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 88 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
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #include "graph/preprocess/graph_preprocess.h"
  17. #include <map>
  18. #include <set>
  19. #include <string>
  20. #include "common/formats/format_transfers/format_transfer_fractal_nz.h"
  21. #include "common/formats/format_transfers/format_transfer_nchw_nc1hwc0.h"
  22. #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h"
  23. #include "common/formats/format_transfers/format_transfer_transpose.h"
  24. #include "common/formats/utils/formats_trans_utils.h"
  25. #include "common/util/error_manager/error_manager.h"
  26. #include "common/helper/model_helper.h"
  27. #include "common/math/math_util.h"
  28. #include "common/op/ge_op_utils.h"
  29. #include "graph/common/ge_call_wrapper.h"
  30. #include "graph/common/local_context.h"
  31. #include "graph/common/transop_util.h"
  32. #include "graph/ge_context.h"
  33. #include "graph/shape_refiner.h"
  34. #include "graph/manager/graph_var_manager.h"
  35. #include "graph/manager/util/rt_context_util.h"
  36. #include "graph/optimize/graph_optimize.h"
  37. #include "graph/passes/addn_pass.h"
  38. #include "graph/passes/aicpu_constant_folding_pass.h"
  39. #include "graph/passes/assert_pass.h"
  40. #include "ge/ge_api_types.h"
  41. #include "graph/passes/common_subexpression_elimination_pass.h"
  42. #include "graph/passes/cond_pass.h"
  43. #include "graph/passes/cond_remove_pass.h"
  44. #include "graph/passes/constant_folding_pass.h"
  45. #include "graph/passes/dimension_adjust_pass.h"
  46. #include "graph/passes/dimension_compute_pass.h"
  47. #include "graph/passes/dropout_pass.h"
  48. #include "graph/passes/enter_pass.h"
  49. #include "graph/passes/for_pass.h"
  50. #include "graph/passes/guarantee_const_pass.h"
  51. #include "graph/passes/hccl_memcpy_pass.h"
  52. #include "graph/passes/hccl_group_pass.h"
  53. #include "graph/passes/identity_pass.h"
  54. #include "graph/passes/infershape_pass.h"
  55. #include "graph/passes/merge_pass.h"
  56. #include "graph/passes/net_output_pass.h"
  57. #include "graph/passes/no_use_reshape_remove_pass.h"
  58. #include "graph/passes/parallel_concat_start_op_pass.h"
  59. #include "graph/passes/placeholder_with_default_pass.h"
  60. #include "graph/passes/prevent_gradient_pass.h"
  61. #include "graph/passes/print_op_pass.h"
  62. #include "graph/passes/prune_pass.h"
  63. #include "graph/passes/replace_transshape_pass.h"
  64. #include "graph/passes/replace_with_empty_const_pass.h"
  65. #include "graph/passes/resource_pair_add_control_pass.h"
  66. #include "graph/passes/resource_pair_remove_control_pass.h"
  67. #include "graph/passes/save_pass.h"
  68. #include "graph/passes/shape_operate_op_remove_pass.h"
  69. #include "graph/passes/snapshot_pass.h"
  70. #include "graph/passes/stop_gradient_pass.h"
  71. #include "graph/passes/switch_dead_branch_elimination.h"
  72. #include "graph/passes/unused_const_pass.h"
  73. #include "graph/passes/var_is_initialized_op_pass.h"
  74. #include "graph/passes/variable_prepare_op_pass.h"
  75. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  76. #include "graph/utils/type_utils.h"
  77. #include "inc/pass_manager.h"
  78. #include "init/gelib.h"
  79. #include "multi_batch_copy_graph.h"
  80. #include "graph/passes/data_pass.h"
  81. #include "graph/passes/mark_agnostic_pass.h"
  82. namespace ge {
  83. namespace {
  84. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  85. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  86. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  87. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  88. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  89. // the size of user defined output datatype or format string after split by ":".
  90. const size_t kUserDefinedElementCount = 2;
  91. const int kDataOutIndex = 0;
  92. const int64_t kInvalidDynaimcDimsType = -1;
  93. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  94. GeTensorPtr tensor = MakeShared<GeTensor>();
  95. if (tensor == nullptr) {
  96. GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed");
  97. return nullptr;
  98. }
  99. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  100. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  101. auto dst_ge_shape = data_tensor.GetShape();
  102. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  103. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  104. tensor->MutableTensorDesc().SetShape(GeShape());
  105. int32_t dst_shape = 1;
  106. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  107. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  108. return nullptr;
  109. }
  110. } else {
  111. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  112. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  113. if (dst_shape == nullptr) {
  114. GELOGE(INTERNAL_ERROR, "Create unique ptr failed");
  115. return nullptr;
  116. }
  117. for (int64_t i = 0; i < dim_cnt; ++i) {
  118. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  119. }
  120. GE_IF_BOOL_EXEC(
  121. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  122. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  123. return nullptr;)
  124. }
  125. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  126. return OpDescUtils::CreateConstOp(tensor);
  127. }
  128. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  129. OpDescPtr &op_desc) {
  130. // For format transfer node, the IR definition has src/dst format attrs
  131. if (node_type == TRANSDATA) {
  132. GE_IF_BOOL_EXEC(
  133. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  134. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  135. GE_IF_BOOL_EXEC(
  136. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  137. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  138. }
  139. // For TransposeD node, the IR definition has perm attrs
  140. if (node_type == TRANSPOSED) {
  141. Format src_format = input.GetFormat();
  142. Format dst_format = output.GetFormat();
  143. std::vector<int64_t> perm_arg;
  144. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  145. "Get perm by foramt failed.");
  146. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  147. "SetStr PERMUTE_ATTR_PERM failed")
  148. }
  149. // For cast node, the IR definition has src/dst attrs
  150. if (node_type == CAST) {
  151. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  152. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  153. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  154. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  155. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  156. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  157. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  158. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  159. }
  160. }
  161. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  162. const GeTensorDesc &output, NodePtr &node) {
  163. if (node == nullptr) {
  164. GELOGE(PARAM_INVALID, "node is null.");
  165. return nullptr;
  166. }
  167. auto graph = node->GetOwnerComputeGraph();
  168. if (graph == nullptr) {
  169. GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str());
  170. return nullptr;
  171. }
  172. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  173. if (index < 0) {
  174. ErrorManager::GetInstance().ATCReportErrMessage(
  175. "E19025", {"situation", "reason"},
  176. {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()});
  177. GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str());
  178. return nullptr;
  179. }
  180. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  181. if (op_desc == nullptr) {
  182. GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed");
  183. return nullptr;
  184. }
  185. // for data dump
  186. GE_IF_BOOL_EXEC(
  187. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  188. GELOGW("CreateTransNode: SetListStr failed");)
  189. // Default single input and single output
  190. auto ret = op_desc->AddInputDesc(input);
  191. if (ret != GRAPH_SUCCESS) {
  192. GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str());
  193. return nullptr;
  194. }
  195. ret = op_desc->AddOutputDesc(output);
  196. if (ret != GRAPH_SUCCESS) {
  197. GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str());
  198. return nullptr;
  199. }
  200. AddTransNodeAttr(node_type, input, output, op_desc);
  201. NodePtr shape_node = nullptr;
  202. if (node_type == RESHAPE) {
  203. auto shape_desc = CreateTensorShape(output);
  204. if (shape_desc == nullptr) {
  205. GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input",
  206. node->GetName().c_str());
  207. return nullptr;
  208. }
  209. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  210. if (ret != GRAPH_SUCCESS) {
  211. GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str());
  212. return nullptr;
  213. }
  214. shape_node = graph->AddNode(shape_desc);
  215. if (shape_node == nullptr) {
  216. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str());
  217. return nullptr;
  218. }
  219. }
  220. auto trans_node = graph->AddNode(op_desc);
  221. if (trans_node == nullptr) {
  222. GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str());
  223. return nullptr;
  224. }
  225. if (node_type == RESHAPE) {
  226. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  227. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str());
  228. return nullptr;
  229. }
  230. }
  231. return trans_node;
  232. }
  233. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  234. NodePtr &trans_node) {
  235. GE_CHECK_NOTNULL(node);
  236. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  237. if (trans_node == nullptr) {
  238. return INTERNAL_ERROR;
  239. }
  240. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  241. if (ret != GRAPH_SUCCESS) {
  242. GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s",
  243. node->GetName().c_str(), node->GetType().c_str());
  244. return INTERNAL_ERROR;
  245. }
  246. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  247. if (ret != GRAPH_SUCCESS) {
  248. GELOGE(INTERNAL_ERROR, "Failed to connect node %s to trans node %s", node->GetName().c_str(),
  249. trans_node->GetName().c_str());
  250. return INTERNAL_ERROR;
  251. }
  252. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  253. if (ret != GRAPH_SUCCESS) {
  254. GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.",
  255. node->GetName().c_str(), trans_node->GetName().c_str());
  256. return INTERNAL_ERROR;
  257. }
  258. return SUCCESS;
  259. }
  260. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  261. NodePtr &trans_node) {
  262. GE_CHECK_NOTNULL(node);
  263. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  264. if (trans_node == nullptr) {
