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.

multi_batch_options.cc 26 kB

4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605
  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 "multi_batch_options.h"
  17. #include "framework/common/debug/ge_log.h"
  18. #include "framework/omg/omg_inner_types.h"
  19. #include "framework/common/util.h"
  20. #include "framework/common/string_util.h"
  21. #include "common/formats/utils/formats_trans_utils.h"
  22. #include "common/util/error_manager/error_manager.h"
  23. #include "graph/debug/ge_attr_define.h"
  24. #include "graph/utils/node_utils.h"
  25. #include "graph/ge_context.h"
  26. #include "graph/common/local_context.h"
  27. #include "framework/common/types.h"
  28. #include "graph/compute_graph.h"
  29. #include "graph/utils/graph_utils.h"
  30. #include "graph/common/omg_util.h"
  31. namespace ge {
  32. namespace multibatch {
  33. constexpr int kDecimal = 10;
  34. constexpr uint8_t kMaxShapesCount = 100;
  35. constexpr uint8_t kMinShapesCount = 2;
  36. const int kDynmaicDims = -1;
  37. const int kDynamicBatchDynamicDimsNum = 1;
  38. const int kDynamicImgSizeDynamciDimsNum = 2;
  39. const size_t kMaxNDDimNum = 4;
  40. const size_t kMinNDDimNum = 1;
  41. const size_t kNumOfGetnextNode = 1;
  42. const int kDivisionConst = 2;
  43. const char *const kSubstrOfGetNextNosinkName = "IteratorGetNext";
  44. const char *const kShapeDataName = "ascend_mbatch_shape_data";
  45. const char *const kGetNextName = "IteratorV2";
  46. const char *const kExtAttrDataNodes = "data_nodes";
  47. const char *const kExtAttrGetNextNoSink = "getnext_no_sink";
  48. inline bool IsGetNextType(const NodePtr &node) {
  49. std::string original_type;
  50. GE_IF_BOOL_EXEC(GetOriginalType(node, original_type) != SUCCESS,
  51. GELOGW("Get original type failed."); return false);
  52. return (original_type == kGetNextName);
  53. }
  54. void ParseDynamicSize(string dynamic_size, vector<vector<int64_t>> &shapes) {
  55. std::vector<std::string> shape_strs = ge::StringUtils::Split(dynamic_size, ';');
  56. for (const auto &shape_str : shape_strs) {
  57. if (shape_str.empty()) {
  58. continue;
  59. }
  60. std::vector<int64_t> shape;
  61. std::vector<std::string> dims = ge::StringUtils::Split(shape_str, ',');
  62. for (const auto &dim : dims) {
  63. if (dim.empty()) {
  64. continue;
  65. }
  66. shape.emplace_back(std::strtol(dim.c_str(), nullptr, kDecimal));
  67. }
  68. if (!shape.empty()) {
  69. shapes.emplace_back(shape);
  70. }
  71. }
  72. }
  73. Status DistinguishGetNextAndData(ComputeGraphPtr &graph, vector<NodePtr> &data_nodes,
  74. vector<NodePtr> &getnext_nosink_nodes, vector<NodePtr> &getnext_sink_nodes) {
  75. GELOGD("Start distinguish getnext and data node.");
  76. for (NodePtr &input_node : graph->GetDirectNode()) {
  77. GE_CHECK_NOTNULL(input_node);
  78. OpDescPtr op_desc = input_node->GetOpDesc();
  79. GE_CHECK_NOTNULL(op_desc);
  80. if (op_desc->GetType() == DATA && op_desc->GetName() != kShapeDataName) {
  81. if (op_desc->GetName().find(kSubstrOfGetNextNosinkName) == string::npos) {
  82. data_nodes.emplace_back(input_node);
  83. GELOGD("Name of data node is %s.", op_desc->GetName().c_str());
  84. } else {
  85. getnext_nosink_nodes.emplace_back(input_node);
  86. GELOGD("Name of getnext nosink is %s.", op_desc->GetName().c_str());
  87. }
