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 86 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
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #include "graph/preprocess/graph_preprocess.h"
  17. #include <map>
  18. #include <set>
  19. #include <string>
  20. #include "common/formats/format_transfers/format_transfer_fractal_nz.h"
  21. #include "common/formats/format_transfers/format_transfer_nchw_nc1hwc0.h"
  22. #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h"
  23. #include "common/formats/format_transfers/format_transfer_transpose.h"
  24. #include "common/formats/utils/formats_trans_utils.h"
  25. #include "common/helper/model_helper.h"
  26. #include "common/math/math_util.h"
  27. #include "common/op/ge_op_utils.h"
  28. #include "graph/common/ge_call_wrapper.h"
  29. #include "graph/common/local_context.h"
  30. #include "graph/common/transop_util.h"
  31. #include "graph/ge_context.h"
  32. #include "graph/shape_refiner.h"
  33. #include "graph/manager/graph_var_manager.h"
  34. #include "graph/manager/util/rt_context_util.h"
  35. #include "graph/optimize/graph_optimize.h"
  36. #include "graph/passes/addn_pass.h"
  37. #include "graph/passes/aicpu_constant_folding_pass.h"
  38. #include "graph/passes/assert_pass.h"
  39. #include "ge/ge_api_types.h"
  40. #include "graph/passes/common_subexpression_elimination_pass.h"
  41. #include "graph/passes/cond_pass.h"
  42. #include "graph/passes/cond_remove_pass.h"
  43. #include "graph/passes/constant_folding_pass.h"
  44. #include "graph/passes/dimension_adjust_pass.h"
  45. #include "graph/passes/dimension_compute_pass.h"
  46. #include "graph/passes/dropout_pass.h"
  47. #include "graph/passes/enter_pass.h"
  48. #include "graph/passes/for_pass.h"
  49. #include "graph/passes/guarantee_const_pass.h"
  50. #include "graph/passes/hccl_group_pass.h"
  51. #include "graph/passes/identity_pass.h"
  52. #include "graph/passes/infershape_pass.h"
  53. #include "graph/passes/merge_pass.h"
  54. #include "graph/passes/net_output_pass.h"
  55. #include "graph/passes/no_use_reshape_remove_pass.h"
  56. #include "graph/passes/parallel_concat_start_op_pass.h"
  57. #include "graph/passes/placeholder_with_default_pass.h"
  58. #include "graph/passes/prevent_gradient_pass.h"
  59. #include "graph/passes/print_op_pass.h"
  60. #include "graph/passes/prune_pass.h"
  61. #include "graph/passes/replace_transshape_pass.h"
  62. #include "graph/passes/replace_with_empty_const_pass.h"
  63. #include "graph/passes/resource_pair_add_control_pass.h"
  64. #include "graph/passes/resource_pair_remove_control_pass.h"
  65. #include "graph/passes/save_pass.h"
  66. #include "graph/passes/shape_operate_op_remove_pass.h"
  67. #include "graph/passes/snapshot_pass.h"
  68. #include "graph/passes/stop_gradient_pass.h"
  69. #include "graph/passes/switch_dead_branch_elimination.h"
  70. #include "graph/passes/unused_const_pass.h"
  71. #include "graph/passes/var_is_initialized_op_pass.h"
  72. #include "graph/passes/variable_prepare_op_pass.h"
  73. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  74. #include "graph/utils/type_utils.h"
  75. #include "inc/pass_manager.h"
  76. #include "init/gelib.h"
  77. #include "multi_batch_copy_graph.h"
  78. #include "graph/passes/data_pass.h"
  79. #include "graph/passes/mark_agnostic_pass.h"
  80. namespace ge {
  81. namespace {
  82. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  83. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  84. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  85. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  86. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  87. // the size of user defined output datatype or format string after split by ":".
  88. const size_t kUserDefinedElementCount = 2;
  89. const int kDataOutIndex = 0;
  90. const int64_t kInvalidDynaimcDimsType = -1;
  91. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  92. GeTensorPtr tensor = MakeShared<GeTensor>();
  93. if (tensor == nullptr) {
  94. GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed");
  95. return nullptr;
  96. }
  97. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  98. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  99. auto dst_ge_shape = data_tensor.GetShape();
  100. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  101. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  102. tensor->MutableTensorDesc().SetShape(GeShape());
  103. int32_t dst_shape = 1;
  104. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  105. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  106. return nullptr;
  107. }
  108. } else {
  109. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  110. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  111. if (dst_shape == nullptr) {
  112. GELOGE(INTERNAL_ERROR, "Create unique ptr failed");
  113. return nullptr;
  114. }
  115. for (int64_t i = 0; i < dim_cnt; ++i) {
  116. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  117. }
  118. GE_IF_BOOL_EXEC(
  119. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  120. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  121. return nullptr;)
  122. }
  123. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  124. return OpDescUtils::CreateConstOp(tensor);
  125. }
  126. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  127. OpDescPtr &op_desc) {
  128. // For format transfer node, the IR definition has src/dst format attrs
  129. if (node_type == TRANSDATA) {
  130. GE_IF_BOOL_EXEC(
  131. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  132. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  133. GE_IF_BOOL_EXEC(
  134. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  135. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  136. }
  137. // For TransposeD node, the IR definition has perm attrs
  138. if (node_type == TRANSPOSED) {
  139. Format src_format = input.GetFormat();
  140. Format dst_format = output.GetFormat();
  141. std::vector<int64_t> perm_arg;
  142. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  143. "Get perm by foramt failed.");
  144. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  145. "SetStr PERMUTE_ATTR_PERM failed")
  146. }
  147. // For cast node, the IR definition has src/dst attrs
  148. if (node_type == CAST) {
  149. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  150. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  151. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  152. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  153. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  154. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  155. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  156. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  157. }
  158. }
  159. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  160. const GeTensorDesc &output, NodePtr &node) {
  161. if (node == nullptr) {
  162. GELOGE(PARAM_INVALID, "node is null.");
  163. return nullptr;
  164. }
  165. auto graph = node->GetOwnerComputeGraph();
  166. if (graph == nullptr) {
  167. GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str());
  168. return nullptr;
  169. }
  170. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  171. if (index < 0) {
  172. ErrorManager::GetInstance().ATCReportErrMessage(
  173. "E19025", {"situation", "reason"},
  174. {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()});
  175. GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str());
  176. return nullptr;
  177. }
  178. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  179. if (op_desc == nullptr) {
  180. GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed");
  181. return nullptr;
  182. }
  183. // for data dump
  184. GE_IF_BOOL_EXEC(
  185. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  186. GELOGW("CreateTransNode: SetListStr failed");)
  187. // Default single input and single output
  188. auto ret = op_desc->AddInputDesc(input);
  189. if (ret != GRAPH_SUCCESS) {
  190. GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str());
  191. return nullptr;
  192. }
  193. ret = op_desc->AddOutputDesc(output);
  194. if (ret != GRAPH_SUCCESS) {
  195. GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str());
  196. return nullptr;
  197. }
  198. AddTransNodeAttr(node_type, input, output, op_desc);
  199. NodePtr shape_node = nullptr;
  200. if (node_type == RESHAPE) {
  201. auto shape_desc = CreateTensorShape(output);
  202. if (shape_desc == nullptr) {
  203. GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input",
  204. node->GetName().c_str());
  205. return nullptr;
  206. }
  207. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  208. if (ret != GRAPH_SUCCESS) {
  209. GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str());
  210. return nullptr;
  211. }
  212. shape_node = graph->AddNode(shape_desc);
  213. if (shape_node == nullptr) {
  214. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str());
  215. return nullptr;
  216. }
  217. }
  218. auto trans_node = graph->AddNode(op_desc);
  219. if (trans_node == nullptr) {
  220. GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str());
  221. return nullptr;
  222. }
  223. if (node_type == RESHAPE) {
  224. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  225. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str());
  226. return nullptr;
  227. }
  228. }
  229. return trans_node;
  230. }
  231. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  232. NodePtr &trans_node) {
  233. GE_CHECK_NOTNULL(node);
  234. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  235. if (trans_node == nullptr) {
  236. return INTERNAL_ERROR;
  237. }
  238. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  239. if (ret != GRAPH_SUCCESS) {
  240. GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s",
  241. node->GetName().c_str(), node->GetType().c_str());
  242. return INTERNAL_ERROR;
  243. }
  244. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  245. if (ret != GRAPH_SUCCESS) {
  246. GELOGE(INTERNAL_ERROR, "Failed to connect node %s to trans node %s", node->GetName().c_str(),
  247. trans_node->GetName().c_str());
  248. return INTERNAL_ERROR;
  249. }
  250. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  251. if (ret != GRAPH_SUCCESS) {
  252. GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.",
  253. node->GetName().c_str(), trans_node->GetName().c_str());
  254. return INTERNAL_ERROR;
  255. }
  256. return SUCCESS;
  257. }
  258. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  259. NodePtr &trans_node) {
  260. GE_CHECK_NOTNULL(node);
  261. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  262. if (trans_node == nullptr) {
  263. return INTERNAL_ERROR;
