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 79 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
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
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
5 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924
  1. /**
  2. * Copyright 2019-2020 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. namespace ge {
  118. namespace {
  119. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  120. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  121. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  122. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  123. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  124. // the size of user defined output datatype or format string after split by ":".
  125. const size_t kUserDefinedElementCount = 2;
  126. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  127. GeTensorPtr tensor = MakeShared<GeTensor>();
  128. if (tensor == nullptr) {
  129. GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed");
  130. return nullptr;
  131. }
  132. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  133. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  134. auto dst_ge_shape = data_tensor.GetShape();
  135. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  136. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  137. tensor->MutableTensorDesc().SetShape(GeShape());
  138. int32_t dst_shape = 1;
  139. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  140. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  141. return nullptr;
  142. }
  143. } else {
  144. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  145. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  146. if (dst_shape == nullptr) {
  147. GELOGE(INTERNAL_ERROR, "Create unique ptr failed");
  148. return nullptr;
  149. }
  150. for (int64_t i = 0; i < dim_cnt; ++i) {
  151. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  152. }
  153. GE_IF_BOOL_EXEC(
  154. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  155. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  156. return nullptr;)
  157. }
  158. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  159. return OpDescUtils::CreateConstOp(tensor);
  160. }
  161. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  162. OpDescPtr &op_desc) {
  163. // For format transfer node, the IR definition has src/dst format attrs
  164. if (node_type == TRANSDATA) {
  165. GE_IF_BOOL_EXEC(
  166. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  167. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  168. GE_IF_BOOL_EXEC(
  169. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  170. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  171. }
  172. // For TransposeD node, the IR definition has perm attrs
  173. if (node_type == TRANSPOSED) {
  174. Format src_format = input.GetFormat();
  175. Format dst_format = output.GetFormat();
  176. std::vector<int64_t> perm_arg;
  177. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  178. "Get perm by foramt failed.");
  179. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  180. "SetStr PERMUTE_ATTR_PERM failed")
  181. }
  182. // For cast node, the IR definition has src/dst attrs
  183. if (node_type == CAST) {
  184. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  185. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  186. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  187. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  188. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  189. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  190. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  191. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  192. }
  193. }
  194. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  195. const GeTensorDesc &output, NodePtr &node) {
  196. if (node == nullptr) {
  197. GELOGE(PARAM_INVALID, "node is null.");
  198. return nullptr;
  199. }
  200. auto graph = node->GetOwnerComputeGraph();
  201. if (graph == nullptr) {
  202. GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str());
  203. return nullptr;
  204. }
  205. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  206. if (index < 0) {
  207. ErrorManager::GetInstance().ATCReportErrMessage(
  208. "E19025", {"situation", "reason"},
  209. {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()});
  210. GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str());
  211. return nullptr;
  212. }
  213. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  214. if (op_desc == nullptr) {
  215. GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed");
  216. return nullptr;
  217. }
  218. // for data dump
  219. GE_IF_BOOL_EXEC(
  220. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  221. GELOGW("CreateTransNode: SetListStr failed");)
  222. // Default single input and single output
  223. auto ret = op_desc->AddInputDesc(input);
  224. if (ret != GRAPH_SUCCESS) {
  225. GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str());
  226. return nullptr;
  227. }
  228. ret = op_desc->AddOutputDesc(output);
  229. if (ret != GRAPH_SUCCESS) {
  230. GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str());
  231. return nullptr;
  232. }
  233. AddTransNodeAttr(node_type, input, output, op_desc);
  234. NodePtr shape_node = nullptr;
  235. if (node_type == RESHAPE) {
  236. auto shape_desc = CreateTensorShape(output);
  237. if (shape_desc == nullptr) {
  238. GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input",
  239. node->GetName().c_str());
  240. return nullptr;
  241. }
  242. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  243. if (ret != GRAPH_SUCCESS) {
  244. GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str());
  245. return nullptr;
  246. }
  247. shape_node = graph->AddNode(shape_desc);
  248. if (shape_node == nullptr) {
  249. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str());
  250. return nullptr;
  251. }
  252. }
  253. auto trans_node = graph->AddNode(op_desc);
  254. if (trans_node == nullptr) {
  255. GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str());
  256. return nullptr;
  257. }
  258. if (node_type == RESHAPE) {
  259. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  260. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str());
  261. return nullptr;
  262. }
  263. }
  264. return trans_node;
  265. }
  266. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  267. NodePtr &trans_node) {
  268. GE_CHECK_NOTNULL(node);
  269. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  270. if (trans_node == nullptr) {
  271. return INTERNAL_ERROR;
  272. }
  273. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  274. if (ret != GRAPH_SUCCESS) {
  275. GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s",
  276. node->GetName().c_str(), node->GetType().c_str());
  277. return INTERNAL_ERROR;
  278. }
  279. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  280. if (ret != GRAPH_SUCCESS) {
  281. GELOGE(INTERNAL_ERROR, "Failed to connect node %s to trans node %s", node->GetName().c_str(),
  282. trans_node->GetName().c_str());
  283. return INTERNAL_ERROR;
  284. }
  285. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  286. if (ret != GRAPH_SUCCESS) {
  287. GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.",
  288. node->GetName().c_str(), trans_node->GetName().c_str());
  289. return INTERNAL_ERROR;
  290. }
  291. return SUCCESS;
  292. }
  293. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  294. NodePtr &trans_node) {
  295. GE_CHECK_NOTNULL(node);
  296. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  297. if (trans_node == nullptr) {
  298. return INTERNAL_ERROR;
  299. }
  300. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  301. if (ret != GRAPH_SUCCESS) {
  302. GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s",
  303. node->GetName().c_str(), node->GetType().c_str());
  304. return INTERNAL_ERROR;
  305. }
  306. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  307. if (ret != GRAPH_SUCCESS) {
  308. GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(),
  309. node->GetName().c_str());
  310. return INTERNAL_ERROR;
  311. }
  312. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  313. if (ret != GRAPH_SUCCESS) {
  314. GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.",
  315. node->GetName().c_str(), trans_node->GetName().c_str());
  316. return INTERNAL_ERROR;
  317. }
  318. return SUCCESS;
  319. }
  320. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  321. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  322. GE_CHECK_NOTNULL(var->GetOpDesc());
  323. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  324. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  325. output_desc.SetFormat(tensor_desc.GetFormat());
  326. output_desc.SetDataType(tensor_desc.GetDataType());
  327. output_desc.SetShape(tensor_desc.GetShape());
  328. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  329. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  330. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  331. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  332. GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed");
  333. return INTERNAL_ERROR;);
  334. }
