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.

block_mem_assigner.cc 71 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770
  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/build/memory/block_mem_assigner.h"
  17. #include <algorithm>
  18. #include <sstream>
  19. #include "external/ge/ge_api_types.h"
  20. #include "framework/common/debug/ge_log.h"
  21. #include "graph/anchor.h"
  22. #include "graph/buffer.h"
  23. #include "graph/ge_attr_value.h"
  24. #include "graph/ge_context.h"
  25. #include "graph/node.h"
  26. #include "graph/utils/graph_utils.h"
  27. #include "graph/utils/node_utils.h"
  28. #include "graph/utils/op_desc_utils.h"
  29. #include "graph/utils/tensor_utils.h"
  30. #include "graph/debug/ge_attr_define.h"
  31. #include "graph/common/local_context.h"
  32. #include "graph/optimize/common/params.h"
  33. #include "omg/omg_inner_types.h"
  34. #include "runtime/mem.h"
  35. using std::map;
  36. using std::set;
  37. using std::list;
  38. using std::pair;
  39. using std::string;
  40. using std::stringstream;
  41. using std::unordered_map;
  42. using std::unordered_set;
  43. using std::vector;
  44. namespace {
  45. const char *const kAttrNameWorkspaceReuseFlag = "workspace_reuse_flag";
  46. const char *const kL2FusionDynamicConvergeOp = "l2fusion_dynamic_converge_op";
  47. const char *const kOpNoReuseMem = "no_reuse_mem_flag";
  48. const char *const OP_NO_REUSE_MEM = "OP_NO_REUSE_MEM";
  49. const int kReuseMaxCount = 10;
  50. const int kReuseMaxOpNum = 10;
  51. const int kReuseMaxCharNum = 2000;
  52. } // namespace
  53. namespace ge {
  54. void AlignMemOffset(size_t &mem_align_size) {
  55. if (mem_align_size <= 0) {
  56. return;
  57. }
  58. mem_align_size = (mem_align_size + MEM_ALIGN_SIZE - 1) / MEM_ALIGN_SIZE * MEM_ALIGN_SIZE;
  59. }
  60. void MemoryBlock::SetHeadOffset(size_t offset) {
  61. head_offset_ = offset;
  62. size_t child_offset = head_offset_;
  63. for (auto block : child_blocks_) {
  64. if (block != nullptr) {
  65. block->SetHeadOffset(child_offset);
  66. child_offset += block->Size();
  67. }
  68. }
  69. }
  70. void MemoryBlock::SetTailOffset(size_t offset) {
  71. tail_offset_ = offset;
  72. size_t child_offset = head_offset_;
  73. for (auto block : child_blocks_) {
  74. if (block != nullptr) {
  75. child_offset += block->Size();
  76. block->SetTailOffset(child_offset - 1);
  77. }
  78. }
  79. }
  80. void MemoryBlock::Resize() {
  81. size_t child_block_size = 0;
  82. for (auto block : child_blocks_) {
  83. if (block != nullptr) {
  84. block->Resize();
  85. child_block_size += block->Size();
  86. }
  87. }
  88. auto iter = std::max_element(real_size_list_.begin(), real_size_list_.end());
  89. if (iter == real_size_list_.end()) {
  90. GELOGW("real_size_list_ is empty");
  91. return;
  92. } else {
  93. size_t block_size = (child_block_size > *iter) ? child_block_size : *iter;
  94. if ((block_size > 0) && (block_size % MEM_ALIGN_SIZE != 0)) {
  95. AlignMemOffset(block_size);
  96. }
  97. block_size_ = block_size;
  98. if (last_continuous_block_) {
  99. block_size_ += MEM_ALIGN_SIZE;
  100. }
  101. }
  102. }
  103. size_t MemoryBlock::AlignSize() const {
  104. size_t align_block_size = 0;
  105. auto iter = std::max_element(real_size_list_.begin(), real_size_list_.end());
  106. if (iter == real_size_list_.end()) {
  107. GELOGW("real_size_list_ is empty");
  108. } else {
  109. align_block_size = *iter;
  110. if ((align_block_size > 0) && (align_block_size % MEM_ALIGN_SIZE != 0)) {
  111. AlignMemOffset(align_block_size);
  112. }
  113. }
  114. return align_block_size;
  115. }
  116. bool MemoryBlock::IsSameLabel(std::string &first_batch_label) {
  117. if (node_type_index_list_.empty()) {
  118. return false;
  119. }
  120. auto node_op_desc = node_type_index_list_[0].node->GetOpDesc();
  121. if (node_op_desc == nullptr) {
  122. return false;
  123. }
  124. // not all op has ATTR_NAME_BATCH_LABEL, no need check return value, only check out parameter
  125. (void)ge::AttrUtils::GetStr(node_op_desc, ATTR_NAME_BATCH_LABEL, first_batch_label);
  126. if (first_batch_label.empty()) {
  127. return false;
  128. }
  129. bool all_same_label = true;
  130. for (size_t index = 1; index < node_type_index_list_.size(); ++index) {
  131. if (node_type_index_list_[index].node == nullptr) {
  132. continue;
  133. }
  134. std::string batch_label;
  135. auto index_op_desc = node_type_index_list_[index].node->GetOpDesc();
  136. GE_IF_BOOL_EXEC(index_op_desc == nullptr, continue);
  137. (void)ge::AttrUtils::GetStr(index_op_desc, ATTR_NAME_BATCH_LABEL, batch_label);
  138. if (first_batch_label != batch_label) {
  139. all_same_label = false;
  140. break;
  141. }
  142. }
  143. return all_same_label;
  144. }
  145. bool CanNotLifeReuse(MemoryBlock *block) {
  146. if ((block == nullptr) || !block->reuse_mem_ || block->deleted_block_) {
  147. return true;
  148. }
  149. return false;
  150. }
  151. void MemoryBlock::AddContinuousLifeReuseBlock(MemoryBlock *block, DependStreamLife &total_node_depend_stream_life) {
  152. // continuous memory case:only real_size is maximum can be reused and only one continuous memory in one block
  153. auto it_block = std::max_element(std::begin(block->NoAlignSizeList()), std::end(block->NoAlignSizeList()));
  154. auto it_this = std::max_element(std::begin(NoAlignSizeList()), std::end(NoAlignSizeList()));
  155. if (it_block != std::end(block->NoAlignSizeList()) && it_this != std::end(NoAlignSizeList())) {
  156. if ((continuous_block_ && block->continuous_block_) ||
  157. (continuous_block_ && (*it_this < *it_block)) || (block->continuous_block_ && (*it_this > *it_block))) {
  158. GELOGD("Conflict current block size:%zu continuous:%d, reuse block max size:%zu continuous:%d",
  159. *it_this, continuous_block_, *it_block, block->continuous_block_);
  160. return;
  161. }
  162. }
  163. MemoryBlock *parent = nullptr;
  164. MemoryBlock *child = nullptr;
  165. // merge small block to large block
  166. if (block->GetDependLifeBegin(stream_id_, total_node_depend_stream_life) > GetLifeEnd()) {
  167. if ((block->child_offset_ + AlignSize()) <= *it_block) {
  168. parent = block;
  169. child = this;
  170. }
  171. }
  172. if ((parent != nullptr) && (child != nullptr) && child->child_blocks_.empty()) {
  173. parent->child_blocks_.emplace_back(child);
  174. parent->child_offset_ += child->AlignSize();
  175. child->deleted_block_ = true;
  176. GELOGI("Add continuous block[%p size:%zu, stream id:%ld life time[begin:%zu, end:%zu]] to"
  177. " block[%p size:%zu, stream id:%ld, life time[begin:%zu, end:%zu]]", child, child->block_size_,
  178. child->stream_id_, child->GetLifeBegin(), child->GetLifeEnd(), parent, parent->block_size_,
  179. parent->stream_id_, parent->GetLifeBegin(), parent->GetLifeEnd());
  180. }
  181. }
  182. void MemoryBlock::AddLifeReuseBlock(MemoryBlock *block, DependStreamLife &total_node_depend_stream_life) {
  183. if (CanNotLifeReuse(this) || CanNotLifeReuse(block)) {
  184. return;
  185. }
  186. if (block->continuous_block_) {
  187. AddContinuousLifeReuseBlock(block, total_node_depend_stream_life);
  188. return;
  189. }
  190. MemoryBlock *parent = nullptr;
  191. MemoryBlock *child = nullptr;
  192. // merge small block to large block
  193. if (block->GetDependLifeBegin(stream_id_, total_node_depend_stream_life) > GetLifeEnd()) {
