::to_string(index) + "]";
  1325. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1326. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1327. GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index);
  1328. return FAILED;
  1329. }
  1330. if (IsDynamicDims(input_node)) {
  1331. continue;
  1332. }
  1333. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1334. auto format = desc.GetFormat();
  1335. auto origin_format = desc.GetOriginFormat();
  1336. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1337. auto tune_flag = (options_.build_mode == BUILD_MODE_TUNING) && (options_.build_step == BUILD_STEP_AFTER_BUILDER);
  1338. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op) && (!tune_flag);
  1339. if (need_check_internal_format) {
  1340. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1341. if (is_internal) {
  1342. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1343. {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" +
  1344. TypeUtils::FormatToSerialString(origin_format) + "]", "it is not support"});
  1345. GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.",
  1346. TypeUtils::FormatToSerialString(format).c_str(),
  1347. TypeUtils::FormatToSerialString(origin_format).c_str());
  1348. return FAILED;
  1349. }
  1350. }
  1351. auto data_type = desc.GetDataType();
  1352. uint32_t length = 1;
  1353. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1354. if (!type_ret) {
  1355. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1356. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1357. GELOGE(PARAM_INVALID, "Input datatype %s is not support.",
  1358. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1359. return FAILED;
  1360. }
  1361. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1362. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1363. int64_t shape_size = desc_shape * length;
  1364. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1365. int64_t size = 0;
  1366. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1367. REPORT_CALL_ERROR("E19999", "Get size of user input tensor failed, index:%ld", index);
  1368. GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed");
  1369. return FAILED);
  1370. bool size_check = (size != 0 && shape_size != size);
  1371. if (size_check) {
  1372. std::string situation = "input data size[" + std::to_string(size) +
  1373. "] and shape_size[" + std::to_string(size) + "]";
  1374. std::string reason = "because size != 0 and shape_size != size";
  1375. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1376. GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size);
  1377. return FAILED;
  1378. }
  1379. ge::TensorUtils::SetSize(desc, shape_size);
  1380. if (!tune_flag) {
  1381. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1382. if (graph_ret != GRAPH_SUCCESS) {
  1383. REPORT_CALL_ERROR("E19999", "Update input desc of op:%s(%s) failed, index:0",
  1384. op->GetName().c_str(), op->GetType().c_str());
  1385. GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
  1386. return graph_ret;
  1387. }
  1388. // Size will be recalculated in the build stage
  1389. ge::TensorUtils::SetSize(desc, 0);
  1390. graph_ret = op->UpdateOutputDesc(0, desc);
  1391. if (graph_ret != GRAPH_SUCCESS) {
  1392. REPORT_CALL_ERROR("E19999", "Update output desc of op:%s(%s) failed, index:0",
  1393. op->GetName().c_str(), op->GetType().c_str());
  1394. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1395. return graph_ret;
  1396. }
  1397. } else {
  1398. GELOGI("data %s skip update info in tune mode", op->GetName().c_str());
  1399. }
  1400. if (!dynamic_shape_range_vec.empty()) {
  1401. ret = UpdateDynamicInputShapeRange(index, dynamic_shape_range_vec, op, desc);
  1402. GE_CHK_STATUS_RET(ret, "Fail to update dynamic input shape range on %s.", op->GetName().c_str());
  1403. continue;
  1404. }
  1405. if (!options_.train_graph_flag) {
  1406. Status ret = AdjustDataOpOutput(input_node);
  1407. GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret); return ret);
  1408. }
  1409. }
  1410. }
  1411. return SUCCESS;
  1412. }
  1413. Status GraphPrepare::TryDoAipp() {
  1414. // infer and with aipp configure file, then call aipp insert
  1415. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1416. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1417. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1418. if (ret != SUCCESS) {
  1419. GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed.");
  1420. return INTERNAL_ERROR;
  1421. }
  1422. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1423. if (ret != SUCCESS) {
  1424. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed",
  1425. options_.insert_op_file.c_str());
  1426. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1427. }
  1428. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1429. if (ret != SUCCESS) {