  335. if (var->GetOpDesc()->GetInputsSize() > 0) {
  336. auto desc = var->GetOpDesc()->GetInputDesc(0);
  337. desc.SetFormat(tensor_desc.GetFormat());
  338. desc.SetDataType(tensor_desc.GetDataType());
  339. desc.SetShape(tensor_desc.GetShape());
  340. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  341. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  342. desc.SetOriginShape(tensor_desc.GetOriginShape());
  343. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  344. GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed");
  345. return INTERNAL_ERROR;)
  346. }
  347. return SUCCESS;
  348. }
  349. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  350. GE_CHECK_NOTNULL(var);
  351. int index = 0;
  352. NodePtr last_node = var;
  353. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  354. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  355. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  356. if (ret != SUCCESS) {
  357. ErrorManager::GetInstance().ATCReportErrMessage("E15001", {"variable", "index", "type"},
  358. {var->GetName(), std::to_string(index), iter->node_type});
  359. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(),
  360. index, iter->node_type.c_str());
  361. return INTERNAL_ERROR;
  362. }
  363. // set stream_label
  364. OpDescPtr var_desc = var->GetOpDesc();
  365. GE_CHECK_NOTNULL(var_desc);
  366. std::string stream_label;
  367. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  368. if (!stream_label.empty()) {
  369. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  370. }
  371. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  372. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  373. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  374. }
  375. if (road.empty()) {
  376. return SUCCESS;
  377. }
  378. return UpdateVarFormats(var, road.rbegin()->output);
  379. }
  380. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  381. for (auto &var : nodes) {
  382. GE_CHECK_NOTNULL(var);
  383. int index = 0;
  384. NodePtr last_node = var;
  385. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  386. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  387. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  388. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  389. if (ret != SUCCESS) {
  390. ErrorManager::GetInstance().ATCReportErrMessage("E15001", {"variable", "index", "type"},
  391. {var->GetName(), std::to_string(index), iter->node_type});
  392. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s",
  393. var->GetName().c_str(), index, iter->node_type.c_str());
  394. return INTERNAL_ERROR;
  395. }
  396. // set stream_label
  397. OpDescPtr var_desc = var->GetOpDesc();
  398. GE_CHECK_NOTNULL(var_desc);
  399. std::string stream_label;
  400. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  401. if (!stream_label.empty()) {
  402. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  403. }
  404. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  405. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  406. }
  407. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  408. return INTERNAL_ERROR;
  409. }
  410. }
  411. return SUCCESS;
  412. }
  413. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  414. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  415. VarNamesToRefs names_to_refs;
  416. std::string var_name;
  417. if (graph == nullptr) {
  418. GELOGE(PARAM_INVALID, "graph is null.");
  419. return names_to_refs;
  420. }
  421. for (auto &node : graph->GetAllNodes()) {
  422. if (node->GetType() != VARIABLE) {
  423. continue;
  424. }
  425. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  426. (void)names_to_refs[var_name].insert(node);
  427. }
  428. }
  429. return names_to_refs;
  430. }
  431. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  432. std::vector<int64_t> &dst_shape) {
  433. if (src_format == FORMAT_NCHW) {
  434. formats::FormatTransferNchwNc1hwc0 transfer;
  435. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  436. GELOGE(INTERNAL_ERROR, "TransShape failed");
  437. return FAILED;
  438. }
  439. } else if (src_format == FORMAT_NHWC) {
  440. formats::FormatTransferNhwcNc1hwc0 transfer;
  441. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  442. GELOGE(INTERNAL_ERROR, "TransShape failed");
  443. return FAILED;
  444. }
  445. }
  446. return SUCCESS;
  447. }
  448. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  449. GE_CHECK_NOTNULL(node_ptr);
  450. auto op_desc = node_ptr->GetOpDesc();
  451. GE_CHECK_NOTNULL(op_desc);
  452. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  453. GE_CHECK_NOTNULL(input);
  454. ge::Format old_format = input->GetFormat();
  455. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  456. ge::DataType dt = input->GetDataType();
  457. std::vector<int64_t> dst_shape_dims;
  458. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  459. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  460. return FAILED;
  461. }
  462. input->SetFormat(FORMAT_NC1HWC0);
  463. input->SetShape(ge::GeShape(dst_shape_dims));
  464. auto output = op_desc->MutableOutputDesc(0);
  465. GE_CHECK_NOTNULL(output);
  466. output->SetFormat(FORMAT_NC1HWC0);
  467. output->SetShape(ge::GeShape(dst_shape_dims));
  468. int64_t size = 0;
  469. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  470. if (graph_status != ge::GRAPH_SUCCESS) {
  471. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  472. return FAILED;
  473. }
  474. ge::TensorUtils::SetSize(*output, size);
  475. ge::TensorUtils::SetSize(*input, size);
  476. return SUCCESS;
  477. }
  478. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  479. GE_CHECK_NOTNULL(input_output);
  480. ge::Format old_format = input_output->GetFormat();
  481. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  482. ge::DataType dt = input_output->GetDataType();
  483. std::vector<int64_t> dst_shape_dims;
  484. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  485. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  486. return FAILED;
  487. }
  488. input_output->SetFormat(FORMAT_NC1HWC0);
  489. input_output->SetShape(ge::GeShape(dst_shape_dims));
  490. return SUCCESS;
  491. }
  492. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  493. vector<int64_t> &dst_shape_dims) {
  494. GE_CHECK_NOTNULL(op_desc);
  495. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  496. GE_CHECK_NOTNULL(input);
  497. ge::Format old_format = input->GetFormat();
  498. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  499. input->SetShape(ge::GeShape(dst_shape_dims));
  500. input->SetFormat(storage_format);
  501. auto output = op_desc->MutableOutputDesc(index);
  502. GE_CHECK_NOTNULL(output);
  503. output->SetShape(ge::GeShape(dst_shape_dims));
  504. output->SetFormat(storage_format);
  505. if (!output->MutableShape().IsUnknownShape()) {
  506. int64_t size = 0;
  507. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  508. if (graph_status != ge::GRAPH_SUCCESS) {
  509. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  510. return FAILED;
  511. }
  512. ge::TensorUtils::SetSize(*input, size);
  513. ge::TensorUtils::SetSize(*output, size);
  514. GELOGI(
  515. "Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  516. "new_shape:%s, new_format:%s, new_size:%lu",
  517. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  518. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  519. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  520. }
  521. return SUCCESS;
  522. }
  523. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) {
  524. is_dynamic_batch = false;
  525. std::string related_node_name;
  526. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  527. if (related_node_name.empty()) {
  528. ErrorManager::GetInstance().ATCReportErrMessage("E15002", {"opname", "value", "reason"},
  529. {data_node->GetName(), "flag", "but the value is empty"});
  530. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty",
  531. data_node->GetName().c_str());
  532. return INTERNAL_ERROR;
  533. }
  534. for (const NodePtr &next_node : data_node->GetOutNodes()) {
  535. if (next_node->GetName() == related_node_name) {
  536. switchn_node = next_node;
  537. break;
  538. }
  539. }
  540. if (switchn_node == nullptr) {
  541. ErrorManager::GetInstance().ATCReportErrMessage(
  542. "E15002", {"opname", "value", "reason"},
  543. {data_node->GetName(), related_node_name, "but can not find it on the graph"});
  544. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph",
  545. data_node->GetName().c_str(), related_node_name.c_str());
  546. return INTERNAL_ERROR;
  547. }
  548. is_dynamic_batch = true;
  549. }
  550. return SUCCESS;
  551. }
  552. bool CheckOpType(const NodePtr &node, const std::string type) {
  553. if (node->GetType() == type) {
  554. return true;
  555. }
  556. return false;
  557. }
  558. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  559. auto op = node_ptr->GetOpDesc();
  560. GE_CHECK_NOTNULL(op);
  561. auto inputDescsPtr = op->GetAllInputsDescPtr();
  562. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  563. ge::Format format = ge::FORMAT_ND;
  564. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  565. // process, otherwise fe will insert transdata.