  265. return INTERNAL_ERROR;
  266. }
  267. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  268. if (ret != GRAPH_SUCCESS) {
  269. GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s",
  270. node->GetName().c_str(), node->GetType().c_str());
  271. return INTERNAL_ERROR;
  272. }
  273. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  274. if (ret != GRAPH_SUCCESS) {
  275. GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(),
  276. node->GetName().c_str());
  277. return INTERNAL_ERROR;
  278. }
  279. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  280. if (ret != GRAPH_SUCCESS) {
  281. GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.",
  282. node->GetName().c_str(), trans_node->GetName().c_str());
  283. return INTERNAL_ERROR;
  284. }
  285. return SUCCESS;
  286. }
  287. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  288. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  289. GE_CHECK_NOTNULL(var->GetOpDesc());
  290. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  291. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  292. output_desc.SetFormat(tensor_desc.GetFormat());
  293. output_desc.SetDataType(tensor_desc.GetDataType());
  294. output_desc.SetShape(tensor_desc.GetShape());
  295. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  296. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  297. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  298. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  299. GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed");
  300. return INTERNAL_ERROR;);
  301. }
  302. if (var->GetOpDesc()->GetInputsSize() > 0) {
  303. auto desc = var->GetOpDesc()->GetInputDesc(0);
  304. desc.SetFormat(tensor_desc.GetFormat());
  305. desc.SetDataType(tensor_desc.GetDataType());
  306. desc.SetShape(tensor_desc.GetShape());
  307. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  308. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  309. desc.SetOriginShape(tensor_desc.GetOriginShape());
  310. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  311. GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed");
  312. return INTERNAL_ERROR;)
  313. }
  314. return SUCCESS;
  315. }
  316. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  317. GE_CHECK_NOTNULL(var);
  318. int index = 0;
  319. NodePtr last_node = var;
  320. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  321. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  322. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  323. if (ret != SUCCESS) {
  324. ErrorManager::GetInstance().ATCReportErrMessage(
  325. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  326. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(),
  327. index, iter->node_type.c_str());
  328. return INTERNAL_ERROR;
  329. }
  330. // set stream_label
  331. OpDescPtr var_desc = var->GetOpDesc();
  332. GE_CHECK_NOTNULL(var_desc);
  333. std::string stream_label;
  334. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  335. if (!stream_label.empty()) {
  336. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  337. }
  338. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  339. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  340. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  341. }
  342. if (road.empty()) {
  343. return SUCCESS;
  344. }
  345. return UpdateVarFormats(var, road.rbegin()->output);
  346. }
  347. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  348. for (auto &var : nodes) {
  349. GE_CHECK_NOTNULL(var);
  350. int index = 0;
  351. NodePtr last_node = var;
  352. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  353. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  354. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  355. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  356. if (ret != SUCCESS) {
  357. ErrorManager::GetInstance().ATCReportErrMessage(
  358. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  359. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s",
  360. var->GetName().c_str(), index, iter->node_type.c_str());
  361. return INTERNAL_ERROR;
  362. }
  363. // set stream_label
  364. OpDescPtr var_desc = var->GetOpDesc();
  365. GE_CHECK_NOTNULL(var_desc);
  366. std::string stream_label;
  367. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  368. if (!stream_label.empty()) {
  369. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  370. }
  371. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  372. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  373. }
  374. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  375. return INTERNAL_ERROR;
  376. }
  377. }
  378. return SUCCESS;
  379. }
  380. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  381. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  382. VarNamesToRefs names_to_refs;
  383. std::string var_name;
  384. if (graph == nullptr) {
  385. GELOGE(PARAM_INVALID, "graph is null.");
  386. return names_to_refs;
  387. }
  388. for (auto &node : graph->GetAllNodes()) {
  389. if (node->GetType() != VARIABLE) {
  390. continue;
  391. }
  392. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  393. (void)names_to_refs[var_name].insert(node);
  394. }
  395. }
  396. return names_to_refs;
  397. }
  398. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  399. std::vector<int64_t> &dst_shape) {
  400. if (src_format == FORMAT_NCHW) {
  401. formats::FormatTransferNchwNc1hwc0 transfer;
  402. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  403. GELOGE(INTERNAL_ERROR, "TransShape failed");
  404. return FAILED;
  405. }
  406. } else if (src_format == FORMAT_NHWC) {
  407. formats::FormatTransferNhwcNc1hwc0 transfer;
  408. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  409. GELOGE(INTERNAL_ERROR, "TransShape failed");
  410. return FAILED;
  411. }
  412. }
  413. return SUCCESS;
  414. }
  415. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  416. GE_CHECK_NOTNULL(node_ptr);
  417. auto op_desc = node_ptr->GetOpDesc();
  418. GE_CHECK_NOTNULL(op_desc);
  419. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  420. GE_CHECK_NOTNULL(input);
  421. ge::Format old_format = input->GetFormat();
  422. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  423. ge::DataType dt = input->GetDataType();
  424. std::vector<int64_t> dst_shape_dims;
  425. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  426. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  427. return FAILED;
  428. }
  429. input->SetFormat(FORMAT_NC1HWC0);
  430. input->SetShape(ge::GeShape(dst_shape_dims));
  431. auto output = op_desc->MutableOutputDesc(0);
  432. GE_CHECK_NOTNULL(output);
  433. output->SetFormat(FORMAT_NC1HWC0);
  434. output->SetShape(ge::GeShape(dst_shape_dims));
  435. int64_t size = 0;
  436. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  437. if (graph_status != ge::GRAPH_SUCCESS) {
  438. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  439. return FAILED;
  440. }
  441. ge::TensorUtils::SetSize(*output, size);
  442. ge::TensorUtils::SetSize(*input, size);
  443. return SUCCESS;
  444. }
  445. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  446. GE_CHECK_NOTNULL(input_output);
  447. ge::Format old_format = input_output->GetFormat();
  448. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  449. ge::DataType dt = input_output->GetDataType();
  450. std::vector<int64_t> dst_shape_dims;
  451. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  452. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  453. return FAILED;
  454. }
  455. input_output->SetFormat(FORMAT_NC1HWC0);
  456. input_output->SetShape(ge::GeShape(dst_shape_dims));
  457. return SUCCESS;
  458. }
  459. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  460. vector<int64_t> &dst_shape_dims) {
  461. GE_CHECK_NOTNULL(op_desc);
  462. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  463. GE_CHECK_NOTNULL(input);
  464. ge::Format old_format = input->GetFormat();
  465. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  466. input->SetShape(ge::GeShape(dst_shape_dims));
  467. input->SetFormat(storage_format);
  468. auto output = op_desc->MutableOutputDesc(index);
  469. GE_CHECK_NOTNULL(output);
  470. output->SetShape(ge::GeShape(dst_shape_dims));
  471. output->SetFormat(storage_format);
  472. if (!output->MutableShape().IsUnknownShape()) {
  473. int64_t size = 0;
  474. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  475. if (graph_status != ge::GRAPH_SUCCESS) {
  476. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  477. return FAILED;
  478. }
  479. ge::TensorUtils::SetSize(*input, size);
  480. ge::TensorUtils::SetSize(*output, size);
  481. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  482. "new_shape:%s, new_format:%s, new_size:%lu",
  483. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  484. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  485. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  486. }
  487. return SUCCESS;
  488. }
  489. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) {
  490. is_dynamic_batch = false;
  491. std::string related_node_name;
  492. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  493. if (related_node_name.empty()) {
  494. ErrorManager::GetInstance().ATCReportErrMessage(
  495. "E15002", {"opname", "value", "reason"}, {data_node->GetName(), "flag", "but the value is empty"});
  496. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty",
  497. data_node->GetName().c_str());
  498. return INTERNAL_ERROR;
  499. }
  500. for (const NodePtr &next_node : data_node->GetOutNodes()) {
  501. if (next_node->GetName() == related_node_name) {
  502. switchn_node = next_node;
  503. break;
  504. }
  505. }
  506. if (switchn_node == nullptr) {
  507. ErrorManager::GetInstance().ATCReportErrMessage(
  508. "E15002", {"opname", "value", "reason"},
  509. {data_node->GetName(), related_node_name, "but can not find it on the graph"});
  510. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph",
  511. data_node->GetName().c_str(), related_node_name.c_str());
  512. return INTERNAL_ERROR;
  513. }
  514. is_dynamic_batch = true;
  515. }
  516. return SUCCESS;
  517. }
  518. bool CheckOpType(const NodePtr &node, const std::string type) {
  519. if (node->GetType() == type) {
  520. return true;
  521. }
  522. return false;
  523. }
  524. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  525. auto op = node_ptr->GetOpDesc();
  526. GE_CHECK_NOTNULL(op);
  527. auto inputDescsPtr = op->GetAllInputsDescPtr();
  528. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  529. ge::Format format = ge::FORMAT_ND;
  530. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  531. // process, otherwise fe will insert transdata.