  1430. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret);
  1431. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1432. }
  1433. }
  1434. return SUCCESS;
  1435. }
  1436. Status GraphPrepare::FormatAndShapeProcess() {
  1437. Status ret = ResourcePairProcess("add");
  1438. if (ret != SUCCESS) {
  1439. GELOGE(ret, "ResourcePairProcess failed");
  1440. return ret;
  1441. }
  1442. GE_TIMESTAMP_START(InferOriginFormat1);
  1443. ret = compute_graph_->InferOriginFormat();
  1444. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1445. GE_DUMP(compute_graph_, "after_first_inferformat");
  1446. if (ret != SUCCESS) {
  1447. GELOGE(ret, "Prepare Graph first inferformat failed");
  1448. return ret;
  1449. }
  1450. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1451. ret = InferShapeForPreprocess();
  1452. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1453. GE_DUMP(compute_graph_, "after_infershape");
  1454. if (ret != SUCCESS) {
  1455. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed");
  1456. return GE_GRAPH_INFERSHAPE_FAILED;
  1457. }
  1458. GE_TIMESTAMP_START(InferOriginFormat2);
  1459. ret = compute_graph_->InferOriginFormat();
  1460. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1461. if (ret != SUCCESS) {
  1462. GELOGE(ret, "Prepare Graph inferformat failed");
  1463. return ret;
  1464. }
  1465. ret = ResourcePairProcess("remove");
  1466. if (ret != SUCCESS) {
  1467. return ret;
  1468. }
  1469. return ret;
  1470. }
  1471. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1472. PassManager control_pass;
  1473. // Graph pass tmp logic for resource infershape
  1474. if (options_.train_graph_flag) {
  1475. try {
  1476. if (action == "add") {
  1477. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1478. } else {
  1479. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1480. new ResourcePairRemoveControlPass);
  1481. }
  1482. } catch (std::bad_alloc &e) {
  1483. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add ResourcePair Pass");
  1484. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str());
  1485. return INTERNAL_ERROR;
  1486. }
  1487. }
  1488. Status ret = control_pass.Run(compute_graph_);
  1489. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1490. GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret);
  1491. return ret;
  1492. }
  1493. return SUCCESS;
  1494. }
  1495. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1496. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1497. GE_CHECK_NOTNULL(node_ptr);
  1498. if (node_ptr->GetType() == DATA) {
  1499. uint32_t index = 0;
  1500. auto op_desc = node_ptr->GetOpDesc();
  1501. GE_CHECK_NOTNULL(op_desc);
  1502. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1503. Format storage_format = FORMAT_RESERVED;
  1504. vector<int64_t> dst_shape_dims;
  1505. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1506. GELOGE(INTERNAL_ERROR, "Get storage format for input failed");
  1507. return FAILED;
  1508. }
  1509. if (storage_format == FORMAT_RESERVED) {
  1510. continue;
  1511. }
  1512. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1513. GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed");
  1514. return FAILED;
  1515. }
  1516. }
  1517. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1518. auto op_desc = node_ptr->GetOpDesc();
  1519. GE_CHECK_NOTNULL(op_desc);
  1520. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1521. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1522. Format storage_format = FORMAT_RESERVED;
  1523. vector<int64_t> dst_shape_dims;
  1524. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1525. GELOGE(INTERNAL_ERROR, "Get storage format from output failed");
  1526. return FAILED;
  1527. }
  1528. if (storage_format == FORMAT_RESERVED) {
  1529. continue;
  1530. }
  1531. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1532. GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed");
  1533. return FAILED;
  1534. }
  1535. }
  1536. }
  1537. }
  1538. return SUCCESS;
  1539. }
  1540. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1541. if (options_.save_original_model == "true") {
  1542. ModelHelper model_helper;
  1543. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1544. options_.original_model_file);
  1545. if (ret != SUCCESS) {
  1546. // If save original model fail, process continue
  1547. GELOGW("SaveOriginalGraphToOmModel fail");
  1548. }
  1549. }
  1550. return SUCCESS;
  1551. }
  1552. #define PP_RUN_AND_DUMP(name, func, ...) \
  1553. do { \
  1554. GE_RUN(Prepare, func, __VA_ARGS__); \
  1555. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1556. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1557. } while (0)
  1558. #define PP_RUN(name, func, ...) \
  1559. do { \
  1560. GE_RUN(Prepare, func, __VA_ARGS__); \