  264. }
  265. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  266. if (ret != GRAPH_SUCCESS) {
  267. GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s",
  268. node->GetName().c_str(), node->GetType().c_str());
  269. return INTERNAL_ERROR;
  270. }
  271. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  272. if (ret != GRAPH_SUCCESS) {
  273. GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(),
  274. node->GetName().c_str());
  275. return INTERNAL_ERROR;
  276. }
  277. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  278. if (ret != GRAPH_SUCCESS) {
  279. GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.",
  280. node->GetName().c_str(), trans_node->GetName().c_str());
  281. return INTERNAL_ERROR;
  282. }
  283. return SUCCESS;
  284. }
  285. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  286. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  287. GE_CHECK_NOTNULL(var->GetOpDesc());
  288. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  289. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  290. output_desc.SetFormat(tensor_desc.GetFormat());
  291. output_desc.SetDataType(tensor_desc.GetDataType());
  292. output_desc.SetShape(tensor_desc.GetShape());
  293. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  294. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  295. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  296. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  297. GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed");
  298. return INTERNAL_ERROR;);
  299. }
  300. if (var->GetOpDesc()->GetInputsSize() > 0) {
  301. auto desc = var->GetOpDesc()->GetInputDesc(0);
  302. desc.SetFormat(tensor_desc.GetFormat());
  303. desc.SetDataType(tensor_desc.GetDataType());
  304. desc.SetShape(tensor_desc.GetShape());
  305. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  306. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  307. desc.SetOriginShape(tensor_desc.GetOriginShape());
  308. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  309. GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed");
  310. return INTERNAL_ERROR;)
  311. }
  312. return SUCCESS;
  313. }
  314. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  315. GE_CHECK_NOTNULL(var);
  316. int index = 0;
  317. NodePtr last_node = var;
  318. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  319. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  320. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  321. if (ret != SUCCESS) {
  322. ErrorManager::GetInstance().ATCReportErrMessage(
  323. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  324. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(),
  325. index, iter->node_type.c_str());
  326. return INTERNAL_ERROR;
  327. }
  328. // set stream_label
  329. OpDescPtr var_desc = var->GetOpDesc();
  330. GE_CHECK_NOTNULL(var_desc);
  331. std::string stream_label;
  332. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  333. if (!stream_label.empty()) {
  334. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  335. }
  336. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  337. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  338. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  339. }
  340. if (road.empty()) {
  341. return SUCCESS;
  342. }
  343. return UpdateVarFormats(var, road.rbegin()->output);
  344. }
  345. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  346. for (auto &var : nodes) {
  347. GE_CHECK_NOTNULL(var);
  348. int index = 0;
  349. NodePtr last_node = var;
  350. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  351. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  352. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  353. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  354. if (ret != SUCCESS) {
  355. ErrorManager::GetInstance().ATCReportErrMessage(
  356. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  357. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s",
  358. var->GetName().c_str(), index, iter->node_type.c_str());
  359. return INTERNAL_ERROR;
  360. }
  361. // set stream_label
  362. OpDescPtr var_desc = var->GetOpDesc();
  363. GE_CHECK_NOTNULL(var_desc);
  364. std::string stream_label;
  365. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  366. if (!stream_label.empty()) {
  367. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  368. }
  369. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  370. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  371. }
  372. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  373. return INTERNAL_ERROR;
  374. }
  375. }
  376. return SUCCESS;
  377. }
  378. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  379. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  380. VarNamesToRefs names_to_refs;
  381. std::string var_name;
  382. if (graph == nullptr) {
  383. GELOGE(PARAM_INVALID, "graph is null.");
  384. return names_to_refs;
  385. }
  386. for (auto &node : graph->GetAllNodes()) {
  387. if (node->GetType() != VARIABLE) {
  388. continue;
  389. }
  390. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  391. (void)names_to_refs[var_name].insert(node);
  392. }
  393. }
  394. return names_to_refs;
  395. }
  396. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  397. std::vector<int64_t> &dst_shape) {
  398. if (src_format == FORMAT_NCHW) {
  399. formats::FormatTransferNchwNc1hwc0 transfer;
  400. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  401. GELOGE(INTERNAL_ERROR, "TransShape failed");
  402. return FAILED;
  403. }
  404. } else if (src_format == FORMAT_NHWC) {
  405. formats::FormatTransferNhwcNc1hwc0 transfer;
  406. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  407. GELOGE(INTERNAL_ERROR, "TransShape failed");
  408. return FAILED;
  409. }
  410. }
  411. return SUCCESS;
  412. }
  413. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  414. GE_CHECK_NOTNULL(node_ptr);
  415. auto op_desc = node_ptr->GetOpDesc();
  416. GE_CHECK_NOTNULL(op_desc);
  417. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  418. GE_CHECK_NOTNULL(input);
  419. ge::Format old_format = input->GetFormat();
  420. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  421. ge::DataType dt = input->GetDataType();
  422. std::vector<int64_t> dst_shape_dims;
  423. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  424. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  425. return FAILED;
  426. }
  427. input->SetFormat(FORMAT_NC1HWC0);
  428. input->SetShape(ge::GeShape(dst_shape_dims));
  429. auto output = op_desc->MutableOutputDesc(0);
  430. GE_CHECK_NOTNULL(output);
  431. output->SetFormat(FORMAT_NC1HWC0);
  432. output->SetShape(ge::GeShape(dst_shape_dims));
  433. int64_t size = 0;
  434. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  435. if (graph_status != ge::GRAPH_SUCCESS) {
  436. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  437. return FAILED;
  438. }
  439. ge::TensorUtils::SetSize(*output, size);
  440. ge::TensorUtils::SetSize(*input, size);
  441. return SUCCESS;
  442. }
  443. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  444. GE_CHECK_NOTNULL(input_output);
  445. ge::Format old_format = input_output->GetFormat();
  446. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  447. ge::DataType dt = input_output->GetDataType();
  448. std::vector<int64_t> dst_shape_dims;
  449. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  450. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  451. return FAILED;
  452. }
  453. input_output->SetFormat(FORMAT_NC1HWC0);
  454. input_output->SetShape(ge::GeShape(dst_shape_dims));
  455. return SUCCESS;
  456. }
  457. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  458. vector<int64_t> &dst_shape_dims) {
  459. GE_CHECK_NOTNULL(op_desc);
  460. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  461. GE_CHECK_NOTNULL(input);
  462. ge::Format old_format = input->GetFormat();
  463. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  464. input->SetShape(ge::GeShape(dst_shape_dims));
  465. input->SetFormat(storage_format);
  466. auto output = op_desc->MutableOutputDesc(index);
  467. GE_CHECK_NOTNULL(output);
  468. output->SetShape(ge::GeShape(dst_shape_dims));
  469. output->SetFormat(storage_format);
  470. if (!output->MutableShape().IsUnknownShape()) {
  471. int64_t size = 0;
  472. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  473. if (graph_status != ge::GRAPH_SUCCESS) {
  474. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  475. return FAILED;
  476. }
  477. ge::TensorUtils::SetSize(*input, size);
  478. ge::TensorUtils::SetSize(*output, size);
  479. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  480. "new_shape:%s, new_format:%s, new_size:%lu",
  481. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  482. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  483. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  484. }
  485. return SUCCESS;
  486. }
  487. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) {
  488. is_dynamic_batch = false;
  489. std::string related_node_name;
  490. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  491. if (related_node_name.empty()) {
  492. ErrorManager::GetInstance().ATCReportErrMessage(
  493. "E15002", {"opname", "value", "reason"}, {data_node->GetName(), "flag", "but the value is empty"});
  494. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty",
  495. data_node->GetName().c_str());
  496. return INTERNAL_ERROR;
  497. }
  498. for (const NodePtr &next_node : data_node->GetOutNodes()) {
  499. if (next_node->GetName() == related_node_name) {
  500. switchn_node = next_node;
  501. break;
  502. }
  503. }
  504. if (switchn_node == nullptr) {
  505. ErrorManager::GetInstance().ATCReportErrMessage(
  506. "E15002", {"opname", "value", "reason"},
  507. {data_node->GetName(), related_node_name, "but can not find it on the graph"});
  508. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph",
  509. data_node->GetName().c_str(), related_node_name.c_str());
  510. return INTERNAL_ERROR;
  511. }
  512. is_dynamic_batch = true;
  513. }
  514. return SUCCESS;
  515. }
  516. bool CheckOpType(const NodePtr &node, const std::string type) {
  517. if (node->GetType() == type) {
  518. return true;
  519. }
  520. return false;
  521. }
  522. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  523. auto op = node_ptr->GetOpDesc();
  524. GE_CHECK_NOTNULL(op);
  525. auto inputDescsPtr = op->GetAllInputsDescPtr();
  526. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  527. ge::Format format = ge::FORMAT_ND;
  528. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  529. // process, otherwise fe will insert transdata.