  532. for (auto &inputDescPtr : inputDescsPtr) {
  533. GE_CHECK_NOTNULL(inputDescPtr);
  534. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  535. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  536. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  537. inputDescPtr->SetFormat(format);
  538. inputDescPtr->SetOriginFormat(format);
  539. }
  540. }
  541. for (auto &outputDescPtr : outputDescsPtr) {
  542. GE_CHECK_NOTNULL(outputDescPtr);
  543. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  544. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  545. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  546. outputDescPtr->SetFormat(format);
  547. outputDescPtr->SetOriginFormat(format);
  548. }
  549. }
  550. return SUCCESS;
  551. }
  552. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  553. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  554. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  555. // data and netoutput, and no longer inserts the transnode.
  556. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) {
  557. GE_CHECK_NOTNULL(node_ptr);
  558. auto op_desc = node_ptr->GetOpDesc();
  559. GE_CHECK_NOTNULL(op_desc);
  560. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  561. GE_CHECK_NOTNULL(input);
  562. ge::DataType src_dtype = input->GetDataType();
  563. if (src_dtype == dt_set) {
  564. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  565. return SUCCESS;
  566. }
  567. input->SetDataType(dt_set);
  568. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  569. GE_CHECK_NOTNULL(output);
  570. output->SetDataType(dt_set);
  571. GeShape shape = input->GetShape();
  572. if (!shape.IsUnknownShape()) {
  573. int64_t input_shape_size = 0;
  574. int64_t output_shape_size = 0;
  575. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  576. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  577. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  578. GELOGE(GRAPH_FAILED, "[Process][InputOp] Get tensor size of op [%s] failed!", node_ptr->GetName().c_str());
  579. return FAILED;
  580. }
  581. ge::TensorUtils::SetSize(*input, input_shape_size);
  582. ge::TensorUtils::SetSize(*output, output_shape_size);
  583. GELOGI("[Process][InputDynShape] Set input and output size of node [%s] success.", node_ptr->GetName().c_str());
  584. }
  585. if (is_dynamic_batch) {
  586. GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str());
  587. auto switchn_op_desc = switchn_node->GetOpDesc();
  588. GE_CHECK_NOTNULL(switchn_op_desc);
  589. auto switchn_input = switchn_op_desc->MutableInputDesc(0);
  590. GE_CHECK_NOTNULL(switchn_input);
  591. switchn_input->SetDataType(dt_set);
  592. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  593. const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i);
  594. GE_CHECK_NOTNULL(switchn_output);
  595. switchn_output->SetDataType(dt_set);
  596. }
  597. }
  598. return SUCCESS;
  599. }
  600. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  601. GE_CHECK_NOTNULL(node_ptr);
  602. auto op_desc = node_ptr->GetOpDesc();
  603. GE_CHECK_NOTNULL(op_desc);
  604. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  605. GE_CHECK_NOTNULL(input);
  606. ge::Format old_format = input->GetFormat();
  607. ge::GeShape old_shape = input->GetShape();
  608. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  609. if (!support) {
  610. ErrorManager::GetInstance().ATCReportErrMessage(
  611. "E19014", {"opname", "value", "reason"},
  612. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  613. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  614. GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str());
  615. return FAILED;
  616. }
  617. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  618. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  619. return FAILED;
  620. }
  621. if (is_dynamic_batch) {
  622. auto switchn_op_desc = switchn_node->GetOpDesc();
  623. GE_CHECK_NOTNULL(switchn_op_desc);
  624. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  625. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  626. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  627. return FAILED;
  628. }
  629. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  630. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  631. GE_CHECK_NOTNULL(switchn_output);
  632. old_format = switchn_output->GetFormat();
  633. old_shape = switchn_output->GetShape();
  634. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  635. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  636. return FAILED;
  637. }
  638. }
  639. }
  640. return SUCCESS;
  641. }
  642. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  643. auto op_desc = node_ptr->GetOpDesc();
  644. GE_CHECK_NOTNULL(op_desc);
  645. string set_dt_str;
  646. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  647. return SUCCESS;
  648. }
  649. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  650. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  651. bool is_dynamic_batch = false;
  652. NodePtr switchn_node = nullptr;
  653. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) {
  654. GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed");
  655. return FAILED;
  656. }
  657. if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) {
  658. GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed");
  659. return FAILED;
  660. }
  661. // check if need to set format
  662. string set_format;
  663. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  664. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  665. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  666. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) {
  667. GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed");
  668. return FAILED;
  669. }
  670. }
  671. return SUCCESS;
  672. }
  673. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  674. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  675. GE_CHECK_NOTNULL(op_desc);
  676. GE_CHECK_NOTNULL(tensor_desc_ptr);
  677. storage_format = FORMAT_RESERVED;
  678. int64_t format = FORMAT_RESERVED;
  679. dst_shape_dims.clear();
  680. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  681. storage_format = static_cast<Format>(format);
  682. vector<int32_t> storage_shape;
  683. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  684. for (auto dim : storage_shape) {
  685. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  686. }
  687. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  688. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  689. formats::JoinToString(storage_shape).c_str());
  690. } else {
  691. ErrorManager::GetInstance().ATCReportErrMessage(
  692. "15003", {"opname", "format"},
  693. {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)});
  694. GELOGE(PARAM_INVALID, "Update node by storage format failed, storage_shape not set. "
  695. "node: [%s], storage_format [%s]",
  696. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  697. return FAILED;
  698. }
  699. ge::Format old_format = tensor_desc_ptr->GetFormat();
  700. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  701. if (old_format == storage_format && old_shape == dst_shape_dims) {
  702. GELOGI("Update node by storage format, not changed.");
  703. storage_format = FORMAT_RESERVED;
  704. return SUCCESS;
  705. }
  706. }
  707. return SUCCESS;
  708. }
  709. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  710. NodePtr &node) {
  711. bool is_dynamic = CheckOpType(node, MERGE);
  712. auto src_op_desc = node->GetOpDesc();
  713. GE_CHECK_NOTNULL(src_op_desc);
  714. ge::GeShape src_shape = src_desc.GetShape();
  715. ge::Format src_format = src_desc.GetFormat();
  716. net_output_input_desc->SetDataType(DT_FLOAT16);
  717. if (is_dynamic) {
  718. auto merge_output = src_op_desc->MutableOutputDesc(0);
  719. GE_CHECK_NOTNULL(merge_output);
  720. merge_output->SetDataType(DT_FLOAT16);
  721. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  722. auto merge_input = src_op_desc->MutableInputDesc(i);
  723. GE_CHECK_NOTNULL(merge_input);
  724. merge_input->SetDataType(DT_FLOAT16);
  725. }
  726. }
  727. std::vector<int64_t> dst_shape_dims;
  728. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  729. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  730. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  731. return FAILED;
  732. }
  733. ge::GeShape dst_shape(dst_shape_dims);
  734. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  735. net_output_input_desc->SetShape(dst_shape);
  736. if (is_dynamic) {
  737. auto merge_out = src_op_desc->MutableOutputDesc(0);
  738. GE_CHECK_NOTNULL(merge_out);
  739. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  740. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  741. return FAILED;
  742. }
  743. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  744. auto merge_in = src_op_desc->MutableInputDesc(i);
  745. GE_CHECK_NOTNULL(merge_in);
  746. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  747. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  748. return FAILED;
  749. }
  750. }
  751. }
  752. return SUCCESS;
  753. }
  754. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  755. GE_CHECK_NOTNULL(netout_desc);
  756. vector<string> output_dt_str;
  757. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  758. for (auto dt_str : output_dt_str) {
  759. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  760. if (dt_str_split.size() == kUserDefinedElementCount) {
  761. if (dt_str_split[0] == to_string(index)) {
  762. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  763. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  764. TypeUtils::DataTypeToSerialString(dt).c_str());
  765. return true;
  766. }
  767. }
  768. }
  769. }
  770. return false;
  771. }
  772. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  773. GE_CHECK_NOTNULL(netout_desc);