  1561. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1562. } while (0)
  1563. Status GraphPrepare::PrepareDynShape(const GraphNodePtr &graph_node, const std::vector<GeTensor> &user_input,
  1564. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1565. GE_CHECK_NOTNULL(graph_node->GetGraph());
  1566. GE_CHECK_NOTNULL(compute_graph);
  1567. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1568. const Graph &const_graph = *graph_node->GetGraph();
  1569. PP_RUN("Init", Init, const_graph, session_id);
  1570. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1571. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input, graph_node->GetOptions());
  1572. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1573. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1574. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1575. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1576. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1577. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1578. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1579. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1580. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1581. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1582. return SUCCESS;
  1583. }
  1584. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1585. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1586. return SUCCESS;
  1587. }
  1588. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1589. auto compute_graph = compute_graph_;
  1590. PassManager pass_manager;
  1591. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1592. new (std::nothrow) VariablePrepareOpPass))
  1593. GE_TIMESTAMP_START(pass_manager);
  1594. auto ret = pass_manager.Run(compute_graph);
  1595. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1596. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1597. GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret);
  1598. return ret;
  1599. }
  1600. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1601. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1602. return SUCCESS;
  1603. }
  1604. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1605. if (compute_graph == nullptr) {
  1606. REPORT_INNER_ERROR("E19999", "Param compute_graph is nullptr, check invalid");
  1607. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1608. return GE_GRAPH_NULL_INPUT;
  1609. }
  1610. GEPass ge_passes(compute_graph);
  1611. NamesToPass hccl_group;
  1612. HcclGroupPass hccl_group_pass;
  1613. GELOGD("Add hccl group pass success");
  1614. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1615. auto ret = ge_passes.Run(hccl_group);
  1616. if (ret != SUCCESS) {
  1617. GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret);
  1618. return ret;
  1619. }
  1620. ret = compute_graph->TopologicalSorting();
  1621. if (ret != SUCCESS) {
  1622. REPORT_CALL_ERROR("E19999", "Topological sorting failed");
  1623. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1624. return ret;
  1625. }
  1626. return SUCCESS;
  1627. }
  1628. #undef PP_RUN_AND_DUMP
  1629. #undef PP_RUN
  1630. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1631. if (graph == nullptr) {
  1632. REPORT_INNER_ERROR("E19999", "Param graph is nullptr, check invalid");
  1633. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1634. return GE_GRAPH_NULL_INPUT;
  1635. }
  1636. const Graph &const_graph = *graph;
  1637. Status ret = Init(const_graph, 0);
  1638. if (ret != SUCCESS) {
  1639. GELOGE(ret, "Init graph_prepare fail, ret:%u", ret);
  1640. return ret;
  1641. }
  1642. GE_DUMP(compute_graph_, "after_parser");
  1643. GELOGI("Start infershape for dump json process.");
  1644. ret = compute_graph_->InferOriginFormat();
  1645. GE_DUMP(compute_graph_, "after_inferformat");
  1646. if (ret != SUCCESS) {
  1647. REPORT_CALL_ERROR("E19999", "Infer OriginFormat failed");
  1648. GELOGE(ret, "Prepare Graph inferformat failed");
  1649. return ret;
  1650. }
  1651. InferShapePass infer_shape_pass;
  1652. NamesToPass names_to_passes;
  1653. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1654. GEPass ge_passes(compute_graph_);
  1655. ret = ge_passes.Run(names_to_passes);
  1656. GE_DUMP(compute_graph_, "after_infershape");
  1657. if (ret != SUCCESS) {
  1658. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1659. return ret;
  1660. }
  1661. ShapeRefiner::ClearContextMap();
  1662. return SUCCESS;
  1663. }
  1664. Status GraphPrepare::CheckConstOp() {
  1665. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1666. GE_CHECK_NOTNULL(node_ptr);
  1667. if (node_ptr->GetType() == CONSTANT) {
  1668. Status ret = VerifyConstOp(node_ptr);
  1669. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1670. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1671. auto op_desc = node_ptr->GetOpDesc();
