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.

util_insert_aipp_op.cc 34 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784
  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/insert_op/util_insert_aipp_op.h"
  17. #include <fstream>
  18. #include <utility>
  19. #include "common/dynamic_aipp.h"
  20. #include "common/formats/utils/formats_trans_utils.h"
  21. #include "common/ge/ge_util.h"
  22. #include "common/op/ge_op_utils.h"
  23. #include "common/util.h"
  24. #include "common/util/error_manager/error_manager.h"
  25. #include "framework/common/debug/ge_log.h"
  26. #include "framework/common/debug/log.h"
  27. #include "framework/common/ge_inner_error_codes.h"
  28. #include "framework/omg/omg_inner_types.h"
  29. #include "graph/debug/ge_attr_define.h"
  30. #include "graph/preprocess/insert_op/ge_aipp_op.h"
  31. #include "graph/utils/attr_utils.h"
  32. #include "graph/utils/graph_utils.h"
  33. #include "graph/utils/op_desc_utils.h"
  34. #include "graph/utils/tensor_utils.h"
  35. #include "graph/utils/type_utils.h"
  36. #include "util_insert_aipp_op.h"
  37. using domi::AippOpParams;
  38. namespace ge {
  39. namespace {
  40. const char *const kMbatchSwitchnName = "mbatch-switch-name";
  41. const int64_t kFormatAgnosticSwitch = 1;
  42. const int64_t kFormatDependInputIndex = 1;
  43. } // namespace
  44. static void ConvertShape2Nhwc(Format &format, vector<int64_t> &shape_vec) {
  45. if ((format == FORMAT_NHWC) || (shape_vec.size() != static_cast<size_t>(NORMAL_TENSOR_SIZE))) {
  46. return;
  47. }
  48. if (format != FORMAT_NCHW) {
  49. GELOGW("The format is not NCHW, current format is %s.", TypeUtils::FormatToSerialString(format).c_str());
  50. return;
  51. }
  52. vector<int64_t> shape_vec_tmp;
  53. shape_vec.swap(shape_vec_tmp);
  54. shape_vec.push_back(shape_vec_tmp[NCHW_DIM_N]);
  55. shape_vec.push_back(shape_vec_tmp[NCHW_DIM_H]);
  56. shape_vec.push_back(shape_vec_tmp[NCHW_DIM_W]);
  57. shape_vec.push_back(shape_vec_tmp[NCHW_DIM_C]);
  58. return;
  59. }
  60. Status InsertNewOpUtil::Init() {
  61. insert_op_conf_.reset((new (std::nothrow) domi::InsertNewOps()));
  62. GE_CHECK_NOTNULL(insert_op_conf_);
  63. return SUCCESS;
  64. }
  65. Status InsertNewOpUtil::Parse(const char *conf_path) {
  66. if (conf_path == nullptr || *conf_path == '\0') {
  67. return SUCCESS;
  68. }
  69. GE_CHK_BOOL_RET_STATUS(ReadProtoFromText(conf_path, insert_op_conf_.get()), FAILED, "Read AIPP conf file error: %s",
  70. conf_path);
  71. GE_CHK_STATUS_RET(CheckPositionNotRepeat(), "Check insert position of op failed");
  72. for (int i = 0; i < insert_op_conf_->aipp_op_size(); i++) {
  73. domi::AippOpParams *aipp_op_params = insert_op_conf_->mutable_aipp_op(i);
  74. std::unique_ptr<AippOp> aipp_op(new (std::nothrow) AippOp());
  75. GE_CHECK_NOTNULL(aipp_op);
  76. GE_CHK_STATUS_RET(aipp_op->Init(aipp_op_params), "Aipp op init failed.");
  77. insert_ops_.push_back(std::move(aipp_op));
  78. }
  79. for (auto &dynamic_op : insert_ops_) {
  80. GE_CHECK_NOTNULL(dynamic_op);
  81. GE_CHK_STATUS_RET(dynamic_op->ValidateParams(), "Validate insert_op config file failed");
  82. GE_CHK_STATUS_RET(dynamic_op->SetDefaultParams(), "Set default value of insert_op failed");
  83. }
  84. return SUCCESS;
  85. }
  86. Status InsertNewOpUtil::InsertAippOps(ComputeGraphPtr &graph, std::string &aippConfigPath) {
  87. GE_CHECK_NOTNULL(graph);
  88. for (uint32_t index = 0; index < insert_ops_.size(); ++index) {
  89. GE_CHK_STATUS_RET(insert_ops_[index]->InsertAippToGraph(graph, aippConfigPath, index), "insert op to graph failed");
  90. }
  91. GE_CHK_STATUS_RET(CheckGraph(graph), "after inserting all ops, check graph failed");
  92. GE_CHK_STATUS_RET(graph->TopologicalSorting(), "after insert dynamic op, sort graph failed");
  93. ClearNewOps();
  94. return SUCCESS;
  95. }
  96. void InsertNewOpUtil::ClearNewOps() {
  97. if (insert_op_conf_ != nullptr) {
  98. insert_op_conf_->Clear();
  99. insert_ops_.clear();
  100. }
  101. }
  102. Status InsertNewOpUtil::CheckInputNamePositionNotRepeat() {
  103. for (int i = 0; i < insert_op_conf_->aipp_op_size(); i++) {
  104. const domi::AippOpParams *item = insert_op_conf_->mutable_aipp_op(i);
  105. GE_CHECK_NOTNULL(item);
  106. for (int j = i + 1; j < insert_op_conf_->aipp_op_size(); j++) {
  107. const domi::AippOpParams *another_item = insert_op_conf_->mutable_aipp_op(j);
  108. GE_CHECK_NOTNULL(another_item);
  109. if (another_item->related_input_name().empty()) {
  110. string error_msg = "Can not both set related_input_name and related_input_rank!"
