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 80 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
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
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #include "graph/preprocess/graph_preprocess.h"
  17. #include <map>
  18. #include <set>
  19. #include <string>
  20. #include <utility>
  21. #include "common/formats/format_transfers/format_transfer_fractal_nz.h"
  22. #include "common/formats/format_transfers/format_transfer_fractal_z.h"
  23. #include "common/formats/format_transfers/format_transfer_nchw_nc1hwc0.h"
  24. #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h"
  25. #include "common/formats/format_transfers/format_transfer_transpose.h"
  26. #include "common/formats/utils/formats_trans_utils.h"
  27. #include "common/helper/model_helper.h"
  28. #include "common/math/math_util.h"
  29. #include "common/op/ge_op_utils.h"
  30. #include "common/util/error_manager/error_manager.h"
  31. #include "common/formats/utils/formats_trans_utils.h"
  32. #include "framework/common/debug/ge_log.h"
  33. #include "graph/common/ge_call_wrapper.h"
  34. #include "graph/common/local_context.h"
  35. #include "graph/common/transop_util.h"
  36. #include "graph/debug/ge_attr_define.h"
  37. #include "graph/ge_context.h"
  38. #include "graph/shape_refiner.h"
  39. #include "graph/manager/graph_var_manager.h"
  40. #include "graph/manager/util/rt_context_util.h"
  41. #include "graph/optimize/graph_optimize.h"
  42. #include "graph/passes/addn_pass.h"
  43. #include "graph/passes/aicpu_constant_folding_pass.h"
  44. #include "graph/passes/assert_pass.h"
  45. #include "graph/passes/assign_pass.h"
  46. #include "graph/passes/base_pass.h"
  47. #include "graph/passes/common_subexpression_elimination_pass.h"
  48. #include "graph/passes/cond_pass.h"
  49. #include "graph/passes/cond_remove_pass.h"
  50. #include "graph/passes/constant_folding_pass.h"
  51. #include "graph/passes/constant_fuse_same_pass.h"
  52. #include "graph/passes/control_trigger_pass.h"
  53. #include "graph/passes/dimension_adjust_pass.h"
  54. #include "graph/passes/dimension_compute_pass.h"
  55. #include "graph/passes/dropout_pass.h"
  56. #include "graph/passes/enter_pass.h"
  57. #include "graph/passes/flow_ctrl_pass.h"
  58. #include "graph/passes/for_pass.h"
  59. #include "graph/passes/get_original_format_pass.h"
  60. #include "graph/passes/guarantee_const_pass.h"
  61. #include "graph/passes/hccl_group_pass.h"
  62. #include "graph/passes/hccl_memcpy_pass.h"
  63. #include "graph/passes/identity_pass.h"
  64. #include "graph/passes/infershape_pass.h"
  65. #include "graph/passes/iterator_op_pass.h"
  66. #include "graph/passes/merge_pass.h"
  67. #include "graph/passes/net_output_pass.h"
  68. #include "graph/passes/next_iteration_pass.h"
  69. #include "graph/passes/no_use_reshape_remove_pass.h"
  70. #include "graph/passes/parallel_concat_start_op_pass.h"
  71. #include "graph/passes/placeholder_with_default_pass.h"
  72. #include "graph/passes/prevent_gradient_pass.h"
  73. #include "graph/passes/print_op_pass.h"
  74. #include "graph/passes/prune_pass.h"
  75. #include "graph/passes/replace_transshape_pass.h"
  76. #include "graph/passes/replace_with_empty_const_pass.h"
  77. #include "graph/passes/resource_pair_add_control_pass.h"
  78. #include "graph/passes/resource_pair_remove_control_pass.h"
  79. #include "graph/passes/save_pass.h"
  80. #include "graph/passes/shape_operate_op_remove_pass.h"
  81. #include "graph/passes/snapshot_pass.h"
  82. #include "graph/passes/stop_gradient_pass.h"
  83. #include "graph/passes/subgraph_pass.h"
  84. #include "graph/passes/switch_data_edges_bypass.h"
  85. #include "graph/passes/switch_dead_branch_elimination.h"
  86. #include "graph/passes/switch_logic_remove_pass.h"
  87. #include "graph/passes/merge_to_stream_merge_pass.h"
  88. #include "graph/passes/switch_to_stream_switch_pass.h"
  89. #include "graph/passes/attach_stream_label_pass.h"
  90. #include "graph/passes/unused_const_pass.h"
  91. #include "graph/passes/unused_op_remove_pass.h"
  92. #include "graph/passes/var_is_initialized_op_pass.h"
  93. #include "graph/passes/variable_prepare_op_pass.h"
  94. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  95. #include "graph/types.h"
  96. #include "graph/utils/tensor_utils.h"
  97. #include "graph/utils/type_utils.h"
  98. #include "inc/pass_manager.h"
  99. #include "init/gelib.h"
  100. #include "multi_batch_copy_graph.h"
  101. #include "runtime/dev.h"
  102. #include "graph/passes/dimension_adjust_pass.h"
  103. #include "graph/passes/link_gen_mask_nodes_pass.h"
  104. #include "graph/passes/permute_pass.h"
  105. #include "graph/passes/reshape_remove_pass.h"
  106. #include "graph/passes/same_transdata_breadth_fusion_pass.h"
  107. #include "graph/passes/transop_breadth_fusion_pass.h"
  108. #include "graph/passes/transop_depth_fusion_pass.h"
  109. #include "graph/passes/transop_nearby_allreduce_fusion_pass.h"
  110. #include "graph/passes/cast_remove_pass.h"
  111. #include "graph/passes/data_pass.h"
  112. #include "graph/passes/transop_without_reshape_fusion_pass.h"
  113. #include "graph/passes/transpose_transdata_pass.h"
  114. #include "graph/passes/variable_op_pass.h"
  115. #include "graph/passes/variable_prepare_op_pass.h"
  116. #include "graph/passes/variable_ref_delete_op_pass.h"
  117. #include "graph/passes/mark_agnostic_pass.h"
  118. namespace ge {
  119. namespace {
  120. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  121. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  122. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  123. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  124. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  125. // the size of user defined output datatype or format string after split by ":".