  1672. if (op_desc == nullptr) {
  1673. REPORT_INNER_ERROR("E19999", "op_desc is nullptr, check invalid");
  1674. GELOGE(PARAM_INVALID, "Get op desc failed");
  1675. return PARAM_INVALID;
  1676. }
  1677. std::string original_type;
  1678. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1679. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1680. GELOGI("original type is %s", original_type.c_str());
  1681. if (original_type == CONSTANT) {
  1682. Status ret = VerifyConstOp(node_ptr);
  1683. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1684. }
  1685. }
  1686. }
  1687. return SUCCESS;
  1688. }
  1689. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1690. GE_CHECK_NOTNULL(node);
  1691. auto op_desc = node->GetOpDesc();
  1692. GE_CHECK_NOTNULL(op_desc);
  1693. ConstGeTensorPtr ge_tensor_ptr;
  1694. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1695. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1696. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  1697. GELOGE(PARAM_INVALID, "Get value from const attr failed");
  1698. return PARAM_INVALID;
  1699. }
  1700. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1701. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1702. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1703. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1704. auto data_type = ge_tensor_desc.GetDataType();
  1705. if (data_type == DT_STRING) {
  1706. return SUCCESS;
  1707. }
  1708. uint32_t length = 1;
  1709. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1710. if (!type_ret) {
  1711. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1712. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1713. GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str());
  1714. return FAILED;
  1715. }
  1716. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1717. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1718. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1719. if (shape_size == 0) {
  1720. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1721. // shape = [], means it's a sclar tensor.
  1722. GE_CHK_BOOL_EXEC(data_size / length == 1,
  1723. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid scalar tensor."});
  1724. return PARAM_INVALID, "Const is invalid scalar tensor.");
  1725. } else {
  1726. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1727. GE_CHK_BOOL_EXEC(data_size == 0,
  1728. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid vector scalar."});
  1729. return PARAM_INVALID, "Const is invalid vector scalar.");
  1730. }
  1731. } else {
  1732. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0,
  1733. ErrorManager::GetInstance().ATCReportErrMessage(
  1734. "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"});
  1735. return PARAM_INVALID, "Const input data size is not equal with tensor desc shape");
  1736. }
  1737. return SUCCESS;
  1738. }
  1739. bool GraphPrepare::IsDynamicDims(const NodePtr &input_node) {
  1740. auto data_shape = NodeUtils::GetOutputDesc(*input_node, kDataOutIndex).GetShape();
  1741. const auto &dims = data_shape.GetDims();
  1742. bool all_is_positive = false;
  1743. if (std::all_of(dims.begin(), dims.end(), [](int64_t val) { return val >= 0; })) {
  1744. all_is_positive = true;
  1745. }
  1746. if (!all_is_positive && !options_.input_shape.empty() && !options_.dynamic_dims.empty() &&
  1747. options_.dynamic_node_type != kInvalidDynaimcDimsType) {
  1748. GELOGI("No need to check and update desc info, the dims of %s is %s.", input_node->GetName().c_str(),
  1749. formats::JoinToString(dims).c_str());
  1750. return true;
  1751. }
  1752. return false;
  1753. }
  1754. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1755. if (GetLocalOmgContext().is_dynamic_input) {
  1756. return SUCCESS;
  1757. }
  1758. unsigned int node_num = 0;
  1759. unsigned int data_num = 0;
  1760. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1761. GE_CHECK_NOTNULL(input_node);
  1762. OpDescPtr op = input_node->GetOpDesc();
  1763. GE_CHECK_NOTNULL(op);
  1764. node_num++;
  1765. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1766. data_num++;
  1767. GeAttrValue::INT index = 0;
  1768. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1769. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed", ATTR_NAME_WEIGHTS.c_str(),
  1770. op->GetName().c_str(), op->GetType().c_str());
  1771. GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed");
  1772. return GE_GRAPH_INIT_FAILED;
  1773. }
  1774. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1775. std::string situation = "data op index[" + std::to_string(index) + "]";
  1776. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1777. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1778. GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index);
