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 75 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
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
4 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
5 years ago
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
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
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868
  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 <utility>
  21. #include "common/formats/format_transfers/format_transfer_fractal_nz.h"
  22. #include "common/formats/format_transfers/format_transfer_fractal_z.h"
  23. #include "common/formats/format_transfers/format_transfer_nchw_nc1hwc0.h"
  24. #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h"
  25. #include "common/formats/format_transfers/format_transfer_transpose.h"
  26. #include "common/formats/utils/formats_trans_utils.h"
  27. #include "common/helper/model_helper.h"
  28. #include "common/math/math_util.h"
  29. #include "common/op/ge_op_utils.h"
  30. #include "common/util/error_manager/error_manager.h"
  31. #include "common/formats/utils/formats_trans_utils.h"
  32. #include "framework/common/debug/ge_log.h"
  33. #include "graph/common/ge_call_wrapper.h"
  34. #include "graph/common/local_context.h"
  35. #include "graph/common/transop_util.h"
  36. #include "graph/debug/ge_attr_define.h"
  37. #include "graph/ge_context.h"
  38. #include "graph/shape_refiner.h"
  39. #include "graph/manager/graph_var_manager.h"
  40. #include "graph/manager/util/rt_context_util.h"
  41. #include "graph/optimize/graph_optimize.h"
  42. #include "graph/passes/addn_pass.h"
  43. #include "graph/passes/aicpu_constant_folding_pass.h"
  44. #include "graph/passes/assert_pass.h"
  45. #include "graph/passes/assign_pass.h"
  46. #include "graph/passes/base_pass.h"
  47. #include "graph/passes/common_subexpression_elimination_pass.h"
  48. #include "graph/passes/cond_pass.h"
  49. #include "graph/passes/cond_remove_pass.h"
  50. #include "graph/passes/constant_folding_pass.h"
  51. #include "graph/passes/constant_fuse_same_pass.h"
  52. #include "graph/passes/control_trigger_pass.h"
  53. #include "graph/passes/dimension_adjust_pass.h"
  54. #include "graph/passes/dimension_compute_pass.h"
  55. #include "graph/passes/dropout_pass.h"
  56. #include "graph/passes/enter_pass.h"
  57. #include "graph/passes/flow_ctrl_pass.h"
  58. #include "graph/passes/for_pass.h"
  59. #include "graph/passes/get_original_format_pass.h"
  60. #include "graph/passes/guarantee_const_pass.h"
  61. #include "graph/passes/hccl_group_pass.h"
  62. #include "graph/passes/hccl_memcpy_pass.h"
  63. #include "graph/passes/identity_pass.h"
  64. #include "graph/passes/infershape_pass.h"
  65. #include "graph/passes/iterator_op_pass.h"
  66. #include "graph/passes/merge_pass.h"
  67. #include "graph/passes/net_output_pass.h"
  68. #include "graph/passes/next_iteration_pass.h"
  69. #include "graph/passes/no_use_reshape_remove_pass.h"
  70. #include "graph/passes/parallel_concat_start_op_pass.h"
  71. #include "graph/passes/placeholder_with_default_pass.h"
  72. #include "graph/passes/prevent_gradient_pass.h"
  73. #include "graph/passes/print_op_pass.h"
  74. #include "graph/passes/prune_pass.h"
  75. #include "graph/passes/replace_transshape_pass.h"
  76. #include "graph/passes/replace_with_empty_const_pass.h"
  77. #include "graph/passes/resource_pair_add_control_pass.h"
  78. #include "graph/passes/resource_pair_remove_control_pass.h"
  79. #include "graph/passes/save_pass.h"
  80. #include "graph/passes/shape_operate_op_remove_pass.h"
  81. #include "graph/passes/snapshot_pass.h"
  82. #include "graph/passes/stop_gradient_pass.h"
  83. #include "graph/passes/subgraph_pass.h"
  84. #include "graph/passes/switch_data_edges_bypass.h"
  85. #include "graph/passes/switch_dead_branch_elimination.h"
  86. #include "graph/passes/switch_logic_remove_pass.h"
  87. #include "graph/passes/merge_to_stream_merge_pass.h"
  88. #include "graph/passes/switch_to_stream_switch_pass.h"
  89. #include "graph/passes/attach_stream_label_pass.h"
  90. #include "graph/passes/unused_const_pass.h"
  91. #include "graph/passes/unused_op_remove_pass.h"
  92. #include "graph/passes/var_is_initialized_op_pass.h"
  93. #include "graph/passes/variable_prepare_op_pass.h"
  94. #include "graph/preprocess/insert_op/util_insert_aipp_op.h"
  95. #include "graph/types.h"
  96. #include "graph/utils/tensor_utils.h"
  97. #include "graph/utils/type_utils.h"
  98. #include "inc/pass_manager.h"
  99. #include "init/gelib.h"
  100. #include "multi_batch_copy_graph.h"
  101. #include "runtime/dev.h"
  102. #include "graph/passes/dimension_adjust_pass.h"
  103. #include "graph/passes/link_gen_mask_nodes_pass.h"
  104. #include "graph/passes/permute_pass.h"
  105. #include "graph/passes/reshape_remove_pass.h"
  106. #include "graph/passes/same_transdata_breadth_fusion_pass.h"
  107. #include "graph/passes/transop_breadth_fusion_pass.h"
  108. #include "graph/passes/transop_depth_fusion_pass.h"
  109. #include "graph/passes/transop_nearby_allreduce_fusion_pass.h"
  110. #include "graph/passes/cast_remove_pass.h"
  111. #include "graph/passes/data_pass.h"
  112. #include "graph/passes/transop_without_reshape_fusion_pass.h"
  113. #include "graph/passes/transpose_transdata_pass.h"
  114. #include "graph/passes/variable_op_pass.h"
  115. #include "graph/passes/variable_prepare_op_pass.h"
  116. #include "graph/passes/variable_ref_delete_op_pass.h"
  117. #include "graph/passes/mark_agnostic_pass.h"
  118. namespace ge {
  119. namespace {
  120. static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
  121. {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16},
  122. {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64},
  123. {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}};
  124. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  125. // the size of user defined output datatype or format string after split by ":".
