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 84 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #include "graph/preprocess/graph_preprocess.h"
  17. #include <map>
  18. #include <set>
  19. #include <string>
  20. #include "common/formats/format_transfers/format_transfer_fractal_nz.h"
  21. #include "common/formats/format_transfers/format_transfer_nchw_nc1hwc0.h"
  22. #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h"
  23. #include "common/formats/format_transfers/format_transfer_transpose.h"
  24. #include "common/formats/utils/formats_trans_utils.h"
  25. #include "common/helper/model_helper.h"
  26. #include "common/math/math_util.h"
  27. #include "common/op/ge_op_utils.h"
  28. #include "graph/common/ge_call_wrapper.h"
  29. #include "graph/common/local_context.h"
  30. #include "graph/common/transop_util.h"
  31. #include "graph/ge_context.h"
  32. #include "graph/shape_refiner.h"
  33. #include "graph/manager/graph_var_manager.h"
  34. #include "graph/manager/util/rt_context_util.h"
  35. #include "graph/optimize/graph_optimize.h"
  36. #include "graph/passes/addn_pass.h"
  37. #include "graph/passes/aicpu_constant_folding_pass.h"
  38. #include "graph/passes/assert_pass.h"
  39. #ifdef ONLY_COMPILE_OPEN_SRC
  40. #include "graph/passes/assign_remove_pass.h"
  41. #endif
  42. #include "graph/passes/common_subexpression_elimination_pass.h"
  43. #include "graph/passes/cond_pass.h"
  44. #include "graph/passes/cond_remove_pass.h"
  45. #include "graph/passes/constant_folding_pass.h"
  46. #include "graph/passes/dimension_adjust_pass.h"
  47. #include "graph/passes/dimension_compute_pass.h"
  48. #include "graph/passes/dropout_pass.h"
  49. #include "graph/passes/enter_pass.h"
  50. #include "graph/passes/for_pass.h"
  51. #include "graph/passes/guarantee_const_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/net_output_pass.h"
  56. #include "graph/passes/no_use_reshape_remove_pass.h"
  57. #include "graph/passes/parallel_concat_start_op_pass.h"
  58. #include "graph/passes/placeholder_with_default_pass.h"
  59. #include "graph/passes/prevent_gradient_pass.h"
  60. #include "graph/passes/print_op_pass.h"
  61. #include "graph/passes/prune_pass.h"
  62. #include "graph/passes/replace_transshape_pass.h"
  63. #include "graph/passes/replace_with_empty_const_pass.h"
  64. #include "graph/passes/resource_pair_add_control_pass.h"
  65. #include "graph/passes/resource_pair_remove_control_pass.h"
  66. #include "graph/passes/save_pass.h"
  67. #include "graph/passes/shape_operate_op_remove_pass.h"
  68. #include "graph/passes/snapshot_pass.h"
  69. #include "graph/passes/stop_gradient_pass.h"
  70. #include "graph/passes/unused_const_pass.h"
  71. #include "graph/passes/var_is_initialized_op_pass.h"
  72. #include "graph/passes/variable_prepare_op_pass.h"
  73. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  74. #include "graph/utils/type_utils.h"
  75. #include "inc/pass_manager.h"
  76. #include "init/gelib.h"
  77. #include "multi_batch_copy_graph.h"
  78. #include "graph/passes/data_pass.h"
  79. #include "graph/passes/mark_agnostic_pass.h"
  80. namespace ge {
  81. namespace {
  82. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  83. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  84. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  85. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  86. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  87. // the size of user defined output datatype or format string after split by ":".
  88. const size_t kUserDefinedElementCount = 2;
  89. const int kDataOutIndex = 0;
  90. const int64_t kInvalidDynaimcDimsType = -1;
  91. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  92. GeTensorPtr tensor = MakeShared<GeTensor>();
  93. if (tensor == nullptr) {
  94. GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed");
  95. return nullptr;
  96. }
  97. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  98. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  99. auto dst_ge_shape = data_tensor.GetShape();
  100. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  101. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  102. tensor->MutableTensorDesc().SetShape(GeShape());
  103. int32_t dst_shape = 1;
  104. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  105. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  106. return nullptr;
  107. }
  108. } else {
  109. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  110. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  111. if (dst_shape == nullptr) {
  112. GELOGE(INTERNAL_ERROR, "Create unique ptr failed");
  113. return nullptr;
  114. }
  115. for (int64_t i = 0; i < dim_cnt; ++i) {
  116. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  117. }
  118. GE_IF_BOOL_EXEC(
  119. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  120. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  121. return nullptr;)
  122. }
  123. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  124. return OpDescUtils::CreateConstOp(tensor);
  125. }
  126. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  127. OpDescPtr &op_desc) {
  128. // For format transfer node, the IR definition has src/dst format attrs
  129. if (node_type == TRANSDATA) {
  130. GE_IF_BOOL_EXEC(
  131. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  132. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  133. GE_IF_BOOL_EXEC(
  134. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  135. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  136. }
  137. // For TransposeD node, the IR definition has perm attrs
  138. if (node_type == TRANSPOSED) {
  139. Format src_format = input.GetFormat();
  140. Format dst_format = output.GetFormat();
  141. std::vector<int64_t> perm_arg;
  142. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  143. "Get perm by foramt failed.");
  144. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  145. "SetStr PERMUTE_ATTR_PERM failed")
  146. }
  147. // For cast node, the IR definition has src/dst attrs
  148. if (node_type == CAST) {
  149. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  150. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  151. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  152. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  153. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  154. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  155. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  156. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  157. }
  158. }
  159. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  160. const GeTensorDesc &output, NodePtr &node) {
  161. if (node == nullptr) {
  162. GELOGE(PARAM_INVALID, "node is null.");
  163. return nullptr;
  164. }
  165. auto graph = node->GetOwnerComputeGraph();
  166. if (graph == nullptr) {
  167. GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str());
  168. return nullptr;
  169. }
  170. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  171. if (index < 0) {
  172. ErrorManager::GetInstance().ATCReportErrMessage(
  173. "E19025", {"situation", "reason"},
  174. {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()});
  175. GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str());
  176. return nullptr;
  177. }
  178. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  179. if (op_desc == nullptr) {
  180. GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed");
  181. return nullptr;
  182. }
  183. // for data dump
  184. GE_IF_BOOL_EXEC(
  185. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  186. GELOGW("CreateTransNode: SetListStr failed");)
  187. // Default single input and single output
  188. auto ret = op_desc->AddInputDesc(input);
  189. if (ret != GRAPH_SUCCESS) {
  190. GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str());
  191. return nullptr;
  192. }
  193. ret = op_desc->AddOutputDesc(output);
  194. if (ret != GRAPH_SUCCESS) {
  195. GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str());
  196. return nullptr;
  197. }
  198. AddTransNodeAttr(node_type, input, output, op_desc);
  199. NodePtr shape_node = nullptr;
  200. if (node_type == RESHAPE) {
  201. auto shape_desc = CreateTensorShape(output);
  202. if (shape_desc == nullptr) {
  203. GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input",
  204. node->GetName().c_str());
  205. return nullptr;
  206. }
  207. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  208. if (ret != GRAPH_SUCCESS) {
  209. GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str());
  210. return nullptr;
  211. }
  212. shape_node = graph->AddNode(shape_desc);
  213. if (shape_node == nullptr) {
  214. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str());
  215. return nullptr;
  216. }
  217. }
  218. auto trans_node = graph->AddNode(op_desc);
  219. if (trans_node == nullptr) {
  220. GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str());
  221. return nullptr;
  222. }
  223. if (node_type == RESHAPE) {
  224. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  225. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str());
  226. return nullptr;
  227. }
  228. }
  229. return trans_node;
  230. }
  231. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  232. NodePtr &trans_node) {
  233. GE_CHECK_NOTNULL(node);
  234. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  235. if (trans_node == nullptr) {
  236. return INTERNAL_ERROR;
  237. }
  238. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  239. if (ret != GRAPH_SUCCESS) {
  240. GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s",
  241. node->GetName().c_str(), node->GetType().c_str());
  242. return INTERNAL_ERROR;
  243. }
  244. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  245. if (ret != GRAPH_SUCCESS) {
  246. GELOGE(INTERNAL_ERROR, "Failed to connect node %s to trans node %s", node->GetName().c_str(),
  247. trans_node->GetName().c_str());
  248. return INTERNAL_ERROR;
  249. }
  250. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  251. if (ret != GRAPH_SUCCESS) {
  252. GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.",
  253. node->GetName().c_str(), trans_node->GetName().c_str());
  254. return INTERNAL_ERROR;
  255. }
  256. return SUCCESS;
  257. }
  258. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  259. NodePtr &trans_node) {
  260. GE_CHECK_NOTNULL(node);
  261. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  262. if (trans_node == nullptr) {
  263. return INTERNAL_ERROR;
  264. }
  265. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  266. if (ret != GRAPH_SUCCESS) {