  111. " Please ensure param is the same with the first aipp config(related_input_name).";
  112. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {error_msg});
  113. GELOGE(PARAM_INVALID,
  114. "Can not both set related_input_rank and related_input_name!"
  115. " Please ensure param is the same with the first aipp config(related_input_name).");
  116. return PARAM_INVALID;
  117. }
  118. if (item->related_input_name() == another_item->related_input_name()) {
  119. string error_msg = "Can not insert aipp to the same postion! Please ensure related_input_name"
  120. " param is different in different aipp config.";
  121. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {error_msg});
  122. GELOGE(PARAM_INVALID,
  123. "Can not insert aipp op to the same postion! Please ensure related_input_rank param "
  124. "is different in different aipp config.");
  125. return PARAM_INVALID;
  126. }
  127. }
  128. }
  129. return SUCCESS;
  130. }
  131. Status InsertNewOpUtil::CheckInputRankPositionNoRepeat() {
  132. for (int i = 0; i < insert_op_conf_->aipp_op_size(); i++) {
  133. const domi::AippOpParams *item = insert_op_conf_->mutable_aipp_op(i);
  134. GE_CHECK_NOTNULL(item);
  135. for (int j = i + 1; j < insert_op_conf_->aipp_op_size(); j++) {
  136. const domi::AippOpParams *another_item = insert_op_conf_->mutable_aipp_op(j);
  137. GE_CHECK_NOTNULL(another_item);
  138. if (!another_item->related_input_name().empty()) {
  139. string error_msg = "Can not both set related_input_rank and related_input_name!"
  140. " Please ensure param is the same with the first aipp config(related_input_rank).";
  141. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {error_msg});
  142. GELOGE(PARAM_INVALID,
  143. "Can not both set related_input_rank and related_input_name!"
  144. " Please ensure param is the same with the first aipp config(related_input_rank).");
  145. return PARAM_INVALID;
  146. }
  147. if (item->related_input_rank() == another_item->related_input_rank()) {
  148. string error_msg = "Can not insert aipp to the same postion! Please ensure related_input_rank"
  149. " param is different in different aipp config.";
  150. ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {error_msg});
  151. GELOGE(PARAM_INVALID,
  152. "Can not insert aipp op to the same postion! Please ensure related_input_rank param "
  153. "is different in different aipp config.");
  154. return PARAM_INVALID;
  155. }
  156. }
  157. }
  158. return SUCCESS;
  159. }
  160. Status InsertNewOpUtil::CheckPositionNotRepeat() {
  161. GE_CHECK_NOTNULL(insert_op_conf_);
  162. if (insert_op_conf_->aipp_op_size() <= 1) {
  163. GELOGI("Aipp op size[%d] less than 2, no need to check position repeat.", insert_op_conf_->aipp_op_size());
  164. return SUCCESS;
  165. }
  166. const domi::AippOpParams *item = insert_op_conf_->mutable_aipp_op(0);
  167. GE_CHECK_NOTNULL(item);
  168. string related_input_name = item->related_input_name();
  169. Status ret = FAILED;
  170. if (related_input_name.empty()) {
  171. ret = CheckInputRankPositionNoRepeat();
  172. } else {
  173. ret = CheckInputNamePositionNotRepeat();
  174. }
  175. if (ret != SUCCESS) {
  176. GELOGE(FAILED, "Check position not repeat failed.");
  177. return FAILED;
  178. }
  179. return SUCCESS;
  180. }
  181. Status InsertNewOpUtil::CheckGraph(const ComputeGraphPtr &graph) {
  182. GE_CHECK_NOTNULL(graph);
  183. domi::AippOpParams::AippMode aippMode = domi::AippOpParams::undefined;
  184. for (const auto &node : graph->GetDirectNode()) {
  185. if (node->GetType() != DATA) {
  186. continue;
  187. }
  188. size_t next_nodes_cnt = 0;
  189. std::vector<NodePtr> aippNodes;
  190. for (const auto &anchor : node->GetAllOutDataAnchors()) {
  191. for (const auto &inAnchor : anchor->GetPeerInDataAnchors()) {
  192. const std::string &nodeType = inAnchor->GetOwnerNode()->GetType();
  193. next_nodes_cnt++;
  194. if (nodeType == AIPP) {
