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 30 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
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
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
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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644
  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 kDynamicImgSizeDynamciDimsNum = 2;
  38. const size_t kNumOfGetnextNode = 1;
  39. const int kDivisionConst = 2;
  40. const char *const kSubstrOfGetNextNosinkName = "IteratorGetNext";
  41. const char *const kShapeDataName = "ascend_mbatch_shape_data";
  42. const char *const kGetNextName = "IteratorV2";
  43. inline bool IsGetNextType(const NodePtr &node) {
  44. std::string original_type;
  45. GE_IF_BOOL_EXEC(GetOriginalType(node, original_type) != SUCCESS,
  46. GELOGW("Get original type failed."); return false);
  47. return (original_type == kGetNextName);
  48. }
  49. void ParseDynamicSize(string dynamic_size, vector<vector<int64_t>> &shapes) {
  50. std::vector<std::string> shape_strs = ge::StringUtils::Split(dynamic_size, ';');
  51. for (const auto &shape_str : shape_strs) {
  52. if (shape_str.empty()) {
  53. continue;
  54. }
  55. std::vector<int64_t> shape;
  56. std::vector<std::string> dims = ge::StringUtils::Split(shape_str, ',');
  57. for (const auto &dim : dims) {
  58. if (dim.empty()) {
  59. continue;
  60. }
  61. shape.emplace_back(std::strtol(dim.c_str(), nullptr, kDecimal));
  62. }
  63. if (!shape.empty()) {
  64. shapes.emplace_back(shape);
  65. }
  66. }
  67. }
  68. Status DistinguishGetNextAndData(ComputeGraphPtr &graph, vector<NodePtr> &data_nodes,
  69. vector<NodePtr> &getnext_nosink_nodes, vector<NodePtr> &getnext_sink_nodes) {
  70. GELOGD("Start distinguish getnext and data node.");
  71. for (NodePtr &input_node : graph->GetDirectNode()) {
  72. GE_CHECK_NOTNULL(input_node);
  73. OpDescPtr op_desc = input_node->GetOpDesc();
  74. GE_CHECK_NOTNULL(op_desc);
  75. if (op_desc->GetType() == DATA && op_desc->GetName() != kShapeDataName) {
  76. if (op_desc->GetName().find(kSubstrOfGetNextNosinkName) == string::npos) {
  77. data_nodes.emplace_back(input_node);
  78. GELOGD("Name of data node is %s.", op_desc->GetName().c_str());
  79. } else {
  80. getnext_nosink_nodes.emplace_back(input_node);
  81. GELOGD("Name of getnext nosink is %s.", op_desc->GetName().c_str());
  82. }
  83. }
  84. if (IsGetNextType(input_node)) {
  85. GELOGD("Name of getnext sink is %s.", op_desc->GetName().c_str());
  86. getnext_sink_nodes.emplace_back(input_node);
  87. }
  88. }
  89. GELOGI("Data count is %zu, getnext nosink count is %zu, getnext sink count is %zu.", data_nodes.size(),
  90. getnext_nosink_nodes.size(), getnext_sink_nodes.size());
  91. GetLocalOmgContext().data_nodes = data_nodes;
  92. GetLocalOmgContext().getnext_nosink_nodes = getnext_nosink_nodes;
  93. return SUCCESS;
  94. }
  95. Status CheckSequenceOfData(ComputeGraphPtr &graph, const vector<NodePtr> &data_nodes) {
  96. GELOGD("Start check input sequence from data nodes and input shape.");
  97. if (data_nodes.size() != GetLocalOmgContext().user_input_dims.size()) {
  98. REPORT_INNER_ERROR("E19999", "Count:%zu of data_nodes in graph:%s should be equal to "
  99. "input_shape count:%zu from option, check invalid",
  100. data_nodes.size(), graph->GetName().c_str(), GetLocalOmgContext().user_input_dims.size());
  101. GELOGE(PARAM_INVALID, "The count of input shape:%zu should be equal to the count of data num:%zu.",