  1779. return GE_GRAPH_INIT_FAILED;
  1780. }
  1781. if (IsDynamicDims(input_node)) {
  1782. continue;
  1783. }
  1784. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1785. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1786. int64_t dim = desc.GetShape().GetDim(i);
  1787. if (dim < UNKNOWN_DIM_NUM) {
  1788. std::string situation = "data dim[" + std::to_string(i) + "][" + std::to_string(dim) + "]" ;
  1789. std::string reason = "it need >= -2";
  1790. REPORT_INPUT_ERROR(
  1791. "E19025", std::vector<std::string>({"situation", "reason"}), std::vector<std::string>({situation, reason}));
  1792. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][InputDim]data dim %zu is not supported, need >= -2, real:%ld.", i, dim);
  1793. return GE_GRAPH_INIT_FAILED;
  1794. }
  1795. }
  1796. }
  1797. }
  1798. if (node_num <= data_num) {
  1799. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1800. }
  1801. return SUCCESS;
  1802. }
  1803. Status GraphPrepare::InferShapeForPreprocess() {
  1804. GELOGI("Start infershape for preprocess.");
  1805. GEPass ge_passes(compute_graph_);
  1806. NamesToPass names_to_passes;
  1807. AssertPass assert_pass;
  1808. if (!options_.train_graph_flag) {
  1809. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1810. }
  1811. SwitchDeadBranchElimination switch_dead_branch_elimination;
  1812. names_to_passes.emplace_back("SwitchDeadBranchElimination", &switch_dead_branch_elimination);
  1813. MergePass merge_pass;
  1814. names_to_passes.emplace_back("MergePass", &merge_pass);
  1815. InferShapePass infer_shape_pass;
  1816. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1817. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1818. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1819. DimensionComputePass dimension_compute_pass;
  1820. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1821. ConstantFoldingPass constant_folding_pass;
  1822. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1823. int32_t dev_count = 0;
  1824. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1825. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1826. rtError_t rt_err = RT_ERROR_NONE;
  1827. if (aicpu_constant_folding_on != nullptr) {
  1828. rt_err = rtGetDeviceCount(&dev_count);
  1829. if (rt_err == RT_ERROR_NONE) {
  1830. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1831. if (result != SUCCESS) {
  1832. GELOGE(result, "Set rt context failed.");
  1833. return result;
  1834. }
  1835. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1836. }
  1837. }
  1838. Status ret = ge_passes.Run(names_to_passes);
  1839. if (aicpu_constant_folding_on != nullptr) {
  1840. if (rt_err == RT_ERROR_NONE) {
  1841. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1842. if (result != SUCCESS) {
  1843. GELOGE(result, "Set rt context failed.");
  1844. return result;
  1845. }
  1846. }
  1847. }
  1848. ShapeRefiner::ClearContextMap();
  1849. if (ret != SUCCESS) {
  1850. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1851. return ret;
  1852. }
  1853. return SUCCESS;
  1854. }
  1855. Status GraphPrepare::PrepareOptimize() {
  1856. GELOGI("Start optimize for preprocess.");
  1857. // check rw type
  1858. GraphOptimize graph_optimize;
  1859. bool has_conflict = false;
  1860. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1861. if (has_conflict) {
  1862. GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize.");
  1863. return FAILED;
  1864. }
  1865. PassManager original_graph_passes;
  1866. // Graph pass
  1867. try {
  1868. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1869. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1870. (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass);
  1871. } catch (std::bad_alloc &e) {
  1872. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass");
  1873. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1874. return INTERNAL_ERROR;
  1875. }
  1876. GE_TIMESTAMP_START(original_graph_passes);
  1877. Status ret = original_graph_passes.Run(compute_graph_);
  1878. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1879. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1880. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1881. return ret;
  1882. }
  1883. // New pass
  1884. GEPass ge_passes(compute_graph_);
  1885. NamesToPass names_to_passes;
  1886. EnterPass enter_pass;
  1887. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1888. CondPass cond_pass;
  1889. names_to_passes.emplace_back("CondPass", &cond_pass);
  1890. PrintOpPass print_pass;
  1891. if (options_.enable_print_op_pass) {
  1892. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1893. }
  1894. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1895. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1896. DropOutPass dropout_pass;
  1897. AssertPass assert_pass;
