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.

transop_without_reshape_fusion_pass.cc 47 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #include "graph/passes/transop_without_reshape_fusion_pass.h"
  17. #include <algorithm>
  18. #include <memory>
  19. #include <sstream>
  20. #include <string>
  21. #include <atomic>
  22. #include "common/ge/ge_util.h"
  23. #include "common/ge_inner_error_codes.h"
  24. #include "common/types.h"
  25. #include "graph/common/transop_util.h"
  26. #include "graph/compute_graph.h"
  27. #include "graph/debug/ge_attr_define.h"
  28. #include "graph/ge_tensor.h"
  29. #include "graph/op_desc.h"
  30. #include "graph/utils/graph_utils.h"
  31. #include "graph/utils/node_utils.h"
  32. #include "graph/utils/op_desc_utils.h"
  33. #include "graph/utils/type_utils.h"
  34. #include "init/gelib.h"
  35. namespace {
  36. const char *const kRemainNode = "node_remain";
  37. const int kInvalidFusionOpCount = -1;
  38. const char *const kAttrNameSrcFormat = "src_format";
  39. const char *const kAttrNameDstFormat = "dst_format";
  40. } // namespace
  41. namespace ge {
  42. void TransOpWithoutReshapeFusionPass::SetRemainNode(
  43. const vector<pair<OutDataAnchorPtr, InDataAnchorPtr>> &nodes_anchor) {
  44. auto iter = nodes_anchor.begin();
  45. while (iter != nodes_anchor.end()) {
  46. auto in_anchor = iter->second;
  47. if (in_anchor == nullptr) {
  48. return;
  49. }
  50. auto in_node = in_anchor->GetOwnerNode();
  51. ++iter;
  52. if (in_node == nullptr) {
  53. return;
  54. }
  55. if (!IsTransOp(in_node)) {
  56. continue;
  57. }
  58. auto op_desc = in_node->GetOpDesc();
  59. if (op_desc == nullptr) {
  60. continue;
  61. }
  62. GELOGI("SetRemainNode node is %s", op_desc->GetName().c_str());
  63. GE_IF_BOOL_EXEC(!op_desc->SetExtAttr(kRemainNode, true), GELOGE(INTERNAL_ERROR, "set ext attr failed"); return);
  64. }
  65. }
  66. bool TransOpWithoutReshapeFusionPass::FormatContinuousCheck(const OutDataAnchorPtr &out_anchor,
  67. const InDataAnchorPtr &in_anchor) {
  68. if (out_anchor == nullptr || in_anchor == nullptr || in_anchor->GetOwnerNode() == nullptr ||
  69. out_anchor->GetOwnerNode() == nullptr) {
  70. return false;
  71. }
  72. auto in_node = in_anchor->GetOwnerNode();
  73. GE_IF_BOOL_EXEC(in_node == nullptr, GELOGE(INTERNAL_ERROR, "in_node is null"); return false);
  74. auto in_op = in_node->GetOpDesc();
  75. auto out_owner_node = out_anchor->GetOwnerNode();
  76. GE_IF_BOOL_EXEC(out_owner_node == nullptr, GELOGE(INTERNAL_ERROR, "out_owner_node is null"); return false);
  77. auto out_op = out_owner_node->GetOpDesc();
  78. GE_IF_BOOL_EXEC(in_op == nullptr, GELOGE(INTERNAL_ERROR, "in_op is null"); return false);
  79. GE_IF_BOOL_EXEC(out_op == nullptr, GELOGE(INTERNAL_ERROR, "out_op is null"); return false);
  80. auto in_op_desc = in_op->GetInputDescPtr(in_anchor->GetIdx());
  81. auto out_op_desc = out_op->GetOutputDescPtr(out_anchor->GetIdx());
  82. GE_IF_BOOL_EXEC(in_op_desc == nullptr, GELOGE(INTERNAL_ERROR, "in_op_desc is null"); return false);
  83. GE_IF_BOOL_EXEC(out_op_desc == nullptr, GELOGE(INTERNAL_ERROR, "out_op_desc is null"); return false);
  84. if (!ShapeEqualCheck(in_op_desc->GetShape(), out_op_desc->GetShape())) {
  85. return false;
  86. }
  87. if (in_op->GetType() == CAST || out_op->GetType() == CAST) {
  88. return TransOpUtil::CheckPrecisionLoss(in_node);
  89. }
  90. if (in_op_desc->GetFormat() == FORMAT_ND) {
  91. return false;
  92. }
  93. if (out_op_desc->GetFormat() == FORMAT_ND) {
  94. return false;
  95. }
  96. if (in_op_desc->GetFormat() != out_op_desc->GetFormat()) {
  97. return false;
  98. }
  99. return FusionFormatSupport(in_op_desc->GetFormat());
  100. }
  101. graphStatus TransOpWithoutReshapeFusionPass::GetSubGraphNodesInfo() {
  102. vector<bool> sub_graph_has_reshape_node(sub_graph_anchors_.size(), false);
  103. vector<int> transop_num_count(sub_graph_anchors_.size(), 0);
  104. vector<vector<NodePtr>> sub_graph_nodes(sub_graph_anchors_.size());
  105. for (size_t i = 0; i < sub_graph_anchors_.size(); ++i) {
  106. auto nodes_anchor = sub_graph_anchors_[i];
  107. vector<NodePtr> nodes_tmp;
  108. auto iter = nodes_anchor.begin();
  109. auto first_out_anchor = iter->first;
  110. if (first_out_anchor == nullptr) {
  111. continue;
  112. }
  113. nodes_tmp.push_back(first_out_anchor->GetOwnerNode());
  114. while (iter != nodes_anchor.end()) {
  115. auto in_anchor = iter->second;
  116. GE_CHECK_NOTNULL(in_anchor);
  117. auto in_node = in_anchor->GetOwnerNode();
  118. GE_CHECK_NOTNULL(in_node);
  119. if (in_node->GetType() == RESHAPE) {
  120. sub_graph_has_reshape_node[i] = true;
  121. break;
  122. }
  123. if (in_node->GetType() == TRANSPOSE || in_node->GetType() == TRANSPOSED) {
  124. auto input_format = in_node->GetOpDesc()->GetInputDescPtr(0)->GetFormat();
  125. auto output_format = in_node->GetOpDesc()->GetOutputDescPtr(0)->GetFormat();
  126. if (input_format == output_format) {
  127. sub_graph_has_reshape_node[i] = true;
  128. break;
  129. }
  130. }
  131. auto out_anchor = iter->first;
  132. GE_CHECK_NOTNULL(out_anchor);
  133. if (!FormatContinuousCheck(out_anchor, in_anchor)) {
  134. sub_graph_has_reshape_node[i] = true;
  135. break;
  136. }
  137. nodes_tmp.push_back(in_node);
  138. if (IsTransOp(in_node)) {
  139. // count transop num
  140. transop_num_count[i]++;
  141. }
  142. ++iter;
  143. }
  144. sub_graph_nodes[i].swap(nodes_tmp);
  145. if (sub_graph_has_reshape_node[i]) {
  146. SetRemainNode(nodes_anchor);
  147. }
  148. }
  149. sub_graph_has_reshape_node_.swap(sub_graph_has_reshape_node);
  150. transop_num_count_.swap(transop_num_count);
  151. sub_graph_nodes_.swap(sub_graph_nodes);
  152. return GRAPH_SUCCESS;
  153. }
  154. void TransOpWithoutReshapeFusionPass::GetOutDataPeerInControlAnchors(
  155. const size_t index, vector<vector<InControlAnchorPtr>> &out_data_peer_in_control_anchors) {
  156. // The caller guarantees that the index is legal.