  126. const size_t kUserDefinedElementCount = 2;
  127. const int kDataOutIndex = 0;
  128. const int64_t kInvalidDynaimcDimsType = -1;
  129. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  130. GeTensorPtr tensor = MakeShared<GeTensor>();
  131. if (tensor == nullptr) {
  132. GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed");
  133. return nullptr;
  134. }
  135. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  136. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  137. auto dst_ge_shape = data_tensor.GetShape();
  138. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  139. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  140. tensor->MutableTensorDesc().SetShape(GeShape());
  141. int32_t dst_shape = 1;
  142. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  143. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  144. return nullptr;
  145. }
  146. } else {
  147. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  148. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  149. if (dst_shape == nullptr) {
  150. GELOGE(INTERNAL_ERROR, "Create unique ptr failed");
  151. return nullptr;
  152. }
  153. for (int64_t i = 0; i < dim_cnt; ++i) {
  154. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  155. }
  156. GE_IF_BOOL_EXEC(
  157. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  158. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  159. return nullptr;)
  160. }
  161. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  162. return OpDescUtils::CreateConstOp(tensor);
  163. }
  164. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  165. OpDescPtr &op_desc) {
  166. // For format transfer node, the IR definition has src/dst format attrs
  167. if (node_type == TRANSDATA) {
  168. GE_IF_BOOL_EXEC(
  169. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  170. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  171. GE_IF_BOOL_EXEC(
  172. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  173. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  174. }
  175. // For TransposeD node, the IR definition has perm attrs
  176. if (node_type == TRANSPOSED) {
  177. Format src_format = input.GetFormat();
  178. Format dst_format = output.GetFormat();
  179. std::vector<int64_t> perm_arg;
  180. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  181. "Get perm by foramt failed.");
  182. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  183. "SetStr PERMUTE_ATTR_PERM failed")
  184. }
  185. // For cast node, the IR definition has src/dst attrs
  186. if (node_type == CAST) {
  187. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  188. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  189. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  190. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  191. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  192. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  193. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  194. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  195. }
  196. }
  197. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  198. const GeTensorDesc &output, NodePtr &node) {
  199. if (node == nullptr) {
  200. GELOGE(PARAM_INVALID, "node is null.");
  201. return nullptr;
  202. }
  203. auto graph = node->GetOwnerComputeGraph();
  204. if (graph == nullptr) {
  205. GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str());
  206. return nullptr;
  207. }
  208. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  209. if (index < 0) {
  210. ErrorManager::GetInstance().ATCReportErrMessage(
  211. "E19025", {"situation", "reason"},
  212. {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()});
  213. GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str());
  214. return nullptr;
  215. }
  216. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  217. if (op_desc == nullptr) {
  218. GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed");
  219. return nullptr;
  220. }
  221. // for data dump
  222. GE_IF_BOOL_EXEC(
  223. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  224. GELOGW("CreateTransNode: SetListStr failed");)
  225. // Default single input and single output
  226. auto ret = op_desc->AddInputDesc(input);
  227. if (ret != GRAPH_SUCCESS) {
  228. GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str());
  229. return nullptr;
  230. }
  231. ret = op_desc->AddOutputDesc(output);
  232. if (ret != GRAPH_SUCCESS) {
  233. GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str());
  234. return nullptr;
  235. }
  236. AddTransNodeAttr(node_type, input, output, op_desc);
  237. NodePtr shape_node = nullptr;
  238. if (node_type == RESHAPE) {
  239. auto shape_desc = CreateTensorShape(output);
  240. if (shape_desc == nullptr) {
  241. GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input",
  242. node->GetName().c_str());
  243. return nullptr;
  244. }
  245. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  246. if (ret != GRAPH_SUCCESS) {
  247. GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str());
  248. return nullptr;
  249. }
  250. shape_node = graph->AddNode(shape_desc);
  251. if (shape_node == nullptr) {
  252. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str());
  253. return nullptr;
  254. }
  255. }
  256. auto trans_node = graph->AddNode(op_desc);
  257. if (trans_node == nullptr) {
  258. GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str());
  259. return nullptr;
  260. }
  261. if (node_type == RESHAPE) {
  262. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  263. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str());
  264. return nullptr;
  265. }
  266. }
  267. return trans_node;
  268. }
  269. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  270. NodePtr &trans_node) {
  271. GE_CHECK_NOTNULL(node);
  272. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  273. if (trans_node == nullptr) {
  274. return INTERNAL_ERROR;
  275. }
  276. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  277. if (ret != GRAPH_SUCCESS) {
  278. GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s",
  279. node->GetName().c_str(), node->GetType().c_str());
  280. return INTERNAL_ERROR;
  281. }
  282. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  283. if (ret != GRAPH_SUCCESS) {
  284. GELOGE(INTERNAL_ERROR, "Failed to connect node %s to trans node %s", node->GetName().c_str(),
  285. trans_node->GetName().c_str());
  286. return INTERNAL_ERROR;
  287. }
  288. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  289. if (ret != GRAPH_SUCCESS) {
  290. GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.",
  291. node->GetName().c_str(), trans_node->GetName().c_str());
  292. return INTERNAL_ERROR;
  293. }
  294. return SUCCESS;
  295. }
  296. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  297. NodePtr &trans_node) {
  298. GE_CHECK_NOTNULL(node);
  299. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  300. if (trans_node == nullptr) {
  301. return INTERNAL_ERROR;
  302. }
  303. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  304. if (ret != GRAPH_SUCCESS) {
  305. GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s",
  306. node->GetName().c_str(), node->GetType().c_str());
  307. return INTERNAL_ERROR;
  308. }
  309. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  310. if (ret != GRAPH_SUCCESS) {
  311. GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(),
  312. node->GetName().c_str());
  313. return INTERNAL_ERROR;
  314. }
  315. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  316. if (ret != GRAPH_SUCCESS) {
  317. GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.",
  318. node->GetName().c_str(), trans_node->GetName().c_str());
  319. return INTERNAL_ERROR;
  320. }
  321. return SUCCESS;
  322. }
  323. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  324. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  325. GE_CHECK_NOTNULL(var->GetOpDesc());
  326. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  327. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  328. output_desc.SetFormat(tensor_desc.GetFormat());
  329. output_desc.SetDataType(tensor_desc.GetDataType());
  330. output_desc.SetShape(tensor_desc.GetShape());
  331. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  332. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  333. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  334. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  335. GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed");
  336. return INTERNAL_ERROR;);
  337. }
  338. if (var->GetOpDesc()->GetInputsSize() > 0) {
  339. auto desc = var->GetOpDesc()->GetInputDesc(0);
  340. desc.SetFormat(tensor_desc.GetFormat());
  341. desc.SetDataType(tensor_desc.GetDataType());
  342. desc.SetShape(tensor_desc.GetShape());
  343. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  344. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  345. desc.SetOriginShape(tensor_desc.GetOriginShape());
  346. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  347. GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed");
  348. return INTERNAL_ERROR;)
  349. }
  350. return SUCCESS;
  351. }
  352. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  353. GE_CHECK_NOTNULL(var);
  354. int index = 0;
  355. NodePtr last_node = var;