  774. vector<string> output_format_str;
  775. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  776. for (auto format_str : output_format_str) {
  777. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  778. if (format_str_split.size() == kUserDefinedElementCount) {
  779. if (format_str_split[0] == to_string(index)) {
  780. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  781. return true;
  782. }
  783. }
  784. }
  785. }
  786. return false;
  787. }
  788. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  789. auto op_desc = node->GetOpDesc();
  790. GE_CHECK_NOTNULL(op_desc);
  791. ge::DataType output_data_type = ge::DT_FLOAT;
  792. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  793. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  794. auto peer_out = in_anchor->GetPeerOutAnchor();
  795. GE_CHECK_NOTNULL(peer_out);
  796. auto src_node = peer_out->GetOwnerNode();
  797. GE_CHECK_NOTNULL(src_node);
  798. bool is_dynamic = CheckOpType(src_node, MERGE);
  799. OpDescPtr src_op_desc = src_node->GetOpDesc();
  800. GE_CHECK_NOTNULL(src_op_desc);
  801. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  802. GE_CHECK_NOTNULL(net_output_input_desc);
  803. ge::GeShape old_shape = net_output_input_desc->GetShape();
  804. ge::Format old_format = net_output_input_desc->GetFormat();
  805. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  806. // Update datatype
  807. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  808. GELOGI("Enter into process output_type schedule");
  809. net_output_input_desc->SetDataType(output_data_type);
  810. if (is_dynamic) {
  811. auto merge_output = src_op_desc->MutableOutputDesc(0);
  812. GE_CHECK_NOTNULL(merge_output);
  813. merge_output->SetDataType(output_data_type);
  814. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  815. auto merge_input = src_op_desc->MutableInputDesc(i);
  816. GE_CHECK_NOTNULL(merge_input);
  817. merge_input->SetDataType(output_data_type);
  818. }
  819. }
  820. }
  821. // check if is_output_adjust_hw_layout is set
  822. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  823. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  824. ErrorManager::GetInstance().ATCReportErrMessage(
  825. "E19014", {"opname", "value", "reason"},
  826. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  827. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  828. GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0.");
  829. return FAILED;
  830. }
  831. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  832. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  833. GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0.");
  834. return FAILED;
  835. }
  836. }
  837. }
  838. return SUCCESS;
  839. }
  840. long StringToLongNoThrow(const string &str) {
  841. try {
  842. return std::stol(str);
  843. } catch (const std::invalid_argument) {
  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. } catch (const std::out_of_range) {
  850. GELOGE(PARAM_INVALID,
  851. "Parse shape range of input failed when transfer from string to int64. Given %s, while correct example: "
  852. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  853. str.c_str());
  854. return PARAM_INVALID;
  855. }
  856. }
  857. /**
  858. * Parser shape_range from string to vector
  859. * shape_range from option normally is "[1~20,3,3~6,-1],[1~20,3,3~6,-1]"
  860. * @param shape_range
  861. */
  862. Status ParseDynamicInputShapeRange(const std::string &shape_range,
  863. std::vector<std::vector<std::pair<int64_t, int64_t>>> &range) {
  864. if (shape_range.size() < 2) {
  865. GELOGE(PARAM_INVALID, "Shape range %s is invalid.", shape_range.c_str());
  866. return PARAM_INVALID;
  867. }
  868. // different shape_range of single input are split by ']'
  869. vector<string> shape_range_set = ge::StringUtils::Split(shape_range, ']');
  870. if (shape_range_set.empty()) {
  871. GELOGE(PARAM_INVALID, "Shape range %s is not valid. Correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  872. shape_range.c_str());
  873. return PARAM_INVALID;
  874. }
  875. for (auto &shape_range_str : shape_range_set) {
  876. if (shape_range_str.size() < 3) {
  877. // shape_range_str should be "[2~3,1"
  878. // or ",[2~3,1". because we should trim '[' or ',['
  879. // so shape_range_str.size() < 3 is invalid
  880. continue;
  881. }
  882. // trim start bytes, after that, single input should be "1~20,3,3~6,-1"
  883. if (ge::StringUtils::StartWith(shape_range_str, "[")) {
  884. shape_range_str = shape_range_str.substr(1, shape_range_str.size());
  885. }
  886. if (ge::StringUtils::StartWith(shape_range_str, ",")) {
  887. shape_range_str = shape_range_str.substr(2, shape_range_str.size());
  888. }
  889. // parse shape_range of single input. eg. "1~20,3,3~6,-1"
  890. std::vector<std::pair<int64_t, int64_t>> range_of_single_input;
  891. vector<string> dim_range_set = ge::StringUtils::Split(shape_range_str, ',');
  892. for (const auto &range_pair_str : dim_range_set) {
  893. vector<string> range_pair_set = ge::StringUtils::Split(range_pair_str, '~');
  894. pair<int64_t, int64_t> range_pair;
  895. if (range_pair_set.size() == 1) {
  896. // fix dim
  897. auto range_value = StringToLongNoThrow(range_pair_set.at(0).c_str());
  898. if (range_value < 0) {
  899. range_pair = std::make_pair(1, range_value);
  900. } else {
  901. range_pair = std::make_pair(range_value, range_value);
  902. }
  903. } else if (range_pair_set.size() == 2) {
  904. // unknown dim, should get range.
  905. auto range_left = StringToLongNoThrow(range_pair_set.at(0).c_str());
  906. auto range_right = StringToLongNoThrow(range_pair_set.at(1).c_str());
  907. if (range_left < 0 || range_right < 0) {
  908. GELOGE(PARAM_INVALID,
  909. "Shape range of input is invalid. Given range pair [%ld,%ld], while correct example: "
  910. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  911. range_left, range_right);
  912. return PARAM_INVALID;
  913. }
  914. range_pair = std::make_pair(range_left, range_right);
  915. } else {
  916. GELOGE(PARAM_INVALID,
  917. "Shape range of input is invalid. Given %s, while correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  918. shape_range.c_str());
  919. return PARAM_INVALID;
  920. }
  921. range_of_single_input.emplace_back(range_pair);
  922. }
  923. range.emplace_back(range_of_single_input);
  924. }
  925. return SUCCESS;
  926. }
  927. Status GetDynamicInputShapeRange(const std::vector<GeTensor> &user_input, const std::map<string, string> &graph_option,
  928. vector<vector<std::pair<int64_t, int64_t>>> &range_vec) {
  929. // check both mode and shape_range option are all enabled
  930. auto mode_iter = graph_option.find(OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  931. bool enable_dynamic_execute_mode = (mode_iter != graph_option.end()) && (mode_iter->second == "dynamic_execute");
  932. if (!enable_dynamic_execute_mode) {
  933. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  934. OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  935. }
  936. auto iter = graph_option.find(OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  937. bool enable_input_shape_range = (iter != graph_option.end()) && (!iter->second.empty());
  938. if (!enable_input_shape_range) {
  939. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  940. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  941. }
  942. if (enable_dynamic_execute_mode && enable_input_shape_range) {
  943. GELOGD("GraphOption: %s value is dynamic_execute, %s value is %s.", OPTION_EXEC_DYNAMIC_EXECUTE_MODE,
  944. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE, iter->second.c_str());
  945. } else if (!enable_dynamic_execute_mode && !enable_input_shape_range) {
  946. return SUCCESS;
  947. } else {
  948. GELOGE(PARAM_INVALID, "Graph option: %s and %s should be enabled at the same time.",
  949. OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  950. return PARAM_INVALID;
  951. }
  952. auto ret = ParseDynamicInputShapeRange(iter->second, range_vec);
  953. GE_CHK_STATUS_RET(ret, "Parse dynamic input shape range failed.");
  954. if (range_vec.size() != user_input.size()) {
  955. GELOGE(PARAM_INVALID, "Dynamic input shape range size is %zu, inputs size is %zu. Not match.", range_vec.size(),
  956. user_input.size());
  957. return PARAM_INVALID;
  958. }
  959. return SUCCESS;
  960. }
  961. Status UpdateDynamicInputShapeRange(const ge::GeAttrValue::INT index,
  962. const vector<vector<std::pair<int64_t, int64_t>>> &range_vec, OpDescPtr &op,
  963. GeTensorDesc &desc) {
  964. auto origin_shape = desc.GetShape();
  965. auto current_shape_range_vec = range_vec.at(index);
  966. if (current_shape_range_vec.size() != origin_shape.GetDimNum()) {
  967. GELOGE(PARAM_INVALID, "Given shape_range dim num is %zu, current dim num is %zu, not match.Pleace Check.",
  968. current_shape_range_vec.size(), origin_shape.GetDimNum());
  969. return PARAM_INVALID;
  970. }
  971. for (size_t i = 0; i < origin_shape.GetDimNum(); ++i) {
  972. auto curr_dim = origin_shape.GetDim(i);
  973. auto left_range = current_shape_range_vec.at(i).first;
  974. auto right_range = current_shape_range_vec.at(i).second;
  975. if (left_range == right_range) {
  976. // given shape_range is known dim, check is same as origin or not
  977. if (curr_dim != left_range) {
  978. GELOGE(PARAM_INVALID, "Given shape range is %ld, current dim shape is %ld, not match.Pleace Check.",
  979. left_range, curr_dim);
  980. return PARAM_INVALID;
  981. }
  982. origin_shape.SetDim(i, left_range);
  983. } else {
  984. // given shape_range is fix range, check input_shape is in this range or not
  985. if (right_range != UNKNOWN_DIM) {
  986. if ((curr_dim < left_range) || (curr_dim > right_range)) {
  987. GELOGE(PARAM_INVALID, "Given shape range is [%ld~%ld], current dim shape is %ld, out of range.Pleace Check.",
  988. left_range, right_range, curr_dim);