  195. aippNodes.push_back(inAnchor->GetOwnerNode());
  196. continue;
  197. }
  198. }
  199. }
  200. GE_CHK_BOOL_RET_STATUS((aippNodes.size() == 0) || (aippNodes.size() == next_nodes_cnt), PARAM_INVALID,
  201. "Can not config part of outputs of Data node to support AIPP, config all "
  202. "of the outputs of Data to support AIPP, or config none of them");
  203. std::unique_ptr<domi::AippOpParams> aippParams(new (std::nothrow) domi::AippOpParams());
  204. GE_CHECK_NOTNULL(aippParams);
  205. GE_IF_BOOL_EXEC(
  206. aippNodes.size() > 1, for (decltype(aippNodes)::size_type i = 1; i < aippNodes.size(); i++) {
  207. std::unique_ptr<domi::AippOpParams> currAippParam(new (std::nothrow) domi::AippOpParams());
  208. GE_CHECK_NOTNULL(currAippParam);
  209. GE_CHK_STATUS(GetAippParams(currAippParam, aippNodes[i]));
  210. if (aippMode == domi::AippOpParams::static_) {
  211. GE_CHK_BOOL_RET_STATUS(aippParams->input_format() == currAippParam->input_format(), PARAM_INVALID,
  212. "The input_format of all aipp_ops after one Data should be the same");
  213. GE_CHK_BOOL_RET_STATUS(aippParams->src_image_size_w() == currAippParam->src_image_size_w(), PARAM_INVALID,
  214. "The src_image_size_w of all aipp_ops after one Data should be the same");
  215. GE_CHK_BOOL_RET_STATUS(aippParams->src_image_size_h() == currAippParam->src_image_size_h(), PARAM_INVALID,
  216. "The src_image_size_h of all aipp_ops after one Data should be the same");
  217. } else {
  218. GE_CHK_BOOL_RET_STATUS(aippParams->max_src_image_size() == currAippParam->max_src_image_size(),
  219. PARAM_INVALID,
  220. "The max_src_image_size of all aipp_ops after one Data should be the same");
  221. }
  222. });
  223. }
  224. return SUCCESS;
  225. }
  226. Status InsertNewOpUtil::GetAippParams(const std::unique_ptr<domi::AippOpParams> &aippParams, const NodePtr &aipp_node) {
  227. GE_CHECK_NOTNULL(aipp_node);
  228. ge::GeAttrValue::NAMED_ATTRS aipp_attr;
  229. const OpDescPtr tmpOpPtr = aipp_node->GetOpDesc();
  230. GE_CHECK_NOTNULL(tmpOpPtr);
  231. GE_CHK_BOOL_RET_STATUS(AttrUtils::GetNamedAttrs(tmpOpPtr, ATTR_NAME_AIPP, aipp_attr), FAILED,
  232. "Aipp node should contain param aipp!");
  233. GE_CHK_STATUS_RET(OpUtils::ConvertAippParams(aipp_attr, aippParams.get()), "get aipp params failed");
  234. return SUCCESS;
  235. }
  236. Status InsertNewOpUtil::AddFormatAgnosticAttrToSwitchn(const NodePtr &aipp_node) {
  237. GE_CHECK_NOTNULL(aipp_node);
  238. auto next_nodes = aipp_node->GetOutDataNodes();
  239. for (const auto next_node : next_nodes) {
  240. GE_CHECK_NOTNULL(next_node);
  241. auto op_desc = next_node->GetOpDesc();
  242. GE_CHECK_NOTNULL(op_desc);
  243. if (op_desc->GetType() == SWITCHN) {
  244. GELOGI("Find switchn node [%s] after aipp [%s]", op_desc->GetName().c_str(), aipp_node->GetName().c_str());
  245. (void)AttrUtils::SetInt(op_desc, "_format_agnostic", kFormatAgnosticSwitch);
  246. (void)AttrUtils::SetListInt(op_desc, "_format_agnostic_except_input",
  247. std::vector<int64_t>({kFormatDependInputIndex}));
  248. }
  249. }
  250. return SUCCESS;
  251. }
  252. Status InsertNewOpUtil::UpdateDataNodeByAipp(const ComputeGraphPtr &graph) {
  253. std::map<std::string, NodePtr> switchn_names_to_data;
  254. std::set<NodePtr> updated_switchn;
  255. NodePtr multbatch_case;
  256. for (auto &node : graph->GetDirectNode()) {
  257. if (node->GetType() == DATA) {
  258. std::string switchn_name;
  259. if (AttrUtils::GetStr(node->GetOpDesc(), kMbatchSwitchnName, switchn_name)) {
  260. switchn_names_to_data[switchn_name] = node;
  261. }
  262. }
  263. if (node->GetType() == AIPP) {
  264. GE_RETURN_IF_ERROR(UpdatePrevNodeByAipp(node, updated_switchn));
  265. // In dynamic batch/HW and dynamic aipp scend, switchn should be set format agnostic, otherwise transdata maybe
  266. // inserted between aipp and switchn which introduce performance and memory increase problem.