  102. GetLocalOmgContext().user_input_dims.size(), data_nodes.size());
  103. return PARAM_INVALID;
  104. }
  105. for (size_t i = 0; i < data_nodes.size(); ++i) {
  106. auto data_node = data_nodes.at(i);
  107. GE_CHECK_NOTNULL(data_node);
  108. GE_CHECK_NOTNULL(data_node->GetOpDesc());
  109. auto output_shape = data_node->GetOpDesc()->GetOutputDesc(0).GetShape().GetDims();
  110. auto dynamic_dims = GetLocalOmgContext().user_input_dims.at(i).second;
  111. GELOGD("The %zu data node is %s, node shape is %s, dynamic dim is %s.", i, data_node->GetName().c_str(),
  112. formats::JoinToString(output_shape).c_str(), formats::JoinToString(dynamic_dims).c_str());
  113. if (output_shape.empty() && dynamic_dims.size() == 1 && dynamic_dims.at(0) == 0) {
  114. GELOGI("No need to check sequence for constant.");
  115. continue;
  116. }
  117. if (dynamic_dims.size() != output_shape.size()) {
  118. REPORT_INNER_ERROR("E19999", "The output shape of %s is %s, the input shape from options of %s is %s, graph:%s,"
  119. "check invalid", data_node->GetName().c_str(),
  120. formats::JoinToString(output_shape).c_str(),
  121. GetLocalOmgContext().user_input_dims.at(i).first.c_str(),
  122. formats::JoinToString(dynamic_dims).c_str(), graph->GetName().c_str());
  123. GELOGE(PARAM_INVALID, "The output shape of %s is %s, the input shape from options of %s is %s.",
  124. data_node->GetName().c_str(), formats::JoinToString(output_shape).c_str(),
  125. GetLocalOmgContext().user_input_dims.at(i).first.c_str(), formats::JoinToString(dynamic_dims).c_str());
  126. return PARAM_INVALID;
  127. }
  128. for (size_t j = 0; j < dynamic_dims.size(); ++j) {
  129. if (dynamic_dims.at(j) != kDynmaicDims && dynamic_dims.at(j) != output_shape.at(j)) {
  130. REPORT_INNER_ERROR("E19999", "Value of input shape %s from option and output shape %s of data op:%s "
  131. "should be equal to %d, index:%zu, graph:%s, check invalid",
  132. formats::JoinToString(dynamic_dims).c_str(),
  133. formats::JoinToString(output_shape).c_str(), data_node->GetName().c_str(), kDynmaicDims,
  134. j, graph->GetName().c_str());
  135. GELOGE(INTERNAL_ERROR, "Value of input shape %s should be equal to %s.",
  136. formats::JoinToString(dynamic_dims).c_str(), formats::JoinToString(output_shape).c_str());
  137. return INTERNAL_ERROR;
  138. }
  139. }
  140. }
  141. return SUCCESS;
  142. }
  143. Status CheckSequenceOfGetnext(ComputeGraphPtr &graph, const vector<NodePtr> &getnext_sink_node) {
  144. GELOGD("Start check input sequence from getnext sink nodes and input shape.");
  145. if (getnext_sink_node.size() != kNumOfGetnextNode) {
  146. REPORT_INNER_ERROR("E19999", "Not support dynamic dims when a graph with multi getnext nodes, graph:%s, "
  147. "num of getnext node:%zu, check invalid",
  148. graph->GetName().c_str(), getnext_sink_node.size());
  149. GELOGE(PARAM_INVALID, "Not support dynamic dims when a graph with multi getnext nodes.");
  150. return PARAM_INVALID;
  151. }
  152. auto data_node = getnext_sink_node.at(0);
  153. GE_CHECK_NOTNULL(data_node);
  154. auto op_desc = data_node->GetOpDesc();
  155. GE_CHECK_NOTNULL(op_desc);
  156. size_t data_count = data_node->GetAllOutDataAnchors().size() / kDivisionConst;
  157. if (data_count != GetLocalOmgContext().user_input_dims.size()) {
  158. REPORT_INNER_ERROR("E19999", "Output desc count of %s is %zu, should be equal to count of input shape: %zu, "
  159. "graph:%s, check invalid", op_desc->GetName().c_str(), data_count,