  530. for (auto &inputDescPtr : inputDescsPtr) {
  531. GE_CHECK_NOTNULL(inputDescPtr);
  532. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  533. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  534. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  535. inputDescPtr->SetFormat(format);
  536. inputDescPtr->SetOriginFormat(format);
  537. }
  538. }
  539. for (auto &outputDescPtr : outputDescsPtr) {
  540. GE_CHECK_NOTNULL(outputDescPtr);
  541. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  542. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  543. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  544. outputDescPtr->SetFormat(format);
  545. outputDescPtr->SetOriginFormat(format);
  546. }
  547. }
  548. return SUCCESS;
  549. }
  550. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  551. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  552. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  553. // data and netoutput, and no longer inserts the transnode.
  554. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) {
  555. GE_CHECK_NOTNULL(node_ptr);
  556. auto op_desc = node_ptr->GetOpDesc();
  557. GE_CHECK_NOTNULL(op_desc);
  558. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  559. GE_CHECK_NOTNULL(input);
  560. ge::DataType src_dtype = input->GetDataType();
  561. if (src_dtype == dt_set) {
  562. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  563. return SUCCESS;
  564. }
  565. input->SetDataType(dt_set);
  566. int64_t input_shape_size = 0;
  567. int64_t output_shape_size = 0;
  568. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  569. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  570. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  571. GELOGE(GRAPH_FAILED, "GetTensorSize failed!");
  572. return FAILED;
  573. }
  574. ge::TensorUtils::SetSize(*input, input_shape_size);
  575. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  576. GE_CHECK_NOTNULL(output);
  577. output->SetDataType(dt_set);
  578. ge::TensorUtils::SetSize(*output, output_shape_size);
  579. if (is_dynamic_batch) {
  580. GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str());
  581. auto switchn_op_desc = switchn_node->GetOpDesc();
  582. GE_CHECK_NOTNULL(switchn_op_desc);
  583. auto switchn_input = switchn_op_desc->MutableInputDesc(0);
  584. GE_CHECK_NOTNULL(switchn_input);
  585. switchn_input->SetDataType(dt_set);
  586. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  587. const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i);
  588. GE_CHECK_NOTNULL(switchn_output);
  589. switchn_output->SetDataType(dt_set);
  590. }
  591. }
  592. return SUCCESS;
  593. }
  594. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  595. GE_CHECK_NOTNULL(node_ptr);
  596. auto op_desc = node_ptr->GetOpDesc();
  597. GE_CHECK_NOTNULL(op_desc);
  598. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  599. GE_CHECK_NOTNULL(input);
  600. ge::Format old_format = input->GetFormat();
  601. ge::GeShape old_shape = input->GetShape();
  602. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  603. if (!support) {
  604. ErrorManager::GetInstance().ATCReportErrMessage(
  605. "E19014", {"opname", "value", "reason"},
  606. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  607. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  608. GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str());
  609. return FAILED;
  610. }
  611. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  612. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  613. return FAILED;
  614. }
  615. if (is_dynamic_batch) {
  616. auto switchn_op_desc = switchn_node->GetOpDesc();
  617. GE_CHECK_NOTNULL(switchn_op_desc);
  618. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  619. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  620. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  621. return FAILED;
  622. }
  623. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  624. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  625. GE_CHECK_NOTNULL(switchn_output);
  626. old_format = switchn_output->GetFormat();
  627. old_shape = switchn_output->GetShape();
  628. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  629. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  630. return FAILED;
  631. }
  632. }
  633. }
  634. return SUCCESS;
  635. }
  636. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  637. auto op_desc = node_ptr->GetOpDesc();
  638. GE_CHECK_NOTNULL(op_desc);
  639. string set_dt_str;
  640. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  641. return SUCCESS;
  642. }
  643. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  644. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  645. bool is_dynamic_batch = false;
  646. NodePtr switchn_node = nullptr;
  647. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) {
  648. GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed");
  649. return FAILED;
  650. }
  651. if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) {
  652. GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed");
  653. return FAILED;
  654. }
  655. // check if need to set format
  656. string set_format;
  657. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  658. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  659. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  660. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) {
  661. GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed");
  662. return FAILED;
  663. }
  664. }
  665. return SUCCESS;
  666. }
  667. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  668. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  669. GE_CHECK_NOTNULL(op_desc);
  670. GE_CHECK_NOTNULL(tensor_desc_ptr);
  671. storage_format = FORMAT_RESERVED;
  672. int64_t format = FORMAT_RESERVED;
  673. dst_shape_dims.clear();
  674. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  675. storage_format = static_cast<Format>(format);
  676. vector<int32_t> storage_shape;
  677. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  678. for (auto dim : storage_shape) {
  679. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  680. }
  681. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  682. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  683. formats::JoinToString(storage_shape).c_str());
  684. } else {
  685. ErrorManager::GetInstance().ATCReportErrMessage(
  686. "15003", {"opname", "format"},
  687. {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)});
  688. GELOGE(PARAM_INVALID, "Update node by storage format failed, storage_shape not set. "
  689. "node: [%s], storage_format [%s]",
  690. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  691. return FAILED;
  692. }
  693. ge::Format old_format = tensor_desc_ptr->GetFormat();
  694. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  695. if (old_format == storage_format && old_shape == dst_shape_dims) {
  696. GELOGI("Update node by storage format, not changed.");
  697. storage_format = FORMAT_RESERVED;
  698. return SUCCESS;
  699. }
  700. }
  701. return SUCCESS;
  702. }
  703. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  704. NodePtr &node) {
  705. bool is_dynamic = CheckOpType(node, MERGE);
  706. auto src_op_desc = node->GetOpDesc();
  707. GE_CHECK_NOTNULL(src_op_desc);
  708. ge::GeShape src_shape = src_desc.GetShape();
  709. ge::Format src_format = src_desc.GetFormat();
  710. net_output_input_desc->SetDataType(DT_FLOAT16);
  711. if (is_dynamic) {
  712. auto merge_output = src_op_desc->MutableOutputDesc(0);
  713. GE_CHECK_NOTNULL(merge_output);
  714. merge_output->SetDataType(DT_FLOAT16);
  715. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  716. auto merge_input = src_op_desc->MutableInputDesc(i);
  717. GE_CHECK_NOTNULL(merge_input);
  718. merge_input->SetDataType(DT_FLOAT16);
  719. }
  720. }
  721. std::vector<int64_t> dst_shape_dims;
  722. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  723. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  724. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  725. return FAILED;
  726. }
  727. ge::GeShape dst_shape(dst_shape_dims);
  728. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  729. net_output_input_desc->SetShape(dst_shape);
  730. if (is_dynamic) {
  731. auto merge_out = src_op_desc->MutableOutputDesc(0);
  732. GE_CHECK_NOTNULL(merge_out);
  733. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  734. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  735. return FAILED;
  736. }
  737. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  738. auto merge_in = src_op_desc->MutableInputDesc(i);
  739. GE_CHECK_NOTNULL(merge_in);
  740. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  741. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  742. return FAILED;
  743. }
  744. }
  745. }
  746. return SUCCESS;
  747. }
  748. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  749. GE_CHECK_NOTNULL(netout_desc);
  750. vector<string> output_dt_str;
  751. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  752. for (auto dt_str : output_dt_str) {
  753. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  754. if (dt_str_split.size() == kUserDefinedElementCount) {
  755. if (dt_str_split[0] == to_string(index)) {
  756. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  757. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  758. TypeUtils::DataTypeToSerialString(dt).c_str());
  759. return true;
  760. }
  761. }
  762. }
  763. }
  764. return false;
  765. }
  766. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  767. GE_CHECK_NOTNULL(netout_desc);