  267. GE_RETURN_IF_ERROR(AddFormatAgnosticAttrToSwitchn(node));
  268. }
  269. if (node->GetType() == CASE && node->GetOpDesc()->HasAttr(ATTR_NAME_BATCH_NUM)) {
  270. multbatch_case = node;
  271. }
  272. }
  273. for (auto &switchn : updated_switchn) {
  274. auto data_iter = switchn_names_to_data.find(switchn->GetName());
  275. if (data_iter == switchn_names_to_data.end()) {
  276. GELOGE(INTERNAL_ERROR, "Failed to find relative data node by switchn %s", switchn->GetName().c_str());
  277. return INTERNAL_ERROR;
  278. }
  279. GE_RETURN_IF_ERROR(UpdateDataBySwitchN(switchn, data_iter->second));
  280. }
  281. if (multbatch_case != nullptr) {
  282. GE_RETURN_IF_ERROR(UpdateCaseNode(graph, multbatch_case));
  283. }
  284. return SUCCESS;
  285. }
  286. Status InsertNewOpUtil::FindMaxSizeNode(const ComputeGraphPtr &graph, const NodePtr &case_node,
  287. map<uint32_t, int64_t> &max_sizes,
  288. map<uint32_t, GeTensorDescPtr> &aipp_inputs) {
  289. const auto &func_desc = case_node->GetOpDesc();
  290. for (const auto &name : func_desc->GetSubgraphInstanceNames()) {
  291. const auto &subgraph = graph->GetSubgraph(name);
  292. if (subgraph == nullptr) {
  293. GELOGE(GE_GRAPH_EMPTY_SUBGRAPH, "Subgraph not found, name: %s", name.c_str());
  294. return GE_GRAPH_EMPTY_SUBGRAPH;
  295. }
  296. std::set<NodePtr> updated_switchn; // fix interface
  297. for (auto &node : subgraph->GetDirectNode()) {
  298. if (node->GetType() == AIPP) {
  299. GE_RETURN_IF_ERROR(UpdatePrevNodeByAipp(node, updated_switchn));
  300. int64_t size = 0;
  301. auto in_data_anchor = node->GetInDataAnchor(0);
  302. GE_CHECK_NOTNULL(in_data_anchor);
  303. auto peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  304. GE_CHECK_NOTNULL(peer_out_anchor);
  305. const auto &src_node = peer_out_anchor->GetOwnerNode();
  306. const auto &src_op = src_node->GetOpDesc();
  307. GE_CHECK_NOTNULL(src_op);
  308. uint32_t parent_index = 0;
  309. if (!AttrUtils::GetInt(src_op, ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
  310. GELOGE(FAILED, "Parent index not found, name: %s", src_op->GetName().c_str());
  311. return FAILED;
  312. }
  313. auto aipp_op_desc = node->GetOpDesc();
  314. GE_CHECK_NOTNULL(aipp_op_desc);
  315. auto input = aipp_op_desc->MutableInputDesc(0);
  316. GE_CHECK_NOTNULL(input);
  317. if (TensorUtils::GetSize(*input, size) == GRAPH_SUCCESS) {
  318. if (max_sizes[parent_index] < size) {
  319. max_sizes[parent_index] = size;
  320. aipp_inputs[parent_index] = input;
  321. }
  322. }
  323. }
  324. }
  325. }
  326. return SUCCESS;
  327. }
  328. Status InsertNewOpUtil::UpdateCaseNode(const ComputeGraphPtr &graph, const NodePtr &case_node) {
  329. const auto &func_desc = case_node->GetOpDesc();
  330. map<uint32_t, int64_t> max_sizes;
  331. map<uint32_t, GeTensorDescPtr> aipp_inputs;
  332. GE_RETURN_IF_ERROR(FindMaxSizeNode(graph, case_node, max_sizes, aipp_inputs));
  333. for (const auto &item : aipp_inputs) {
  334. uint32_t parent_index = item.first;
  335. const GeTensorDescPtr &aipp_input = item.second;
  336. GE_CHECK_NOTNULL(aipp_input);
  337. const GeTensorDescPtr &input_desc = func_desc->MutableInputDesc(parent_index);
  338. GE_CHECK_NOTNULL(input_desc);
  339. input_desc->SetDataType(aipp_input->GetDataType());
  340. input_desc->SetOriginDataType(aipp_input->GetOriginDataType());
  341. input_desc->SetShape(aipp_input->GetShape());
  342. input_desc->SetOriginShape(aipp_input->GetShape());
  343. input_desc->SetFormat(aipp_input->GetFormat());
  344. input_desc->SetOriginFormat(aipp_input->GetFormat());
  345. ge::TensorUtils::SetSize(*input_desc, max_sizes[item.first]);
  346. const auto &in_anchor = case_node->GetInDataAnchor(parent_index);
  347. const auto &out_anchor = in_anchor->GetPeerOutAnchor();
  348. const auto &data = out_anchor->GetOwnerNode();
  349. auto data_opdesc = data->GetOpDesc();
  350. GE_CHECK_NOTNULL(data_opdesc);
  351. Format old_format = data_opdesc->MutableOutputDesc(0)->GetFormat();
  352. auto ret = data_opdesc->UpdateOutputDesc(0, *input_desc);
  353. if (ret != GRAPH_SUCCESS) {
  354. GELOGE(INTERNAL_ERROR, "Failed to update data %s output using case %s", data->GetName().c_str(),
  355. case_node->GetName().c_str());
  356. return INTERNAL_ERROR;
  357. }