  157. for (size_t j = 1; j < sub_graph_anchors_[index].size(); ++j) {
  158. auto nodes_anchor = sub_graph_anchors_[index][j];
  159. auto out_data_anchor = nodes_anchor.first;
  160. GE_CHECK_NOTNULL_JUST_RETURN(out_data_anchor);
  161. for (const auto &peer_in_control_anchor : out_data_anchor->GetPeerInControlAnchors()) {
  162. GE_CHECK_NOTNULL_JUST_RETURN(peer_in_control_anchor);
  163. auto peer_node = peer_in_control_anchor->GetOwnerNode();
  164. if (peer_node == nullptr) {
  165. continue;
  166. }
  167. auto iter = std::find(sub_graph_nodes_[index].begin(), sub_graph_nodes_[index].end(), peer_node);
  168. if (iter == sub_graph_nodes_[index].end()) {
  169. out_data_peer_in_control_anchors[index].push_back(peer_in_control_anchor);
  170. } else {
  171. sub_graph_has_out_data_peer_in_control_edge_[index] = true;
  172. }
  173. }
  174. }
  175. }
  176. void TransOpWithoutReshapeFusionPass::GetInControlPeerOutControlAnchors(
  177. const size_t index, vector<vector<OutControlAnchorPtr>> &in_control_peer_out_control_anchors) {
  178. // The caller guarantees that the index is legal.
  179. for (size_t j = 1; j < (sub_graph_nodes_[index].size() - 1); ++j) {
  180. auto node = sub_graph_nodes_[index][j];
  181. GE_CHECK_NOTNULL_JUST_RETURN(node);
  182. auto in_control_anchor = node->GetInControlAnchor();
  183. if (in_control_anchor == nullptr) {
  184. continue;
  185. }
  186. for (const auto &peer_out_anchor : in_control_anchor->GetPeerOutControlAnchors()) {
  187. GE_CHECK_NOTNULL_JUST_RETURN(peer_out_anchor);
  188. auto peer_node = peer_out_anchor->GetOwnerNode();
  189. if (peer_node == nullptr) {
  190. continue;
  191. }
  192. auto iter = std::find(sub_graph_nodes_[index].begin(), sub_graph_nodes_[index].end(), peer_node);
  193. if (iter == sub_graph_nodes_[index].end()) {
  194. in_control_peer_out_control_anchors[index].push_back(peer_out_anchor);
  195. } else {
  196. sub_graph_has_control_edge_[index] = true;
  197. }
  198. }
  199. }
  200. }
  201. void TransOpWithoutReshapeFusionPass::GetOutControlPeerAnchors(
  202. const size_t index, vector<vector<InControlAnchorPtr>> &out_control_peer_in_control_anchors,
  203. vector<vector<InDataAnchorPtr>> &out_control_peer_in_data_anchors) {
  204. for (size_t j = 0; j < sub_graph_nodes_[index].size() - 1; ++j) {
  205. auto node = sub_graph_nodes_[index][j];
  206. GE_CHECK_NOTNULL_JUST_RETURN(node);
  207. auto out_control_anchor = node->GetOutControlAnchor();
  208. GE_CHECK_NOTNULL_JUST_RETURN(out_control_anchor);
  209. for (const auto &peer_in_anchor : out_control_anchor->GetPeerInControlAnchors()) {
  210. GE_CHECK_NOTNULL_JUST_RETURN(peer_in_anchor);
  211. auto peer_node = peer_in_anchor->GetOwnerNode();
  212. if (peer_node == nullptr) {
  213. continue;
  214. }
  215. auto iter = std::find(sub_graph_nodes_[index].begin(), sub_graph_nodes_[index].end(), peer_node);
  216. if (iter == sub_graph_nodes_[index].end()) {
  217. if (j > 0) {
  218. out_control_peer_in_control_anchors[index].push_back(peer_in_anchor);
  219. }
  220. } else {
  221. sub_graph_has_control_edge_[index] = true;
  222. }
  223. }
  224. for (const auto &peer_in_anchor : out_control_anchor->GetPeerInDataAnchors()) {
  225. GE_CHECK_NOTNULL_JUST_RETURN(peer_in_anchor);
  226. auto peer_node = peer_in_anchor->GetOwnerNode();
  227. if (peer_node == nullptr) {
  228. continue;
  229. }
  230. auto iter = std::find(sub_graph_nodes_[index].begin(), sub_graph_nodes_[index].end(), peer_node);
  231. if (iter == sub_graph_nodes_[index].end()) {
  232. if (j > 0) {
  233. out_control_peer_in_data_anchors[index].push_back(peer_in_anchor);
  234. }
  235. } else {
  236. sub_graph_has_control_edge_[index] = true;
  237. }
  238. }
  239. }
  240. }
  241. void TransOpWithoutReshapeFusionPass::GetControlAnchors() {
  242. vector<vector<OutControlAnchorPtr>> in_control_peer_out_control_anchors(sub_graph_nodes_.size());
  243. vector<vector<InControlAnchorPtr>> out_control_peer_in_control_anchors(sub_graph_nodes_.size());
  244. vector<vector<InDataAnchorPtr>> out_control_peer_in_data_anchors(sub_graph_nodes_.size());
  245. vector<vector<InControlAnchorPtr>> out_data_peer_in_control_anchors(sub_graph_nodes_.size());
  246. vector<bool> sub_graph_has_control_edge(sub_graph_nodes_.size(), false);
  247. sub_graph_has_control_edge_.swap(sub_graph_has_control_edge);
  248. vector<bool> sub_graph_has_out_data_peer_in_control_edge(sub_graph_nodes_.size(), false);
  249. sub_graph_has_out_data_peer_in_control_edge_.swap(sub_graph_has_out_data_peer_in_control_edge);
  250. for (size_t i = 0; i < sub_graph_nodes_.size(); ++i) {
  251. if (sub_graph_has_reshape_node_[i]) {
  252. continue;
  253. }
  254. GetOutDataPeerInControlAnchors(i, out_data_peer_in_control_anchors);
  255. GetInControlPeerOutControlAnchors(i, in_control_peer_out_control_anchors);
  256. GetOutControlPeerAnchors(i, out_control_peer_in_control_anchors, out_control_peer_in_data_anchors);
  257. }
  258. in_control_peer_out_control_anchors_.swap(in_control_peer_out_control_anchors);
  259. out_control_peer_in_control_anchors_.swap(out_control_peer_in_control_anchors);
  260. out_control_peer_in_data_anchors_.swap(out_control_peer_in_data_anchors);
  261. out_data_peer_in_control_anchors_.swap(out_data_peer_in_control_anchors);
  262. }
  263. void TransOpWithoutReshapeFusionPass::EraseInvalidAnchorsPair() {
  264. auto sub_graph_iter = sub_graph_anchors_.begin();
  265. while (sub_graph_iter != sub_graph_anchors_.end()) {
  266. if (sub_graph_iter->size() <= 1) {
  267. sub_graph_iter = sub_graph_anchors_.erase(sub_graph_iter);
  268. } else {
  269. ++sub_graph_iter;
  270. }
  271. }
  272. }
  273. void TransOpWithoutReshapeFusionPass::UpdateOutputName(const OutDataAnchorPtr &out_anchor,
  274. const InDataAnchorPtr &old_peer_in_anchor,
  275. const NodePtr &in_owner_node) {
  276. if (out_anchor == nullptr || old_peer_in_anchor == nullptr || in_owner_node == nullptr) {
