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

5 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647
  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 when %s",
  100. data_nodes.size(), graph->GetName().c_str(), GetLocalOmgContext().user_input_dims.size(),
  101. __FUNCTION__);
  102. GELOGE(PARAM_INVALID, "The count of input shape:%zu should be equal to the count of data num:%zu.",
  103. GetLocalOmgContext().user_input_dims.size(), data_nodes.size());
  104. return PARAM_INVALID;
  105. }
  106. for (size_t i = 0; i < data_nodes.size(); ++i) {
  107. auto data_node = data_nodes.at(i);
  108. GE_CHECK_NOTNULL(data_node);
  109. GE_CHECK_NOTNULL(data_node->GetOpDesc());
  110. auto output_shape = data_node->GetOpDesc()->GetOutputDesc(0).GetShape().GetDims();
  111. auto dynamic_dims = GetLocalOmgContext().user_input_dims.at(i).second;
  112. GELOGD("The %zu data node is %s, node shape is %s, dynamic dim is %s.", i, data_node->GetName().c_str(),
  113. formats::JoinToString(output_shape).c_str(), formats::JoinToString(dynamic_dims).c_str());
  114. if (output_shape.empty() && dynamic_dims.size() == 1 && dynamic_dims.at(0) == 0) {
  115. GELOGI("No need to check sequence for constant.");
  116. continue;
  117. }
  118. if (dynamic_dims.size() != output_shape.size()) {
  119. REPORT_INNER_ERROR("E19999", "The output shape of %s is %s, the input shape from options of %s is %s, graph:%s,"
  120. "check invalid when %s", data_node->GetName().c_str(),
  121. formats::JoinToString(output_shape).c_str(),
  122. GetLocalOmgContext().user_input_dims.at(i).first.c_str(),
  123. formats::JoinToString(dynamic_dims).c_str(), graph->GetName().c_str(), __FUNCTION__);
  124. GELOGE(PARAM_INVALID, "The output shape of %s is %s, the input shape from options of %s is %s.",
  125. data_node->GetName().c_str(), formats::JoinToString(output_shape).c_str(),
  126. GetLocalOmgContext().user_input_dims.at(i).first.c_str(), formats::JoinToString(dynamic_dims).c_str());
  127. return PARAM_INVALID;
  128. }
  129. for (size_t j = 0; j < dynamic_dims.size(); ++j) {
  130. if (dynamic_dims.at(j) != kDynmaicDims && dynamic_dims.at(j) != output_shape.at(j)) {
  131. REPORT_INNER_ERROR("E19999", "Value of input shape %s from option and output shape %s of data op:%s "
  132. "should be equal to %d, index:%zu, graph:%s, check invalid when %s",
  133. formats::JoinToString(dynamic_dims).c_str(),
  134. formats::JoinToString(output_shape).c_str(), data_node->GetName().c_str(), kDynmaicDims,
  135. j, graph->GetName().c_str(), __FUNCTION__);
  136. GELOGE(INTERNAL_ERROR, "Value of input shape %s should be equal to %s.",
  137. formats::JoinToString(dynamic_dims).c_str(), formats::JoinToString(output_shape).c_str());
  138. return INTERNAL_ERROR;
  139. }
  140. }
  141. }
  142. return SUCCESS;
  143. }
  144. Status CheckSequenceOfGetnext(ComputeGraphPtr &graph, const vector<NodePtr> &getnext_sink_node) {
  145. GELOGD("Start check input sequence from getnext sink nodes and input shape.");
  146. if (getnext_sink_node.size() != kNumOfGetnextNode) {
  147. REPORT_INNER_ERROR("E19999", "Not support dynamic dims when a graph with multi getnext nodes, graph:%s, "
  148. "num of getnext node:%zu, check invalid when %s",
