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 100 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
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
4 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #include "graph/preprocess/graph_preprocess.h"
  17. #include <map>
  18. #include <set>
  19. #include <string>
  20. #include "common/formats/format_transfers/format_transfer_fractal_nz.h"
  21. #include "common/formats/format_transfers/format_transfer_nchw_nc1hwc0.h"
  22. #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h"
  23. #include "common/formats/format_transfers/format_transfer_transpose.h"
  24. #include "common/formats/utils/formats_trans_utils.h"
  25. #include "common/util/error_manager/error_manager.h"
  26. #include "common/helper/model_helper.h"
  27. #include "common/math/math_util.h"
  28. #include "common/op/ge_op_utils.h"
  29. #include "graph/common/ge_call_wrapper.h"
  30. #include "graph/common/local_context.h"
  31. #include "graph/common/transop_util.h"
  32. #include "graph/ge_context.h"
  33. #include "graph/shape_refiner.h"
  34. #include "graph/manager/graph_var_manager.h"
  35. #include "graph/manager/util/rt_context_util.h"
  36. #include "graph/optimize/graph_optimize.h"
  37. #include "graph/passes/addn_pass.h"
  38. #include "graph/passes/aicpu_constant_folding_pass.h"
  39. #include "graph/passes/assert_pass.h"
  40. #include "ge/ge_api_types.h"
  41. #include "graph/passes/common_subexpression_elimination_pass.h"
  42. #include "graph/passes/cond_pass.h"
  43. #include "graph/passes/cond_remove_pass.h"
  44. #include "graph/passes/constant_folding_pass.h"
  45. #include "graph/passes/dimension_adjust_pass.h"
  46. #include "graph/passes/dimension_compute_pass.h"
  47. #include "graph/passes/dropout_pass.h"
  48. #include "graph/passes/enter_pass.h"
  49. #include "graph/passes/for_pass.h"
  50. #include "graph/passes/guarantee_const_pass.h"
  51. #include "graph/passes/hccl_memcpy_pass.h"
  52. #include "graph/passes/hccl_group_pass.h"
  53. #include "graph/passes/identity_pass.h"
  54. #include "graph/passes/infershape_pass.h"
  55. #include "graph/passes/merge_pass.h"
  56. #include "graph/passes/net_output_pass.h"
  57. #include "graph/passes/no_use_reshape_remove_pass.h"
  58. #include "graph/passes/parallel_concat_start_op_pass.h"
  59. #include "graph/passes/placeholder_with_default_pass.h"
  60. #include "graph/passes/prevent_gradient_pass.h"
  61. #include "graph/passes/print_op_pass.h"
  62. #include "graph/passes/prune_pass.h"
  63. #include "graph/passes/replace_transshape_pass.h"
  64. #include "graph/passes/replace_with_empty_const_pass.h"
  65. #include "graph/passes/resource_pair_add_control_pass.h"
  66. #include "graph/passes/resource_pair_remove_control_pass.h"
  67. #include "graph/passes/save_pass.h"
  68. #include "graph/passes/shape_operate_op_remove_pass.h"
  69. #include "graph/passes/snapshot_pass.h"
  70. #include "graph/passes/stop_gradient_pass.h"
  71. #include "graph/passes/switch_dead_branch_elimination.h"
  72. #include "graph/passes/unused_const_pass.h"
  73. #include "graph/passes/var_is_initialized_op_pass.h"
  74. #include "graph/passes/variable_prepare_op_pass.h"
  75. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  76. #include "graph/utils/type_utils.h"
  77. #include "inc/pass_manager.h"
  78. #include "init/gelib.h"
  79. #include "multi_batch_copy_graph.h"
  80. #include "graph/passes/data_pass.h"
  81. #include "graph/passes/mark_agnostic_pass.h"
  82. namespace ge {
  83. namespace {
  84. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  85. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  86. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  87. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  88. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  89. // the size of user defined output datatype or format string after split by ":".
  90. const size_t kUserDefinedElementCount = 2;
  91. const int kDataOutIndex = 0;
  92. const int64_t kInvalidDynaimcDimsType = -1;
  93. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  94. GeTensorPtr tensor = MakeShared<GeTensor>();
  95. if (tensor == nullptr) {
  96. REPORT_CALL_ERROR("E19999", "New GeTensor failed when %s", __FUNCTION__);
  97. GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed");
  98. return nullptr;
  99. }
  100. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  101. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  102. auto dst_ge_shape = data_tensor.GetShape();
  103. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  104. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  105. tensor->MutableTensorDesc().SetShape(GeShape());
  106. int32_t dst_shape = 1;
  107. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  108. REPORT_CALL_ERROR("E19999", "Set data to tensor failed when %s", __FUNCTION__);
  109. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  110. return nullptr;
  111. }
  112. } else {
  113. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  114. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  115. if (dst_shape == nullptr) {
  116. REPORT_CALL_ERROR("E19999", "Malloc buffer failed, size:%zu, when %s", dim_cnt, __FUNCTION__);
  117. GELOGE(INTERNAL_ERROR, "Create unique ptr failed");
  118. return nullptr;
  119. }
  120. for (int64_t i = 0; i < dim_cnt; ++i) {
  121. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  122. }
  123. GE_IF_BOOL_EXEC(
  124. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  125. REPORT_CALL_ERROR("E19999", "Set data to tensor failed when %s", __FUNCTION__);
  126. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  127. return nullptr;)
  128. }
  129. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  130. return OpDescUtils::CreateConstOp(tensor);
  131. }
  132. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  133. OpDescPtr &op_desc) {
  134. // For format transfer node, the IR definition has src/dst format attrs
  135. if (node_type == TRANSDATA) {
  136. GE_IF_BOOL_EXEC(
  137. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  138. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  139. GE_IF_BOOL_EXEC(
  140. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  141. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  142. }
  143. // For TransposeD node, the IR definition has perm attrs
  144. if (node_type == TRANSPOSED) {
  145. Format src_format = input.GetFormat();
  146. Format dst_format = output.GetFormat();
  147. std::vector<int64_t> perm_arg;
  148. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  149. "Get perm by foramt failed.");
  150. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  151. "SetStr PERMUTE_ATTR_PERM failed")
  152. }
  153. // For cast node, the IR definition has src/dst attrs
  154. if (node_type == CAST) {
  155. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  156. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  157. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  158. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  159. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  160. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  161. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  162. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  163. }
  164. }
  165. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  166. const GeTensorDesc &output, NodePtr &node) {
  167. if (node == nullptr) {
  168. REPORT_INNER_ERROR("E19999", "Param node is nullptr, trans_name:%s, trans_type:%s, check invalid when %s",
  169. name.c_str(), node_type.c_str(), __FUNCTION__);
  170. GELOGE(PARAM_INVALID, "node is null.");
  171. return nullptr;
  172. }
  173. auto graph = node->GetOwnerComputeGraph();
  174. if (graph == nullptr) {
  175. REPORT_INNER_ERROR("E19999", "Owner graph in node is nullptr, trans_name:%s, trans_type:%s, check invalid when %s",
  176. name.c_str(), node_type.c_str(), __FUNCTION__);
  177. GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str());
  178. return nullptr;
  179. }
  180. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  181. if (index < 0) {
  182. ErrorManager::GetInstance().ATCReportErrMessage(
  183. "E19025", {"situation", "reason"},
  184. {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()});
  185. GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str());
  186. return nullptr;
  187. }
  188. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  189. if (op_desc == nullptr) {
  190. REPORT_CALL_ERROR("E19999", "New OpDesc failed, trans_name:%s, trans_type:%s, when %s",
  191. name.c_str(), node_type.c_str(), __FUNCTION__);
  192. GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed");
  193. return nullptr;
  194. }
  195. // for data dump
  196. GE_IF_BOOL_EXEC(
  197. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  198. GELOGW("CreateTransNode: SetListStr failed");)
  199. // Default single input and single output
  200. auto ret = op_desc->AddInputDesc(input);
  201. if (ret != GRAPH_SUCCESS) {
  202. REPORT_CALL_ERROR("E19999", "Add input desc into op:%s(%s) failed when %s",
  203. op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  204. GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str());
  205. return nullptr;
  206. }
  207. ret = op_desc->AddOutputDesc(output);
  208. if (ret != GRAPH_SUCCESS) {
  209. REPORT_CALL_ERROR("E19999", "Add output desc into op:%s(%s) failed when %s",
  210. op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  211. GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str());
  212. return nullptr;
  213. }
  214. AddTransNodeAttr(node_type, input, output, op_desc);
  215. NodePtr shape_node = nullptr;
  216. if (node_type == RESHAPE) {
  217. auto shape_desc = CreateTensorShape(output);
  218. if (shape_desc == nullptr) {
  219. GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input",
  220. node->GetName().c_str());
  221. return nullptr;
  222. }
  223. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  224. if (ret != GRAPH_SUCCESS) {
  225. REPORT_CALL_ERROR("E19999", "Add input desc into op:%s(%s) failed when %s",
  226. op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  227. GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str());
  228. return nullptr;
  229. }
  230. shape_node = graph->AddNode(shape_desc);
  231. if (shape_node == nullptr) {
  232. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed when %s",
  233. shape_desc->GetName().c_str(), shape_desc->GetType().c_str(),
  234. graph->GetName().c_str(), __FUNCTION__);
  235. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str());
  236. return nullptr;
  237. }
  238. }
  239. auto trans_node = graph->AddNode(op_desc);
  240. if (trans_node == nullptr) {
  241. REPORT_CALL_ERROR("E19999", "Add node:%s(%s) to graph:%s failed when %s",
  242. op_desc->GetName().c_str(), op_desc->GetType().c_str(),
  243. graph->GetName().c_str(), __FUNCTION__);
  244. GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str());
  245. return nullptr;
  246. }
  247. if (node_type == RESHAPE) {
  248. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  249. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed "
  250. "when %s", shape_node->GetName().c_str(), shape_node->GetType().c_str(),
  251. trans_node->GetName().c_str(), trans_node->GetType().c_str(), __FUNCTION__);
  252. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str());
  253. return nullptr;
  254. }
  255. }
  256. return trans_node;
  257. }
  258. Status RecoverOneTransNodeForVar(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.output, trans_node_info.input, 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. REPORT_CALL_ERROR("E19999", "Replace out anchors of node:%s(%s) by node:%s(%s) failed when %s",
  268. node->GetName().c_str(), node->GetType().c_str(),
  269. trans_node->GetName().c_str(), trans_node->GetType().c_str(), __FUNCTION__);
  270. GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s",
  271. node->GetName().c_str(), node->GetType().c_str());
  272. return INTERNAL_ERROR;
  273. }
  274. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  275. if (ret != GRAPH_SUCCESS) {
  276. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed when %s",
  277. node->GetName().c_str(), node->GetType().c_str(),
  278. trans_node->GetName().c_str(), trans_node->GetType().c_str(), __FUNCTION__);
  279. GELOGE(INTERNAL_ERROR, "Failed to connect node %s to trans node %s", node->GetName().c_str(),
  280. trans_node->GetName().c_str());
  281. return INTERNAL_ERROR;
  282. }
  283. ret = GraphUtils::MoveOutCtrlEdges(node, trans_node);
  284. if (ret != GRAPH_SUCCESS) {
  285. REPORT_CALL_ERROR("E19999", "Move out control edges from node:%s(%s) to node:%s(%s) failed when %s",
  286. node->GetName().c_str(), node->GetType().c_str(),
  287. trans_node->GetName().c_str(), trans_node->GetType().c_str(), __FUNCTION__);
  288. GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.",
  289. node->GetName().c_str(), trans_node->GetName().c_str());
  290. return INTERNAL_ERROR;
  291. }
  292. return SUCCESS;