  358. ret = data_opdesc->UpdateInputDesc(0, *input_desc);
  359. if (ret != GRAPH_SUCCESS) {
  360. GELOGE(INTERNAL_ERROR, "Failed to update data %s input using case %s", data->GetName().c_str(),
  361. case_node->GetName().c_str());
  362. return INTERNAL_ERROR;
  363. }
  364. // Update attr _mbatch_origin_input_dims for data when it is linked to aipp
  365. UpdateMultiBatchInputDims(data_opdesc, old_format);
  366. }
  367. return SUCCESS;
  368. }
  369. Status InsertNewOpUtil::UpdatePrevNodeByAipp(NodePtr &node, std::set<NodePtr> &switchns) {
  370. GELOGI("Start to update prev node size by aipp %s.", node->GetName().c_str());
  371. auto aipp_op_desc = node->GetOpDesc();
  372. GE_CHECK_NOTNULL(aipp_op_desc);
  373. auto aipp_input = aipp_op_desc->MutableInputDesc(0);
  374. GE_CHECK_NOTNULL(aipp_input);
  375. int64_t size = 0;
  376. graphStatus graph_ret = ge::TensorUtils::GetSize(*aipp_input, size);
  377. if (graph_ret != GRAPH_SUCCESS) {
  378. GELOGE(FAILED, "UpdateOutputDesc fail, graph_ret:%d", graph_ret);
  379. return FAILED;
  380. }
  381. GELOGI("Get input size [%ld] from aipp [%s].", size, aipp_op_desc->GetName().c_str());
  382. if (size == 0) {
  383. GELOGE(FAILED, "Can not get size from aipp [%s]", aipp_op_desc->GetName().c_str());
  384. return FAILED;
  385. }
  386. (void)AttrUtils::SetInt(aipp_input, ATTR_NAME_INPUT_ORIGIN_SIZE, size);
  387. auto in_data_anchor = node->GetInDataAnchor(0);
  388. GE_CHECK_NOTNULL(in_data_anchor);
  389. auto peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  390. GE_CHECK_NOTNULL(peer_out_anchor);
  391. const auto &src_node = peer_out_anchor->GetOwnerNode();
  392. const auto &src_op = src_node->GetOpDesc();
  393. GE_CHECK_NOTNULL(src_op);
  394. // if the type of src_node is SwitchN, the input of it may be updated to a size not the max one
  395. // the correct size will be updated in function `UpdateDataBySwitchN`
  396. DataType aipp_dt = aipp_input->GetDataType();
  397. aipp_input->SetOriginDataType(aipp_dt);
  398. DataType aipp_origni_dt = aipp_input->GetOriginDataType();
  399. GeShape aipp_shape = aipp_input->GetShape();
  400. Format aipp_format = aipp_input->GetFormat();
  401. GELOGI("Aipp [%s] input datatype is %s, origin datatype is %s, input shape is %s", aipp_op_desc->GetName().c_str(),
  402. TypeUtils::DataTypeToSerialString(aipp_dt).c_str(), TypeUtils::DataTypeToSerialString(aipp_origni_dt).c_str(),
  403. ge::formats::ShapeToString(aipp_shape.GetDims()).c_str());
  404. const GeTensorDescPtr &input = src_op->MutableInputDesc(0);
  405. GE_CHECK_NOTNULL(input);
  406. input->SetDataType(aipp_dt);
  407. input->SetOriginDataType(aipp_origni_dt);
  408. input->SetShape(aipp_shape);
  409. input->SetOriginShape(aipp_shape);
  410. input->SetFormat(aipp_format);
  411. input->SetOriginFormat(aipp_format);
  412. ge::TensorUtils::SetSize(*input, size);
  413. const GeTensorDescPtr &output = src_op->MutableOutputDesc(peer_out_anchor->GetIdx());
  414. GE_CHECK_NOTNULL(output);
  415. output->SetDataType(aipp_dt);
  416. output->SetOriginDataType(aipp_origni_dt);
  417. output->SetShape(aipp_shape);
  418. output->SetOriginShape(aipp_shape);
  419. output->SetFormat(aipp_format);
  420. output->SetOriginFormat(aipp_format);
  421. ge::TensorUtils::SetSize(*output, size);
  422. if (src_node->GetType() == SWITCHN) {
  423. switchns.insert(src_node);
  424. }
  425. GELOGI("Set node %s output %d size %ld by aipp.", src_node->GetName().c_str(), peer_out_anchor->GetIdx(), size);
  426. return SUCCESS;
  427. }
  428. Status InsertNewOpUtil::UpdateDataBySwitchN(const NodePtr &switchn, const NodePtr &data) {
  429. size_t max_index = switchn->GetOpDesc()->GetOutputsSize();
  430. int64_t max_size = 0;
  431. for (size_t i = 0; i < switchn->GetOpDesc()->GetOutputsSize(); ++i) {
  432. int64_t size = 0;
  433. auto output_desc = switchn->GetOpDesc()->MutableOutputDesc(i);
  434. GE_CHECK_NOTNULL(output_desc);
  435. if (TensorUtils::GetSize(*output_desc, size) == GRAPH_SUCCESS) {
  436. if (max_size < size) {
  437. max_size = size;
  438. max_index = i;
  439. }
  440. }
  441. }
  442. if (max_index >= switchn->GetOpDesc()->GetOutputsSize()) {
  443. GELOGE(INTERNAL_ERROR, "No max size found from switchn node %s", switchn->GetName().c_str());
  444. return INTERNAL_ERROR;
  445. }