  768. vector<string> output_format_str;
  769. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  770. for (auto format_str : output_format_str) {
  771. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  772. if (format_str_split.size() == kUserDefinedElementCount) {
  773. if (format_str_split[0] == to_string(index)) {
  774. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  775. return true;
  776. }
  777. }
  778. }
  779. }
  780. return false;
  781. }
  782. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  783. auto op_desc = node->GetOpDesc();
  784. GE_CHECK_NOTNULL(op_desc);
  785. ge::DataType output_data_type = ge::DT_FLOAT;
  786. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  787. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  788. auto peer_out = in_anchor->GetPeerOutAnchor();
  789. GE_CHECK_NOTNULL(peer_out);
  790. auto src_node = peer_out->GetOwnerNode();
  791. GE_CHECK_NOTNULL(src_node);
  792. bool is_dynamic = CheckOpType(src_node, MERGE);
  793. OpDescPtr src_op_desc = src_node->GetOpDesc();
  794. GE_CHECK_NOTNULL(src_op_desc);
  795. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  796. GE_CHECK_NOTNULL(net_output_input_desc);
  797. ge::GeShape old_shape = net_output_input_desc->GetShape();
  798. ge::Format old_format = net_output_input_desc->GetFormat();
  799. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  800. // Update datatype
  801. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  802. GELOGI("Enter into process output_type schedule");
  803. net_output_input_desc->SetDataType(output_data_type);
  804. if (is_dynamic) {
  805. auto merge_output = src_op_desc->MutableOutputDesc(0);
  806. GE_CHECK_NOTNULL(merge_output);
  807. merge_output->SetDataType(output_data_type);
  808. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  809. auto merge_input = src_op_desc->MutableInputDesc(i);
  810. GE_CHECK_NOTNULL(merge_input);
  811. merge_input->SetDataType(output_data_type);
  812. }
  813. }
  814. }
  815. // check if is_output_adjust_hw_layout is set
  816. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  817. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  818. ErrorManager::GetInstance().ATCReportErrMessage(
  819. "E19014", {"opname", "value", "reason"},
  820. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  821. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  822. GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0.");
  823. return FAILED;
  824. }
  825. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  826. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  827. GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0.");
  828. return FAILED;
  829. }
  830. }
  831. }
  832. return SUCCESS;
  833. }
  834. long StringToLongNoThrow(const string &str) {
  835. try {
  836. return std::stol(str);
  837. } catch (const std::invalid_argument) {
  838. GELOGE(PARAM_INVALID,
  839. "Parse shape range of input failed when transfer from string to int64. Given %s, while correct example: "
  840. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  841. str.c_str());
  842. return PARAM_INVALID;
  843. } catch (const std::out_of_range) {
  844. GELOGE(PARAM_INVALID,
  845. "Parse shape range of input failed when transfer from string to int64. Given %s, while correct example: "
  846. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  847. str.c_str());
  848. return PARAM_INVALID;
  849. }
  850. }
  851. /**
  852. * Parser shape_range from string to vector
  853. * shape_range from option normally is "[1~20,3,3~6,-1],[1~20,3,3~6,-1]"
  854. * @param shape_range
  855. */
  856. Status ParseDynamicInputShapeRange(const std::string &shape_range,
  857. std::vector<std::vector<std::pair<int64_t, int64_t>>> &range) {
  858. if (shape_range.size() < 2) {
  859. GELOGE(PARAM_INVALID, "Shape range %s is invalid.", shape_range.c_str());
  860. return PARAM_INVALID;
  861. }
  862. // different shape_range of single input are split by ']'
  863. vector<string> shape_range_set = ge::StringUtils::Split(shape_range, ']');
  864. if (shape_range_set.empty()) {
  865. GELOGE(PARAM_INVALID, "Shape range %s is not valid. Correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  866. shape_range.c_str());
  867. return PARAM_INVALID;
  868. }
  869. for (auto &shape_range_str : shape_range_set) {
  870. if (shape_range_str.size() < 3) {
  871. // shape_range_str should be "[2~3,1"
  872. // or ",[2~3,1". because we should trim '[' or ',['
  873. // so shape_range_str.size() < 3 is invalid
  874. continue;
  875. }
  876. // trim start bytes, after that, single input should be "1~20,3,3~6,-1"
  877. if (ge::StringUtils::StartWith(shape_range_str, "[")) {
  878. shape_range_str = shape_range_str.substr(1, shape_range_str.size());
  879. }
  880. if (ge::StringUtils::StartWith(shape_range_str, ",")) {
  881. shape_range_str = shape_range_str.substr(2, shape_range_str.size());
  882. }
  883. // parse shape_range of single input. eg. "1~20,3,3~6,-1"
  884. std::vector<std::pair<int64_t, int64_t>> range_of_single_input;
  885. vector<string> dim_range_set = ge::StringUtils::Split(shape_range_str, ',');
  886. for (const auto &range_pair_str : dim_range_set) {
  887. vector<string> range_pair_set = ge::StringUtils::Split(range_pair_str, '~');
  888. pair<int64_t, int64_t> range_pair;
  889. if (range_pair_set.size() == 1) {
  890. // fix dim
  891. auto range_value = StringToLongNoThrow(range_pair_set.at(0).c_str());
  892. if (range_value < 0) {
  893. range_pair = std::make_pair(1, range_value);
  894. } else {
  895. range_pair = std::make_pair(range_value, range_value);
  896. }
  897. } else if (range_pair_set.size() == 2) {
  898. // unknown dim, should get range.
  899. auto range_left = StringToLongNoThrow(range_pair_set.at(0).c_str());
  900. auto range_right = StringToLongNoThrow(range_pair_set.at(1).c_str());
  901. range_pair = std::make_pair(range_left, range_right);
  902. } else {
  903. GELOGE(PARAM_INVALID,
  904. "Shape range of input is invalid. Given %s, while correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  905. shape_range.c_str());
  906. return PARAM_INVALID;
  907. }
  908. range_of_single_input.emplace_back(range_pair);
  909. }
  910. range.emplace_back(range_of_single_input);
  911. }
  912. return SUCCESS;
  913. }
  914. Status GetDynamicInputShapeRange(const std::vector<GeTensor> &user_input, const std::map<string, string> &graph_option,
  915. vector<vector<std::pair<int64_t, int64_t>>> &range_vec) {
  916. auto mode_iter = graph_option.find(OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  917. if (mode_iter == graph_option.end()) {
  918. GELOGD("Graph Option: Can not find %s option in graph options.", OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  919. return SUCCESS;
  920. }
  921. GELOGD("Graph Option: dynamic_input_mode value is %s.", mode_iter->second.c_str());
  922. if (mode_iter->second != "dynamic_execute") {
  923. return SUCCESS;
  924. }
  925. auto iter = graph_option.find(OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  926. if (iter == graph_option.end()) {
  927. GELOGE(PARAM_INVALID, "Graph option %s is required when %s is dynamic_execute", OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE,
  928. OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  929. return PARAM_INVALID;
  930. }
  931. GELOGD("GraphOption: dynamic_inputs_shape_range value is %s.", iter->second.c_str());
  932. auto ret = ParseDynamicInputShapeRange(iter->second, range_vec);
  933. GE_CHK_STATUS_RET(ret, "Parse dynamic input shape range failed.");
  934. if (range_vec.size() != user_input.size()) {
  935. GELOGE(PARAM_INVALID, "Dynamic input shape range size is %zu, inputs size is %zu. Not match.", range_vec.size(),
  936. user_input.size());
  937. return PARAM_INVALID;
  938. }
  939. return SUCCESS;
  940. }
  941. Status UpdateDynamicInputShapeRange(const ge::GeAttrValue::INT index,
  942. const vector<vector<std::pair<int64_t, int64_t>>> &range_vec, OpDescPtr &op,
  943. GeTensorDesc &desc) {
  944. auto origin_shape = desc.GetShape();
  945. auto current_shape_range_vec = range_vec.at(index);
  946. if (current_shape_range_vec.size() != origin_shape.GetDimNum()) {
  947. GELOGE(PARAM_INVALID, "Given shape_range dim num is %zu, current dim num is %zu, not match.Pleace Check.",
  948. current_shape_range_vec.size(), origin_shape.GetDimNum());
  949. return PARAM_INVALID;
  950. }
  951. for (size_t i = 0; i < origin_shape.GetDimNum(); ++i) {
  952. auto curr_dim = origin_shape.GetDim(i);
  953. auto left_range = current_shape_range_vec.at(i).first;
  954. auto right_range = current_shape_range_vec.at(i).second;
  955. if (left_range == right_range) {
  956. // given shape_range is known dim, check is same as origin or not
  957. if (curr_dim != left_range) {
  958. GELOGE(PARAM_INVALID, "Given shape range is %ld, current dim shape is %ld, not match.Pleace Check.",
  959. left_range, curr_dim);
  960. return PARAM_INVALID;
  961. }
  962. origin_shape.SetDim(i, left_range);
  963. } else {
  964. // given shape_range is fix range, check input_shape is in this range or not
  965. if (right_range != UNKNOWN_DIM) {
  966. if ((curr_dim < left_range) || (curr_dim > right_range)) {
  967. GELOGE(PARAM_INVALID, "Given shape range is [%ld~%ld], current dim shape is %ld, out of range.Pleace Check.",
  968. left_range, right_range, curr_dim);
  969. return PARAM_INVALID;
  970. }
  971. }
  972. origin_shape.SetDim(i, UNKNOWN_DIM);
  973. }
  974. }
  975. desc.SetShape(origin_shape);
  976. desc.SetShapeRange(current_shape_range_vec);