  267. GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s",
  268. node->GetName().c_str(), node->GetType().c_str());
  269. return INTERNAL_ERROR;
  270. }
  271. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  272. if (ret != GRAPH_SUCCESS) {
  273. GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(),
  274. node->GetName().c_str());
  275. return INTERNAL_ERROR;
  276. }
  277. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  278. if (ret != GRAPH_SUCCESS) {
  279. GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.",
  280. node->GetName().c_str(), trans_node->GetName().c_str());
  281. return INTERNAL_ERROR;
  282. }
  283. return SUCCESS;
  284. }
  285. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  286. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  287. GE_CHECK_NOTNULL(var->GetOpDesc());
  288. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  289. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  290. output_desc.SetFormat(tensor_desc.GetFormat());
  291. output_desc.SetDataType(tensor_desc.GetDataType());
  292. output_desc.SetShape(tensor_desc.GetShape());
  293. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  294. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  295. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  296. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  297. GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed");
  298. return INTERNAL_ERROR;);
  299. }
  300. if (var->GetOpDesc()->GetInputsSize() > 0) {
  301. auto desc = var->GetOpDesc()->GetInputDesc(0);
  302. desc.SetFormat(tensor_desc.GetFormat());
  303. desc.SetDataType(tensor_desc.GetDataType());
  304. desc.SetShape(tensor_desc.GetShape());
  305. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  306. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  307. desc.SetOriginShape(tensor_desc.GetOriginShape());
  308. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  309. GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed");
  310. return INTERNAL_ERROR;)
  311. }
  312. return SUCCESS;
  313. }
  314. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  315. GE_CHECK_NOTNULL(var);
  316. int index = 0;
  317. NodePtr last_node = var;
  318. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  319. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  320. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  321. if (ret != SUCCESS) {
  322. ErrorManager::GetInstance().ATCReportErrMessage(
  323. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  324. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(),
  325. index, iter->node_type.c_str());
  326. return INTERNAL_ERROR;
  327. }
  328. // set stream_label
  329. OpDescPtr var_desc = var->GetOpDesc();
  330. GE_CHECK_NOTNULL(var_desc);
  331. std::string stream_label;
  332. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  333. if (!stream_label.empty()) {
  334. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  335. }
  336. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  337. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  338. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  339. }
  340. if (road.empty()) {
  341. return SUCCESS;
  342. }
  343. return UpdateVarFormats(var, road.rbegin()->output);
  344. }
  345. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  346. for (auto &var : nodes) {
  347. GE_CHECK_NOTNULL(var);
  348. int index = 0;
  349. NodePtr last_node = var;
  350. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  351. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  352. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  353. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  354. if (ret != SUCCESS) {
  355. ErrorManager::GetInstance().ATCReportErrMessage(
  356. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  357. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s",
  358. var->GetName().c_str(), index, iter->node_type.c_str());
  359. return INTERNAL_ERROR;
  360. }
  361. // set stream_label
  362. OpDescPtr var_desc = var->GetOpDesc();
  363. GE_CHECK_NOTNULL(var_desc);
  364. std::string stream_label;
  365. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  366. if (!stream_label.empty()) {
  367. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  368. }
  369. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  370. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  371. }
  372. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  373. return INTERNAL_ERROR;
  374. }
  375. }
  376. return SUCCESS;
  377. }
  378. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  379. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  380. VarNamesToRefs names_to_refs;
  381. std::string var_name;
  382. if (graph == nullptr) {
  383. GELOGE(PARAM_INVALID, "graph is null.");
  384. return names_to_refs;
  385. }
  386. for (auto &node : graph->GetAllNodes()) {
  387. if (node->GetType() != VARIABLE) {
  388. continue;
  389. }
  390. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  391. (void)names_to_refs[var_name].insert(node);
  392. }
  393. }
  394. return names_to_refs;
  395. }
  396. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  397. std::vector<int64_t> &dst_shape) {
  398. if (src_format == FORMAT_NCHW) {
  399. formats::FormatTransferNchwNc1hwc0 transfer;
  400. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  401. GELOGE(INTERNAL_ERROR, "TransShape failed");
  402. return FAILED;
  403. }
  404. } else if (src_format == FORMAT_NHWC) {
  405. formats::FormatTransferNhwcNc1hwc0 transfer;
  406. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  407. GELOGE(INTERNAL_ERROR, "TransShape failed");
  408. return FAILED;
  409. }
  410. }
  411. return SUCCESS;
  412. }
  413. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  414. GE_CHECK_NOTNULL(node_ptr);
  415. auto op_desc = node_ptr->GetOpDesc();
  416. GE_CHECK_NOTNULL(op_desc);
  417. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  418. GE_CHECK_NOTNULL(input);
  419. ge::Format old_format = input->GetFormat();
  420. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  421. ge::DataType dt = input->GetDataType();
  422. std::vector<int64_t> dst_shape_dims;
  423. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  424. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  425. return FAILED;
  426. }
  427. input->SetFormat(FORMAT_NC1HWC0);
  428. input->SetShape(ge::GeShape(dst_shape_dims));
  429. auto output = op_desc->MutableOutputDesc(0);
  430. GE_CHECK_NOTNULL(output);
  431. output->SetFormat(FORMAT_NC1HWC0);
  432. output->SetShape(ge::GeShape(dst_shape_dims));
  433. int64_t size = 0;
  434. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  435. if (graph_status != ge::GRAPH_SUCCESS) {
  436. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  437. return FAILED;
  438. }
  439. ge::TensorUtils::SetSize(*output, size);
  440. ge::TensorUtils::SetSize(*input, size);
  441. return SUCCESS;
  442. }
  443. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  444. GE_CHECK_NOTNULL(input_output);
  445. ge::Format old_format = input_output->GetFormat();
  446. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  447. ge::DataType dt = input_output->GetDataType();
  448. std::vector<int64_t> dst_shape_dims;
  449. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  450. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  451. return FAILED;
  452. }
  453. input_output->SetFormat(FORMAT_NC1HWC0);
  454. input_output->SetShape(ge::GeShape(dst_shape_dims));
  455. return SUCCESS;
  456. }
  457. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  458. vector<int64_t> &dst_shape_dims) {
  459. GE_CHECK_NOTNULL(op_desc);
  460. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  461. GE_CHECK_NOTNULL(input);
  462. ge::Format old_format = input->GetFormat();
  463. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  464. input->SetShape(ge::GeShape(dst_shape_dims));
  465. input->SetFormat(storage_format);
  466. auto output = op_desc->MutableOutputDesc(index);
  467. GE_CHECK_NOTNULL(output);
  468. output->SetShape(ge::GeShape(dst_shape_dims));
  469. output->SetFormat(storage_format);
  470. if (!output->MutableShape().IsUnknownShape()) {
  471. int64_t size = 0;
  472. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  473. if (graph_status != ge::GRAPH_SUCCESS) {
  474. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  475. return FAILED;
  476. }
  477. ge::TensorUtils::SetSize(*input, size);
  478. ge::TensorUtils::SetSize(*output, size);
  479. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  480. "new_shape:%s, new_format:%s, new_size:%lu",
  481. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  482. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  483. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  484. }
  485. return SUCCESS;
  486. }
  487. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) {
  488. is_dynamic_batch = false;
  489. std::string related_node_name;
  490. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  491. if (related_node_name.empty()) {
  492. ErrorManager::GetInstance().ATCReportErrMessage(
  493. "E15002", {"opname", "value", "reason"}, {data_node->GetName(), "flag", "but the value is empty"});
  494. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty",
  495. data_node->GetName().c_str());
  496. return INTERNAL_ERROR;
  497. }
  498. for (const NodePtr &next_node : data_node->GetOutNodes()) {
  499. if (next_node->GetName() == related_node_name) {
  500. switchn_node = next_node;
  501. break;
  502. }
  503. }
  504. if (switchn_node == nullptr) {
  505. ErrorManager::GetInstance().ATCReportErrMessage(
  506. "E15002", {"opname", "value", "reason"},
  507. {data_node->GetName(), related_node_name, "but can not find it on the graph"});
  508. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph",
  509. data_node->GetName().c_str(), related_node_name.c_str());
  510. return INTERNAL_ERROR;
  511. }
  512. is_dynamic_batch = true;
  513. }
  514. return SUCCESS;
  515. }
  516. bool CheckOpType(const NodePtr &node, const std::string type) {
  517. if (node->GetType() == type) {
  518. return true;
  519. }
  520. return false;
  521. }
  522. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  523. auto op = node_ptr->GetOpDesc();
  524. GE_CHECK_NOTNULL(op);
  525. auto inputDescsPtr = op->GetAllInputsDescPtr();
  526. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  527. ge::Format format = ge::FORMAT_ND;
  528. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  529. // process, otherwise fe will insert transdata.