  356. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  357. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  358. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  359. if (ret != SUCCESS) {
  360. ErrorManager::GetInstance().ATCReportErrMessage(
  361. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  362. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(),
  363. index, iter->node_type.c_str());
  364. return INTERNAL_ERROR;
  365. }
  366. // set stream_label
  367. OpDescPtr var_desc = var->GetOpDesc();
  368. GE_CHECK_NOTNULL(var_desc);
  369. std::string stream_label;
  370. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  371. if (!stream_label.empty()) {
  372. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  373. }
  374. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  375. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  376. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  377. }
  378. if (road.empty()) {
  379. return SUCCESS;
  380. }
  381. return UpdateVarFormats(var, road.rbegin()->output);
  382. }
  383. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  384. for (auto &var : nodes) {
  385. GE_CHECK_NOTNULL(var);
  386. int index = 0;
  387. NodePtr last_node = var;
  388. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  389. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  390. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  391. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  392. if (ret != SUCCESS) {
  393. ErrorManager::GetInstance().ATCReportErrMessage(
  394. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  395. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s",
  396. var->GetName().c_str(), index, iter->node_type.c_str());
  397. return INTERNAL_ERROR;
  398. }
  399. // set stream_label
  400. OpDescPtr var_desc = var->GetOpDesc();
  401. GE_CHECK_NOTNULL(var_desc);
  402. std::string stream_label;
  403. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  404. if (!stream_label.empty()) {
  405. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  406. }
  407. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  408. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  409. }
  410. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  411. return INTERNAL_ERROR;
  412. }
  413. }
  414. return SUCCESS;
  415. }
  416. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  417. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  418. VarNamesToRefs names_to_refs;
  419. std::string var_name;
  420. if (graph == nullptr) {
  421. GELOGE(PARAM_INVALID, "graph is null.");
  422. return names_to_refs;
  423. }
  424. for (auto &node : graph->GetAllNodes()) {
  425. if (node->GetType() != VARIABLE) {
  426. continue;
  427. }
  428. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  429. (void)names_to_refs[var_name].insert(node);
  430. }
  431. }
  432. return names_to_refs;
  433. }
  434. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  435. std::vector<int64_t> &dst_shape) {
  436. if (src_format == FORMAT_NCHW) {
  437. formats::FormatTransferNchwNc1hwc0 transfer;
  438. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  439. GELOGE(INTERNAL_ERROR, "TransShape failed");
  440. return FAILED;
  441. }
  442. } else if (src_format == FORMAT_NHWC) {
  443. formats::FormatTransferNhwcNc1hwc0 transfer;
  444. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  445. GELOGE(INTERNAL_ERROR, "TransShape failed");
  446. return FAILED;
  447. }
  448. }
  449. return SUCCESS;
  450. }
  451. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  452. GE_CHECK_NOTNULL(node_ptr);
  453. auto op_desc = node_ptr->GetOpDesc();
  454. GE_CHECK_NOTNULL(op_desc);
  455. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  456. GE_CHECK_NOTNULL(input);
  457. ge::Format old_format = input->GetFormat();
  458. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  459. ge::DataType dt = input->GetDataType();
  460. std::vector<int64_t> dst_shape_dims;
  461. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  462. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  463. return FAILED;
  464. }
  465. input->SetFormat(FORMAT_NC1HWC0);
  466. input->SetShape(ge::GeShape(dst_shape_dims));
  467. auto output = op_desc->MutableOutputDesc(0);
  468. GE_CHECK_NOTNULL(output);
  469. output->SetFormat(FORMAT_NC1HWC0);
  470. output->SetShape(ge::GeShape(dst_shape_dims));
  471. int64_t size = 0;
  472. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  473. if (graph_status != ge::GRAPH_SUCCESS) {
  474. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  475. return FAILED;
  476. }
  477. ge::TensorUtils::SetSize(*output, size);
  478. ge::TensorUtils::SetSize(*input, size);
  479. return SUCCESS;
  480. }
  481. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  482. GE_CHECK_NOTNULL(input_output);
  483. ge::Format old_format = input_output->GetFormat();
  484. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  485. ge::DataType dt = input_output->GetDataType();
  486. std::vector<int64_t> dst_shape_dims;
  487. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  488. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  489. return FAILED;
  490. }
  491. input_output->SetFormat(FORMAT_NC1HWC0);
  492. input_output->SetShape(ge::GeShape(dst_shape_dims));
  493. return SUCCESS;
  494. }
  495. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  496. vector<int64_t> &dst_shape_dims) {
  497. GE_CHECK_NOTNULL(op_desc);
  498. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  499. GE_CHECK_NOTNULL(input);
  500. ge::Format old_format = input->GetFormat();
  501. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  502. input->SetShape(ge::GeShape(dst_shape_dims));
  503. input->SetFormat(storage_format);
  504. auto output = op_desc->MutableOutputDesc(index);
  505. GE_CHECK_NOTNULL(output);
  506. output->SetShape(ge::GeShape(dst_shape_dims));
  507. output->SetFormat(storage_format);
  508. if (!output->MutableShape().IsUnknownShape()) {
  509. int64_t size = 0;
  510. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  511. if (graph_status != ge::GRAPH_SUCCESS) {
  512. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  513. return FAILED;
  514. }
  515. ge::TensorUtils::SetSize(*input, size);
  516. ge::TensorUtils::SetSize(*output, size);
  517. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  518. "new_shape:%s, new_format:%s, new_size:%lu",
  519. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  520. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  521. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  522. }
  523. return SUCCESS;
  524. }
  525. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) {
  526. is_dynamic_batch = false;
  527. std::string related_node_name;
  528. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  529. if (related_node_name.empty()) {
  530. ErrorManager::GetInstance().ATCReportErrMessage(
  531. "E15002", {"opname", "value", "reason"}, {data_node->GetName(), "flag", "but the value is empty"});
  532. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty",
  533. data_node->GetName().c_str());
  534. return INTERNAL_ERROR;
  535. }
  536. for (const NodePtr &next_node : data_node->GetOutNodes()) {
  537. if (next_node->GetName() == related_node_name) {
  538. switchn_node = next_node;
  539. break;
  540. }
  541. }
  542. if (switchn_node == nullptr) {
  543. ErrorManager::GetInstance().ATCReportErrMessage(
  544. "E15002", {"opname", "value", "reason"},
  545. {data_node->GetName(), related_node_name, "but can not find it on the graph"});
  546. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph",
  547. data_node->GetName().c_str(), related_node_name.c_str());
  548. return INTERNAL_ERROR;
  549. }
  550. is_dynamic_batch = true;
  551. }
  552. return SUCCESS;
  553. }
  554. bool CheckOpType(const NodePtr &node, const std::string type) {
  555. if (node->GetType() == type) {
  556. return true;
  557. }
  558. return false;
  559. }
  560. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  561. auto op = node_ptr->GetOpDesc();
  562. GE_CHECK_NOTNULL(op);
  563. auto inputDescsPtr = op->GetAllInputsDescPtr();
  564. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  565. ge::Format format = ge::FORMAT_ND;
  566. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  567. // process, otherwise fe will insert transdata.
  568. for (auto &inputDescPtr : inputDescsPtr) {
  569. GE_CHECK_NOTNULL(inputDescPtr);
  570. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  571. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  572. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  573. inputDescPtr->SetFormat(format);
  574. inputDescPtr->SetOriginFormat(format);
  575. }
  576. }
  577. for (auto &outputDescPtr : outputDescsPtr) {
  578. GE_CHECK_NOTNULL(outputDescPtr);
  579. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  580. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  581. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  582. outputDescPtr->SetFormat(format);
  583. outputDescPtr->SetOriginFormat(format);
  584. }
  585. }
  586. return SUCCESS;
  587. }
  588. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  589. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  590. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  591. // data and netoutput, and no longer inserts the transnode.