  160. GetLocalOmgContext().user_input_dims.size(), graph->GetName().c_str());
  161. GELOGE(PARAM_INVALID, "Output count of %s is %zu, should be equal to count of input shape: %zu",
  162. op_desc->GetName().c_str(), data_count, GetLocalOmgContext().user_input_dims.size());
  163. return PARAM_INVALID;
  164. }
  165. for (size_t i = 0; i < data_count; ++i) {
  166. auto output_shape = data_node->GetOpDesc()->GetOutputDesc(i).GetShape().GetDims();
  167. auto dynamic_dims = GetLocalOmgContext().user_input_dims.at(i).second;
  168. GELOGD("The %zu getnext node is %s, node shape is %s, dynamic dim is %s.", i, data_node->GetName().c_str(),
  169. formats::JoinToString(output_shape).c_str(), formats::JoinToString(dynamic_dims).c_str());
  170. if (output_shape.empty() && dynamic_dims.size() == 1 && dynamic_dims.at(0) == 0) {
  171. GELOGI("No need to check sequence for constant.");
  172. continue;
  173. }
  174. if (dynamic_dims.size() != output_shape.size()) {
  175. REPORT_INNER_ERROR("E19999", "The %zu output_shape of %s is %s not equal to the input_shape:%s "
  176. "from options of %s, graph:%s, check invalid", i,
  177. data_node->GetName().c_str(), formats::JoinToString(output_shape).c_str(),
  178. formats::JoinToString(dynamic_dims).c_str(),
  179. GetLocalOmgContext().user_input_dims.at(i).first.c_str(),
  180. graph->GetName().c_str());
  181. GELOGE(PARAM_INVALID, "the output_shape of %s is %s, the input_shape from options of %s is %s.",
  182. data_node->GetName().c_str(), formats::JoinToString(output_shape).c_str(),
  183. GetLocalOmgContext().user_input_dims.at(i).first.c_str(), formats::JoinToString(dynamic_dims).c_str());
  184. return PARAM_INVALID;
  185. }
  186. for (size_t j = 0; j < dynamic_dims.size(); ++j) {
  187. if (dynamic_dims.at(j) != kDynmaicDims && dynamic_dims.at(j) != output_shape.at(j)) {
  188. REPORT_INNER_ERROR("E19999", "Value of input shape %s from option and output shape %s of data op:%s "
  189. "should be equal to %d, index:%zu, graph:%s, check invalid",
  190. formats::JoinToString(dynamic_dims).c_str(),
  191. formats::JoinToString(output_shape).c_str(), data_node->GetName().c_str(), kDynmaicDims,
  192. j, graph->GetName().c_str());
  193. GELOGE(INTERNAL_ERROR, "value of input_shape %s should be equal to %s.",
  194. formats::JoinToString(dynamic_dims).c_str(), formats::JoinToString(output_shape).c_str());
  195. return INTERNAL_ERROR;
  196. }
  197. }
  198. }
  199. return SUCCESS;
  200. }
  201. Status CheckSequenceOfOptions(ComputeGraphPtr &graph, vector<NodePtr> &data_nodes,
  202. vector<NodePtr> &getnext_nosink_nodes, vector<NodePtr> &getnext_sink_nodes) {
  203. if (GetLocalOmgContext().dynamic_node_type.empty()) {
  204. GELOGI("No need to CheckSequenceOfOptions.");
  205. return SUCCESS;
  206. }
  207. if (DistinguishGetNextAndData(graph, data_nodes, getnext_nosink_nodes, getnext_sink_nodes) != SUCCESS) {
  208. GELOGE(PARAM_INVALID, "DistinguishGetNextAndData failed.");
  209. return PARAM_INVALID;
  210. }
  211. if (GetLocalOmgContext().dynamic_node_type == DATA) {
  212. GELOGD("Users want data nodes to be dynamic.");
  213. if (CheckSequenceOfData(graph, data_nodes) != SUCCESS) {
  214. GELOGE(PARAM_INVALID, "Failed to check sequence of data nodes.");
  215. return PARAM_INVALID;
  216. }
  217. } else {
  218. GELOGD("Users want getnext nodes to be dynamic.");
  219. if (!getnext_nosink_nodes.empty()) {
  220. if (CheckSequenceOfData(graph, getnext_nosink_nodes) != SUCCESS) {
  221. GELOGE(PARAM_INVALID, "Failed to check sequence of getnext nosink nodes.");