  977. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  978. GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  979. graph_ret = op->UpdateOutputDesc(0, desc);
  980. GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  981. return SUCCESS;
  982. }
  983. } // namespace
  984. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  985. GraphPrepare::~GraphPrepare() {}
  986. /**
  987. * @param graph
  988. * @return
  989. */
  990. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  991. GE_CHECK_NOTNULL(graph);
  992. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  993. for (auto &node : graph->GetAllNodes()) {
  994. if (node == nullptr) {
  995. continue;
  996. }
  997. if (node->GetType() != VARIABLE) {
  998. continue;
  999. }
  1000. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  1001. if (trans_road == nullptr) {
  1002. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  1003. continue;
  1004. }
  1005. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  1006. auto ret = RecoverTransRoadForVar(node, *trans_road);
  1007. if (ret != SUCCESS) {
  1008. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str());
  1009. return INTERNAL_ERROR;
  1010. }
  1011. auto iter = var_names_to_refs.find(node->GetName());
  1012. if (iter != var_names_to_refs.end()) {
  1013. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  1014. if (ret != SUCCESS) {
  1015. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str());
  1016. return INTERNAL_ERROR;
  1017. }
  1018. }
  1019. }
  1020. return SUCCESS;
  1021. }
  1022. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  1023. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  1024. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  1025. if (compute_graph_ != nullptr) {
  1026. compute_graph_->SetSessionID(session_id);
  1027. }
  1028. session_id_ = session_id;
  1029. Status ret = CheckGraph();
  1030. if (ret != SUCCESS) {
  1031. GELOGE(ret, "RunGraph graph check fail, ret:%u", ret);
  1032. return ret;
  1033. }
  1034. (void)compute_graph_->TopologicalSorting();
  1035. ret = CheckRefOp();
  1036. if (ret != SUCCESS) {
  1037. GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret);
  1038. return ret;
  1039. }
  1040. return SUCCESS;
  1041. }
  1042. Status GraphPrepare::CheckGraph() {
  1043. if (compute_graph_ == nullptr) {
  1044. GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR");
  1045. return GE_GRAPH_INIT_FAILED;
  1046. }
  1047. auto nodes = compute_graph_->GetAllNodes();
  1048. if (nodes.empty()) {
  1049. GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph.");
  1050. return GE_GRAPH_INIT_FAILED;
  1051. }
  1052. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  1053. GE_CHECK_NOTNULL(node);
  1054. if (node->GetOpDesc() == nullptr) {
  1055. GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL");
  1056. return GE_GRAPH_INIT_FAILED;
  1057. }
  1058. }
  1059. return SUCCESS;
  1060. }
  1061. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  1062. const std::set<NodePtr> &ref_nodes) {
  1063. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  1064. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  1065. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  1066. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  1067. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  1068. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH};
  1069. GE_CHECK_NOTNULL(node);
  1070. const auto &op_desc = node->GetOpDesc();
  1071. GE_CHECK_NOTNULL(op_desc);
  1072. const auto input_index = op_desc->GetInputIndexByName(input_name);
  1073. const auto &in_anchor = node->GetInDataAnchor(input_index);
  1074. GE_CHECK_NOTNULL(in_anchor);
  1075. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  1076. GE_CHECK_NOTNULL(peer_out_anchor);
  1077. const auto &input_node = peer_out_anchor->GetOwnerNode();
  1078. GE_CHECK_NOTNULL(input_node);
  1079. const auto &input_op_desc = input_node->GetOpDesc();
  1080. GE_CHECK_NOTNULL(input_op_desc);
  1081. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  1082. if (is_ref) {
  1083. return SUCCESS;
  1084. }
  1085. auto input_type = input_op_desc->GetType();
  1086. if (input_type == ge::FRAMEWORKOP) {
  1087. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  1088. GELOGE(PARAM_INVALID, "Get original type failed.");
  1089. return PARAM_INVALID;
  1090. }
  1091. }
  1092. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  1093. if (!is_acceptable) {
  1094. ErrorManager::GetInstance().ATCReportErrMessage(
  1095. "E15005", {"opname", "optype", "opname1", "optype1"},
  1096. {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()});
  1097. GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.",
  1098. node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  1099. input_op_desc->GetType().c_str());
  1100. return PARAM_INVALID;
  1101. }
  1102. return SUCCESS;
  1103. }
  1104. Status GraphPrepare::CheckRefOp() {
  1105. GE_CHECK_NOTNULL(compute_graph_);
  1106. std::set<NodePtr> ref_nodes;
  1107. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  1108. if (node == nullptr) {
  1109. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  1110. return PARAM_INVALID;
  1111. }
  1112. auto op_desc = node->GetOpDesc();
  1113. if (op_desc == nullptr) {
  1114. GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null.");
  1115. return PARAM_INVALID;
  1116. }
  1117. auto input_name_index = op_desc->GetAllInputName();
  1118. auto outputs = op_desc->GetAllOutputName();
  1119. for (const auto &name_index : input_name_index) {
  1120. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  1121. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  1122. GELOGE(PARAM_INVALID, "CheckRefInputNode failed.");
  1123. return PARAM_INVALID;
  1124. }
  1125. (void)ref_nodes.insert(node); // no need to check value
  1126. }
  1127. }
  1128. }
  1129. return SUCCESS;
  1130. };
  1131. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  1132. GE_CHECK_NOTNULL(compute_graph_);
  1133. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  1134. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  1135. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  1136. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  1137. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  1138. return SUCCESS;
  1139. }
  1140. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1141. if (node == nullptr) {
  1142. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL");
  1143. return GE_GRAPH_GRAPH_NODE_NULL;
  1144. }
  1145. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1146. if (op_desc_ptr == nullptr) {
  1147. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL");
  1148. return GE_GRAPH_GRAPH_NODE_NULL;
  1149. }
  1150. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1151. int64_t tensor_size = 0;
  1152. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1153. if (graph_status != GRAPH_SUCCESS) {
  1154. ErrorManager::GetInstance().ATCReportErrMessage(
  1155. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1156. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  1157. return FAILED;
  1158. }
  1159. TensorUtils::SetSize(output, tensor_size);
  1160. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1161. if (graph_ret != GRAPH_SUCCESS) {
  1162. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1163. return graph_ret;
  1164. }
  1165. return SUCCESS;
  1166. }
  1167. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input, const std::map<string,string> &graph_option) {
  1168. // Get shape range of input in dynamic_execute mode
  1169. vector<vector<std::pair<int64_t,int64_t>>> dynamic_shape_range_vec;
  1170. auto ret = GetDynamicInputShapeRange(user_input, graph_option, dynamic_shape_range_vec);
  1171. GE_CHK_STATUS_RET(ret, "Graph option is not right on Dynamic execute mode.");
  1172. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1173. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1174. GE_CHECK_NOTNULL(input_node);
  1175. OpDescPtr op = input_node->GetOpDesc();
  1176. GE_CHECK_NOTNULL(op);
  1177. if (op->GetType() == DATA) {
  1178. GeAttrValue::INT index = 0;
  1179. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1180. GELOGW("Get index from data attr failed");
  1181. continue;
  1182. }
  1183. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1184. std::string situation = "data op index[" + std::to_string(index) + "]";
  1185. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1186. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1187. GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index);
  1188. return FAILED;
  1189. }
  1190. if (IsDynamicDims(input_node)) {
  1191. continue;
  1192. }
  1193. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1194. auto format = desc.GetFormat();
  1195. auto origin_format = desc.GetOriginFormat();
  1196. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1197. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op);
  1198. if (need_check_internal_format) {
  1199. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1200. if (is_internal) {
  1201. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1202. {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" +
  1203. TypeUtils::FormatToSerialString(origin_format) + "]", "it is not support"});
  1204. GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.",
  1205. TypeUtils::FormatToSerialString(format).c_str(),
  1206. TypeUtils::FormatToSerialString(origin_format).c_str());
  1207. return FAILED;
  1208. }
  1209. }
  1210. auto data_type = desc.GetDataType();
  1211. uint32_t length = 1;
  1212. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1213. if (!type_ret) {