  566. for (auto &inputDescPtr : inputDescsPtr) {
  567. GE_CHECK_NOTNULL(inputDescPtr);
  568. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  569. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  570. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  571. inputDescPtr->SetFormat(format);
  572. inputDescPtr->SetOriginFormat(format);
  573. }
  574. }
  575. for (auto &outputDescPtr : outputDescsPtr) {
  576. GE_CHECK_NOTNULL(outputDescPtr);
  577. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  578. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  579. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  580. outputDescPtr->SetFormat(format);
  581. outputDescPtr->SetOriginFormat(format);
  582. }
  583. }
  584. return SUCCESS;
  585. }
  586. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  587. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  588. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  589. // data and netoutput, and no longer inserts the transnode.
  590. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) {
  591. GE_CHECK_NOTNULL(node_ptr);
  592. auto op_desc = node_ptr->GetOpDesc();
  593. GE_CHECK_NOTNULL(op_desc);
  594. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  595. GE_CHECK_NOTNULL(input);
  596. ge::DataType src_dtype = input->GetDataType();
  597. if (src_dtype == dt_set) {
  598. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  599. return SUCCESS;
  600. }
  601. input->SetDataType(dt_set);
  602. int64_t input_shape_size = 0;
  603. int64_t output_shape_size = 0;
  604. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  605. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  606. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  607. GELOGE(GRAPH_FAILED, "GetTensorSize failed!");
  608. return FAILED;
  609. }
  610. ge::TensorUtils::SetSize(*input, input_shape_size);
  611. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  612. GE_CHECK_NOTNULL(output);
  613. output->SetDataType(dt_set);
  614. ge::TensorUtils::SetSize(*output, output_shape_size);
  615. if (is_dynamic_batch) {
  616. GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str());
  617. auto switchn_op_desc = switchn_node->GetOpDesc();
  618. GE_CHECK_NOTNULL(switchn_op_desc);
  619. auto switchn_input = switchn_op_desc->MutableInputDesc(0);
  620. GE_CHECK_NOTNULL(switchn_input);
  621. switchn_input->SetDataType(dt_set);
  622. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  623. const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i);
  624. GE_CHECK_NOTNULL(switchn_output);
  625. switchn_output->SetDataType(dt_set);
  626. }
  627. }
  628. return SUCCESS;
  629. }
  630. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  631. GE_CHECK_NOTNULL(node_ptr);
  632. auto op_desc = node_ptr->GetOpDesc();
  633. GE_CHECK_NOTNULL(op_desc);
  634. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  635. GE_CHECK_NOTNULL(input);
  636. ge::Format old_format = input->GetFormat();
  637. ge::GeShape old_shape = input->GetShape();
  638. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  639. if (!support) {
  640. ErrorManager::GetInstance().ATCReportErrMessage(
  641. "E19014", {"opname", "value", "reason"},
  642. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  643. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  644. GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str());
  645. return FAILED;
  646. }
  647. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  648. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  649. return FAILED;
  650. }
  651. if (is_dynamic_batch) {
  652. auto switchn_op_desc = switchn_node->GetOpDesc();
  653. GE_CHECK_NOTNULL(switchn_op_desc);
  654. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  655. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  656. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  657. return FAILED;
  658. }
  659. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  660. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  661. GE_CHECK_NOTNULL(switchn_output);
  662. old_format = switchn_output->GetFormat();
  663. old_shape = switchn_output->GetShape();
  664. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  665. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  666. return FAILED;
  667. }
  668. }
  669. }
  670. return SUCCESS;
  671. }
  672. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  673. auto op_desc = node_ptr->GetOpDesc();
  674. GE_CHECK_NOTNULL(op_desc);
  675. string set_dt_str;
  676. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  677. return SUCCESS;
  678. }
  679. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  680. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  681. bool is_dynamic_batch = false;
  682. NodePtr switchn_node = nullptr;
  683. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) {
  684. GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed");
  685. return FAILED;
  686. }
  687. if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) {
  688. GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed");
  689. return FAILED;
  690. }
  691. // check if need to set format
  692. string set_format;
  693. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  694. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  695. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  696. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) {
  697. GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed");
  698. return FAILED;
  699. }
  700. }
  701. return SUCCESS;
  702. }
  703. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr, Format &storage_format,
  704. vector<int64_t> &dst_shape_dims) {
  705. GE_CHECK_NOTNULL(op_desc);
  706. GE_CHECK_NOTNULL(tensor_desc_ptr);
  707. storage_format = FORMAT_RESERVED;
  708. int64_t format = FORMAT_RESERVED;
  709. dst_shape_dims.clear();
  710. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  711. storage_format = static_cast<Format>(format);
  712. vector<int32_t> storage_shape;
  713. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  714. for (auto dim : storage_shape) {
  715. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  716. }
  717. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  718. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  719. formats::JoinToString(storage_shape).c_str());
  720. } else {
  721. ErrorManager::GetInstance().ATCReportErrMessage(
  722. "15003", {"opname", "format"}, {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)});
  723. GELOGE(PARAM_INVALID,
  724. "Update node by storage format failed, storage_shape not set. "
  725. "node: [%s], storage_format [%s]",
  726. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  727. return FAILED;
  728. }
  729. ge::Format old_format = tensor_desc_ptr->GetFormat();
  730. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  731. if (old_format == storage_format && old_shape == dst_shape_dims) {
  732. GELOGI("Update node by storage format, not changed.");
  733. storage_format = FORMAT_RESERVED;
  734. return SUCCESS;
  735. }
  736. }
  737. return SUCCESS;
  738. }
  739. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  740. NodePtr &node) {
  741. bool is_dynamic = CheckOpType(node, MERGE);
  742. auto src_op_desc = node->GetOpDesc();
  743. GE_CHECK_NOTNULL(src_op_desc);
  744. ge::GeShape src_shape = src_desc.GetShape();
  745. ge::Format src_format = src_desc.GetFormat();
  746. net_output_input_desc->SetDataType(DT_FLOAT16);
  747. if (is_dynamic) {
  748. auto merge_output = src_op_desc->MutableOutputDesc(0);
  749. GE_CHECK_NOTNULL(merge_output);
  750. merge_output->SetDataType(DT_FLOAT16);
  751. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  752. auto merge_input = src_op_desc->MutableInputDesc(i);
  753. GE_CHECK_NOTNULL(merge_input);
  754. merge_input->SetDataType(DT_FLOAT16);
  755. }
  756. }
  757. std::vector<int64_t> dst_shape_dims;
  758. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  759. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  760. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  761. return FAILED;
  762. }