  88. }
  89. if (IsGetNextType(input_node)) {
  90. GELOGD("Name of getnext sink is %s.", op_desc->GetName().c_str());
  91. getnext_sink_nodes.emplace_back(input_node);
  92. }
  93. }
  94. GELOGI("Data count is %zu, getnext nosink count is %zu, getnext sink count is %zu.", data_nodes.size(),
  95. getnext_nosink_nodes.size(), getnext_sink_nodes.size());
  96. GE_IF_BOOL_EXEC(!graph->SetExtAttr(kExtAttrDataNodes, data_nodes), GELOGW("Set data nodes attr failed.");)
  97. GE_IF_BOOL_EXEC(!graph->SetExtAttr(kExtAttrGetNextNoSink, getnext_nosink_nodes),
  98. GELOGW("Set getnext nosink nodes attr failed.");)
  99. return SUCCESS;
  100. }
  101. Status CheckSequenceOfData(ComputeGraphPtr &graph, const vector<NodePtr> &data_nodes) {
  102. GELOGD("Start check input sequence from data nodes and input shape.");
  103. if (data_nodes.size() != GetLocalOmgContext().user_input_dims.size()) {
  104. GELOGE(PARAM_INVALID, "The count of input shape:%zu should be equal to the count of data num:%zu.",
  105. GetLocalOmgContext().user_input_dims.size(), data_nodes.size());
  106. return PARAM_INVALID;
  107. }
  108. for (size_t i = 0; i < data_nodes.size(); ++i) {
  109. auto data_node = data_nodes.at(i);
  110. GE_CHECK_NOTNULL(data_node);
  111. GE_CHECK_NOTNULL(data_node->GetOpDesc());
  112. auto output_shape = data_node->GetOpDesc()->GetOutputDesc(0).GetShape().GetDims();
  113. auto dynamic_dims = GetLocalOmgContext().user_input_dims.at(i).second;
  114. GELOGD("The %zu data node is %s, node shape is %s, dynamic dim is %s.", i, data_node->GetName().c_str(),
  115. formats::JoinToString(output_shape).c_str(), formats::JoinToString(dynamic_dims).c_str());
  116. if (output_shape.empty() && dynamic_dims.size() == 1 && dynamic_dims.at(0) == 0) {
  117. GELOGI("No need to check sequence for constant.");
  118. continue;
  119. }
  120. if (dynamic_dims.size() != output_shape.size()) {
  121. GELOGE(PARAM_INVALID, "The output shape of %s is %s, the input shape from options of %s is %s.",
  122. data_node->GetName().c_str(), formats::JoinToString(output_shape).c_str(),
  123. GetLocalOmgContext().user_input_dims.at(i).first.c_str(), formats::JoinToString(dynamic_dims).c_str());
  124. return PARAM_INVALID;
  125. }
  126. for (size_t j = 0; j < dynamic_dims.size(); ++j) {
  127. if (dynamic_dims.at(j) != kDynmaicDims && dynamic_dims.at(j) != output_shape.at(j)) {
  128. GELOGE(INTERNAL_ERROR, "Value of input shape %s should be equal to %s.",
  129. formats::JoinToString(dynamic_dims).c_str(), formats::JoinToString(output_shape).c_str());
  130. return INTERNAL_ERROR;
  131. }
  132. }
  133. }
  134. return SUCCESS;
  135. }
  136. Status CheckSequenceOfGetnext(ComputeGraphPtr &graph, const vector<NodePtr> &getnext_sink_node) {
  137. GELOGD("Start check input sequence from getnext sink nodes and input shape.");
  138. if (getnext_sink_node.size() != kNumOfGetnextNode) {
  139. GELOGE(PARAM_INVALID, "Not support dynamic dims when a graph with multi getnext nodes.");
  140. return PARAM_INVALID;
  141. }
  142. auto data_node = getnext_sink_node.at(0);
  143. GE_CHECK_NOTNULL(data_node);
  144. auto op_desc = data_node->GetOpDesc();
  145. GE_CHECK_NOTNULL(op_desc);
  146. size_t data_count = data_node->GetAllOutDataAnchors().size() / kDivisionConst;
  147. if (data_count != GetLocalOmgContext().user_input_dims.size()) {