  293. }
  294. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  295. NodePtr &trans_node) {
  296. GE_CHECK_NOTNULL(node);
  297. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  298. if (trans_node == nullptr) {
  299. return INTERNAL_ERROR;
  300. }
  301. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  302. if (ret != GRAPH_SUCCESS) {
  303. REPORT_CALL_ERROR("E19999", "Replace out anchors of node:%s(%s) by node:%s(%s) failed when %s",
  304. node->GetName().c_str(), node->GetType().c_str(),
  305. trans_node->GetName().c_str(), trans_node->GetType().c_str(), __FUNCTION__);
  306. GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s",
  307. node->GetName().c_str(), node->GetType().c_str());
  308. return INTERNAL_ERROR;
  309. }
  310. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  311. if (ret != GRAPH_SUCCESS) {
  312. REPORT_CALL_ERROR("E19999", "Add edge between op:%s(%s)(out_index:0) and op:%s(%s)(in_index:0) failed when %s",
  313. trans_node->GetName().c_str(), trans_node->GetType().c_str(),
  314. node->GetName().c_str(), node->GetType().c_str(), __FUNCTION__);
  315. GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(),
  316. node->GetName().c_str());
  317. return INTERNAL_ERROR;
  318. }
  319. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  320. if (ret != GRAPH_SUCCESS) {
  321. REPORT_CALL_ERROR("E19999", "Move out control edges from node:%s(%s) to node:%s(%s) failed when %s",
  322. node->GetName().c_str(), node->GetType().c_str(),
  323. trans_node->GetName().c_str(), trans_node->GetType().c_str(), __FUNCTION__);
  324. GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.",
  325. node->GetName().c_str(), trans_node->GetName().c_str());
  326. return INTERNAL_ERROR;
  327. }
  328. return SUCCESS;
  329. }
  330. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  331. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  332. GE_CHECK_NOTNULL(var->GetOpDesc());
  333. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  334. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  335. output_desc.SetFormat(tensor_desc.GetFormat());
  336. output_desc.SetDataType(tensor_desc.GetDataType());
  337. output_desc.SetShape(tensor_desc.GetShape());
  338. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  339. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  340. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  341. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  342. REPORT_CALL_ERROR("E19999", "Update output desc of node:%s(%s) failed, index:0, when %s",
  343. var->GetName().c_str(), var->GetType().c_str(), __FUNCTION__);
  344. GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed");
  345. return INTERNAL_ERROR;);
  346. }
  347. if (var->GetOpDesc()->GetInputsSize() > 0) {
  348. auto desc = var->GetOpDesc()->GetInputDesc(0);
  349. desc.SetFormat(tensor_desc.GetFormat());
  350. desc.SetDataType(tensor_desc.GetDataType());
  351. desc.SetShape(tensor_desc.GetShape());
  352. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  353. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  354. desc.SetOriginShape(tensor_desc.GetOriginShape());
  355. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  356. REPORT_CALL_ERROR("E19999", "Update input desc of node:%s(%s) failed, index:0, when %s",
  357. var->GetName().c_str(), var->GetType().c_str(), __FUNCTION__);
  358. GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed");
  359. return INTERNAL_ERROR;)
  360. }
  361. return SUCCESS;
  362. }
  363. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  364. GE_CHECK_NOTNULL(var);
  365. int index = 0;
  366. NodePtr last_node = var;
  367. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  368. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  369. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  370. if (ret != SUCCESS) {
  371. ErrorManager::GetInstance().ATCReportErrMessage(
  372. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  373. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(),
  374. index, iter->node_type.c_str());
  375. return INTERNAL_ERROR;
  376. }
  377. // set stream_label
  378. OpDescPtr var_desc = var->GetOpDesc();
  379. GE_CHECK_NOTNULL(var_desc);
  380. std::string stream_label;
  381. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  382. if (!stream_label.empty()) {
  383. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  384. }
  385. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  386. REPORT_CALL_ERROR("E19999", "Set Attr:%s of node:%s(%s) failed when %s",
  387. ge::ATTR_INSERTED_BY_GE.c_str(),
  388. last_node->GetName().c_str(), last_node->GetType().c_str(), __FUNCTION__);
  389. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  390. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  391. }
  392. if (road.empty()) {
  393. return SUCCESS;
  394. }
  395. return UpdateVarFormats(var, road.rbegin()->output);
  396. }
  397. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  398. for (auto &var : nodes) {
  399. GE_CHECK_NOTNULL(var);
  400. int index = 0;
  401. NodePtr last_node = var;
  402. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  403. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  404. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  405. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  406. if (ret != SUCCESS) {
  407. ErrorManager::GetInstance().ATCReportErrMessage(
  408. "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type});
  409. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s",
  410. var->GetName().c_str(), index, iter->node_type.c_str());
  411. return INTERNAL_ERROR;
  412. }
  413. // set stream_label
  414. OpDescPtr var_desc = var->GetOpDesc();
  415. GE_CHECK_NOTNULL(var_desc);
  416. std::string stream_label;
  417. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  418. if (!stream_label.empty()) {
  419. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  420. }
  421. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  422. REPORT_CALL_ERROR("E19999", "Set Attr:%s of node:%s(%s) failed when %s",
  423. ge::ATTR_INSERTED_BY_GE.c_str(),
  424. last_node->GetName().c_str(), last_node->GetType().c_str(), __FUNCTION__);
  425. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  426. }
  427. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  428. return INTERNAL_ERROR;
  429. }
  430. }
  431. return SUCCESS;
  432. }
  433. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  434. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  435. VarNamesToRefs names_to_refs;
  436. std::string var_name;
  437. if (graph == nullptr) {
  438. REPORT_INNER_ERROR("E19999", "Param graph is nullptr, check invalid when %s", __FUNCTION__);
  439. GELOGE(PARAM_INVALID, "graph is null.");
  440. return names_to_refs;
  441. }
  442. for (auto &node : graph->GetAllNodes()) {
  443. if (node->GetType() != VARIABLE) {
  444. continue;
  445. }
  446. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  447. (void)names_to_refs[var_name].insert(node);
  448. }
  449. }
  450. return names_to_refs;
  451. }
  452. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  453. std::vector<int64_t> &dst_shape) {
  454. if (src_format == FORMAT_NCHW) {
  455. formats::FormatTransferNchwNc1hwc0 transfer;
  456. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  457. GELOGE(INTERNAL_ERROR, "TransShape failed");
  458. return FAILED;
  459. }
  460. } else if (src_format == FORMAT_NHWC) {
  461. formats::FormatTransferNhwcNc1hwc0 transfer;
  462. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  463. GELOGE(INTERNAL_ERROR, "TransShape failed");
  464. return FAILED;
  465. }
  466. }
  467. return SUCCESS;
  468. }
  469. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  470. GE_CHECK_NOTNULL(node_ptr);
  471. auto op_desc = node_ptr->GetOpDesc();
  472. GE_CHECK_NOTNULL(op_desc);
  473. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  474. GE_CHECK_NOTNULL(input);
  475. ge::Format old_format = input->GetFormat();
  476. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  477. ge::DataType dt = input->GetDataType();
  478. std::vector<int64_t> dst_shape_dims;
  479. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  480. REPORT_CALL_ERROR("E19999", "Transfer shape to NC1HWC0 failed, op:%s(%s), when %s",
  481. op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  482. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  483. return FAILED;
  484. }
  485. input->SetFormat(FORMAT_NC1HWC0);
  486. input->SetShape(ge::GeShape(dst_shape_dims));
  487. auto output = op_desc->MutableOutputDesc(0);
  488. GE_CHECK_NOTNULL(output);
  489. output->SetFormat(FORMAT_NC1HWC0);
  490. output->SetShape(ge::GeShape(dst_shape_dims));
  491. int64_t size = 0;
  492. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  493. if (graph_status != ge::GRAPH_SUCCESS) {
  494. REPORT_CALL_ERROR("E19999", "Get output tensor size failed, op:%s(%s), index:0 when %s",
  495. op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  496. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  497. return FAILED;
  498. }
  499. ge::TensorUtils::SetSize(*output, size);
  500. ge::TensorUtils::SetSize(*input, size);
  501. return SUCCESS;
  502. }
  503. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  504. GE_CHECK_NOTNULL(input_output);
  505. ge::Format old_format = input_output->GetFormat();
  506. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  507. ge::DataType dt = input_output->GetDataType();
  508. std::vector<int64_t> dst_shape_dims;
  509. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  510. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  511. return FAILED;
  512. }
  513. input_output->SetFormat(FORMAT_NC1HWC0);
  514. input_output->SetShape(ge::GeShape(dst_shape_dims));
  515. return SUCCESS;
  516. }
  517. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  518. vector<int64_t> &dst_shape_dims) {
  519. GE_CHECK_NOTNULL(op_desc);
  520. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  521. GE_CHECK_NOTNULL(input);
  522. ge::Format old_format = input->GetFormat();
  523. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  524. input->SetShape(ge::GeShape(dst_shape_dims));
  525. input->SetFormat(storage_format);
  526. auto output = op_desc->MutableOutputDesc(index);
  527. GE_CHECK_NOTNULL(output);
  528. output->SetShape(ge::GeShape(dst_shape_dims));
  529. output->SetFormat(storage_format);
  530. if (!output->MutableShape().IsUnknownShape()) {
  531. int64_t size = 0;
  532. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  533. if (graph_status != ge::GRAPH_SUCCESS) {
  534. REPORT_CALL_ERROR("E19999", "Get output tensor size failed, op:%s(%s), index:%u when %s",
  535. op_desc->GetName().c_str(), op_desc->GetType().c_str(), index, __FUNCTION__);
  536. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  537. return FAILED;
  538. }
  539. ge::TensorUtils::SetSize(*input, size);
  540. ge::TensorUtils::SetSize(*output, size);
  541. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  542. "new_shape:%s, new_format:%s, new_size:%lu",
  543. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  544. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  545. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  546. }
  547. return SUCCESS;
  548. }
  549. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) {
  550. is_dynamic_batch = false;
  551. std::string related_node_name;
  552. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  553. if (related_node_name.empty()) {
  554. ErrorManager::GetInstance().ATCReportErrMessage(
  555. "E15002", {"opname", "value", "reason"}, {data_node->GetName(), "flag", "but the value is empty"});
  556. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty",
  557. data_node->GetName().c_str());
  558. return INTERNAL_ERROR;
  559. }
  560. for (const NodePtr &next_node : data_node->GetOutNodes()) {
  561. if (next_node->GetName() == related_node_name) {
  562. switchn_node = next_node;
  563. break;
  564. }
  565. }
  566. if (switchn_node == nullptr) {
  567. ErrorManager::GetInstance().ATCReportErrMessage(
  568. "E15002", {"opname", "value", "reason"},
  569. {data_node->GetName(), related_node_name, "but can not find it on the graph"});
  570. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph",
  571. data_node->GetName().c_str(), related_node_name.c_str());
  572. return INTERNAL_ERROR;
  573. }
  574. is_dynamic_batch = true;
  575. }
  576. return SUCCESS;
  577. }
  578. bool CheckOpType(const NodePtr &node, const std::string type) {
  579. if (node->GetType() == type) {
  580. return true;
  581. }
  582. return false;
  583. }
  584. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  585. auto op = node_ptr->GetOpDesc();
  586. GE_CHECK_NOTNULL(op);
  587. auto inputDescsPtr = op->GetAllInputsDescPtr();
  588. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  589. ge::Format format = ge::FORMAT_ND;
  590. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  591. // process, otherwise fe will insert transdata.