  277. GELOGI("out_anchor or old_peer_in_anchor or in_owner_node is nullptr");
  278. return;
  279. }
  280. auto out_owner_node = out_anchor->GetOwnerNode();
  281. GE_CHECK_NOTNULL_JUST_RETURN(out_owner_node);
  282. GE_CHECK_NOTNULL_JUST_RETURN(old_peer_in_anchor->GetOwnerNode());
  283. auto old_peer_in_name = old_peer_in_anchor->GetOwnerNode()->GetName();
  284. auto output_op = out_owner_node->GetOpDesc();
  285. GE_CHECK_NOTNULL_JUST_RETURN(output_op);
  286. auto output_names = output_op->GetAllOutputName();
  287. auto old_peer_in_name_iter = output_names.find(old_peer_in_name);
  288. if (old_peer_in_name_iter != output_names.end()) {
  289. output_names.erase(old_peer_in_name_iter);
  290. }
  291. output_names[in_owner_node->GetName()] = out_anchor->GetIdx();
  292. if (!output_op->UpdateOutputName(output_names)) {
  293. GELOGW("output_op UpdateOutputName failed");
  294. }
  295. }
  296. void TransOpWithoutReshapeFusionPass::UpdateInputName(const OutDataAnchorPtr &old_peer_out_anchor,
  297. const InDataAnchorPtr &in_anchor, const NodePtr &out_owner_node) {
  298. if (old_peer_out_anchor == nullptr || in_anchor == nullptr || out_owner_node == nullptr) {
  299. GELOGI("old_peer_out_anchor or in_anchor or out_owner_node is nullptr");
  300. return;
  301. }
  302. auto old_node = old_peer_out_anchor->GetOwnerNode();
  303. GE_CHECK_NOTNULL_JUST_RETURN(old_node);
  304. auto old_peer_out_name = old_node->GetName();
  305. auto in_owner_node = in_anchor->GetOwnerNode();
  306. GE_CHECK_NOTNULL_JUST_RETURN(in_owner_node);
  307. auto input_op = in_owner_node->GetOpDesc();
  308. GE_CHECK_NOTNULL_JUST_RETURN(input_op);
  309. auto input_names = input_op->GetAllInputName();
  310. auto old_peer_out_name_iter = input_names.find(old_peer_out_name);
  311. if (old_peer_out_name_iter != input_names.end()) {
  312. input_names.erase(old_peer_out_name_iter);
  313. }
  314. input_names[out_owner_node->GetName()] = in_anchor->GetIdx();
  315. input_op->UpdateInputName(input_names);
  316. }
  317. graphStatus TransOpWithoutReshapeFusionPass::RelinkSubGraphControlEdges(
  318. const pair<OutDataAnchorPtr, InDataAnchorPtr> &begin_anchors_pair,
  319. const pair<OutDataAnchorPtr, InDataAnchorPtr> &end_anchors_pair, const int index) {
  320. auto out_anchor = begin_anchors_pair.first;
  321. GE_CHECK_NOTNULL(out_anchor);
  322. auto out_owner_node = out_anchor->GetOwnerNode();
  323. GE_CHECK_NOTNULL(out_owner_node);
  324. auto in_anchor = end_anchors_pair.second;
  325. GE_CHECK_NOTNULL(in_anchor);
  326. auto in_owner_node = in_anchor->GetOwnerNode();
  327. GE_CHECK_NOTNULL(in_owner_node);
  328. if (sub_graph_has_control_edge_[index]) {
  329. GELOGI("add control edge.src:%s, dst:%s", out_owner_node->GetName().c_str(), in_owner_node->GetName().c_str());
  330. if (GraphUtils::AddEdge(out_owner_node->GetOutControlAnchor(), in_owner_node->GetInControlAnchor()) !=
  331. GRAPH_SUCCESS) {
  332. return GRAPH_FAILED;
  333. }
  334. }
  335. if (sub_graph_has_out_data_peer_in_control_edge_[index]) {
  336. GELOGI("add out data 2 in contorl edge.src:%s, dst:%s", out_owner_node->GetName().c_str(),
  337. in_owner_node->GetName().c_str());
  338. if (GraphUtils::AddEdge(out_anchor, in_owner_node->GetInControlAnchor()) != GRAPH_SUCCESS) {
  339. return GRAPH_FAILED;
  340. }
  341. }
  342. return GRAPH_SUCCESS;
  343. }
  344. graphStatus TransOpWithoutReshapeFusionPass::RelinkControlEdgesWhenDescNotChanged(
  345. const pair<OutDataAnchorPtr, InDataAnchorPtr> &begin_anchors_pair,
  346. const pair<OutDataAnchorPtr, InDataAnchorPtr> &end_anchors_pair, const int index) {
  347. if (RelinkSubGraphControlEdges(begin_anchors_pair, end_anchors_pair, index) != GRAPH_SUCCESS) {
  348. return GRAPH_FAILED;
  349. }
  350. auto out_anchor = begin_anchors_pair.first;
  351. GE_CHECK_NOTNULL(out_anchor);
  352. auto out_owner_node = out_anchor->GetOwnerNode();
  353. GE_CHECK_NOTNULL(out_owner_node);
  354. auto in_anchor = end_anchors_pair.second;
  355. GE_CHECK_NOTNULL(in_anchor);
  356. auto in_owner_node = in_anchor->GetOwnerNode();
  357. GE_CHECK_NOTNULL(in_owner_node);
  358. // can not remove old control edge
  359. for (const auto &peer_in_anchor : out_control_peer_in_control_anchors_[index]) {
  360. GE_CHECK_NOTNULL(peer_in_anchor);
  361. GELOGI("add control edge.src:%s, dst:%s, dst idx:%d", out_owner_node->GetName().c_str(),
  362. peer_in_anchor->GetOwnerNode()->GetName().c_str(), peer_in_anchor->GetIdx());
  363. if (GraphUtils::AddEdge(out_owner_node->GetOutControlAnchor(), peer_in_anchor) != GRAPH_SUCCESS) {
  364. return GRAPH_FAILED;
  365. }
  366. }
  367. for (const auto &peer_out_anchor : in_control_peer_out_control_anchors_[index]) {
  368. GE_CHECK_NOTNULL(peer_out_anchor);
  369. GELOGI("add control edge.src:%s, src idx:%d, dst:%s", peer_out_anchor->GetOwnerNode()->GetName().c_str(),
  370. peer_out_anchor->GetIdx(), in_owner_node->GetName().c_str());
  371. if (GraphUtils::AddEdge(peer_out_anchor, in_owner_node->GetInControlAnchor()) != GRAPH_SUCCESS) {
  372. return GRAPH_FAILED;
  373. }
  374. }
  375. for (const auto &peer_in_anchor : out_control_peer_in_data_anchors_[index]) {
  376. GE_CHECK_NOTNULL(peer_in_anchor);
  377. GELOGI("add out control 2 in data edge.src:%s, dst:%s, dst idx:%d", out_owner_node->GetName().c_str(),
  378. peer_in_anchor->GetOwnerNode()->GetName().c_str(), peer_in_anchor->GetIdx());
  379. if (GraphUtils::AddEdge(out_owner_node->GetOutControlAnchor(), peer_in_anchor) != GRAPH_SUCCESS) {
  380. return GRAPH_FAILED;
  381. }
  382. }
  383. for (const auto &peer_in_anchor : out_data_peer_in_control_anchors_[index]) {
  384. GE_CHECK_NOTNULL(peer_in_anchor);
  385. GELOGI("add out data 2 in control edge.src:%s, dst:%s, dst idx:%d", out_owner_node->GetName().c_str(),
  386. peer_in_anchor->GetOwnerNode()->GetName().c_str(), peer_in_anchor->GetIdx());
  387. if (GraphUtils::AddEdge(out_anchor, peer_in_anchor) != GRAPH_SUCCESS) {