  989. return PARAM_INVALID;
  990. }
  991. }
  992. origin_shape.SetDim(i, UNKNOWN_DIM);
  993. }
  994. }
  995. desc.SetShape(origin_shape);
  996. desc.SetShapeRange(current_shape_range_vec);
  997. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  998. GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  999. graph_ret = op->UpdateOutputDesc(0, desc);
  1000. GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  1001. return SUCCESS;
  1002. }
  1003. } // namespace
  1004. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  1005. GraphPrepare::~GraphPrepare() {}
  1006. /**
  1007. * @param graph
  1008. * @return
  1009. */
  1010. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  1011. GE_CHECK_NOTNULL(graph);
  1012. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  1013. for (auto &node : graph->GetAllNodes()) {
  1014. if (node == nullptr) {
  1015. continue;
  1016. }
  1017. if (node->GetType() != VARIABLE) {
  1018. continue;
  1019. }
  1020. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  1021. if (trans_road == nullptr) {
  1022. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  1023. continue;
  1024. }
  1025. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  1026. auto ret = RecoverTransRoadForVar(node, *trans_road);
  1027. if (ret != SUCCESS) {
  1028. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str());
  1029. return INTERNAL_ERROR;
  1030. }
  1031. auto iter = var_names_to_refs.find(node->GetName());
  1032. if (iter != var_names_to_refs.end()) {
  1033. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  1034. if (ret != SUCCESS) {
  1035. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str());
  1036. return INTERNAL_ERROR;
  1037. }
  1038. }
  1039. }
  1040. return SUCCESS;
  1041. }
  1042. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  1043. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  1044. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  1045. if (compute_graph_ != nullptr) {
  1046. compute_graph_->SetSessionID(session_id);
  1047. }
  1048. session_id_ = session_id;
  1049. Status ret = CheckGraph();
  1050. if (ret != SUCCESS) {
  1051. GELOGE(ret, "RunGraph graph check fail, ret:%u", ret);
  1052. return ret;
  1053. }
  1054. (void)compute_graph_->TopologicalSorting();
  1055. ret = CheckRefOp();
  1056. if (ret != SUCCESS) {
  1057. GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret);
  1058. return ret;
  1059. }
  1060. return SUCCESS;
  1061. }
  1062. Status GraphPrepare::CheckGraph() {
  1063. if (compute_graph_ == nullptr) {
  1064. GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR");
  1065. return GE_GRAPH_INIT_FAILED;
  1066. }
  1067. auto nodes = compute_graph_->GetAllNodes();
  1068. if (nodes.empty()) {
  1069. GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph.");
  1070. return GE_GRAPH_INIT_FAILED;
  1071. }
  1072. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  1073. GE_CHECK_NOTNULL(node);
  1074. if (node->GetOpDesc() == nullptr) {
  1075. GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL");
  1076. return GE_GRAPH_INIT_FAILED;
  1077. }
  1078. }
  1079. return SUCCESS;
  1080. }
  1081. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  1082. const std::set<NodePtr> &ref_nodes) {
  1083. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  1084. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  1085. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  1086. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  1087. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  1088. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH};
  1089. GE_CHECK_NOTNULL(node);
  1090. const auto &op_desc = node->GetOpDesc();
  1091. GE_CHECK_NOTNULL(op_desc);
  1092. const auto input_index = op_desc->GetInputIndexByName(input_name);
  1093. const auto &in_anchor = node->GetInDataAnchor(input_index);
  1094. GE_CHECK_NOTNULL(in_anchor);
  1095. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  1096. GE_CHECK_NOTNULL(peer_out_anchor);
  1097. const auto &input_node = peer_out_anchor->GetOwnerNode();
  1098. GE_CHECK_NOTNULL(input_node);
  1099. const auto &input_op_desc = input_node->GetOpDesc();
  1100. GE_CHECK_NOTNULL(input_op_desc);
  1101. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  1102. if (is_ref) {
  1103. return SUCCESS;
  1104. }
  1105. auto input_type = input_op_desc->GetType();
  1106. if (input_type == ge::FRAMEWORKOP) {
  1107. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  1108. GELOGE(PARAM_INVALID, "Get original type failed.");
  1109. return PARAM_INVALID;
  1110. }
  1111. }
  1112. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  1113. if (!is_acceptable) {
  1114. ErrorManager::GetInstance().ATCReportErrMessage(
  1115. "E15005", {"opname", "optype", "opname1", "optype1"},
  1116. {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()});
  1117. GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.",
  1118. node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  1119. input_op_desc->GetType().c_str());
  1120. return PARAM_INVALID;
  1121. }
  1122. return SUCCESS;
  1123. }
  1124. Status GraphPrepare::CheckRefOp() {
  1125. GE_CHECK_NOTNULL(compute_graph_);
  1126. std::set<NodePtr> ref_nodes;
  1127. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  1128. if (node == nullptr) {
  1129. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  1130. return PARAM_INVALID;
  1131. }
  1132. auto op_desc = node->GetOpDesc();
  1133. if (op_desc == nullptr) {
  1134. GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null.");
  1135. return PARAM_INVALID;
  1136. }
  1137. auto input_name_index = op_desc->GetAllInputName();
  1138. auto outputs = op_desc->GetAllOutputName();
  1139. for (const auto &name_index : input_name_index) {
  1140. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  1141. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  1142. GELOGE(PARAM_INVALID, "CheckRefInputNode failed.");
  1143. return PARAM_INVALID;
  1144. }
  1145. (void)ref_nodes.insert(node); // no need to check value
  1146. }
  1147. }
  1148. }
  1149. return SUCCESS;
  1150. };
  1151. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  1152. GE_CHECK_NOTNULL(compute_graph_);
  1153. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  1154. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  1155. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  1156. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  1157. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  1158. return SUCCESS;
  1159. }
  1160. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1161. if (node == nullptr) {
  1162. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL");
  1163. return GE_GRAPH_GRAPH_NODE_NULL;
  1164. }
  1165. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1166. if (op_desc_ptr == nullptr) {
  1167. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL");
  1168. return GE_GRAPH_GRAPH_NODE_NULL;
  1169. }
  1170. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1171. GeShape output_shape = output.GetShape();
  1172. if (output_shape.IsUnknownShape()) {
  1173. GELOGD("[Adjust][DataOpOutput] Shape of op [%s] output is unknown.", node->GetName().c_str());
  1174. return SUCCESS;
  1175. }
  1176. int64_t tensor_size = 0;
  1177. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1178. if (graph_status != GRAPH_SUCCESS) {
  1179. ErrorManager::GetInstance().ATCReportErrMessage(
  1180. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1181. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  1182. return FAILED;
  1183. }
  1184. TensorUtils::SetSize(output, tensor_size);
  1185. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1186. if (graph_ret != GRAPH_SUCCESS) {
  1187. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1188. return graph_ret;
  1189. }
  1190. return SUCCESS;
  1191. }
  1192. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input,
  1193. const std::map<string, string> &graph_option) {
  1194. // Get shape range of input in dynamic_execute mode
  1195. vector<vector<std::pair<int64_t, int64_t>>> dynamic_shape_range_vec;
  1196. auto ret = GetDynamicInputShapeRange(user_input, graph_option, dynamic_shape_range_vec);
  1197. GE_CHK_STATUS_RET(ret, "Graph option is not right on Dynamic execute mode.");
  1198. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1199. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1200. GE_CHECK_NOTNULL(input_node);
  1201. OpDescPtr op = input_node->GetOpDesc();
  1202. GE_CHECK_NOTNULL(op);
  1203. if (op->GetType() == DATA) {
  1204. GeAttrValue::INT index = 0;
  1205. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1206. GELOGW("Get index from data attr failed");
  1207. continue;
  1208. }
  1209. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1210. std::string situation = "data op index[" + std::to_string(index) + "]";
  1211. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1212. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1213. GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index);
  1214. return FAILED;
  1215. }
  1216. if (IsDynamicDims(input_node)) {
  1217. continue;
  1218. }
  1219. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1220. auto format = desc.GetFormat();
  1221. auto origin_format = desc.GetOriginFormat();
  1222. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1223. auto tune_flag = (options_.build_mode == BUILD_MODE_TUNING) && (options_.build_step == BUILD_STEP_AFTER_BUILDER);
  1224. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op) && (!tune_flag);
  1225. if (need_check_internal_format) {
  1226. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1227. if (is_internal) {
  1228. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1229. {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" +
  1230. TypeUtils::FormatToSerialString(origin_format) + "]", "it is not support"});