  194. if ((child_offset_ + block->AlignSize()) <= AlignSize()) {
  195. parent = this;
  196. child = block;
  197. } else if ((block->child_offset_ + AlignSize()) <= block->AlignSize()) {
  198. parent = block;
  199. child = this;
  200. }
  201. }
  202. if ((parent != nullptr) && (child != nullptr) && child->child_blocks_.empty()) {
  203. parent->child_blocks_.emplace_back(child);
  204. parent->child_offset_ += child->AlignSize();
  205. child->deleted_block_ = true;
  206. GELOGI("Add block[%p size:%zu, stream id:%ld life time[begin:%zu, end:%zu]] to"
  207. " block[%p size:%zu, stream id:%ld, life time[begin:%zu, end:%zu]]", child, child->block_size_,
  208. child->stream_id_, child->GetLifeBegin(), child->GetLifeEnd(), parent, parent->block_size_,
  209. parent->stream_id_, parent->GetLifeBegin(), parent->GetLifeEnd());
  210. }
  211. }
  212. size_t MemoryBlock::GetLifeBegin() {
  213. size_t life_time = 0;
  214. if (!node_type_index_list_.empty()) {
  215. if (node_type_index_list_.front().node != nullptr) {
  216. auto node_op_desc = node_type_index_list_.front().node->GetOpDesc();
  217. if (node_op_desc != nullptr) {
  218. life_time = node_op_desc->GetId();
  219. }
  220. }
  221. }
  222. return life_time;
  223. }
  224. /// |-stream 1-| |-stream 2-|
  225. /// |--block1--| |--block---|
  226. /// |--block2--| |--block---|
  227. /// |--block3--|\ |--block---|
  228. /// |--block---| \ |--block---|
  229. /// |--block---| \|--block---|
  230. /// |--block---| |--block7--|
  231. /// |--block---| |--block---|
  232. /// block7's first node's input node's life begin > block2's life end, block7 can reuse block1~block2
  233. size_t MemoryBlock::GetDependLifeBegin(int64_t stream_id, DependStreamLife &total_node_depend_stream_life) {
  234. AddDependLifeBegin(total_node_depend_stream_life);
  235. auto it = depend_stream_life_.find(stream_id);
  236. if (it == depend_stream_life_.end()) {
  237. return 0;
  238. }
  239. return it->second;
  240. }
  241. void AddDependLife(const ge::NodePtr &org_node, const ge::NodePtr &node, int64_t stream_id,
  242. std::map<int64_t, size_t> &depend_stream_life, DependStreamLife &total_node_depend_stream_life) {
  243. GE_CHECK_NOTNULL_EXEC(node, return);
  244. auto node_desc = node->GetOpDesc();
  245. GE_CHECK_NOTNULL_EXEC(node_desc, return);
  246. auto node_id = node_desc->GetId();
  247. auto stream_life = total_node_depend_stream_life.find(node_id);
  248. if (stream_life != total_node_depend_stream_life.end()) {
  249. for (auto &it : stream_life->second) {
  250. if (depend_stream_life.find(it.first) == depend_stream_life.end()) {
  251. depend_stream_life[it.first] = it.second;
  252. }
  253. }
  254. return;
  255. }
  256. for (const auto &in_anchor : node->GetAllInAnchors()) {
  257. GE_CHECK_NOTNULL_EXEC(in_anchor, continue);
  258. for (auto peer_out_anchor : in_anchor->GetPeerAnchors()) {
  259. GE_CHECK_NOTNULL_EXEC(peer_out_anchor, continue);
  260. auto peer_node = peer_out_anchor->GetOwnerNode();
  261. GE_CHECK_NOTNULL_EXEC(peer_node, continue);
  262. auto peer_node_desc = peer_node->GetOpDesc();
  263. GE_CHECK_NOTNULL_EXEC(peer_node_desc, continue);
  264. auto peer_node_stream_id = peer_node_desc->GetStreamId();
  265. if (peer_node_stream_id < 0) {
  266. continue;
  267. }
  268. size_t peer_node_life_time = peer_node_desc->GetId();
  269. auto it = depend_stream_life.find(peer_node_stream_id);
  270. if (it == depend_stream_life.end() || peer_node_life_time > it->second) {
  271. depend_stream_life[peer_node_stream_id] = peer_node_life_time;
  272. if (peer_node_stream_id != stream_id) {
  273. GELOGI("Node:%s stream id:%ld depend node:%s stream id:%ld index[%d] life time[%zu].",
  274. org_node->GetName().c_str(), stream_id, peer_node_desc->GetName().c_str(),
  275. peer_node_stream_id, peer_out_anchor->GetIdx(), peer_node_life_time);
  276. }
  277. AddDependLife(org_node, peer_node, stream_id, depend_stream_life, total_node_depend_stream_life);
  278. }
  279. }
  280. }
  281. // save on node to save next calculation
  282. for (auto &it : depend_stream_life) {
  283. if (total_node_depend_stream_life[node_id].find(it.first) == total_node_depend_stream_life[node_id].end()) {
  284. total_node_depend_stream_life[node_id][it.first] = it.second;
  285. }
  286. }
  287. }
  288. void MemoryBlock::AddDependLifeBegin(DependStreamLife &total_node_depend_stream_life) {
  289. if (!depend_stream_life_.empty()) {
  290. return;
  291. }
  292. if (!node_type_index_list_.empty()) {
  293. auto node = node_type_index_list_.front().node;
  294. if (node != nullptr) {
  295. AddDependLife(node, node, stream_id_, depend_stream_life_, total_node_depend_stream_life);
  296. }
  297. }
  298. depend_stream_life_[stream_id_] = GetLifeBegin();
  299. }
  300. size_t MemoryBlock::GetLifeEnd() {
  301. if (!node_type_index_list_.empty()) {
  302. return node_type_index_list_.back().life_time_end;
  303. }
  304. return kMaxLifeTime;
  305. }
  306. void MemoryBlock::SetLifeTimeEnd(size_t time) {
  307. if (!node_type_index_list_.empty()) {
  308. node_type_index_list_.back().life_time_end = time;
  309. }
  310. }
  311. void SetLastUsedInputMemAttr(NodePtr &node, int input_index) {
  312. if (node == nullptr) {
  313. return;
  314. }
  315. auto node_op_desc = node->GetOpDesc();
  316. if (node_op_desc != nullptr) {
  317. auto input_desc = node_op_desc->GetInputDesc(input_index);
  318. if (!ge::AttrUtils::SetInt(input_desc, ATTR_NAME_IS_END_OF_INPUTMEM_LIFECYCLE, true)) {
  319. GELOGW("Set %s input[%d] ATTR_NAME_IS_END_OF_INPUTMEM_LIFECYCLE to true failed.", node_op_desc->GetName().c_str(),
  320. input_index);
  321. return;
  322. }
  323. GELOGD("Set %s input[%d] ATTR_NAME_IS_END_OF_INPUTMEM_LIFECYCLE to true success.", node_op_desc->GetName().c_str(),
  324. input_index);
  325. if (node_op_desc->UpdateInputDesc(input_index, input_desc) != GRAPH_SUCCESS) {
  326. GELOGW("Update %s input[%d] desc failed.", node_op_desc->GetName().c_str(), input_index);
  327. }
  328. }
  329. }
  330. Status GetNoAlignSize(const ge::OpDesc &desc, uint32_t index, size_t &size) {
  331. // calculate tensor real size
  332. auto output_op_desc = desc.GetOutputDescPtr(index);
  333. if (output_op_desc == nullptr) {
  334. GELOGI("GetNoAlignSize failed. OpName: %s, OpType: %s, index: %d",
  335. desc.GetName().c_str(), desc.GetType().c_str(), index);
  336. return FAILED;
  337. }
  338. int64_t tensor_size = 0;
  339. GeShape shape = output_op_desc->GetShape();
  340. Format format = output_op_desc->GetFormat();
  341. DataType data_type = output_op_desc->GetDataType();
  342. graphStatus graph_status = TensorUtils::CalcTensorMemSize(shape, format, data_type, tensor_size);
  343. if (graph_status != GRAPH_SUCCESS) {
  344. GELOGE(graph_status, "CalcTensorMemSize failed!");
  345. return FAILED;
  346. }
  347. size = static_cast<size_t>(tensor_size);
  348. return SUCCESS;
  349. }
  350. string ToString(ge::NodeTypeIndex &x) {
  351. stringstream ss;
  352. ss << "[" << x.node->GetName() << "(" << x.node->GetType() << "), ";
  353. if (x.mem_type == kOutput) {
  354. ss << "Output, ";
  355. } else {
  356. ss << "Workspace, ";
  357. }
  358. ss << x.index << "]";
  359. return ss.str();
  360. }
  361. string MemoryBlock::String() {
  362. stringstream ss;
  363. ss << "Block size: " << Size() << " from " << HeadOffset() << " to " << TailOffset() << " ";