  126. const size_t kUserDefinedElementCount = 2;
  127. OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) {
  128. GeTensorPtr tensor = MakeShared<GeTensor>();
  129. if (tensor == nullptr) {
  130. GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed");
  131. return nullptr;
  132. }
  133. tensor->MutableTensorDesc().SetDataType(DT_INT32);
  134. tensor->MutableTensorDesc().SetFormat(FORMAT_ND);
  135. auto dst_ge_shape = data_tensor.GetShape();
  136. auto dim_cnt = static_cast<int64_t>(dst_ge_shape.GetDimNum());
  137. if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar
  138. tensor->MutableTensorDesc().SetShape(GeShape());
  139. int32_t dst_shape = 1;
  140. if (tensor->SetData(reinterpret_cast<const uint8_t *>(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) {
  141. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  142. return nullptr;
  143. }
  144. } else {
  145. tensor->MutableTensorDesc().SetShape(GeShape(std::vector<int64_t>({dim_cnt})));
  146. unique_ptr<int32_t[]> dst_shape(new (std::nothrow) int32_t[dim_cnt]());
  147. if (dst_shape == nullptr) {
  148. GELOGE(INTERNAL_ERROR, "Create unique ptr failed");
  149. return nullptr;
  150. }
  151. for (int64_t i = 0; i < dim_cnt; ++i) {
  152. dst_shape[i] = dst_ge_shape.GetDim(static_cast<size_t>(i));
  153. }
  154. GE_IF_BOOL_EXEC(
  155. tensor->SetData(reinterpret_cast<const uint8_t *>(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS,
  156. GELOGE(INTERNAL_ERROR, "tensor set data failed");
  157. return nullptr;)
  158. }
  159. GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str());
  160. return OpDescUtils::CreateConstOp(tensor);
  161. }
  162. void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output,
  163. OpDescPtr &op_desc) {
  164. // For format transfer node, the IR definition has src/dst format attrs
  165. if (node_type == TRANSDATA) {
  166. GE_IF_BOOL_EXEC(
  167. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())),
  168. GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");)
  169. GE_IF_BOOL_EXEC(
  170. !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())),
  171. GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");)
  172. }
  173. // For TransposeD node, the IR definition has perm attrs
  174. if (node_type == TRANSPOSED) {
  175. Format src_format = input.GetFormat();
  176. Format dst_format = output.GetFormat();
  177. std::vector<int64_t> perm_arg;
  178. GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return,
  179. "Get perm by foramt failed.");
  180. GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return,
  181. "SetStr PERMUTE_ATTR_PERM failed")
  182. }
  183. // For cast node, the IR definition has src/dst attrs
  184. if (node_type == CAST) {
  185. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast<int64_t>(input.GetDataType())),
  186. GELOGW("SetInt CAST_ATTR_SRCT failed");)
  187. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast<int64_t>(output.GetDataType())),
  188. GELOGW("SetInt CAST_ATTR_DSTT failed");)
  189. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast<int64_t>(output.GetDataType())),
  190. GELOGW("SetInt CAST_ATTR_DST_TYPE failed");)
  191. GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false),
  192. GELOGW("SetBool CAST_ATTR_TRUNCATE failed");)
  193. }
  194. }
  195. NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input,
  196. const GeTensorDesc &output, NodePtr &node) {
  197. if (node == nullptr) {
  198. GELOGE(PARAM_INVALID, "node is null.");
  199. return nullptr;
  200. }
  201. auto graph = node->GetOwnerComputeGraph();
  202. if (graph == nullptr) {
  203. GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str());
  204. return nullptr;
  205. }
  206. auto index = TransOpUtil::GetTransOpDataIndex(node_type);
  207. if (index < 0) {
  208. GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str());
  209. return nullptr;
  210. }
  211. OpDescPtr op_desc = MakeShared<OpDesc>(name, node_type);
  212. if (op_desc == nullptr) {
  213. GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed");
  214. return nullptr;
  215. }
  216. // for data dump
  217. GE_IF_BOOL_EXEC(
  218. !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector<std::string>())),
  219. GELOGW("CreateTransNode: SetListStr failed");)
  220. // Default single input and single output
  221. auto ret = op_desc->AddInputDesc(input);
  222. if (ret != GRAPH_SUCCESS) {
  223. GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str());
  224. return nullptr;
  225. }
  226. ret = op_desc->AddOutputDesc(output);
  227. if (ret != GRAPH_SUCCESS) {
  228. GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str());
  229. return nullptr;
  230. }
  231. AddTransNodeAttr(node_type, input, output, op_desc);
  232. NodePtr shape_node = nullptr;
  233. if (node_type == RESHAPE) {
  234. auto shape_desc = CreateTensorShape(output);
  235. if (shape_desc == nullptr) {
  236. GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input",
  237. node->GetName().c_str());
  238. return nullptr;
  239. }
  240. ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0));
  241. if (ret != GRAPH_SUCCESS) {
  242. GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str());
  243. return nullptr;
  244. }
  245. shape_node = graph->AddNode(shape_desc);
  246. if (shape_node == nullptr) {
  247. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str());
  248. return nullptr;
  249. }
  250. }
  251. auto trans_node = graph->AddNode(op_desc);
  252. if (trans_node == nullptr) {
  253. GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str());
  254. return nullptr;
  255. }
  256. if (node_type == RESHAPE) {
  257. if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) {
  258. GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str());
  259. return nullptr;
  260. }
  261. }
  262. return trans_node;
  263. }
  264. Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  265. NodePtr &trans_node) {
  266. GE_CHECK_NOTNULL(node);
  267. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node);
  268. if (trans_node == nullptr) {
  269. return INTERNAL_ERROR;
  270. }
  271. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0});
  272. if (ret != GRAPH_SUCCESS) {
  273. GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s",
  274. node->GetName().c_str(), node->GetType().c_str());
  275. return INTERNAL_ERROR;
  276. }
  277. ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0));
  278. if (ret != GRAPH_SUCCESS) {
  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. GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.",
  286. node->GetName().c_str(), trans_node->GetName().c_str());
  287. return INTERNAL_ERROR;
  288. }
  289. return SUCCESS;
  290. }
  291. Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node,
  292. NodePtr &trans_node) {
  293. GE_CHECK_NOTNULL(node);
  294. trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node);
  295. if (trans_node == nullptr) {
  296. return INTERNAL_ERROR;
  297. }
  298. auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {});
  299. if (ret != GRAPH_SUCCESS) {
  300. GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s",
  301. node->GetName().c_str(), node->GetType().c_str());
  302. return INTERNAL_ERROR;
  303. }
  304. ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0));
  305. if (ret != GRAPH_SUCCESS) {
  306. GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(),
  307. node->GetName().c_str());
  308. return INTERNAL_ERROR;
  309. }
  310. ret = GraphUtils::MoveInCtrlEdges(node, trans_node);
  311. if (ret != GRAPH_SUCCESS) {
  312. GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.",
  313. node->GetName().c_str(), trans_node->GetName().c_str());
  314. return INTERNAL_ERROR;
  315. }
  316. return SUCCESS;
  317. }
  318. Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) {
  319. GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR);
  320. GE_CHECK_NOTNULL(var->GetOpDesc());
  321. if (var->GetOpDesc()->GetOutputsSize() > 0) {
  322. auto output_desc = var->GetOpDesc()->GetOutputDesc(0);
  323. output_desc.SetFormat(tensor_desc.GetFormat());
  324. output_desc.SetDataType(tensor_desc.GetDataType());
  325. output_desc.SetShape(tensor_desc.GetShape());
  326. output_desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  327. output_desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  328. output_desc.SetOriginShape(tensor_desc.GetOriginShape());
  329. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS,
  330. GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed");
  331. return INTERNAL_ERROR;);
  332. }
  333. if (var->GetOpDesc()->GetInputsSize() > 0) {
  334. auto desc = var->GetOpDesc()->GetInputDesc(0);
  335. desc.SetFormat(tensor_desc.GetFormat());
  336. desc.SetDataType(tensor_desc.GetDataType());
  337. desc.SetShape(tensor_desc.GetShape());
  338. desc.SetOriginFormat(tensor_desc.GetOriginFormat());
  339. desc.SetOriginDataType(tensor_desc.GetOriginDataType());
  340. desc.SetOriginShape(tensor_desc.GetOriginShape());
  341. GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS,
  342. GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed");
  343. return INTERNAL_ERROR;)
  344. }
  345. return SUCCESS;
  346. }
  347. Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) {
  348. GE_CHECK_NOTNULL(var);
  349. int index = 0;