  1231. GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.",
  1232. TypeUtils::FormatToSerialString(format).c_str(),
  1233. TypeUtils::FormatToSerialString(origin_format).c_str());
  1234. return FAILED;
  1235. }
  1236. }
  1237. auto data_type = desc.GetDataType();
  1238. uint32_t length = 1;
  1239. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1240. if (!type_ret) {
  1241. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1242. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1243. GELOGE(PARAM_INVALID, "Input datatype %s is not support.",
  1244. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1245. return FAILED;
  1246. }
  1247. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1248. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1249. int64_t shape_size = desc_shape * length;
  1250. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1251. int64_t size = 0;
  1252. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1253. GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed");
  1254. return FAILED);
  1255. bool size_check = (size != 0 && shape_size != size);
  1256. if (size_check) {
  1257. std::string situation = "input data size[" + std::to_string(size) +
  1258. "] and shape_size[" + std::to_string(size) + "]";
  1259. std::string reason = "because size != 0 and shape_size != size";
  1260. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1261. GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size);
  1262. return FAILED;
  1263. }
  1264. ge::TensorUtils::SetSize(desc, shape_size);
  1265. if (!tune_flag) {
  1266. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1267. if (graph_ret != GRAPH_SUCCESS) {
  1268. GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
  1269. return graph_ret;
  1270. }
  1271. // Size will be recalculated in the build stage
  1272. ge::TensorUtils::SetSize(desc, 0);
  1273. graph_ret = op->UpdateOutputDesc(0, desc);
  1274. if (graph_ret != GRAPH_SUCCESS) {
  1275. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1276. return graph_ret;
  1277. }
  1278. } else {
  1279. GELOGI("data %s skip update info in tune mode", op->GetName().c_str());
  1280. }
  1281. if (!dynamic_shape_range_vec.empty()) {
  1282. ret = UpdateDynamicInputShapeRange(index, dynamic_shape_range_vec, op, desc);
  1283. GE_CHK_STATUS_RET(ret, "Fail to update dynamic input shape range on %s.", op->GetName().c_str());
  1284. continue;
  1285. }
  1286. if (!options_.train_graph_flag) {
  1287. Status ret = AdjustDataOpOutput(input_node);
  1288. GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret); return ret);
  1289. }
  1290. }
  1291. }
  1292. return SUCCESS;
  1293. }
  1294. Status GraphPrepare::TryDoAipp() {
  1295. // infer and with aipp configure file, then call aipp insert
  1296. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1297. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1298. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1299. if (ret != SUCCESS) {
  1300. GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed.");
  1301. return INTERNAL_ERROR;
  1302. }
  1303. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1304. if (ret != SUCCESS) {
  1305. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed",
  1306. options_.insert_op_file.c_str());
  1307. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1308. }
  1309. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1310. if (ret != SUCCESS) {
  1311. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret);
  1312. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1313. }
  1314. }
  1315. return SUCCESS;
  1316. }
  1317. Status GraphPrepare::FormatAndShapeProcess() {
  1318. Status ret = ResourcePairProcess("add");
  1319. if (ret != SUCCESS) {
  1320. GELOGE(ret, "ResourcePairProcess failed");
  1321. return ret;
  1322. }
  1323. GE_TIMESTAMP_START(InferOriginFormat1);
  1324. ret = compute_graph_->InferOriginFormat();
  1325. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1326. GE_DUMP(compute_graph_, "after_first_inferformat");
  1327. if (ret != SUCCESS) {
  1328. GELOGE(ret, "Prepare Graph first inferformat failed");
  1329. return ret;
  1330. }
  1331. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1332. ret = InferShapeForPreprocess();
  1333. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1334. GE_DUMP(compute_graph_, "after_infershape");
  1335. if (ret != SUCCESS) {
  1336. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed");
  1337. return GE_GRAPH_INFERSHAPE_FAILED;
  1338. }
  1339. GE_TIMESTAMP_START(InferOriginFormat2);
  1340. ret = compute_graph_->InferOriginFormat();
  1341. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1342. if (ret != SUCCESS) {
  1343. GELOGE(ret, "Prepare Graph inferformat failed");
  1344. return ret;
  1345. }
  1346. ret = ResourcePairProcess("remove");
  1347. if (ret != SUCCESS) {
  1348. return ret;
  1349. }
  1350. return ret;
  1351. }
  1352. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1353. PassManager control_pass;
  1354. // Graph pass tmp logic for resource infershape
  1355. if (options_.train_graph_flag) {
  1356. try {
  1357. if (action == "add") {
  1358. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1359. } else {
  1360. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1361. new ResourcePairRemoveControlPass);
  1362. }
  1363. } catch (std::bad_alloc &e) {
  1364. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str());
  1365. return INTERNAL_ERROR;
  1366. }
  1367. }
  1368. Status ret = control_pass.Run(compute_graph_);
  1369. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1370. GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret);
  1371. return ret;
  1372. }
  1373. return SUCCESS;
  1374. }
  1375. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1376. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1377. GE_CHECK_NOTNULL(node_ptr);
  1378. if (node_ptr->GetType() == DATA) {
  1379. uint32_t index = 0;
  1380. auto op_desc = node_ptr->GetOpDesc();
  1381. GE_CHECK_NOTNULL(op_desc);
  1382. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1383. Format storage_format = FORMAT_RESERVED;
  1384. vector<int64_t> dst_shape_dims;
  1385. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1386. GELOGE(INTERNAL_ERROR, "Get storage format for input failed");
  1387. return FAILED;
  1388. }
  1389. if (storage_format == FORMAT_RESERVED) {
  1390. continue;
  1391. }
  1392. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1393. GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed");
  1394. return FAILED;
  1395. }
  1396. }
  1397. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1398. auto op_desc = node_ptr->GetOpDesc();
  1399. GE_CHECK_NOTNULL(op_desc);
  1400. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1401. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1402. Format storage_format = FORMAT_RESERVED;
  1403. vector<int64_t> dst_shape_dims;
  1404. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1405. GELOGE(INTERNAL_ERROR, "Get storage format from output failed");
  1406. return FAILED;
  1407. }
  1408. if (storage_format == FORMAT_RESERVED) {
  1409. continue;
  1410. }
  1411. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1412. GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed");
  1413. return FAILED;
  1414. }
  1415. }
  1416. }
  1417. }
  1418. return SUCCESS;
  1419. }
  1420. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1421. if (options_.save_original_model == "true") {
  1422. ModelHelper model_helper;
  1423. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1424. options_.original_model_file);
  1425. if (ret != SUCCESS) {
  1426. // If save original model fail, process continue
  1427. GELOGW("SaveOriginalGraphToOmModel fail");
  1428. }
  1429. }
  1430. return SUCCESS;
  1431. }
  1432. #define PP_RUN_AND_DUMP(name, func, ...) \
  1433. do { \
  1434. GE_RUN(Prepare, func, __VA_ARGS__); \
  1435. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1436. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1437. } while (0)
  1438. #define PP_RUN(name, func, ...) \
  1439. do { \
  1440. GE_RUN(Prepare, func, __VA_ARGS__); \
  1441. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1442. } while (0)
  1443. Status GraphPrepare::PrepareDynShape(const GraphNodePtr &graph_node, const std::vector<GeTensor> &user_input,
  1444. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1445. GE_CHECK_NOTNULL(graph_node->GetGraph());
  1446. GE_CHECK_NOTNULL(compute_graph);
  1447. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1448. const Graph &const_graph = *graph_node->GetGraph();
  1449. PP_RUN("Init", Init, const_graph, session_id);
  1450. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1451. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input, graph_node->GetOptions());
  1452. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1453. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1454. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1455. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1456. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1457. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1458. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1459. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1460. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1461. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1462. return SUCCESS;
  1463. }
  1464. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1465. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1466. return SUCCESS;
  1467. }
  1468. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1469. auto compute_graph = compute_graph_;
  1470. PassManager pass_manager;
  1471. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1472. new (std::nothrow) VariablePrepareOpPass))
  1473. GE_TIMESTAMP_START(pass_manager);