  222. return PARAM_INVALID;
  223. }
  224. } else {
  225. if (CheckSequenceOfGetnext(graph, getnext_sink_nodes) != SUCCESS) {
  226. GELOGE(PARAM_INVALID, "Failed to check sequence of getnext sink nodes.");
  227. return PARAM_INVALID;
  228. }
  229. }
  230. }
  231. return SUCCESS;
  232. }
  233. Status UpdateNameOfData(ComputeGraphPtr &graph, const vector<NodePtr> &data_nodes) {
  234. GELOGD("Update first value of input shape by data nodes.");
  235. if (data_nodes.size() != GetLocalOmgContext().user_input_dims.size()) {
  236. REPORT_INNER_ERROR("E19999", "Count:%zu of data_nodes in graph:%s should be equal to "
  237. "input_shape count:%zu from option, check invalid",
  238. data_nodes.size(), graph->GetName().c_str(), GetLocalOmgContext().user_input_dims.size());
  239. GELOGE(PARAM_INVALID, "count of data_nodes: %zu should be equal to input_shape count: %zu.",
  240. data_nodes.size(), GetLocalOmgContext().user_input_dims.size());
  241. return PARAM_INVALID;
  242. }
  243. for (size_t i = 0; i < data_nodes.size(); ++i) {
  244. GELOGD("The %zu data name is %s.", i, data_nodes.at(i)->GetOpDesc()->GetName().c_str());
  245. GetLocalOmgContext().user_input_dims.at(i).first = data_nodes.at(i)->GetOpDesc()->GetName();
  246. }
  247. return SUCCESS;
  248. }
  249. Status UpdateNameOfGetnext(ComputeGraphPtr &graph, const vector<NodePtr> &getnext_sink_nodes) {
  250. GELOGD("Update first value of input shape by getnext sink nodes.");
  251. if (getnext_sink_nodes.size() != kNumOfGetnextNode) {
  252. REPORT_INNER_ERROR("E19999", "Not support dynamic dims when a graph with multi getnext nodes, graph:%s, "
  253. "num of getnext node:%zu, check invalid",
  254. graph->GetName().c_str(), getnext_sink_nodes.size());
  255. GELOGE(PARAM_INVALID, "Not support dynamic dims when a graph with multi getnext nodes.");
  256. return PARAM_INVALID;
  257. }
  258. auto input_node = getnext_sink_nodes.at(0);
  259. GE_CHECK_NOTNULL(input_node);
  260. auto op_desc = input_node->GetOpDesc();
  261. GE_CHECK_NOTNULL(op_desc);
  262. // user want getnext dynamic, just getnext or data+getnext_sink
  263. size_t data_count = input_node->GetAllOutDataAnchors().size() / kDivisionConst;
  264. if (data_count != GetLocalOmgContext().user_input_dims.size()) {
  265. REPORT_INNER_ERROR("E19999", "Output desc count of %s is %zu, should be equal to count of input shape: %zu, "
  266. "graph:%s, check invalid", op_desc->GetName().c_str(), data_count,
  267. GetLocalOmgContext().user_input_dims.size(), graph->GetName().c_str());
  268. GELOGE(PARAM_INVALID, "Output count of %s is %zu, should be equal to count of input shape: %zu",
  269. op_desc->GetName().c_str(), data_count, GetLocalOmgContext().user_input_dims.size());
  270. return PARAM_INVALID;
  271. }
  272. for (size_t i = 0; i < data_count; ++i) {
  273. string data_name = op_desc->GetName() + "_" + std::to_string(i);
  274. GELOGD("Data just from getnext sink is %s.", data_name.c_str());
  275. GetLocalOmgContext().user_input_dims.at(i).first = data_name;
  276. }
  277. return SUCCESS;
  278. }
  279. // need to distinguish online and offline, offline no need to update the name of input_shape
  280. Status UpdateNameOfInputShape(ComputeGraphPtr &graph, const vector<NodePtr> &data_nodes,
  281. const vector<NodePtr> &getnext_nosink_nodes, const vector<NodePtr> &getnext_sink_nodes) {
  282. if (GetLocalOmgContext().dynamic_node_type.empty()) {
  283. GELOGI("No need to update first value of input shape when offline infer.");