  1898. UnusedConstPass unused_const_pass;
  1899. StopGradientPass stop_gradient_pass;
  1900. PreventGradientPass prevent_gradient_pass;
  1901. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1902. GuaranteeConstPass guarantee_const_pass;
  1903. VarIsInitializedOpPass var_is_initialized_pass;
  1904. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1905. IdentityPass identity_pass(false);
  1906. SnapshotPass snapshot_pass;
  1907. if (!options_.train_graph_flag) {
  1908. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1909. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1910. }
  1911. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1912. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1913. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1914. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1915. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1916. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1917. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1918. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1919. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1920. GE_TIMESTAMP_START(names_to_passes);
  1921. ret = ge_passes.Run(names_to_passes);
  1922. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1923. if (ret != SUCCESS) {
  1924. GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret);
  1925. return ret;
  1926. }
  1927. PassManager graph_pass;
  1928. try {
  1929. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1930. // can't move to optimize1/2 directly, may cause more identity insert, cause CI fail
  1931. (void)graph_pass.AddPass("PrepareOptimize::HcclMemcpyPass", new HcclMemcpyPass);
  1932. } catch (std::bad_alloc &e) {
  1933. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass");
  1934. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1935. return INTERNAL_ERROR;
  1936. }
  1937. GE_TIMESTAMP_START(graph_passes);
  1938. ret = graph_pass.Run(compute_graph_);
  1939. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1940. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1941. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1942. return ret;
  1943. }
  1944. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1945. TypeConversionOfConstant();
  1946. ret = compute_graph_->TopologicalSorting();
  1947. if (ret != SUCCESS) {
  1948. REPORT_CALL_ERROR("E19999", "Topological sorting failed");
  1949. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1950. return ret;
  1951. }
  1952. GELOGI("End optimize for preprocess.");
  1953. return SUCCESS;
  1954. }
  1955. void GraphPrepare::TypeConversionOfConstant() {
  1956. bool is_acl_compile = false;
  1957. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1958. // This can ensure that n is not a null pointer
  1959. // No Conversion when called by aclOpCompile
  1960. (void)AttrUtils::GetBool(n->GetOpDesc(), ATTR_SINGLE_OP_SCENE, is_acl_compile);
  1961. if (is_acl_compile) {
  1962. return;
  1963. }
  1964. }
  1965. if (options_.train_graph_flag) {
  1966. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1967. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1968. // This can ensure that n is not a null pointer
  1969. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1970. n->GetOpDesc()->SetType(CONSTANTOP);
  1971. }
  1972. }
  1973. } else {
  1974. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  1975. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1976. // This can ensure that n is not a null pointer
  1977. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  1978. n->GetOpDesc()->SetType(CONSTANT);
  1979. }
  1980. }
  1981. }
  1982. }
  1983. Status GraphPrepare::GraphEquivalentTransformation() {
  1984. NamesToPass names_to_pass;
  1985. ForPass for_pass;
  1986. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  1987. return GEPass(compute_graph_).Run(names_to_pass);
  1988. }
  1989. Status GraphPrepare::ProcessBeforeInfershape() {
  1990. NamesToPass names_to_passes;
  1991. CondRemovePass condition_remove_pass;
  1992. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  1993. GE_TIMESTAMP_START(ProcessCondRemove);
  1994. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  1995. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  1996. if (ret != SUCCESS) {
  1997. GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  1998. return ret;
  1999. }
  2000. return SUCCESS;
  2001. }
  2002. Status GraphPrepare::ProcessNetOutput() {
  2003. PassManager graph_passes_before_infershape;
  2004. try {
  2005. if (options_.train_graph_flag) {
  2006. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  2007. }
  2008. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  2009. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  2010. new (std::nothrow) DataPass); // Add NetOutput first.