  364. ss << "real_size_list: " << ToString(real_size_list_) << " ";
  365. ss << "ref_count: " << ref_count_ << " ";
  366. ss << "members: ";
  367. for (auto x : NodeTypeIndexList()) {
  368. ss << "__node: " << ToString(x) << " ";
  369. }
  370. for (const auto& symbol : SymbolList()) {
  371. ss << "__symbol: " << symbol << " ";
  372. }
  373. ss << "memory_type: " << memory_type_ << " ";
  374. return ss.str();
  375. }
  376. BlockMemAssigner::BlockMemAssigner(ComputeGraphPtr compute_graph, const map<string, string> &anchor_to_symbol,
  377. const map<string, list<NodeIndexIO>> &symbol_to_anchors)
  378. : mem_offset_(0), p2p_mem_offset_(0), compute_graph_(std::move(compute_graph)),
  379. symbol_to_anchors_(symbol_to_anchors), anchor_to_symbol_(anchor_to_symbol), life_time_(0) {}
  380. BlockMemAssigner::~BlockMemAssigner() {
  381. GELOGD("blocks_store_ size : %lu", blocks_store_.size());
  382. for (MemoryBlock *memory_block : blocks_store_) {
  383. GE_DELETE_NEW_SINGLE(memory_block);
  384. }
  385. }
  386. void BlockMemAssigner::GetOutAndWorkSpaceMem(vector<int64_t> &all_memory_size) {
  387. vector<int64_t> temp;
  388. for (const NodePtr &n : compute_graph_->GetAllNodes()) {
  389. auto node_op_desc = n->GetOpDesc();
  390. GE_IF_BOOL_EXEC(node_op_desc == nullptr, continue);
  391. if (node_op_desc->GetType() == ATOMICADDRCLEAN) {
  392. atomic_addr_clean_id_ = node_op_desc->GetId();
  393. }
  394. for (auto &out_anchor : n->GetAllOutDataAnchors()) {
  395. GeTensorDesc output_desc = node_op_desc->GetOutputDesc(out_anchor->GetIdx());
  396. bool reuse_input = false;
  397. GE_IF_BOOL_EXEC(ge::TensorUtils::GetReuseInput(output_desc, reuse_input) != SUCCESS,
  398. GELOGI("Get reuse_input failed"));
  399. if (!reuse_input) {
  400. int64_t size = 0;
  401. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(output_desc, size) != SUCCESS, GELOGI("Get size failed"));
  402. if (anchor_to_symbol_.empty()) {
  403. all_memory_size.emplace_back(size);
  404. } else {
  405. auto iter1 = anchor_to_symbol_.find(NodeIndexIO(n, out_anchor->GetIdx(), kOut).ToString());
  406. if (iter1 == anchor_to_symbol_.end()) {
  407. continue;
  408. }
  409. const std::string &symbol = iter1->second;
  410. auto iter2 = symbol_size_.find(symbol);
  411. if (iter2 == symbol_size_.end()) {
  412. symbol_size_[symbol] = size;
  413. } else if (size > static_cast<int64_t>(iter2->second)) {
  414. iter2->second = size;
  415. }
  416. }
  417. }
  418. }
  419. temp.clear();
  420. GetNodeWorkSpaceSize(n, temp);
  421. all_memory_size.insert(all_memory_size.end(), temp.begin(), temp.end());
  422. }
  423. GELOGI("The last atomic_addr_clean node id: %ld", atomic_addr_clean_id_);
  424. for (const auto &pair : symbol_size_) {
  425. all_memory_size.emplace_back(pair.second);
  426. }
  427. sort(all_memory_size.begin(), all_memory_size.end());
  428. GELOGI("All memory size: %s", ToString(all_memory_size).c_str());
  429. for (auto iter = all_memory_size.begin(); iter != all_memory_size.end();) {
  430. if (*iter == 0) {
  431. iter = all_memory_size.erase(iter);
  432. } else {
  433. ++iter;
  434. }
  435. }
  436. InitReuseFlag();
  437. PrintSymbolMap();
  438. }
  439. ///
  440. /// @ingroup domi
  441. /// @brief decide memory size based on actual input memory size
  442. /// @param [in] size actual memory size in need
  443. /// @param [in] ranges memory size provided
  444. /// @return size_t memory size to apply
  445. ///
  446. size_t GetBlockSize(size_t size, const vector<int64_t> &ranges) {
  447. for (int64_t x : ranges) {
  448. auto x_temp = static_cast<size_t>(x);
  449. if (size <= x_temp) {
  450. return x_temp;
  451. }
  452. }
  453. GELOGW("Memory needed size:%zu is beyond the biggest block in memory ranges.", size);
  454. return size;
  455. }
  456. bool IsDirectOutputNode(const NodePtr &node, int idx) {
  457. if ((node != nullptr) && (node->GetOpDesc() != nullptr) && (node->GetOpDesc()->GetType() == NETOUTPUT)) {
  458. GELOGI("This is netoutput node, the input node mem can not be reused");
  459. return true;
  460. }
  461. return false;
  462. }
  463. void AddReusableBlockCount(const MemoryBlock &mem_block, map<string, uint64_t> &reusable_block_counts) {
  464. string key = std::to_string(mem_block.Size());
  465. key += "_" + std::to_string(mem_block.stream_id_);
  466. key += "_" + std::to_string(mem_block.memory_type_);
  467. auto it = reusable_block_counts.find(key);
  468. if (it != reusable_block_counts.end()) {
  469. it->second++;
  470. } else {
  471. reusable_block_counts[key] = 1;
  472. }
  473. }
  474. void ReduceReusableBlockCount(const MemoryBlock &mem_block, map<string, uint64_t> &reusable_block_counts) {
  475. string key = std::to_string(mem_block.Size());
  476. key += "_" + std::to_string(mem_block.stream_id_);
  477. key += "_" + std::to_string(mem_block.memory_type_);
  478. auto it = reusable_block_counts.find(key);
  479. if (it != reusable_block_counts.end()) {
  480. if (it->second > 0) {
  481. it->second--;
  482. }
  483. }
  484. }
  485. bool CanReuseBySize(const map<string, uint64_t> &reusable_block_counts, const MemoryBlock &reusable_block,
  486. size_t block_size, size_t real_size, bool continuous) {
  487. bool can_reuse = false;
  488. if (reusable_block.Size() == block_size) {
  489. can_reuse = true;
  490. } else {
  491. string key = std::to_string(reusable_block.Size());
  492. key += "_" + std::to_string(reusable_block.stream_id_);
  493. key += "_" + std::to_string(reusable_block.memory_type_);
  494. auto it = reusable_block_counts.find(key);
  495. GE_IF_BOOL_EXEC((it != reusable_block_counts.end() && (it->second > kReuseMaxCount)) &&
  496. (reusable_block.Size() > block_size),
  497. can_reuse = true;
  498. GELOGD("Less size mem reuse, reuse block size:%zu, current block size:%zu",
  499. reusable_block.Size(), block_size););
  500. }
  501. return can_reuse;
  502. }
  503. bool BlockMemAssigner::IsOutNodeSetContinuousInput(const NodePtr &n, uint32_t out_index, std::string &peer_name,
  504. uint32_t &peer_input_index,
  505. bool &no_need_assign_memory, bool &reset_zero_copy_flag) {
  506. if (n == nullptr || n->GetAllOutDataAnchors().size() <= 0) {
  507. return false;
  508. }
  509. if (static_cast<size_t>(out_index) < n->GetAllOutDataAnchors().size()) {
  510. auto out_anchor = n->GetOutDataAnchor(out_index);
  511. GE_IF_BOOL_EXEC(out_anchor == nullptr,
  512. GELOGE(FAILED, "Node[%s] output[%u] anchor is null.", n->GetName().c_str(), out_index);
  513. return false;);
  514. for (auto const &peer_in_anchor : out_anchor->GetPeerInDataAnchors()) {
  515. GE_IF_BOOL_EXEC(peer_in_anchor == nullptr,
  516. GELOGE(FAILED, "Node[%s] output[%u] peer_in_anchor 0 is null.", n->GetName().c_str(), out_index);
  517. return false;);
  518. auto peer_node = peer_in_anchor->GetOwnerNode();
  519. GE_IF_BOOL_EXEC(peer_node == nullptr,
  520. GELOGE(FAILED, "Node[%s] output[%u] node is null.", n->GetName().c_str(), out_index);
  521. return false;);
  522. // Get the continuous input type of the node, default is false
  523. bool is_input_continuous = false;
  524. auto peer_in_node_desc = peer_node->GetOpDesc();
  525. GE_IF_BOOL_EXEC(peer_in_node_desc == nullptr,
  526. GELOGE(FAILED, "Node[%s] output[%u] nodedesc is null.", n->GetName().c_str(), out_index);
  527. return false;);
  528. // If GetBool fail, is_input_continuous is false.