  350. NodePtr last_node = var;
  351. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  352. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  353. auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node);
  354. if (ret != SUCCESS) {
  355. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(),
  356. index, iter->node_type.c_str());
  357. return INTERNAL_ERROR;
  358. }
  359. // set stream_label
  360. OpDescPtr var_desc = var->GetOpDesc();
  361. GE_CHECK_NOTNULL(var_desc);
  362. std::string stream_label;
  363. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  364. if (!stream_label.empty()) {
  365. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  366. }
  367. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  368. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  369. GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str());
  370. }
  371. if (road.empty()) {
  372. return SUCCESS;
  373. }
  374. return UpdateVarFormats(var, road.rbegin()->output);
  375. }
  376. Status RecoverTransRoadForVarRef(const std::set<NodePtr> &nodes, const VarTransRoad &road) {
  377. for (auto &var : nodes) {
  378. GE_CHECK_NOTNULL(var);
  379. int index = 0;
  380. NodePtr last_node = var;
  381. GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str());
  382. for (auto iter = road.rbegin(); iter != road.rend(); ++iter) {
  383. auto trans_name = var->GetName() + "_trans_" + std::to_string(index++);
  384. auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node);
  385. if (ret != SUCCESS) {
  386. GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s",
  387. var->GetName().c_str(), index, iter->node_type.c_str());
  388. return INTERNAL_ERROR;
  389. }
  390. // set stream_label
  391. OpDescPtr var_desc = var->GetOpDesc();
  392. GE_CHECK_NOTNULL(var_desc);
  393. std::string stream_label;
  394. (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label);
  395. if (!stream_label.empty()) {
  396. GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed");
  397. }
  398. GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)),
  399. return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed.");
  400. }
  401. if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) {
  402. return INTERNAL_ERROR;
  403. }
  404. }
  405. return SUCCESS;
  406. }
  407. using VarNamesToRefs = std::map<std::string, std::set<NodePtr>>;
  408. VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) {
  409. VarNamesToRefs names_to_refs;
  410. std::string var_name;
  411. if (graph == nullptr) {
  412. GELOGE(PARAM_INVALID, "graph is null.");
  413. return names_to_refs;
  414. }
  415. for (auto &node : graph->GetAllNodes()) {
  416. if (node->GetType() != VARIABLE) {
  417. continue;
  418. }
  419. if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) {
  420. (void)names_to_refs[var_name].insert(node);
  421. }
  422. }
  423. return names_to_refs;
  424. }
  425. Status TransferShape2NC1HWC0(Format src_format, const std::vector<int64_t> &src_shape, DataType dt, Format dst_format,
  426. std::vector<int64_t> &dst_shape) {
  427. if (src_format == FORMAT_NCHW) {
  428. formats::FormatTransferNchwNc1hwc0 transfer;
  429. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  430. GELOGE(INTERNAL_ERROR, "TransShape failed");
  431. return FAILED;
  432. }
  433. } else if (src_format == FORMAT_NHWC) {
  434. formats::FormatTransferNhwcNc1hwc0 transfer;
  435. if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) {
  436. GELOGE(INTERNAL_ERROR, "TransShape failed");
  437. return FAILED;
  438. }
  439. }
  440. return SUCCESS;
  441. }
  442. Status ModifyInputFormatAndShape(NodePtr &node_ptr) {
  443. GE_CHECK_NOTNULL(node_ptr);
  444. auto op_desc = node_ptr->GetOpDesc();
  445. GE_CHECK_NOTNULL(op_desc);
  446. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  447. GE_CHECK_NOTNULL(input);
  448. ge::Format old_format = input->GetFormat();
  449. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  450. ge::DataType dt = input->GetDataType();
  451. std::vector<int64_t> dst_shape_dims;
  452. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  453. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  454. return FAILED;
  455. }
  456. input->SetFormat(FORMAT_NC1HWC0);
  457. input->SetShape(ge::GeShape(dst_shape_dims));
  458. auto output = op_desc->MutableOutputDesc(0);
  459. GE_CHECK_NOTNULL(output);
  460. output->SetFormat(FORMAT_NC1HWC0);
  461. output->SetShape(ge::GeShape(dst_shape_dims));
  462. int64_t size = 0;
  463. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  464. if (graph_status != ge::GRAPH_SUCCESS) {
  465. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  466. return FAILED;
  467. }
  468. ge::TensorUtils::SetSize(*output, size);
  469. ge::TensorUtils::SetSize(*input, size);
  470. return SUCCESS;
  471. }
  472. Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) {
  473. GE_CHECK_NOTNULL(input_output);
  474. ge::Format old_format = input_output->GetFormat();
  475. std::vector<int64_t> old_shape = input_output->GetShape().GetDims();
  476. ge::DataType dt = input_output->GetDataType();
  477. std::vector<int64_t> dst_shape_dims;
  478. if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  479. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  480. return FAILED;
  481. }
  482. input_output->SetFormat(FORMAT_NC1HWC0);
  483. input_output->SetShape(ge::GeShape(dst_shape_dims));
  484. return SUCCESS;
  485. }
  486. Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format,
  487. vector<int64_t> &dst_shape_dims) {
  488. GE_CHECK_NOTNULL(op_desc);
  489. const GeTensorDescPtr &input = op_desc->MutableInputDesc(index);
  490. GE_CHECK_NOTNULL(input);
  491. ge::Format old_format = input->GetFormat();
  492. std::vector<int64_t> old_shape = input->GetShape().GetDims();
  493. input->SetShape(ge::GeShape(dst_shape_dims));
  494. input->SetFormat(storage_format);
  495. auto output = op_desc->MutableOutputDesc(index);
  496. GE_CHECK_NOTNULL(output);
  497. output->SetShape(ge::GeShape(dst_shape_dims));
  498. output->SetFormat(storage_format);
  499. if (!output->MutableShape().IsUnknownShape()) {
  500. int64_t size = 0;
  501. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size);
  502. if (graph_status != ge::GRAPH_SUCCESS) {
  503. GELOGE(graph_status, "GetTensorSizeInBytes failed!");
  504. return FAILED;
  505. }
  506. ge::TensorUtils::SetSize(*input, size);
  507. ge::TensorUtils::SetSize(*output, size);
  508. GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, "
  509. "new_shape:%s, new_format:%s, new_size:%lu",
  510. op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(),
  511. ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(),
  512. ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size);
  513. }
  514. return SUCCESS;
  515. }
  516. Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) {
  517. is_dynamic_batch = false;
  518. std::string related_node_name;
  519. if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) {
  520. if (related_node_name.empty()) {
  521. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty",
  522. data_node->GetName().c_str());
  523. return INTERNAL_ERROR;
  524. }
  525. for (const NodePtr &next_node : data_node->GetOutNodes()) {
  526. if (next_node->GetName() == related_node_name) {
  527. switchn_node = next_node;
  528. break;
  529. }
  530. }
  531. if (switchn_node == nullptr) {
  532. GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph",
  533. data_node->GetName().c_str(), related_node_name.c_str());
  534. return INTERNAL_ERROR;
  535. }
  536. is_dynamic_batch = true;
  537. }
  538. return SUCCESS;
  539. }
  540. bool CheckOpType(const NodePtr &node, const std::string type) {
  541. if (node->GetType() == type) {
  542. return true;
  543. }
  544. return false;
  545. }
  546. Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) {
  547. auto op = node_ptr->GetOpDesc();
  548. GE_CHECK_NOTNULL(op);
  549. auto inputDescsPtr = op->GetAllInputsDescPtr();
  550. auto outputDescsPtr = op->GetAllOutputsDescPtr();
  551. ge::Format format = ge::FORMAT_ND;
  552. // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE
  553. // process, otherwise fe will insert transdata.
  554. for (auto &inputDescPtr : inputDescsPtr) {
  555. GE_CHECK_NOTNULL(inputDescPtr);
  556. if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  557. ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  558. GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str());
  559. inputDescPtr->SetFormat(format);
  560. inputDescPtr->SetOriginFormat(format);
  561. }
  562. }
  563. for (auto &outputDescPtr : outputDescsPtr) {
  564. GE_CHECK_NOTNULL(outputDescPtr);
  565. if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) &&
  566. ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) {
  567. GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str());
  568. outputDescPtr->SetFormat(format);
  569. outputDescPtr->SetOriginFormat(format);
  570. }
  571. }
  572. return SUCCESS;
  573. }
  574. // A new function ending in 'DynShape' has been added for the dynamic shape processing.
  575. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole
  576. // graph optimization, GE only sets the final data_type/format/shape information for variable,
  577. // data and netoutput, and no longer inserts the transnode.