  149. graph->GetName().c_str(), getnext_sink_node.size(), __FUNCTION__);
  150. GELOGE(PARAM_INVALID, "Not support dynamic dims when a graph with multi getnext nodes.");
  151. return PARAM_INVALID;
  152. }
  153. auto data_node = getnext_sink_node.at(0);
  154. GE_CHECK_NOTNULL(data_node);
  155. auto op_desc = data_node->GetOpDesc();
  156. GE_CHECK_NOTNULL(op_desc);
  157. size_t data_count = data_node->GetAllOutDataAnchors().size() / kDivisionConst;
  158. if (data_count != GetLocalOmgContext().user_input_dims.size()) {
  159. REPORT_INNER_ERROR("E19999", "Output desc count of %s is %zu, should be equal to count of input shape: %zu, "
  160. "graph:%s, check invalid when %s", op_desc->GetName().c_str(), data_count,
  161. GetLocalOmgContext().user_input_dims.size(), graph->GetName().c_str(), __FUNCTION__);
  162. GELOGE(PARAM_INVALID, "Output count of %s is %zu, should be equal to count of input shape: %zu",
  163. op_desc->GetName().c_str(), data_count, GetLocalOmgContext().user_input_dims.size());
  164. return PARAM_INVALID;
  165. }
  166. for (size_t i = 0; i < data_count; ++i) {
  167. auto output_shape = data_node->GetOpDesc()->GetOutputDesc(i).GetShape().GetDims();
  168. auto dynamic_dims = GetLocalOmgContext().user_input_dims.at(i).second;
  169. GELOGD("The %zu getnext node is %s, node shape is %s, dynamic dim is %s.", i, data_node->GetName().c_str(),
  170. formats::JoinToString(output_shape).c_str(), formats::JoinToString(dynamic_dims).c_str());
  171. if (output_shape.empty() && dynamic_dims.size() == 1 && dynamic_dims.at(0) == 0) {
  172. GELOGI("No need to check sequence for constant.");
  173. continue;
  174. }
  175. if (dynamic_dims.size() != output_shape.size()) {
  176. REPORT_INNER_ERROR("E19999", "The %zu output_shape of %s is %s not equal to the input_shape:%s "
  177. "from options of %s, graph:%s, check invalid when %s", i,
  178. data_node->GetName().c_str(), formats::JoinToString(output_shape).c_str(),
  179. formats::JoinToString(dynamic_dims).c_str(),
  180. GetLocalOmgContext().user_input_dims.at(i).first.c_str(),
  181. graph->GetName().c_str(), __FUNCTION__);
  182. GELOGE(PARAM_INVALID, "the output_shape of %s is %s, the input_shape from options of %s is %s.",
  183. data_node->GetName().c_str(), formats::JoinToString(output_shape).c_str(),
  184. GetLocalOmgContext().user_input_dims.at(i).first.c_str(), formats::JoinToString(dynamic_dims).c_str());
  185. return PARAM_INVALID;
  186. }
  187. for (size_t j = 0; j < dynamic_dims.size(); ++j) {
  188. if (dynamic_dims.at(j) != kDynmaicDims && dynamic_dims.at(j) != output_shape.at(j)) {
  189. REPORT_INNER_ERROR("E19999", "Value of input shape %s from option and output shape %s of data op:%s "
  190. "should be equal to %d, index:%zu, graph:%s, check invalid when %s",
  191. formats::JoinToString(dynamic_dims).c_str(),
  192. formats::JoinToString(output_shape).c_str(), data_node->GetName().c_str(), kDynmaicDims,
  193. j, graph->GetName().c_str(), __FUNCTION__);
  194. GELOGE(INTERNAL_ERROR, "value of input_shape %s should be equal to %s.",
  195. formats::JoinToString(dynamic_dims).c_str(), formats::JoinToString(output_shape).c_str());
  196. return INTERNAL_ERROR;
  197. }
  198. }
  199. }
  200. return SUCCESS;
  201. }