  592. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) {
  593. GE_CHECK_NOTNULL(node_ptr);
  594. auto op_desc = node_ptr->GetOpDesc();
  595. GE_CHECK_NOTNULL(op_desc);
  596. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  597. GE_CHECK_NOTNULL(input);
  598. ge::DataType src_dtype = input->GetDataType();
  599. if (src_dtype == dt_set) {
  600. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  601. return SUCCESS;
  602. }
  603. input->SetDataType(dt_set);
  604. int64_t input_shape_size = 0;
  605. int64_t output_shape_size = 0;
  606. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  607. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  608. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  609. GELOGE(GRAPH_FAILED, "GetTensorSize failed!");
  610. return FAILED;
  611. }
  612. ge::TensorUtils::SetSize(*input, input_shape_size);
  613. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  614. GE_CHECK_NOTNULL(output);
  615. output->SetDataType(dt_set);
  616. ge::TensorUtils::SetSize(*output, output_shape_size);
  617. if (is_dynamic_batch) {
  618. GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str());
  619. auto switchn_op_desc = switchn_node->GetOpDesc();
  620. GE_CHECK_NOTNULL(switchn_op_desc);
  621. auto switchn_input = switchn_op_desc->MutableInputDesc(0);
  622. GE_CHECK_NOTNULL(switchn_input);
  623. switchn_input->SetDataType(dt_set);
  624. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  625. const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i);
  626. GE_CHECK_NOTNULL(switchn_output);
  627. switchn_output->SetDataType(dt_set);
  628. }
  629. }
  630. return SUCCESS;
  631. }
  632. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  633. GE_CHECK_NOTNULL(node_ptr);
  634. auto op_desc = node_ptr->GetOpDesc();
  635. GE_CHECK_NOTNULL(op_desc);
  636. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  637. GE_CHECK_NOTNULL(input);
  638. ge::Format old_format = input->GetFormat();
  639. ge::GeShape old_shape = input->GetShape();
  640. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  641. if (!support) {
  642. ErrorManager::GetInstance().ATCReportErrMessage(
  643. "E19014", {"opname", "value", "reason"},
  644. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  645. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  646. GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str());
  647. return FAILED;
  648. }
  649. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  650. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  651. return FAILED;
  652. }
  653. if (is_dynamic_batch) {
  654. auto switchn_op_desc = switchn_node->GetOpDesc();
  655. GE_CHECK_NOTNULL(switchn_op_desc);
  656. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  657. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  658. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  659. return FAILED;
  660. }
  661. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  662. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  663. GE_CHECK_NOTNULL(switchn_output);
  664. old_format = switchn_output->GetFormat();
  665. old_shape = switchn_output->GetShape();
  666. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  667. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  668. return FAILED;
  669. }
  670. }
  671. }
  672. return SUCCESS;
  673. }
  674. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  675. auto op_desc = node_ptr->GetOpDesc();
  676. GE_CHECK_NOTNULL(op_desc);
  677. string set_dt_str;
  678. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  679. return SUCCESS;
  680. }
  681. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  682. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  683. bool is_dynamic_batch = false;
  684. NodePtr switchn_node = nullptr;
  685. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) {
  686. GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed");
  687. return FAILED;
  688. }
  689. if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) {
  690. GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed");
  691. return FAILED;
  692. }
  693. // check if need to set format
  694. string set_format;
  695. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  696. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  697. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  698. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) {
  699. GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed");
  700. return FAILED;
  701. }
  702. }
  703. return SUCCESS;
  704. }
  705. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  706. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  707. GE_CHECK_NOTNULL(op_desc);
  708. GE_CHECK_NOTNULL(tensor_desc_ptr);
  709. storage_format = FORMAT_RESERVED;
  710. int64_t format = FORMAT_RESERVED;
  711. dst_shape_dims.clear();
  712. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  713. storage_format = static_cast<Format>(format);
  714. vector<int32_t> storage_shape;
  715. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  716. for (auto dim : storage_shape) {
  717. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  718. }
  719. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  720. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  721. formats::JoinToString(storage_shape).c_str());
  722. } else {
  723. ErrorManager::GetInstance().ATCReportErrMessage(
  724. "15003", {"opname", "format"},
  725. {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)});
  726. GELOGE(PARAM_INVALID, "Update node by storage format failed, storage_shape not set. "
  727. "node: [%s], storage_format [%s]",
  728. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  729. return FAILED;
  730. }
  731. ge::Format old_format = tensor_desc_ptr->GetFormat();
  732. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  733. if (old_format == storage_format && old_shape == dst_shape_dims) {
  734. GELOGI("Update node by storage format, not changed.");
  735. storage_format = FORMAT_RESERVED;
  736. return SUCCESS;
  737. }
  738. }
  739. return SUCCESS;
  740. }
  741. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  742. NodePtr &node) {
  743. bool is_dynamic = CheckOpType(node, MERGE);
  744. auto src_op_desc = node->GetOpDesc();
  745. GE_CHECK_NOTNULL(src_op_desc);
  746. ge::GeShape src_shape = src_desc.GetShape();
  747. ge::Format src_format = src_desc.GetFormat();
  748. net_output_input_desc->SetDataType(DT_FLOAT16);
  749. if (is_dynamic) {
  750. auto merge_output = src_op_desc->MutableOutputDesc(0);
  751. GE_CHECK_NOTNULL(merge_output);
  752. merge_output->SetDataType(DT_FLOAT16);
  753. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  754. auto merge_input = src_op_desc->MutableInputDesc(i);
  755. GE_CHECK_NOTNULL(merge_input);
  756. merge_input->SetDataType(DT_FLOAT16);
  757. }
  758. }
  759. std::vector<int64_t> dst_shape_dims;
  760. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  761. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  762. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  763. return FAILED;
  764. }
  765. ge::GeShape dst_shape(dst_shape_dims);
  766. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  767. net_output_input_desc->SetShape(dst_shape);
  768. if (is_dynamic) {
  769. auto merge_out = src_op_desc->MutableOutputDesc(0);
  770. GE_CHECK_NOTNULL(merge_out);
  771. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  772. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  773. return FAILED;
  774. }
  775. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  776. auto merge_in = src_op_desc->MutableInputDesc(i);
  777. GE_CHECK_NOTNULL(merge_in);
  778. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  779. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  780. return FAILED;
  781. }
  782. }
  783. }
  784. return SUCCESS;
  785. }
  786. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  787. GE_CHECK_NOTNULL(netout_desc);
  788. vector<string> output_dt_str;
  789. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  790. for (auto dt_str : output_dt_str) {
  791. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  792. if (dt_str_split.size() == kUserDefinedElementCount) {
  793. if (dt_str_split[0] == to_string(index)) {
  794. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  795. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  796. TypeUtils::DataTypeToSerialString(dt).c_str());
  797. return true;
  798. }
  799. }
  800. }
  801. }
  802. return false;
  803. }
  804. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  805. GE_CHECK_NOTNULL(netout_desc);
  806. vector<string> output_format_str;
  807. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  808. for (auto format_str : output_format_str) {
  809. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  810. if (format_str_split.size() == kUserDefinedElementCount) {
  811. if (format_str_split[0] == to_string(index)) {
  812. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  813. return true;
  814. }
  815. }
  816. }
  817. }
  818. return false;
  819. }
  820. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  821. auto op_desc = node->GetOpDesc();
  822. GE_CHECK_NOTNULL(op_desc);
  823. ge::DataType output_data_type = ge::DT_FLOAT;
  824. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  825. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  826. auto peer_out = in_anchor->GetPeerOutAnchor();
  827. GE_CHECK_NOTNULL(peer_out);
  828. auto src_node = peer_out->GetOwnerNode();
  829. GE_CHECK_NOTNULL(src_node);
  830. bool is_dynamic = CheckOpType(src_node, MERGE);
  831. OpDescPtr src_op_desc = src_node->GetOpDesc();
  832. GE_CHECK_NOTNULL(src_op_desc);
  833. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  834. GE_CHECK_NOTNULL(net_output_input_desc);