  530. for (auto &inputDescPtr : inputDescsPtr) {
  531. GE_CHECK_NOTNULL(inputDescPtr);
  532. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  533. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  534. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  535. inputDescPtr->SetFormat(format);
  536. inputDescPtr->SetOriginFormat(format);
  537. }
  538. }
  539. for (auto &outputDescPtr : outputDescsPtr) {
  540. GE_CHECK_NOTNULL(outputDescPtr);
  541. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  542. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  543. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  544. outputDescPtr->SetFormat(format);
  545. outputDescPtr->SetOriginFormat(format);
  546. }
  547. }
  548. return SUCCESS;
  549. }
  550. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  551. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  552. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  553. // data and netoutput, and no longer inserts the transnode.
  554. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) {
  555. GE_CHECK_NOTNULL(node_ptr);
  556. auto op_desc = node_ptr->GetOpDesc();
  557. GE_CHECK_NOTNULL(op_desc);
  558. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  559. GE_CHECK_NOTNULL(input);
  560. ge::DataType src_dtype = input->GetDataType();
  561. if (src_dtype == dt_set) {
  562. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  563. return SUCCESS;
  564. }
  565. input->SetDataType(dt_set);
  566. int64_t input_shape_size = 0;
  567. int64_t output_shape_size = 0;
  568. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  569. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  570. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  571. GELOGE(GRAPH_FAILED, "GetTensorSize failed!");
  572. return FAILED;
  573. }
  574. ge::TensorUtils::SetSize(*input, input_shape_size);
  575. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  576. GE_CHECK_NOTNULL(output);
  577. output->SetDataType(dt_set);
  578. ge::TensorUtils::SetSize(*output, output_shape_size);
  579. if (is_dynamic_batch) {
  580. GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str());
  581. auto switchn_op_desc = switchn_node->GetOpDesc();
  582. GE_CHECK_NOTNULL(switchn_op_desc);
  583. auto switchn_input = switchn_op_desc->MutableInputDesc(0);
  584. GE_CHECK_NOTNULL(switchn_input);
  585. switchn_input->SetDataType(dt_set);
  586. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  587. const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i);
  588. GE_CHECK_NOTNULL(switchn_output);
  589. switchn_output->SetDataType(dt_set);
  590. }
  591. }
  592. return SUCCESS;
  593. }
  594. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  595. GE_CHECK_NOTNULL(node_ptr);
  596. auto op_desc = node_ptr->GetOpDesc();
  597. GE_CHECK_NOTNULL(op_desc);
  598. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  599. GE_CHECK_NOTNULL(input);
  600. ge::Format old_format = input->GetFormat();
  601. ge::GeShape old_shape = input->GetShape();
  602. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  603. if (!support) {
  604. ErrorManager::GetInstance().ATCReportErrMessage(
  605. "E19014", {"opname", "value", "reason"},
  606. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  607. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  608. GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str());
  609. return FAILED;
  610. }
  611. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  612. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  613. return FAILED;
  614. }
  615. if (is_dynamic_batch) {
  616. auto switchn_op_desc = switchn_node->GetOpDesc();
  617. GE_CHECK_NOTNULL(switchn_op_desc);
  618. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  619. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  620. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  621. return FAILED;
  622. }
  623. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  624. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  625. GE_CHECK_NOTNULL(switchn_output);
  626. old_format = switchn_output->GetFormat();
  627. old_shape = switchn_output->GetShape();
  628. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  629. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  630. return FAILED;
  631. }
  632. }
  633. }
  634. return SUCCESS;
  635. }
  636. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  637. auto op_desc = node_ptr->GetOpDesc();
  638. GE_CHECK_NOTNULL(op_desc);
  639. string set_dt_str;
  640. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  641. return SUCCESS;
  642. }
  643. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  644. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  645. bool is_dynamic_batch = false;
  646. NodePtr switchn_node = nullptr;
  647. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) {
  648. GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed");
  649. return FAILED;
  650. }
  651. if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) {
  652. GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed");
  653. return FAILED;
  654. }
  655. // check if need to set format
  656. string set_format;
  657. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  658. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  659. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  660. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) {
  661. GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed");
  662. return FAILED;
  663. }
  664. }
  665. return SUCCESS;
  666. }
  667. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  668. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  669. GE_CHECK_NOTNULL(op_desc);
  670. GE_CHECK_NOTNULL(tensor_desc_ptr);
  671. storage_format = FORMAT_RESERVED;
  672. int64_t format = FORMAT_RESERVED;
  673. dst_shape_dims.clear();
  674. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  675. storage_format = static_cast<Format>(format);
  676. vector<int32_t> storage_shape;
  677. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  678. for (auto dim : storage_shape) {
  679. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  680. }
  681. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  682. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  683. formats::JoinToString(storage_shape).c_str());
  684. } else {
  685. ErrorManager::GetInstance().ATCReportErrMessage(
  686. "15003", {"opname", "format"},
  687. {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)});
  688. GELOGE(PARAM_INVALID, "Update node by storage format failed, storage_shape not set. "
  689. "node: [%s], storage_format [%s]",
  690. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  691. return FAILED;
  692. }
  693. ge::Format old_format = tensor_desc_ptr->GetFormat();
  694. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  695. if (old_format == storage_format && old_shape == dst_shape_dims) {
  696. GELOGI("Update node by storage format, not changed.");
  697. storage_format = FORMAT_RESERVED;
  698. return SUCCESS;
  699. }
  700. }
  701. return SUCCESS;
  702. }
  703. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  704. NodePtr &node) {
  705. bool is_dynamic = CheckOpType(node, MERGE);
  706. auto src_op_desc = node->GetOpDesc();
  707. GE_CHECK_NOTNULL(src_op_desc);
  708. ge::GeShape src_shape = src_desc.GetShape();
  709. ge::Format src_format = src_desc.GetFormat();
  710. net_output_input_desc->SetDataType(DT_FLOAT16);
  711. if (is_dynamic) {
  712. auto merge_output = src_op_desc->MutableOutputDesc(0);
  713. GE_CHECK_NOTNULL(merge_output);
  714. merge_output->SetDataType(DT_FLOAT16);
  715. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  716. auto merge_input = src_op_desc->MutableInputDesc(i);
  717. GE_CHECK_NOTNULL(merge_input);
  718. merge_input->SetDataType(DT_FLOAT16);
  719. }
  720. }
  721. std::vector<int64_t> dst_shape_dims;
  722. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  723. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  724. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  725. return FAILED;
  726. }
  727. ge::GeShape dst_shape(dst_shape_dims);
  728. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  729. net_output_input_desc->SetShape(dst_shape);
  730. if (is_dynamic) {
  731. auto merge_out = src_op_desc->MutableOutputDesc(0);
  732. GE_CHECK_NOTNULL(merge_out);
  733. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  734. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  735. return FAILED;
  736. }
  737. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  738. auto merge_in = src_op_desc->MutableInputDesc(i);
  739. GE_CHECK_NOTNULL(merge_in);
  740. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  741. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  742. return FAILED;
  743. }
  744. }
  745. }
  746. return SUCCESS;
  747. }
  748. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  749. GE_CHECK_NOTNULL(netout_desc);
  750. vector<string> output_dt_str;
  751. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  752. for (auto dt_str : output_dt_str) {
  753. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  754. if (dt_str_split.size() == kUserDefinedElementCount) {