  202. Status CheckSequenceOfOptions(ComputeGraphPtr &graph, vector<NodePtr> &data_nodes,
  203. vector<NodePtr> &getnext_nosink_nodes, vector<NodePtr> &getnext_sink_nodes) {
  204. if (GetLocalOmgContext().dynamic_node_type.empty()) {
  205. GELOGI("No need to CheckSequenceOfOptions.");
  206. return SUCCESS;
  207. }
  208. if (DistinguishGetNextAndData(graph, data_nodes, getnext_nosink_nodes, getnext_sink_nodes) != SUCCESS) {
  209. GELOGE(PARAM_INVALID, "DistinguishGetNextAndData failed.");
  210. return PARAM_INVALID;
  211. }
  212. if (GetLocalOmgContext().dynamic_node_type == DATA) {
  213. GELOGD("Users want data nodes to be dynamic.");
  214. if (CheckSequenceOfData(graph, data_nodes) != SUCCESS) {
  215. GELOGE(PARAM_INVALID, "Failed to check sequence of data nodes.");
  216. return PARAM_INVALID;
  217. }
  218. } else {
  219. GELOGD("Users want getnext nodes to be dynamic.");
  220. if (!getnext_nosink_nodes.empty()) {
  221. if (CheckSequenceOfData(graph, getnext_nosink_nodes) != SUCCESS) {
  222. GELOGE(PARAM_INVALID, "Failed to check sequence of getnext nosink nodes.");
  223. return PARAM_INVALID;
  224. }
  225. } else {
  226. if (CheckSequenceOfGetnext(graph, getnext_sink_nodes) != SUCCESS) {
  227. GELOGE(PARAM_INVALID, "Failed to check sequence of getnext sink nodes.");
  228. return PARAM_INVALID;
  229. }
  230. }
  231. }
  232. return SUCCESS;
  233. }
  234. Status UpdateNameOfData(ComputeGraphPtr &graph, const vector<NodePtr> &data_nodes) {
  235. GELOGD("Update first value of input shape by data nodes.");
  236. if (data_nodes.size() != GetLocalOmgContext().user_input_dims.size()) {
  237. REPORT_INNER_ERROR("E19999", "Count:%zu of data_nodes in graph:%s should be equal to "
  238. "input_shape count:%zu from option, check invalid when %s",
  239. data_nodes.size(), graph->GetName().c_str(), GetLocalOmgContext().user_input_dims.size(),
  240. __FUNCTION__);
  241. GELOGE(PARAM_INVALID, "count of data_nodes: %zu should be equal to input_shape count: %zu.",
  242. data_nodes.size(), GetLocalOmgContext().user_input_dims.size());
  243. return PARAM_INVALID;
  244. }
  245. for (size_t i = 0; i < data_nodes.size(); ++i) {
  246. GELOGD("The %zu data name is %s.", i, data_nodes.at(i)->GetOpDesc()->GetName().c_str());
  247. GetLocalOmgContext().user_input_dims.at(i).first = data_nodes.at(i)->GetOpDesc()->GetName();
  248. }
  249. return SUCCESS;
  250. }
  251. Status UpdateNameOfGetnext(ComputeGraphPtr &graph, const vector<NodePtr> &getnext_sink_nodes) {
  252. GELOGD("Update first value of input shape by getnext sink nodes.");
  253. if (getnext_sink_nodes.size() != kNumOfGetnextNode) {
  254. REPORT_INNER_ERROR("E19999", "Not support dynamic dims when a graph with multi getnext nodes, graph:%s, "
  255. "num of getnext node:%zu, check invalid when %s",
  256. graph->GetName().c_str(), getnext_sink_nodes.size(), __FUNCTION__);
  257. GELOGE(PARAM_INVALID, "Not support dynamic dims when a graph with multi getnext nodes.");
  258. return PARAM_INVALID;
  259. }
  260. auto input_node = getnext_sink_nodes.at(0);
  261. GE_CHECK_NOTNULL(input_node);
  262. auto op_desc = input_node->GetOpDesc();
  263. GE_CHECK_NOTNULL(op_desc);
  264. // user want getnext dynamic, just getnext or data+getnext_sink
  265. size_t data_count = input_node->GetAllOutDataAnchors().size() / kDivisionConst;