  1474. auto ret = pass_manager.Run(compute_graph);
  1475. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1476. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1477. GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret);
  1478. return ret;
  1479. }
  1480. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1481. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1482. return SUCCESS;
  1483. }
  1484. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1485. if (compute_graph == nullptr) {
  1486. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1487. return GE_GRAPH_NULL_INPUT;
  1488. }
  1489. GEPass ge_passes(compute_graph);
  1490. NamesToPass hccl_group;
  1491. HcclGroupPass hccl_group_pass;
  1492. GELOGD("Add hccl group pass success");
  1493. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1494. auto ret = ge_passes.Run(hccl_group);
  1495. if (ret != SUCCESS) {
  1496. GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret);
  1497. return ret;
  1498. }
  1499. ret = compute_graph->TopologicalSorting();
  1500. if (ret != SUCCESS) {
  1501. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1502. return ret;
  1503. }
  1504. return SUCCESS;
  1505. }
  1506. #undef PP_RUN_AND_DUMP
  1507. #undef PP_RUN
  1508. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1509. if (graph == nullptr) {
  1510. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1511. return GE_GRAPH_NULL_INPUT;
  1512. }
  1513. const Graph &const_graph = *graph;
  1514. Status ret = Init(const_graph, 0);
  1515. if (ret != SUCCESS) {
  1516. GELOGE(ret, "Init graph_prepare fail, ret:%u", ret);
  1517. return ret;
  1518. }
  1519. GE_DUMP(compute_graph_, "after_parser");
  1520. GELOGI("Start infershape for dump json process.");
  1521. ret = compute_graph_->InferOriginFormat();
  1522. GE_DUMP(compute_graph_, "after_inferformat");
  1523. if (ret != SUCCESS) {
  1524. GELOGE(ret, "Prepare Graph inferformat failed");
  1525. return ret;
  1526. }
  1527. InferShapePass infer_shape_pass;
  1528. NamesToPass names_to_passes;
  1529. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1530. GEPass ge_passes(compute_graph_);
  1531. ret = ge_passes.Run(names_to_passes);
  1532. GE_DUMP(compute_graph_, "after_infershape");
  1533. if (ret != SUCCESS) {
  1534. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1535. return ret;
  1536. }
  1537. ShapeRefiner::ClearContextMap();
  1538. return SUCCESS;
  1539. }
  1540. Status GraphPrepare::CheckConstOp() {
  1541. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1542. GE_CHECK_NOTNULL(node_ptr);
  1543. if (node_ptr->GetType() == CONSTANT) {
  1544. Status ret = VerifyConstOp(node_ptr);
  1545. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1546. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1547. auto op_desc = node_ptr->GetOpDesc();
  1548. if (op_desc == nullptr) {
  1549. GELOGE(PARAM_INVALID, "Get op desc failed");
  1550. return PARAM_INVALID;
  1551. }
  1552. std::string original_type;
  1553. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1554. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1555. GELOGI("original type is %s", original_type.c_str());
  1556. if (original_type == CONSTANT) {
  1557. Status ret = VerifyConstOp(node_ptr);
  1558. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1559. }
  1560. }
  1561. }
  1562. return SUCCESS;
  1563. }
  1564. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1565. GE_CHECK_NOTNULL(node);
  1566. auto op_desc = node->GetOpDesc();
  1567. GE_CHECK_NOTNULL(op_desc);
  1568. ConstGeTensorPtr ge_tensor_ptr;
  1569. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1570. GELOGE(PARAM_INVALID, "Get value from const attr failed");
  1571. return PARAM_INVALID;
  1572. }
  1573. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1574. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1575. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1576. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1577. auto data_type = ge_tensor_desc.GetDataType();
  1578. if (data_type == DT_STRING) {
  1579. return SUCCESS;
  1580. }
  1581. uint32_t length = 1;
  1582. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1583. if (!type_ret) {
  1584. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1585. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1586. GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str());
  1587. return FAILED;
  1588. }
  1589. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1590. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1591. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1592. if (shape_size == 0) {
  1593. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1594. // shape = [], means it's a sclar tensor.
  1595. GE_CHK_BOOL_EXEC(data_size / length == 1,
  1596. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid scalar tensor."});
  1597. return PARAM_INVALID, "Const is invalid scalar tensor.");
  1598. } else {
  1599. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1600. GE_CHK_BOOL_EXEC(data_size == 0,
  1601. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid vector scalar."});
  1602. return PARAM_INVALID, "Const is invalid vector scalar.");
  1603. }
  1604. } else {
  1605. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0,
  1606. ErrorManager::GetInstance().ATCReportErrMessage(
  1607. "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"});
  1608. return PARAM_INVALID, "Const input data size is not equal with tensor desc shape");
  1609. }
  1610. return SUCCESS;
  1611. }
  1612. bool GraphPrepare::IsDynamicDims(const NodePtr &input_node) {
  1613. auto data_shape = NodeUtils::GetOutputDesc(*input_node, kDataOutIndex).GetShape();
  1614. const auto &dims = data_shape.GetDims();
  1615. bool all_is_positive = false;
  1616. if (std::all_of(dims.begin(), dims.end(), [](int64_t val) { return val >= 0; })) {
  1617. all_is_positive = true;
  1618. }
  1619. if (!all_is_positive && !options_.input_shape.empty() && !options_.dynamic_dims.empty() &&
  1620. options_.dynamic_node_type != kInvalidDynaimcDimsType) {
  1621. GELOGI("No need to check and update desc info, the dims of %s is %s.", input_node->GetName().c_str(),
  1622. formats::JoinToString(dims).c_str());
  1623. return true;
  1624. }
  1625. return false;
  1626. }
  1627. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1628. if (GetLocalOmgContext().is_dynamic_input) {
  1629. return SUCCESS;
  1630. }
  1631. unsigned int node_num = 0;
  1632. unsigned int data_num = 0;
  1633. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1634. GE_CHECK_NOTNULL(input_node);
  1635. OpDescPtr op = input_node->GetOpDesc();
  1636. GE_CHECK_NOTNULL(op);
  1637. node_num++;
  1638. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1639. data_num++;
  1640. GeAttrValue::INT index = 0;
  1641. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1642. GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed");
  1643. return GE_GRAPH_INIT_FAILED;
  1644. }
  1645. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1646. std::string situation = "data op index[" + std::to_string(index) + "]";
  1647. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1648. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1649. GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index);
  1650. return GE_GRAPH_INIT_FAILED;
  1651. }
  1652. if (IsDynamicDims(input_node)) {
  1653. continue;
  1654. }
  1655. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1656. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1657. int64_t dim = desc.GetShape().GetDim(i);
  1658. if (dim < UNKNOWN_DIM_NUM) {
  1659. std::string situation = "data dim[" + std::to_string(i) + "][" + std::to_string(dim) + "]" ;
  1660. std::string reason = "it need >= -2";
  1661. REPORT_INPUT_ERROR(
  1662. "E19025", std::vector<std::string>({"situation", "reason"}), std::vector<std::string>({situation, reason}));
  1663. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][InputDim]data dim %zu is not supported, need >= -2, real:%ld.", i, dim);
  1664. return GE_GRAPH_INIT_FAILED;
  1665. }
  1666. }
  1667. }
  1668. }
  1669. if (node_num <= data_num) {
  1670. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1671. }
  1672. return SUCCESS;
  1673. }
  1674. Status GraphPrepare::InferShapeForPreprocess() {
  1675. GELOGI("Start infershape for preprocess.");
  1676. GEPass ge_passes(compute_graph_);
  1677. NamesToPass names_to_passes;
  1678. AssertPass assert_pass;
  1679. if (!options_.train_graph_flag) {
  1680. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1681. }
  1682. SwitchDeadBranchElimination switch_dead_branch_elimination;
  1683. names_to_passes.emplace_back("SwitchDeadBranchElimination", &switch_dead_branch_elimination);
  1684. MergePass merge_pass;
  1685. names_to_passes.emplace_back("MergePass", &merge_pass);
  1686. InferShapePass infer_shape_pass;
  1687. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1688. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1689. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1690. DimensionComputePass dimension_compute_pass;
  1691. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1692. ConstantFoldingPass constant_folding_pass;
  1693. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1694. int32_t dev_count = 0;
  1695. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1696. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1697. rtError_t rt_err = RT_ERROR_NONE;
  1698. if (aicpu_constant_folding_on != nullptr) {
  1699. rt_err = rtGetDeviceCount(&dev_count);
  1700. if (rt_err == RT_ERROR_NONE) {
  1701. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1702. if (result != SUCCESS) {
  1703. GELOGE(result, "Set rt context failed.");
  1704. return result;
  1705. }
  1706. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1707. }
  1708. }
  1709. Status ret = ge_passes.Run(names_to_passes);
  1710. if (aicpu_constant_folding_on != nullptr) {