  388. return GRAPH_FAILED;
  389. }
  390. }
  391. return GRAPH_SUCCESS;
  392. }
  393. graphStatus TransOpWithoutReshapeFusionPass::RelinkNodesWhenDescNotChanged(
  394. const pair<OutDataAnchorPtr, InDataAnchorPtr> &begin_anchors_pair,
  395. const pair<OutDataAnchorPtr, InDataAnchorPtr> &end_anchors_pair, const int index) {
  396. auto out_anchor = begin_anchors_pair.first;
  397. GE_CHECK_NOTNULL(out_anchor);
  398. auto out_owner_node = out_anchor->GetOwnerNode();
  399. GE_CHECK_NOTNULL(out_owner_node);
  400. auto in_anchor = end_anchors_pair.second;
  401. GE_CHECK_NOTNULL(in_anchor);
  402. auto in_owner_node = in_anchor->GetOwnerNode();
  403. GE_CHECK_NOTNULL(in_owner_node);
  404. GELOGI("remove edge.src %s, src idx:%d, dst:%s, dst idx:%d",
  405. end_anchors_pair.first->GetOwnerNode()->GetName().c_str(), end_anchors_pair.first->GetIdx(),
  406. in_owner_node->GetName().c_str(), in_anchor->GetIdx());
  407. GE_CHK_STATUS_RET(GraphUtils::RemoveEdge(end_anchors_pair.first, in_anchor), "remove edge failed");
  408. GELOGI("relink node.src node:%s, src idx:%d, dst node:%s, dst idx:%d", out_owner_node->GetName().c_str(),
  409. out_anchor->GetIdx(), in_owner_node->GetName().c_str(), in_anchor->GetIdx());
  410. if (GraphUtils::AddEdge(out_anchor, in_anchor) != GRAPH_SUCCESS) {
  411. GELOGE(GRAPH_FAILED, "add edge failed!src:%s, src idx:%d, dst:%s, dst idx:%d", out_owner_node->GetName().c_str(),
  412. out_anchor->GetIdx(), in_owner_node->GetName().c_str(), in_anchor->GetIdx());
  413. return GRAPH_FAILED;
  414. } else {
  415. auto old_peer_in_anchor = begin_anchors_pair.second;
  416. UpdateOutputName(out_anchor, old_peer_in_anchor, in_owner_node);
  417. auto old_peer_out_anchor = end_anchors_pair.first;
  418. UpdateInputName(old_peer_out_anchor, in_anchor, out_owner_node);
  419. }
  420. return RelinkControlEdgesWhenDescNotChanged(begin_anchors_pair, end_anchors_pair, index);
  421. }
  422. OpDescPtr TransOpWithoutReshapeFusionPass::GetFormatTransferOp(const GeTensorDesc &format_trans_input_desc,
  423. const GeTensorDesc &format_trans_output_desc) {
  424. static std::atomic_long atomic_fusion_format_transfer_op_count(1);
  425. auto fusion_format_transfer_op_count = atomic_fusion_format_transfer_op_count.fetch_add(1);
  426. std::stringstream format_transfer_op_name;
  427. format_transfer_op_name << "fusion_format_transfer_" << fusion_format_transfer_op_count;
  428. OpDescPtr format_transfer_op = MakeShared<OpDesc>(format_transfer_op_name.str().c_str(), TRANSDATA);
  429. if (format_transfer_op == nullptr) {
  430. GELOGE(INTERNAL_ERROR, "new format transfer op failed!");
  431. return nullptr;
  432. }
  433. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(format_transfer_op, ATTR_NAME_INPUT_FORMAT,
  434. static_cast<int64_t>(format_trans_input_desc.GetFormat())),
  435. GELOGE(INTERNAL_ERROR, "set ATTR_NAME_INPUT_FORMAT failed");
  436. return nullptr);
  437. GE_IF_BOOL_EXEC(!AttrUtils::SetInt(format_transfer_op, ATTR_NAME_OUTPUT_FORMAT,
  438. static_cast<int64_t>(format_trans_output_desc.GetFormat())),
  439. GELOGE(INTERNAL_ERROR, "set ATTR_NAME_OUTPUT_FORMAT failed");
  440. return nullptr);
  441. string src_format = TypeUtils::FormatToSerialString(format_trans_input_desc.GetFormat());
  442. string dst_format = TypeUtils::FormatToSerialString(format_trans_output_desc.GetFormat());
  443. GE_IF_BOOL_EXEC(!AttrUtils::SetStr(format_transfer_op, kAttrNameSrcFormat, src_format),
  444. GELOGE(INTERNAL_ERROR, "set kAttrNameSrcFormat failed");
  445. return nullptr);
  446. GE_IF_BOOL_EXEC(!AttrUtils::SetStr(format_transfer_op, kAttrNameDstFormat, dst_format),
  447. GELOGE(INTERNAL_ERROR, "set kAttrNameDstFormat failed");
  448. return nullptr);
  449. GE_IF_BOOL_EXEC(format_transfer_op->AddInputDesc(format_trans_input_desc) != GRAPH_SUCCESS,
  450. GELOGE(INTERNAL_ERROR, "add input desc failed");
  451. return nullptr);
  452. GE_IF_BOOL_EXEC(format_transfer_op->AddOutputDesc(format_trans_output_desc) != GRAPH_SUCCESS,
  453. GELOGE(INTERNAL_ERROR, "add output desc failed");
  454. return nullptr);
  455. GE_IF_BOOL_EXEC(!ge::AttrUtils::SetBool(format_transfer_op, ATTR_NEED_COMPILE, true),
  456. GELOGE(INTERNAL_ERROR, "set ext attr failed");
  457. return nullptr);
  458. return format_transfer_op;
  459. }
  460. OpDescPtr TransOpWithoutReshapeFusionPass::GetCastOp(const GeTensorDesc &cast_input_desc,
  461. const GeTensorDesc &cast_output_desc) {
  462. static std::atomic_long atomic_fusion_cast_op_count(1);
  463. auto fusion_cast_op_count = atomic_fusion_cast_op_count.fetch_add(1);
  464. std::stringstream cast_op_name;
  465. cast_op_name << "fusion_cast_op_" << fusion_cast_op_count;
  466. auto node_op = ge::OperatorFactory::CreateOperator(cast_op_name.str(), CAST);
  467. auto cast_op = ge::OpDescUtils::GetOpDescFromOperator(node_op);
  468. node_op.BreakConnect();
  469. if (cast_op == nullptr) {
  470. GELOGE(INTERNAL_ERROR, "new cast op failed!");
  471. return nullptr;
  472. }
  473. const int default_input_index = 0;
  474. const int default_output_index = 0;
  475. if (cast_op->GetInputsSize() == 0) {
  476. GE_IF_BOOL_EXEC(cast_op->AddInputDesc(cast_input_desc) != GRAPH_SUCCESS,
  477. GELOGE(INTERNAL_ERROR, "add input desc failed");
  478. return nullptr);
  479. } else {
  480. GE_IF_BOOL_EXEC(cast_op->UpdateInputDesc(default_input_index, cast_input_desc) != GRAPH_SUCCESS,
  481. GELOGE(INTERNAL_ERROR, "update input desc failed");
  482. return nullptr);
  483. }
  484. if (cast_op->GetOutputsSize() == 0) {
  485. GE_IF_BOOL_EXEC(cast_op->AddOutputDesc(cast_output_desc) != GRAPH_SUCCESS,
  486. GELOGE(INTERNAL_ERROR, "add output desc failed");
  487. return nullptr);
  488. } else {
  489. GE_IF_BOOL_EXEC(cast_op->UpdateOutputDesc(default_output_index, cast_output_desc) != GRAPH_SUCCESS,
  490. GELOGE(INTERNAL_ERROR, "update output desc failed");
  491. return nullptr);
  492. }