  529. bool is_input_continuous_no_padding = false;
  530. (void)ge::AttrUtils::GetBool(peer_in_node_desc, ATTR_NAME_NOPADDING_CONTINUOUS_INPUT,
  531. is_input_continuous_no_padding);
  532. if (is_input_continuous_no_padding) {
  533. reset_zero_copy_flag = true;
  534. return false;
  535. }
  536. (void)ge::AttrUtils::GetBool(peer_in_node_desc, ATTR_NAME_CONTINUOUS_INPUT, is_input_continuous);
  537. GE_IF_BOOL_EXEC(is_input_continuous && CheckIsZeroMemNodeType(peer_node->GetType()),
  538. GELOGI("Node[%s] output[%u] no_need_assign_memory.", n->GetName().c_str(), out_index);
  539. no_need_assign_memory = true;
  540. return false;);
  541. if (is_input_continuous) {
  542. if (n->GetOwnerComputeGraph() != nullptr) {
  543. string graph_name = n->GetOwnerComputeGraph()->GetName();
  544. GELOGI("%s name[%s] output[%u] node[%s] set input[%d] continuous, input size[%u].", graph_name.c_str(),
  545. n->GetName().c_str(), out_index, peer_in_node_desc->GetName().c_str(), peer_in_anchor->GetIdx(),
  546. peer_node->GetAllInDataAnchorsSize());
  547. // Only set attr one times.
  548. if (node_continuous_input_blocks_[peer_in_node_desc->GetName()].size() == 0) {
  549. (void)ge::AttrUtils::SetBool(peer_in_node_desc, ATTR_NAME_CONTINUOUS_INPUT_ALLOC, true);
  550. node_continuous_input_counts_[peer_in_node_desc->GetName()] = peer_node->GetAllInDataAnchorsSize();
  551. }
  552. peer_input_index = peer_in_anchor->GetIdx();
  553. peer_name = peer_in_node_desc->GetName();
  554. return true;
  555. }
  556. }
  557. }
  558. }
  559. return false;
  560. }
  561. ///
  562. /// @ingroup GE
  563. /// @brief Check pre_reuse flag & post_reuse glag for each symbol
  564. /// @return void
  565. ///
  566. void BlockMemAssigner::InitReuseFlag() {
  567. static const std::set<std::string> kPreReuseTypes = { ge::DATA_TYPE, ge::AIPP_DATA_TYPE, ge::ANN_DATA_TYPE,
  568. ge::NETOUTPUT, ge::PROPOSAL, ge::ZEROSLIKE,
  569. ge::CONSTANT, ge::CONSTANTOP };
  570. static const std::set<std::string> kPostReuseTypes = { ge::DATA_TYPE, ge::AIPP_DATA_TYPE, ge::ENTER, ge::REFENTER,
  571. ge::NEXTITERATION, ge::REFNEXTITERATION };
  572. for (const auto &pair : symbol_to_anchors_) {
  573. std::string symbol = pair.first;
  574. bool pre_reuse_flag = true;
  575. bool post_reuse_flag = true;
  576. // default memory type
  577. int64_t mem_type = RT_MEMORY_HBM;
  578. GetSymbolMemType(pair.second, mem_type);
  579. GELOGD("The memory type of symbol[%s] is [%ld]].", symbol.c_str(), mem_type);
  580. if (mem_type == RT_MEMORY_P2P_DDR) {
  581. UpdateOpTensorMemType(pair.second, mem_type);
  582. }
  583. // Only the memory with special requirements is processed. The HBM uses the default processing mode.
  584. if (mem_type == RT_MEMORY_P2P_DDR) {
  585. symbol_to_mem_type_[symbol] = mem_type;
  586. }
  587. for (const auto &node_index_io : pair.second) {
  588. if (node_index_io.io_type_ == kIn) {
  589. continue;
  590. }
  591. OutDataAnchorPtr out_anchor = node_index_io.node_->GetOutDataAnchor(node_index_io.index_);
  592. if (out_anchor == nullptr) {
  593. continue;
  594. }
  595. bool out_flg = false;
  596. if (node_index_io.node_->GetOutDataNodes().empty()) {
  597. out_flg = true;
  598. }
  599. for (const auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  600. if (IsDirectOutputNode(in_anchor->GetOwnerNode(), in_anchor->GetIdx())) {
  601. out_flg = true;
  602. break;
  603. }
  604. }
  605. const std::string &type = out_anchor->GetOwnerNode()->GetType();
  606. pre_reuse_flag = pre_reuse_flag && !out_flg && (kPreReuseTypes.count(type) == 0);
  607. post_reuse_flag = post_reuse_flag && (kPostReuseTypes.count(type) == 0);
  608. if (!pre_reuse_flag && !post_reuse_flag) {
  609. break;
  610. }
  611. }
  612. pre_reuse_flag_[symbol] = pre_reuse_flag;
  613. post_reuse_flag_[symbol] = post_reuse_flag;
  614. }
  615. }
  616. ///
  617. /// @ingroup GE
  618. /// @brief get pre_reuse flag
  619. /// @param [in] node
  620. /// @param [in] out_index
  621. /// @return bool
  622. ///
  623. bool BlockMemAssigner::IsPreReuse(const NodePtr &node, uint32_t out_index) const {
  624. OutDataAnchorPtr out_data_anchor = nullptr;
  625. if (static_cast<size_t>(out_index) < node->GetAllOutDataAnchors().size()) {
  626. out_data_anchor = node->GetOutDataAnchor(out_index);
  627. }
  628. if (out_data_anchor == nullptr) {
  629. return false;
  630. }
  631. NodeIndexIO cur_node_index_io(out_data_anchor->GetOwnerNode(), out_data_anchor->GetIdx(), kOut);
  632. auto iter1 = anchor_to_symbol_.find(cur_node_index_io.ToString());
  633. if (iter1 == anchor_to_symbol_.end()) {
  634. return false;
  635. }
  636. const std::string &symbol = iter1->second;
  637. auto iter2 = pre_reuse_flag_.find(symbol);
  638. if (iter2 == pre_reuse_flag_.end()) {
  639. return false;
  640. }
  641. return iter2->second;
  642. }
  643. ///
  644. /// @ingroup GE
  645. /// @brief get post_reuse flag
  646. /// @param [in] mem_block
  647. /// @return bool
  648. ///
  649. bool BlockMemAssigner::IsPostReuse(const MemoryBlock *mem_block) const {
  650. if (mem_block == nullptr) {
  651. return false;
  652. }
  653. for (const auto &symbol : mem_block->SymbolList()) {
  654. auto iter = post_reuse_flag_.find(symbol);
  655. if (iter == post_reuse_flag_.end()) {
  656. continue;
  657. }
  658. if (!iter->second) {
  659. return false;
  660. }
  661. }
  662. return true;
  663. }
  664. ///
  665. /// @ingroup GE
  666. /// @brief check if symbol of cur node_index_io has block
  667. /// @param [in] node_index_io
  668. /// @param [out] symbol
  669. /// @return bool
  670. ///
  671. bool BlockMemAssigner::IsSymbolExist(const NodeIndexIO &node_index_io, string &symbol) {
  672. auto iter = anchor_to_symbol_.find(node_index_io.ToString());
  673. if (iter == anchor_to_symbol_.end()) {
  674. return false;
  675. }
  676. symbol = iter->second;
  677. return symbol_blocks_.find(iter->second) != symbol_blocks_.end();
  678. }
  679. ///
  680. /// @ingroup GE
  681. /// @brief Print symbol
  682. /// @return void
  683. ///
  684. void BlockMemAssigner::PrintSymbolMap() {
  685. for (const auto &pair : symbol_to_anchors_) {
  686. GELOGD("symbol=%s, max_size=%zu, pre_reuse=%s, post_reuse=%s", pair.first.c_str(), symbol_size_[pair.first],
  687. pre_reuse_flag_[pair.first] ? "true" : "false", post_reuse_flag_[pair.first] ? "true" : "false");
  688. for (const auto &node_index_io : pair.second) {
  689. GELOGD("anchor:%s", node_index_io.ToString().c_str());
  690. }
  691. }
  692. }
  693. void BlockMemAssigner::GetSymbolMemType(std::list<NodeIndexIO> node_index_io_list, int64_t &memory_type) {
  694. memory_type = RT_MEMORY_HBM;
  695. vector<int64_t> memory_types;
  696. for (auto &node_index_io : node_index_io_list) {
  697. auto op_desc = node_index_io.node_->GetOpDesc();
  698. if (op_desc == nullptr) {
  699. GELOGW("Node[%s] op desc is null.", node_index_io.node_->GetName().c_str());
  700. return;
  701. }
  702. if (node_index_io.io_type_ == kIn) {