  592. for (auto &inputDescPtr : inputDescsPtr) {
  593. GE_CHECK_NOTNULL(inputDescPtr);
  594. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  595. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  596. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  597. inputDescPtr->SetFormat(format);
  598. inputDescPtr->SetOriginFormat(format);
  599. }
  600. }
  601. for (auto &outputDescPtr : outputDescsPtr) {
  602. GE_CHECK_NOTNULL(outputDescPtr);
  603. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  604. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  605. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  606. outputDescPtr->SetFormat(format);
  607. outputDescPtr->SetOriginFormat(format);
  608. }
  609. }
  610. return SUCCESS;
  611. }
  612. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  613. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  614. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  615. // data and netoutput, and no longer inserts the transnode.
  616. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) {
  617. GE_CHECK_NOTNULL(node_ptr);
  618. auto op_desc = node_ptr->GetOpDesc();
  619. GE_CHECK_NOTNULL(op_desc);
  620. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  621. GE_CHECK_NOTNULL(input);
  622. ge::DataType src_dtype = input->GetDataType();
  623. if (src_dtype == dt_set) {
  624. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  625. return SUCCESS;
  626. }
  627. input->SetDataType(dt_set);
  628. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  629. GE_CHECK_NOTNULL(output);
  630. output->SetDataType(dt_set);
  631. GeShape shape = input->GetShape();
  632. if (!shape.IsUnknownShape()) {
  633. int64_t input_shape_size = 0;
  634. int64_t output_shape_size = 0;
  635. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  636. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  637. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  638. REPORT_CALL_ERROR("E19999", "Get input tensor size failed, op:%s(%s), index:0 when %s",
  639. op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  640. GELOGE(GRAPH_FAILED, "[Process][InputOp] Get tensor size of op [%s] failed!", node_ptr->GetName().c_str());
  641. return FAILED;
  642. }
  643. ge::TensorUtils::SetSize(*input, input_shape_size);
  644. ge::TensorUtils::SetSize(*output, output_shape_size);
  645. GELOGI("[Process][InputDynShape] Set input and output size of node [%s] success.", node_ptr->GetName().c_str());
  646. }
  647. if (is_dynamic_batch) {
  648. GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str());
  649. auto switchn_op_desc = switchn_node->GetOpDesc();
  650. GE_CHECK_NOTNULL(switchn_op_desc);
  651. auto switchn_input = switchn_op_desc->MutableInputDesc(0);
  652. GE_CHECK_NOTNULL(switchn_input);
  653. switchn_input->SetDataType(dt_set);
  654. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  655. const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i);
  656. GE_CHECK_NOTNULL(switchn_output);
  657. switchn_output->SetDataType(dt_set);
  658. }
  659. }
  660. return SUCCESS;
  661. }
  662. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  663. GE_CHECK_NOTNULL(node_ptr);
  664. auto op_desc = node_ptr->GetOpDesc();
  665. GE_CHECK_NOTNULL(op_desc);
  666. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  667. GE_CHECK_NOTNULL(input);
  668. ge::Format old_format = input->GetFormat();
  669. ge::GeShape old_shape = input->GetShape();
  670. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  671. if (!support) {
  672. ErrorManager::GetInstance().ATCReportErrMessage(
  673. "E19014", {"opname", "value", "reason"},
  674. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  675. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  676. GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str());
  677. return FAILED;
  678. }
  679. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  680. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  681. return FAILED;
  682. }
  683. if (is_dynamic_batch) {
  684. auto switchn_op_desc = switchn_node->GetOpDesc();
  685. GE_CHECK_NOTNULL(switchn_op_desc);
  686. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  687. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  688. REPORT_CALL_ERROR("E19999", "Modify format and shape of input:0 in op:%s(%s) failed when %s",
  689. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str(), __FUNCTION__);
  690. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  691. return FAILED;
  692. }
  693. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  694. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  695. GE_CHECK_NOTNULL(switchn_output);
  696. old_format = switchn_output->GetFormat();
  697. old_shape = switchn_output->GetShape();
  698. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  699. REPORT_CALL_ERROR("E19999", "Modify format and shape of output:%u in op:%s(%s) failed when %s", i,
  700. switchn_op_desc->GetName().c_str(), switchn_op_desc->GetType().c_str(), __FUNCTION__);
  701. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  702. return FAILED;
  703. }
  704. }
  705. }
  706. return SUCCESS;
  707. }
  708. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  709. auto op_desc = node_ptr->GetOpDesc();
  710. GE_CHECK_NOTNULL(op_desc);
  711. string set_dt_str;
  712. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  713. return SUCCESS;
  714. }
  715. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  716. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  717. bool is_dynamic_batch = false;
  718. NodePtr switchn_node = nullptr;
  719. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) {
  720. GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed");
  721. return FAILED;
  722. }
  723. if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) {
  724. GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed");
  725. return FAILED;
  726. }
  727. // check if need to set format
  728. string set_format;
  729. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  730. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  731. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  732. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) {
  733. GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed");
  734. return FAILED;
  735. }
  736. }
  737. return SUCCESS;
  738. }
  739. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  740. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  741. GE_CHECK_NOTNULL(op_desc);
  742. GE_CHECK_NOTNULL(tensor_desc_ptr);
  743. storage_format = FORMAT_RESERVED;
  744. int64_t format = FORMAT_RESERVED;
  745. dst_shape_dims.clear();
  746. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  747. storage_format = static_cast<Format>(format);
  748. vector<int32_t> storage_shape;
  749. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  750. for (auto dim : storage_shape) {
  751. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  752. }
  753. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  754. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  755. formats::JoinToString(storage_shape).c_str());
  756. } else {
  757. ErrorManager::GetInstance().ATCReportErrMessage(
  758. "15003", {"opname", "format"},
  759. {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)});
  760. GELOGE(PARAM_INVALID, "Update node by storage format failed, storage_shape not set. "
  761. "node: [%s], storage_format [%s]",
  762. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  763. return FAILED;
  764. }
  765. ge::Format old_format = tensor_desc_ptr->GetFormat();
  766. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  767. if (old_format == storage_format && old_shape == dst_shape_dims) {
  768. GELOGI("Update node by storage format, not changed.");
  769. storage_format = FORMAT_RESERVED;
  770. return SUCCESS;
  771. }
  772. }
  773. return SUCCESS;
  774. }
  775. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  776. NodePtr &node) {
  777. bool is_dynamic = CheckOpType(node, MERGE);
  778. auto src_op_desc = node->GetOpDesc();
  779. GE_CHECK_NOTNULL(src_op_desc);
  780. ge::GeShape src_shape = src_desc.GetShape();
  781. ge::Format src_format = src_desc.GetFormat();
  782. net_output_input_desc->SetDataType(DT_FLOAT16);
  783. if (is_dynamic) {
  784. auto merge_output = src_op_desc->MutableOutputDesc(0);
  785. GE_CHECK_NOTNULL(merge_output);
  786. merge_output->SetDataType(DT_FLOAT16);
  787. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  788. auto merge_input = src_op_desc->MutableInputDesc(i);