  493. if (!AttrUtils::SetInt(cast_op, CAST_ATTR_DST_TYPE, static_cast<int64_t>(cast_output_desc.GetDataType()))) {
  494. GELOGE(INTERNAL_ERROR, "set dst_type attr failed");
  495. return nullptr;
  496. }
  497. if (!AttrUtils::SetBool(cast_op, ATTR_NEED_COMPILE, true)) {
  498. GELOGE(INTERNAL_ERROR, "set need_compile attr failed");
  499. return nullptr;
  500. }
  501. return cast_op;
  502. }
  503. bool TransOpWithoutReshapeFusionPass::InsertCastFirstCheck(const GeTensorDesc &out_desc,
  504. const GeTensorDesc &in_desc) const {
  505. return out_desc.GetDataType() != in_desc.GetDataType() && out_desc.GetDataType() != DT_FLOAT16 &&
  506. in_desc.GetDataType() == DT_FLOAT16;
  507. }
  508. void TransOpWithoutReshapeFusionPass::GetFormatTransferDesc(const GeTensorDesc &out_desc, const GeTensorDesc &in_desc,
  509. GeTensorDesc &format_transfer_input,
  510. GeTensorDesc &format_transfer_output) {
  511. bool insert_cast_first = InsertCastFirstCheck(out_desc, in_desc);
  512. if (insert_cast_first) {
  513. format_transfer_input = out_desc;
  514. format_transfer_input.SetDataType(in_desc.GetDataType());
  515. format_transfer_output = in_desc;
  516. } else {
  517. format_transfer_input = out_desc;
  518. format_transfer_output = in_desc;
  519. format_transfer_output.SetDataType(out_desc.GetDataType());
  520. }
  521. }
  522. void TransOpWithoutReshapeFusionPass::GetCastOpDesc(const GeTensorDesc &out_desc, const GeTensorDesc &in_desc,
  523. GeTensorDesc &cast_input, GeTensorDesc &cast_output) {
  524. bool insert_cast_first = InsertCastFirstCheck(out_desc, in_desc);
  525. if (insert_cast_first) {
  526. cast_input = out_desc;
  527. cast_output = out_desc;
  528. cast_output.SetDataType(in_desc.GetDataType());
  529. } else {
  530. cast_input = in_desc;
  531. cast_input.SetDataType(out_desc.GetDataType());
  532. cast_output = in_desc;
  533. }
  534. }
  535. void TransOpWithoutReshapeFusionPass::GetBeginOutDescAndEndInDesc(const int index, GeTensorDesc &out_desc,
  536. GeTensorDesc &in_desc) {
  537. auto nodes_anchor = sub_graph_anchors_[index];
  538. auto out_peer_anchor = nodes_anchor.front().second;
  539. GE_CHECK_NOTNULL_JUST_RETURN(out_peer_anchor);
  540. auto out_owner_node = out_peer_anchor->GetOwnerNode();
  541. GE_CHECK_NOTNULL_JUST_RETURN(out_owner_node);
  542. auto out_peer_op_desc = out_owner_node->GetOpDesc();
  543. GE_IF_BOOL_EXEC(out_peer_op_desc == nullptr, GELOGE(INTERNAL_ERROR, "out_peer_op_desc is nullptr"); return);
  544. out_desc = out_peer_op_desc->GetInputDesc(out_peer_anchor->GetIdx());
  545. auto in_peer_anchor = nodes_anchor.back().first;
  546. GE_CHECK_NOTNULL_JUST_RETURN(in_peer_anchor);
  547. auto in_owner_node = in_peer_anchor->GetOwnerNode();
  548. GE_CHECK_NOTNULL_JUST_RETURN(in_owner_node);
  549. auto in_peer_op_desc = in_owner_node->GetOpDesc();
  550. GE_IF_BOOL_EXEC(in_peer_op_desc == nullptr, GELOGE(INTERNAL_ERROR, "in_peer_op_desc is nullptr"); return);
  551. in_desc = in_peer_op_desc->GetOutputDesc(in_peer_anchor->GetIdx());
  552. }
  553. graphStatus TransOpWithoutReshapeFusionPass::FormatFusion(const int index, OpDescPtr &format_transfer_op,
  554. int32_t &fusion_op_count, bool &fusion_continue) {
  555. GeTensorDesc out_desc;
  556. GeTensorDesc in_desc;
  557. GetBeginOutDescAndEndInDesc(index, out_desc, in_desc);
  558. GeTensorDesc format_transfer_input;
  559. GeTensorDesc format_transfer_output;
  560. GetFormatTransferDesc(out_desc, in_desc, format_transfer_input, format_transfer_output);
  561. if (out_desc.GetFormat() == in_desc.GetFormat() &&
  562. (!ShapeEqualCheck(out_desc.GetShape(), in_desc.GetShape()) ||
  563. !ShapeEqualCheck(out_desc.GetOriginShape(), in_desc.GetOriginShape()))) {
  564. SetRemainNode(sub_graph_anchors_[index]);
  565. return GRAPH_SUCCESS;
  566. }
  567. if (out_desc.GetFormat() != in_desc.GetFormat() && FusionFormatSupport(out_desc.GetFormat()) &&
  568. FusionFormatSupport(in_desc.GetFormat())) {
  569. // create format transop
  570. format_transfer_op = GetFormatTransferOp(format_transfer_input, format_transfer_output);
  571. if (format_transfer_op == nullptr) {
  572. return GRAPH_FAILED;
  573. }
  574. if (OpAccuracyAbilityCheck(format_transfer_op)) {
  575. ++fusion_op_count;
  576. GELOGI("support format transfer op %s", format_transfer_op->GetName().c_str());
  577. } else {
  578. GELOGW("ability not support.src format:%d, src datatype:%d, dst format:%d, dst datatype:%d",
  579. format_transfer_input.GetFormat(), format_transfer_input.GetDataType(), format_transfer_output.GetFormat(),
  580. format_transfer_output.GetDataType());
  581. fusion_op_count = kInvalidFusionOpCount;
  582. }
  583. } else if (out_desc.GetFormat() != in_desc.GetFormat()) {
  584. SetRemainNode(sub_graph_anchors_[index]);
  585. return GRAPH_SUCCESS;
  586. }
  587. fusion_continue = true;
  588. return GRAPH_SUCCESS;
  589. }
  590. graphStatus TransOpWithoutReshapeFusionPass::DataTypeFusion(const int index, OpDescPtr &cast_op,
  591. int32_t &fusion_op_count) {
  592. GeTensorDesc out_desc;
  593. GeTensorDesc in_desc;
  594. GetBeginOutDescAndEndInDesc(index, out_desc, in_desc);
  595. GeTensorDesc cast_input;
  596. GeTensorDesc cast_output;
  597. GetCastOpDesc(out_desc, in_desc, cast_input, cast_output);
  598. if (fusion_op_count != kInvalidFusionOpCount && out_desc.GetDataType() != in_desc.GetDataType()) {
  599. // create cast op
  600. cast_op = GetCastOp(cast_input, cast_output);
  601. if (cast_op == nullptr) {
  602. fusion_op_count = kInvalidFusionOpCount;
  603. return GRAPH_FAILED;
  604. }
  605. if (OpAccuracyAbilityCheck(cast_op)) {
  606. ++fusion_op_count;
  607. GELOGI("support cast op %s. src format:%d, src datatype:%d, dst format:%d, dst datatype:%d",
  608. cast_op->GetName().c_str(), cast_input.GetFormat(), cast_input.GetDataType(), cast_output.GetFormat(),
  609. cast_output.GetDataType());
  610. } else {
  611. GELOGW("ability not support.src format:%d, src datatype:%d, dst format:%d, dst datatype:%d",