  266. if (data_count != GetLocalOmgContext().user_input_dims.size()) {
  267. REPORT_INNER_ERROR("E19999", "Output desc count of %s is %zu, should be equal to count of input shape: %zu, "
  268. "graph:%s, check invalid when %s", op_desc->GetName().c_str(), data_count,
  269. GetLocalOmgContext().user_input_dims.size(), graph->GetName().c_str(), __FUNCTION__);
  270. GELOGE(PARAM_INVALID, "Output count of %s is %zu, should be equal to count of input shape: %zu",
  271. op_desc->GetName().c_str(), data_count, GetLocalOmgContext().user_input_dims.size());
  272. return PARAM_INVALID;
  273. }
  274. for (size_t i = 0; i < data_count; ++i) {
  275. string data_name = op_desc->GetName() + "_" + std::to_string(i);
  276. GELOGD("Data just from getnext sink is %s.", data_name.c_str());
  277. GetLocalOmgContext().user_input_dims.at(i).first = data_name;
  278. }
  279. return SUCCESS;
  280. }
  281. // need to distinguish online and offline, offline no need to update the name of input_shape
  282. Status UpdateNameOfInputShape(ComputeGraphPtr &graph, const vector<NodePtr> &data_nodes,
  283. const vector<NodePtr> &getnext_nosink_nodes, const vector<NodePtr> &getnext_sink_nodes) {
  284. if (GetLocalOmgContext().dynamic_node_type.empty()) {
  285. GELOGI("No need to update first value of input shape when offline infer.");
  286. return SUCCESS;
  287. }
  288. if (GetLocalOmgContext().dynamic_node_type == DATA) {
  289. GELOGD("Users want data nodes to be dynamic.");
  290. if (UpdateNameOfData(graph, data_nodes) != SUCCESS) {
  291. GELOGE(PARAM_INVALID, "Failed to update first value of input shape of data nodes.");
  292. return PARAM_INVALID;
  293. }
  294. } else {
  295. GELOGD("Users want getnext nodes to be dynamic.");
  296. if (!getnext_nosink_nodes.empty()) {
  297. if (UpdateNameOfData(graph, getnext_nosink_nodes) != SUCCESS) {
  298. GELOGE(PARAM_INVALID, "Failed to update first value of input shape of getnext nosink nodes.");
  299. return PARAM_INVALID;
  300. }
  301. } else {
  302. if (UpdateNameOfGetnext(graph, getnext_sink_nodes) != SUCCESS) {
  303. GELOGE(PARAM_INVALID, "Failed to update first value of input shape of getnext sink nodes.");
  304. return PARAM_INVALID;
  305. }
  306. }
  307. }
  308. return SUCCESS;
  309. }
  310. Status DeleteIdentityInsertByAdapter(ComputeGraphPtr &graph) {
  311. GELOGD("Start delete identity node inserted by adapter.");
  312. for (NodePtr &node : graph->GetDirectNode()) {
  313. GE_CHECK_NOTNULL(node);
  314. OpDescPtr op_desc = node->GetOpDesc();
  315. GE_CHECK_NOTNULL(op_desc);
  316. if (IsGetNextType(node)) {
  317. for (auto &out_data_anchor : node->GetAllOutDataAnchors()) {
  318. GE_IF_BOOL_EXEC(out_data_anchor == nullptr, continue);
  319. for (auto &peer_in_anchor : out_data_anchor->GetPeerInDataAnchors()) {
  320. GE_IF_BOOL_EXEC(peer_in_anchor == nullptr, continue);
  321. auto dst_node = peer_in_anchor->GetOwnerNode();
  322. GE_IF_BOOL_EXEC(dst_node == nullptr, continue);
  323. if (dst_node->GetType() == IDENTITY) {
  324. GELOGI("Need to remove %s.", dst_node->GetName().c_str());
  325. if (ge::GraphUtils::RemoveNodeWithoutRelink(graph, dst_node) != GRAPH_SUCCESS) {
  326. REPORT_CALL_ERROR("E19999", "Remove node:%s(%s) from graph:%s failed when %s",