  578. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) {
  579. GE_CHECK_NOTNULL(node_ptr);
  580. auto op_desc = node_ptr->GetOpDesc();
  581. GE_CHECK_NOTNULL(op_desc);
  582. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  583. GE_CHECK_NOTNULL(input);
  584. ge::DataType src_dtype = input->GetDataType();
  585. if (src_dtype == dt_set) {
  586. GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str());
  587. return SUCCESS;
  588. }
  589. input->SetDataType(dt_set);
  590. int64_t input_shape_size = 0;
  591. int64_t output_shape_size = 0;
  592. ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size);
  593. ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size);
  594. if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) {
  595. GELOGE(GRAPH_FAILED, "GetTensorSize failed!");
  596. return FAILED;
  597. }
  598. ge::TensorUtils::SetSize(*input, input_shape_size);
  599. const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0);
  600. GE_CHECK_NOTNULL(output);
  601. output->SetDataType(dt_set);
  602. ge::TensorUtils::SetSize(*output, output_shape_size);
  603. if (is_dynamic_batch) {
  604. GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str());
  605. auto switchn_op_desc = switchn_node->GetOpDesc();
  606. GE_CHECK_NOTNULL(switchn_op_desc);
  607. auto switchn_input = switchn_op_desc->MutableInputDesc(0);
  608. GE_CHECK_NOTNULL(switchn_input);
  609. switchn_input->SetDataType(dt_set);
  610. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  611. const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i);
  612. GE_CHECK_NOTNULL(switchn_output);
  613. switchn_output->SetDataType(dt_set);
  614. }
  615. }
  616. return SUCCESS;
  617. }
  618. Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) {
  619. GE_CHECK_NOTNULL(node_ptr);
  620. auto op_desc = node_ptr->GetOpDesc();
  621. GE_CHECK_NOTNULL(op_desc);
  622. const GeTensorDescPtr &input = op_desc->MutableInputDesc(0);
  623. GE_CHECK_NOTNULL(input);
  624. ge::Format old_format = input->GetFormat();
  625. ge::GeShape old_shape = input->GetShape();
  626. bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC));
  627. if (!support) {
  628. GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str());
  629. return FAILED;
  630. }
  631. if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) {
  632. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  633. return FAILED;
  634. }
  635. if (is_dynamic_batch) {
  636. auto switchn_op_desc = switchn_node->GetOpDesc();
  637. GE_CHECK_NOTNULL(switchn_op_desc);
  638. const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0);
  639. if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) {
  640. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  641. return FAILED;
  642. }
  643. for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) {
  644. auto switchn_output = switchn_op_desc->MutableOutputDesc(i);
  645. GE_CHECK_NOTNULL(switchn_output);
  646. old_format = switchn_output->GetFormat();
  647. old_shape = switchn_output->GetShape();
  648. if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) {
  649. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  650. return FAILED;
  651. }
  652. }
  653. }
  654. return SUCCESS;
  655. }
  656. Status ProcessDataNodeDynShape(NodePtr &node_ptr) {
  657. auto op_desc = node_ptr->GetOpDesc();
  658. GE_CHECK_NOTNULL(op_desc);
  659. string set_dt_str;
  660. if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) {
  661. return SUCCESS;
  662. }
  663. DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str);
  664. GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str());
  665. bool is_dynamic_batch = false;
  666. NodePtr switchn_node = nullptr;
  667. if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) {
  668. GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed");
  669. return FAILED;
  670. }
  671. if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) {
  672. GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed");
  673. return FAILED;
  674. }
  675. // check if need to set format
  676. string set_format;
  677. bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format);
  678. if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) {
  679. GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str());
  680. if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) {
  681. GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed");
  682. return FAILED;
  683. }
  684. }
  685. return SUCCESS;
  686. }
  687. Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr,
  688. Format &storage_format, vector<int64_t> &dst_shape_dims) {
  689. GE_CHECK_NOTNULL(op_desc);
  690. GE_CHECK_NOTNULL(tensor_desc_ptr);
  691. storage_format = FORMAT_RESERVED;
  692. int64_t format = FORMAT_RESERVED;
  693. dst_shape_dims.clear();
  694. if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) {
  695. storage_format = static_cast<Format>(format);
  696. vector<int32_t> storage_shape;
  697. if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) {
  698. for (auto dim : storage_shape) {
  699. dst_shape_dims.push_back(static_cast<int64_t>(dim));
  700. }
  701. GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]",
  702. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(),
  703. formats::JoinToString(storage_shape).c_str());
  704. } else {
  705. GELOGE(PARAM_INVALID, "Update node by storage format failed, storage_shape not set. "
  706. "node: [%s], storage_format [%s]",
  707. op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str());
  708. return FAILED;
  709. }
  710. ge::Format old_format = tensor_desc_ptr->GetFormat();
  711. auto old_shape = tensor_desc_ptr->GetShape().GetDims();
  712. if (old_format == storage_format && old_shape == dst_shape_dims) {
  713. GELOGI("Update node by storage format, not changed.");
  714. storage_format = FORMAT_RESERVED;
  715. return SUCCESS;
  716. }
  717. }
  718. return SUCCESS;
  719. }
  720. Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc,
  721. NodePtr &node) {
  722. bool is_dynamic = CheckOpType(node, MERGE);
  723. auto src_op_desc = node->GetOpDesc();
  724. GE_CHECK_NOTNULL(src_op_desc);
  725. ge::GeShape src_shape = src_desc.GetShape();
  726. ge::Format src_format = src_desc.GetFormat();
  727. net_output_input_desc->SetDataType(DT_FLOAT16);
  728. if (is_dynamic) {
  729. auto merge_output = src_op_desc->MutableOutputDesc(0);
  730. GE_CHECK_NOTNULL(merge_output);
  731. merge_output->SetDataType(DT_FLOAT16);
  732. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  733. auto merge_input = src_op_desc->MutableInputDesc(i);
  734. GE_CHECK_NOTNULL(merge_input);
  735. merge_input->SetDataType(DT_FLOAT16);
  736. }
  737. }
  738. std::vector<int64_t> dst_shape_dims;
  739. std::vector<int64_t> src_shape_dims = src_shape.GetDims();
  740. if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) {
  741. GELOGE(INTERNAL_ERROR, "Trans shape failed");
  742. return FAILED;
  743. }
  744. ge::GeShape dst_shape(dst_shape_dims);
  745. net_output_input_desc->SetFormat(FORMAT_NC1HWC0);
  746. net_output_input_desc->SetShape(dst_shape);
  747. if (is_dynamic) {
  748. auto merge_out = src_op_desc->MutableOutputDesc(0);
  749. GE_CHECK_NOTNULL(merge_out);
  750. if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) {
  751. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  752. return FAILED;
  753. }
  754. for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) {
  755. auto merge_in = src_op_desc->MutableInputDesc(i);
  756. GE_CHECK_NOTNULL(merge_in);
  757. if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) {
  758. GELOGE(INTERNAL_ERROR, "modify format and shape failed");
  759. return FAILED;
  760. }
  761. }
  762. }
  763. return SUCCESS;
  764. }
  765. bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) {
  766. GE_CHECK_NOTNULL(netout_desc);
  767. vector<string> output_dt_str;
  768. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) {
  769. for (auto dt_str : output_dt_str) {
  770. vector<string> dt_str_split = StringUtils::Split(dt_str, ':');
  771. if (dt_str_split.size() == kUserDefinedElementCount) {
  772. if (dt_str_split[0] == to_string(index)) {
  773. dt = TypeUtils::SerialStringToDataType(dt_str_split[1]);
  774. GELOGI("Find netoutput node output %u datatype should be set %s .", index,
  775. TypeUtils::DataTypeToSerialString(dt).c_str());
  776. return true;
  777. }
  778. }
  779. }
  780. }
  781. return false;
  782. }
  783. bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) {
  784. GE_CHECK_NOTNULL(netout_desc);
  785. vector<string> output_format_str;
  786. if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) {
  787. for (auto format_str : output_format_str) {
  788. vector<string> format_str_split = StringUtils::Split(format_str, ':');
  789. if (format_str_split.size() == kUserDefinedElementCount) {
  790. if (format_str_split[0] == to_string(index)) {
  791. GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index);
  792. return true;
  793. }
  794. }
  795. }
  796. }
  797. return false;
  798. }
  799. Status ProcessNetoutputNodeDynShape(NodePtr &node) {
  800. auto op_desc = node->GetOpDesc();
  801. GE_CHECK_NOTNULL(op_desc);
  802. ge::DataType output_data_type = ge::DT_FLOAT;
  803. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  804. auto index = static_cast<uint32_t>(in_anchor->GetIdx());
  805. auto peer_out = in_anchor->GetPeerOutAnchor();
  806. GE_CHECK_NOTNULL(peer_out);
  807. auto src_node = peer_out->GetOwnerNode();
  808. GE_CHECK_NOTNULL(src_node);
  809. bool is_dynamic = CheckOpType(src_node, MERGE);
  810. OpDescPtr src_op_desc = src_node->GetOpDesc();
  811. GE_CHECK_NOTNULL(src_op_desc);
  812. auto net_output_input_desc = op_desc->MutableInputDesc(index);
  813. GE_CHECK_NOTNULL(net_output_input_desc);
  814. ge::GeShape old_shape = net_output_input_desc->GetShape();
  815. ge::Format old_format = net_output_input_desc->GetFormat();
  816. ge::DataType old_dtype = net_output_input_desc->GetDataType();
  817. // Update datatype
  818. if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) {
  819. GELOGI("Enter into process output_type schedule");
  820. net_output_input_desc->SetDataType(output_data_type);
  821. if (is_dynamic) {
  822. auto merge_output = src_op_desc->MutableOutputDesc(0);
  823. GE_CHECK_NOTNULL(merge_output);
  824. merge_output->SetDataType(output_data_type);
  825. for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) {
  826. auto merge_input = src_op_desc->MutableInputDesc(i);
  827. GE_CHECK_NOTNULL(merge_input);
  828. merge_input->SetDataType(output_data_type);
  829. }
  830. }
  831. }
  832. // check if is_output_adjust_hw_layout is set
  833. if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) {
  834. if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) {
  835. GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0.");
  836. return FAILED;
  837. }
  838. GeTensorDesc old_desc(old_shape, old_format, old_dtype);
  839. if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) {
  840. GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0.");
  841. return FAILED;
  842. }
  843. }
  844. }
  845. return SUCCESS;
  846. }
  847. } // namespace
  848. GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {}
  849. GraphPrepare::~GraphPrepare() {}
  850. /**
  851. * @param graph
  852. * @return
  853. */
  854. Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) {
  855. GE_CHECK_NOTNULL(graph);
  856. auto var_names_to_refs = CollectVarNamesToRefs(graph);
  857. for (auto &node : graph->GetAllNodes()) {
  858. if (node == nullptr) {
  859. continue;
  860. }
  861. if (node->GetType() != VARIABLE) {
  862. continue;
  863. }
  864. auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName());
  865. if (trans_road == nullptr) {
  866. GELOGD("The variable %s does not have any trans road", node->GetName().c_str());
  867. continue;
  868. }
  869. GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str());
  870. auto ret = RecoverTransRoadForVar(node, *trans_road);
  871. if (ret != SUCCESS) {
  872. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str());
  873. return INTERNAL_ERROR;
  874. }
  875. auto iter = var_names_to_refs.find(node->GetName());
  876. if (iter != var_names_to_refs.end()) {
  877. ret = RecoverTransRoadForVarRef(iter->second, *trans_road);
  878. if (ret != SUCCESS) {
  879. GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str());
  880. return INTERNAL_ERROR;
  881. }
  882. }
  883. }
  884. return SUCCESS;
  885. }
  886. void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; }
  887. Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) {
  888. compute_graph_ = GraphUtils::GetComputeGraph(graph);
  889. if (compute_graph_ != nullptr) {
  890. compute_graph_->SetSessionID(session_id);
  891. }
  892. session_id_ = session_id;
  893. Status ret = CheckGraph();
  894. if (ret != SUCCESS) {
  895. GELOGE(ret, "RunGraph graph check fail, ret:%u", ret);
  896. return ret;
  897. }
  898. (void)compute_graph_->TopologicalSorting();
  899. ret = CheckRefOp();
  900. if (ret != SUCCESS) {
  901. GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret);
  902. return ret;
  903. }
  904. return SUCCESS;
  905. }
  906. Status GraphPrepare::CheckGraph() {
  907. if (compute_graph_ == nullptr) {
  908. GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR");
  909. return GE_GRAPH_INIT_FAILED;
  910. }
  911. auto nodes = compute_graph_->GetAllNodes();
  912. if (nodes.empty()) {
  913. GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph.");
  914. return GE_GRAPH_INIT_FAILED;
  915. }
  916. for (const NodePtr &node : compute_graph_->GetAllNodes()) {
  917. GE_CHECK_NOTNULL(node);
  918. if (node->GetOpDesc() == nullptr) {
  919. GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL");
  920. return GE_GRAPH_INIT_FAILED;
  921. }
  922. }
  923. return SUCCESS;
  924. }
  925. Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name,
  926. const std::set<NodePtr> &ref_nodes) {
  927. // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic
  928. // lossscale and would be optimized in SwitchToStreamSwitchPass.