  763. ge::GeShape dst_shape(dst_shape_dims);
  764. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  765. net_output_input_desc->SetShape(dst_shape);
  766. if (is_dynamic) {
  767. auto merge_out = src_op_desc->MutableOutputDesc(0);
  768. GE_CHECK_NOTNULL(merge_out);
  769. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  770. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  771. return FAILED;
  772. }
  773. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  774. auto merge_in = src_op_desc->MutableInputDesc(i);
  775. GE_CHECK_NOTNULL(merge_in);
  776. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  777. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  778. return FAILED;
  779. }
  780. }
  781. }
  782. return SUCCESS;
  783. }
  784. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  785. GE_CHECK_NOTNULL(netout_desc);
  786. vector<string> output_dt_str;
  787. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  788. for (auto dt_str : output_dt_str) {
  789. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  790. if (dt_str_split.size() == kUserDefinedElementCount) {
  791. if (dt_str_split[0] == to_string(index)) {
  792. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  793. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  794. TypeUtils::DataTypeToSerialString(dt).c_str());
  795. return true;
  796. }
  797. }
  798. }
  799. }
  800. return false;
  801. }
  802. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  803. GE_CHECK_NOTNULL(netout_desc);
  804. vector<string> output_format_str;
  805. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  806. for (auto format_str : output_format_str) {
  807. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  808. if (format_str_split.size() == kUserDefinedElementCount) {
  809. if (format_str_split[0] == to_string(index)) {
  810. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  811. return true;
  812. }
  813. }
  814. }
  815. }
  816. return false;
  817. }
  818. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  819. auto op_desc = node->GetOpDesc();
  820. GE_CHECK_NOTNULL(op_desc);
  821. ge::DataType output_data_type = ge::DT_FLOAT;
  822. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  823. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  824. auto peer_out = in_anchor->GetPeerOutAnchor();
  825. GE_CHECK_NOTNULL(peer_out);
  826. auto src_node = peer_out->GetOwnerNode();
  827. GE_CHECK_NOTNULL(src_node);
  828. bool is_dynamic = CheckOpType(src_node, MERGE);
  829. OpDescPtr src_op_desc = src_node->GetOpDesc();
  830. GE_CHECK_NOTNULL(src_op_desc);
  831. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  832. GE_CHECK_NOTNULL(net_output_input_desc);
  833. ge::GeShape old_shape = net_output_input_desc->GetShape();
  834. ge::Format old_format = net_output_input_desc->GetFormat();
  835. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  836. // Update datatype
  837. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  838. GELOGI("Enter into process output_type schedule");
  839. net_output_input_desc->SetDataType(output_data_type);
  840. if (is_dynamic) {
  841. auto merge_output = src_op_desc->MutableOutputDesc(0);
  842. GE_CHECK_NOTNULL(merge_output);
  843. merge_output->SetDataType(output_data_type);
  844. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  845. auto merge_input = src_op_desc->MutableInputDesc(i);
  846. GE_CHECK_NOTNULL(merge_input);
  847. merge_input->SetDataType(output_data_type);
  848. }
  849. }
  850. }
  851. // check if is_output_adjust_hw_layout is set
  852. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  853. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  854. ErrorManager::GetInstance().ATCReportErrMessage(
  855. "E19014", {"opname", "value", "reason"},
  856. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  857. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  858. GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0.");
  859. return FAILED;
  860. }
  861. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  862. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  863. GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0.");
  864. return FAILED;
  865. }
  866. }
  867. }
  868. return SUCCESS;
  869. }
  870. } // namespace
  871. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  872. GraphPrepare::~GraphPrepare() {}
  873. /**
  874. * @param graph
  875. * @return
  876. */
  877. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  878. GE_CHECK_NOTNULL(graph);
  879. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  880. for (auto &node : graph->GetAllNodes()) {
  881. if (node == nullptr) {
  882. continue;
  883. }
  884. if (node->GetType() != VARIABLE) {
  885. continue;
  886. }
  887. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  888. if (trans_road == nullptr) {
  889. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  890. continue;
  891. }
  892. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  893. auto ret = RecoverTransRoadForVar(node, *trans_road);
  894. if (ret != SUCCESS) {
  895. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str());
  896. return INTERNAL_ERROR;
  897. }
  898. auto iter = var_names_to_refs.find(node->GetName());
  899. if (iter != var_names_to_refs.end()) {
  900. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  901. if (ret != SUCCESS) {
  902. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str());
  903. return INTERNAL_ERROR;
  904. }
  905. }
  906. }
  907. return SUCCESS;
  908. }
  909. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  910. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  911. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  912. if (compute_graph_ != nullptr) {
  913. compute_graph_->SetSessionID(session_id);
  914. }
  915. session_id_ = session_id;
  916. Status ret = CheckGraph();
  917. if (ret != SUCCESS) {
  918. GELOGE(ret, "RunGraph graph check fail, ret:%u", ret);
  919. return ret;
  920. }
  921. (void)compute_graph_->TopologicalSorting();
  922. ret = CheckRefOp();
  923. if (ret != SUCCESS) {
  924. GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret);
  925. return ret;
  926. }
  927. return SUCCESS;
  928. }
  929. Status GraphPrepare::CheckGraph() {
  930. if (compute_graph_ == nullptr) {
  931. GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR");
  932. return GE_GRAPH_INIT_FAILED;
  933. }
  934. auto nodes = compute_graph_->GetAllNodes();
  935. if (nodes.empty()) {
  936. GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph.");
  937. return GE_GRAPH_INIT_FAILED;
  938. }
  939. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  940. GE_CHECK_NOTNULL(node);
  941. if (node->GetOpDesc() == nullptr) {
  942. GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL");
  943. return GE_GRAPH_INIT_FAILED;
  944. }
  945. }
  946. return SUCCESS;
  947. }
  948. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  949. const std::set<NodePtr> &ref_nodes) {
  950. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  951. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  952. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  953. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  954. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  955. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH};
  956. GE_CHECK_NOTNULL(node);
  957. const auto &op_desc = node->GetOpDesc();
  958. GE_CHECK_NOTNULL(op_desc);
  959. const auto input_index = op_desc->GetInputIndexByName(input_name);
  960. const auto &in_anchor = node->GetInDataAnchor(input_index);
  961. GE_CHECK_NOTNULL(in_anchor);
  962. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  963. GE_CHECK_NOTNULL(peer_out_anchor);
  964. const auto &input_node = peer_out_anchor->GetOwnerNode();
  965. GE_CHECK_NOTNULL(input_node);
  966. const auto &input_op_desc = input_node->GetOpDesc();
  967. GE_CHECK_NOTNULL(input_op_desc);
  968. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  969. if (is_ref) {
  970. return SUCCESS;
  971. }
  972. auto input_type = input_op_desc->GetType();
  973. if (input_type == ge::FRAMEWORKOP) {