  284. return SUCCESS;
  285. }
  286. if (GetLocalOmgContext().dynamic_node_type == DATA) {
  287. GELOGD("Users want data nodes to be dynamic.");
  288. if (UpdateNameOfData(graph, data_nodes) != SUCCESS) {
  289. GELOGE(PARAM_INVALID, "Failed to update first value of input shape of data nodes.");
  290. return PARAM_INVALID;
  291. }
  292. } else {
  293. GELOGD("Users want getnext nodes to be dynamic.");
  294. if (!getnext_nosink_nodes.empty()) {
  295. if (UpdateNameOfData(graph, getnext_nosink_nodes) != SUCCESS) {
  296. GELOGE(PARAM_INVALID, "Failed to update first value of input shape of getnext nosink nodes.");
  297. return PARAM_INVALID;
  298. }
  299. } else {
  300. if (UpdateNameOfGetnext(graph, getnext_sink_nodes) != SUCCESS) {
  301. GELOGE(PARAM_INVALID, "Failed to update first value of input shape of getnext sink nodes.");
  302. return PARAM_INVALID;
  303. }
  304. }
  305. }
  306. return SUCCESS;
  307. }
  308. Status DeleteIdentityInsertByAdapter(ComputeGraphPtr &graph) {
  309. GELOGD("Start delete identity node inserted by adapter.");
  310. for (NodePtr &node : graph->GetDirectNode()) {
  311. GE_CHECK_NOTNULL(node);
  312. OpDescPtr op_desc = node->GetOpDesc();
  313. GE_CHECK_NOTNULL(op_desc);
  314. if (IsGetNextType(node)) {
  315. for (auto &out_data_anchor : node->GetAllOutDataAnchors()) {
  316. GE_IF_BOOL_EXEC(out_data_anchor == nullptr, continue);
  317. for (auto &peer_in_anchor : out_data_anchor->GetPeerInDataAnchors()) {
  318. GE_IF_BOOL_EXEC(peer_in_anchor == nullptr, continue);
  319. auto dst_node = peer_in_anchor->GetOwnerNode();
  320. GE_IF_BOOL_EXEC(dst_node == nullptr, continue);
  321. if (dst_node->GetType() == IDENTITY && dst_node->GetAllOutDataAnchors().empty()) {
  322. GELOGI("Need to remove %s.", dst_node->GetName().c_str());
  323. if (GraphUtils::RemoveNodeWithoutRelink(graph, dst_node) != GRAPH_SUCCESS) {
  324. REPORT_CALL_ERROR("E19999", "Remove node:%s(%s) from graph:%s failed",
  325. dst_node->GetName().c_str(), dst_node->GetType().c_str(), graph->GetName().c_str());
  326. GELOGE(FAILED, "Remove Identity node %s failed.", dst_node->GetName().c_str());
  327. return FAILED;
  328. }
  329. }
  330. }
  331. }
  332. }
  333. }
  334. return SUCCESS;
  335. }
  336. Status CheckNegativeCountOfOptions(const std::vector<std::vector<int64_t>> &shapes) {
  337. if (!GetLocalOmgContext().dynamic_dims.empty()) {
  338. size_t negative_count = 0;
  339. for (size_t i = 0; i < GetLocalOmgContext().user_input_dims.size(); ++i) {
  340. for (size_t j = 0; j < GetLocalOmgContext().user_input_dims.at(i).second.size(); ++j) {
  341. if (GetLocalOmgContext().user_input_dims.at(i).second.at(j) == kDynmaicDims) {
  342. negative_count++;
  343. }
  344. }
  345. }
  346. for (size_t i = 0; i < shapes.size(); ++i) {
  347. if (shapes.at(i).size() != negative_count) {
  348. REPORT_INNER_ERROR("E19999", "gear num of dynamic_dims is %zu should be equal to num:%zu from option, "
  349. "check invalid", shapes.at(i).size(), negative_count);
  350. GELOGE(PARAM_INVALID, "Each gear num of dynamic_dims is %zu should be equal to %zu.", shapes.at(i).size(),
  351. negative_count);
  352. return PARAM_INVALID;
  353. }
  354. }
  355. }
  356. return SUCCESS;
  357. }
  358. ///
  359. /// @ingroup ge
  360. /// @brief Init Dynamic Param from Options.
  361. /// @param [out] std::vector<std::vector<int64_t>> &shapes: Result for Params.
  362. /// @return true: Configed for Multi batch / false: Not configed for Multi batch.