  929. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch.
  930. static std::set<std::string> acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP,
  931. ge::REFSWITCH, ge::REFMERGE, ge::REFENTER,
  932. ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH};
  933. GE_CHECK_NOTNULL(node);
  934. const auto &op_desc = node->GetOpDesc();
  935. GE_CHECK_NOTNULL(op_desc);
  936. const auto input_index = op_desc->GetInputIndexByName(input_name);
  937. const auto &in_anchor = node->GetInDataAnchor(input_index);
  938. GE_CHECK_NOTNULL(in_anchor);
  939. const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor();
  940. GE_CHECK_NOTNULL(peer_out_anchor);
  941. const auto &input_node = peer_out_anchor->GetOwnerNode();
  942. GE_CHECK_NOTNULL(input_node);
  943. const auto &input_op_desc = input_node->GetOpDesc();
  944. GE_CHECK_NOTNULL(input_op_desc);
  945. bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end());
  946. if (is_ref) {
  947. return SUCCESS;
  948. }
  949. auto input_type = input_op_desc->GetType();
  950. if (input_type == ge::FRAMEWORKOP) {
  951. if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) {
  952. GELOGE(PARAM_INVALID, "Get original type failed.");
  953. return PARAM_INVALID;
  954. }
  955. }
  956. bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end());
  957. if (!is_acceptable) {
  958. GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.",
  959. node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(),
  960. input_op_desc->GetType().c_str());
  961. return PARAM_INVALID;
  962. }
  963. return SUCCESS;
  964. }
  965. Status GraphPrepare::CheckRefOp() {
  966. GE_CHECK_NOTNULL(compute_graph_);
  967. std::set<NodePtr> ref_nodes;
  968. for (const NodePtr &node : compute_graph_->GetDirectNode()) {
  969. if (node == nullptr) {
  970. GELOGE(PARAM_INVALID, "param [node] must not be null.");
  971. return PARAM_INVALID;
  972. }
  973. auto op_desc = node->GetOpDesc();
  974. if (op_desc == nullptr) {
  975. GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null.");
  976. return PARAM_INVALID;
  977. }
  978. auto input_name_index = op_desc->GetAllInputName();
  979. auto outputs = op_desc->GetAllOutputName();
  980. for (const auto &name_index : input_name_index) {
  981. if (op_desc->GetOutputIndexByName(name_index.first) != -1) {
  982. if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) {
  983. GELOGE(PARAM_INVALID, "CheckRefInputNode failed.");
  984. return PARAM_INVALID;
  985. }
  986. (void)ref_nodes.insert(node); // no need to check value
  987. }
  988. }
  989. }
  990. return SUCCESS;
  991. };
  992. Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) {
  993. GE_CHECK_NOTNULL(compute_graph_);
  994. GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_,
  995. compute_graph_->GetGraphID(), static_cast<int>(mode), ge::GetContext().DeviceId());
  996. GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId()));
  997. GE_CHK_RT_RET(rtCtxSetCurrent(rt_context));
  998. RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context);
  999. return SUCCESS;
  1000. }
  1001. Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) {
  1002. if (node == nullptr) {
  1003. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL");
  1004. return GE_GRAPH_GRAPH_NODE_NULL;
  1005. }
  1006. OpDescPtr op_desc_ptr = node->GetOpDesc();
  1007. if (op_desc_ptr == nullptr) {
  1008. GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL");
  1009. return GE_GRAPH_GRAPH_NODE_NULL;
  1010. }
  1011. GeTensorDesc output = op_desc_ptr->GetOutputDesc(0);
  1012. int64_t tensor_size = 0;
  1013. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size);
  1014. if (graph_status != GRAPH_SUCCESS) {
  1015. ErrorManager::GetInstance().ATCReportErrMessage(
  1016. "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()});
  1017. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  1018. return FAILED;
  1019. }
  1020. TensorUtils::SetSize(output, tensor_size);
  1021. graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output);
  1022. if (graph_ret != GRAPH_SUCCESS) {
  1023. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1024. return graph_ret;
  1025. }
  1026. return SUCCESS;
  1027. }
  1028. Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input) {
  1029. compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format));
  1030. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1031. GE_CHECK_NOTNULL(input_node);
  1032. OpDescPtr op = input_node->GetOpDesc();
  1033. GE_CHECK_NOTNULL(op);
  1034. if (op->GetType() == DATA) {
  1035. GeAttrValue::INT index = 0;
  1036. if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) {
  1037. GELOGW("Get index from data attr failed");
  1038. continue;
  1039. }
  1040. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1041. GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index);
  1042. return FAILED;
  1043. }
  1044. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1045. auto format = desc.GetFormat();
  1046. auto origin_format = desc.GetOriginFormat();
  1047. // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
  1048. bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op);
  1049. if (need_check_internal_format) {
  1050. bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
  1051. if (is_internal) {
  1052. GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.",
  1053. TypeUtils::FormatToSerialString(format).c_str(),
  1054. TypeUtils::FormatToSerialString(origin_format).c_str());
  1055. return FAILED;
  1056. }
  1057. }
  1058. auto data_type = desc.GetDataType();
  1059. uint32_t length = 1;
  1060. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1061. if (!type_ret) {
  1062. GELOGE(PARAM_INVALID, "Input datatype %s is not support.",
  1063. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1064. return FAILED;
  1065. }
  1066. int64_t desc_shape = desc.GetShape().GetShapeSize();
  1067. FMK_INT64_UINT32_MULCHECK(desc_shape, length);
  1068. int64_t shape_size = desc_shape * length;
  1069. GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast<int64_t>(length));
  1070. int64_t size = 0;
  1071. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS,
  1072. GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed");
  1073. return FAILED);
  1074. bool size_check = (size != 0 && shape_size != size);
  1075. if (size_check) {
  1076. GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size);
  1077. return FAILED;
  1078. }
  1079. ge::TensorUtils::SetSize(desc, shape_size);
  1080. graphStatus graph_ret = op->UpdateInputDesc(0, desc);
  1081. if (graph_ret != GRAPH_SUCCESS) {
  1082. GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
  1083. return graph_ret;
  1084. }
  1085. // Size will be recalculated in the build stage
  1086. ge::TensorUtils::SetSize(desc, 0);
  1087. graph_ret = op->UpdateOutputDesc(0, desc);
  1088. if (graph_ret != GRAPH_SUCCESS) {
  1089. GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
  1090. return graph_ret;
  1091. }
  1092. if (!options_.train_graph_flag) {
  1093. Status ret = AdjustDataOpOutput(input_node);
  1094. if (ret != SUCCESS) {
  1095. GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret);
  1096. return ret;
  1097. }
  1098. }
  1099. }
  1100. }
  1101. return SUCCESS;
  1102. }
  1103. Status GraphPrepare::TryDoAipp() {
  1104. // infer and with aipp configure file, then call aipp insert
  1105. if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) {
  1106. GE_DUMP(compute_graph_, "Before_insert_aipp");
  1107. Status ret = ge::InsertNewOpUtil::Instance().Init();
  1108. if (ret != SUCCESS) {
  1109. GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed.");
  1110. return INTERNAL_ERROR;
  1111. }
  1112. ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str());
  1113. if (ret != SUCCESS) {
  1114. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed",
  1115. options_.insert_op_file.c_str());
  1116. return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED;
  1117. }
  1118. ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file);
  1119. if (ret != SUCCESS) {
  1120. GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret);
  1121. return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED;
  1122. }
  1123. }