  974. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  975. GELOGE(PARAM_INVALID, "Get original type failed.");
  976. return PARAM_INVALID;
  977. }
  978. }
  979. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  980. if (!is_acceptable) {
  981. ErrorManager::GetInstance().ATCReportErrMessage(
  982. "E15005", {"opname", "optype", "opname1", "optype1"},
  983. {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()});
  984. GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.",
  985. node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  986. input_op_desc->GetType().c_str());
  987. return PARAM_INVALID;
  988. }
  989. return SUCCESS;
  990. }
  991. Status GraphPrepare::CheckRefOp() {
  992. GE_CHECK_NOTNULL(compute_graph_);
  993. std::set<NodePtr> ref_nodes;
  994. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  995. if (node == nullptr) {
  996. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  997. return PARAM_INVALID;
  998. }
  999. auto op_desc = node->GetOpDesc();
  1000. if (op_desc == nullptr) {
  1001. GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null.");
  1002. return PARAM_INVALID;
  1003. }
  1004. auto input_name_index = op_desc->GetAllInputName();
  1005. auto outputs = op_desc->GetAllOutputName();
  1006. for (const auto &name_index : input_name_index) {
  1007. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  1008. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  1009. GELOGE(PARAM_INVALID, "CheckRefInputNode failed.");
  1010. return PARAM_INVALID;
  1011. }
  1012. (void)ref_nodes.insert(node); // no need to check value
  1013. }
  1014. }
  1015. }
  1016. return SUCCESS;
  1017. };
  1018. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  1019. GE_CHECK_NOTNULL(compute_graph_);
  1020. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  1021. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  1022. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  1023. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  1024. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  1025. return SUCCESS;
  1026. }
  1027. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1028. if (node == nullptr) {
  1029. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL");
  1030. return GE_GRAPH_GRAPH_NODE_NULL;
  1031. }
  1032. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1033. if (op_desc_ptr == nullptr) {
  1034. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL");
  1035. return GE_GRAPH_GRAPH_NODE_NULL;
  1036. }
  1037. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1038. int64_t tensor_size = 0;
  1039. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1040. if (graph_status != GRAPH_SUCCESS) {
  1041. ErrorManager::GetInstance().ATCReportErrMessage("E19012", {"function", "reason"},
  1042. {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1043. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  1044. return FAILED;
  1045. }
  1046. TensorUtils::SetSize(output, tensor_size);
  1047. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1048. if (graph_ret != GRAPH_SUCCESS) {
  1049. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1050. return graph_ret;
  1051. }
  1052. return SUCCESS;
  1053. }
  1054. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input) {
  1055. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1056. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1057. GE_CHECK_NOTNULL(input_node);
  1058. OpDescPtr op = input_node->GetOpDesc();
  1059. GE_CHECK_NOTNULL(op);
  1060. if (op->GetType() == DATA) {
  1061. GeAttrValue::INT index = 0;
  1062. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1063. GELOGW("Get index from data attr failed");
  1064. continue;
  1065. }
  1066. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1067. std::string situation = "data op index[" + std::to_string(index) + "]";
  1068. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1069. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1070. GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index);
  1071. return FAILED;
  1072. }
  1073. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1074. auto format = desc.GetFormat();
  1075. auto origin_format = desc.GetOriginFormat();
  1076. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1077. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op);
  1078. if (need_check_internal_format) {
  1079. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1080. if (is_internal) {
  1081. ErrorManager::GetInstance().ATCReportErrMessage(
  1082. "E19025", {"situation", "reason"},
  1083. {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" +
  1084. TypeUtils::FormatToSerialString(origin_format) + "]",
  1085. "it is not support"});
  1086. GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.",
  1087. TypeUtils::FormatToSerialString(format).c_str(),
  1088. TypeUtils::FormatToSerialString(origin_format).c_str());
  1089. return FAILED;
  1090. }
  1091. }
  1092. auto data_type = desc.GetDataType();
  1093. uint32_t length = 1;
  1094. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1095. if (!type_ret) {
  1096. ErrorManager::GetInstance().ATCReportErrMessage(
  1097. "E19025", {"situation", "reason"},
  1098. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1099. GELOGE(PARAM_INVALID, "Input datatype %s is not support.",
  1100. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1101. return FAILED;
  1102. }
  1103. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1104. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1105. int64_t shape_size = desc_shape * length;
  1106. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1107. int64_t size = 0;
  1108. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1109. GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed");
  1110. return FAILED);
  1111. bool size_check = (size != 0 && shape_size != size);
  1112. if (size_check) {
  1113. std::string situation =
  1114. "input data size[" + std::to_string(size) + "] and shape_size[" + std::to_string(size) + "]";
  1115. std::string reason = "because size != 0 and shape_size != size";
  1116. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1117. GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size);
  1118. return FAILED;
  1119. }
  1120. ge::TensorUtils::SetSize(desc, shape_size);
  1121. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1122. if (graph_ret != GRAPH_SUCCESS) {
  1123. GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
  1124. return graph_ret;
  1125. }
  1126. // Size will be recalculated in the build stage
  1127. ge::TensorUtils::SetSize(desc, 0);
  1128. graph_ret = op->UpdateOutputDesc(0, desc);
  1129. if (graph_ret != GRAPH_SUCCESS) {
  1130. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1131. return graph_ret;
  1132. }
  1133. if (!options_.train_graph_flag) {
  1134. Status ret = AdjustDataOpOutput(input_node);
  1135. if (ret != SUCCESS) {
  1136. GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret);
  1137. return ret;
  1138. }
  1139. }
  1140. }
  1141. }
  1142. return SUCCESS;
  1143. }
  1144. Status GraphPrepare::TryDoAipp() {
  1145. // infer and with aipp configure file, then call aipp insert
  1146. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1147. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1148. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1149. if (ret != SUCCESS) {
  1150. GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed.");
  1151. return INTERNAL_ERROR;
  1152. }
  1153. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1154. if (ret != SUCCESS) {
  1155. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed",
  1156. options_.insert_op_file.c_str());
  1157. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1158. }
  1159. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1160. if (ret != SUCCESS) {
  1161. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret);
  1162. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1163. }
  1164. }
  1165. return SUCCESS;
  1166. }
  1167. Status GraphPrepare::FormatAndShapeProcess() {
  1168. Status ret = ResourcePairProcess("add");