  446. auto output_desc = switchn->GetOpDesc()->MutableOutputDesc(max_index);
  447. auto input_desc = switchn->GetOpDesc()->MutableInputDesc(0);
  448. GE_CHECK_NOTNULL(input_desc);
  449. input_desc->SetDataType(output_desc->GetDataType());
  450. input_desc->SetOriginDataType(output_desc->GetOriginDataType());
  451. input_desc->SetShape(output_desc->GetShape());
  452. input_desc->SetOriginShape(output_desc->GetOriginShape());
  453. input_desc->SetFormat(output_desc->GetFormat());
  454. input_desc->SetOriginFormat(output_desc->GetOriginFormat());
  455. TensorUtils::SetSize(*input_desc, max_size);
  456. auto data_opdesc = data->GetOpDesc();
  457. GE_CHECK_NOTNULL(data_opdesc);
  458. Format old_format = data_opdesc->MutableOutputDesc(0)->GetFormat();
  459. auto ret = data_opdesc->UpdateOutputDesc(0, *input_desc);
  460. if (ret != GRAPH_SUCCESS) {
  461. GELOGE(INTERNAL_ERROR, "Failed to update data %s output using switchn %s", data->GetName().c_str(),
  462. switchn->GetName().c_str());
  463. return INTERNAL_ERROR;
  464. }
  465. ret = data_opdesc->UpdateInputDesc(0, *input_desc);
  466. if (ret != GRAPH_SUCCESS) {
  467. GELOGE(INTERNAL_ERROR, "Failed to update data %s input using switchn %s", data->GetName().c_str(),
  468. switchn->GetName().c_str());
  469. return INTERNAL_ERROR;
  470. }
  471. // Update attr _mbatch_origin_input_dims for data when it is linked to aipp
  472. UpdateMultiBatchInputDims(data_opdesc, old_format);
  473. return SUCCESS;
  474. }
  475. void InsertNewOpUtil::UpdateMultiBatchInputDims(const OpDescPtr &data_opdesc, Format &old_format) {
  476. if (!data_opdesc->HasAttr(ATTR_MBATCH_ORIGIN_INPUT_DIMS)) {
  477. GELOGW("Failed to acquire _mbatch_origin_input_dims attr from node [%s]", data_opdesc->GetName().c_str());
  478. return;
  479. }
  480. auto new_data_dims = data_opdesc->GetOutputDesc(0).GetShape().GetDims();
  481. vector<int64_t> origin_input_dims;
  482. (void)AttrUtils::GetListInt(data_opdesc, ATTR_MBATCH_ORIGIN_INPUT_DIMS, origin_input_dims);
  483. // Convert origin_input_dims to NHWC because data format is set to NHWC when it is linked to aipp.
  484. ConvertShape2Nhwc(old_format, origin_input_dims);
  485. if (new_data_dims.size() != origin_input_dims.size()) {
  486. return;
  487. }
  488. for (size_t i = 0; i < origin_input_dims.size(); ++i) {
  489. // Need to update shape when aipp has crop function because H,W is different, ignore -1.
  490. if (origin_input_dims[i] > 0) {
  491. origin_input_dims[i] = new_data_dims[i];
  492. }
  493. }
  494. (void)AttrUtils::SetListInt(data_opdesc, ATTR_MBATCH_ORIGIN_INPUT_DIMS, origin_input_dims);
  495. return;
  496. }
  497. Status InsertNewOpUtil::GetDataRelatedNode(NodePtr &node, std::map<NodePtr, std::set<NodePtr>> &data_next_node_map) {
  498. GELOGI("Start to get data and next node %s.", node->GetName().c_str());
  499. OpDescPtr data_op = node->GetOpDesc();
  500. GE_CHECK_NOTNULL(data_op);
  501. if (!data_op->HasAttr(ATTR_NAME_AIPP)) {
  502. GELOGI("there is not AIPP info for Data: %s.", data_op->GetName().c_str());
  503. return SUCCESS;
  504. }
  505. std::unique_ptr<domi::AippOpParams> aipp_params(new (std::nothrow) domi::AippOpParams());
  506. ge::GeAttrValue::NAMED_ATTRS aipp_attr;
  507. GE_CHK_BOOL_RET_STATUS(AttrUtils::GetNamedAttrs(data_op, ATTR_NAME_AIPP, aipp_attr), GE_AIPP_NOT_EXIST,
  508. "Data node do not contain param aipp!");
  509. GE_CHK_STATUS_RET(OpUtils::ConvertAippParams(aipp_attr, aipp_params.get()), "get aipp params failed");
  510. for (auto out_data_anchor : node->GetAllOutDataAnchors()) {
  511. GE_CHECK_NOTNULL(out_data_anchor);
  512. auto peer_in_anchors = out_data_anchor->GetPeerInDataAnchors();
  513. for (auto peer_in_data_anchor : peer_in_anchors) {
  514. GE_CHECK_NOTNULL(peer_in_data_anchor);
  515. const auto &dst_node = peer_in_data_anchor->GetOwnerNode();
  516. const auto &dst_op = dst_node->GetOpDesc();
  517. GE_CHECK_NOTNULL(dst_op);
  518. if (dst_op->GetType() == AIPP || dst_op->GetType() == SWITCHN || dst_op->GetType() == CASE) {
  519. auto data_iter = data_next_node_map.find(node);
  520. if (data_iter == data_next_node_map.end()) {
  521. std::set<NodePtr> next_node_set;
  522. next_node_set.insert(dst_node);
  523. data_next_node_map[node] = next_node_set;
  524. } else {