  789. GE_CHECK_NOTNULL(merge_input);
  790. merge_input->SetDataType(DT_FLOAT16);
  791. }
  792. }
  793. std::vector<int64_t> dst_shape_dims;
  794. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  795. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  796. REPORT_CALL_ERROR("E19999", "Transfer output:0 shape of op:%s(%s) to NC1HWC0 format failed, shape:%s, format:%s, "
  797. "when %s", src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str(),
  798. src_shape.ToString().c_str(), TypeUtils::FormatToSerialString(src_format).c_str(), __FUNCTION__);
  799. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  800. return FAILED;
  801. }
  802. ge::GeShape dst_shape(dst_shape_dims);
  803. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  804. net_output_input_desc->SetShape(dst_shape);
  805. if (is_dynamic) {
  806. auto merge_out = src_op_desc->MutableOutputDesc(0);
  807. GE_CHECK_NOTNULL(merge_out);
  808. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  809. REPORT_CALL_ERROR("E19999", "Modify format and shape of output:0 in op:%s(%s) failed when %s",
  810. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str(), __FUNCTION__);
  811. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  812. return FAILED;
  813. }
  814. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  815. auto merge_in = src_op_desc->MutableInputDesc(i);
  816. GE_CHECK_NOTNULL(merge_in);
  817. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  818. REPORT_CALL_ERROR("E19999", "Modify format and shape of input:%u in op:%s(%s) failed when %s", i,
  819. src_op_desc->GetName().c_str(), src_op_desc->GetType().c_str(), __FUNCTION__);
  820. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  821. return FAILED;
  822. }
  823. }
  824. }
  825. return SUCCESS;
  826. }
  827. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  828. GE_CHECK_NOTNULL(netout_desc);
  829. vector<string> output_dt_str;
  830. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  831. for (auto dt_str : output_dt_str) {
  832. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  833. if (dt_str_split.size() == kUserDefinedElementCount) {
  834. if (dt_str_split[0] == to_string(index)) {
  835. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  836. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  837. TypeUtils::DataTypeToSerialString(dt).c_str());
  838. return true;
  839. }
  840. }
  841. }
  842. }
  843. return false;
  844. }
  845. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  846. GE_CHECK_NOTNULL(netout_desc);
  847. vector<string> output_format_str;
  848. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  849. for (auto format_str : output_format_str) {
  850. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  851. if (format_str_split.size() == kUserDefinedElementCount) {
  852. if (format_str_split[0] == to_string(index)) {
  853. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  854. return true;
  855. }
  856. }
  857. }
  858. }
  859. return false;
  860. }
  861. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  862. auto op_desc = node->GetOpDesc();
  863. GE_CHECK_NOTNULL(op_desc);
  864. ge::DataType output_data_type = ge::DT_FLOAT;
  865. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  866. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  867. auto peer_out = in_anchor->GetPeerOutAnchor();
  868. GE_CHECK_NOTNULL(peer_out);
  869. auto src_node = peer_out->GetOwnerNode();
  870. GE_CHECK_NOTNULL(src_node);
  871. bool is_dynamic = CheckOpType(src_node, MERGE);
  872. OpDescPtr src_op_desc = src_node->GetOpDesc();
  873. GE_CHECK_NOTNULL(src_op_desc);
  874. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  875. GE_CHECK_NOTNULL(net_output_input_desc);
  876. ge::GeShape old_shape = net_output_input_desc->GetShape();
  877. ge::Format old_format = net_output_input_desc->GetFormat();
  878. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  879. // Update datatype
  880. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  881. GELOGI("Enter into process output_type schedule");
  882. net_output_input_desc->SetDataType(output_data_type);
  883. if (is_dynamic) {
  884. auto merge_output = src_op_desc->MutableOutputDesc(0);
  885. GE_CHECK_NOTNULL(merge_output);
  886. merge_output->SetDataType(output_data_type);
  887. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  888. auto merge_input = src_op_desc->MutableInputDesc(i);
  889. GE_CHECK_NOTNULL(merge_input);
  890. merge_input->SetDataType(output_data_type);
  891. }
  892. }
  893. }
  894. // check if is_output_adjust_hw_layout is set
  895. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  896. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  897. ErrorManager::GetInstance().ATCReportErrMessage(
  898. "E19014", {"opname", "value", "reason"},
  899. {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]",
  900. "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"});
  901. GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0.");
  902. return FAILED;
  903. }
  904. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  905. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  906. GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0.");
  907. return FAILED;
  908. }
  909. }
  910. }
  911. return SUCCESS;
  912. }
  913. long StringToLongNoThrow(const string &str) {
  914. try {
  915. return std::stol(str);
  916. } catch (const std::invalid_argument) {
  917. GELOGE(PARAM_INVALID,
  918. "Parse shape range of input failed when transfer from string to int64. Given %s, while correct example: "
  919. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  920. str.c_str());
  921. return PARAM_INVALID;
  922. } catch (const std::out_of_range) {
  923. GELOGE(PARAM_INVALID,
  924. "Parse shape range of input failed when transfer from string to int64. Given %s, while correct example: "
  925. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  926. str.c_str());
  927. return PARAM_INVALID;
  928. }
  929. }
  930. /**
  931. * Parser shape_range from string to vector
  932. * shape_range from option normally is "[1~20,3,3~6,-1],[1~20,3,3~6,-1]"
  933. * @param shape_range
  934. */
  935. Status ParseDynamicInputShapeRange(const std::string &shape_range,
  936. std::vector<std::vector<std::pair<int64_t, int64_t>>> &range) {
  937. if (shape_range.size() < 2) {
  938. REPORT_INNER_ERROR("E19999", "shape_range.size:%zu < 2, check invalid when %s", shape_range.size(), __FUNCTION__);
  939. GELOGE(PARAM_INVALID, "Shape range %s is invalid.", shape_range.c_str());
  940. return PARAM_INVALID;
  941. }
  942. // different shape_range of single input are split by ']'
  943. vector<string> shape_range_set = ge::StringUtils::Split(shape_range, ']');
  944. if (shape_range_set.empty()) {
  945. REPORT_INNER_ERROR("E19999", "Shape range %s is not valid. Correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  946. shape_range.c_str());
  947. GELOGE(PARAM_INVALID, "Shape range %s is not valid. Correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  948. shape_range.c_str());
  949. return PARAM_INVALID;
  950. }
  951. for (auto &shape_range_str : shape_range_set) {
  952. if (shape_range_str.size() < 3) {
  953. // shape_range_str should be "[2~3,1"
  954. // or ",[2~3,1". because we should trim '[' or ',['
  955. // so shape_range_str.size() < 3 is invalid
  956. continue;
  957. }
  958. // trim start bytes, after that, single input should be "1~20,3,3~6,-1"
  959. if (ge::StringUtils::StartWith(shape_range_str, "[")) {
  960. shape_range_str = shape_range_str.substr(1, shape_range_str.size());
  961. }
  962. if (ge::StringUtils::StartWith(shape_range_str, ",")) {
  963. shape_range_str = shape_range_str.substr(2, shape_range_str.size());
  964. }
  965. // parse shape_range of single input. eg. "1~20,3,3~6,-1"
  966. std::vector<std::pair<int64_t, int64_t>> range_of_single_input;
  967. vector<string> dim_range_set = ge::StringUtils::Split(shape_range_str, ',');
  968. for (const auto &range_pair_str : dim_range_set) {
  969. vector<string> range_pair_set = ge::StringUtils::Split(range_pair_str, '~');
  970. pair<int64_t, int64_t> range_pair;
  971. if (range_pair_set.size() == 1) {
  972. // fix dim
  973. auto range_value = StringToLongNoThrow(range_pair_set.at(0).c_str());
  974. if (range_value < 0) {
  975. range_pair = std::make_pair(1, range_value);
  976. } else {
  977. range_pair = std::make_pair(range_value, range_value);
  978. }
  979. } else if (range_pair_set.size() == 2) {
  980. // unknown dim, should get range.