  1169. if (ret != SUCCESS) {
  1170. GELOGE(ret, "ResourcePairProcess failed");
  1171. return ret;
  1172. }
  1173. GE_TIMESTAMP_START(InferOriginFormat1);
  1174. ret = compute_graph_->InferOriginFormat();
  1175. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1176. GE_DUMP(compute_graph_, "after_first_inferformat");
  1177. if (ret != SUCCESS) {
  1178. GELOGE(ret, "Prepare Graph first inferformat failed");
  1179. return ret;
  1180. }
  1181. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1182. ret = InferShapeForPreprocess();
  1183. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1184. GE_DUMP(compute_graph_, "after_infershape");
  1185. if (ret != SUCCESS) {
  1186. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed");
  1187. return GE_GRAPH_INFERSHAPE_FAILED;
  1188. }
  1189. GE_TIMESTAMP_START(InferOriginFormat2);
  1190. ret = compute_graph_->InferOriginFormat();
  1191. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1192. if (ret != SUCCESS) {
  1193. GELOGE(ret, "Prepare Graph inferformat failed");
  1194. return ret;
  1195. }
  1196. ret = ResourcePairProcess("remove");
  1197. if (ret != SUCCESS) {
  1198. return ret;
  1199. }
  1200. return ret;
  1201. }
  1202. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1203. PassManager control_pass;
  1204. // Graph pass tmp logic for resource infershape
  1205. if (options_.train_graph_flag) {
  1206. try {
  1207. if (action == "add") {
  1208. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1209. } else {
  1210. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1211. new ResourcePairRemoveControlPass);
  1212. }
  1213. } catch (std::bad_alloc &e) {
  1214. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str());
  1215. return INTERNAL_ERROR;
  1216. }
  1217. }
  1218. Status ret = control_pass.Run(compute_graph_);
  1219. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1220. GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret);
  1221. return ret;
  1222. }
  1223. return SUCCESS;
  1224. }
  1225. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1226. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1227. GE_CHECK_NOTNULL(node_ptr);
  1228. if (node_ptr->GetType() == DATA) {
  1229. uint32_t index = 0;
  1230. auto op_desc = node_ptr->GetOpDesc();
  1231. GE_CHECK_NOTNULL(op_desc);
  1232. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1233. Format storage_format = FORMAT_RESERVED;
  1234. vector<int64_t> dst_shape_dims;
  1235. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1236. GELOGE(INTERNAL_ERROR, "Get storage format for input failed");
  1237. return FAILED;
  1238. }
  1239. if (storage_format == FORMAT_RESERVED) {
  1240. continue;
  1241. }
  1242. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1243. GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed");
  1244. return FAILED;
  1245. }
  1246. }
  1247. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1248. auto op_desc = node_ptr->GetOpDesc();
  1249. GE_CHECK_NOTNULL(op_desc);
  1250. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1251. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1252. Format storage_format = FORMAT_RESERVED;
  1253. vector<int64_t> dst_shape_dims;
  1254. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1255. GELOGE(INTERNAL_ERROR, "Get storage format from output failed");
  1256. return FAILED;
  1257. }
  1258. if (storage_format == FORMAT_RESERVED) {
  1259. continue;
  1260. }
  1261. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1262. GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed");
  1263. return FAILED;
  1264. }
  1265. }
  1266. }
  1267. }
  1268. return SUCCESS;
  1269. }
  1270. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1271. if (options_.save_original_model == "true") {
  1272. ModelHelper model_helper;
  1273. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1274. options_.original_model_file);
  1275. if (ret != SUCCESS) {
  1276. // If save original model fail, process continue
  1277. GELOGW("SaveOriginalGraphToOmModel fail");
  1278. }
  1279. }
  1280. return SUCCESS;
  1281. }
  1282. #define PP_RUN_AND_DUMP(name, func, ...) \
  1283. do { \
  1284. GE_RUN(Prepare, func, __VA_ARGS__); \
  1285. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1286. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1287. } while (0)
  1288. #define PP_RUN(name, func, ...) \
  1289. do { \
  1290. GE_RUN(Prepare, func, __VA_ARGS__); \
  1291. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1292. } while (0)
  1293. Status GraphPrepare::PrepareDynShape(ConstGraphPtr graph, const std::vector<GeTensor> &user_input,
  1294. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1295. GE_CHECK_NOTNULL(graph);
  1296. GE_CHECK_NOTNULL(compute_graph);
  1297. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1298. const Graph &const_graph = *graph;
  1299. PP_RUN("Init", Init, const_graph, session_id);
  1300. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1301. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input);
  1302. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1303. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1304. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1305. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1306. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1307. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1308. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1309. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1310. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1311. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1312. return SUCCESS;
  1313. }
  1314. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1315. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1316. return SUCCESS;
  1317. }
  1318. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1319. auto compute_graph = compute_graph_;
  1320. PassManager pass_manager;
  1321. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1322. new (std::nothrow) VariablePrepareOpPass))
  1323. GE_TIMESTAMP_START(pass_manager);
  1324. auto ret = pass_manager.Run(compute_graph);
  1325. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1326. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1327. GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret);
  1328. return ret;
  1329. }
  1330. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1331. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1332. return SUCCESS;
  1333. }
  1334. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1335. if (compute_graph == nullptr) {
  1336. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1337. return GE_GRAPH_NULL_INPUT;
  1338. }
  1339. GEPass ge_passes(compute_graph);
  1340. NamesToPass hccl_group;
  1341. HcclGroupPass hccl_group_pass;
  1342. GELOGD("Add hccl group pass success");
  1343. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1344. auto ret = ge_passes.Run(hccl_group);
  1345. if (ret != SUCCESS) {
  1346. GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret);
  1347. return ret;
  1348. }
  1349. ret = compute_graph->TopologicalSorting();
  1350. if (ret != SUCCESS) {
  1351. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1352. return ret;
  1353. }
  1354. return SUCCESS;
  1355. }
  1356. #undef PP_RUN_AND_DUMP
  1357. #undef PP_RUN
  1358. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1359. if (graph == nullptr) {
  1360. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1361. return GE_GRAPH_NULL_INPUT;
  1362. }
  1363. const Graph &const_graph = *graph;
  1364. Status ret = Init(const_graph, 0);
  1365. if (ret != SUCCESS) {
  1366. GELOGE(ret, "Init graph_prepare fail, ret:%u", ret);
  1367. return ret;
  1368. }
  1369. GE_DUMP(compute_graph_, "after_parser");
  1370. GELOGI("Start infershape for dump json process.");
  1371. ret = compute_graph_->InferOriginFormat();
  1372. GE_DUMP(compute_graph_, "after_inferformat");
  1373. if (ret != SUCCESS) {
  1374. GELOGE(ret, "Prepare Graph inferformat failed");
  1375. return ret;
  1376. }
  1377. InferShapePass infer_shape_pass;
  1378. NamesToPass names_to_passes;