  612. cast_input.GetFormat(), cast_input.GetDataType(), cast_output.GetFormat(), cast_output.GetDataType());
  613. fusion_op_count = kInvalidFusionOpCount;
  614. }
  615. }
  616. return GRAPH_SUCCESS;
  617. }
  618. graphStatus TransOpWithoutReshapeFusionPass::TransOpFuseHandle(const ComputeGraphPtr &graph, const int index) {
  619. bool fusion_continue = false;
  620. OpDescPtr format_transfer_op = nullptr;
  621. int32_t fusion_op_count = 0;
  622. auto fortmat_fusion_ret = FormatFusion(index, format_transfer_op, fusion_op_count, fusion_continue);
  623. if (fortmat_fusion_ret != GRAPH_SUCCESS || !fusion_continue) {
  624. SetRemainNode(sub_graph_anchors_[index]);
  625. return GRAPH_SUCCESS;
  626. }
  627. OpDescPtr cast_op = nullptr;
  628. if (DataTypeFusion(index, cast_op, fusion_op_count) != GRAPH_SUCCESS) {
  629. SetRemainNode(sub_graph_anchors_[index]);
  630. return GRAPH_SUCCESS;
  631. }
  632. if (fusion_op_count != kInvalidFusionOpCount && fusion_op_count < transop_num_count_[index]) {
  633. GeTensorDesc out_desc;
  634. GeTensorDesc in_desc;
  635. GetBeginOutDescAndEndInDesc(index, out_desc, in_desc);
  636. bool insert_cast_first = InsertCastFirstCheck(out_desc, in_desc);
  637. if (InsertNewTransOp(graph, cast_op, format_transfer_op, index, insert_cast_first) != GRAPH_SUCCESS) {
  638. return GRAPH_FAILED;
  639. }
  640. } else {
  641. // remain all nodes
  642. SetRemainNode(sub_graph_anchors_[index]);
  643. }
  644. return GRAPH_SUCCESS;
  645. }
  646. void TransOpWithoutReshapeFusionPass::RemoveNousedNodes(const ComputeGraphPtr &graph) {
  647. if (graph == nullptr) {
  648. return;
  649. }
  650. for (size_t i = 0; i < sub_graph_nodes_.size(); ++i) {
  651. if (sub_graph_has_reshape_node_[i]) {
  652. continue;
  653. }
  654. for (const auto &node : sub_graph_nodes_[i]) {
  655. GE_CHECK_NOTNULL_JUST_RETURN(node);
  656. // remove nodes
  657. if (!IsTransOp(node)) {
  658. continue;
  659. }
  660. auto op_desc = node->GetOpDesc();
  661. GE_CHECK_NOTNULL_JUST_RETURN(op_desc);
  662. bool node_remain_flag = op_desc->TryGetExtAttr(kRemainNode, false);
  663. if (node_remain_flag) {
  664. continue;
  665. }
  666. GE_IF_BOOL_EXEC(!op_desc->SetExtAttr(kRemainNode, true), GELOGE(INTERNAL_ERROR, "set ext attr failed"); return);
  667. GELOGI("remove node:%s", node->GetName().c_str());
  668. if (GraphUtils::IsolateNode(node, {0}) != GRAPH_SUCCESS) {
  669. GELOGW("Isolate node: %s failed.", node->GetName().c_str());
  670. continue;
  671. }
  672. if (GraphUtils::RemoveNodeWithoutRelink(graph, node) != GRAPH_SUCCESS) {
  673. GELOGW("Remove node: %s failed.", node->GetName().c_str());
  674. continue;
  675. }
  676. }
  677. }
  678. }
  679. graphStatus TransOpWithoutReshapeFusionPass::Run(ComputeGraphPtr graph) {
  680. GELOGI("[TransOpWithoutReshapeFusionPass]: optimize begin.");
  681. if (graph == nullptr) {
  682. return GRAPH_SUCCESS;
  683. }
  684. for (const auto &node : graph->GetDirectNode()) {
  685. GE_CHECK_NOTNULL(node);
  686. if (IsTransOp(node)) {
  687. continue;
  688. }
  689. bool is_unknown = false;
  690. auto ret = NodeUtils::GetNodeUnknownShapeStatus(*node, is_unknown);
  691. if (ret != GRAPH_SUCCESS) {
  692. GELOGW("Get node unknown status failed, node name:%s, type:%s.", node->GetName().c_str(),
  693. node->GetType().c_str());
  694. continue;
  695. }
  696. if (is_unknown) {
  697. GELOGI("Current node %s, type %s is unknown shape which should be skip.", node->GetName().c_str(),
  698. node->GetType().c_str());
  699. continue;
  700. }
  701. GELOGI("Current normal node name: %s, type: %s.", node->GetName().c_str(), node->GetType().c_str());
  702. for (const auto &out_anchor : node->GetAllOutDataAnchors()) {
  703. GE_CHECK_NOTNULL(out_anchor);
  704. vector<vector<pair<OutDataAnchorPtr, InDataAnchorPtr>>> sub_graph_anchors;
  705. vector<pair<OutDataAnchorPtr, InDataAnchorPtr>> nodes_list;
  706. if (GetSubGraphsBetweenNormalNode(out_anchor, sub_graph_anchors, nodes_list) != GRAPH_SUCCESS) {
  707. GELOGW("get transops failed!");
  708. continue;
  709. }
  710. sub_graph_anchors_.swap(sub_graph_anchors);
  711. EraseInvalidAnchorsPair();
  712. if (sub_graph_anchors_.empty()) {
  713. continue;
  714. }
  715. // check reshape node
  716. if (GetSubGraphNodesInfo() != GRAPH_SUCCESS) {
  717. continue;
  718. }
  719. // save control edge
  720. GetControlAnchors();
  721. if (TransOpFuse(graph) != GRAPH_SUCCESS) {
  722. return GRAPH_FAILED;
  723. }
  724. }
  725. }
  726. GELOGI("[TransOpWithoutReshapeFusionPass]: Optimize end.");
  727. return GRAPH_SUCCESS;
  728. }
  729. bool TransOpWithoutReshapeFusionPass::DescEqualCheck(ConstGeTensorDescPtr &desc_src,
  730. ConstGeTensorDescPtr &desc_dst) const {
  731. if (desc_src == nullptr || desc_dst == nullptr) {
  732. return false;
  733. }
  734. if (desc_src->GetFormat() != desc_dst->GetFormat() || desc_src->GetDataType() != desc_dst->GetDataType()) {
  735. return false;
  736. }
  737. if (!ShapeEqualCheck(desc_src->GetShape(), desc_dst->GetShape())) {
  738. return false;
  739. }
  740. return ShapeEqualCheck(desc_src->GetOriginShape(), desc_dst->GetOriginShape());
  741. }
  742. bool TransOpWithoutReshapeFusionPass::ShapeEqualCheck(const GeShape &src, const GeShape &dst) const {
  743. if (src.GetDims().size() != dst.GetDims().size()) {
  744. return false;
  745. }
  746. for (size_t i = 0; i < src.GetDims().size(); ++i) {
  747. if (src.GetDim(i) != dst.GetDim(i)) {
  748. return false;
  749. }
  750. }
  751. return true;
  752. }
  753. graphStatus TransOpWithoutReshapeFusionPass::TransOpFuse(const ComputeGraphPtr &graph) {
  754. for (size_t i = 0; i < sub_graph_anchors_.size(); ++i) {
  755. if (sub_graph_has_reshape_node_[i]) {
  756. continue;
  757. }
  758. auto nodes_anchor = sub_graph_anchors_[i];
  759. auto out_anchor = nodes_anchor.front().first;
  760. GE_CHECK_NOTNULL(out_anchor);