  981. auto range_left = StringToLongNoThrow(range_pair_set.at(0).c_str());
  982. auto range_right = StringToLongNoThrow(range_pair_set.at(1).c_str());
  983. if (range_left < 0 || range_right < 0) {
  984. REPORT_INNER_ERROR("E19999", "Shape range of input is invalid. Given range pair [%ld,%ld], "
  985. "while correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"", range_left, range_right);
  986. GELOGE(PARAM_INVALID,
  987. "Shape range of input is invalid. Given range pair [%ld,%ld], while correct example: "
  988. "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  989. range_left, range_right);
  990. return PARAM_INVALID;
  991. }
  992. range_pair = std::make_pair(range_left, range_right);
  993. } else {
  994. REPORT_INNER_ERROR("E19999", "Shape range of input is invalid. Given %s, "
  995. "while correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"", shape_range.c_str());
  996. GELOGE(PARAM_INVALID,
  997. "Shape range of input is invalid. Given %s, while correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"",
  998. shape_range.c_str());
  999. return PARAM_INVALID;
  1000. }
  1001. range_of_single_input.emplace_back(range_pair);
  1002. }
  1003. range.emplace_back(range_of_single_input);
  1004. }
  1005. return SUCCESS;
  1006. }
  1007. Status GetDynamicInputShapeRange(const std::vector<GeTensor> &user_input, const std::map<string, string> &graph_option,
  1008. vector<vector<std::pair<int64_t, int64_t>>> &range_vec) {
  1009. // check both mode and shape_range option are all enabled
  1010. auto mode_iter = graph_option.find(OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  1011. bool enable_dynamic_execute_mode = (mode_iter != graph_option.end()) && (mode_iter->second == "dynamic_execute");
  1012. if (!enable_dynamic_execute_mode) {
  1013. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  1014. OPTION_EXEC_DYNAMIC_EXECUTE_MODE);
  1015. }
  1016. auto iter = graph_option.find(OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1017. bool enable_input_shape_range = (iter != graph_option.end()) && (!iter->second.empty());
  1018. if (!enable_input_shape_range) {
  1019. GELOGD("Graph Option: Can not find %s option in graph options or option value is empty",
  1020. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1021. }
  1022. if (enable_dynamic_execute_mode && enable_input_shape_range) {
  1023. GELOGD("GraphOption: %s value is dynamic_execute, %s value is %s.", OPTION_EXEC_DYNAMIC_EXECUTE_MODE,
  1024. OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE, iter->second.c_str());
  1025. } else if (!enable_dynamic_execute_mode && !enable_input_shape_range) {
  1026. return SUCCESS;
  1027. } else {
  1028. REPORT_INNER_ERROR("E19999", "Graph option: %s and %s should be enabled at the same time, check invalid when %s",
  1029. OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE, __FUNCTION__);
  1030. GELOGE(PARAM_INVALID, "Graph option: %s and %s should be enabled at the same time.",
  1031. OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE);
  1032. return PARAM_INVALID;
  1033. }
  1034. auto ret = ParseDynamicInputShapeRange(iter->second, range_vec);
  1035. GE_CHK_STATUS_RET(ret, "Parse dynamic input shape range failed.");
  1036. if (range_vec.size() != user_input.size()) {
  1037. GELOGE(PARAM_INVALID, "Dynamic input shape range size is %zu, inputs size is %zu. Not match.", range_vec.size(),
  1038. user_input.size());
  1039. return PARAM_INVALID;
  1040. }
  1041. return SUCCESS;
  1042. }
  1043. Status UpdateDynamicInputShapeRange(const ge::GeAttrValue::INT index,
  1044. const vector<vector<std::pair<int64_t, int64_t>>> &range_vec, OpDescPtr &op,
  1045. GeTensorDesc &desc) {
  1046. auto origin_shape = desc.GetShape();
  1047. auto current_shape_range_vec = range_vec.at(index);
  1048. if (current_shape_range_vec.size() != origin_shape.GetDimNum()) {
  1049. REPORT_INNER_ERROR("E19999", "Given shape_range dim num is %zu, current dim:%s num is %zu, not match, "
  1050. "check invalid when %s", current_shape_range_vec.size(), origin_shape.ToString().c_str(),
  1051. origin_shape.GetDimNum(), __FUNCTION__);
  1052. GELOGE(PARAM_INVALID, "Given shape_range dim num is %zu, current dim num is %zu, not match.Pleace Check.",
  1053. current_shape_range_vec.size(), origin_shape.GetDimNum());
  1054. return PARAM_INVALID;
  1055. }
  1056. for (size_t i = 0; i < origin_shape.GetDimNum(); ++i) {
  1057. auto curr_dim = origin_shape.GetDim(i);
  1058. auto left_range = current_shape_range_vec.at(i).first;
  1059. auto right_range = current_shape_range_vec.at(i).second;
  1060. if (left_range == right_range) {
  1061. // given shape_range is known dim, check is same as origin or not
  1062. if (curr_dim != left_range) {
  1063. REPORT_INNER_ERROR("E19999", "Given shape range is %ld, current dim shape is %ld, not match, dim_index:%zu, "
  1064. "check invalid when %s", left_range, curr_dim, i, __FUNCTION__);
  1065. GELOGE(PARAM_INVALID, "Given shape range is %ld, current dim shape is %ld, not match.Pleace Check.",
  1066. left_range, curr_dim);
  1067. return PARAM_INVALID;
  1068. }
  1069. origin_shape.SetDim(i, left_range);
  1070. } else {
  1071. // given shape_range is fix range, check input_shape is in this range or not
  1072. if (right_range != UNKNOWN_DIM) {
  1073. if ((curr_dim < left_range) || (curr_dim > right_range)) {
  1074. REPORT_INNER_ERROR("E19999", "Given shape range is [%ld~%ld], current dim shape is %ld, out of range, "
  1075. "dim_index:%zu, check invalid when %s",
  1076. left_range, right_range, curr_dim, i, __FUNCTION__);
  1077. GELOGE(PARAM_INVALID, "Given shape range is [%ld~%ld], current dim shape is %ld, out of range.Pleace Check.",
  1078. left_range, right_range, curr_dim);
  1079. return PARAM_INVALID;
  1080. }
  1081. }
  1082. origin_shape.SetDim(i, UNKNOWN_DIM);
  1083. }
  1084. }
  1085. desc.SetShape(origin_shape);
  1086. desc.SetShapeRange(current_shape_range_vec);
  1087. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1088. GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  1089. graph_ret = op->UpdateOutputDesc(0, desc);
  1090. GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret);
  1091. return SUCCESS;
  1092. }
  1093. } // namespace
  1094. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  1095. GraphPrepare::~GraphPrepare() {}
  1096. /**
  1097. * @param graph
  1098. * @return
  1099. */
  1100. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  1101. GE_CHECK_NOTNULL(graph);
  1102. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  1103. for (auto &node : graph->GetAllNodes()) {
  1104. if (node == nullptr) {
  1105. continue;
  1106. }
  1107. if (node->GetType() != VARIABLE) {
  1108. continue;
  1109. }
  1110. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  1111. if (trans_road == nullptr) {
  1112. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  1113. continue;
  1114. }
  1115. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  1116. auto ret = RecoverTransRoadForVar(node, *trans_road);
  1117. if (ret != SUCCESS) {
  1118. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str());
  1119. return INTERNAL_ERROR;
  1120. }
  1121. auto iter = var_names_to_refs.find(node->GetName());
  1122. if (iter != var_names_to_refs.end()) {
  1123. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  1124. if (ret != SUCCESS) {
  1125. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str());
  1126. return INTERNAL_ERROR;
  1127. }
  1128. }
  1129. }
  1130. return SUCCESS;
  1131. }
  1132. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  1133. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  1134. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  1135. if (compute_graph_ != nullptr) {
  1136. compute_graph_->SetSessionID(session_id);
  1137. }
  1138. session_id_ = session_id;
  1139. Status ret = CheckGraph();
  1140. if (ret != SUCCESS) {
  1141. GELOGE(ret, "RunGraph graph check fail, ret:%u", ret);
  1142. return ret;
  1143. }
  1144. (void)compute_graph_->TopologicalSorting();
  1145. ret = CheckRefOp();
  1146. if (ret != SUCCESS) {
  1147. GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret);
  1148. return ret;
  1149. }
  1150. return SUCCESS;
  1151. }
  1152. Status GraphPrepare::CheckGraph() {
  1153. if (compute_graph_ == nullptr) {
  1154. REPORT_INNER_ERROR("E19999", "compute_graph_ is nullptr, check invalid when GraphPrepare %s", __FUNCTION__);
  1155. GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR");
  1156. return GE_GRAPH_INIT_FAILED;
  1157. }
  1158. auto nodes = compute_graph_->GetAllNodes();
  1159. if (nodes.empty()) {
  1160. REPORT_INNER_ERROR("E19999", "nodes in graph is empty, check invalid when GraphPrepare %s", __FUNCTION__);
  1161. GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph.");
  1162. return GE_GRAPH_INIT_FAILED;
  1163. }
  1164. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  1165. GE_CHECK_NOTNULL(node);
  1166. if (node->GetOpDesc() == nullptr) {
  1167. REPORT_INNER_ERROR("E19999", "node without opdesc exist in graph, check invalid when GraphPrepare %s",
  1168. __FUNCTION__);
  1169. GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL");
  1170. return GE_GRAPH_INIT_FAILED;
  1171. }
  1172. }
  1173. return SUCCESS;
  1174. }
  1175. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  1176. const std::set<NodePtr> &ref_nodes) {
  1177. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  1178. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  1179. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  1180. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  1181. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  1182. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH};
  1183. GE_CHECK_NOTNULL(node);
  1184. const auto &op_desc = node->GetOpDesc();
  1185. GE_CHECK_NOTNULL(op_desc);
  1186. const auto input_index = op_desc->GetInputIndexByName(input_name);
  1187. const auto &in_anchor = node->GetInDataAnchor(input_index);
  1188. GE_CHECK_NOTNULL(in_anchor);
  1189. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  1190. GE_CHECK_NOTNULL(peer_out_anchor);
  1191. const auto &input_node = peer_out_anchor->GetOwnerNode();
  1192. GE_CHECK_NOTNULL(input_node);
  1193. const auto &input_op_desc = input_node->GetOpDesc();
  1194. GE_CHECK_NOTNULL(input_op_desc);
  1195. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  1196. if (is_ref) {
  1197. return SUCCESS;
  1198. }
  1199. auto input_type = input_op_desc->GetType();
  1200. if (input_type == ge::FRAMEWORKOP) {
  1201. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  1202. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed when GraphPrepare %s",
  1203. ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE.c_str(),
  1204. input_op_desc->GetName().c_str(), input_op_desc->GetType().c_str(), __FUNCTION__);
  1205. GELOGE(PARAM_INVALID, "Get original type failed.");
  1206. return PARAM_INVALID;
  1207. }
  1208. }
  1209. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  1210. if (!is_acceptable) {
  1211. ErrorManager::GetInstance().ATCReportErrMessage(
  1212. "E15005", {"opname", "optype", "opname1", "optype1"},
  1213. {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()});
  1214. GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.",
  1215. node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  1216. input_op_desc->GetType().c_str());
  1217. return PARAM_INVALID;
  1218. }
  1219. return SUCCESS;
  1220. }
  1221. Status GraphPrepare::CheckRefOp() {
  1222. GE_CHECK_NOTNULL(compute_graph_);
  1223. std::set<NodePtr> ref_nodes;
  1224. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  1225. if (node == nullptr) {
  1226. REPORT_INNER_ERROR("E19999", "nullptr node exist in graph, check invalid when GraphPrepare %s",
  1227. __FUNCTION__);
  1228. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  1229. return PARAM_INVALID;
  1230. }
  1231. auto op_desc = node->GetOpDesc();
  1232. if (op_desc == nullptr) {
  1233. REPORT_INNER_ERROR("E19999", "node without opdesc exist in graph, check invalid when GraphPrepare %s",
  1234. __FUNCTION__);
  1235. GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null.");
  1236. return PARAM_INVALID;
  1237. }