  363. ///
  364. bool InitDynamicParams(vector<vector<int64_t>> &shapes) {
  365. if (!GetLocalOmgContext().dynamic_batch_size.empty()) {
  366. GELOGD("Found dynamic batch option, value %s", GetLocalOmgContext().dynamic_batch_size.c_str());
  367. std::vector<std::string> dims = ge::StringUtils::Split(GetLocalOmgContext().dynamic_batch_size, ',');
  368. for (const auto &dim : dims) {
  369. if (dim.empty()) {
  370. continue;
  371. }
  372. shapes.emplace_back(std::vector<int64_t>({std::strtol(dim.c_str(), nullptr, kDecimal)}));
  373. GELOGI("Found dynamic batch, shape %s", formats::JoinToString(*shapes.rbegin()).c_str());
  374. }
  375. }
  376. if (!GetLocalOmgContext().dynamic_image_size.empty()) {
  377. GELOGD("Found dynamic image size option, value %s", GetLocalOmgContext().dynamic_image_size.c_str());
  378. ParseDynamicSize(GetLocalOmgContext().dynamic_image_size, shapes);
  379. for (const auto &shape : shapes) {
  380. GELOGI("Found dynamic image size, shape %s", formats::JoinToString(shape).c_str());
  381. }
  382. }
  383. if (!GetLocalOmgContext().dynamic_dims.empty()) {
  384. GELOGD("Found dynamic dims option, value %s", GetLocalOmgContext().dynamic_dims.c_str());
  385. ParseDynamicSize(GetLocalOmgContext().dynamic_dims, shapes);
  386. for (const auto &shape : shapes) {
  387. GELOGI("Found dynamic dims, shape %s", formats::JoinToString(shape).c_str());
  388. }
  389. }
  390. return !shapes.empty();
  391. }
  392. ///
  393. /// @ingroup ge
  394. /// @brief parse each data's own dynamic dims.
  395. /// @param [out] map<string, vector<vector<int64_t>>> &data_to_dynamic_info: key:data_name. value:dynamic dims.
  396. /// @return true: Configed for Multi batch / false: Not configed for Multi batch.
  397. ///
  398. Status ParserDataToDynamicInfo(const vector<vector<int64_t>> &shapes,
  399. vector<pair<string, vector<int64_t>>> &data_name_and_shape,
  400. map<string, vector<vector<int64_t>> > &data_to_dynamic_info) {
  401. size_t cur_data_index = 0;
  402. for (size_t index = 0; index < data_name_and_shape.size(); ++index) {
  403. auto &cur_item = data_name_and_shape[index];
  404. auto &data_name = cur_item.first;
  405. auto &data_shape = cur_item.second;
  406. auto dynamic_dims_num = std::count_if(data_shape.begin(), data_shape.end(),
  407. [&data_shape](int64_t dim){ return dim < 0; });
  408. GELOGI("Train_Dynamic dynamic_dims_num of %s is %zu", data_name.c_str(), dynamic_dims_num);
  409. vector<vector<int64_t> > dynamic_info;
  410. for (auto &dynamic_gear_info : shapes) {
  411. GELOGI("Train_Dynamic dynamic_gear_info is %s", formats::JoinToString(dynamic_gear_info).c_str());
  412. vector<int64_t> one_gear;
  413. if (dynamic_gear_info.size() == static_cast<size_t>(dynamic_dims_num)) {
  414. one_gear = dynamic_gear_info;
  415. } else if (dynamic_gear_info.size() > static_cast<size_t>(dynamic_dims_num)) {
  416. auto tmp_index = cur_data_index;
  417. for (size_t i = 0; i < static_cast<size_t>(dynamic_dims_num); ++i) {
  418. if (tmp_index >= dynamic_gear_info.size()) {
  419. ErrorManager::GetInstance().ATCReportErrMessage(
  420. "E10045", {"name", "shape"}, {data_name, formats::JoinToString(data_shape)});
  421. GELOGE(PARAM_INVALID, "Data: %s shape: %s make dynamic dims overflow", data_name.c_str(),
  422. formats::JoinToString(data_shape).c_str());
  423. return FAILED;
  424. }
  425. one_gear.push_back(dynamic_gear_info[tmp_index++]);
  426. }
  427. } else {
  428. ErrorManager::GetInstance().ATCReportErrMessage(
  429. "E10046", {"name", "shape"}, {data_name, formats::JoinToString(data_shape)});
  430. GELOGE(PARAM_INVALID, "Dynamic dims num of data: %s shape: %s can not be more than one gear dynamic info size",
  431. data_name.c_str(), formats::JoinToString(data_shape).c_str());
  432. return FAILED;
  433. }
  434. GELOGI("Train_Dynamic one_gear is %s.", formats::JoinToString(one_gear).c_str());
  435. dynamic_info.push_back(one_gear);
  436. }
  437. cur_data_index += dynamic_dims_num;
  438. data_to_dynamic_info[data_name] = dynamic_info;
  439. }
  440. return SUCCESS;
  441. }
  442. ///
  443. /// @ingroup ge
  444. /// @brief Check Dynamic Param is invalid.
  445. /// @param [in] const vector<vector<int64_t>> &shapes: Params for check.