  327. dst_node->GetName().c_str(), dst_node->GetType().c_str(), graph->GetName().c_str(),
  328. __FUNCTION__);
  329. GELOGE(FAILED, "Remove Identity node %s failed.", dst_node->GetName().c_str());
  330. return FAILED;
  331. }
  332. }
  333. }
  334. }
  335. }
  336. }
  337. return SUCCESS;
  338. }
  339. Status CheckNegativeCountOfOptions(const std::vector<std::vector<int64_t>> &shapes) {
  340. if (!GetLocalOmgContext().dynamic_dims.empty()) {
  341. size_t negative_count = 0;
  342. for (size_t i = 0; i < GetLocalOmgContext().user_input_dims.size(); ++i) {
  343. for (size_t j = 0; j < GetLocalOmgContext().user_input_dims.at(i).second.size(); ++j) {
  344. if (GetLocalOmgContext().user_input_dims.at(i).second.at(j) == kDynmaicDims) {
  345. negative_count++;
  346. }
  347. }
  348. }
  349. for (size_t i = 0; i < shapes.size(); ++i) {
  350. if (shapes.at(i).size() != negative_count) {
  351. REPORT_INNER_ERROR("E19999", "gear num of dynamic_dims is %zu should be equal to num:%zu from option, "
  352. "check invalid when %s", shapes.at(i).size(), negative_count, __FUNCTION__);
  353. GELOGE(PARAM_INVALID, "Each gear num of dynamic_dims is %zu should be equal to %zu.", shapes.at(i).size(),
  354. negative_count);
  355. return PARAM_INVALID;
  356. }
  357. }
  358. }
  359. return SUCCESS;
  360. }
  361. ///
  362. /// @ingroup ge
  363. /// @brief Init Dynamic Param from Options.
  364. /// @param [out] std::vector<std::vector<int64_t>> &shapes: Result for Params.
  365. /// @return true: Configed for Multi batch / false: Not configed for Multi batch.
  366. ///
  367. bool InitDynamicParams(vector<vector<int64_t>> &shapes) {
  368. if (!GetLocalOmgContext().dynamic_batch_size.empty()) {
  369. GELOGD("Found dynamic batch option, value %s", GetLocalOmgContext().dynamic_batch_size.c_str());
  370. std::vector<std::string> dims = ge::StringUtils::Split(GetLocalOmgContext().dynamic_batch_size, ',');
  371. for (const auto &dim : dims) {
  372. if (dim.empty()) {
  373. continue;
  374. }
  375. shapes.emplace_back(std::vector<int64_t>({std::strtol(dim.c_str(), nullptr, kDecimal)}));
  376. GELOGI("Found dynamic batch, shape %s", formats::JoinToString(*shapes.rbegin()).c_str());
  377. }
  378. }
  379. if (!GetLocalOmgContext().dynamic_image_size.empty()) {
  380. GELOGD("Found dynamic image size option, value %s", GetLocalOmgContext().dynamic_image_size.c_str());
  381. ParseDynamicSize(GetLocalOmgContext().dynamic_image_size, shapes);
  382. for (const auto &shape : shapes) {
  383. GELOGI("Found dynamic image size, shape %s", formats::JoinToString(shape).c_str());
  384. }
  385. }
  386. if (!GetLocalOmgContext().dynamic_dims.empty()) {
  387. GELOGD("Found dynamic dims option, value %s", GetLocalOmgContext().dynamic_dims.c_str());
  388. ParseDynamicSize(GetLocalOmgContext().dynamic_dims, shapes);
  389. for (const auto &shape : shapes) {
  390. GELOGI("Found dynamic dims, shape %s", formats::JoinToString(shape).c_str());
  391. }
  392. }
  393. return !shapes.empty();
  394. }
  395. ///
  396. /// @ingroup ge
  397. /// @brief parse each data's own dynamic dims.
  398. /// @param [out] map<string, vector<vector<int64_t>>> &data_to_dynamic_info: key:data_name. value:dynamic dims.