  148. GELOGE(PARAM_INVALID, "Output count of %s is %zu, should be equal to count of input shape: %zu",
  149. op_desc->GetName().c_str(), data_count, GetLocalOmgContext().user_input_dims.size());
  150. return PARAM_INVALID;
  151. }
  152. for (size_t i = 0; i < data_count; ++i) {
  153. auto output_shape = data_node->GetOpDesc()->GetOutputDesc(i).GetShape().GetDims();
  154. auto dynamic_dims = GetLocalOmgContext().user_input_dims.at(i).second;
  155. GELOGD("The %zu getnext node is %s, node shape is %s, dynamic dim is %s.", i, data_node->GetName().c_str(),
  156. formats::JoinToString(output_shape).c_str(), formats::JoinToString(dynamic_dims).c_str());
  157. if (output_shape.empty() && dynamic_dims.size() == 1 && dynamic_dims.at(0) == 0) {
  158. GELOGI("No need to check sequence for constant.");
  159. continue;
  160. }
  161. if (dynamic_dims.size() != output_shape.size()) {
  162. GELOGE(PARAM_INVALID, "the output_shape of %s is %s, the input_shape from options of %s is %s.",
  163. data_node->GetName().c_str(), formats::JoinToString(output_shape).c_str(),
  164. GetLocalOmgContext().user_input_dims.at(i).first.c_str(), formats::JoinToString(dynamic_dims).c_str());
  165. return PARAM_INVALID;
  166. }
  167. for (size_t j = 0; j < dynamic_dims.size(); ++j) {
  168. if (dynamic_dims.at(j) != kDynmaicDims && dynamic_dims.at(j) != output_shape.at(j)) {
  169. GELOGE(INTERNAL_ERROR, "value of input_shape %s should be equal to %s.",
  170. formats::JoinToString(dynamic_dims).c_str(), formats::JoinToString(output_shape).c_str());
  171. return INTERNAL_ERROR;
  172. }
  173. }
  174. }
  175. return SUCCESS;
  176. }
  177. Status CheckSequenceOfOptions(ComputeGraphPtr &graph, vector<NodePtr> &data_nodes,
  178. vector<NodePtr> &getnext_nosink_nodes, vector<NodePtr> &getnext_sink_nodes) {
  179. if (GetLocalOmgContext().dynamic_node_type.empty()) {
  180. GELOGI("No need to CheckSequenceOfOptions.");
  181. return SUCCESS;
  182. }
  183. if (DistinguishGetNextAndData(graph, data_nodes, getnext_nosink_nodes, getnext_sink_nodes) != SUCCESS) {
  184. GELOGE(PARAM_INVALID, "DistinguishGetNextAndData failed.");
  185. return PARAM_INVALID;
  186. }
  187. if (GetLocalOmgContext().dynamic_node_type == DATA) {
  188. GELOGD("Users want data nodes to be dynamic.");
  189. if (CheckSequenceOfData(graph, data_nodes) != SUCCESS) {
  190. GELOGE(PARAM_INVALID, "Failed to check sequence of data nodes.");
  191. return PARAM_INVALID;
  192. }
  193. } else {
  194. GELOGD("Users want getnext nodes to be dynamic.");
  195. if (!getnext_nosink_nodes.empty()) {
  196. if (CheckSequenceOfData(graph, getnext_nosink_nodes) != SUCCESS) {
  197. GELOGE(PARAM_INVALID, "Failed to check sequence of getnext nosink nodes.");
  198. return PARAM_INVALID;
  199. }
  200. } else {
  201. if (CheckSequenceOfGetnext(graph, getnext_sink_nodes) != SUCCESS) {
  202. GELOGE(PARAM_INVALID, "Failed to check sequence of getnext sink nodes.");
  203. return PARAM_INVALID;
  204. }
  205. }
  206. }
  207. return SUCCESS;
  208. }
  209. Status UpdateNameOfData(ComputeGraphPtr &graph, const vector<NodePtr> &data_nodes) {
  210. GELOGD("Update first value of input shape by data nodes.");
  211. if (data_nodes.size() != GetLocalOmgContext().user_input_dims.size()) {
  212. GELOGE(PARAM_INVALID, "count of data_nodes: %zu should be equal to input_shape count: %zu.",
  213. data_nodes.size(), GetLocalOmgContext().user_input_dims.size());