  755. if (dt_str_split[0] == to_string(index)) {
  756. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  757. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  758. TypeUtils::DataTypeToSerialString(dt).c_str());
  759. return true;
  760. }
  761. }
  762. }
  763. }
  764. return false;
  765. }
  766. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  767. GE_CHECK_NOTNULL(netout_desc);
  768. vector<string> output_format_str;
  769. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  770. for (auto format_str : output_format_str) {
  771. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  772. if (format_str_split.size() == kUserDefinedElementCount) {
  773. if (format_str_split[0] == to_string(index)) {
  774. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  775. return true;
  776. }
  777. }
  778. }
  779. }
  780. return false;
  781. }
  782. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  783. auto op_desc = node->GetOpDesc();
  784. GE_CHECK_NOTNULL(op_desc);
  785. ge::DataType output_data_type = ge::DT_FLOAT;
  786. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  787. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  788. auto peer_out = in_anchor->GetPeerOutAnchor();
  789. GE_CHECK_NOTNULL(peer_out);
  790. auto src_node = peer_out->GetOwnerNode();
  791. GE_CHECK_NOTNULL(src_node);
  792. bool is_dynamic = CheckOpType(src_node, MERGE);
  793. OpDescPtr src_op_desc = src_node->GetOpDesc();
  794. GE_CHECK_NOTNULL(src_op_desc);
  795. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  796. GE_CHECK_NOTNULL(net_output_input_desc);
  797. ge::GeShape old_shape = net_output_input_desc->GetShape();
  798. ge::Format old_format = net_output_input_desc->GetFormat();
  799. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  800. // Update datatype
  801. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  802. GELOGI("Enter into process output_type schedule");
  803. net_output_input_desc->SetDataType(output_data_type);
  804. if (is_dynamic) {
  805. auto merge_output = src_op_desc->MutableOutputDesc(0);
  806. GE_CHECK_NOTNULL(merge_output);
  807. merge_output->SetDataType(output_data_type);
  808. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  809. auto merge_input = src_op_desc->MutableInputDesc(i);
  810. GE_CHECK_NOTNULL(merge_input);
  811. merge_input->SetDataType(output_data_type);
  812. }
  813. }
  814. }
  815. // check if is_output_adjust_hw_layout is set
  816. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  817. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  818. ErrorManager::GetInstance().ATCReportErrMessage(
  819. "E19014", {"opname", "value", "reason"},
  820. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  821. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  822. GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0.");
  823. return FAILED;
  824. }
  825. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  826. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  827. GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0.");
  828. return FAILED;
  829. }
  830. }
  831. }
  832. return SUCCESS;
  833. }
  834. /**
  835. * Parser shape_range from string to vector
  836. * shape_range from option normally is "[1~20],[3],[3~6],[-1]"
  837. * @param shape_range
  838. */
  839. void ParseDynamicInputShapeRange(const std::string &shape_range,
  840. std::vector<std::vector<std::pair<int64_t, int64_t>>> &range) {
  841. if (shape_range.empty() || shape_range.size() < 2) {
  842. GELOGW("Shape range %s is invalid.", shape_range);
  843. return;
  844. }
  845. // different parameter sets are split by ';'
  846. vector<string> shape_set = ge::StringUtils::Split(shape_range, ']');
  847. if (shape_set.empty()) {
  848. return;
  849. }
  850. for (auto shape_str : shape_set) {
  851. if (shape_str.empty()) {
  852. continue;
  853. }
  854. if (ge::StringUtils::StartWith(shape_str, "[")) {
  855. shape_str = shape_str.substr(1, shape_str.size());
  856. }
  857. if (ge::StringUtils::StartWith(shape_str, ",")) {
  858. shape_str = shape_str.substr(2, shape_str.size());
  859. }
  860. std::vector<std::pair<int64_t, int64_t>> range_of_single;
  861. vector<string> range_set = ge::StringUtils::Split(shape_str, ',');
  862. for (auto range_str : range_set) {
  863. vector<string> pair_set = ge::StringUtils::Split(range_str, '~');
  864. pair<int64_t, int64_t> range_pair;
  865. if (pair_set.size() == 1) {
  866. auto range_value = atoi(pair_set.at(0).c_str());
  867. if (range_value < 0) {
  868. range_pair = std::make_pair(1, range_value);
  869. } else {
  870. range_pair = std::make_pair(range_value, range_value);
  871. }
  872. } else if (pair_set.size() == 2) {
  873. auto range_left = atoi(pair_set.at(0).c_str());
  874. auto range_right = atoi(pair_set.at(1).c_str());
  875. range_pair = std::make_pair(range_left, range_right);
  876. }
  877. range_of_single.emplace_back(range_pair);
  878. }
  879. range.emplace_back(range_of_single);
  880. }
  881. }
  882. Status GetDynamicInputShapeRange(const std::vector<GeTensor> &user_input, const std::map<string, string> &graph_option,
  883. vector<vector<std::pair<int64_t, int64_t>>> &range_vec) {
  884. auto mode_iter = graph_option.find(OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  885. if (mode_iter == graph_option.end()) {
  886. GELOGD("Graph Option: Can not find %s option in graph options.", OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  887. return SUCCESS;
  888. }
  889. GELOGD("Graph Option: dynamic_input_mode value is %s.", mode_iter->second.c_str());
  890. if (mode_iter->second != "dynamic_execute") {
  891. return SUCCESS;
  892. }
  893. auto iter = graph_option.find(OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  894. if (iter == graph_option.end()) {
  895. GELOGE(PARAM_INVALID, "Graph option %s is required when %s is dynamic_execute", OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE,
  896. OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  897. return PARAM_INVALID;
  898. }
  899. GELOGD("GraphOption: dynamic_inputs_shape_range value is %s.", iter->second.c_str());
  900. ParseDynamicInputShapeRange(iter->second, range_vec);
  901. if (range_vec.size() != user_input.size()) {
  902. GELOGE(PARAM_INVALID, "Dynamic input shape range size is %zu, inputs size is %zu. Not match.", range_vec.size(),
  903. user_input.size());
  904. return PARAM_INVALID;
  905. }
  906. return SUCCESS;
  907. }
  908. Status UpdateDynamicInputShapeRange(const ge::GeAttrValue::INT index,
  909. const ector<vector<std::pair<int64_t, int64_t>>> &range_vec, OpDescPtr &op,
  910. GeTensorDesc &desc) {
  911. auto unkown_shape = desc.GetShape();
  912. auto shape_range = range_vec.at(index);
  913. for (size_t i = 0; i < unkown_shape.GetDimNum(); ++i) {
  914. if (shape_range.at(i).first == shape_range.at(i).second) {
  915. unkown_shape.SetDim(i, shape_range.at(i).first);
  916. } else {
  917. unkown_shape.SetDim(i, -1);
  918. }
  919. }
  920. desc.SetShape(unkown_shape);
  921. desc.SetShapeRange(shape_range);
  922. int64_t dynamic_shape_size = 1;
  923. for (const auto range_pair : range_vec.at(index)) {
  924. FMK_INT64_MULCHECK(dynamic_shape_size, range_pair.second);
  925. dynamic_shape_size *= range_pair.second;
  926. }
  927. auto data_type_size = GetSizeByDataType(desc.GetDataType());
  928. if (data_type_size < 0) {
  929. GELOGE(PARAM_INVALID, "Input data type is %s, is not supported.",
  930. TypeUtils::DataTypeToSerialString(desc.GetDataType()).c_str());
  931. return PARAM_INVALID;
  932. }
  933. FMK_INT64_MULCHECK(dynamic_shape_size, data_type_size);
  934. dynamic_shape_size *= data_type_size;
  935. GELOGI("In dynamic_execute mode ,set input %s shape range size %ld", op->GetName().c_str(), dynamic_shape_size);
  936. ge::TensorUtils::SetSize(desc, dynamic_shape_size);
  937. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  938. GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  939. graph_ret = op->UpdateOutputDesc(0, desc);
  940. GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  941. return SUCCESS;
  942. }
  943. } // namespace
  944. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  945. GraphPrepare::~GraphPrepare() {}
  946. /**
  947. * @param graph
  948. * @return
  949. */
  950. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  951. GE_CHECK_NOTNULL(graph);
  952. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  953. for (auto &node : graph->GetAllNodes()) {
  954. if (node == nullptr) {
  955. continue;
  956. }
  957. if (node->GetType() != VARIABLE) {
  958. continue;
  959. }
  960. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  961. if (trans_road == nullptr) {
  962. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  963. continue;
  964. }
  965. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  966. auto ret = RecoverTransRoadForVar(node, *trans_road);
  967. if (ret != SUCCESS) {
  968. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str());
  969. return INTERNAL_ERROR;
  970. }
  971. auto iter = var_names_to_refs.find(node->GetName());
  972. if (iter != var_names_to_refs.end()) {
  973. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  974. if (ret != SUCCESS) {
  975. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str());