  1238. auto input_name_index = op_desc->GetAllInputName();
  1239. auto outputs = op_desc->GetAllOutputName();
  1240. for (const auto &name_index : input_name_index) {
  1241. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  1242. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  1243. GELOGE(PARAM_INVALID, "CheckRefInputNode failed.");
  1244. return PARAM_INVALID;
  1245. }
  1246. (void)ref_nodes.insert(node); // no need to check value
  1247. }
  1248. }
  1249. }
  1250. return SUCCESS;
  1251. };
  1252. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  1253. GE_CHECK_NOTNULL(compute_graph_);
  1254. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  1255. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  1256. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  1257. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  1258. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  1259. return SUCCESS;
  1260. }
  1261. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1262. if (node == nullptr) {
  1263. REPORT_INNER_ERROR("E19999", "Param node is nullptr, check invalid when GraphPrepare %s",
  1264. __FUNCTION__);
  1265. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL");
  1266. return GE_GRAPH_GRAPH_NODE_NULL;
  1267. }
  1268. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1269. if (op_desc_ptr == nullptr) {
  1270. REPORT_INNER_ERROR("E19999", "Param node's op_desc is nullptr, check invalid when GraphPrepare %s",
  1271. __FUNCTION__);
  1272. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL");
  1273. return GE_GRAPH_GRAPH_NODE_NULL;
  1274. }
  1275. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1276. GeShape output_shape = output.GetShape();
  1277. if (output_shape.IsUnknownShape()) {
  1278. GELOGD("[Adjust][DataOpOutput] Shape of op [%s] output is unknown.", node->GetName().c_str());
  1279. return SUCCESS;
  1280. }
  1281. int64_t tensor_size = 0;
  1282. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1283. if (graph_status != GRAPH_SUCCESS) {
  1284. ErrorManager::GetInstance().ATCReportErrMessage(
  1285. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1286. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  1287. return FAILED;
  1288. }
  1289. TensorUtils::SetSize(output, tensor_size);
  1290. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1291. if (graph_ret != GRAPH_SUCCESS) {
  1292. REPORT_CALL_ERROR("E19999", "Update output desc of op:%s(%s) failed, index:0, when GraphPrepare %s",
  1293. op_desc_ptr->GetName().c_str(), op_desc_ptr->GetType().c_str(), __FUNCTION__);
  1294. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1295. return graph_ret;
  1296. }
  1297. return SUCCESS;
  1298. }
  1299. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input,
  1300. const std::map<string, string> &graph_option) {
  1301. // Get shape range of input in dynamic_execute mode
  1302. vector<vector<std::pair<int64_t, int64_t>>> dynamic_shape_range_vec;
  1303. auto ret = GetDynamicInputShapeRange(user_input, graph_option, dynamic_shape_range_vec);
  1304. GE_CHK_STATUS_RET(ret, "Graph option is not right on Dynamic execute mode.");
  1305. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1306. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1307. GE_CHECK_NOTNULL(input_node);
  1308. OpDescPtr op = input_node->GetOpDesc();
  1309. GE_CHECK_NOTNULL(op);
  1310. if (op->GetType() == DATA) {
  1311. GeAttrValue::INT index = 0;
  1312. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1313. GELOGW("Get index from data attr failed");
  1314. continue;
  1315. }
  1316. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1317. std::string situation = "data op index[" + std::to_string(index) + "]";
  1318. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1319. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1320. GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index);
  1321. return FAILED;
  1322. }
  1323. if (IsDynamicDims(input_node)) {
  1324. continue;
  1325. }
  1326. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1327. auto format = desc.GetFormat();
  1328. auto origin_format = desc.GetOriginFormat();
  1329. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1330. auto tune_flag = (options_.build_mode == BUILD_MODE_TUNING) && (options_.build_step == BUILD_STEP_AFTER_BUILDER);
  1331. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op) && (!tune_flag);
  1332. if (need_check_internal_format) {
  1333. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1334. if (is_internal) {
  1335. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1336. {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" +
  1337. TypeUtils::FormatToSerialString(origin_format) + "]", "it is not support"});
  1338. GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.",
  1339. TypeUtils::FormatToSerialString(format).c_str(),
  1340. TypeUtils::FormatToSerialString(origin_format).c_str());
  1341. return FAILED;
  1342. }
  1343. }
  1344. auto data_type = desc.GetDataType();
  1345. uint32_t length = 1;
  1346. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1347. if (!type_ret) {
  1348. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1349. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1350. GELOGE(PARAM_INVALID, "Input datatype %s is not support.",
  1351. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1352. return FAILED;
  1353. }
  1354. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1355. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1356. int64_t shape_size = desc_shape * length;
  1357. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1358. int64_t size = 0;
  1359. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1360. REPORT_CALL_ERROR("E19999", "Get size of user input tensor failed, index:%ld, "
  1361. "when GraphPrepare %s", index, __FUNCTION__);
  1362. GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed");
  1363. return FAILED);
  1364. bool size_check = (size != 0 && shape_size != size);
  1365. if (size_check) {
  1366. std::string situation = "input data size[" + std::to_string(size) +
  1367. "] and shape_size[" + std::to_string(size) + "]";
  1368. std::string reason = "because size != 0 and shape_size != size";
  1369. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1370. GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size);
  1371. return FAILED;
  1372. }
  1373. ge::TensorUtils::SetSize(desc, shape_size);
  1374. if (!tune_flag) {
  1375. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1376. if (graph_ret != GRAPH_SUCCESS) {
  1377. REPORT_CALL_ERROR("E19999", "Update input desc of op:%s(%s) failed, index:0, when GraphPrepare %s",
  1378. op->GetName().c_str(), op->GetType().c_str(), __FUNCTION__);
  1379. GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
  1380. return graph_ret;
  1381. }
  1382. // Size will be recalculated in the build stage
  1383. ge::TensorUtils::SetSize(desc, 0);
  1384. graph_ret = op->UpdateOutputDesc(0, desc);
  1385. if (graph_ret != GRAPH_SUCCESS) {
  1386. REPORT_CALL_ERROR("E19999", "Update output desc of op:%s(%s) failed, index:0, when GraphPrepare %s",
  1387. op->GetName().c_str(), op->GetType().c_str(), __FUNCTION__);
  1388. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1389. return graph_ret;
  1390. }
  1391. } else {
  1392. GELOGI("data %s skip update info in tune mode", op->GetName().c_str());
  1393. }
  1394. if (!dynamic_shape_range_vec.empty()) {
  1395. ret = UpdateDynamicInputShapeRange(index, dynamic_shape_range_vec, op, desc);
  1396. GE_CHK_STATUS_RET(ret, "Fail to update dynamic input shape range on %s.", op->GetName().c_str());
  1397. continue;
  1398. }
  1399. if (!options_.train_graph_flag) {
  1400. Status ret = AdjustDataOpOutput(input_node);
  1401. GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret); return ret);
  1402. }
  1403. }
  1404. }
  1405. return SUCCESS;
  1406. }
  1407. Status GraphPrepare::TryDoAipp() {
  1408. // infer and with aipp configure file, then call aipp insert
  1409. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1410. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1411. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1412. if (ret != SUCCESS) {
  1413. GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed.");
  1414. return INTERNAL_ERROR;
  1415. }
  1416. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1417. if (ret != SUCCESS) {
  1418. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed",
  1419. options_.insert_op_file.c_str());
  1420. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1421. }
  1422. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1423. if (ret != SUCCESS) {
  1424. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret);
  1425. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1426. }
  1427. }
  1428. return SUCCESS;
  1429. }
  1430. Status GraphPrepare::FormatAndShapeProcess() {
  1431. Status ret = ResourcePairProcess("add");
  1432. if (ret != SUCCESS) {
  1433. GELOGE(ret, "ResourcePairProcess failed");
  1434. return ret;
  1435. }
  1436. GE_TIMESTAMP_START(InferOriginFormat1);
  1437. ret = compute_graph_->InferOriginFormat();
  1438. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1439. GE_DUMP(compute_graph_, "after_first_inferformat");
  1440. if (ret != SUCCESS) {
  1441. GELOGE(ret, "Prepare Graph first inferformat failed");
  1442. return ret;
  1443. }
  1444. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1445. ret = InferShapeForPreprocess();
  1446. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1447. GE_DUMP(compute_graph_, "after_infershape");
  1448. if (ret != SUCCESS) {
  1449. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed");
  1450. return GE_GRAPH_INFERSHAPE_FAILED;
  1451. }
  1452. GE_TIMESTAMP_START(InferOriginFormat2);
  1453. ret = compute_graph_->InferOriginFormat();
  1454. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1455. if (ret != SUCCESS) {
  1456. GELOGE(ret, "Prepare Graph inferformat failed");
  1457. return ret;
  1458. }
  1459. ret = ResourcePairProcess("remove");
  1460. if (ret != SUCCESS) {
  1461. return ret;
  1462. }
  1463. return ret;
  1464. }
  1465. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1466. PassManager control_pass;
  1467. // Graph pass tmp logic for resource infershape
  1468. if (options_.train_graph_flag) {
  1469. try {
  1470. if (action == "add") {
  1471. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1472. } else {
  1473. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1474. new ResourcePairRemoveControlPass);
  1475. }
  1476. } catch (std::bad_alloc &e) {
  1477. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add ResourcePair Pass, when GraphPrepare %s",
  1478. __FUNCTION__);
  1479. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str());
  1480. return INTERNAL_ERROR;
  1481. }
  1482. }
  1483. Status ret = control_pass.Run(compute_graph_);
  1484. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1485. GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret);
  1486. return ret;
  1487. }
  1488. return SUCCESS;
  1489. }
  1490. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1491. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1492. GE_CHECK_NOTNULL(node_ptr);
  1493. if (node_ptr->GetType() == DATA) {
  1494. uint32_t index = 0;
  1495. auto op_desc = node_ptr->GetOpDesc();
  1496. GE_CHECK_NOTNULL(op_desc);
  1497. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1498. Format storage_format = FORMAT_RESERVED;
  1499. vector<int64_t> dst_shape_dims;
  1500. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1501. GELOGE(INTERNAL_ERROR, "Get storage format for input failed");
  1502. return FAILED;
  1503. }
  1504. if (storage_format == FORMAT_RESERVED) {
  1505. continue;
  1506. }
  1507. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1508. GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed");
  1509. return FAILED;
  1510. }
  1511. }
  1512. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1513. auto op_desc = node_ptr->GetOpDesc();
  1514. GE_CHECK_NOTNULL(op_desc);
  1515. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1516. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1517. Format storage_format = FORMAT_RESERVED;