  214. return PARAM_INVALID;
  215. }
  216. for (size_t i = 0; i < data_nodes.size(); ++i) {
  217. GELOGD("The %zu data name is %s.", i, data_nodes.at(i)->GetOpDesc()->GetName().c_str());
  218. GetLocalOmgContext().user_input_dims.at(i).first = data_nodes.at(i)->GetOpDesc()->GetName();
  219. }
  220. return SUCCESS;
  221. }
  222. Status UpdateNameOfGetnext(ComputeGraphPtr &graph, const vector<NodePtr> &getnext_sink_nodes) {
  223. GELOGD("Update first value of input shape by getnext sink nodes.");
  224. if (getnext_sink_nodes.size() != kNumOfGetnextNode) {
  225. GELOGE(PARAM_INVALID, "Not support dynamic dims when a graph with multi getnext nodes.");
  226. return PARAM_INVALID;
  227. }
  228. auto input_node = getnext_sink_nodes.at(0);
  229. GE_CHECK_NOTNULL(input_node);
  230. auto op_desc = input_node->GetOpDesc();
  231. GE_CHECK_NOTNULL(op_desc);
  232. // user want getnext dynamic, just getnext or data+getnext_sink
  233. size_t data_count = input_node->GetAllOutDataAnchors().size() / kDivisionConst;
  234. if (data_count != GetLocalOmgContext().user_input_dims.size()) {
  235. GELOGE(PARAM_INVALID, "Output count of %s is %zu, should be equal to count of input shape: %zu",
  236. op_desc->GetName().c_str(), data_count, GetLocalOmgContext().user_input_dims.size());
  237. return PARAM_INVALID;
  238. }
  239. for (size_t i = 0; i < data_count; ++i) {
  240. string data_name = op_desc->GetName() + "_" + std::to_string(i);
  241. GELOGD("Data just from getnext sink is %s.", data_name.c_str());
  242. GetLocalOmgContext().user_input_dims.at(i).first = data_name;
  243. }
  244. return SUCCESS;
  245. }
  246. // need to distinguish online and offline, offline no need to update the name of input_shape
  247. Status UpdateNameOfInputShape(ComputeGraphPtr &graph, const vector<NodePtr> &data_nodes,
  248. const vector<NodePtr> &getnext_nosink_nodes, const vector<NodePtr> &getnext_sink_nodes) {
  249. if (GetLocalOmgContext().dynamic_node_type.empty()) {
  250. GELOGI("No need to update first value of input shape when offline infer.");
  251. return SUCCESS;
  252. }
  253. if (GetLocalOmgContext().dynamic_node_type == DATA) {
  254. GELOGD("Users want data nodes to be dynamic.");
  255. if (UpdateNameOfData(graph, data_nodes) != SUCCESS) {
  256. GELOGE(PARAM_INVALID, "Failed to update first value of input shape of data nodes.");
  257. return PARAM_INVALID;
  258. }
  259. } else {
  260. GELOGD("Users want getnext nodes to be dynamic.");
  261. if (!getnext_nosink_nodes.empty()) {
  262. if (UpdateNameOfData(graph, getnext_nosink_nodes) != SUCCESS) {
  263. GELOGE(PARAM_INVALID, "Failed to update first value of input shape of getnext nosink nodes.");
  264. return PARAM_INVALID;
  265. }
  266. } else {
  267. if (UpdateNameOfGetnext(graph, getnext_sink_nodes) != SUCCESS) {
  268. GELOGE(PARAM_INVALID, "Failed to update first value of input shape of getnext sink nodes.");
  269. return PARAM_INVALID;
  270. }
  271. }
  272. }
  273. return SUCCESS;
  274. }
  275. Status DeleteIdentityInsertByAdapter(ComputeGraphPtr &graph) {
  276. GELOGD("Start delete identity node inserted by adapter.");
  277. for (NodePtr &node : graph->GetDirectNode()) {
  278. GE_CHECK_NOTNULL(node);
  279. OpDescPtr op_desc = node->GetOpDesc();
  280. GE_CHECK_NOTNULL(op_desc);
  281. if (IsGetNextType(node)) {