  1124. return SUCCESS;
  1125. }
  1126. Status GraphPrepare::FormatAndShapeProcess() {
  1127. Status ret = ResourcePairProcess("add");
  1128. if (ret != SUCCESS) {
  1129. GELOGE(ret, "ResourcePairProcess failed");
  1130. return ret;
  1131. }
  1132. GE_TIMESTAMP_START(InferOriginFormat1);
  1133. ret = compute_graph_->InferOriginFormat();
  1134. GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1");
  1135. GE_DUMP(compute_graph_, "after_first_inferformat");
  1136. if (ret != SUCCESS) {
  1137. GELOGE(ret, "Prepare Graph first inferformat failed");
  1138. return ret;
  1139. }
  1140. GE_TIMESTAMP_START(InferShapeForPreprocess);
  1141. ret = InferShapeForPreprocess();
  1142. GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess");
  1143. GE_DUMP(compute_graph_, "after_infershape");
  1144. if (ret != SUCCESS) {
  1145. GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed");
  1146. return GE_GRAPH_INFERSHAPE_FAILED;
  1147. }
  1148. GE_TIMESTAMP_START(InferOriginFormat2);
  1149. ret = compute_graph_->InferOriginFormat();
  1150. GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2");
  1151. if (ret != SUCCESS) {
  1152. GELOGE(ret, "Prepare Graph inferformat failed");
  1153. return ret;
  1154. }
  1155. ret = ResourcePairProcess("remove");
  1156. if (ret != SUCCESS) {
  1157. return ret;
  1158. }
  1159. return ret;
  1160. }
  1161. Status GraphPrepare::ResourcePairProcess(const std::string &action) {
  1162. PassManager control_pass;
  1163. // Graph pass tmp logic for resource infershape
  1164. if (options_.train_graph_flag) {
  1165. try {
  1166. if (action == "add") {
  1167. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass);
  1168. } else {
  1169. (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass",
  1170. new ResourcePairRemoveControlPass);
  1171. }
  1172. } catch (std::bad_alloc &e) {
  1173. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str());
  1174. return INTERNAL_ERROR;
  1175. }
  1176. }
  1177. Status ret = control_pass.Run(compute_graph_);
  1178. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1179. GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret);
  1180. return ret;
  1181. }
  1182. return SUCCESS;
  1183. }
  1184. Status GraphPrepare::UpdateDataNetOutputByStorageFormat() {
  1185. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1186. GE_CHECK_NOTNULL(node_ptr);
  1187. if (node_ptr->GetType() == DATA) {
  1188. uint32_t index = 0;
  1189. auto op_desc = node_ptr->GetOpDesc();
  1190. GE_CHECK_NOTNULL(op_desc);
  1191. const GeTensorDescPtr input = op_desc->MutableInputDesc(index);
  1192. Format storage_format = FORMAT_RESERVED;
  1193. vector<int64_t> dst_shape_dims;
  1194. if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) {
  1195. GELOGE(INTERNAL_ERROR, "Get storage format for input failed");
  1196. return FAILED;
  1197. }
  1198. if (storage_format == FORMAT_RESERVED) {
  1199. continue;
  1200. }
  1201. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1202. GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed");
  1203. return FAILED;
  1204. }
  1205. }
  1206. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1207. auto op_desc = node_ptr->GetOpDesc();
  1208. GE_CHECK_NOTNULL(op_desc);
  1209. for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) {
  1210. const GeTensorDescPtr output = op_desc->MutableOutputDesc(index);
  1211. Format storage_format = FORMAT_RESERVED;
  1212. vector<int64_t> dst_shape_dims;
  1213. if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) {
  1214. GELOGE(INTERNAL_ERROR, "Get storage format from output failed");
  1215. return FAILED;
  1216. }
  1217. if (storage_format == FORMAT_RESERVED) {
  1218. continue;
  1219. }
  1220. if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) {
  1221. GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed");
  1222. return FAILED;
  1223. }
  1224. }
  1225. }
  1226. }
  1227. return SUCCESS;
  1228. }
  1229. Status GraphPrepare::SaveOriginalGraphToOmModel() {
  1230. if (options_.save_original_model == "true") {
  1231. ModelHelper model_helper;
  1232. Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_),
  1233. options_.original_model_file);
  1234. if (ret != SUCCESS) {
  1235. // If save original model fail, process continue
  1236. GELOGW("SaveOriginalGraphToOmModel fail");
  1237. }
  1238. }
  1239. return SUCCESS;
  1240. }
  1241. #define PP_RUN_AND_DUMP(name, func, ...) \
  1242. do { \
  1243. GE_RUN(Prepare, func, __VA_ARGS__); \
  1244. GE_DUMP(compute_graph, "PrepareAfter" name); \
  1245. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1246. } while (0)
  1247. #define PP_RUN(name, func, ...) \
  1248. do { \
  1249. GE_RUN(Prepare, func, __VA_ARGS__); \
  1250. GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \
  1251. } while (0)
  1252. Status GraphPrepare::PrepareDynShape(ConstGraphPtr graph, const std::vector<GeTensor> &user_input,
  1253. ge::ComputeGraphPtr &compute_graph, uint64_t session_id) {
  1254. GE_CHECK_NOTNULL(graph);
  1255. GE_CHECK_NOTNULL(compute_graph);
  1256. GetLocalOmgContext().type = static_cast<domi::FrameworkType>(options_.framework_type);
  1257. const Graph &const_graph = *graph;
  1258. PP_RUN("Init", Init, const_graph, session_id);
  1259. PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE);
  1260. PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input);
  1261. PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation);
  1262. PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput);
  1263. PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_);
  1264. PP_RUN_AND_DUMP("InsertAipp", TryDoAipp);
  1265. PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape);
  1266. PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess);
  1267. PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_);
  1268. PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_);
  1269. PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel);
  1270. PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize);
  1271. return SUCCESS;
  1272. }
  1273. Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) {
  1274. PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_);
  1275. return SUCCESS;
  1276. }
  1277. Status GraphPrepare::PrepareRunningFormatRefiner() {
  1278. auto compute_graph = compute_graph_;
  1279. PassManager pass_manager;
  1280. GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass",
  1281. new (std::nothrow) VariablePrepareOpPass))
  1282. GE_TIMESTAMP_START(pass_manager);
  1283. auto ret = pass_manager.Run(compute_graph);
  1284. GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner");
  1285. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1286. GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret);
  1287. return ret;
  1288. }
  1289. PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions);
  1290. PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_);
  1291. return SUCCESS;
  1292. }
  1293. Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) {
  1294. if (compute_graph == nullptr) {
  1295. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1296. return GE_GRAPH_NULL_INPUT;
  1297. }
  1298. GEPass ge_passes(compute_graph);
  1299. NamesToPass hccl_group;
  1300. HcclGroupPass hccl_group_pass;
  1301. GELOGD("Add hccl group pass success");
  1302. hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass);
  1303. auto ret = ge_passes.Run(hccl_group);
  1304. if (ret != SUCCESS) {
  1305. GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret);
  1306. return ret;
  1307. }
  1308. ret = compute_graph->TopologicalSorting();
  1309. if (ret != SUCCESS) {
  1310. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1311. return ret;
  1312. }
  1313. return SUCCESS;
  1314. }
  1315. #undef PP_RUN_AND_DUMP
  1316. #undef PP_RUN
  1317. Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) {
  1318. if (graph == nullptr) {
  1319. GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL");
  1320. return GE_GRAPH_NULL_INPUT;
  1321. }
  1322. const Graph &const_graph = *graph;
  1323. Status ret = Init(const_graph, 0);
  1324. if (ret != SUCCESS) {
  1325. GELOGE(ret, "Init graph_prepare fail, ret:%u", ret);
  1326. return ret;
  1327. }