  835. ge::GeShape old_shape = net_output_input_desc->GetShape();
  836. ge::Format old_format = net_output_input_desc->GetFormat();
  837. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  838. // Update datatype
  839. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  840. GELOGI("Enter into process output_type schedule");
  841. net_output_input_desc->SetDataType(output_data_type);
  842. if (is_dynamic) {
  843. auto merge_output = src_op_desc->MutableOutputDesc(0);
  844. GE_CHECK_NOTNULL(merge_output);
  845. merge_output->SetDataType(output_data_type);
  846. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  847. auto merge_input = src_op_desc->MutableInputDesc(i);
  848. GE_CHECK_NOTNULL(merge_input);
  849. merge_input->SetDataType(output_data_type);
  850. }
  851. }
  852. }
  853. // check if is_output_adjust_hw_layout is set
  854. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  855. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  856. ErrorManager::GetInstance().ATCReportErrMessage(
  857. "E19014", {"opname", "value", "reason"},
  858. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  859. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  860. GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0.");
  861. return FAILED;
  862. }
  863. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  864. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  865. GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0.");
  866. return FAILED;
  867. }
  868. }
  869. }
  870. return SUCCESS;
  871. }
  872. } // namespace
  873. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  874. GraphPrepare::~GraphPrepare() {}
  875. /**
  876. * @param graph
  877. * @return
  878. */
  879. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  880. GE_CHECK_NOTNULL(graph);
  881. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  882. for (auto &node : graph->GetAllNodes()) {
  883. if (node == nullptr) {
  884. continue;
  885. }
  886. if (node->GetType() != VARIABLE) {
  887. continue;
  888. }
  889. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  890. if (trans_road == nullptr) {
  891. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  892. continue;
  893. }
  894. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  895. auto ret = RecoverTransRoadForVar(node, *trans_road);
  896. if (ret != SUCCESS) {
  897. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str());
  898. return INTERNAL_ERROR;
  899. }
  900. auto iter = var_names_to_refs.find(node->GetName());
  901. if (iter != var_names_to_refs.end()) {
  902. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  903. if (ret != SUCCESS) {
  904. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str());
  905. return INTERNAL_ERROR;
  906. }
  907. }
  908. }
  909. return SUCCESS;
  910. }
  911. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  912. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  913. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  914. if (compute_graph_ != nullptr) {
  915. compute_graph_->SetSessionID(session_id);
  916. }
  917. session_id_ = session_id;
  918. Status ret = CheckGraph();
  919. if (ret != SUCCESS) {
  920. GELOGE(ret, "RunGraph graph check fail, ret:%u", ret);
  921. return ret;
  922. }
  923. (void)compute_graph_->TopologicalSorting();
  924. ret = CheckRefOp();
  925. if (ret != SUCCESS) {
  926. GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret);
  927. return ret;
  928. }
  929. return SUCCESS;
  930. }
  931. Status GraphPrepare::CheckGraph() {
  932. if (compute_graph_ == nullptr) {
  933. GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR");
  934. return GE_GRAPH_INIT_FAILED;
  935. }
  936. auto nodes = compute_graph_->GetAllNodes();
  937. if (nodes.empty()) {
  938. GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph.");
  939. return GE_GRAPH_INIT_FAILED;
  940. }
  941. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  942. GE_CHECK_NOTNULL(node);
  943. if (node->GetOpDesc() == nullptr) {
  944. GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL");
  945. return GE_GRAPH_INIT_FAILED;
  946. }
  947. }
  948. return SUCCESS;
  949. }
  950. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  951. const std::set<NodePtr> &ref_nodes) {
  952. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  953. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  954. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  955. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  956. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  957. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH};
  958. GE_CHECK_NOTNULL(node);
  959. const auto &op_desc = node->GetOpDesc();
  960. GE_CHECK_NOTNULL(op_desc);
  961. const auto input_index = op_desc->GetInputIndexByName(input_name);
  962. const auto &in_anchor = node->GetInDataAnchor(input_index);
  963. GE_CHECK_NOTNULL(in_anchor);
  964. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  965. GE_CHECK_NOTNULL(peer_out_anchor);
  966. const auto &input_node = peer_out_anchor->GetOwnerNode();
  967. GE_CHECK_NOTNULL(input_node);
  968. const auto &input_op_desc = input_node->GetOpDesc();
  969. GE_CHECK_NOTNULL(input_op_desc);
  970. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  971. if (is_ref) {
  972. return SUCCESS;
  973. }
  974. auto input_type = input_op_desc->GetType();
  975. if (input_type == ge::FRAMEWORKOP) {
  976. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  977. GELOGE(PARAM_INVALID, "Get original type failed.");
  978. return PARAM_INVALID;
  979. }
  980. }
  981. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  982. if (!is_acceptable) {
  983. ErrorManager::GetInstance().ATCReportErrMessage(
  984. "E15005", {"opname", "optype", "opname1", "optype1"},
  985. {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()});
  986. GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.",
  987. node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  988. input_op_desc->GetType().c_str());
  989. return PARAM_INVALID;
  990. }
  991. return SUCCESS;
  992. }
  993. Status GraphPrepare::CheckRefOp() {
  994. GE_CHECK_NOTNULL(compute_graph_);
  995. std::set<NodePtr> ref_nodes;
  996. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  997. if (node == nullptr) {
  998. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  999. return PARAM_INVALID;
  1000. }
  1001. auto op_desc = node->GetOpDesc();
  1002. if (op_desc == nullptr) {
  1003. GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null.");
  1004. return PARAM_INVALID;
  1005. }
  1006. auto input_name_index = op_desc->GetAllInputName();
  1007. auto outputs = op_desc->GetAllOutputName();
  1008. for (const auto &name_index : input_name_index) {
  1009. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  1010. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  1011. GELOGE(PARAM_INVALID, "CheckRefInputNode failed.");
  1012. return PARAM_INVALID;
  1013. }
  1014. (void)ref_nodes.insert(node); // no need to check value
  1015. }
  1016. }
  1017. }
  1018. return SUCCESS;
  1019. };
  1020. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  1021. GE_CHECK_NOTNULL(compute_graph_);
  1022. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  1023. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  1024. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  1025. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  1026. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  1027. return SUCCESS;
  1028. }
  1029. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1030. if (node == nullptr) {
  1031. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL");
  1032. return GE_GRAPH_GRAPH_NODE_NULL;
  1033. }
  1034. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1035. if (op_desc_ptr == nullptr) {
  1036. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL");
  1037. return GE_GRAPH_GRAPH_NODE_NULL;
  1038. }
  1039. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1040. int64_t tensor_size = 0;
  1041. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1042. if (graph_status != GRAPH_SUCCESS) {
  1043. ErrorManager::GetInstance().ATCReportErrMessage(
  1044. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1045. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  1046. return FAILED;
  1047. }
  1048. TensorUtils::SetSize(output, tensor_size);
  1049. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1050. if (graph_ret != GRAPH_SUCCESS) {
  1051. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1052. return graph_ret;
  1053. }
  1054. return SUCCESS;
  1055. }
  1056. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input) {
  1057. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1058. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1059. GE_CHECK_NOTNULL(input_node);
  1060. OpDescPtr op = input_node->GetOpDesc();
  1061. GE_CHECK_NOTNULL(op);
  1062. if (op->GetType() == DATA) {
  1063. GeAttrValue::INT index = 0;
  1064. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1065. GELOGW("Get index from data attr failed");
  1066. continue;
  1067. }
  1068. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1069. std::string situation = "data op index[" + std::to_string(index) + "]";
  1070. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1071. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1072. GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index);
  1073. return FAILED;
  1074. }
  1075. if (IsDynamicDims(input_node)) {
  1076. continue;
  1077. }
  1078. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1079. auto format = desc.GetFormat();
  1080. auto origin_format = desc.GetOriginFormat();
  1081. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1082. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op);