  282. for (auto &out_data_anchor : node->GetAllOutDataAnchors()) {
  283. GE_IF_BOOL_EXEC(out_data_anchor == nullptr, continue);
  284. for (auto &peer_in_anchor : out_data_anchor->GetPeerInDataAnchors()) {
  285. GE_IF_BOOL_EXEC(peer_in_anchor == nullptr, continue);
  286. auto dst_node = peer_in_anchor->GetOwnerNode();
  287. GE_IF_BOOL_EXEC(dst_node == nullptr, continue);
  288. if (dst_node->GetType() == IDENTITY) {
  289. GELOGI("Need to remove %s.", dst_node->GetName().c_str());
  290. if (ge::GraphUtils::RemoveNodeWithoutRelink(graph, dst_node) != GRAPH_SUCCESS) {
  291. GELOGE(FAILED, "Remove Identity node %s failed.", dst_node->GetName().c_str());
  292. return FAILED;
  293. }
  294. }
  295. }
  296. }
  297. }
  298. }
  299. return SUCCESS;
  300. }
  301. Status CheckNegativeCountOfOptions(const std::vector<std::vector<int64_t>> &shapes) {
  302. if (!GetLocalOmgContext().dynamic_dims.empty()) {
  303. size_t negative_count = 0;
  304. for (size_t i = 0; i < GetLocalOmgContext().user_input_dims.size(); ++i) {
  305. for (size_t j = 0; j < GetLocalOmgContext().user_input_dims.at(i).second.size(); ++j) {
  306. if (GetLocalOmgContext().user_input_dims.at(i).second.at(j) == kDynmaicDims) {
  307. negative_count++;
  308. }
  309. }
  310. }
  311. for (size_t i = 0; i < shapes.size(); ++i) {
  312. if (shapes.at(i).size() != negative_count) {
  313. GELOGE(PARAM_INVALID, "Each gear num of dynamic_dims is %zu should be equal to %zu.", shapes.at(i).size(),
  314. negative_count);
  315. return PARAM_INVALID;
  316. }
  317. }
  318. }
  319. return SUCCESS;
  320. }
  321. ///
  322. /// @ingroup ge
  323. /// @brief Init Dynamic Param from Options.
  324. /// @param [out] std::vector<std::vector<int64_t>> &shapes: Result for Params.
  325. /// @return true: Configed for Multi batch / false: Not configed for Multi batch.
  326. ///
  327. bool InitDynamicParams(vector<vector<int64_t>> &shapes) {
  328. if (!GetLocalOmgContext().dynamic_batch_size.empty()) {
  329. GELOGD("Found dynamic batch option, value %s", GetLocalOmgContext().dynamic_batch_size.c_str());
  330. std::vector<std::string> dims = ge::StringUtils::Split(GetLocalOmgContext().dynamic_batch_size, ',');
  331. for (const auto &dim : dims) {
  332. if (dim.empty()) {
  333. continue;
  334. }
  335. shapes.emplace_back(std::vector<int64_t>({std::strtol(dim.c_str(), nullptr, kDecimal)}));
  336. GELOGI("Found dynamic batch, shape %s", formats::JoinToString(*shapes.rbegin()).c_str());
  337. }
  338. }
  339. if (!GetLocalOmgContext().dynamic_image_size.empty()) {
  340. GELOGD("Found dynamic image size option, value %s", GetLocalOmgContext().dynamic_image_size.c_str());
  341. ParseDynamicSize(GetLocalOmgContext().dynamic_image_size, shapes);
  342. for (const auto &shape : shapes) {
  343. GELOGI("Found dynamic image size, shape %s", formats::JoinToString(shape).c_str());
  344. }
  345. }
  346. if (!GetLocalOmgContext().dynamic_dims.empty()) {
  347. GELOGD("Found dynamic dims option, value %s", GetLocalOmgContext().dynamic_dims.c_str());
  348. ParseDynamicSize(GetLocalOmgContext().dynamic_dims, shapes);
  349. for (const auto &shape : shapes) {
  350. GELOGI("Found dynamic dims, shape %s", formats::JoinToString(shape).c_str());
  351. }
  352. }
  353. return !shapes.empty();
  354. }
  355. ///
  356. /// @ingroup ge
  357. /// @brief parse each data's own dynamic dims.