  446. /// @return SUCCESS: valid / PARAM_INVALID: invalid.
  447. ///
  448. Status CheckDynamicParams(const vector<vector<int64_t>> &shapes) {
  449. if (shapes.size() < kMinShapesCount) {
  450. ErrorManager::GetInstance().ATCReportErrMessage(
  451. "E10035", {"shapesize", "minshapesize"}, {std::to_string(shapes.size()), std::to_string(kMinShapesCount - 1)});
  452. GELOGE(PARAM_INVALID,
  453. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims]'s "
  454. "value size [%zu] must be greater than [%d].",
  455. shapes.size(), kMinShapesCount - 1);
  456. return PARAM_INVALID;
  457. }
  458. if (shapes.size() > kMaxShapesCount) {
  459. ErrorManager::GetInstance().ATCReportErrMessage(
  460. "E10036", {"shapesize", "maxshapesize"}, {std::to_string(shapes.size()), std::to_string(kMaxShapesCount + 1)});
  461. GELOGE(PARAM_INVALID,
  462. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims]'s "
  463. "value size [%zu] must be less than [%d].",
  464. shapes.size(), kMaxShapesCount + 1);
  465. return PARAM_INVALID;
  466. }
  467. std::set<std::vector<int64_t>> shapes_set;
  468. size_t shape_size = shapes.at(0).size();
  469. for (auto &shape : shapes) {
  470. if (shape_size != shape.size()) {
  471. ErrorManager::GetInstance().ATCReportErrMessage("E10037", {"shapesize1", "shapesize2"},
  472. {std::to_string(shape_size), std::to_string(shape.size())});
  473. GELOGE(PARAM_INVALID,
  474. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims]'s "
  475. "value size must be same, first group's size is %zu and another's is %zu.",
  476. shape_size, shape.size());
  477. return PARAM_INVALID;
  478. }
  479. for (auto dim : shape) {
  480. if (dim <= 0) {
  481. ErrorManager::GetInstance().ATCReportErrMessage("E10038", {"dim"}, {std::to_string(dim)});
  482. GELOGE(PARAM_INVALID, "Invalid dim %ld, all dims must be greater than 0", dim);
  483. return PARAM_INVALID;
  484. }
  485. }
  486. shapes_set.insert(shape);
  487. }
  488. if (shapes_set.size() != shapes.size()) {
  489. ErrorManager::GetInstance().ATCReportErrMessage("E10039");
  490. GELOGE(PARAM_INVALID,
  491. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims] exist duplicate shapes.");
  492. return PARAM_INVALID;
  493. }
  494. return SUCCESS;
  495. }
  496. ///
  497. /// @ingroup ge
  498. /// @brief Get GeShape from configed shape.
  499. /// @param [in] const std::vector<int64_t> &batch_shape: Configed shape.
  500. /// @param [out] GeShape &data_shape: GeShape for configed shape.
  501. /// @return SUCCESS / PARAM_INVALID
  502. ///
  503. Status CalcShape(const std::vector<int64_t> &batch_shape, GeShape &data_shape) {
  504. size_t batch_shape_index = 0;
  505. for (size_t i = 0; i < data_shape.GetDimNum(); ++i) {
  506. if (data_shape.GetDim(i) < 0) {
  507. if (batch_shape_index >= batch_shape.size()) {
  508. ErrorManager::GetInstance().ATCReportErrMessage(
  509. "E19012", {"function", "reason"},
  510. {"CalcShape", "the batch shape count " + std::to_string(batch_shape.size()) +
  511. " does not match the data shape " + data_shape.ToString()});
  512. GELOGE(PARAM_INVALID,
  513. "Failed to calc tensor shape, the batch shape count %zu, does not match the data shape %s",
  514. batch_shape.size(), data_shape.ToString().c_str());
  515. return PARAM_INVALID;
  516. }
  517. data_shape.SetDim(i, batch_shape[batch_shape_index++]);
  518. }
  519. }
  520. GELOGI("CalcShape size of batch_shape is %zu, batch_shape_index is %zu.", batch_shape.size(), batch_shape_index);
  521. if (batch_shape_index != batch_shape.size()) {
  522. ErrorManager::GetInstance().ATCReportErrMessage(
  523. "E19012", {"function", "reason"}, {"CalcShape", "the batch shape count " + std::to_string(batch_shape.size()) +
  524. " does not match the data shape " + data_shape.ToString()});
  525. GELOGE(PARAM_INVALID, "Failed to calc tensor shape, the batch shape count %zu, does not match the data shape %s",
  526. batch_shape.size(), data_shape.ToString().c_str());
  527. return PARAM_INVALID;
  528. }
  529. return SUCCESS;
  530. }
  531. ///
  532. /// @ingroup ge
  533. /// @brief Set mbatch_dynamic_type on node.