  976. return INTERNAL_ERROR;
  977. }
  978. }
  979. }
  980. return SUCCESS;
  981. }
  982. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  983. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  984. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  985. if (compute_graph_ != nullptr) {
  986. compute_graph_->SetSessionID(session_id);
  987. }
  988. session_id_ = session_id;
  989. Status ret = CheckGraph();
  990. if (ret != SUCCESS) {
  991. GELOGE(ret, "RunGraph graph check fail, ret:%u", ret);
  992. return ret;
  993. }
  994. (void)compute_graph_->TopologicalSorting();
  995. ret = CheckRefOp();
  996. if (ret != SUCCESS) {
  997. GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret);
  998. return ret;
  999. }
  1000. return SUCCESS;
  1001. }
  1002. Status GraphPrepare::CheckGraph() {
  1003. if (compute_graph_ == nullptr) {
  1004. GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR");
  1005. return GE_GRAPH_INIT_FAILED;
  1006. }
  1007. auto nodes = compute_graph_->GetAllNodes();
  1008. if (nodes.empty()) {
  1009. GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph.");
  1010. return GE_GRAPH_INIT_FAILED;
  1011. }
  1012. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  1013. GE_CHECK_NOTNULL(node);
  1014. if (node->GetOpDesc() == nullptr) {
  1015. GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL");
  1016. return GE_GRAPH_INIT_FAILED;
  1017. }
  1018. }
  1019. return SUCCESS;
  1020. }
  1021. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  1022. const std::set<NodePtr> &ref_nodes) {
  1023. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  1024. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  1025. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  1026. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  1027. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  1028. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH};
  1029. GE_CHECK_NOTNULL(node);
  1030. const auto &op_desc = node->GetOpDesc();
  1031. GE_CHECK_NOTNULL(op_desc);
  1032. const auto input_index = op_desc->GetInputIndexByName(input_name);
  1033. const auto &in_anchor = node->GetInDataAnchor(input_index);
  1034. GE_CHECK_NOTNULL(in_anchor);
  1035. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  1036. GE_CHECK_NOTNULL(peer_out_anchor);
  1037. const auto &input_node = peer_out_anchor->GetOwnerNode();
  1038. GE_CHECK_NOTNULL(input_node);
  1039. const auto &input_op_desc = input_node->GetOpDesc();
  1040. GE_CHECK_NOTNULL(input_op_desc);
  1041. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  1042. if (is_ref) {
  1043. return SUCCESS;
  1044. }
  1045. auto input_type = input_op_desc->GetType();
  1046. if (input_type == ge::FRAMEWORKOP) {
  1047. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  1048. GELOGE(PARAM_INVALID, "Get original type failed.");
  1049. return PARAM_INVALID;
  1050. }
  1051. }
  1052. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  1053. if (!is_acceptable) {
  1054. ErrorManager::GetInstance().ATCReportErrMessage(
  1055. "E15005", {"opname", "optype", "opname1", "optype1"},
  1056. {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()});
  1057. GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.",
  1058. node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  1059. input_op_desc->GetType().c_str());
  1060. return PARAM_INVALID;
  1061. }
  1062. return SUCCESS;
  1063. }
  1064. Status GraphPrepare::CheckRefOp() {
  1065. GE_CHECK_NOTNULL(compute_graph_);
  1066. std::set<NodePtr> ref_nodes;
  1067. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  1068. if (node == nullptr) {
  1069. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  1070. return PARAM_INVALID;
  1071. }
  1072. auto op_desc = node->GetOpDesc();
  1073. if (op_desc == nullptr) {
  1074. GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null.");
  1075. return PARAM_INVALID;
  1076. }
  1077. auto input_name_index = op_desc->GetAllInputName();
  1078. auto outputs = op_desc->GetAllOutputName();
  1079. for (const auto &name_index : input_name_index) {
  1080. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  1081. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  1082. GELOGE(PARAM_INVALID, "CheckRefInputNode failed.");
  1083. return PARAM_INVALID;
  1084. }
  1085. (void)ref_nodes.insert(node); // no need to check value
  1086. }
  1087. }
  1088. }
  1089. return SUCCESS;
  1090. };
  1091. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  1092. GE_CHECK_NOTNULL(compute_graph_);
  1093. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  1094. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  1095. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  1096. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  1097. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  1098. return SUCCESS;
  1099. }
  1100. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1101. if (node == nullptr) {
  1102. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL");
  1103. return GE_GRAPH_GRAPH_NODE_NULL;
  1104. }
  1105. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1106. if (op_desc_ptr == nullptr) {
  1107. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL");
  1108. return GE_GRAPH_GRAPH_NODE_NULL;
  1109. }
  1110. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1111. int64_t tensor_size = 0;
  1112. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1113. if (graph_status != GRAPH_SUCCESS) {
  1114. ErrorManager::GetInstance().ATCReportErrMessage(
  1115. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1116. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  1117. return FAILED;
  1118. }
  1119. TensorUtils::SetSize(output, tensor_size);
  1120. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1121. if (graph_ret != GRAPH_SUCCESS) {
  1122. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1123. return graph_ret;
  1124. }
  1125. return SUCCESS;
  1126. }
  1127. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input, const std::map<string,string> &graph_option) {
  1128. // Get shape range of input in dynamic_execute mode
  1129. vector<vector<std::pair<int64_t,int64_t>>> dynamic_shape_range_vec;
  1130. auto ret = GetDynamicInputShapeRange(user_input, graph_option, dynamic_shape_range_vec);
  1131. GE_CHK_STATUS_RET(ret, "Graph option is not right on Dynamic execute mode.");
  1132. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1133. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1134. GE_CHECK_NOTNULL(input_node);
  1135. OpDescPtr op = input_node->GetOpDesc();
  1136. GE_CHECK_NOTNULL(op);
  1137. if (op->GetType() == DATA) {
  1138. GeAttrValue::INT index = 0;
  1139. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1140. GELOGW("Get index from data attr failed");
  1141. continue;
  1142. }
  1143. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1144. std::string situation = "data op index[" + std::to_string(index) + "]";
  1145. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1146. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1147. GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index);
  1148. return FAILED;
  1149. }
  1150. if (IsDynamicDims(input_node)) {
  1151. continue;
  1152. }
  1153. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1154. auto format = desc.GetFormat();
  1155. auto origin_format = desc.GetOriginFormat();
  1156. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1157. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op);
  1158. if (need_check_internal_format) {
  1159. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1160. if (is_internal) {
  1161. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1162. {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" +
  1163. TypeUtils::FormatToSerialString(origin_format) + "]", "it is not support"});
  1164. GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.",
  1165. TypeUtils::FormatToSerialString(format).c_str(),
  1166. TypeUtils::FormatToSerialString(origin_format).c_str());
  1167. return FAILED;
  1168. }
  1169. }
  1170. auto data_type = desc.GetDataType();
  1171. uint32_t length = 1;
  1172. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1173. if (!type_ret) {
  1174. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1175. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1176. GELOGE(PARAM_INVALID, "Input datatype %s is not support.",
  1177. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1178. return FAILED;
  1179. }
  1180. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1181. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1182. int64_t shape_size = desc_shape * length;
  1183. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1184. int64_t size = 0;
  1185. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1186. GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed");
  1187. return FAILED);
  1188. bool size_check = (size != 0 && shape_size != size);
  1189. if (size_check) {
  1190. std::string situation = "input data size[" + std::to_string(size) +
  1191. "] and shape_size[" + std::to_string(size) + "]";
  1192. std::string reason = "because size != 0 and shape_size != size";
  1193. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1194. GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size);