  2011. } catch (std::bad_alloc) {
  2012. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass");
  2013. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  2014. return INTERNAL_ERROR;
  2015. }
  2016. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  2017. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  2018. GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret);
  2019. return ret;
  2020. }
  2021. return SUCCESS;
  2022. }
  2023. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input,
  2024. const std::map<string, string> &graph_option) {
  2025. compute_graph_->SetInputSize(user_input.size());
  2026. if (user_input.empty()) {
  2027. return SUCCESS;
  2028. }
  2029. auto ret = CheckUserInput(user_input);
  2030. if (ret != SUCCESS) {
  2031. GELOGE(ret, "Check user input failed.");
  2032. return ret;
  2033. }
  2034. ret = UpdateInput(user_input, graph_option);
  2035. if (ret != SUCCESS) {
  2036. GELOGE(ret, "UpdateInput fail, ret:%u", ret);
  2037. return ret;
  2038. }
  2039. if (user_input.size() != 0) {
  2040. ret = CheckConstOp();
  2041. if (ret != SUCCESS) {
  2042. GELOGE(ret, "CheckConstOp fail, ret:%u", ret);
  2043. return ret;
  2044. }
  2045. } else {
  2046. ret = compute_graph_->TopologicalSorting();
  2047. if (ret != SUCCESS) {
  2048. REPORT_CALL_ERROR("E19999", "Topological sorting failed");
  2049. GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting");
  2050. return FAILED;
  2051. }
  2052. }
  2053. return SUCCESS;
  2054. }
  2055. Status GraphPrepare::UpdateInputOutputByOptions() {
  2056. auto ret = UpdateDataNetOutputByStorageFormat();
  2057. if (ret != SUCCESS) {
  2058. GELOGE(ret, "Update format acoording to storage format failed.");
  2059. return ret;
  2060. }
  2061. if (options_.train_graph_flag) {
  2062. GELOGI("This is train mode, no need to do this schedule.");
  2063. return SUCCESS;
  2064. }
  2065. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  2066. GE_CHECK_NOTNULL(node_ptr);
  2067. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  2068. GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str());
  2069. return FAILED;
  2070. }
  2071. if (node_ptr->GetType() == DATA) {
  2072. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  2073. GELOGE(INTERNAL_ERROR, "Process data node failed");
  2074. return FAILED;
  2075. }
  2076. }
  2077. if (node_ptr->GetType() == ge::NETOUTPUT) {
  2078. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  2079. GELOGE(INTERNAL_ERROR, "Process netoutput node failed");
  2080. return FAILED;
  2081. }
  2082. }
  2083. }
  2084. return SUCCESS;
  2085. }
  2086. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  2087. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  2088. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  2089. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  2090. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  2091. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  2092. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  2093. if (dst_node->GetType() == TRANSDATA) {
  2094. return true;
  2095. }
  2096. }
  2097. }
  2098. return false;
  2099. }
  2100. } // 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