  399. /// @return true: Configed for Multi batch / false: Not configed for Multi batch.
  400. ///
  401. Status ParserDataToDynamicInfo(const vector<vector<int64_t>> &shapes,
  402. vector<pair<string, vector<int64_t>>> &data_name_and_shape,
  403. map<string, vector<vector<int64_t>> > &data_to_dynamic_info) {
  404. size_t cur_data_index = 0;
  405. for (size_t index = 0; index < data_name_and_shape.size(); ++index) {
  406. auto &cur_item = data_name_and_shape[index];
  407. auto &data_name = cur_item.first;
  408. auto &data_shape = cur_item.second;
  409. auto dynamic_dims_num = std::count_if(data_shape.begin(), data_shape.end(),
  410. [&data_shape](int64_t dim){ return dim < 0; });
  411. GELOGI("Train_Dynamic dynamic_dims_num of %s is %zu", data_name.c_str(), dynamic_dims_num);
  412. vector<vector<int64_t> > dynamic_info;
  413. for (auto &dynamic_gear_info : shapes) {
  414. GELOGI("Train_Dynamic dynamic_gear_info is %s", formats::JoinToString(dynamic_gear_info).c_str());
  415. vector<int64_t> one_gear;
  416. if (dynamic_gear_info.size() == static_cast<size_t>(dynamic_dims_num)) {
  417. one_gear = dynamic_gear_info;
  418. } else if (dynamic_gear_info.size() > static_cast<size_t>(dynamic_dims_num)) {
  419. auto tmp_index = cur_data_index;
  420. for (size_t i = 0; i < static_cast<size_t>(dynamic_dims_num); ++i) {
  421. if (tmp_index >= dynamic_gear_info.size()) {
  422. ErrorManager::GetInstance().ATCReportErrMessage(
  423. "E10045", {"name", "shape"}, {data_name, formats::JoinToString(data_shape)});
  424. GELOGE(PARAM_INVALID, "Data: %s shape: %s make dynamic dims overflow", data_name.c_str(),
  425. formats::JoinToString(data_shape).c_str());
  426. return FAILED;
  427. }
  428. one_gear.push_back(dynamic_gear_info[tmp_index++]);
  429. }
  430. } else {
  431. ErrorManager::GetInstance().ATCReportErrMessage(
  432. "E10046", {"name", "shape"}, {data_name, formats::JoinToString(data_shape)});
  433. GELOGE(PARAM_INVALID, "Dynamic dims num of data: %s shape: %s can not be more than one gear dynamic info size",
  434. data_name.c_str(), formats::JoinToString(data_shape).c_str());
  435. return FAILED;
  436. }
  437. GELOGI("Train_Dynamic one_gear is %s.", formats::JoinToString(one_gear).c_str());
  438. dynamic_info.push_back(one_gear);
  439. }
  440. cur_data_index += dynamic_dims_num;
  441. data_to_dynamic_info[data_name] = dynamic_info;
  442. }
  443. return SUCCESS;
  444. }
  445. ///
  446. /// @ingroup ge
  447. /// @brief Check Dynamic Param is invalid.
  448. /// @param [in] const vector<vector<int64_t>> &shapes: Params for check.
  449. /// @return SUCCESS: valid / PARAM_INVALID: invalid.