  1195. return FAILED;
  1196. }
  1197. ge::TensorUtils::SetSize(desc, shape_size);
  1198. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1199. if (graph_ret != GRAPH_SUCCESS) {
  1200. GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
  1201. return graph_ret;
  1202. }
  1203. // Size will be recalculated in the build stage
  1204. ge::TensorUtils::SetSize(desc, 0);
  1205. graph_ret = op->UpdateOutputDesc(0, desc);
  1206. if (graph_ret != GRAPH_SUCCESS) {
  1207. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1208. return graph_ret;
  1209. }
  1210. if (!dynamic_shape_range_vec.empty()) {
  1211. ret = UpdateDynamicInputShapeRange(index, dynamic_shape_range_vec, op, desc);
  1212. GE_CHK_STATUS_RET(ret, "Fail to update dynamic input shape range on %s.", op->GetName().c_str());
  1213. continue;
  1214. }
  1215. if (!options_.train_graph_flag) {
  1216. Status ret = AdjustDataOpOutput(input_node);
  1217. GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret); return ret);
  1218. }
  1219. }
  1220. }
  1221. return SUCCESS;
  1222. }
  1223. Status GraphPrepare::TryDoAipp() {
  1224. // infer and with aipp configure file, then call aipp insert
  1225. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1226. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1227. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1228. if (ret != SUCCESS) {
  1229. GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed.");
  1230. return INTERNAL_ERROR;
  1231. }
  1232. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1233. if (ret != SUCCESS) {
  1234. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed",
  1235. options_.insert_op_file.c_str());
  1236. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1237. }
  1238. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1239. if (ret != SUCCESS) {
  1240. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret);
  1241. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1242. }
  1243. }
  1244. return SUCCESS;
  1245. }
  1246. Status GraphPrepare::FormatAndShapeProcess() {
  1247. Status ret = ResourcePairProcess("add");
  1248. if (ret != SUCCESS) {
  1249. GELOGE(ret, "ResourcePairProcess failed");
  1250. return ret;
  1251. }
  1252. GE_TIMESTAMP_START(InferOriginFormat1);
  1253. ret = compute_graph_->InferOriginFormat();
  1254. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1255. GE_DUMP(compute_graph_, "after_first_inferformat");
  1256. if (ret != SUCCESS) {
  1257. GELOGE(ret, "Prepare Graph first inferformat failed");
  1258. return ret;
  1259. }
  1260. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1261. ret = InferShapeForPreprocess();
  1262. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1263. GE_DUMP(compute_graph_, "after_infershape");
  1264. if (ret != SUCCESS) {
  1265. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed");
  1266. return GE_GRAPH_INFERSHAPE_FAILED;
  1267. }
  1268. GE_TIMESTAMP_START(InferOriginFormat2);
  1269. ret = compute_graph_->InferOriginFormat();
  1270. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1271. if (ret != SUCCESS) {
  1272. GELOGE(ret, "Prepare Graph inferformat failed");
  1273. return ret;
  1274. }
  1275. ret = ResourcePairProcess("remove");
  1276. if (ret != SUCCESS) {
  1277. return ret;
  1278. }
  1279. return ret;
  1280. }
  1281. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1282. PassManager control_pass;
  1283. // Graph pass tmp logic for resource infershape
  1284. if (options_.train_graph_flag) {
  1285. try {
  1286. if (action == "add") {
  1287. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1288. } else {
  1289. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1290. new ResourcePairRemoveControlPass);
  1291. }
  1292. } catch (std::bad_alloc &e) {
  1293. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str());
  1294. return INTERNAL_ERROR;
  1295. }
  1296. }
  1297. Status ret = control_pass.Run(compute_graph_);
  1298. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1299. GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret);
  1300. return ret;
  1301. }
  1302. return SUCCESS;
  1303. }
  1304. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1305. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1306. GE_CHECK_NOTNULL(node_ptr);
  1307. if (node_ptr->GetType() == DATA) {
  1308. uint32_t index = 0;
  1309. auto op_desc = node_ptr->GetOpDesc();
  1310. GE_CHECK_NOTNULL(op_desc);
  1311. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1312. Format storage_format = FORMAT_RESERVED;
  1313. vector<int64_t> dst_shape_dims;
  1314. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1315. GELOGE(INTERNAL_ERROR, "Get storage format for input failed");
  1316. return FAILED;
  1317. }
  1318. if (storage_format == FORMAT_RESERVED) {
  1319. continue;
  1320. }
  1321. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1322. GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed");
  1323. return FAILED;
  1324. }
  1325. }
  1326. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1327. auto op_desc = node_ptr->GetOpDesc();
  1328. GE_CHECK_NOTNULL(op_desc);
  1329. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1330. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1331. Format storage_format = FORMAT_RESERVED;
  1332. vector<int64_t> dst_shape_dims;
  1333. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1334. GELOGE(INTERNAL_ERROR, "Get storage format from output failed");
  1335. return FAILED;
  1336. }
  1337. if (storage_format == FORMAT_RESERVED) {
  1338. continue;
  1339. }
  1340. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1341. GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed");
  1342. return FAILED;
  1343. }
  1344. }
  1345. }
  1346. }
  1347. return SUCCESS;
  1348. }
  1349. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1350. if (options_.save_original_model == "true") {
  1351. ModelHelper model_helper;
  1352. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1353. options_.original_model_file);
  1354. if (ret != SUCCESS) {
  1355. // If save original model fail, process continue
  1356. GELOGW("SaveOriginalGraphToOmModel fail");
  1357. }
  1358. }
  1359. return SUCCESS;
  1360. }
  1361. #define PP_RUN_AND_DUMP(name, func, ...) \
  1362. do { \
  1363. GE_RUN(Prepare, func, __VA_ARGS__); \
  1364. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1365. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1366. } while (0)
  1367. #define PP_RUN(name, func, ...) \
  1368. do { \
  1369. GE_RUN(Prepare, func, __VA_ARGS__); \
  1370. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1371. } while (0)
  1372. Status GraphPrepare::PrepareDynShape(const GraphNodePtr &graph_node, const std::vector<GeTensor> &user_input,
  1373. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1374. GE_CHECK_NOTNULL(graph_node->GetGraph());
  1375. GE_CHECK_NOTNULL(compute_graph);
  1376. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1377. const Graph &const_graph = *graph_node->GetGraph();
  1378. PP_RUN("Init", Init, const_graph, session_id);
  1379. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1380. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input, graph_node->GetGraph());
  1381. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1382. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1383. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1384. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1385. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1386. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1387. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1388. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1389. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1390. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1391. return SUCCESS;
  1392. }
  1393. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1394. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1395. return SUCCESS;
  1396. }
  1397. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1398. auto compute_graph = compute_graph_;
  1399. PassManager pass_manager;
  1400. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1401. new (std::nothrow) VariablePrepareOpPass))
  1402. GE_TIMESTAMP_START(pass_manager);
  1403. auto ret = pass_manager.Run(compute_graph);
  1404. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1405. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1406. GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret);
  1407. return ret;
  1408. }
  1409. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1410. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1411. return SUCCESS;
  1412. }
  1413. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1414. if (compute_graph == nullptr) {
  1415. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1416. return GE_GRAPH_NULL_INPUT;
  1417. }
  1418. GEPass ge_passes(compute_graph);
  1419. NamesToPass hccl_group;
  1420. HcclGroupPass hccl_group_pass;
  1421. GELOGD("Add hccl group pass success");
  1422. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1423. auto ret = ge_passes.Run(hccl_group);
  1424. if (ret != SUCCESS) {
  1425. GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret);