  1214. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1215. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1216. GELOGE(PARAM_INVALID, "Input datatype %s is not support.",
  1217. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1218. return FAILED;
  1219. }
  1220. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1221. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1222. int64_t shape_size = desc_shape * length;
  1223. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1224. int64_t size = 0;
  1225. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1226. GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed");
  1227. return FAILED);
  1228. bool size_check = (size != 0 && shape_size != size);
  1229. if (size_check) {
  1230. std::string situation = "input data size[" + std::to_string(size) +
  1231. "] and shape_size[" + std::to_string(size) + "]";
  1232. std::string reason = "because size != 0 and shape_size != size";
  1233. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1234. GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size);
  1235. return FAILED;
  1236. }
  1237. ge::TensorUtils::SetSize(desc, shape_size);
  1238. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1239. if (graph_ret != GRAPH_SUCCESS) {
  1240. GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
  1241. return graph_ret;
  1242. }
  1243. // Size will be recalculated in the build stage
  1244. ge::TensorUtils::SetSize(desc, 0);
  1245. graph_ret = op->UpdateOutputDesc(0, desc);
  1246. if (graph_ret != GRAPH_SUCCESS) {
  1247. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1248. return graph_ret;
  1249. }
  1250. if (!dynamic_shape_range_vec.empty()) {
  1251. ret = UpdateDynamicInputShapeRange(index, dynamic_shape_range_vec, op, desc);
  1252. GE_CHK_STATUS_RET(ret, "Fail to update dynamic input shape range on %s.", op->GetName().c_str());
  1253. continue;
  1254. }
  1255. if (!options_.train_graph_flag) {
  1256. Status ret = AdjustDataOpOutput(input_node);
  1257. GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret); return ret);
  1258. }
  1259. }
  1260. }
  1261. return SUCCESS;
  1262. }
  1263. Status GraphPrepare::TryDoAipp() {
  1264. // infer and with aipp configure file, then call aipp insert
  1265. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1266. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1267. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1268. if (ret != SUCCESS) {
  1269. GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed.");
  1270. return INTERNAL_ERROR;
  1271. }
  1272. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1273. if (ret != SUCCESS) {
  1274. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed",
  1275. options_.insert_op_file.c_str());
  1276. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1277. }
  1278. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1279. if (ret != SUCCESS) {
  1280. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret);
  1281. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1282. }
  1283. }
  1284. return SUCCESS;
  1285. }
  1286. Status GraphPrepare::FormatAndShapeProcess() {
  1287. Status ret = ResourcePairProcess("add");
  1288. if (ret != SUCCESS) {
  1289. GELOGE(ret, "ResourcePairProcess failed");
  1290. return ret;
  1291. }
  1292. GE_TIMESTAMP_START(InferOriginFormat1);
  1293. ret = compute_graph_->InferOriginFormat();
  1294. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1295. GE_DUMP(compute_graph_, "after_first_inferformat");
  1296. if (ret != SUCCESS) {
  1297. GELOGE(ret, "Prepare Graph first inferformat failed");
  1298. return ret;
  1299. }
  1300. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1301. ret = InferShapeForPreprocess();
  1302. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1303. GE_DUMP(compute_graph_, "after_infershape");
  1304. if (ret != SUCCESS) {
  1305. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed");
  1306. return GE_GRAPH_INFERSHAPE_FAILED;
  1307. }
  1308. GE_TIMESTAMP_START(InferOriginFormat2);
  1309. ret = compute_graph_->InferOriginFormat();
  1310. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1311. if (ret != SUCCESS) {
  1312. GELOGE(ret, "Prepare Graph inferformat failed");
  1313. return ret;
  1314. }
  1315. ret = ResourcePairProcess("remove");
  1316. if (ret != SUCCESS) {
  1317. return ret;
  1318. }
  1319. return ret;
  1320. }
  1321. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1322. PassManager control_pass;
  1323. // Graph pass tmp logic for resource infershape
  1324. if (options_.train_graph_flag) {
  1325. try {
  1326. if (action == "add") {
  1327. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1328. } else {
  1329. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1330. new ResourcePairRemoveControlPass);
  1331. }
  1332. } catch (std::bad_alloc &e) {
  1333. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str());
  1334. return INTERNAL_ERROR;
  1335. }
  1336. }
  1337. Status ret = control_pass.Run(compute_graph_);
  1338. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1339. GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret);
  1340. return ret;
  1341. }
  1342. return SUCCESS;
  1343. }
  1344. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1345. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1346. GE_CHECK_NOTNULL(node_ptr);
  1347. if (node_ptr->GetType() == DATA) {
  1348. uint32_t index = 0;
  1349. auto op_desc = node_ptr->GetOpDesc();
  1350. GE_CHECK_NOTNULL(op_desc);
  1351. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1352. Format storage_format = FORMAT_RESERVED;
  1353. vector<int64_t> dst_shape_dims;
  1354. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1355. GELOGE(INTERNAL_ERROR, "Get storage format for input failed");
  1356. return FAILED;
  1357. }
  1358. if (storage_format == FORMAT_RESERVED) {
  1359. continue;
  1360. }
  1361. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1362. GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed");
  1363. return FAILED;
  1364. }
  1365. }
  1366. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1367. auto op_desc = node_ptr->GetOpDesc();
  1368. GE_CHECK_NOTNULL(op_desc);
  1369. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1370. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1371. Format storage_format = FORMAT_RESERVED;
  1372. vector<int64_t> dst_shape_dims;
  1373. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1374. GELOGE(INTERNAL_ERROR, "Get storage format from output failed");
  1375. return FAILED;
  1376. }
  1377. if (storage_format == FORMAT_RESERVED) {
  1378. continue;
  1379. }
  1380. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1381. GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed");
  1382. return FAILED;
  1383. }
  1384. }
  1385. }
  1386. }
  1387. return SUCCESS;
  1388. }
  1389. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1390. if (options_.save_original_model == "true") {
  1391. ModelHelper model_helper;
  1392. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1393. options_.original_model_file);
  1394. if (ret != SUCCESS) {
  1395. // If save original model fail, process continue
  1396. GELOGW("SaveOriginalGraphToOmModel fail");
  1397. }
  1398. }
  1399. return SUCCESS;
  1400. }
  1401. #define PP_RUN_AND_DUMP(name, func, ...) \
  1402. do { \
  1403. GE_RUN(Prepare, func, __VA_ARGS__); \
  1404. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1405. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1406. } while (0)
  1407. #define PP_RUN(name, func, ...) \
  1408. do { \
  1409. GE_RUN(Prepare, func, __VA_ARGS__); \
  1410. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1411. } while (0)
  1412. Status GraphPrepare::PrepareDynShape(const GraphNodePtr &graph_node, const std::vector<GeTensor> &user_input,
  1413. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1414. GE_CHECK_NOTNULL(graph_node->GetGraph());
  1415. GE_CHECK_NOTNULL(compute_graph);
  1416. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1417. const Graph &const_graph = *graph_node->GetGraph();
  1418. PP_RUN("Init", Init, const_graph, session_id);
  1419. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1420. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input, graph_node->GetOptions());
  1421. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1422. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1423. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1424. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1425. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1426. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1427. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1428. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1429. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1430. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1431. return SUCCESS;
  1432. }
  1433. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1434. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1435. return SUCCESS;
  1436. }
  1437. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1438. auto compute_graph = compute_graph_;
  1439. PassManager pass_manager;
  1440. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1441. new (std::nothrow) VariablePrepareOpPass))
  1442. GE_TIMESTAMP_START(pass_manager);
  1443. auto ret = pass_manager.Run(compute_graph);
  1444. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1445. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1446. GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret);
  1447. return ret;
  1448. }