  534. /// @param [in] const OpDescPtr &op_desc: Node for set attribute.
  535. /// @return 0: SUCCESS / others: INTERNAL_ERROR
  536. ///
  537. Status StampDynamicType(const OpDescPtr &op_desc) {
  538. GE_CHECK_NOTNULL(op_desc);
  539. int32_t dynamic_type = static_cast<int32_t>(FIXED);
  540. if (!GetLocalOmgContext().dynamic_batch_size.empty()) {
  541. dynamic_type = static_cast<int32_t>(DYNAMIC_BATCH);
  542. }
  543. if (!GetLocalOmgContext().dynamic_image_size.empty()) {
  544. dynamic_type = static_cast<int32_t>(DYNAMIC_IMAGE);
  545. }
  546. if (!GetLocalOmgContext().dynamic_dims.empty()) {
  547. dynamic_type = static_cast<int32_t>(DYNAMIC_DIMS);
  548. }
  549. if (!AttrUtils::SetInt(op_desc, ATTR_DYNAMIC_TYPE, dynamic_type)) {
  550. REPORT_CALL_ERROR("E19999", "Set Attr:%s to node:%s(%s) failed",
  551. ATTR_DYNAMIC_TYPE.c_str(), op_desc->GetName().c_str(), op_desc->GetType().c_str());
  552. GELOGE(INTERNAL_ERROR, "Failed to add dynamic type attr for node %s", op_desc->GetName().c_str());
  553. return INTERNAL_ERROR;
  554. }
  555. return SUCCESS;
  556. }
  557. ///
  558. /// @ingroup ge
  559. /// @brief Check dynamic batch Shape.
  560. /// @param [in] const vector<int64_t> &shape: data_shape to be checked.
  561. /// @param [in] const string &data_name: cur data name.
  562. /// @return 0: true/false
  563. ///
  564. bool CheckDynamicBatchShape(const vector<int64_t> &shape, const string &data_name) {
  565. if (shape[0] == kDynmaicDims) {
  566. for (size_t i = 1; i < shape.size(); ++i) {
  567. if (shape[i] < 1) {
  568. ErrorManager::GetInstance().ATCReportErrMessage("E10018", {"index", "shape"},
  569. {std::to_string(i), std::to_string(shape[i])});
  570. GELOGE(ge::PARAM_INVALID,
  571. "Only batch N can be -1 when set --dynamic_batch_size, current data: %s shape[%zu] is %ld",
  572. data_name.c_str(), i, shape[i]);
  573. return false;
  574. }
  575. }
  576. return true;
  577. } else {
  578. return false;
  579. }
  580. }
  581. ///
  582. /// @ingroup ge
  583. /// @brief Check Dynamic image size shape.
  584. /// @param [in] unordered_map<string, vector<int64_t>> &shape_map: map of data_name and data_shape.
  585. /// @param [in] const std::string &input_format: format of input.
  586. /// @return 0: true/false
  587. ///
  588. bool CheckDynamicImageSizeShape(const vector<int64_t> &shape, const string &data_name,
  589. const std::string &input_format) {
  590. int64_t height = 0;
  591. int64_t width = 0;
  592. if (input_format == "NCHW") {
  593. height = shape[NCHW_DIM_H];
  594. width = shape[NCHW_DIM_W];
  595. }
  596. if (input_format == "NHWC") {
  597. height = shape[NHWC_DIM_H];
  598. width = shape[NHWC_DIM_W];
  599. }
  600. if (height == kDynmaicDims && width == kDynmaicDims &&
  601. std::count(shape.begin(), shape.end(), kDynmaicDims) == kDynamicImgSizeDynamciDimsNum) {
  602. return true;
  603. } else {
  604. ErrorManager::GetInstance().ATCReportErrMessage("E10019");
  605. GELOGE(ge::PARAM_INVALID,
  606. "--input_shape's shape is invalid, only height and width can be -1 when set --dynamic_image_size.");
  607. return false;
  608. }
  609. }
  610. } // namespace multibatch
  611. } // 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