  1379. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1380. GEPass ge_passes(compute_graph_);
  1381. ret = ge_passes.Run(names_to_passes);
  1382. GE_DUMP(compute_graph_, "after_infershape");
  1383. if (ret != SUCCESS) {
  1384. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1385. return ret;
  1386. }
  1387. ShapeRefiner::ClearContextMap();
  1388. return SUCCESS;
  1389. }
  1390. Status GraphPrepare::CheckConstOp() {
  1391. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1392. GE_CHECK_NOTNULL(node_ptr);
  1393. if (node_ptr->GetType() == CONSTANT) {
  1394. Status ret = VerifyConstOp(node_ptr);
  1395. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1396. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1397. auto op_desc = node_ptr->GetOpDesc();
  1398. if (op_desc == nullptr) {
  1399. GELOGE(PARAM_INVALID, "Get op desc failed");
  1400. return PARAM_INVALID;
  1401. }
  1402. std::string original_type;
  1403. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1404. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1405. GELOGI("original type is %s", original_type.c_str());
  1406. if (original_type == CONSTANT) {
  1407. Status ret = VerifyConstOp(node_ptr);
  1408. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1409. }
  1410. }
  1411. }
  1412. return SUCCESS;
  1413. }
  1414. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1415. GE_CHECK_NOTNULL(node);
  1416. auto op_desc = node->GetOpDesc();
  1417. GE_CHECK_NOTNULL(op_desc);
  1418. ConstGeTensorPtr ge_tensor_ptr;
  1419. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1420. GELOGE(PARAM_INVALID, "Get value from const attr failed");
  1421. return PARAM_INVALID;
  1422. }
  1423. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1424. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1425. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1426. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1427. auto data_type = ge_tensor_desc.GetDataType();
  1428. uint32_t length = 1;
  1429. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1430. if (!type_ret) {
  1431. ErrorManager::GetInstance().ATCReportErrMessage(
  1432. "E19025", {"situation", "reason"},
  1433. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1434. GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str());
  1435. return FAILED;
  1436. }
  1437. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1438. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1439. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1440. if (shape_size == 0) {
  1441. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1442. // shape = [], means it's a sclar tensor.
  1443. GE_CHK_BOOL_EXEC(data_size / length == 1, ErrorManager::GetInstance().ATCReportErrMessage(
  1444. "E10043", {"reason"}, {"Const is invalid scalar tensor."});
  1445. return PARAM_INVALID, "Const is invalid scalar tensor.");
  1446. } else {
  1447. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1448. GE_CHK_BOOL_EXEC(data_size == 0, ErrorManager::GetInstance().ATCReportErrMessage(
  1449. "E10043", {"reason"}, {"Const is invalid vector scalar."});
  1450. return PARAM_INVALID, "Const is invalid vector scalar.");
  1451. }
  1452. } else {
  1453. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0,
  1454. ErrorManager::GetInstance().ATCReportErrMessage(
  1455. "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"});
  1456. return PARAM_INVALID, "Const input data size is not equal with tensor desc shape");
  1457. }
  1458. return SUCCESS;
  1459. }
  1460. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1461. if (GetLocalOmgContext().is_dynamic_input) {
  1462. return SUCCESS;
  1463. }
  1464. unsigned int node_num = 0;
  1465. unsigned int data_num = 0;
  1466. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1467. GE_CHECK_NOTNULL(input_node);
  1468. OpDescPtr op = input_node->GetOpDesc();
  1469. GE_CHECK_NOTNULL(op);
  1470. node_num++;
  1471. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1472. data_num++;
  1473. GeAttrValue::INT index = 0;
  1474. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1475. GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed");
  1476. return GE_GRAPH_INIT_FAILED;
  1477. }
  1478. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1479. std::string situation = "data op index[" + std::to_string(index) + "]";
  1480. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1481. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1482. GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index);
  1483. return GE_GRAPH_INIT_FAILED;
  1484. }
  1485. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1486. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1487. if (desc.GetShape().GetDim(i) < 0) {
  1488. std::string situation =
  1489. "data dim[" + std::to_string(i) + "][" + std::to_string(desc.GetShape().GetDim(i)) + "]";
  1490. std::string reason = "it need >= 0";
  1491. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1492. GELOGE(GE_GRAPH_INIT_FAILED, "data dim %zu is not supported, need >= 0, real:%ld.", i,
  1493. desc.GetShape().GetDim(i));
  1494. return GE_GRAPH_INIT_FAILED;
  1495. }
  1496. }
  1497. }
  1498. }
  1499. if (node_num <= data_num) {
  1500. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1501. }
  1502. return SUCCESS;
  1503. }
  1504. Status GraphPrepare::InferShapeForPreprocess() {
  1505. GELOGI("Start infershape for preprocess.");
  1506. GEPass ge_passes(compute_graph_);
  1507. NamesToPass names_to_passes;
  1508. AssertPass assert_pass;
  1509. if (!options_.train_graph_flag) {
  1510. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1511. }
  1512. InferShapePass infer_shape_pass;
  1513. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1514. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1515. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1516. DimensionComputePass dimension_compute_pass;
  1517. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1518. ConstantFoldingPass constant_folding_pass;
  1519. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1520. int32_t dev_count = 0;
  1521. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1522. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1523. rtError_t rt_err = RT_ERROR_NONE;
  1524. if (aicpu_constant_folding_on != nullptr) {
  1525. rt_err = rtGetDeviceCount(&dev_count);
  1526. if (rt_err == RT_ERROR_NONE) {
  1527. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1528. if (result != SUCCESS) {
  1529. GELOGE(result, "Set rt context failed.");
  1530. return result;
  1531. }
  1532. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1533. }
  1534. }
  1535. Status ret = ge_passes.Run(names_to_passes);
  1536. if (aicpu_constant_folding_on != nullptr) {
  1537. if (rt_err == RT_ERROR_NONE) {
  1538. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1539. if (result != SUCCESS) {
  1540. GELOGE(result, "Set rt context failed.");
  1541. return result;
  1542. }
  1543. }
  1544. }
  1545. ShapeRefiner::ClearContextMap();
  1546. if (ret != SUCCESS) {
  1547. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1548. return ret;
  1549. }
  1550. return SUCCESS;
  1551. }
  1552. Status GraphPrepare::PrepareOptimize() {
  1553. GELOGI("Start optimize for preprocess.");
  1554. // check rw type
  1555. GraphOptimize graph_optimize;
  1556. bool has_conflict = false;
  1557. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1558. if (has_conflict) {
  1559. GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize.");
  1560. return FAILED;
  1561. }
  1562. PassManager original_graph_passes;
  1563. // Graph pass
  1564. try {
  1565. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1566. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1567. } catch (std::bad_alloc &e) {
  1568. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1569. return INTERNAL_ERROR;
  1570. }
  1571. GE_TIMESTAMP_START(original_graph_passes);
  1572. Status ret = original_graph_passes.Run(compute_graph_);
  1573. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1574. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1575. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1576. return ret;