  761. auto out_op_desc = out_anchor->GetOwnerNode()->GetOpDesc();
  762. GE_CHECK_NOTNULL(out_op_desc);
  763. auto out_desc = out_op_desc->GetOutputDescPtr(out_anchor->GetIdx());
  764. GE_CHECK_NOTNULL(out_desc);
  765. auto in_anchor = nodes_anchor.back().second;
  766. GE_CHECK_NOTNULL(in_anchor);
  767. auto in_op_desc = in_anchor->GetOwnerNode()->GetOpDesc();
  768. GE_CHECK_NOTNULL(in_op_desc);
  769. auto in_desc = in_op_desc->GetInputDescPtr(in_anchor->GetIdx());
  770. GE_CHECK_NOTNULL(in_desc);
  771. if (FusionFormatSupport(out_desc->GetFormat()) && DescEqualCheck(out_desc, in_desc)) {
  772. // relink begin_out to end_in
  773. if (RelinkNodesWhenDescNotChanged(nodes_anchor.front(), nodes_anchor.back(), static_cast<int>(i)) !=
  774. GRAPH_SUCCESS) {
  775. return GRAPH_FAILED;
  776. }
  777. } else {
  778. if (TransOpFuseHandle(graph, static_cast<int>(i)) != GRAPH_SUCCESS) {
  779. return GRAPH_FAILED;
  780. }
  781. }
  782. }
  783. RemoveNousedNodes(graph);
  784. return GRAPH_SUCCESS;
  785. }
  786. graphStatus TransOpWithoutReshapeFusionPass::AddTransNode(const ComputeGraphPtr &graph, const OpDescPtr &transop,
  787. NodePtr &trans_node) {
  788. if (graph == nullptr) {
  789. return GRAPH_SUCCESS;
  790. }
  791. if (transop == nullptr) {
  792. return GRAPH_SUCCESS;
  793. }
  794. trans_node = graph->AddNode(transop);
  795. if (trans_node == nullptr) {
  796. GELOGE(GRAPH_FAILED, "add node failed!");
  797. return GRAPH_FAILED;
  798. }
  799. return GRAPH_SUCCESS;
  800. }
  801. graphStatus TransOpWithoutReshapeFusionPass::GetTransNode(const ComputeGraphPtr &graph, const OpDescPtr &cast_op,
  802. const OpDescPtr &format_transfer_op,
  803. const bool insert_cast_first,
  804. std::vector<NodePtr> &new_trans_nodes) {
  805. NodePtr format_transfer_node;
  806. if (AddTransNode(graph, format_transfer_op, format_transfer_node) != GRAPH_SUCCESS) {
  807. return GRAPH_FAILED;
  808. }
  809. NodePtr cast_node;
  810. if (AddTransNode(graph, cast_op, cast_node) != GRAPH_SUCCESS) {
  811. return GRAPH_FAILED;
  812. }
  813. if (insert_cast_first) {
  814. if (cast_node != nullptr) {
  815. new_trans_nodes.push_back(cast_node);
  816. }
  817. if (format_transfer_node != nullptr) {
  818. new_trans_nodes.push_back(format_transfer_node);
  819. }
  820. } else {
  821. if (format_transfer_node != nullptr) {
  822. new_trans_nodes.push_back(format_transfer_node);
  823. }
  824. if (cast_node != nullptr) {
  825. new_trans_nodes.push_back(cast_node);
  826. }
  827. }
  828. return GRAPH_SUCCESS;
  829. }
  830. graphStatus TransOpWithoutReshapeFusionPass::InsertNewTransOp(const ComputeGraphPtr &graph, const OpDescPtr &cast_op,
  831. const OpDescPtr &format_transfer_op, const int index,
  832. const bool insert_cast_first) {
  833. std::vector<NodePtr> new_trans_nodes;
  834. if (GetTransNode(graph, cast_op, format_transfer_op, insert_cast_first, new_trans_nodes) != GRAPH_SUCCESS) {
  835. return GRAPH_FAILED;
  836. }
  837. if (new_trans_nodes.empty()) {
  838. GELOGI("No new trans node. Do not need insert new transop.");
  839. return GRAPH_SUCCESS;
  840. }
  841. pair<OutDataAnchorPtr, InDataAnchorPtr> begin_out = sub_graph_anchors_[index].front();
  842. pair<OutDataAnchorPtr, InDataAnchorPtr> end_in = sub_graph_anchors_[index].back();
  843. auto out_anchor = begin_out.first;
  844. GE_CHECK_NOTNULL(out_anchor);
  845. auto out_owner_node = out_anchor->GetOwnerNode();
  846. GE_CHECK_NOTNULL(out_owner_node);
  847. auto in_anchor = end_in.second;
  848. GE_CHECK_NOTNULL(in_anchor);
  849. auto in_owner_node = in_anchor->GetOwnerNode();
  850. GE_CHECK_NOTNULL(in_owner_node);
  851. GELOGI("remove edge.src:%s, src idx:%d, dst:%s, dst idx:%d", end_in.first->GetOwnerNode()->GetName().c_str(),
  852. end_in.first->GetIdx(), in_anchor->GetOwnerNode()->GetName().c_str(), in_anchor->GetIdx());
  853. GE_CHK_STATUS_RET(GraphUtils::RemoveEdge(end_in.first, in_anchor), "remove edge failed");
  854. GELOGI("add edge.src:%s, src idx:%d, dst:%s", out_anchor->GetOwnerNode()->GetName().c_str(), out_anchor->GetIdx(),
  855. new_trans_nodes.front()->GetName().c_str());
  856. if (GraphUtils::AddEdge(out_anchor, new_trans_nodes.front()->GetInAnchor(0)) != GRAPH_SUCCESS) {
  857. return GRAPH_FAILED;
  858. } else {
  859. auto old_peer_in_anchor = begin_out.second;
  860. GE_CHECK_NOTNULL(old_peer_in_anchor);
  861. UpdateOutputName(out_anchor, old_peer_in_anchor, in_owner_node);
  862. }
  863. if (new_trans_nodes.size() > 1) {
  864. GELOGI("add edge.src:%s, dst:%s", new_trans_nodes.front()->GetName().c_str(),
  865. new_trans_nodes.back()->GetName().c_str());
  866. if (GraphUtils::AddEdge(new_trans_nodes.front()->GetOutAnchor(0), new_trans_nodes.back()->GetInAnchor(0)) !=
  867. GRAPH_SUCCESS) {
  868. return GRAPH_FAILED;
  869. } else {
  870. auto old_peer_out_anchor = end_in.first;
  871. GE_CHECK_NOTNULL(old_peer_out_anchor);
  872. UpdateInputName(old_peer_out_anchor, in_anchor, out_owner_node);
  873. }
  874. }
  875. GELOGI("add edge.src:%s, dst:%s, dst idx:%d", new_trans_nodes.back()->GetName().c_str(),
  876. in_anchor->GetOwnerNode()->GetName().c_str(), in_anchor->GetIdx());
  877. if (GraphUtils::AddEdge(new_trans_nodes.back()->GetOutAnchor(0), in_anchor) != GRAPH_SUCCESS) {
  878. return GRAPH_FAILED;
  879. }
  880. return RelinkControlEdge(index, out_anchor, new_trans_nodes);
  881. }
  882. graphStatus TransOpWithoutReshapeFusionPass::RelinkControlEdge(const int index, const OutDataAnchorPtr &out_anchor,
  883. const vector<NodePtr> &new_trans_nodes) {
  884. GE_CHECK_NOTNULL(out_anchor);
  885. if (new_trans_nodes.front() == nullptr || new_trans_nodes.back() == nullptr) {
  886. return GRAPH_FAILED;
  887. }
  888. if (sub_graph_has_control_edge_[index]) {
  889. GELOGI("add control edge.src:%s, dst:%s", out_anchor->GetOwnerNode()->GetName().c_str(),
  890. new_trans_nodes.front()->GetName().c_str());