  1449. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1450. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1451. return SUCCESS;
  1452. }
  1453. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1454. if (compute_graph == nullptr) {
  1455. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1456. return GE_GRAPH_NULL_INPUT;
  1457. }
  1458. GEPass ge_passes(compute_graph);
  1459. NamesToPass hccl_group;
  1460. HcclGroupPass hccl_group_pass;
  1461. GELOGD("Add hccl group pass success");
  1462. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1463. auto ret = ge_passes.Run(hccl_group);
  1464. if (ret != SUCCESS) {
  1465. GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret);
  1466. return ret;
  1467. }
  1468. ret = compute_graph->TopologicalSorting();
  1469. if (ret != SUCCESS) {
  1470. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1471. return ret;
  1472. }
  1473. return SUCCESS;
  1474. }
  1475. #undef PP_RUN_AND_DUMP
  1476. #undef PP_RUN
  1477. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1478. if (graph == nullptr) {
  1479. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1480. return GE_GRAPH_NULL_INPUT;
  1481. }
  1482. const Graph &const_graph = *graph;
  1483. Status ret = Init(const_graph, 0);
  1484. if (ret != SUCCESS) {
  1485. GELOGE(ret, "Init graph_prepare fail, ret:%u", ret);
  1486. return ret;
  1487. }
  1488. GE_DUMP(compute_graph_, "after_parser");
  1489. GELOGI("Start infershape for dump json process.");
  1490. ret = compute_graph_->InferOriginFormat();
  1491. GE_DUMP(compute_graph_, "after_inferformat");
  1492. if (ret != SUCCESS) {
  1493. GELOGE(ret, "Prepare Graph inferformat failed");
  1494. return ret;
  1495. }
  1496. InferShapePass infer_shape_pass;
  1497. NamesToPass names_to_passes;
  1498. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1499. GEPass ge_passes(compute_graph_);
  1500. ret = ge_passes.Run(names_to_passes);
  1501. GE_DUMP(compute_graph_, "after_infershape");
  1502. if (ret != SUCCESS) {
  1503. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1504. return ret;
  1505. }
  1506. ShapeRefiner::ClearContextMap();
  1507. return SUCCESS;
  1508. }
  1509. Status GraphPrepare::CheckConstOp() {
  1510. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1511. GE_CHECK_NOTNULL(node_ptr);
  1512. if (node_ptr->GetType() == CONSTANT) {
  1513. Status ret = VerifyConstOp(node_ptr);
  1514. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1515. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1516. auto op_desc = node_ptr->GetOpDesc();
  1517. if (op_desc == nullptr) {
  1518. GELOGE(PARAM_INVALID, "Get op desc failed");
  1519. return PARAM_INVALID;
  1520. }
  1521. std::string original_type;
  1522. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1523. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1524. GELOGI("original type is %s", original_type.c_str());
  1525. if (original_type == CONSTANT) {
  1526. Status ret = VerifyConstOp(node_ptr);
  1527. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1528. }
  1529. }
  1530. }
  1531. return SUCCESS;
  1532. }
  1533. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1534. GE_CHECK_NOTNULL(node);
  1535. auto op_desc = node->GetOpDesc();
  1536. GE_CHECK_NOTNULL(op_desc);
  1537. ConstGeTensorPtr ge_tensor_ptr;
  1538. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1539. GELOGE(PARAM_INVALID, "Get value from const attr failed");
  1540. return PARAM_INVALID;
  1541. }
  1542. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1543. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1544. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1545. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1546. auto data_type = ge_tensor_desc.GetDataType();
  1547. if (data_type == DT_STRING) {
  1548. return SUCCESS;
  1549. }
  1550. uint32_t length = 1;
  1551. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1552. if (!type_ret) {
  1553. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1554. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1555. GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str());
  1556. return FAILED;
  1557. }
  1558. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1559. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1560. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1561. if (shape_size == 0) {
  1562. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1563. // shape = [], means it's a sclar tensor.
  1564. GE_CHK_BOOL_EXEC(data_size / length == 1,
  1565. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid scalar tensor."});
  1566. return PARAM_INVALID, "Const is invalid scalar tensor.");
  1567. } else {
  1568. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1569. GE_CHK_BOOL_EXEC(data_size == 0,
  1570. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid vector scalar."});
  1571. return PARAM_INVALID, "Const is invalid vector scalar.");
  1572. }
  1573. } else {
  1574. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0,
  1575. ErrorManager::GetInstance().ATCReportErrMessage(
  1576. "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"});
  1577. return PARAM_INVALID, "Const input data size is not equal with tensor desc shape");
  1578. }
  1579. return SUCCESS;
  1580. }
  1581. bool GraphPrepare::IsDynamicDims(const NodePtr &input_node) {
  1582. auto data_shape = NodeUtils::GetOutputDesc(*input_node, kDataOutIndex).GetShape();
  1583. const auto &dims = data_shape.GetDims();
  1584. bool all_is_positive = false;
  1585. if (std::all_of(dims.begin(), dims.end(), [](int64_t val) { return val >= 0; })) {
  1586. all_is_positive = true;
  1587. }
  1588. if (!all_is_positive && !options_.input_shape.empty() && !options_.dynamic_dims.empty() &&
  1589. options_.dynamic_node_type != kInvalidDynaimcDimsType) {
  1590. GELOGI("No need to check and update desc info, the dims of %s is %s.", input_node->GetName().c_str(),
  1591. formats::JoinToString(dims).c_str());
  1592. return true;
  1593. }
  1594. return false;
  1595. }
  1596. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1597. if (GetLocalOmgContext().is_dynamic_input) {
  1598. return SUCCESS;
  1599. }
  1600. unsigned int node_num = 0;
  1601. unsigned int data_num = 0;
  1602. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1603. GE_CHECK_NOTNULL(input_node);
  1604. OpDescPtr op = input_node->GetOpDesc();
  1605. GE_CHECK_NOTNULL(op);
  1606. node_num++;
  1607. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1608. data_num++;
  1609. GeAttrValue::INT index = 0;
  1610. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1611. GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed");
  1612. return GE_GRAPH_INIT_FAILED;
  1613. }
  1614. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1615. std::string situation = "data op index[" + std::to_string(index) + "]";
  1616. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1617. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1618. GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index);
  1619. return GE_GRAPH_INIT_FAILED;
  1620. }
  1621. if (IsDynamicDims(input_node)) {
  1622. continue;
  1623. }
  1624. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1625. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1626. if (desc.GetShape().GetDim(i) < 0) {
  1627. std::string situation = "data dim[" + std::to_string(i) + "][" +
  1628. std::to_string(desc.GetShape().GetDim(i)) + "]" ;
  1629. std::string reason = "it need >= 0";
  1630. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1631. GELOGE(GE_GRAPH_INIT_FAILED, "data dim %zu is not supported, need >= 0, real:%ld.", i,
  1632. desc.GetShape().GetDim(i));
  1633. return GE_GRAPH_INIT_FAILED;
  1634. }
  1635. }
  1636. }
  1637. }
  1638. if (node_num <= data_num) {
  1639. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1640. }
  1641. return SUCCESS;
  1642. }
  1643. Status GraphPrepare::InferShapeForPreprocess() {
  1644. GELOGI("Start infershape for preprocess.");
  1645. GEPass ge_passes(compute_graph_);
  1646. NamesToPass names_to_passes;
  1647. AssertPass assert_pass;
  1648. if (!options_.train_graph_flag) {
  1649. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1650. }
  1651. SwitchDeadBranchElimination switch_dead_branch_elimination;
  1652. names_to_passes.emplace_back("SwitchDeadBranchElimination", &switch_dead_branch_elimination);
  1653. MergePass merge_pass;
  1654. names_to_passes.emplace_back("MergePass", &merge_pass);
  1655. InferShapePass infer_shape_pass;
  1656. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1657. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1658. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1659. DimensionComputePass dimension_compute_pass;
  1660. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1661. ConstantFoldingPass constant_folding_pass;
  1662. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1663. int32_t dev_count = 0;
  1664. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1665. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1666. rtError_t rt_err = RT_ERROR_NONE;
  1667. if (aicpu_constant_folding_on != nullptr) {
  1668. rt_err = rtGetDeviceCount(&dev_count);
  1669. if (rt_err == RT_ERROR_NONE) {
  1670. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1671. if (result != SUCCESS) {
  1672. GELOGE(result, "Set rt context failed.");
  1673. return result;
  1674. }
  1675. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1676. }
  1677. }
  1678. Status ret = ge_passes.Run(names_to_passes);
  1679. if (aicpu_constant_folding_on != nullptr) {
  1680. if (rt_err == RT_ERROR_NONE) {