  525. if (data_next_node_map[node].find(dst_node) == data_next_node_map[node].end()) {
  526. data_next_node_map[node].insert(dst_node);
  527. }
  528. }
  529. }
  530. }
  531. }
  532. return SUCCESS;
  533. }
  534. Status InsertNewOpUtil::GetAllAipps(const NodePtr &data_node, const NodePtr &node, std::vector<NodePtr> &aipps) {
  535. GE_CHECK_NOTNULL(node);
  536. OpDescPtr op = node->GetOpDesc();
  537. GE_CHECK_NOTNULL(op);
  538. GELOGI("Get all aipp node from this node %s.", op->GetName().c_str());
  539. if (op->GetType() == AIPP) {
  540. aipps.emplace_back(node);
  541. } else if (op->GetType() == SWITCHN) {
  542. for (auto out_data_anchor : node->GetAllOutDataAnchors()) {
  543. GE_CHECK_NOTNULL(out_data_anchor);
  544. auto peer_in_anchors = out_data_anchor->GetPeerInDataAnchors();
  545. if (peer_in_anchors.size() > 0) {
  546. auto peer_in_anchor = peer_in_anchors.at(0);
  547. GE_CHECK_NOTNULL(peer_in_anchor);
  548. auto dst_aipp_node = peer_in_anchor->GetOwnerNode();
  549. if (dst_aipp_node->GetType() == AIPP) {
  550. aipps.emplace_back(dst_aipp_node);
  551. }
  552. }
  553. }
  554. } else if (op->GetType() == CASE) {
  555. const ComputeGraphPtr &graph = node->GetOwnerComputeGraph();
  556. for (const auto &name : op->GetSubgraphInstanceNames()) {
  557. const auto &subgraph = graph->GetSubgraph(name);
  558. if (subgraph == nullptr) {
  559. GELOGE(GE_GRAPH_EMPTY_SUBGRAPH, "Subgraph not found, name: %s", name.c_str());
  560. return GE_GRAPH_EMPTY_SUBGRAPH;
  561. }
  562. for (auto &subgraph_node : subgraph->GetDirectNode()) {
  563. if (subgraph_node->GetType() == AIPP) {
  564. auto src_node = subgraph_node->GetInDataNodes().at(0);
  565. const auto &src_op = src_node->GetOpDesc();
  566. GE_CHECK_NOTNULL(src_op);
  567. uint32_t parent_index = 0;
  568. if (!AttrUtils::GetInt(src_op, ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
  569. GELOGE(FAILED, "Parent index not found, name: %s", src_op->GetName().c_str());
  570. return FAILED;
  571. }
  572. auto data = node->GetInDataNodes().at(parent_index);
  573. if (data->GetName() == data_node->GetName()) {
  574. aipps.emplace_back(subgraph_node);
  575. }
  576. }
  577. }
  578. }
  579. }
  580. return SUCCESS;
  581. }
  582. Status InsertNewOpUtil::RecordAIPPInfoToData(const ComputeGraphPtr &graph) {
  583. GELOGI("Start to record aipp info to Data.");
  584. std::map<NodePtr, std::set<NodePtr> data_next_node_map;
  585. for (auto &node : graph->GetDirectNode()) {
  586. if (node->GetType() == DATA) {
  587. GE_RETURN_IF_ERROR(GetDataRelatedNode(node, data_next_node_map));
  588. }
  589. }
  590. for (auto it : data_next_node_map) {
  591. std::vector<std::string> input_dims;
  592. std::vector<std::string> output_dims;
  593. auto data_node = it.first;
  594. auto data_op_desc = data_node->GetOpDesc();
  595. GE_CHECK_NOTNULL(data_op_desc);
  596. std::set<NodePtr> aipps_or_switchs_or_case = it.second;
  597. if (aipps_or_switchs_or_case.size() != 1) {
  598. GELOGW("The number of successors swith or aipp of data is more than 1");
  599. continue;
  600. }
  601. std::vector<NodePtr> aipps;
  602. GE_RETURN_IF_ERROR(GetAllAipps(data_node, *aipps_or_switchs_or_case.begin(), aipps));
  603. GELOGI("RecordAIPPInfoToData: Data: name[%s], type[%s], batch size[%u]", data_node->GetName().c_str(),
  604. data_node->GetType().c_str(), aipps.size());
  605. for (auto aipp_it : aipps) {
  606. string input;
  607. string output;
  608. GetInputOutputInfo(data_node, aipp_it, input, output);
  609. input_dims.emplace_back(input);
  610. output_dims.emplace_back(output);
  611. // When static aipp is set, need to get the model input dims which processed by aipp
  612. GE_RETURN_IF_ERROR(SetModelInputDims(data_node, aipp_it));
  613. }
  614. // if _all_origin_gears_inputs is set, use its value directly.
  615. if (data_op_desc->HasAttr("_all_origin_gears_inputs")) {
  616. std::vector<std::string> input_dims_str;
  617. (void)AttrUtils::GetListStr(data_op_desc, "_all_origin_gears_inputs", input_dims_str);
  618. (void)AttrUtils::SetListStr(data_op_desc, ATTR_NAME_AIPP_INPUTS, input_dims_str);
  619. if ((input_dims_str.size() > output_dims.size()) && !output_dims.empty()) {
  620. // make sure output and input counts is equal, appears in dynamic aipp and dynamic shape/batch scene.