  1083. if (need_check_internal_format) {
  1084. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1085. if (is_internal) {
  1086. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1087. {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" +
  1088. TypeUtils::FormatToSerialString(origin_format) + "]", "it is not support"});
  1089. GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.",
  1090. TypeUtils::FormatToSerialString(format).c_str(),
  1091. TypeUtils::FormatToSerialString(origin_format).c_str());
  1092. return FAILED;
  1093. }
  1094. }
  1095. auto data_type = desc.GetDataType();
  1096. uint32_t length = 1;
  1097. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1098. if (!type_ret) {
  1099. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1100. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1101. GELOGE(PARAM_INVALID, "Input datatype %s is not support.",
  1102. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1103. return FAILED;
  1104. }
  1105. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1106. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1107. int64_t shape_size = desc_shape * length;
  1108. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1109. int64_t size = 0;
  1110. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1111. GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed");
  1112. return FAILED);
  1113. bool size_check = (size != 0 && shape_size != size);
  1114. if (size_check) {
  1115. std::string situation = "input data size[" + std::to_string(size) +
  1116. "] and shape_size[" + std::to_string(size) + "]";
  1117. std::string reason = "because size != 0 and shape_size != size";
  1118. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1119. GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size);
  1120. return FAILED;
  1121. }
  1122. ge::TensorUtils::SetSize(desc, shape_size);
  1123. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1124. if (graph_ret != GRAPH_SUCCESS) {
  1125. GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
  1126. return graph_ret;
  1127. }
  1128. // Size will be recalculated in the build stage
  1129. ge::TensorUtils::SetSize(desc, 0);
  1130. graph_ret = op->UpdateOutputDesc(0, desc);
  1131. if (graph_ret != GRAPH_SUCCESS) {
  1132. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1133. return graph_ret;
  1134. }
  1135. if (!options_.train_graph_flag) {
  1136. Status ret = AdjustDataOpOutput(input_node);
  1137. GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret); return ret);
  1138. }
  1139. }
  1140. }
  1141. return SUCCESS;
  1142. }
  1143. Status GraphPrepare::TryDoAipp() {
  1144. // infer and with aipp configure file, then call aipp insert
  1145. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1146. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1147. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1148. if (ret != SUCCESS) {
  1149. GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed.");
  1150. return INTERNAL_ERROR;
  1151. }
  1152. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1153. if (ret != SUCCESS) {
  1154. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed",
  1155. options_.insert_op_file.c_str());
  1156. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1157. }
  1158. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1159. if (ret != SUCCESS) {
  1160. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret);
  1161. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1162. }
  1163. }
  1164. return SUCCESS;
  1165. }
  1166. Status GraphPrepare::FormatAndShapeProcess() {
  1167. Status ret = ResourcePairProcess("add");
  1168. if (ret != SUCCESS) {
  1169. GELOGE(ret, "ResourcePairProcess failed");
  1170. return ret;
  1171. }
  1172. GE_TIMESTAMP_START(InferOriginFormat1);
  1173. ret = compute_graph_->InferOriginFormat();
  1174. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1175. GE_DUMP(compute_graph_, "after_first_inferformat");
  1176. if (ret != SUCCESS) {
  1177. GELOGE(ret, "Prepare Graph first inferformat failed");
  1178. return ret;
  1179. }
  1180. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1181. ret = InferShapeForPreprocess();
  1182. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1183. GE_DUMP(compute_graph_, "after_infershape");
  1184. if (ret != SUCCESS) {
  1185. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed");
  1186. return GE_GRAPH_INFERSHAPE_FAILED;
  1187. }
  1188. GE_TIMESTAMP_START(InferOriginFormat2);
  1189. ret = compute_graph_->InferOriginFormat();
  1190. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1191. if (ret != SUCCESS) {
  1192. GELOGE(ret, "Prepare Graph inferformat failed");
  1193. return ret;
  1194. }
  1195. ret = ResourcePairProcess("remove");
  1196. if (ret != SUCCESS) {
  1197. return ret;
  1198. }
  1199. return ret;
  1200. }
  1201. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1202. PassManager control_pass;
  1203. // Graph pass tmp logic for resource infershape
  1204. if (options_.train_graph_flag) {
  1205. try {
  1206. if (action == "add") {
  1207. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1208. } else {
  1209. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1210. new ResourcePairRemoveControlPass);
  1211. }
  1212. } catch (std::bad_alloc &e) {
  1213. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str());
  1214. return INTERNAL_ERROR;
  1215. }
  1216. }
  1217. Status ret = control_pass.Run(compute_graph_);
  1218. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1219. GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret);
  1220. return ret;
  1221. }
  1222. return SUCCESS;
  1223. }
  1224. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1225. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1226. GE_CHECK_NOTNULL(node_ptr);
  1227. if (node_ptr->GetType() == DATA) {
  1228. uint32_t index = 0;
  1229. auto op_desc = node_ptr->GetOpDesc();
  1230. GE_CHECK_NOTNULL(op_desc);
  1231. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1232. Format storage_format = FORMAT_RESERVED;
  1233. vector<int64_t> dst_shape_dims;
  1234. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1235. GELOGE(INTERNAL_ERROR, "Get storage format for input failed");
  1236. return FAILED;
  1237. }
  1238. if (storage_format == FORMAT_RESERVED) {
  1239. continue;
  1240. }
  1241. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1242. GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed");
  1243. return FAILED;
  1244. }
  1245. }
  1246. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1247. auto op_desc = node_ptr->GetOpDesc();
  1248. GE_CHECK_NOTNULL(op_desc);
  1249. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1250. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1251. Format storage_format = FORMAT_RESERVED;
  1252. vector<int64_t> dst_shape_dims;
  1253. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1254. GELOGE(INTERNAL_ERROR, "Get storage format from output failed");
  1255. return FAILED;
  1256. }
  1257. if (storage_format == FORMAT_RESERVED) {
  1258. continue;
  1259. }
  1260. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1261. GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed");
  1262. return FAILED;
  1263. }
  1264. }
  1265. }
  1266. }
  1267. return SUCCESS;
  1268. }
  1269. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1270. if (options_.save_original_model == "true") {
  1271. ModelHelper model_helper;
  1272. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1273. options_.original_model_file);
  1274. if (ret != SUCCESS) {
  1275. // If save original model fail, process continue
  1276. GELOGW("SaveOriginalGraphToOmModel fail");
  1277. }
  1278. }
  1279. return SUCCESS;
  1280. }
  1281. #define PP_RUN_AND_DUMP(name, func, ...) \
  1282. do { \
  1283. GE_RUN(Prepare, func, __VA_ARGS__); \
  1284. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1285. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1286. } while (0)
  1287. #define PP_RUN(name, func, ...) \
  1288. do { \
  1289. GE_RUN(Prepare, func, __VA_ARGS__); \
  1290. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1291. } while (0)
  1292. Status GraphPrepare::PrepareDynShape(ConstGraphPtr graph, const std::vector<GeTensor> &user_input,
  1293. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1294. GE_CHECK_NOTNULL(graph);
  1295. GE_CHECK_NOTNULL(compute_graph);
  1296. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1297. const Graph &const_graph = *graph;
  1298. PP_RUN("Init", Init, const_graph, session_id);
  1299. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1300. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input);
  1301. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1302. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1303. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1304. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1305. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1306. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1307. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1308. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1309. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1310. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1311. return SUCCESS;
  1312. }
  1313. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1314. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1315. return SUCCESS;
  1316. }
  1317. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1318. auto compute_graph = compute_graph_;
  1319. PassManager pass_manager;
  1320. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1321. new (std::nothrow) VariablePrepareOpPass))
  1322. GE_TIMESTAMP_START(pass_manager);
  1323. auto ret = pass_manager.Run(compute_graph);