  1681. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1682. if (result != SUCCESS) {
  1683. GELOGE(result, "Set rt context failed.");
  1684. return result;
  1685. }
  1686. }
  1687. }
  1688. ShapeRefiner::ClearContextMap();
  1689. if (ret != SUCCESS) {
  1690. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1691. return ret;
  1692. }
  1693. return SUCCESS;
  1694. }
  1695. Status GraphPrepare::PrepareOptimize() {
  1696. GELOGI("Start optimize for preprocess.");
  1697. // check rw type
  1698. GraphOptimize graph_optimize;
  1699. bool has_conflict = false;
  1700. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1701. if (has_conflict) {
  1702. GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize.");
  1703. return FAILED;
  1704. }
  1705. PassManager original_graph_passes;
  1706. // Graph pass
  1707. try {
  1708. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1709. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1710. (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass);
  1711. } catch (std::bad_alloc &e) {
  1712. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1713. return INTERNAL_ERROR;
  1714. }
  1715. GE_TIMESTAMP_START(original_graph_passes);
  1716. Status ret = original_graph_passes.Run(compute_graph_);
  1717. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1718. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1719. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1720. return ret;
  1721. }
  1722. // New pass
  1723. GEPass ge_passes(compute_graph_);
  1724. NamesToPass names_to_passes;
  1725. EnterPass enter_pass;
  1726. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1727. CondPass cond_pass;
  1728. names_to_passes.emplace_back("CondPass", &cond_pass);
  1729. PrintOpPass print_pass;
  1730. if (options_.enable_print_op_pass) {
  1731. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1732. }
  1733. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1734. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1735. DropOutPass dropout_pass;
  1736. AssertPass assert_pass;
  1737. UnusedConstPass unused_const_pass;
  1738. StopGradientPass stop_gradient_pass;
  1739. PreventGradientPass prevent_gradient_pass;
  1740. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1741. GuaranteeConstPass guarantee_const_pass;
  1742. VarIsInitializedOpPass var_is_initialized_pass;
  1743. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1744. IdentityPass identity_pass(false);
  1745. SnapshotPass snapshot_pass;
  1746. if (!options_.train_graph_flag) {
  1747. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1748. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1749. }
  1750. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1751. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1752. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1753. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1754. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1755. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1756. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1757. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1758. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1759. GE_TIMESTAMP_START(names_to_passes);
  1760. ret = ge_passes.Run(names_to_passes);
  1761. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1762. if (ret != SUCCESS) {
  1763. GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret);
  1764. return ret;
  1765. }
  1766. PassManager graph_pass;
  1767. try {
  1768. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1769. } catch (std::bad_alloc &e) {
  1770. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1771. return INTERNAL_ERROR;
  1772. }
  1773. GE_TIMESTAMP_START(graph_passes);
  1774. ret = graph_pass.Run(compute_graph_);
  1775. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1776. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1777. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1778. return ret;
  1779. }
  1780. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1781. TypeConversionOfConstant();
  1782. ret = compute_graph_->TopologicalSorting();
  1783. if (ret != SUCCESS) {
  1784. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1785. return ret;
  1786. }
  1787. GELOGI("End optimize for preprocess.");
  1788. return SUCCESS;
  1789. }
  1790. void GraphPrepare::TypeConversionOfConstant() {
  1791. bool is_acl_compile = false;
  1792. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1793. // This can ensure that n is not a null pointer
  1794. // No Conversion when called by aclOpCompile
  1795. (void)AttrUtils::GetBool(n->GetOpDesc(), ATTR_DYNAMIC_SHAPE_SINGLE_AICPU, is_acl_compile);
  1796. if (is_acl_compile) {
  1797. return;
  1798. }
  1799. }
  1800. if (options_.train_graph_flag) {
  1801. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1802. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1803. // This can ensure that n is not a null pointer
  1804. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1805. n->GetOpDesc()->SetType(CONSTANTOP);
  1806. }
  1807. }
  1808. } else {
  1809. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  1810. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1811. // This can ensure that n is not a null pointer
  1812. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  1813. n->GetOpDesc()->SetType(CONSTANT);
  1814. }
  1815. }
  1816. }
  1817. }
  1818. Status GraphPrepare::GraphEquivalentTransformation() {
  1819. NamesToPass names_to_pass;
  1820. ForPass for_pass;
  1821. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  1822. return GEPass(compute_graph_).Run(names_to_pass);
  1823. }
  1824. Status GraphPrepare::ProcessBeforeInfershape() {
  1825. NamesToPass names_to_passes;
  1826. CondRemovePass condition_remove_pass;
  1827. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  1828. GE_TIMESTAMP_START(ProcessCondRemove);
  1829. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  1830. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  1831. if (ret != SUCCESS) {
  1832. GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  1833. return ret;
  1834. }
  1835. return SUCCESS;
  1836. }
  1837. Status GraphPrepare::ProcessNetOutput() {
  1838. PassManager graph_passes_before_infershape;
  1839. try {
  1840. if (options_.train_graph_flag) {
  1841. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  1842. }
  1843. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  1844. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  1845. new (std::nothrow) DataPass); // Add NetOutput first.
  1846. } catch (std::bad_alloc) {
  1847. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1848. return INTERNAL_ERROR;
  1849. }
  1850. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  1851. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  1852. GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret);
  1853. return ret;
  1854. }
  1855. return SUCCESS;
  1856. }
  1857. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input,const std::map<string,string> &graph_option) {
  1858. compute_graph_->SetInputSize(user_input.size());
  1859. if (user_input.empty()) {
  1860. return SUCCESS;
  1861. }
  1862. auto ret = CheckUserInput(user_input);
  1863. if (ret != SUCCESS) {
  1864. GELOGE(ret, "Check user input failed.");
  1865. return ret;
  1866. }
  1867. ret = UpdateInput(user_input, graph_option);
  1868. if (ret != SUCCESS) {
  1869. GELOGE(ret, "UpdateInput fail, ret:%u", ret);
  1870. return ret;
  1871. }
  1872. if (user_input.size() != 0) {
  1873. ret = CheckConstOp();
  1874. if (ret != SUCCESS) {
  1875. GELOGE(ret, "CheckConstOp fail, ret:%u", ret);
  1876. return ret;
  1877. }
  1878. } else {
  1879. ret = compute_graph_->TopologicalSorting();
  1880. if (ret != SUCCESS) {
  1881. GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting");
  1882. return FAILED;
  1883. }
  1884. }
  1885. return SUCCESS;
  1886. }
  1887. Status GraphPrepare::UpdateInputOutputByOptions() {
  1888. auto ret = UpdateDataNetOutputByStorageFormat();
  1889. if (ret != SUCCESS) {
  1890. GELOGE(ret, "Update format acoording to storage format failed.");
  1891. return ret;
  1892. }
  1893. if (options_.train_graph_flag) {
  1894. GELOGI("This is train mode, no need to do this schedule.");
  1895. return SUCCESS;
  1896. }
  1897. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  1898. GE_CHECK_NOTNULL(node_ptr);
  1899. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  1900. GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str());
  1901. return FAILED;
  1902. }
  1903. if (node_ptr->GetType() == DATA) {
  1904. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  1905. GELOGE(INTERNAL_ERROR, "Process data node failed");
  1906. return FAILED;
  1907. }
  1908. }
  1909. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1910. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  1911. GELOGE(INTERNAL_ERROR, "Process netoutput node failed");
  1912. return FAILED;
  1913. }
  1914. }
  1915. }
  1916. return SUCCESS;
  1917. }
  1918. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  1919. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  1920. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  1921. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  1922. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  1923. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  1924. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  1925. if (dst_node->GetType() == TRANSDATA) {
  1926. return true;
  1927. }
  1928. }
  1929. }
  1930. return false;
  1931. }
  1932. } // 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