  703. vector<int64_t> input_memory_types;
  704. (void) ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_INPUT_MEM_TYPE_LIST, input_memory_types);
  705. if (!input_memory_types.empty() && node_index_io.index_ < input_memory_types.size()) {
  706. int64_t input_memory_type = input_memory_types[node_index_io.index_];
  707. GELOGD("Node[%s]: the memory type of input index [%u] is [%ld]].", op_desc->GetName().c_str(),
  708. node_index_io.index_, input_memory_type);
  709. memory_types.emplace_back(input_memory_type);
  710. }
  711. }
  712. if (node_index_io.io_type_ == kOut) {
  713. vector<int64_t> output_memory_types;
  714. (void) ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_OUTPUT_MEM_TYPE_LIST, output_memory_types);
  715. if (!output_memory_types.empty() && node_index_io.index_ < output_memory_types.size()) {
  716. int64_t output_memory_type = output_memory_types[node_index_io.index_];
  717. GELOGD("Node[%s]: the memory type of output index [%u] is [%ld]].", op_desc->GetName().c_str(),
  718. node_index_io.index_, output_memory_type);
  719. memory_types.emplace_back(output_memory_type);
  720. }
  721. }
  722. }
  723. // memory priority
  724. for (auto node_memory_type : memory_types) {
  725. if (node_memory_type > memory_type) {
  726. memory_type = node_memory_type;
  727. }
  728. }
  729. }
  730. void BlockMemAssigner::UpdateOpTensorMemType(std::list<NodeIndexIO> node_index_io_list, int64_t memory_type) {
  731. for (auto &node_index_io : node_index_io_list) {
  732. auto op_desc = node_index_io.node_->GetOpDesc();
  733. if (op_desc == nullptr) {
  734. GELOGW("Node[%s] op desc is null.", node_index_io.node_->GetName().c_str());
  735. return;
  736. }
  737. if (node_index_io.io_type_ == kIn) {
  738. auto input_desc = op_desc->MutableInputDesc(node_index_io.index_);
  739. (void) AttrUtils::SetInt(input_desc, ATTR_NAME_TENSOR_MEM_TYPE, memory_type);
  740. }
  741. if (node_index_io.io_type_ == kOut) {
  742. auto output_desc = op_desc->MutableOutputDesc(node_index_io.index_);
  743. (void) AttrUtils::SetInt(output_desc, ATTR_NAME_TENSOR_MEM_TYPE, memory_type);
  744. }
  745. }
  746. }
  747. bool BlockMemAssigner::IsContinuousOutput(const NodePtr &n) {
  748. if (n == nullptr) {
  749. GELOGE(FAILED, "Node is null.");
  750. return false;
  751. }
  752. // Get the continuous output type of the node, default is false
  753. bool is_output_continuous = false;
  754. auto node_desc = n->GetOpDesc();
  755. if (node_desc == nullptr) {
  756. GELOGE(FAILED, "Node[%s] nodedesc is null.", n->GetName().c_str());
  757. return false;
  758. }
  759. // If GetBool fail, is_output_continuous is false.
  760. (void)ge::AttrUtils::GetBool(node_desc, ATTR_NAME_CONTINUOUS_OUTPUT, is_output_continuous);
  761. if (is_output_continuous) {
  762. if (n->GetOwnerComputeGraph() != nullptr) {
  763. string graph_name = n->GetOwnerComputeGraph()->GetName();
  764. GELOGI("%s name[%s] set continuous, output size[%u].", graph_name.c_str(),
  765. n->GetName().c_str(), n->GetAllOutDataAnchorsSize());
  766. return true;
  767. }
  768. }
  769. return false;
  770. }
  771. bool BlockMemAssigner::IsZeroCopyBlock(const NodePtr &node, bool continuous) {
  772. if (NodeUtils::IsDynamicShape(node)) {
  773. return ((node->GetType() == DATA_TYPE) && !continuous) || (node->GetType() == NETOUTPUT);
  774. }
  775. if ((node->GetType() == DATA_TYPE) && !continuous) {
  776. return !node->GetOpDesc()->HasAttr(ATTR_NAME_PARENT_NODE_INDEX);
  777. }
  778. if (node->GetType() == NETOUTPUT) {
  779. const auto &owner = node->GetOwnerComputeGraph();
  780. return owner->GetParentGraph() == nullptr;
  781. }
  782. return false;
  783. }
  784. MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size, size_t no_align_size,
  785. OpMemoryType mem_type, const NodePtr &n, uint32_t out_index,
  786. const vector<bool> &workspace_reuse_flag, const bool is_op_reuse_mem,
  787. const bool continuous, int64_t memory_type) {
  788. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr, return nullptr, "Input parameter n is null.");
  789. auto node_op_desc = n->GetOpDesc();
  790. GE_IF_BOOL_EXEC(node_op_desc == nullptr, return nullptr);
  791. bool is_reuse_memory = false;
  792. string ge_disable_reuse_mem_env = "0";
  793. (void)ge::GetContext().GetOption(OPTION_EXEC_DISABLE_REUSED_MEMORY, ge_disable_reuse_mem_env);
  794. if (ge_disable_reuse_mem_env != "1") {
  795. bool reuse_mem_flag = !((workspace_reuse_flag.size() > out_index) && !workspace_reuse_flag[out_index]);
  796. is_reuse_memory = !node_op_desc->HasAttr(kL2FusionDynamicConvergeOp) &&
  797. !node_op_desc->HasAttr(kOpNoReuseMem) &&
  798. reuse_mem_flag &&
  799. is_op_reuse_mem &&
  800. (IsPreReuse(n, out_index));
  801. auto stream_id = node_op_desc->GetStreamId();
  802. if (is_reuse_memory && !continuous && !reusable_blocks_[memory_type].empty()) {
  803. for (auto it = reusable_blocks_[memory_type][stream_id].begin();
  804. it != reusable_blocks_[memory_type][stream_id].end(); ++it) {
  805. MemoryBlock *reusable_block = *it;
  806. if (!IsPostReuse(reusable_block)) {
  807. reusable_block->reuse_mem_ = false;
  808. GELOGI("Unreusable block.");
  809. continue;
  810. }
  811. // A node can reuse blocks of the same stream and preorder streams
  812. if (CanReuseBySize(reusable_block_counts_, *reusable_block, block_size, real_size, continuous)) {
  813. reusable_block->AddNodeTypeIndex({n, mem_type, out_index, false}, real_size, no_align_size);
  814. if (mem_type == kOutput) {
  815. auto iter = anchor_to_symbol_.find(NodeIndexIO(n, out_index, kOut).ToString());
  816. if (iter != anchor_to_symbol_.end()) {
  817. reusable_block->AddSymbol(iter->second);
  818. }
  819. }
  820. reusable_block->continuous_block_ = continuous;
  821. reusable_block->ref_count_++;
  822. ReduceReusableBlockCount(*reusable_block, reusable_block_counts_);
  823. reusable_blocks_[memory_type][stream_id].erase(it);
  824. return reusable_block;
  825. }
  826. }
  827. }
  828. }
  829. auto block = new (std::nothrow) MemoryBlock(block_size, node_op_desc->GetStreamId(), is_reuse_memory, memory_type);
  830. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(block == nullptr, return nullptr, "new an object failed.");
  831. // Data and netoutput need zero copy block
  832. block->is_zero_copy_ = IsZeroCopyBlock(n, continuous);
  833. block->Init(real_size, mem_type, n, out_index, no_align_size);
  834. block->stream_id_ = node_op_desc->GetStreamId();
  835. block->ref_count_++;
  836. block->continuous_block_ = continuous;
  837. if (mem_type == kOutput) {
  838. auto iter = anchor_to_symbol_.find(NodeIndexIO(n, out_index, kOut).ToString());
  839. if (iter != anchor_to_symbol_.end()) {
  840. block->AddSymbol(iter->second);
  841. }
  842. }
  843. memory_blocks_.emplace_back(block);
  844. // cause memory_blocks_ may reduce when swap after,
  845. // create blocks_store_ to assure blocks deleted finally
  846. blocks_store_.emplace_back(block);