  1324. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1325. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1326. GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret);
  1327. return ret;
  1328. }
  1329. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1330. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1331. return SUCCESS;
  1332. }
  1333. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1334. if (compute_graph == nullptr) {
  1335. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1336. return GE_GRAPH_NULL_INPUT;
  1337. }
  1338. GEPass ge_passes(compute_graph);
  1339. NamesToPass hccl_group;
  1340. HcclGroupPass hccl_group_pass;
  1341. GELOGD("Add hccl group pass success");
  1342. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1343. auto ret = ge_passes.Run(hccl_group);
  1344. if (ret != SUCCESS) {
  1345. GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret);
  1346. return ret;
  1347. }
  1348. ret = compute_graph->TopologicalSorting();
  1349. if (ret != SUCCESS) {
  1350. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1351. return ret;
  1352. }
  1353. return SUCCESS;
  1354. }
  1355. #undef PP_RUN_AND_DUMP
  1356. #undef PP_RUN
  1357. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1358. if (graph == nullptr) {
  1359. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1360. return GE_GRAPH_NULL_INPUT;
  1361. }
  1362. const Graph &const_graph = *graph;
  1363. Status ret = Init(const_graph, 0);
  1364. if (ret != SUCCESS) {
  1365. GELOGE(ret, "Init graph_prepare fail, ret:%u", ret);
  1366. return ret;
  1367. }
  1368. GE_DUMP(compute_graph_, "after_parser");
  1369. GELOGI("Start infershape for dump json process.");
  1370. ret = compute_graph_->InferOriginFormat();
  1371. GE_DUMP(compute_graph_, "after_inferformat");
  1372. if (ret != SUCCESS) {
  1373. GELOGE(ret, "Prepare Graph inferformat failed");
  1374. return ret;
  1375. }
  1376. InferShapePass infer_shape_pass;
  1377. NamesToPass names_to_passes;
  1378. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1379. GEPass ge_passes(compute_graph_);
  1380. ret = ge_passes.Run(names_to_passes);
  1381. GE_DUMP(compute_graph_, "after_infershape");
  1382. if (ret != SUCCESS) {
  1383. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1384. return ret;
  1385. }
  1386. ShapeRefiner::ClearContextMap();
  1387. return SUCCESS;
  1388. }
  1389. Status GraphPrepare::CheckConstOp() {
  1390. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1391. GE_CHECK_NOTNULL(node_ptr);
  1392. if (node_ptr->GetType() == CONSTANT) {
  1393. Status ret = VerifyConstOp(node_ptr);
  1394. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1395. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1396. auto op_desc = node_ptr->GetOpDesc();
  1397. if (op_desc == nullptr) {
  1398. GELOGE(PARAM_INVALID, "Get op desc failed");
  1399. return PARAM_INVALID;
  1400. }
  1401. std::string original_type;
  1402. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1403. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1404. GELOGI("original type is %s", original_type.c_str());
  1405. if (original_type == CONSTANT) {
  1406. Status ret = VerifyConstOp(node_ptr);
  1407. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1408. }
  1409. }
  1410. }
  1411. return SUCCESS;
  1412. }
  1413. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1414. GE_CHECK_NOTNULL(node);
  1415. auto op_desc = node->GetOpDesc();
  1416. GE_CHECK_NOTNULL(op_desc);
  1417. ConstGeTensorPtr ge_tensor_ptr;
  1418. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1419. GELOGE(PARAM_INVALID, "Get value from const attr failed");
  1420. return PARAM_INVALID;
  1421. }
  1422. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1423. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1424. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1425. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1426. auto data_type = ge_tensor_desc.GetDataType();
  1427. uint32_t length = 1;
  1428. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1429. if (!type_ret) {
  1430. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1431. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1432. GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str());
  1433. return FAILED;
  1434. }
  1435. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1436. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1437. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1438. if (shape_size == 0) {
  1439. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1440. // shape = [], means it's a sclar tensor.
  1441. GE_CHK_BOOL_EXEC(data_size / length == 1,
  1442. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid scalar tensor."});
  1443. return PARAM_INVALID, "Const is invalid scalar tensor.");
  1444. } else {
  1445. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1446. GE_CHK_BOOL_EXEC(data_size == 0,
  1447. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid vector scalar."});
  1448. return PARAM_INVALID, "Const is invalid vector scalar.");
  1449. }
  1450. } else {
  1451. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0,
  1452. ErrorManager::GetInstance().ATCReportErrMessage(
  1453. "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"});
  1454. return PARAM_INVALID, "Const input data size is not equal with tensor desc shape");
  1455. }
  1456. return SUCCESS;
  1457. }
  1458. bool GraphPrepare::IsDynamicDims(const NodePtr &input_node) {
  1459. auto data_shape = NodeUtils::GetOutputDesc(*input_node, kDataOutIndex).GetShape();
  1460. const auto &dims = data_shape.GetDims();
  1461. bool all_is_positive = false;
  1462. if (std::all_of(dims.begin(), dims.end(), [](int64_t val) { return val >= 0; })) {
  1463. all_is_positive = true;
  1464. }
  1465. if (!all_is_positive && !options_.input_shape.empty() && !options_.dynamic_dims.empty() &&
  1466. options_.dynamic_node_type != kInvalidDynaimcDimsType) {
  1467. GELOGI("No need to check and update desc info, the dims of %s is %s.", input_node->GetName().c_str(),
  1468. formats::JoinToString(dims).c_str());
  1469. return true;
  1470. }
  1471. return false;
  1472. }
  1473. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1474. if (GetLocalOmgContext().is_dynamic_input) {
  1475. return SUCCESS;
  1476. }
  1477. unsigned int node_num = 0;
  1478. unsigned int data_num = 0;
  1479. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1480. GE_CHECK_NOTNULL(input_node);
  1481. OpDescPtr op = input_node->GetOpDesc();
  1482. GE_CHECK_NOTNULL(op);
  1483. node_num++;
  1484. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1485. data_num++;
  1486. GeAttrValue::INT index = 0;
  1487. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1488. GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed");
  1489. return GE_GRAPH_INIT_FAILED;
  1490. }
  1491. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1492. std::string situation = "data op index[" + std::to_string(index) + "]";
  1493. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1494. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1495. GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index);
  1496. return GE_GRAPH_INIT_FAILED;
  1497. }
  1498. if (IsDynamicDims(input_node)) {
  1499. continue;
  1500. }
  1501. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1502. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1503. if (desc.GetShape().GetDim(i) < 0) {
  1504. std::string situation = "data dim[" + std::to_string(i) + "][" +
  1505. std::to_string(desc.GetShape().GetDim(i)) + "]" ;
  1506. std::string reason = "it need >= 0";
  1507. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1508. GELOGE(GE_GRAPH_INIT_FAILED, "data dim %zu is not supported, need >= 0, real:%ld.", i,
  1509. desc.GetShape().GetDim(i));
  1510. return GE_GRAPH_INIT_FAILED;
  1511. }
  1512. }
  1513. }
  1514. }
  1515. if (node_num <= data_num) {
  1516. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1517. }
  1518. return SUCCESS;
  1519. }
  1520. Status GraphPrepare::InferShapeForPreprocess() {
  1521. GELOGI("Start infershape for preprocess.");
  1522. GEPass ge_passes(compute_graph_);
  1523. NamesToPass names_to_passes;
  1524. AssertPass assert_pass;
  1525. if (!options_.train_graph_flag) {
  1526. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1527. }
  1528. SwitchDeadBranchElimination switch_dead_branch_elimination;
  1529. names_to_passes.emplace_back("SwitchDeadBranchElimination", &switch_dead_branch_elimination);
  1530. MergePass merge_pass;
  1531. names_to_passes.emplace_back("MergePass", &merge_pass);
  1532. InferShapePass infer_shape_pass;
  1533. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1534. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1535. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1536. DimensionComputePass dimension_compute_pass;
  1537. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1538. ConstantFoldingPass constant_folding_pass;
  1539. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1540. int32_t dev_count = 0;
  1541. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1542. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1543. rtError_t rt_err = RT_ERROR_NONE;
  1544. if (aicpu_constant_folding_on != nullptr) {
  1545. rt_err = rtGetDeviceCount(&dev_count);
  1546. if (rt_err == RT_ERROR_NONE) {
  1547. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1548. if (result != SUCCESS) {
  1549. GELOGE(result, "Set rt context failed.");
  1550. return result;
  1551. }
  1552. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1553. }
  1554. }
  1555. Status ret = ge_passes.Run(names_to_passes);
  1556. if (aicpu_constant_folding_on != nullptr) {
  1557. if (rt_err == RT_ERROR_NONE) {
  1558. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1559. if (result != SUCCESS) {