  358. /// @param [out] map<string, vector<vector<int64_t>>> &data_to_dynamic_info: key:data_name. value:dynamic dims.
  359. /// @return true: Configed for Multi batch / false: Not configed for Multi batch.
  360. ///
  361. Status ParserDataToDynmaicInfo(const vector<vector<int64_t>> &shapes,
  362. vector<pair<string, vector<int64_t>>> &data_name_and_shape,
  363. map<string, vector<vector<int64_t>> > &data_to_dynamic_info) {
  364. size_t cur_data_index = 0;
  365. for (size_t index = 0; index < data_name_and_shape.size(); ++index) {
  366. auto &cur_item = data_name_and_shape[index];
  367. auto &data_name = cur_item.first;
  368. auto &data_shape = cur_item.second;
  369. auto dynamic_dims_num = std::count_if(data_shape.begin(), data_shape.end(),
  370. [&data_shape](int64_t dim){ return dim < 0; });
  371. GELOGI("Train_Dynamic dynamic_dims_num of %s is %zu", data_name.c_str(), dynamic_dims_num);
  372. vector<vector<int64_t> > dynamic_info;
  373. for (auto &dynamic_gear_info : shapes) {
  374. GELOGI("Train_Dynamic dynamic_gear_info is %s", formats::JoinToString(dynamic_gear_info).c_str());
  375. vector<int64_t> one_gear;
  376. if (dynamic_gear_info.size() == static_cast<size_t>(dynamic_dims_num)) {
  377. one_gear = dynamic_gear_info;
  378. } else if (dynamic_gear_info.size() > static_cast<size_t>(dynamic_dims_num)) {
  379. auto tmp_index = cur_data_index;
  380. for (size_t i = 0; i < static_cast<size_t>(dynamic_dims_num); ++i) {
  381. if (tmp_index >= dynamic_gear_info.size()) {
  382. ErrorManager::GetInstance().ATCReportErrMessage(
  383. "E10045", {"name", "shape"}, {data_name, formats::JoinToString(data_shape)});
  384. GELOGE(PARAM_INVALID, "Data: %s shape: %s make dynamic dims overflow", data_name.c_str(),
  385. formats::JoinToString(data_shape).c_str());
  386. return FAILED;
  387. }
  388. one_gear.push_back(dynamic_gear_info[tmp_index++]);
  389. }
  390. } else {
  391. ErrorManager::GetInstance().ATCReportErrMessage(
  392. "E10046", {"name", "shape"}, {data_name, formats::JoinToString(data_shape)});
  393. GELOGE(PARAM_INVALID, "Dynamic dims num of data: %s shape: %s can not be more than one gear dynamic info size",
  394. data_name.c_str(), formats::JoinToString(data_shape).c_str());
  395. return FAILED;
  396. }
  397. GELOGI("Train_Dynamic one_gear is %s.", formats::JoinToString(one_gear).c_str());
  398. dynamic_info.push_back(one_gear);
  399. }
  400. cur_data_index += dynamic_dims_num;
  401. data_to_dynamic_info[data_name] = dynamic_info;
  402. }
  403. return SUCCESS;
  404. }
  405. ///
  406. /// @ingroup ge
  407. /// @brief Check Dynamic Param is invalid.
  408. /// @param [in] const vector<vector<int64_t>> &shapes: Params for check.
  409. /// @return SUCCESS: valid / PARAM_INVALID: invalid.