  847. return block;
  848. }
  849. MemoryBlock *BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<int64_t> &ranges,
  850. const bool is_op_reuse_mem) {
  851. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr, return nullptr, "input node is null.");
  852. auto node_op_desc = n->GetOpDesc();
  853. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node_op_desc == nullptr, return nullptr, "node_op_desc is null.");
  854. MemoryBlock *block = nullptr;
  855. int64_t total_size = 0;
  856. int64_t memory_type = RT_MEMORY_HBM;
  857. for (uint32_t index = 0; index < static_cast<uint32_t>(node_op_desc->GetOutputsSize()); index++) {
  858. auto output_op_desc = node_op_desc->GetOutputDescPtr(index);
  859. if (output_op_desc == nullptr) {
  860. return nullptr;
  861. }
  862. int64_t size = 0;
  863. if (ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS) {
  864. GELOGI("Get size failed");
  865. return nullptr;
  866. }
  867. size_t align_size = static_cast<size_t>(size);
  868. AlignMemOffset(align_size);
  869. total_size += align_size;
  870. // only apply total size in first block
  871. if (index != 0) {
  872. zero_memory_list_.emplace_back(n, kOutput, index);
  873. }
  874. if (index == 0) {
  875. NodeIndexIO node_index_io(n, index, kOut);
  876. auto iter = anchor_to_symbol_.find(node_index_io.ToString());
  877. if (iter != anchor_to_symbol_.end()) {
  878. string symbol = iter->second;
  879. if (symbol_to_mem_type_.find(symbol) != symbol_to_mem_type_.end()) {
  880. memory_type = symbol_to_mem_type_[symbol];
  881. GELOGD("Continuous out memory symbol is [%s], memory type is [%ld]", symbol.c_str(), memory_type);
  882. }
  883. }
  884. }
  885. }
  886. auto block_size = GetBlockSize(total_size, ranges);
  887. GELOGI("Node[%s] continuous out memory size[%ld] block size[%zu]", node_op_desc->GetName().c_str(),
  888. total_size, block_size);
  889. vector<bool> workspace_reuse_flag;
  890. block = ApplyMemory(block_size, total_size, total_size, kOutput, n, 0, workspace_reuse_flag, is_op_reuse_mem, true,
  891. memory_type);
  892. if (block != nullptr) {
  893. // hccl task need align header and tail
  894. block->first_continuous_block_ = true;
  895. block->last_continuous_block_ = true;
  896. }
  897. return block;
  898. }
  899. MemoryBlock *BlockMemAssigner::ApplyOutMemory(const NodePtr &n, uint32_t index, const vector<int64_t> &ranges,
  900. const bool is_op_reuse_mem, const bool continuous) {
  901. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr, return nullptr, "input node is null.");
  902. auto node_op_desc = n->GetOpDesc();
  903. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node_op_desc == nullptr, return nullptr, "node_op_desc is null.");
  904. MemoryBlock *block = nullptr;
  905. NodeIndexIO node_index_io(n, index, kOut);
  906. int64_t size = 0;
  907. auto output_op_desc = node_op_desc->GetOutputDescPtr(index);
  908. if (output_op_desc != nullptr) {
  909. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS, GELOGI("Get size failed"));
  910. }
  911. size_t no_align_size = 0;
  912. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(GetNoAlignSize(*node_op_desc, index, no_align_size) != SUCCESS,
  913. return nullptr, "Get no align size failed");
  914. std::string symbol;
  915. if (IsSymbolExist(node_index_io, symbol)) {
  916. block = symbol_blocks_[symbol];
  917. block->AddNodeTypeIndex({n, kOutput, index, true}, size, no_align_size);
  918. block->ref_count_++;
  919. } else {
  920. int64_t max_size = size;
  921. int64_t memory_type = RT_MEMORY_HBM;
  922. auto iter1 = anchor_to_symbol_.find(node_index_io.ToString());
  923. if (iter1 != anchor_to_symbol_.end()) {
  924. auto iter2 = symbol_size_.find(iter1->second);
  925. if (iter2 != symbol_size_.end()) {
  926. max_size = iter2->second;
  927. }
  928. auto iter3 = symbol_to_mem_type_.find(iter1->second);
  929. if (iter3 != symbol_to_mem_type_.end()) {
  930. memory_type = iter3->second;
  931. }
  932. }
  933. auto block_size = GetBlockSize(max_size, ranges);
  934. vector<bool> workspace_reuse_flag;
  935. block = ApplyMemory(block_size, size, no_align_size, kOutput, n, index,
  936. workspace_reuse_flag, is_op_reuse_mem, continuous, memory_type);
  937. }
  938. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(block == nullptr, return nullptr, "Block is nullptr.");
  939. int out_count_reuse_input = block->ref_count_;
  940. int out_count = 0;
  941. GE_IF_BOOL_EXEC(index >= n->GetAllOutDataAnchors().size(), GELOGE(FAILED, "index is out of range."); return nullptr);
  942. auto out_data_anchor = n->GetOutDataAnchor(index);
  943. GE_IF_BOOL_EXEC(out_data_anchor == nullptr, GELOGE(FAILED, "Out data anchor is nullptr."); return nullptr);
  944. for (const auto &in_anchor : out_data_anchor->GetPeerInDataAnchors()) {
  945. auto owner_node = in_anchor->GetOwnerNode();
  946. auto op_desc = owner_node->GetOpDesc();
  947. GE_IF_BOOL_EXEC(op_desc == nullptr, continue);
  948. Params *instance = Params::Instance();
  949. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(instance == nullptr, return nullptr, "Params instance is nullptr.");
  950. if (!((instance->GetTarget() == TARGET_TYPE_TINY) && (op_desc->GetType() == NETOUTPUT))) {
  951. out_count++;
  952. }
  953. }
  954. bool reuse_input = false;
  955. for (const auto &in_anchor : out_data_anchor->GetPeerInDataAnchors()) {
  956. auto owner_node = in_anchor->GetOwnerNode();
  957. GE_IF_BOOL_EXEC(owner_node == nullptr, continue);
  958. auto op_desc = owner_node->GetOpDesc();
  959. GE_IF_BOOL_EXEC(op_desc == nullptr, continue);
  960. for (uint32_t i = 0; i < static_cast<uint32_t>(op_desc->GetOutputsSize()); i++) {
  961. bool dst_reuse_input = false;
  962. uint32_t dst_reuse_input_index = 0;
  963. auto owner_node_op_desc = op_desc->GetOutputDescPtr(i);
  964. GE_IF_BOOL_EXEC(owner_node_op_desc == nullptr, continue);
  965. GE_IF_BOOL_EXEC(ge::TensorUtils::GetReuseInput(*owner_node_op_desc, dst_reuse_input) != SUCCESS,
  966. GELOGI("Get dst_reuse_input failed"));
  967. GE_IF_BOOL_EXEC(ge::TensorUtils::GetReuseInputIndex(*owner_node_op_desc, dst_reuse_input_index) != SUCCESS,
  968. GELOGI("Get dst_reuse_input_index failed"));
  969. if (dst_reuse_input && (dst_reuse_input_index == static_cast<uint32_t>(in_anchor->GetIdx()))) {
  970. block->AddNodeTypeIndex({owner_node, kOutput, i, true}, block->Size(), block->Size());
  971. out_count_reuse_input += 1;
  972. reuse_input = true;
  973. }
  974. }
  975. }
  976. block->ref_count_ = reuse_input ? out_count_reuse_input + out_count - 1 : out_count;
  977. return block;
  978. }
  979. bool IsOutputBlock(const ge::InDataAnchorPtr &in_data_anchor) {
  980. auto peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  981. GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, GELOGE(FAILED, "Peer out anchor is nullptr."); return false);
  982. auto src = peer_out_anchor->GetOwnerNode();
  983. int32_t index = peer_out_anchor->GetIdx();