  621. std::vector<std::string> output_dims_str{input_dims_str.size(), output_dims[0]};
  622. (void)AttrUtils::SetListStr(data_op_desc, ATTR_NAME_AIPP_OUTPUTS, output_dims_str);
  623. } else {
  624. (void)AttrUtils::SetListStr(data_op_desc, ATTR_NAME_AIPP_OUTPUTS, output_dims);
  625. }
  626. } else {
  627. (void)AttrUtils::SetListStr(data_op_desc, ATTR_NAME_AIPP_INPUTS, input_dims);
  628. (void)AttrUtils::SetListStr(data_op_desc, ATTR_NAME_AIPP_OUTPUTS, output_dims);
  629. }
  630. }
  631. return SUCCESS;
  632. }
  633. Status InsertNewOpUtil::GetInputOutputInfo(NodePtr &data_node, NodePtr &aipp_node, std::string &input,
  634. std::string &output) {
  635. GE_CHECK_NOTNULL(data_node);
  636. GE_CHECK_NOTNULL(aipp_node);
  637. OpDescPtr data_op = data_node->GetOpDesc();
  638. GE_CHECK_NOTNULL(data_op);
  639. OpDescPtr aipp_op = aipp_node->GetOpDesc();
  640. GE_CHECK_NOTNULL(aipp_op);
  641. // aipp node's original output shape equals to original model data's shape
  642. ConstGeTensorDescPtr output_desc = aipp_op->GetOutputDescPtr(0);
  643. Format orig_format = output_desc->GetOriginFormat();
  644. DataType orig_data_type = output_desc->GetOriginDataType();
  645. std::string tensor_name = data_op->GetName();
  646. size_t dim_num = output_desc->GetOriginShape().GetDimNum();
  647. int64_t tensor_size = 0;
  648. (void)TensorUtils::CalcTensorMemSize(output_desc->GetOriginShape(), orig_format, orig_data_type, tensor_size);
  649. int64_t input_size = tensor_size;
  650. input = TypeUtils::FormatToSerialString(orig_format) + ":" + TypeUtils::DataTypeToSerialString(orig_data_type) + ":" +
  651. tensor_name + ":" + std::to_string(input_size) + ":" + std::to_string(dim_num) + ":" +
  652. formats::JoinToString(output_desc->GetOriginShape().GetDims());
  653. Format format = output_desc->GetFormat();
  654. DataType data_type = output_desc->GetDataType();
  655. std::string output_name = aipp_op->GetOutputNameByIndex(0);
  656. size_t output_dim_num = output_desc->GetShape().GetDimNum();
  657. (void)TensorUtils::CalcTensorMemSize(output_desc->GetShape(), output_desc->GetFormat(), output_desc->GetDataType(),
  658. tensor_size);
  659. int64_t output_size = tensor_size;
  660. output = TypeUtils::FormatToSerialString(format) + ":" + TypeUtils::DataTypeToSerialString(data_type) + ":" +
  661. output_name + ":" + std::to_string(output_size) + ":" + std::to_string(output_dim_num) + ":" +
  662. formats::JoinToString(output_desc->GetShape().GetDims());
  663. GELOGI("GetInputOutputInfo: get data[%s] node related aipp[%s] node info, input[%s], output[%s].",
  664. data_node->GetName().c_str(), aipp_node->GetName().c_str(), input.c_str(), output.c_str());
  665. return SUCCESS;
  666. }
  667. Status InsertNewOpUtil::SetModelInputDims(NodePtr &data_node, NodePtr &aipp_node) {
  668. GE_CHECK_NOTNULL(data_node);
  669. GE_CHECK_NOTNULL(aipp_node);
  670. OpDescPtr data_opdesc = data_node->GetOpDesc();
  671. GE_CHECK_NOTNULL(data_opdesc);
  672. OpDescPtr aipp_opdesc = aipp_node->GetOpDesc();
  673. GE_CHECK_NOTNULL(aipp_opdesc);
  674. // In dynamic bacth/hw scenario, the new model input dims only need be set once
  675. if (data_node->GetOpDesc()->HasAttr(ATTR_NAME_INPUT_DIMS)) {
  676. GELOGD("Data %s already has attribute %s", data_node->GetOpDesc()->GetName().c_str(), ATTR_NAME_INPUT_DIMS.c_str());
  677. return SUCCESS;
  678. }
  679. vector<int64_t> model_input_dims;
  680. vector<int64_t> origin_input_dims;
  681. if (AttrUtils::GetListInt(aipp_opdesc, ATTR_NAME_INPUT_DIMS, model_input_dims) && !model_input_dims.empty()) {
  682. // When dynamic bacth/hw is set, N or HW need to be set to -1
  683. if (AttrUtils::GetListInt(data_opdesc, ATTR_MBATCH_ORIGIN_INPUT_DIMS, origin_input_dims) &&
  684. !origin_input_dims.empty()) {
  685. GELOGI("In dynamic bacth/hw scenario, N or HW need to be set to -1. model_input_dims: %s, origin_input_dims: %s",
  686. formats::JoinToString(model_input_dims).c_str(), formats::JoinToString(origin_input_dims).c_str());
  687. for (size_t i = 0; i < origin_input_dims.size(); ++i) {
  688. // N or HW need to be set to -1
  689. if (origin_input_dims[i] < 0) {
  690. model_input_dims[i] = origin_input_dims[i];
  691. }
  692. }
  693. }
  694. GELOGD("After set N or H/W to -1, the model input dims: %s.", formats::JoinToString(model_input_dims).c_str());
  695. if (!AttrUtils::SetListInt(data_opdesc, ATTR_NAME_INPUT_DIMS, model_input_dims)) {
  696. GELOGE(FAILED, "SetListInt of %s failed.", ATTR_NAME_INPUT_DIMS.c_str());
  697. return FAILED;
  698. }
  699. }
  700. return SUCCESS;
  701. }
  702. } // 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