  410. ///
  411. Status CheckDynamicParams(const vector<vector<int64_t>> &shapes) {
  412. if (shapes.size() < kMinShapesCount) {
  413. ErrorManager::GetInstance().ATCReportErrMessage(
  414. "E10035", {"shapesize", "minshapesize"}, {std::to_string(shapes.size()), std::to_string(kMinShapesCount - 1)});
  415. GELOGE(PARAM_INVALID,
  416. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims]'s "
  417. "value size [%zu] must be greater than [%zu].",
  418. shapes.size(), kMinShapesCount - 1);
  419. return PARAM_INVALID;
  420. }
  421. if (shapes.size() > kMaxShapesCount) {
  422. ErrorManager::GetInstance().ATCReportErrMessage(
  423. "E10036", {"shapesize", "maxshapesize"}, {std::to_string(shapes.size()), std::to_string(kMaxShapesCount + 1)});
  424. GELOGE(PARAM_INVALID,
  425. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims]'s "
  426. "value size [%zu] must be less than [%zu].",
  427. shapes.size(), kMaxShapesCount + 1);
  428. return PARAM_INVALID;
  429. }
  430. std::set<std::vector<int64_t>> shapes_set;
  431. size_t shape_size = shapes.at(0).size();
  432. for (auto &shape : shapes) {
  433. if (shape_size != shape.size()) {
  434. ErrorManager::GetInstance().ATCReportErrMessage("E10037", {"shapesize1", "shapesize2"},
  435. {std::to_string(shape_size), std::to_string(shape.size())});
  436. GELOGE(PARAM_INVALID,
  437. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims]'s "
  438. "value size must be same, first group's size is %zu and another's is %zu.",
  439. shape_size, shape.size());
  440. return PARAM_INVALID;
  441. }
  442. for (auto dim : shape) {
  443. if (dim <= 0) {
  444. ErrorManager::GetInstance().ATCReportErrMessage("E10038", {"dim"}, {std::to_string(dim)});
  445. GELOGE(PARAM_INVALID, "Invalid dim %ld, all dims must be greater than 0", dim);
  446. return PARAM_INVALID;
  447. }
  448. }
  449. shapes_set.insert(shape);
  450. }
  451. if (shapes_set.size() != shapes.size()) {
  452. ErrorManager::GetInstance().ATCReportErrMessage("E10039");
  453. GELOGE(PARAM_INVALID,
  454. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims] exist duplicate shapes.");
  455. return PARAM_INVALID;
  456. }
  457. return SUCCESS;
  458. }
  459. ///
  460. /// @ingroup ge
  461. /// @brief Get GeShape from configed shape.
  462. /// @param [in] const std::vector<int64_t> &batch_shape: Configed shape.
  463. /// @param [out] GeShape &data_shape: GeShape for configed shape.
  464. /// @return SUCCESS / PARAM_INVALID
  465. ///
  466. Status CalcShape(const std::vector<int64_t> &batch_shape, GeShape &data_shape) {
  467. size_t batch_shape_index = 0;
  468. for (size_t i = 0; i < data_shape.GetDimNum(); ++i) {
  469. if (data_shape.GetDim(i) < 0) {
  470. if (batch_shape_index >= batch_shape.size()) {
  471. ErrorManager::GetInstance().ATCReportErrMessage(
  472. "E19012", {"function", "reason"},
  473. {"CalcShape", "the batch shape count " + std::to_string(batch_shape.size()) +
  474. " does not match the data shape " + data_shape.ToString()});
  475. GELOGE(PARAM_INVALID,
  476. "Failed to calc tensor shape, the batch shape count %zu, does not match the data shape %s",
  477. batch_shape.size(), data_shape.ToString().c_str());
  478. return PARAM_INVALID;
  479. }
  480. data_shape.SetDim(i, batch_shape[batch_shape_index++]);
  481. }
  482. }
  483. GELOGI("CalcShape size of batch_shape is %zu, batch_shape_index is %zu.", batch_shape.size(), batch_shape_index);
  484. if (batch_shape_index != batch_shape.size()) {
  485. ErrorManager::GetInstance().ATCReportErrMessage(
  486. "E19012", {"function", "reason"}, {"CalcShape", "the batch shape count " + std::to_string(batch_shape.size()) +
  487. " does not match the data shape " + data_shape.ToString()});
  488. GELOGE(PARAM_INVALID, "Failed to calc tensor shape, the batch shape count %zu, does not match the data shape %s",
  489. batch_shape.size(), data_shape.ToString().c_str());
  490. return PARAM_INVALID;
  491. }
  492. return SUCCESS;
  493. }
  494. ///
  495. /// @ingroup ge
  496. /// @brief Set mbatch_dynamic_type on node.