  984. auto iter = GetLocalOmgContext().out_nodes_map.find(src->GetName());
  985. if (iter != GetLocalOmgContext().out_nodes_map.end()) {
  986. for (auto id : iter->second) {
  987. if (index == id) {
  988. return true;
  989. }
  990. }
  991. }
  992. return false;
  993. }
  994. // atomic out memory will be reassigned
  995. bool IsAtomicOutputMemory(const ge::NodePtr &node, uint32_t output_index, bool is_atomic,
  996. bool out_node_set_continuous_input) {
  997. auto op_desc = node->GetOpDesc();
  998. if (op_desc == nullptr) {
  999. return false;
  1000. }
  1001. vector<int64_t> atomic_output_index;
  1002. // If GetListInt fail, atomic_output_index is empty.
  1003. (void)ge::AttrUtils::GetListInt(op_desc, ATOMIC_ATTR_OUTPUT_INDEX, atomic_output_index);
  1004. if (!out_node_set_continuous_input && is_atomic) {
  1005. for (auto &index : atomic_output_index) {
  1006. if (static_cast<uint32_t>(index) == output_index) {
  1007. if (node->GetOwnerComputeGraph() != nullptr) {
  1008. string graph_name = node->GetOwnerComputeGraph()->GetName();
  1009. GELOGD("[IMAS]Atomic no assign %s name[%s] output[%d] streamid[%ld].", graph_name.c_str(),
  1010. op_desc->GetName().c_str(), index, op_desc->GetStreamId());
  1011. }
  1012. return true;
  1013. }
  1014. }
  1015. }
  1016. return false;
  1017. }
  1018. bool IsKnownSubgraphData(const NodePtr &node) {
  1019. if (NodeUtils::IsDynamicShape(node)) {
  1020. return false;
  1021. }
  1022. return node->GetOpDesc()->HasAttr(ATTR_NAME_PARENT_NODE_INDEX);
  1023. }
  1024. void BlockMemAssigner::ReleaseMemory(MemoryBlock *to_release, vector<MemoryBlock *> &reusable_memory) {
  1025. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(to_release == nullptr, return, "Input parameter to_release is null.");
  1026. GE_CHK_TRUE_EXEC_INFO(to_release->ref_count_ <= 0, return, "Release memory");
  1027. GE_CHK_TRUE_EXEC_INFO(!to_release->reuse_mem_, return, "doesn't reuse memory");
  1028. --to_release->ref_count_;
  1029. if (to_release->ref_count_ == 0) {
  1030. to_release->SetLifeTimeEnd(life_time_);
  1031. reusable_memory.emplace_back(to_release);
  1032. AddReusableBlockCount(*to_release, reusable_block_counts_);
  1033. }
  1034. }
  1035. void BlockMemAssigner::ReleaseMemorys(const vector<MemoryBlock *> &to_releases,
  1036. vector<MemoryBlock *> &reusable_memory) {
  1037. for (auto mem_block : to_releases) {
  1038. ReleaseMemory(mem_block, reusable_memory);
  1039. }
  1040. }
  1041. void BlockMemAssigner::ReleaseInputNodeOutMemory(const unordered_map<string, vector<MemoryBlock *>> &node_out_blocks,
  1042. vector<MemoryBlock *> &reusable_memory, NodePtr &node) {
  1043. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  1044. if ((in_anchor->GetPeerOutAnchor() == nullptr) ||
  1045. (in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetOpDesc() == nullptr) || (node->GetOpDesc() == nullptr)) {
  1046. return;
  1047. }
  1048. GE_IF_BOOL_EXEC(IsOutputBlock(in_anchor), continue);
  1049. auto node_name = in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetName();
  1050. GE_IF_BOOL_EXEC((in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetType() == CONSTANT) ||
  1051. (in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetType() == FASTRCNNPREDICTIONS) ||
  1052. (in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetType() == CONSTANTOP),
  1053. continue);
  1054. auto it = node_out_blocks.find(node_name);
  1055. if (it == node_out_blocks.end()) {
  1056. continue;
  1057. }
  1058. for (auto block : it->second) {
  1059. const vector<NodeTypeIndex> &node_type_indexs = block->NodeTypeIndexList();
  1060. if (node_type_indexs.empty()) {
  1061. continue;
  1062. }
  1063. GELOGD("node_type_indexs: %d, %s", node_type_indexs.back().index,
  1064. node_type_indexs.back().node->GetName().c_str());
  1065. if ((node_type_indexs.back().node == in_anchor->GetPeerOutAnchor()->GetOwnerNode()) &&
  1066. (node_type_indexs.back().index == static_cast<uint32_t>(in_anchor->GetPeerOutAnchor()->GetIdx())) &&
  1067. (node->GetOpDesc()->GetStreamId() == block->stream_id_)) {
  1068. ReleaseMemory(block, reusable_memory);
  1069. if (block->ref_count_ == 0) {
  1070. SetLastUsedInputMemAttr(node, in_anchor->GetIdx());
  1071. }
  1072. }
  1073. }
  1074. }
  1075. }
  1076. void SplitStringByComma(const string &str, vector<string> &sub_str_vec) {
  1077. std::string tmp_string = str + ",";
  1078. std::string::size_type start_pos = 0;
  1079. std::string::size_type cur_pos = tmp_string.find(',', 0);
  1080. while (cur_pos != std::string::npos) {
  1081. std::string sub_str = tmp_string.substr(start_pos, cur_pos - start_pos);
  1082. if (!sub_str.empty()) {
  1083. vector<string>::iterator ret = std::find(sub_str_vec.begin(), sub_str_vec.end(), sub_str);
  1084. if (ret == sub_str_vec.end()) {
  1085. sub_str_vec.push_back(sub_str);
  1086. }
  1087. }
  1088. start_pos = cur_pos + 1;
  1089. cur_pos = tmp_string.find(',', start_pos);
  1090. }
  1091. }
  1092. void CheckAndGetOpReuseEnv(const string &env, vector<string> &env_vec, bool &op_reuse_env_valid) {
  1093. string env_str;
  1094. env_str = string(env);
  1095. if (env_str.size() > kReuseMaxCharNum) {
  1096. GELOGE(FAILED, "The OP_NO_REUSE_MEM has more than %d characters.", kReuseMaxCharNum);
  1097. return;
  1098. }
  1099. SplitStringByComma(env_str, env_vec);
  1100. if (env_vec.size() > kReuseMaxOpNum) {
  1101. GELOGE(FAILED, "The OP_NO_REUSE_MEM has more than %d nodes.", kReuseMaxOpNum);
  1102. return;
  1103. }
  1104. op_reuse_env_valid = true;
  1105. return;
  1106. }
  1107. Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector<int64_t> &ranges) {
  1108. auto op_desc = node->GetOpDesc();
  1109. int64_t stream_id = op_desc->GetStreamId();
  1110. vector<int64_t> memorys_type;
  1111. bool has_mem_type_attr = ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_OUTPUT_MEM_TYPE_LIST, memorys_type);
  1112. GELOGI("Assign memory node[%s], output size[%d], output memory type size[%d]", op_desc->GetName().c_str(),
  1113. op_desc->GetOutputsSize(), memorys_type.size());
  1114. if (has_mem_type_attr && (memorys_type.size() != op_desc->GetOutputsSize())) {
  1115. GELOGE(INTERNAL_ERROR, "fusion: node[%s], output memory size err[outputsize:%zu, memorysize:%zu]",
  1116. op_desc->GetName().c_str(), op_desc->GetOutputsSize(), memorys_type.size());
  1117. return INTERNAL_ERROR;
  1118. }
  1119. is_op_reuse_mem_ = true;
  1120. if (op_reuse_env_valid_ == true) {
  1121. vector<string>::iterator it_name =
  1122. std::find(op_no_reuse_mem_vec_.begin(), op_no_reuse_mem_vec_.end(), op_desc->GetName());
  1123. vector<string>::iterator it_type =
  1124. std::find(op_no_reuse_mem_vec_.begin(), op_no_reuse_mem_vec_.end(), op_desc->GetType());
  1125. GE_IF_BOOL_EXEC(it_name != op_no_reuse_mem_vec_.end() || it_type != op_no_reuse_mem_vec_.end(),
  1126. is_op_reuse_mem_ = false;);
  1127. }
  1128. bool is_atomic = false;
  1129. // If GetBool fail, is_atomic is false.