  891. if (GraphUtils::AddEdge(out_anchor->GetOwnerNode()->GetOutControlAnchor(),
  892. new_trans_nodes.front()->GetInControlAnchor()) != GRAPH_SUCCESS) {
  893. return GRAPH_FAILED;
  894. }
  895. }
  896. for (const auto &peer_in_anchor : out_control_peer_in_control_anchors_[index]) {
  897. GE_CHECK_NOTNULL(peer_in_anchor);
  898. GELOGI("add control edge.src:%s, dst:%s", new_trans_nodes.back()->GetName().c_str(),
  899. peer_in_anchor->GetOwnerNode()->GetName().c_str());
  900. if (GraphUtils::AddEdge(new_trans_nodes.back()->GetOutControlAnchor(), peer_in_anchor) != GRAPH_SUCCESS) {
  901. return GRAPH_FAILED;
  902. }
  903. }
  904. for (const auto &peer_out_anchor : in_control_peer_out_control_anchors_[index]) {
  905. GE_CHECK_NOTNULL(peer_out_anchor);
  906. GELOGI("add control edge.src:%s, dst:%s", peer_out_anchor->GetOwnerNode()->GetName().c_str(),
  907. new_trans_nodes.front()->GetName().c_str());
  908. if (GraphUtils::AddEdge(peer_out_anchor, new_trans_nodes.front()->GetInControlAnchor()) != GRAPH_SUCCESS) {
  909. return GRAPH_FAILED;
  910. }
  911. }
  912. for (const auto &peer_in_anchor : out_control_peer_in_data_anchors_[index]) {
  913. GE_CHECK_NOTNULL(peer_in_anchor);
  914. GELOGI("add control edge.src:%s, dst:%s", new_trans_nodes.back()->GetName().c_str(),
  915. peer_in_anchor->GetOwnerNode()->GetName().c_str());
  916. if (GraphUtils::AddEdge(new_trans_nodes.back()->GetOutControlAnchor(), peer_in_anchor) != GRAPH_SUCCESS) {
  917. return GRAPH_FAILED;
  918. }
  919. }
  920. for (const auto &peer_in_anchor : out_data_peer_in_control_anchors_[index]) {
  921. GE_CHECK_NOTNULL(peer_in_anchor);
  922. GELOGI("add control edge.src:%s, dst:%s", new_trans_nodes.back()->GetName().c_str(),
  923. peer_in_anchor->GetOwnerNode()->GetName().c_str());
  924. if (GraphUtils::AddEdge(new_trans_nodes.back()->GetOutDataAnchor(0), peer_in_anchor) != GRAPH_SUCCESS) {
  925. return GRAPH_FAILED;
  926. }
  927. }
  928. if (sub_graph_has_out_data_peer_in_control_edge_[index]) {
  929. auto in_anchor = sub_graph_anchors_[index].back().second;
  930. GELOGI("add control edge.src:%s, dst:%s", new_trans_nodes.back()->GetName().c_str(),
  931. in_anchor->GetOwnerNode()->GetName().c_str());
  932. if (GraphUtils::AddEdge(new_trans_nodes.back()->GetOutDataAnchor(0),
  933. in_anchor->GetOwnerNode()->GetInControlAnchor()) != GRAPH_SUCCESS) {
  934. return GRAPH_FAILED;
  935. }
  936. }
  937. return GRAPH_SUCCESS;
  938. }
  939. bool TransOpWithoutReshapeFusionPass::OpAccuracyAbilityCheck(const OpDescPtr &op_desc) {
  940. auto instance = GELib::GetInstance();
  941. if ((instance == nullptr) || (!instance->InitFlag())) {
  942. GELOGW("GELib is not initialized!");
  943. return false;
  944. }
  945. if (op_desc == nullptr) {
  946. return false;
  947. }
  948. OpsKernelManager &ops_kernel_manager = instance->OpsKernelManagerObj();
  949. vector<OpInfo> op_infos = ops_kernel_manager.GetOpsKernelInfo(op_desc->GetType());
  950. if (op_infos.empty()) {
  951. GELOGI("Can not get op info by op type:%s", op_desc->GetType().c_str());
  952. return false;
  953. }
  954. std::string unsupported_reason;
  955. for (const auto &it : op_infos) {
  956. auto kernel_map = ops_kernel_manager.GetAllOpsKernelInfoStores();
  957. auto &kernel_name = it.opKernelLib;
  958. auto kernel_info_store = kernel_map.find(kernel_name);
  959. if (kernel_info_store != kernel_map.end()) {
  960. if (kernel_info_store->second != nullptr &&
  961. kernel_info_store->second->CheckAccuracySupported(op_desc, unsupported_reason)) {
  962. op_desc->SetOpEngineName(it.engine);
  963. op_desc->SetOpKernelLibName(kernel_name);
  964. GELOGI("Set OpKernelLibName %s and engine name %s into op_desc %s", kernel_name.c_str(), it.engine.c_str(),
  965. op_desc->GetName().c_str());
  966. return true;
  967. }
  968. }
  969. }
  970. GELOGI("op %s CheckAccuracySupported failed!reason:%s", op_desc->GetType().c_str(), unsupported_reason.c_str());
  971. return false;
  972. }
  973. bool TransOpWithoutReshapeFusionPass::FusionFormatSupport(Format format) {
  974. return format == FORMAT_NCHW || format == FORMAT_NHWC || format == FORMAT_FRACTAL_Z || format == FORMAT_NC1HWC0;
  975. }
  976. graphStatus TransOpWithoutReshapeFusionPass::GetSubGraphsBetweenNormalNode(
  977. const OutDataAnchorPtr &out_anchor, std::vector<vector<std::pair<OutDataAnchorPtr, InDataAnchorPtr>>> &sub_graphs_out,
  978. vector<std::pair<OutDataAnchorPtr, InDataAnchorPtr>> &nodes_list) {
  979. graphStatus ret = GRAPH_SUCCESS;
  980. if (out_anchor == nullptr) {
  981. return GRAPH_FAILED;
  982. }
  983. for (const auto &peer_in_anchor : out_anchor->GetPeerInDataAnchors()) {
  984. if (peer_in_anchor == nullptr || peer_in_anchor->GetOwnerNode() == nullptr ||
  985. peer_in_anchor->GetOwnerNode()->GetOpDesc() == nullptr) {
  986. continue;
  987. }
  988. nodes_list.emplace_back(out_anchor, peer_in_anchor);
  989. auto peer_in_node = peer_in_anchor->GetOwnerNode();
  990. GE_CHECK_NOTNULL(peer_in_node);
  991. if (!IsTransOp(peer_in_node)) {
  992. sub_graphs_out.push_back(nodes_list);
  993. nodes_list.pop_back();
  994. } else {
  995. for (const auto &peer_out_anchor : peer_in_node->GetAllOutDataAnchors()) {
  996. ret = GetSubGraphsBetweenNormalNode(peer_out_anchor, sub_graphs_out, nodes_list);
  997. if (ret != GRAPH_SUCCESS) {
  998. GELOGE(GRAPH_FAILED, "get all transops between normal node failed!node:%s", peer_in_node->GetName().c_str());
  999. return GRAPH_FAILED;
  1000. }
  1001. }
  1002. nodes_list.pop_back();
  1003. }
  1004. }
  1005. return GRAPH_SUCCESS;
  1006. }
  1007. bool TransOpWithoutReshapeFusionPass::IsTransOp(const NodePtr &node) {
  1008. // The caller guarantees that the pointer is not null.
  1009. return node->GetType() == CAST || node->GetType() == RESHAPE || node->GetType() == TRANSPOSE ||
  1010. node->GetType() == TRANSPOSED || node->GetType() == TRANSDATA;
  1011. }
  1012. } // 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