  1426. return ret;
  1427. }
  1428. ret = compute_graph->TopologicalSorting();
  1429. if (ret != SUCCESS) {
  1430. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1431. return ret;
  1432. }
  1433. return SUCCESS;
  1434. }
  1435. #undef PP_RUN_AND_DUMP
  1436. #undef PP_RUN
  1437. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1438. if (graph == nullptr) {
  1439. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1440. return GE_GRAPH_NULL_INPUT;
  1441. }
  1442. const Graph &const_graph = *graph;
  1443. Status ret = Init(const_graph, 0);
  1444. if (ret != SUCCESS) {
  1445. GELOGE(ret, "Init graph_prepare fail, ret:%u", ret);
  1446. return ret;
  1447. }
  1448. GE_DUMP(compute_graph_, "after_parser");
  1449. GELOGI("Start infershape for dump json process.");
  1450. ret = compute_graph_->InferOriginFormat();
  1451. GE_DUMP(compute_graph_, "after_inferformat");
  1452. if (ret != SUCCESS) {
  1453. GELOGE(ret, "Prepare Graph inferformat failed");
  1454. return ret;
  1455. }
  1456. InferShapePass infer_shape_pass;
  1457. NamesToPass names_to_passes;
  1458. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1459. GEPass ge_passes(compute_graph_);
  1460. ret = ge_passes.Run(names_to_passes);
  1461. GE_DUMP(compute_graph_, "after_infershape");
  1462. if (ret != SUCCESS) {
  1463. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1464. return ret;
  1465. }
  1466. ShapeRefiner::ClearContextMap();
  1467. return SUCCESS;
  1468. }
  1469. Status GraphPrepare::CheckConstOp() {
  1470. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1471. GE_CHECK_NOTNULL(node_ptr);
  1472. if (node_ptr->GetType() == CONSTANT) {
  1473. Status ret = VerifyConstOp(node_ptr);
  1474. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1475. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1476. auto op_desc = node_ptr->GetOpDesc();
  1477. if (op_desc == nullptr) {
  1478. GELOGE(PARAM_INVALID, "Get op desc failed");
  1479. return PARAM_INVALID;
  1480. }
  1481. std::string original_type;
  1482. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1483. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1484. GELOGI("original type is %s", original_type.c_str());
  1485. if (original_type == CONSTANT) {
  1486. Status ret = VerifyConstOp(node_ptr);
  1487. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1488. }
  1489. }
  1490. }
  1491. return SUCCESS;
  1492. }
  1493. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1494. GE_CHECK_NOTNULL(node);
  1495. auto op_desc = node->GetOpDesc();
  1496. GE_CHECK_NOTNULL(op_desc);
  1497. ConstGeTensorPtr ge_tensor_ptr;
  1498. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1499. GELOGE(PARAM_INVALID, "Get value from const attr failed");
  1500. return PARAM_INVALID;
  1501. }
  1502. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1503. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1504. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1505. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1506. auto data_type = ge_tensor_desc.GetDataType();
  1507. uint32_t length = 1;
  1508. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1509. if (!type_ret) {
  1510. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1511. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1512. GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str());
  1513. return FAILED;
  1514. }
  1515. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1516. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1517. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1518. if (shape_size == 0) {
  1519. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1520. // shape = [], means it's a sclar tensor.
  1521. GE_CHK_BOOL_EXEC(data_size / length == 1,
  1522. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid scalar tensor."});
  1523. return PARAM_INVALID, "Const is invalid scalar tensor.");
  1524. } else {
  1525. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1526. GE_CHK_BOOL_EXEC(data_size == 0,
  1527. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid vector scalar."});
  1528. return PARAM_INVALID, "Const is invalid vector scalar.");
  1529. }
  1530. } else {
  1531. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0,
  1532. ErrorManager::GetInstance().ATCReportErrMessage(
  1533. "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"});
  1534. return PARAM_INVALID, "Const input data size is not equal with tensor desc shape");
  1535. }
  1536. return SUCCESS;
  1537. }
  1538. bool GraphPrepare::IsDynamicDims(const NodePtr &input_node) {
  1539. auto data_shape = NodeUtils::GetOutputDesc(*input_node, kDataOutIndex).GetShape();
  1540. const auto &dims = data_shape.GetDims();
  1541. bool all_is_positive = false;
  1542. if (std::all_of(dims.begin(), dims.end(), [](int64_t val) { return val >= 0; })) {
  1543. all_is_positive = true;
  1544. }
  1545. if (!all_is_positive && !options_.input_shape.empty() && !options_.dynamic_dims.empty() &&
  1546. options_.dynamic_node_type != kInvalidDynaimcDimsType) {
  1547. GELOGI("No need to check and update desc info, the dims of %s is %s.", input_node->GetName().c_str(),
  1548. formats::JoinToString(dims).c_str());
  1549. return true;
  1550. }
  1551. return false;
  1552. }
  1553. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1554. if (GetLocalOmgContext().is_dynamic_input) {
  1555. return SUCCESS;
  1556. }
  1557. unsigned int node_num = 0;
  1558. unsigned int data_num = 0;
  1559. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1560. GE_CHECK_NOTNULL(input_node);
  1561. OpDescPtr op = input_node->GetOpDesc();
  1562. GE_CHECK_NOTNULL(op);
  1563. node_num++;
  1564. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1565. data_num++;
  1566. GeAttrValue::INT index = 0;
  1567. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1568. GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed");
  1569. return GE_GRAPH_INIT_FAILED;
  1570. }
  1571. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1572. std::string situation = "data op index[" + std::to_string(index) + "]";
  1573. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1574. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1575. GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index);
  1576. return GE_GRAPH_INIT_FAILED;
  1577. }
  1578. if (IsDynamicDims(input_node)) {
  1579. continue;
  1580. }
  1581. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1582. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1583. if (desc.GetShape().GetDim(i) < 0) {
  1584. std::string situation = "data dim[" + std::to_string(i) + "][" +
  1585. std::to_string(desc.GetShape().GetDim(i)) + "]" ;
  1586. std::string reason = "it need >= 0";
  1587. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1588. GELOGE(GE_GRAPH_INIT_FAILED, "data dim %zu is not supported, need >= 0, real:%ld.", i,
  1589. desc.GetShape().GetDim(i));
  1590. return GE_GRAPH_INIT_FAILED;
  1591. }
  1592. }
  1593. }
  1594. }
  1595. if (node_num <= data_num) {
  1596. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1597. }
  1598. return SUCCESS;
  1599. }
  1600. Status GraphPrepare::InferShapeForPreprocess() {
  1601. GELOGI("Start infershape for preprocess.");
  1602. GEPass ge_passes(compute_graph_);
  1603. NamesToPass names_to_passes;
  1604. AssertPass assert_pass;
  1605. if (!options_.train_graph_flag) {
  1606. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1607. }
  1608. InferShapePass infer_shape_pass;
  1609. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1610. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1611. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1612. DimensionComputePass dimension_compute_pass;
  1613. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1614. ConstantFoldingPass constant_folding_pass;
  1615. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1616. int32_t dev_count = 0;
  1617. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1618. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1619. rtError_t rt_err = RT_ERROR_NONE;
  1620. if (aicpu_constant_folding_on != nullptr) {
  1621. rt_err = rtGetDeviceCount(&dev_count);
  1622. if (rt_err == RT_ERROR_NONE) {
  1623. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1624. if (result != SUCCESS) {
  1625. GELOGE(result, "Set rt context failed.");
  1626. return result;
  1627. }
  1628. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1629. }
  1630. }
  1631. Status ret = ge_passes.Run(names_to_passes);
  1632. if (aicpu_constant_folding_on != nullptr) {
  1633. if (rt_err == RT_ERROR_NONE) {
  1634. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1635. if (result != SUCCESS) {
  1636. GELOGE(result, "Set rt context failed.");
  1637. return result;
  1638. }
  1639. }
  1640. }
  1641. ShapeRefiner::ClearContextMap();
  1642. if (ret != SUCCESS) {
  1643. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1644. return ret;
  1645. }
  1646. return SUCCESS;
  1647. }
  1648. Status GraphPrepare::PrepareOptimize() {
  1649. GELOGI("Start optimize for preprocess.");
  1650. // check rw type
  1651. GraphOptimize graph_optimize;
  1652. bool has_conflict = false;
  1653. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1654. if (has_conflict) {