  497. /// @param [in] const OpDescPtr &op_desc: Node for set attribute.
  498. /// @return 0: SUCCESS / others: INTERNAL_ERROR
  499. ///
  500. Status StampDynamicType(const OpDescPtr &op_desc) {
  501. GE_CHECK_NOTNULL(op_desc);
  502. int32_t dynamic_type = static_cast<int32_t>(FIXED);
  503. if (!GetLocalOmgContext().dynamic_batch_size.empty()) {
  504. dynamic_type = static_cast<int32_t>(DYNAMIC_BATCH);
  505. }
  506. if (!GetLocalOmgContext().dynamic_image_size.empty()) {
  507. dynamic_type = static_cast<int32_t>(DYNAMIC_IMAGE);
  508. }
  509. if (!GetLocalOmgContext().dynamic_dims.empty()) {
  510. dynamic_type = static_cast<int32_t>(DYNAMIC_DIMS);
  511. }
  512. if (!AttrUtils::SetInt(op_desc, ATTR_DYNAMIC_TYPE, dynamic_type)) {
  513. GELOGE(INTERNAL_ERROR, "Failed to add dynamic type attr for node %s", op_desc->GetName().c_str());
  514. return INTERNAL_ERROR;
  515. }
  516. return SUCCESS;
  517. }
  518. ///
  519. /// @ingroup ge
  520. /// @brief Check dynamic batch Shape.
  521. /// @param [in] const vector<int64_t> &shape: data_shape to be checked.
  522. /// @param [in] const string &data_name: cur data name.
  523. /// @return 0: true/false
  524. ///
  525. bool CheckDynamicBatchShape(const vector<int64_t> &shape, const string &data_name) {
  526. if (shape[0] == kDynmaicDims) {
  527. for (size_t i = 1; i < shape.size(); ++i) {
  528. if (shape[i] < 1) {
  529. ErrorManager::GetInstance().ATCReportErrMessage("E10018", {"index", "shape"},
  530. {std::to_string(i), std::to_string(shape[i])});
  531. GELOGE(ge::PARAM_INVALID,
  532. "Only batch N can be -1 when set --dynamic_batch_size, current data: %s shape[%zu] is %ld",
  533. data_name.c_str(), i, shape[i]);
  534. return false;
  535. }
  536. }
  537. return true;
  538. } else {
  539. return false;
  540. }
  541. }
  542. ///
  543. /// @ingroup ge
  544. /// @brief Check Dynamic image size shape.
  545. /// @param [in] unordered_map<string, vector<int64_t>> &shape_map: map of data_name and data_shape.
  546. /// @param [in] const std::string &input_format: format of input.
  547. /// @return 0: true/false
  548. ///
  549. bool CheckDynamicImageSizeShape(const vector<int64_t> &shape, const string &data_name,
  550. const std::string &input_format) {
  551. int64_t height = 0;
  552. int64_t width = 0;
  553. if (input_format == "NCHW") {
  554. height = shape[NCHW_DIM_H];
  555. width = shape[NCHW_DIM_W];
  556. }
  557. if (input_format == "NHWC") {
  558. height = shape[NHWC_DIM_H];
  559. width = shape[NHWC_DIM_W];
  560. }
  561. if (height == kDynmaicDims && width == kDynmaicDims &&
  562. std::count(shape.begin(), shape.end(), kDynmaicDims) == kDynamicImgSizeDynamciDimsNum) {
  563. return true;
  564. } else {
  565. ErrorManager::GetInstance().ATCReportErrMessage("E10019");
  566. GELOGE(ge::PARAM_INVALID,
  567. "--input_shape's shape is invalid, only height and width can be -1 when set --dynamic_image_size.");
  568. return false;
  569. }
  570. }
  571. } // namespace multibatch
  572. } // 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