  1560. GELOGE(result, "Set rt context failed.");
  1561. return result;
  1562. }
  1563. }
  1564. }
  1565. ShapeRefiner::ClearContextMap();
  1566. if (ret != SUCCESS) {
  1567. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1568. return ret;
  1569. }
  1570. return SUCCESS;
  1571. }
  1572. Status GraphPrepare::PrepareOptimize() {
  1573. GELOGI("Start optimize for preprocess.");
  1574. // check rw type
  1575. GraphOptimize graph_optimize;
  1576. bool has_conflict = false;
  1577. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1578. if (has_conflict) {
  1579. GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize.");
  1580. return FAILED;
  1581. }
  1582. PassManager original_graph_passes;
  1583. // Graph pass
  1584. try {
  1585. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1586. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1587. (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass);
  1588. } catch (std::bad_alloc &e) {
  1589. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1590. return INTERNAL_ERROR;
  1591. }
  1592. GE_TIMESTAMP_START(original_graph_passes);
  1593. Status ret = original_graph_passes.Run(compute_graph_);
  1594. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1595. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1596. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1597. return ret;
  1598. }
  1599. // New pass
  1600. GEPass ge_passes(compute_graph_);
  1601. NamesToPass names_to_passes;
  1602. EnterPass enter_pass;
  1603. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1604. CondPass cond_pass;
  1605. names_to_passes.emplace_back("CondPass", &cond_pass);
  1606. PrintOpPass print_pass;
  1607. if (options_.enable_print_op_pass) {
  1608. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1609. }
  1610. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1611. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1612. DropOutPass dropout_pass;
  1613. AssertPass assert_pass;
  1614. UnusedConstPass unused_const_pass;
  1615. StopGradientPass stop_gradient_pass;
  1616. PreventGradientPass prevent_gradient_pass;
  1617. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1618. GuaranteeConstPass guarantee_const_pass;
  1619. VarIsInitializedOpPass var_is_initialized_pass;
  1620. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1621. IdentityPass identity_pass(false);
  1622. AssignPass assign_pass;
  1623. SnapshotPass snapshot_pass;
  1624. if (!options_.train_graph_flag) {
  1625. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1626. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1627. }
  1628. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1629. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1630. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1631. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1632. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1633. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1634. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1635. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1636. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1637. if (GetContext().GetHostExecFlag()) {
  1638. names_to_passes.emplace_back("AssignPass", &assign_pass);
  1639. }
  1640. GE_TIMESTAMP_START(names_to_passes);
  1641. ret = ge_passes.Run(names_to_passes);
  1642. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1643. if (ret != SUCCESS) {
  1644. GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret);
  1645. return ret;
  1646. }
  1647. PassManager graph_pass;
  1648. try {
  1649. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1650. // todo 临时把hccl的memcpy插入放到图准备,为了防止其多插memcpy
  1651. (void)graph_pass.AddPass("PrepareOptimize::HcclMemcpyPass", new (std::nothrow) HcclMemcpyPass);
  1652. } catch (std::bad_alloc &e) {
  1653. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1654. return INTERNAL_ERROR;
  1655. }
  1656. GE_TIMESTAMP_START(graph_passes);
  1657. ret = graph_pass.Run(compute_graph_);
  1658. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1659. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1660. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1661. return ret;
  1662. }
  1663. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1664. TypeConversionOfConstant();
  1665. ret = compute_graph_->TopologicalSorting();
  1666. if (ret != SUCCESS) {
  1667. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1668. return ret;
  1669. }
  1670. GELOGI("End optimize for preprocess.");
  1671. return SUCCESS;
  1672. }
  1673. void GraphPrepare::TypeConversionOfConstant() {
  1674. bool is_acl_compile = false;
  1675. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1676. // This can ensure that n is not a null pointer
  1677. // No Conversion when called by aclOpCompile
  1678. (void)AttrUtils::GetBool(n->GetOpDesc(), ATTR_DYNAMIC_SHAPE_SINGLE_AICPU, is_acl_compile);
  1679. if (is_acl_compile) {
  1680. return;
  1681. }
  1682. }
  1683. if (options_.train_graph_flag) {
  1684. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1685. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1686. // This can ensure that n is not a null pointer
  1687. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1688. n->GetOpDesc()->SetType(CONSTANTOP);
  1689. }
  1690. }
  1691. } else {
  1692. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  1693. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1694. // This can ensure that n is not a null pointer
  1695. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  1696. n->GetOpDesc()->SetType(CONSTANT);
  1697. }
  1698. }
  1699. }
  1700. }
  1701. Status GraphPrepare::GraphEquivalentTransformation() {
  1702. NamesToPass names_to_pass;
  1703. ForPass for_pass;
  1704. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  1705. return GEPass(compute_graph_).Run(names_to_pass);
  1706. }
  1707. Status GraphPrepare::ProcessBeforeInfershape() {
  1708. NamesToPass names_to_passes;
  1709. CondRemovePass condition_remove_pass;
  1710. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  1711. GE_TIMESTAMP_START(ProcessCondRemove);
  1712. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  1713. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  1714. if (ret != SUCCESS) {
  1715. GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  1716. return ret;
  1717. }
  1718. return SUCCESS;
  1719. }
  1720. Status GraphPrepare::ProcessNetOutput() {
  1721. PassManager graph_passes_before_infershape;
  1722. try {
  1723. if (options_.train_graph_flag) {
  1724. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  1725. }
  1726. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  1727. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  1728. new (std::nothrow) DataPass); // Add NetOutput first.
  1729. } catch (std::bad_alloc) {
  1730. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1731. return INTERNAL_ERROR;
  1732. }
  1733. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  1734. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  1735. GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret);
  1736. return ret;
  1737. }
  1738. return SUCCESS;
  1739. }
  1740. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input) {
  1741. compute_graph_->SetInputSize(user_input.size());
  1742. if (user_input.empty()) {
  1743. return SUCCESS;
  1744. }
  1745. auto ret = CheckUserInput(user_input);
  1746. if (ret != SUCCESS) {
  1747. GELOGE(ret, "Check user input failed.");
  1748. return ret;
  1749. }
  1750. ret = UpdateInput(user_input);
  1751. if (ret != SUCCESS) {
  1752. GELOGE(ret, "UpdateInput fail, ret:%u", ret);
  1753. return ret;
  1754. }
  1755. if (user_input.size() != 0) {
  1756. ret = CheckConstOp();
  1757. if (ret != SUCCESS) {
  1758. GELOGE(ret, "CheckConstOp fail, ret:%u", ret);
  1759. return ret;
  1760. }
  1761. } else {
  1762. ret = compute_graph_->TopologicalSorting();
  1763. if (ret != SUCCESS) {
  1764. GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting");
  1765. return FAILED;
  1766. }
  1767. }
  1768. return SUCCESS;
  1769. }
  1770. Status GraphPrepare::UpdateInputOutputByOptions() {
  1771. auto ret = UpdateDataNetOutputByStorageFormat();
  1772. if (ret != SUCCESS) {
  1773. GELOGE(ret, "Update format acoording to storage format failed.");
  1774. return ret;
  1775. }
  1776. if (options_.train_graph_flag) {
  1777. GELOGI("This is train mode, no need to do this schedule.");
  1778. return SUCCESS;
  1779. }
  1780. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  1781. GE_CHECK_NOTNULL(node_ptr);
  1782. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  1783. GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str());
  1784. return FAILED;
  1785. }
  1786. if (node_ptr->GetType() == DATA) {
  1787. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  1788. GELOGE(INTERNAL_ERROR, "Process data node failed");
  1789. return FAILED;
  1790. }
  1791. }
  1792. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1793. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  1794. GELOGE(INTERNAL_ERROR, "Process netoutput node failed");
  1795. return FAILED;
  1796. }
  1797. }
  1798. }
  1799. return SUCCESS;
  1800. }
  1801. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  1802. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  1803. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  1804. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  1805. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  1806. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  1807. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  1808. if (dst_node->GetType() == TRANSDATA) {
  1809. return true;
  1810. }
  1811. }
  1812. }
  1813. return false;
  1814. }
  1815. } // 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