  1518. vector<int64_t> dst_shape_dims;
  1519. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1520. GELOGE(INTERNAL_ERROR, "Get storage format from output failed");
  1521. return FAILED;
  1522. }
  1523. if (storage_format == FORMAT_RESERVED) {
  1524. continue;
  1525. }
  1526. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1527. GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed");
  1528. return FAILED;
  1529. }
  1530. }
  1531. }
  1532. }
  1533. return SUCCESS;
  1534. }
  1535. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1536. if (options_.save_original_model == "true") {
  1537. ModelHelper model_helper;
  1538. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1539. options_.original_model_file);
  1540. if (ret != SUCCESS) {
  1541. // If save original model fail, process continue
  1542. GELOGW("SaveOriginalGraphToOmModel fail");
  1543. }
  1544. }
  1545. return SUCCESS;
  1546. }
  1547. #define PP_RUN_AND_DUMP(name, func, ...) \
  1548. do { \
  1549. GE_RUN(Prepare, func, __VA_ARGS__); \
  1550. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1551. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1552. } while (0)
  1553. #define PP_RUN(name, func, ...) \
  1554. do { \
  1555. GE_RUN(Prepare, func, __VA_ARGS__); \
  1556. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1557. } while (0)
  1558. Status GraphPrepare::PrepareDynShape(const GraphNodePtr &graph_node, const std::vector<GeTensor> &user_input,
  1559. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1560. GE_CHECK_NOTNULL(graph_node->GetGraph());
  1561. GE_CHECK_NOTNULL(compute_graph);
  1562. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1563. const Graph &const_graph = *graph_node->GetGraph();
  1564. PP_RUN("Init", Init, const_graph, session_id);
  1565. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1566. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input, graph_node->GetOptions());
  1567. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1568. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1569. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1570. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1571. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1572. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1573. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1574. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1575. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1576. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1577. return SUCCESS;
  1578. }
  1579. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1580. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1581. return SUCCESS;
  1582. }
  1583. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1584. auto compute_graph = compute_graph_;
  1585. PassManager pass_manager;
  1586. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1587. new (std::nothrow) VariablePrepareOpPass))
  1588. GE_TIMESTAMP_START(pass_manager);
  1589. auto ret = pass_manager.Run(compute_graph);
  1590. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1591. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1592. GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret);
  1593. return ret;
  1594. }
  1595. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1596. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1597. return SUCCESS;
  1598. }
  1599. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1600. if (compute_graph == nullptr) {
  1601. REPORT_INNER_ERROR("E19999", "Param compute_graph is nullptr, check invalid when GraphPrepare %s", __FUNCTION__);
  1602. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1603. return GE_GRAPH_NULL_INPUT;
  1604. }
  1605. GEPass ge_passes(compute_graph);
  1606. NamesToPass hccl_group;
  1607. HcclGroupPass hccl_group_pass;
  1608. GELOGD("Add hccl group pass success");
  1609. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1610. auto ret = ge_passes.Run(hccl_group);
  1611. if (ret != SUCCESS) {
  1612. GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret);
  1613. return ret;
  1614. }
  1615. ret = compute_graph->TopologicalSorting();
  1616. if (ret != SUCCESS) {
  1617. REPORT_CALL_ERROR("E19999", "Topological sorting failed when GraphPrepare %s", __FUNCTION__);
  1618. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1619. return ret;
  1620. }
  1621. return SUCCESS;
  1622. }
  1623. #undef PP_RUN_AND_DUMP
  1624. #undef PP_RUN
  1625. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1626. if (graph == nullptr) {
  1627. REPORT_INNER_ERROR("E19999", "Param graph is nullptr, check invalid when GraphPrepare %s", __FUNCTION__);
  1628. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1629. return GE_GRAPH_NULL_INPUT;
  1630. }
  1631. const Graph &const_graph = *graph;
  1632. Status ret = Init(const_graph, 0);
  1633. if (ret != SUCCESS) {
  1634. GELOGE(ret, "Init graph_prepare fail, ret:%u", ret);
  1635. return ret;
  1636. }
  1637. GE_DUMP(compute_graph_, "after_parser");
  1638. GELOGI("Start infershape for dump json process.");
  1639. ret = compute_graph_->InferOriginFormat();
  1640. GE_DUMP(compute_graph_, "after_inferformat");
  1641. if (ret != SUCCESS) {
  1642. REPORT_CALL_ERROR("E19999", "Infer OriginFormat failed when GraphPrepare %s", __FUNCTION__);
  1643. GELOGE(ret, "Prepare Graph inferformat failed");
  1644. return ret;
  1645. }
  1646. InferShapePass infer_shape_pass;
  1647. NamesToPass names_to_passes;
  1648. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1649. GEPass ge_passes(compute_graph_);
  1650. ret = ge_passes.Run(names_to_passes);
  1651. GE_DUMP(compute_graph_, "after_infershape");
  1652. if (ret != SUCCESS) {
  1653. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1654. return ret;
  1655. }
  1656. ShapeRefiner::ClearContextMap();
  1657. return SUCCESS;
  1658. }
  1659. Status GraphPrepare::CheckConstOp() {
  1660. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1661. GE_CHECK_NOTNULL(node_ptr);
  1662. if (node_ptr->GetType() == CONSTANT) {
  1663. Status ret = VerifyConstOp(node_ptr);
  1664. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1665. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1666. auto op_desc = node_ptr->GetOpDesc();
  1667. if (op_desc == nullptr) {
  1668. REPORT_INNER_ERROR("E19999", "op_desc is nullptr, check invalid when GraphPrepare %s", __FUNCTION__);
  1669. GELOGE(PARAM_INVALID, "Get op desc failed");
  1670. return PARAM_INVALID;
  1671. }
  1672. std::string original_type;
  1673. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1674. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1675. GELOGI("original type is %s", original_type.c_str());
  1676. if (original_type == CONSTANT) {
  1677. Status ret = VerifyConstOp(node_ptr);
  1678. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1679. }
  1680. }
  1681. }
  1682. return SUCCESS;
  1683. }
  1684. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1685. GE_CHECK_NOTNULL(node);
  1686. auto op_desc = node->GetOpDesc();
  1687. GE_CHECK_NOTNULL(op_desc);
  1688. ConstGeTensorPtr ge_tensor_ptr;
  1689. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1690. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed when GraphPrepare %s", ATTR_NAME_WEIGHTS.c_str(),
  1691. op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  1692. GELOGE(PARAM_INVALID, "Get value from const attr failed");
  1693. return PARAM_INVALID;
  1694. }
  1695. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1696. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1697. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1698. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1699. auto data_type = ge_tensor_desc.GetDataType();
  1700. if (data_type == DT_STRING) {
  1701. return SUCCESS;
  1702. }
  1703. uint32_t length = 1;
  1704. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1705. if (!type_ret) {
  1706. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"},
  1707. {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"});
  1708. GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str());
  1709. return FAILED;
  1710. }
  1711. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1712. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1713. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1714. if (shape_size == 0) {
  1715. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1716. // shape = [], means it's a sclar tensor.
  1717. GE_CHK_BOOL_EXEC(data_size / length == 1,
  1718. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid scalar tensor."});
  1719. return PARAM_INVALID, "Const is invalid scalar tensor.");
  1720. } else {
  1721. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1722. GE_CHK_BOOL_EXEC(data_size == 0,
  1723. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid vector scalar."});
  1724. return PARAM_INVALID, "Const is invalid vector scalar.");
  1725. }
  1726. } else {
  1727. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0,
  1728. ErrorManager::GetInstance().ATCReportErrMessage(
  1729. "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"});
  1730. return PARAM_INVALID, "Const input data size is not equal with tensor desc shape");
  1731. }
  1732. return SUCCESS;
  1733. }
  1734. bool GraphPrepare::IsDynamicDims(const NodePtr &input_node) {
  1735. auto data_shape = NodeUtils::GetOutputDesc(*input_node, kDataOutIndex).GetShape();
  1736. const auto &dims = data_shape.GetDims();
  1737. bool all_is_positive = false;
  1738. if (std::all_of(dims.begin(), dims.end(), [](int64_t val) { return val >= 0; })) {
  1739. all_is_positive = true;
  1740. }
  1741. if (!all_is_positive && !options_.input_shape.empty() && !options_.dynamic_dims.empty() &&
  1742. options_.dynamic_node_type != kInvalidDynaimcDimsType) {
  1743. GELOGI("No need to check and update desc info, the dims of %s is %s.", input_node->GetName().c_str(),
  1744. formats::JoinToString(dims).c_str());
  1745. return true;
  1746. }
  1747. return false;
  1748. }
  1749. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1750. if (GetLocalOmgContext().is_dynamic_input) {
  1751. return SUCCESS;
  1752. }
  1753. unsigned int node_num = 0;
  1754. unsigned int data_num = 0;
  1755. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1756. GE_CHECK_NOTNULL(input_node);
  1757. OpDescPtr op = input_node->GetOpDesc();
  1758. GE_CHECK_NOTNULL(op);
  1759. node_num++;
  1760. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1761. data_num++;
  1762. GeAttrValue::INT index = 0;
  1763. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1764. REPORT_INNER_ERROR("E19999", "Get Attr:%s of op:%s(%s) failed when GraphPrepare %s", ATTR_NAME_WEIGHTS.c_str(),
  1765. op->GetName().c_str(), op->GetType().c_str(), __FUNCTION__);
  1766. GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed");
  1767. return GE_GRAPH_INIT_FAILED;
  1768. }
  1769. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1770. std::string situation = "data op index[" + std::to_string(index) + "]";
  1771. std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]";
  1772. ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
  1773. GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index);
  1774. return GE_GRAPH_INIT_FAILED;
  1775. }
  1776. if (IsDynamicDims(input_node)) {
  1777. continue;
  1778. }
  1779. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1780. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1781. int64_t dim = desc.GetShape().GetDim(i);
  1782. if (dim < UNKNOWN_DIM_NUM) {
  1783. std::string situation = "data dim[" + std::to_string(i) + "][" + std::to_string(dim) + "]" ;
  1784. std::string reason = "it need >= -2";
  1785. REPORT_INPUT_ERROR(
  1786. "E19025", std::vector<std::string>({"situation", "reason"}), std::vector<std::string>({situation, reason}));
  1787. GELOGE(GE_GRAPH_INIT_FAILED, "[Check][InputDim]data dim %zu is not supported, need >= -2, real:%ld.", i, dim);
  1788. return GE_GRAPH_INIT_FAILED;
  1789. }
  1790. }
  1791. }
  1792. }
  1793. if (node_num <= data_num) {
  1794. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1795. }
  1796. return SUCCESS;
  1797. }
  1798. Status GraphPrepare::InferShapeForPreprocess() {