  1577. }
  1578. // New pass
  1579. GEPass ge_passes(compute_graph_);
  1580. NamesToPass names_to_passes;
  1581. EnterPass enter_pass;
  1582. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1583. CondPass cond_pass;
  1584. names_to_passes.emplace_back("CondPass", &cond_pass);
  1585. PrintOpPass print_pass;
  1586. if (options_.enable_print_op_pass) {
  1587. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1588. }
  1589. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1590. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1591. DropOutPass dropout_pass;
  1592. AssertPass assert_pass;
  1593. UnusedConstPass unused_const_pass;
  1594. StopGradientPass stop_gradient_pass;
  1595. PreventGradientPass prevent_gradient_pass;
  1596. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1597. GuaranteeConstPass guarantee_const_pass;
  1598. VarIsInitializedOpPass var_is_initialized_pass;
  1599. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1600. IdentityPass identity_pass(false);
  1601. AssignPass assign_pass;
  1602. SnapshotPass snapshot_pass;
  1603. if (!options_.train_graph_flag) {
  1604. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1605. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1606. }
  1607. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1608. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1609. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1610. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1611. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1612. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1613. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1614. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1615. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1616. if (GetContext().GetHostExecFlag()) {
  1617. names_to_passes.emplace_back("AssignPass", &assign_pass);
  1618. }
  1619. GE_TIMESTAMP_START(names_to_passes);
  1620. ret = ge_passes.Run(names_to_passes);
  1621. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1622. if (ret != SUCCESS) {
  1623. GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret);
  1624. return ret;
  1625. }
  1626. PassManager graph_pass;
  1627. try {
  1628. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1629. // todo 临时把hccl的memcpy插入放到图准备,为了防止其多插memcpy
  1630. (void)graph_pass.AddPass("PrepareOptimize::HcclMemcpyPass", new (std::nothrow) HcclMemcpyPass);
  1631. } catch (std::bad_alloc &e) {
  1632. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1633. return INTERNAL_ERROR;
  1634. }
  1635. GE_TIMESTAMP_START(graph_passes);
  1636. ret = graph_pass.Run(compute_graph_);
  1637. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1638. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1639. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1640. return ret;
  1641. }
  1642. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1643. TypeConversionOfConstant();
  1644. ret = compute_graph_->TopologicalSorting();
  1645. if (ret != SUCCESS) {
  1646. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1647. return ret;
  1648. }
  1649. GELOGI("End optimize for preprocess.");
  1650. return SUCCESS;
  1651. }
  1652. void GraphPrepare::TypeConversionOfConstant() {
  1653. if (options_.train_graph_flag) {
  1654. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1655. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1656. // This can ensure that n is not a null pointer
  1657. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1658. n->GetOpDesc()->SetType(CONSTANTOP);
  1659. }
  1660. }
  1661. } else {
  1662. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  1663. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1664. // This can ensure that n is not a null pointer
  1665. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  1666. n->GetOpDesc()->SetType(CONSTANT);
  1667. }
  1668. }
  1669. }
  1670. }
  1671. Status GraphPrepare::GraphEquivalentTransformation() {
  1672. NamesToPass names_to_pass;
  1673. ForPass for_pass;
  1674. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  1675. return GEPass(compute_graph_).Run(names_to_pass);
  1676. }
  1677. Status GraphPrepare::ProcessBeforeInfershape() {
  1678. NamesToPass names_to_passes;
  1679. CondRemovePass condition_remove_pass;
  1680. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  1681. GE_TIMESTAMP_START(ProcessCondRemove);
  1682. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  1683. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  1684. if (ret != SUCCESS) {
  1685. GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  1686. return ret;
  1687. }
  1688. return SUCCESS;
  1689. }
  1690. Status GraphPrepare::ProcessNetOutput() {
  1691. PassManager graph_passes_before_infershape;
  1692. try {
  1693. if (options_.train_graph_flag) {
  1694. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  1695. }
  1696. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  1697. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  1698. new (std::nothrow) DataPass); // Add NetOutput first.
  1699. } catch (std::bad_alloc) {
  1700. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1701. return INTERNAL_ERROR;
  1702. }
  1703. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  1704. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  1705. GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret);
  1706. return ret;
  1707. }
  1708. return SUCCESS;
  1709. }
  1710. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input) {
  1711. compute_graph_->SetInputSize(user_input.size());
  1712. if (user_input.empty()) {
  1713. return SUCCESS;
  1714. }
  1715. auto ret = CheckUserInput(user_input);
  1716. if (ret != SUCCESS) {
  1717. GELOGE(ret, "Check user input failed.");
  1718. return ret;
  1719. }
  1720. ret = UpdateInput(user_input);
  1721. if (ret != SUCCESS) {
  1722. GELOGE(ret, "UpdateInput fail, ret:%u", ret);
  1723. return ret;
  1724. }
  1725. if (user_input.size() != 0) {
  1726. ret = CheckConstOp();
  1727. if (ret != SUCCESS) {
  1728. GELOGE(ret, "CheckConstOp fail, ret:%u", ret);
  1729. return ret;
  1730. }
  1731. } else {
  1732. ret = compute_graph_->TopologicalSorting();
  1733. if (ret != SUCCESS) {
  1734. GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting");
  1735. return FAILED;
  1736. }
  1737. }
  1738. return SUCCESS;
  1739. }
  1740. Status GraphPrepare::UpdateInputOutputByOptions() {
  1741. auto ret = UpdateDataNetOutputByStorageFormat();
  1742. if (ret != SUCCESS) {
  1743. GELOGE(ret, "Update format acoording to storage format failed.");
  1744. return ret;
  1745. }
  1746. if (options_.train_graph_flag) {
  1747. GELOGI("This is train mode, no need to do this schedule.");
  1748. return SUCCESS;
  1749. }
  1750. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  1751. GE_CHECK_NOTNULL(node_ptr);
  1752. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  1753. GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str());
  1754. return FAILED;
  1755. }
  1756. if (node_ptr->GetType() == DATA) {
  1757. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  1758. GELOGE(INTERNAL_ERROR, "Process data node failed");
  1759. return FAILED;
  1760. }
  1761. }
  1762. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1763. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  1764. GELOGE(INTERNAL_ERROR, "Process netoutput node failed");
  1765. return FAILED;
  1766. }
  1767. }
  1768. }
  1769. return SUCCESS;
  1770. }
  1771. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  1772. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  1773. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  1774. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  1775. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  1776. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  1777. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  1778. if (dst_node->GetType() == TRANSDATA) {
  1779. return true;
  1780. }
  1781. }
  1782. }
  1783. return false;
  1784. }
  1785. } // 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