  1655. GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize.");
  1656. return FAILED;
  1657. }
  1658. PassManager original_graph_passes;
  1659. // Graph pass
  1660. try {
  1661. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1662. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1663. (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass);
  1664. } catch (std::bad_alloc &e) {
  1665. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1666. return INTERNAL_ERROR;
  1667. }
  1668. GE_TIMESTAMP_START(original_graph_passes);
  1669. Status ret = original_graph_passes.Run(compute_graph_);
  1670. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1671. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1672. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1673. return ret;
  1674. }
  1675. // New pass
  1676. GEPass ge_passes(compute_graph_);
  1677. NamesToPass names_to_passes;
  1678. EnterPass enter_pass;
  1679. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1680. CondPass cond_pass;
  1681. names_to_passes.emplace_back("CondPass", &cond_pass);
  1682. PrintOpPass print_pass;
  1683. if (options_.enable_print_op_pass) {
  1684. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1685. }
  1686. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1687. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1688. DropOutPass dropout_pass;
  1689. AssertPass assert_pass;
  1690. UnusedConstPass unused_const_pass;
  1691. StopGradientPass stop_gradient_pass;
  1692. PreventGradientPass prevent_gradient_pass;
  1693. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1694. GuaranteeConstPass guarantee_const_pass;
  1695. VarIsInitializedOpPass var_is_initialized_pass;
  1696. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1697. IdentityPass identity_pass(false);
  1698. #ifdef ONLY_COMPILE_OPEN_SRC
  1699. AssignRemovePass assign_remove_pass;
  1700. #endif
  1701. SnapshotPass snapshot_pass;
  1702. if (!options_.train_graph_flag) {
  1703. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1704. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1705. }
  1706. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1707. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1708. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1709. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1710. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1711. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1712. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1713. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1714. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1715. #ifdef ONLY_COMPILE_OPEN_SRC
  1716. if (GetContext().GetHostExecFlag()) {
  1717. names_to_passes.emplace_back("AssignRemovePass", &assign_remove_pass);
  1718. }
  1719. #endif
  1720. GE_TIMESTAMP_START(names_to_passes);
  1721. ret = ge_passes.Run(names_to_passes);
  1722. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1723. if (ret != SUCCESS) {
  1724. GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret);
  1725. return ret;
  1726. }
  1727. PassManager graph_pass;
  1728. try {
  1729. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1730. } catch (std::bad_alloc &e) {
  1731. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1732. return INTERNAL_ERROR;
  1733. }
  1734. GE_TIMESTAMP_START(graph_passes);
  1735. ret = graph_pass.Run(compute_graph_);
  1736. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1737. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1738. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1739. return ret;
  1740. }
  1741. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1742. TypeConversionOfConstant();
  1743. ret = compute_graph_->TopologicalSorting();
  1744. if (ret != SUCCESS) {
  1745. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1746. return ret;
  1747. }
  1748. GELOGI("End optimize for preprocess.");
  1749. return SUCCESS;
  1750. }
  1751. void GraphPrepare::TypeConversionOfConstant() {
  1752. bool is_acl_compile = false;
  1753. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1754. // This can ensure that n is not a null pointer
  1755. // No Conversion when called by aclOpCompile
  1756. (void)AttrUtils::GetBool(n->GetOpDesc(), ATTR_DYNAMIC_SHAPE_SINGLE_AICPU, is_acl_compile);
  1757. if (is_acl_compile) {
  1758. return;
  1759. }
  1760. }
  1761. if (options_.train_graph_flag) {
  1762. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1763. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1764. // This can ensure that n is not a null pointer
  1765. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1766. n->GetOpDesc()->SetType(CONSTANTOP);
  1767. }
  1768. }
  1769. } else {
  1770. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  1771. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1772. // This can ensure that n is not a null pointer
  1773. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  1774. n->GetOpDesc()->SetType(CONSTANT);
  1775. }
  1776. }
  1777. }
  1778. }
  1779. Status GraphPrepare::GraphEquivalentTransformation() {
  1780. NamesToPass names_to_pass;
  1781. ForPass for_pass;
  1782. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  1783. return GEPass(compute_graph_).Run(names_to_pass);
  1784. }
  1785. Status GraphPrepare::ProcessBeforeInfershape() {
  1786. NamesToPass names_to_passes;
  1787. CondRemovePass condition_remove_pass;
  1788. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  1789. GE_TIMESTAMP_START(ProcessCondRemove);
  1790. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  1791. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  1792. if (ret != SUCCESS) {
  1793. GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  1794. return ret;
  1795. }
  1796. return SUCCESS;
  1797. }
  1798. Status GraphPrepare::ProcessNetOutput() {
  1799. PassManager graph_passes_before_infershape;
  1800. try {
  1801. if (options_.train_graph_flag) {
  1802. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  1803. }
  1804. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  1805. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  1806. new (std::nothrow) DataPass); // Add NetOutput first.
  1807. } catch (std::bad_alloc) {
  1808. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1809. return INTERNAL_ERROR;
  1810. }
  1811. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  1812. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  1813. GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret);
  1814. return ret;
  1815. }
  1816. return SUCCESS;
  1817. }
  1818. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input,const std::map<string,string> &graph_option) {
  1819. compute_graph_->SetInputSize(user_input.size());
  1820. if (user_input.empty()) {
  1821. return SUCCESS;
  1822. }
  1823. auto ret = CheckUserInput(user_input);
  1824. if (ret != SUCCESS) {
  1825. GELOGE(ret, "Check user input failed.");
  1826. return ret;
  1827. }
  1828. ret = UpdateInput(user_input, graph_option);
  1829. if (ret != SUCCESS) {
  1830. GELOGE(ret, "UpdateInput fail, ret:%u", ret);
  1831. return ret;
  1832. }
  1833. if (user_input.size() != 0) {
  1834. ret = CheckConstOp();
  1835. if (ret != SUCCESS) {
  1836. GELOGE(ret, "CheckConstOp fail, ret:%u", ret);
  1837. return ret;
  1838. }
  1839. } else {
  1840. ret = compute_graph_->TopologicalSorting();
  1841. if (ret != SUCCESS) {
  1842. GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting");
  1843. return FAILED;
  1844. }
  1845. }
  1846. return SUCCESS;
  1847. }
  1848. Status GraphPrepare::UpdateInputOutputByOptions() {
  1849. auto ret = UpdateDataNetOutputByStorageFormat();
  1850. if (ret != SUCCESS) {
  1851. GELOGE(ret, "Update format acoording to storage format failed.");
  1852. return ret;
  1853. }
  1854. if (options_.train_graph_flag) {
  1855. GELOGI("This is train mode, no need to do this schedule.");
  1856. return SUCCESS;
  1857. }
  1858. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  1859. GE_CHECK_NOTNULL(node_ptr);
  1860. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  1861. GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str());
  1862. return FAILED;
  1863. }
  1864. if (node_ptr->GetType() == DATA) {
  1865. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  1866. GELOGE(INTERNAL_ERROR, "Process data node failed");
  1867. return FAILED;
  1868. }
  1869. }
  1870. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1871. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  1872. GELOGE(INTERNAL_ERROR, "Process netoutput node failed");
  1873. return FAILED;
  1874. }
  1875. }
  1876. }
  1877. return SUCCESS;
  1878. }
  1879. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  1880. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  1881. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  1882. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  1883. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  1884. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  1885. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  1886. if (dst_node->GetType() == TRANSDATA) {
  1887. return true;
  1888. }
  1889. }
  1890. }
  1891. return false;
  1892. }
  1893. } // 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.