  450. ///
  451. Status CheckDynamicParams(const vector<vector<int64_t>> &shapes) {
  452. if (shapes.size() < kMinShapesCount) {
  453. ErrorManager::GetInstance().ATCReportErrMessage(
  454. "E10035", {"shapesize", "minshapesize"}, {std::to_string(shapes.size()), std::to_string(kMinShapesCount - 1)});
  455. GELOGE(PARAM_INVALID,
  456. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims]'s "
  457. "value size [%zu] must be greater than [%d].",
  458. shapes.size(), kMinShapesCount - 1);
  459. return PARAM_INVALID;
  460. }
  461. if (shapes.size() > kMaxShapesCount) {
  462. ErrorManager::GetInstance().ATCReportErrMessage(
  463. "E10036", {"shapesize", "maxshapesize"}, {std::to_string(shapes.size()), std::to_string(kMaxShapesCount + 1)});
  464. GELOGE(PARAM_INVALID,
  465. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims]'s "
  466. "value size [%zu] must be less than [%d].",
  467. shapes.size(), kMaxShapesCount + 1);
  468. return PARAM_INVALID;
  469. }
  470. std::set<std::vector<int64_t>> shapes_set;
  471. size_t shape_size = shapes.at(0).size();
  472. for (auto &shape : shapes) {
  473. if (shape_size != shape.size()) {
  474. ErrorManager::GetInstance().ATCReportErrMessage("E10037", {"shapesize1", "shapesize2"},
  475. {std::to_string(shape_size), std::to_string(shape.size())});
  476. GELOGE(PARAM_INVALID,
  477. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims]'s "
  478. "value size must be same, first group's size is %zu and another's is %zu.",
  479. shape_size, shape.size());
  480. return PARAM_INVALID;
  481. }
  482. for (auto dim : shape) {
  483. if (dim <= 0) {
  484. ErrorManager::GetInstance().ATCReportErrMessage("E10038", {"dim"}, {std::to_string(dim)});
  485. GELOGE(PARAM_INVALID, "Invalid dim %ld, all dims must be greater than 0", dim);
  486. return PARAM_INVALID;
  487. }
  488. }
  489. shapes_set.insert(shape);
  490. }
  491. if (shapes_set.size() != shapes.size()) {
  492. ErrorManager::GetInstance().ATCReportErrMessage("E10039");
  493. GELOGE(PARAM_INVALID,
  494. "Input parameter[--dynamic_batch_size, --dynamic_image_size or --dynamic_dims] exist duplicate shapes.");
  495. return PARAM_INVALID;
  496. }
  497. return SUCCESS;
  498. }
  499. ///
  500. /// @ingroup ge
  501. /// @brief Get GeShape from configed shape.
  502. /// @param [in] const std::vector<int64_t> &batch_shape: Configed shape.
  503. /// @param [out] GeShape &data_shape: GeShape for configed shape.
  504. /// @return SUCCESS / PARAM_INVALID
  505. ///
  506. Status CalcShape(const std::vector<int64_t> &batch_shape, GeShape &data_shape) {
  507. size_t batch_shape_index = 0;
  508. for (size_t i = 0; i < data_shape.GetDimNum(); ++i) {
  509. if (data_shape.GetDim(i) < 0) {
  510. if (batch_shape_index >= batch_shape.size()) {
  511. ErrorManager::GetInstance().ATCReportErrMessage(
  512. "E19012", {"function", "reason"},
  513. {"CalcShape", "the batch shape count " + std::to_string(batch_shape.size()) +
  514. " does not match the data shape " + data_shape.ToString()});
  515. GELOGE(PARAM_INVALID,
  516. "Failed to calc tensor shape, the batch shape count %zu, does not match the data shape %s",
  517. batch_shape.size(), data_shape.ToString().c_str());
  518. return PARAM_INVALID;
  519. }
  520. data_shape.SetDim(i, batch_shape[batch_shape_index++]);
  521. }
  522. }
  523. GELOGI("CalcShape size of batch_shape is %zu, batch_shape_index is %zu.", batch_shape.size(), batch_shape_index);
  524. if (batch_shape_index != batch_shape.size()) {
  525. ErrorManager::GetInstance().ATCReportErrMessage(
  526. "E19012", {"function", "reason"}, {"CalcShape", "the batch shape count " + std::to_string(batch_shape.size()) +
  527. " does not match the data shape " + data_shape.ToString()});
  528. GELOGE(PARAM_INVALID, "Failed to calc tensor shape, the batch shape count %zu, does not match the data shape %s",
  529. batch_shape.size(), data_shape.ToString().c_str());
  530. return PARAM_INVALID;
  531. }
  532. return SUCCESS;
  533. }
  534. ///
  535. /// @ingroup ge
  536. /// @brief Set mbatch_dynamic_type on node.