  1328. GE_DUMP(compute_graph_, "after_parser");
  1329. GELOGI("Start infershape for dump json process.");
  1330. ret = compute_graph_->InferOriginFormat();
  1331. GE_DUMP(compute_graph_, "after_inferformat");
  1332. if (ret != SUCCESS) {
  1333. GELOGE(ret, "Prepare Graph inferformat failed");
  1334. return ret;
  1335. }
  1336. InferShapePass infer_shape_pass;
  1337. NamesToPass names_to_passes;
  1338. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1339. GEPass ge_passes(compute_graph_);
  1340. ret = ge_passes.Run(names_to_passes);
  1341. GE_DUMP(compute_graph_, "after_infershape");
  1342. if (ret != SUCCESS) {
  1343. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1344. return ret;
  1345. }
  1346. ShapeRefiner::ClearContextMap();
  1347. return SUCCESS;
  1348. }
  1349. Status GraphPrepare::CheckConstOp() {
  1350. for (auto &node_ptr : compute_graph_->GetAllNodes()) {
  1351. GE_CHECK_NOTNULL(node_ptr);
  1352. if (node_ptr->GetType() == CONSTANT) {
  1353. Status ret = VerifyConstOp(node_ptr);
  1354. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1355. } else if (node_ptr->GetType() == FRAMEWORKOP) {
  1356. auto op_desc = node_ptr->GetOpDesc();
  1357. if (op_desc == nullptr) {
  1358. GELOGE(PARAM_INVALID, "Get op desc failed");
  1359. return PARAM_INVALID;
  1360. }
  1361. std::string original_type;
  1362. GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type),
  1363. GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str()));
  1364. GELOGI("original type is %s", original_type.c_str());
  1365. if (original_type == CONSTANT) {
  1366. Status ret = VerifyConstOp(node_ptr);
  1367. GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed");
  1368. }
  1369. }
  1370. }
  1371. return SUCCESS;
  1372. }
  1373. Status GraphPrepare::VerifyConstOp(const NodePtr &node) {
  1374. GE_CHECK_NOTNULL(node);
  1375. auto op_desc = node->GetOpDesc();
  1376. GE_CHECK_NOTNULL(op_desc);
  1377. ConstGeTensorPtr ge_tensor_ptr;
  1378. if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) {
  1379. GELOGE(PARAM_INVALID, "Get value from const attr failed");
  1380. return PARAM_INVALID;
  1381. }
  1382. GE_CHECK_NOTNULL(ge_tensor_ptr);
  1383. auto data_size = ge_tensor_ptr->GetData().GetSize();
  1384. auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc();
  1385. int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize();
  1386. auto data_type = ge_tensor_desc.GetDataType();
  1387. uint32_t length = 1;
  1388. bool type_ret = TypeUtils::GetDataTypeLength(data_type, length);
  1389. if (!type_ret) {
  1390. GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str());
  1391. return FAILED;
  1392. }
  1393. FMK_INT64_UINT32_MULCHECK(shape_size, length);
  1394. GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length,
  1395. TypeUtils::DataTypeToSerialString(data_type).c_str());
  1396. if (shape_size == 0) {
  1397. if (ge_tensor_desc.GetShape().GetDims().size() == 0) {
  1398. // shape = [], means it's a sclar tensor.
  1399. GE_CHK_BOOL_EXEC(data_size / length == 1, return PARAM_INVALID, "Const is invalid scalar tensor.");
  1400. } else {
  1401. // shape = [x, y, 0,...], means it's a vector tensor that value is [].
  1402. GE_CHK_BOOL_EXEC(data_size == 0, return PARAM_INVALID, "Const is invalid vector scalar.");
  1403. }
  1404. } else {
  1405. GE_CHK_BOOL_EXEC(data_size == static_cast<size_t>(shape_size * length) && data_size != 0, return PARAM_INVALID,
  1406. "Const input data size is not equal with tensor desc shape");
  1407. }
  1408. return SUCCESS;
  1409. }
  1410. Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
  1411. if (GetLocalOmgContext().is_dynamic_input) {
  1412. return SUCCESS;
  1413. }
  1414. unsigned int node_num = 0;
  1415. unsigned int data_num = 0;
  1416. for (NodePtr &input_node : compute_graph_->GetDirectNode()) {
  1417. GE_CHECK_NOTNULL(input_node);
  1418. OpDescPtr op = input_node->GetOpDesc();
  1419. GE_CHECK_NOTNULL(op);
  1420. node_num++;
  1421. if (op->GetType() == DATA || op->GetType() == AIPPDATA) {
  1422. data_num++;
  1423. GeAttrValue::INT index = 0;
  1424. if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) {
  1425. GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed");
  1426. return GE_GRAPH_INIT_FAILED;
  1427. }
  1428. if ((index < 0) || (static_cast<size_t>(index) >= user_input.size())) {
  1429. GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index);
  1430. return GE_GRAPH_INIT_FAILED;
  1431. }
  1432. GeTensorDesc desc(user_input[index].GetTensorDesc());
  1433. for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
  1434. if (desc.GetShape().GetDim(i) < 0) {
  1435. GELOGE(GE_GRAPH_INIT_FAILED, "data dim %zu is not supported, need >= 0, real:%ld.", i,
  1436. desc.GetShape().GetDim(i));
  1437. return GE_GRAPH_INIT_FAILED;
  1438. }
  1439. }
  1440. }
  1441. }
  1442. if (node_num <= data_num) {
  1443. GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num);
  1444. }
  1445. return SUCCESS;
  1446. }
  1447. Status GraphPrepare::InferShapeForPreprocess() {
  1448. GELOGI("Start infershape for preprocess.");
  1449. GEPass ge_passes(compute_graph_);
  1450. NamesToPass names_to_passes;
  1451. AssertPass assert_pass;
  1452. if (!options_.train_graph_flag) {
  1453. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1454. }
  1455. InferShapePass infer_shape_pass;
  1456. names_to_passes.emplace_back("InferShapePass", &infer_shape_pass);
  1457. ReplaceWithEmptyConstPass replace_with_empty_const_pass;
  1458. names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass);
  1459. DimensionComputePass dimension_compute_pass;
  1460. names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass);
  1461. ConstantFoldingPass constant_folding_pass;
  1462. names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
  1463. int32_t dev_count = 0;
  1464. AicpuConstantFoldingPass aicpu_constant_folding_pass;
  1465. const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON");
  1466. rtError_t rt_err = RT_ERROR_NONE;
  1467. if (aicpu_constant_folding_on != nullptr) {
  1468. rt_err = rtGetDeviceCount(&dev_count);
  1469. if (rt_err == RT_ERROR_NONE) {
  1470. Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE);
  1471. if (result != SUCCESS) {
  1472. GELOGE(result, "Set rt context failed.");
  1473. return result;
  1474. }
  1475. names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass);
  1476. }
  1477. }
  1478. Status ret = ge_passes.Run(names_to_passes);
  1479. if (aicpu_constant_folding_on != nullptr) {
  1480. if (rt_err == RT_ERROR_NONE) {
  1481. Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE);
  1482. if (result != SUCCESS) {
  1483. GELOGE(result, "Set rt context failed.");
  1484. return result;
  1485. }
  1486. }
  1487. }
  1488. ShapeRefiner::ClearContextMap();
  1489. if (ret != SUCCESS) {
  1490. GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret);
  1491. return ret;
  1492. }
  1493. return SUCCESS;
  1494. }
  1495. Status GraphPrepare::PrepareOptimize() {
  1496. GELOGI("Start optimize for preprocess.");
  1497. // check rw type
  1498. GraphOptimize graph_optimize;
  1499. bool has_conflict = false;
  1500. graph_optimize.CheckRWConflict(compute_graph_, has_conflict);
  1501. if (has_conflict) {
  1502. GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize.");
  1503. return FAILED;
  1504. }
  1505. PassManager original_graph_passes;
  1506. // Graph pass
  1507. try {
  1508. (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass);
  1509. (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass);
  1510. (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass);
  1511. } catch (std::bad_alloc &e) {
  1512. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1513. return INTERNAL_ERROR;
  1514. }
  1515. GE_TIMESTAMP_START(original_graph_passes);
  1516. Status ret = original_graph_passes.Run(compute_graph_);
  1517. GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses");
  1518. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1519. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1520. return ret;
  1521. }
  1522. // New pass
  1523. GEPass ge_passes(compute_graph_);
  1524. NamesToPass names_to_passes;
  1525. EnterPass enter_pass;
  1526. names_to_passes.emplace_back("EnterPass", &enter_pass);