  1711. if (rt_err == RT_ERROR_NONE) {
  1712. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1713. if (result != SUCCESS) {
  1714. GELOGE(result, "Set rt context failed.");
  1715. return result;
  1716. }
  1717. }
  1718. }
  1719. ShapeRefiner::ClearContextMap();
  1720. if (ret != SUCCESS) {
  1721. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1722. return ret;
  1723. }
  1724. return SUCCESS;
  1725. }
  1726. Status GraphPrepare::PrepareOptimize() {
  1727. GELOGI("Start optimize for preprocess.");
  1728. // check rw type
  1729. GraphOptimize graph_optimize;
  1730. bool has_conflict = false;
  1731. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1732. if (has_conflict) {
  1733. GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize.");
  1734. return FAILED;
  1735. }
  1736. PassManager original_graph_passes;
  1737. // Graph pass
  1738. try {
  1739. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1740. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1741. (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass);
  1742. } catch (std::bad_alloc &e) {
  1743. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1744. return INTERNAL_ERROR;
  1745. }
  1746. GE_TIMESTAMP_START(original_graph_passes);
  1747. Status ret = original_graph_passes.Run(compute_graph_);
  1748. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1749. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1750. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1751. return ret;
  1752. }
  1753. // New pass
  1754. GEPass ge_passes(compute_graph_);
  1755. NamesToPass names_to_passes;
  1756. EnterPass enter_pass;
  1757. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1758. CondPass cond_pass;
  1759. names_to_passes.emplace_back("CondPass", &cond_pass);
  1760. PrintOpPass print_pass;
  1761. if (options_.enable_print_op_pass) {
  1762. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1763. }
  1764. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1765. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1766. DropOutPass dropout_pass;
  1767. AssertPass assert_pass;
  1768. UnusedConstPass unused_const_pass;
  1769. StopGradientPass stop_gradient_pass;
  1770. PreventGradientPass prevent_gradient_pass;
  1771. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1772. GuaranteeConstPass guarantee_const_pass;
  1773. VarIsInitializedOpPass var_is_initialized_pass;
  1774. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1775. IdentityPass identity_pass(false);
  1776. SnapshotPass snapshot_pass;
  1777. if (!options_.train_graph_flag) {
  1778. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1779. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1780. }
  1781. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1782. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1783. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1784. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1785. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1786. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1787. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1788. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1789. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1790. GE_TIMESTAMP_START(names_to_passes);
  1791. ret = ge_passes.Run(names_to_passes);
  1792. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1793. if (ret != SUCCESS) {
  1794. GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret);
  1795. return ret;
  1796. }
  1797. PassManager graph_pass;
  1798. try {
  1799. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1800. // can't move to optimize1/2 directly, may cause more identity insert, cause CI fail
  1801. (void)graph_pass.AddPass("PrepareOptimize::HcclMemcpyPass", new HcclMemcpyPass);
  1802. } catch (std::bad_alloc &e) {
  1803. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1804. return INTERNAL_ERROR;
  1805. }
  1806. GE_TIMESTAMP_START(graph_passes);
  1807. ret = graph_pass.Run(compute_graph_);
  1808. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1809. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1810. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1811. return ret;
  1812. }
  1813. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1814. TypeConversionOfConstant();
  1815. ret = compute_graph_->TopologicalSorting();
  1816. if (ret != SUCCESS) {
  1817. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1818. return ret;
  1819. }
  1820. GELOGI("End optimize for preprocess.");
  1821. return SUCCESS;
  1822. }
  1823. void GraphPrepare::TypeConversionOfConstant() {
  1824. bool is_acl_compile = false;
  1825. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1826. // This can ensure that n is not a null pointer
  1827. // No Conversion when called by aclOpCompile
  1828. (void)AttrUtils::GetBool(n->GetOpDesc(), ATTR_SINGLE_OP_SCENE, is_acl_compile);
  1829. if (is_acl_compile) {
  1830. return;
  1831. }
  1832. }
  1833. if (options_.train_graph_flag) {
  1834. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1835. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1836. // This can ensure that n is not a null pointer
  1837. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1838. n->GetOpDesc()->SetType(CONSTANTOP);
  1839. }
  1840. }
  1841. } else {
  1842. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  1843. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1844. // This can ensure that n is not a null pointer
  1845. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  1846. n->GetOpDesc()->SetType(CONSTANT);
  1847. }
  1848. }
  1849. }
  1850. }
  1851. Status GraphPrepare::GraphEquivalentTransformation() {
  1852. NamesToPass names_to_pass;
  1853. ForPass for_pass;
  1854. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  1855. return GEPass(compute_graph_).Run(names_to_pass);
  1856. }
  1857. Status GraphPrepare::ProcessBeforeInfershape() {
  1858. NamesToPass names_to_passes;
  1859. CondRemovePass condition_remove_pass;
  1860. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  1861. GE_TIMESTAMP_START(ProcessCondRemove);
  1862. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  1863. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  1864. if (ret != SUCCESS) {
  1865. GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  1866. return ret;
  1867. }
  1868. return SUCCESS;
  1869. }
  1870. Status GraphPrepare::ProcessNetOutput() {
  1871. PassManager graph_passes_before_infershape;
  1872. try {
  1873. if (options_.train_graph_flag) {
  1874. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  1875. }
  1876. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  1877. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  1878. new (std::nothrow) DataPass); // Add NetOutput first.
  1879. } catch (std::bad_alloc) {
  1880. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1881. return INTERNAL_ERROR;
  1882. }
  1883. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  1884. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  1885. GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret);
  1886. return ret;
  1887. }
  1888. return SUCCESS;
  1889. }
  1890. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input,
  1891. const std::map<string, string> &graph_option) {
  1892. compute_graph_->SetInputSize(user_input.size());
  1893. if (user_input.empty()) {
  1894. return SUCCESS;
  1895. }
  1896. auto ret = CheckUserInput(user_input);
  1897. if (ret != SUCCESS) {
  1898. GELOGE(ret, "Check user input failed.");
  1899. return ret;
  1900. }
  1901. ret = UpdateInput(user_input, graph_option);
  1902. if (ret != SUCCESS) {
  1903. GELOGE(ret, "UpdateInput fail, ret:%u", ret);
  1904. return ret;
  1905. }
  1906. if (user_input.size() != 0) {
  1907. ret = CheckConstOp();
  1908. if (ret != SUCCESS) {
  1909. GELOGE(ret, "CheckConstOp fail, ret:%u", ret);
  1910. return ret;
  1911. }
  1912. } else {
  1913. ret = compute_graph_->TopologicalSorting();
  1914. if (ret != SUCCESS) {
  1915. GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting");
  1916. return FAILED;
  1917. }
  1918. }
  1919. return SUCCESS;
  1920. }
  1921. Status GraphPrepare::UpdateInputOutputByOptions() {
  1922. auto ret = UpdateDataNetOutputByStorageFormat();
  1923. if (ret != SUCCESS) {
  1924. GELOGE(ret, "Update format acoording to storage format failed.");
  1925. return ret;
  1926. }
  1927. if (options_.train_graph_flag) {
  1928. GELOGI("This is train mode, no need to do this schedule.");
  1929. return SUCCESS;
  1930. }
  1931. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  1932. GE_CHECK_NOTNULL(node_ptr);
  1933. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  1934. GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str());
  1935. return FAILED;
  1936. }
  1937. if (node_ptr->GetType() == DATA) {
  1938. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  1939. GELOGE(INTERNAL_ERROR, "Process data node failed");
  1940. return FAILED;
  1941. }
  1942. }
  1943. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1944. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  1945. GELOGE(INTERNAL_ERROR, "Process netoutput node failed");
  1946. return FAILED;
  1947. }
  1948. }
  1949. }
  1950. return SUCCESS;
  1951. }
  1952. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  1953. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  1954. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  1955. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  1956. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  1957. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  1958. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  1959. if (dst_node->GetType() == TRANSDATA) {
  1960. return true;
  1961. }
  1962. }
  1963. }
  1964. return false;
  1965. }
  1966. } // 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