  537. /// @param [in] const OpDescPtr &op_desc: Node for set attribute.
  538. /// @return 0: SUCCESS / others: INTERNAL_ERROR
  539. ///
  540. Status StampDynamicType(const OpDescPtr &op_desc) {
  541. GE_CHECK_NOTNULL(op_desc);
  542. int32_t dynamic_type = static_cast<int32_t>(FIXED);
  543. if (!GetLocalOmgContext().dynamic_batch_size.empty()) {
  544. dynamic_type = static_cast<int32_t>(DYNAMIC_BATCH);
  545. }
  546. if (!GetLocalOmgContext().dynamic_image_size.empty()) {
  547. dynamic_type = static_cast<int32_t>(DYNAMIC_IMAGE);
  548. }
  549. if (!GetLocalOmgContext().dynamic_dims.empty()) {
  550. dynamic_type = static_cast<int32_t>(DYNAMIC_DIMS);
  551. }
  552. if (!AttrUtils::SetInt(op_desc, ATTR_DYNAMIC_TYPE, dynamic_type)) {
  553. REPORT_CALL_ERROR("E19999", "Set Attr:%s to node:%s(%s) failed when %s",
  554. ATTR_DYNAMIC_TYPE.c_str(), op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  555. GELOGE(INTERNAL_ERROR, "Failed to add dynamic type attr for node %s", op_desc->GetName().c_str());
  556. return INTERNAL_ERROR;
  557. }
  558. return SUCCESS;
  559. }
  560. ///
  561. /// @ingroup ge
  562. /// @brief Check dynamic batch Shape.
  563. /// @param [in] const vector<int64_t> &shape: data_shape to be checked.
  564. /// @param [in] const string &data_name: cur data name.
  565. /// @return 0: true/false
  566. ///
  567. bool CheckDynamicBatchShape(const vector<int64_t> &shape, const string &data_name) {
  568. if (shape[0] == kDynmaicDims) {
  569. for (size_t i = 1; i < shape.size(); ++i) {
  570. if (shape[i] < 1) {
  571. ErrorManager::GetInstance().ATCReportErrMessage("E10018", {"index", "shape"},
  572. {std::to_string(i), std::to_string(shape[i])});
  573. GELOGE(ge::PARAM_INVALID,
  574. "Only batch N can be -1 when set --dynamic_batch_size, current data: %s shape[%zu] is %ld",
  575. data_name.c_str(), i, shape[i]);
  576. return false;
  577. }
  578. }
  579. return true;
  580. } else {
  581. return false;
  582. }
  583. }
  584. ///
  585. /// @ingroup ge
  586. /// @brief Check Dynamic image size shape.
  587. /// @param [in] unordered_map<string, vector<int64_t>> &shape_map: map of data_name and data_shape.
  588. /// @param [in] const std::string &input_format: format of input.
  589. /// @return 0: true/false
  590. ///
  591. bool CheckDynamicImageSizeShape(const vector<int64_t> &shape, const string &data_name,
  592. const std::string &input_format) {
  593. int64_t height = 0;
  594. int64_t width = 0;
  595. if (input_format == "NCHW") {
  596. height = shape[NCHW_DIM_H];
  597. width = shape[NCHW_DIM_W];
  598. }
  599. if (input_format == "NHWC") {
  600. height = shape[NHWC_DIM_H];
  601. width = shape[NHWC_DIM_W];
  602. }
  603. if (height == kDynmaicDims && width == kDynmaicDims &&
  604. std::count(shape.begin(), shape.end(), kDynmaicDims) == kDynamicImgSizeDynamciDimsNum) {
  605. return true;
  606. } else {
  607. ErrorManager::GetInstance().ATCReportErrMessage("E10019");
  608. GELOGE(ge::PARAM_INVALID,
  609. "--input_shape's shape is invalid, only height and width can be -1 when set --dynamic_image_size.");
  610. return false;
  611. }
  612. }
  613. } // namespace multibatch
  614. } // 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