  1799. GELOGI("Start infershape for preprocess.");
  1800. GEPass ge_passes(compute_graph_);
  1801. NamesToPass names_to_passes;
  1802. AssertPass assert_pass;
  1803. if (!options_.train_graph_flag) {
  1804. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1805. }
  1806. SwitchDeadBranchElimination switch_dead_branch_elimination;
  1807. names_to_passes.emplace_back("SwitchDeadBranchElimination", &switch_dead_branch_elimination);
  1808. MergePass merge_pass;
  1809. names_to_passes.emplace_back("MergePass", &merge_pass);
  1810. InferShapePass infer_shape_pass;
  1811. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1812. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1813. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1814. DimensionComputePass dimension_compute_pass;
  1815. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1816. ConstantFoldingPass constant_folding_pass;
  1817. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1818. int32_t dev_count = 0;
  1819. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1820. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1821. rtError_t rt_err = RT_ERROR_NONE;
  1822. if (aicpu_constant_folding_on != nullptr) {
  1823. rt_err = rtGetDeviceCount(&dev_count);
  1824. if (rt_err == RT_ERROR_NONE) {
  1825. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1826. if (result != SUCCESS) {
  1827. GELOGE(result, "Set rt context failed.");
  1828. return result;
  1829. }
  1830. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1831. }
  1832. }
  1833. Status ret = ge_passes.Run(names_to_passes);
  1834. if (aicpu_constant_folding_on != nullptr) {
  1835. if (rt_err == RT_ERROR_NONE) {
  1836. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1837. if (result != SUCCESS) {
  1838. GELOGE(result, "Set rt context failed.");
  1839. return result;
  1840. }
  1841. }
  1842. }
  1843. ShapeRefiner::ClearContextMap();
  1844. if (ret != SUCCESS) {
  1845. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1846. return ret;
  1847. }
  1848. return SUCCESS;
  1849. }
  1850. Status GraphPrepare::PrepareOptimize() {
  1851. GELOGI("Start optimize for preprocess.");
  1852. // check rw type
  1853. GraphOptimize graph_optimize;
  1854. bool has_conflict = false;
  1855. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1856. if (has_conflict) {
  1857. GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize.");
  1858. return FAILED;
  1859. }
  1860. PassManager original_graph_passes;
  1861. // Graph pass
  1862. try {
  1863. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1864. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1865. (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass);
  1866. } catch (std::bad_alloc &e) {
  1867. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass, when GraphPrepare %s", __FUNCTION__);
  1868. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1869. return INTERNAL_ERROR;
  1870. }
  1871. GE_TIMESTAMP_START(original_graph_passes);
  1872. Status ret = original_graph_passes.Run(compute_graph_);
  1873. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1874. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1875. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1876. return ret;
  1877. }
  1878. // New pass
  1879. GEPass ge_passes(compute_graph_);
  1880. NamesToPass names_to_passes;
  1881. EnterPass enter_pass;
  1882. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1883. CondPass cond_pass;
  1884. names_to_passes.emplace_back("CondPass", &cond_pass);
  1885. PrintOpPass print_pass;
  1886. if (options_.enable_print_op_pass) {
  1887. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1888. }
  1889. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1890. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1891. DropOutPass dropout_pass;
  1892. AssertPass assert_pass;
  1893. UnusedConstPass unused_const_pass;
  1894. StopGradientPass stop_gradient_pass;
  1895. PreventGradientPass prevent_gradient_pass;
  1896. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1897. GuaranteeConstPass guarantee_const_pass;
  1898. VarIsInitializedOpPass var_is_initialized_pass;
  1899. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1900. IdentityPass identity_pass(false);
  1901. SnapshotPass snapshot_pass;
  1902. if (!options_.train_graph_flag) {
  1903. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1904. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1905. }
  1906. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1907. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1908. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1909. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1910. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1911. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1912. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1913. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1914. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1915. GE_TIMESTAMP_START(names_to_passes);
  1916. ret = ge_passes.Run(names_to_passes);
  1917. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1918. if (ret != SUCCESS) {
  1919. GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret);
  1920. return ret;
  1921. }
  1922. PassManager graph_pass;
  1923. try {
  1924. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1925. // can't move to optimize1/2 directly, may cause more identity insert, cause CI fail
  1926. (void)graph_pass.AddPass("PrepareOptimize::HcclMemcpyPass", new HcclMemcpyPass);
  1927. } catch (std::bad_alloc &e) {
  1928. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass, when GraphPrepare %s", __FUNCTION__);
  1929. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1930. return INTERNAL_ERROR;
  1931. }
  1932. GE_TIMESTAMP_START(graph_passes);
  1933. ret = graph_pass.Run(compute_graph_);
  1934. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1935. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1936. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1937. return ret;
  1938. }
  1939. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1940. TypeConversionOfConstant();
  1941. ret = compute_graph_->TopologicalSorting();
  1942. if (ret != SUCCESS) {
  1943. REPORT_CALL_ERROR("E19999", "Topological sorting failed when GraphPrepare %s", __FUNCTION__);
  1944. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1945. return ret;
  1946. }
  1947. GELOGI("End optimize for preprocess.");
  1948. return SUCCESS;
  1949. }
  1950. void GraphPrepare::TypeConversionOfConstant() {
  1951. bool is_acl_compile = false;
  1952. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1953. // This can ensure that n is not a null pointer
  1954. // No Conversion when called by aclOpCompile
  1955. (void)AttrUtils::GetBool(n->GetOpDesc(), ATTR_SINGLE_OP_SCENE, is_acl_compile);
  1956. if (is_acl_compile) {
  1957. return;
  1958. }
  1959. }
  1960. if (options_.train_graph_flag) {
  1961. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1962. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1963. // This can ensure that n is not a null pointer
  1964. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1965. n->GetOpDesc()->SetType(CONSTANTOP);
  1966. }
  1967. }
  1968. } else {
  1969. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  1970. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1971. // This can ensure that n is not a null pointer
  1972. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  1973. n->GetOpDesc()->SetType(CONSTANT);
  1974. }
  1975. }
  1976. }
  1977. }
  1978. Status GraphPrepare::GraphEquivalentTransformation() {
  1979. NamesToPass names_to_pass;
  1980. ForPass for_pass;
  1981. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  1982. return GEPass(compute_graph_).Run(names_to_pass);
  1983. }
  1984. Status GraphPrepare::ProcessBeforeInfershape() {
  1985. NamesToPass names_to_passes;
  1986. CondRemovePass condition_remove_pass;
  1987. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  1988. GE_TIMESTAMP_START(ProcessCondRemove);
  1989. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  1990. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  1991. if (ret != SUCCESS) {
  1992. GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  1993. return ret;
  1994. }
  1995. return SUCCESS;
  1996. }
  1997. Status GraphPrepare::ProcessNetOutput() {
  1998. PassManager graph_passes_before_infershape;
  1999. try {
  2000. if (options_.train_graph_flag) {
  2001. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  2002. }
  2003. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  2004. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  2005. new (std::nothrow) DataPass); // Add NetOutput first.
  2006. } catch (std::bad_alloc) {
  2007. REPORT_INNER_ERROR("E19999", "bad memory allocation occur when add Pass, when GraphPrepare %s", __FUNCTION__);
  2008. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  2009. return INTERNAL_ERROR;
  2010. }
  2011. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  2012. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  2013. GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret);
  2014. return ret;
  2015. }
  2016. return SUCCESS;
  2017. }
  2018. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input,
  2019. const std::map<string, string> &graph_option) {
  2020. compute_graph_->SetInputSize(user_input.size());
  2021. if (user_input.empty()) {
  2022. return SUCCESS;
  2023. }
  2024. auto ret = CheckUserInput(user_input);
  2025. if (ret != SUCCESS) {
  2026. GELOGE(ret, "Check user input failed.");
  2027. return ret;
  2028. }
  2029. ret = UpdateInput(user_input, graph_option);
  2030. if (ret != SUCCESS) {
  2031. GELOGE(ret, "UpdateInput fail, ret:%u", ret);
  2032. return ret;
  2033. }
  2034. if (user_input.size() != 0) {
  2035. ret = CheckConstOp();
  2036. if (ret != SUCCESS) {
  2037. GELOGE(ret, "CheckConstOp fail, ret:%u", ret);
  2038. return ret;
  2039. }
  2040. } else {
  2041. ret = compute_graph_->TopologicalSorting();
  2042. if (ret != SUCCESS) {
  2043. REPORT_CALL_ERROR("E19999", "Topological sorting failed when GraphPrepare %s", __FUNCTION__);
  2044. GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting");
  2045. return FAILED;
  2046. }
  2047. }
  2048. return SUCCESS;
  2049. }
  2050. Status GraphPrepare::UpdateInputOutputByOptions() {
  2051. auto ret = UpdateDataNetOutputByStorageFormat();
  2052. if (ret != SUCCESS) {
  2053. GELOGE(ret, "Update format acoording to storage format failed.");
  2054. return ret;
  2055. }
  2056. if (options_.train_graph_flag) {
  2057. GELOGI("This is train mode, no need to do this schedule.");
  2058. return SUCCESS;
  2059. }
  2060. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  2061. GE_CHECK_NOTNULL(node_ptr);
  2062. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  2063. GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str());
  2064. return FAILED;
  2065. }
  2066. if (node_ptr->GetType() == DATA) {
  2067. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  2068. GELOGE(INTERNAL_ERROR, "Process data node failed");
  2069. return FAILED;
  2070. }
  2071. }
  2072. if (node_ptr->GetType() == ge::NETOUTPUT) {
  2073. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  2074. GELOGE(INTERNAL_ERROR, "Process netoutput node failed");
  2075. return FAILED;
  2076. }
  2077. }
  2078. }
  2079. return SUCCESS;
  2080. }
  2081. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  2082. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  2083. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  2084. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  2085. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  2086. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  2087. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  2088. if (dst_node->GetType() == TRANSDATA) {
  2089. return true;
  2090. }
  2091. }
  2092. }
  2093. return false;
  2094. }
  2095. } // namespace ge

图引擎模块(GE)是MindSpore的一个子模块,其代码由C++实现,位于前端模块ME和底层硬件之间,起到承接作用。图引擎模块以ME下发的图作为输入,然后进行一系列的深度图优化操作,最后输出一张可以在底层硬件上高效运行的图。GE针对昇腾AI处理器的硬件结构特点,做了特定的优化工作,以此来充分发挥出昇腾AI处理器的强大算力。在进行模型训练/推理时,GE会被自动调用而用户并不感知。GE主要由GE API和GE Core两部分组成,详细的架构图如下所示