  1527. CondPass cond_pass;
  1528. names_to_passes.emplace_back("CondPass", &cond_pass);
  1529. PrintOpPass print_pass;
  1530. if (options_.enable_print_op_pass) {
  1531. names_to_passes.emplace_back("PrintOpPass", &print_pass);
  1532. }
  1533. NoUseReshapeRemovePass no_use_reshape_remove_pass;
  1534. names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass);
  1535. DropOutPass dropout_pass;
  1536. AssertPass assert_pass;
  1537. UnusedConstPass unused_const_pass;
  1538. StopGradientPass stop_gradient_pass;
  1539. PreventGradientPass prevent_gradient_pass;
  1540. PlaceholderWithDefaultPass placeholder_with_default_pass;
  1541. GuaranteeConstPass guarantee_const_pass;
  1542. VarIsInitializedOpPass var_is_initialized_pass;
  1543. ParallelConcatStartOpPass parallel_concat_start_op_pass;
  1544. IdentityPass identity_pass(false);
  1545. AssignPass assign_pass;
  1546. SnapshotPass snapshot_pass;
  1547. if (!options_.train_graph_flag) {
  1548. names_to_passes.emplace_back("DropOutPass", &dropout_pass);
  1549. names_to_passes.emplace_back("AssertPass", &assert_pass);
  1550. }
  1551. names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass);
  1552. names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass);
  1553. names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass);
  1554. names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass);
  1555. names_to_passes.emplace_back("SnapshotPass", &snapshot_pass);
  1556. names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass);
  1557. names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
  1558. names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
  1559. names_to_passes.emplace_back("IdentityPass", &identity_pass);
  1560. if (GetContext().GetHostExecFlag()) {
  1561. names_to_passes.emplace_back("AssignPass", &assign_pass);
  1562. }
  1563. GE_TIMESTAMP_START(names_to_passes);
  1564. ret = ge_passes.Run(names_to_passes);
  1565. GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses");
  1566. if (ret != SUCCESS) {
  1567. GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret);
  1568. return ret;
  1569. }
  1570. PassManager graph_pass;
  1571. try {
  1572. (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass);
  1573. // todo 临时把hccl的memcpy插入放到图准备,为了防止其多插memcpy
  1574. (void)graph_pass.AddPass("PrepareOptimize::HcclMemcpyPass", new (std::nothrow) HcclMemcpyPass);
  1575. } catch (std::bad_alloc &e) {
  1576. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1577. return INTERNAL_ERROR;
  1578. }
  1579. GE_TIMESTAMP_START(graph_passes);
  1580. ret = graph_pass.Run(compute_graph_);
  1581. GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses");
  1582. if (ret != SUCCESS && ret != NOT_CHANGED) {
  1583. GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret);
  1584. return ret;
  1585. }
  1586. // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future.
  1587. TypeConversionOfConstant();
  1588. ret = compute_graph_->TopologicalSorting();
  1589. if (ret != SUCCESS) {
  1590. GELOGE(ret, "Graph topological sort failed, ret:%u.", ret);
  1591. return ret;
  1592. }
  1593. GELOGI("End optimize for preprocess.");
  1594. return SUCCESS;
  1595. }
  1596. void GraphPrepare::TypeConversionOfConstant() {
  1597. if (options_.train_graph_flag) {
  1598. GELOGD("trans CONSTANT to CONSTANTOP in train.");
  1599. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1600. // This can ensure that n is not a null pointer
  1601. if (n->GetOpDesc()->GetType() == CONSTANT) {
  1602. n->GetOpDesc()->SetType(CONSTANTOP);
  1603. }
  1604. }
  1605. } else {
  1606. GELOGD("trans CONSTANTOP to CONSTANT in inferrence.");
  1607. for (ge::NodePtr &n : compute_graph_->GetAllNodes()) {
  1608. // This can ensure that n is not a null pointer
  1609. if (n->GetOpDesc()->GetType() == CONSTANTOP) {
  1610. n->GetOpDesc()->SetType(CONSTANT);
  1611. }
  1612. }
  1613. }
  1614. }
  1615. Status GraphPrepare::GraphEquivalentTransformation() {
  1616. NamesToPass names_to_pass;
  1617. ForPass for_pass;
  1618. names_to_pass.emplace_back("ForToWhilePass", &for_pass);
  1619. return GEPass(compute_graph_).Run(names_to_pass);
  1620. }
  1621. Status GraphPrepare::ProcessBeforeInfershape() {
  1622. NamesToPass names_to_passes;
  1623. CondRemovePass condition_remove_pass;
  1624. names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
  1625. GE_TIMESTAMP_START(ProcessCondRemove);
  1626. auto ret = GEPass(compute_graph_).Run(names_to_passes);
  1627. GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove");
  1628. if (ret != SUCCESS) {
  1629. GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
  1630. return ret;
  1631. }
  1632. return SUCCESS;
  1633. }
  1634. Status GraphPrepare::ProcessNetOutput() {
  1635. PassManager graph_passes_before_infershape;
  1636. try {
  1637. if (options_.train_graph_flag) {
  1638. graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass);
  1639. }
  1640. graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass);
  1641. graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass",
  1642. new (std::nothrow) DataPass); // Add NetOutput first.
  1643. } catch (std::bad_alloc) {
  1644. GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs.");
  1645. return INTERNAL_ERROR;
  1646. }
  1647. auto ret = graph_passes_before_infershape.Run(compute_graph_);
  1648. if ((ret != SUCCESS) && (ret != NOT_CHANGED)) {
  1649. GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret);
  1650. return ret;
  1651. }
  1652. return SUCCESS;
  1653. }
  1654. Status GraphPrepare::CheckAndUpdateInput(const std::vector<GeTensor> &user_input) {
  1655. compute_graph_->SetInputSize(user_input.size());
  1656. if (user_input.empty()) {
  1657. return SUCCESS;
  1658. }
  1659. auto ret = CheckUserInput(user_input);
  1660. if (ret != SUCCESS) {
  1661. GELOGE(ret, "Check user input failed.");
  1662. return ret;
  1663. }
  1664. ret = UpdateInput(user_input);
  1665. if (ret != SUCCESS) {
  1666. GELOGE(ret, "UpdateInput fail, ret:%u", ret);
  1667. return ret;
  1668. }
  1669. if (user_input.size() != 0) {
  1670. ret = CheckConstOp();
  1671. if (ret != SUCCESS) {
  1672. GELOGE(ret, "CheckConstOp fail, ret:%u", ret);
  1673. return ret;
  1674. }
  1675. } else {
  1676. ret = compute_graph_->TopologicalSorting();
  1677. if (ret != SUCCESS) {
  1678. GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting");
  1679. return FAILED;
  1680. }
  1681. }
  1682. return SUCCESS;
  1683. }
  1684. Status GraphPrepare::UpdateInputOutputByOptions() {
  1685. auto ret = UpdateDataNetOutputByStorageFormat();
  1686. if (ret != SUCCESS) {
  1687. GELOGE(ret, "Update format acoording to storage format failed.");
  1688. return ret;
  1689. }
  1690. if (options_.train_graph_flag) {
  1691. GELOGI("This is train mode, no need to do this schedule.");
  1692. return SUCCESS;
  1693. }
  1694. for (auto &node_ptr : compute_graph_->GetDirectNode()) {
  1695. GE_CHECK_NOTNULL(node_ptr);
  1696. if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) {
  1697. GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str());
  1698. return FAILED;
  1699. }
  1700. if (node_ptr->GetType() == DATA) {
  1701. if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) {
  1702. GELOGE(INTERNAL_ERROR, "Process data node failed");
  1703. return FAILED;
  1704. }
  1705. }
  1706. if (node_ptr->GetType() == ge::NETOUTPUT) {
  1707. if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) {
  1708. GELOGE(INTERNAL_ERROR, "Process netoutput node failed");
  1709. return FAILED;
  1710. }
  1711. }
  1712. }
  1713. return SUCCESS;
  1714. }
  1715. bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) {
  1716. for (auto &out_anchor : var_node->GetAllOutDataAnchors()) {
  1717. GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
  1718. for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  1719. GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
  1720. ge::NodePtr dst_node = in_anchor->GetOwnerNode();
  1721. GE_RT_FALSE_CHECK_NOTNULL(dst_node);
  1722. if (dst_node->GetType() == TRANSDATA) {
  1723. return true;
  1724. }
  1725. }
  1726. }
  1727. return false;
  1728. }
  1729. } // 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