  1130. (void)ge::AttrUtils::GetBool(op_desc, ATOMIC_ATTR_IS_ATOMIC_NODE, is_atomic);
  1131. // Allocate memory for the current node and release node memory of the same size in the workspace
  1132. GE_IF_BOOL_EXEC(ge_disable_reuse_mem_env_ != "1",
  1133. for (auto iter = stream_workspace_blocks_.begin(); iter != stream_workspace_blocks_.end();
  1134. ++iter) { ReleaseMemorys(iter->second[stream_id], reusable_blocks_[iter->first][stream_id]); });
  1135. if (IsContinuousOutput(node)) {
  1136. (void)ApplyContinuousMemory(node, ranges, is_op_reuse_mem_);
  1137. return SUCCESS;
  1138. }
  1139. for (uint32_t i = 0; i < static_cast<uint32_t>(op_desc->GetOutputsSize()); i++) {
  1140. int64_t size = 0;
  1141. auto output_op_desc = op_desc->GetOutputDescPtr(i);
  1142. if (output_op_desc != nullptr) {
  1143. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS, GELOGI("Get size failed"));
  1144. }
  1145. // fusion: other type's size not means malloc HBM memory
  1146. bool l1_flag = has_mem_type_attr && memorys_type[i] == RT_MEMORY_L1;
  1147. if (l1_flag) {
  1148. GELOGI("fusion: node[%s], output[%s], output memory type [%d]",
  1149. op_desc->GetName().c_str(), op_desc->GetOutputNameByIndex(i).c_str(), memorys_type[i]);
  1150. size = 0;
  1151. }
  1152. std::string peer_name;
  1153. uint32_t peer_input_index = 0;
  1154. bool out_node_set_continuous_input = false;
  1155. bool reset_zero_copy_flag = false;
  1156. bool no_need_assign_memory = ((size == 0) || CheckIsZeroMemNodeType(node->GetType()));
  1157. if (!no_need_assign_memory) {
  1158. out_node_set_continuous_input =
  1159. IsOutNodeSetContinuousInput(node, i, peer_name, peer_input_index, no_need_assign_memory, reset_zero_copy_flag);
  1160. GE_IF_BOOL_EXEC(!no_need_assign_memory,
  1161. no_need_assign_memory = IsAtomicOutputMemory(node, i, is_atomic, out_node_set_continuous_input););
  1162. }
  1163. no_need_assign_memory = (no_need_assign_memory || IsKnownSubgraphData(node));
  1164. if (no_need_assign_memory) {
  1165. zero_memory_list_.emplace_back(node, kOutput, i, false);
  1166. continue;
  1167. }
  1168. // atomic can't be reused
  1169. bool need_change = is_op_reuse_mem_ && out_node_set_continuous_input && is_atomic;
  1170. if (need_change) {
  1171. is_op_reuse_mem_ = false;
  1172. }
  1173. MemoryBlock *mem_block = ApplyOutMemory(node, i, ranges, is_op_reuse_mem_, out_node_set_continuous_input);
  1174. if (mem_block != nullptr) {
  1175. GE_IF_BOOL_EXEC(reset_zero_copy_flag,
  1176. mem_block->is_zero_copy_ = false;
  1177. GELOGI("Node[%s] output[%u] need assign memory before reassign.", op_desc->GetName().c_str(), i););
  1178. node_out_blocks_[node->GetName()].emplace_back(mem_block);
  1179. if (out_node_set_continuous_input) {
  1180. node_continuous_input_blocks_[peer_name][peer_input_index] = mem_block;
  1181. }
  1182. NodeIndexIO node_index_io(node, i, kOut);
  1183. auto iter = anchor_to_symbol_.find(node_index_io.ToString());
  1184. if (iter == anchor_to_symbol_.end()) {
  1185. continue;
  1186. }
  1187. symbol_blocks_[iter->second] = mem_block;
  1188. }
  1189. }
  1190. return SUCCESS;
  1191. }
  1192. ///
  1193. /// @ingroup domi
  1194. /// @brief traverse all nodes outputs and workspace in need, apply memory block considering memory reuse
  1195. /// @param [in/out] ranges memory size provided
  1196. /// @return Status result
  1197. ///
  1198. void BlockMemAssigner::AssignMemoryWithReuse(vector<int64_t> &ranges) {
  1199. (void)ge::GetContext().GetOption(OPTION_EXEC_DISABLE_REUSED_MEMORY, ge_disable_reuse_mem_env_);
  1200. GEEVENT("Reuse memory %s", ge_disable_reuse_mem_env_ == "1" ? "close" : "open");
  1201. string op_no_reuse_mem_str;
  1202. const char *op_no_reuse_mem = std::getenv(OP_NO_REUSE_MEM);
  1203. GE_IF_BOOL_EXEC(op_no_reuse_mem != nullptr, op_no_reuse_mem_str = string(op_no_reuse_mem);
  1204. CheckAndGetOpReuseEnv(op_no_reuse_mem_str, op_no_reuse_mem_vec_, op_reuse_env_valid_););
  1205. for (NodePtr &n : compute_graph_->GetAllNodes()) {
  1206. auto node_op_desc = n->GetOpDesc();
  1207. GE_IF_BOOL_EXEC(node_op_desc == nullptr, continue);
  1208. life_time_ = node_op_desc->GetId();
  1209. int64_t stream_id = node_op_desc->GetStreamId();
  1210. if (AssignOutputMemoryWithReuse(n, ranges) != SUCCESS) {
  1211. return;
  1212. }
  1213. for (auto iter = stream_workspace_blocks_.begin(); iter != stream_workspace_blocks_.end(); ++iter) {
  1214. iter->second[stream_id].clear();
  1215. }
  1216. vector<int64_t> temp;
  1217. GetNodeWorkSpaceSize(n, temp);
  1218. vector<int64_t> workspace_bytes;
  1219. vector<int64_t> tvm_workspace_memory_type;
  1220. bool has_tvm_workspace_mem_type_attr =
  1221. ge::AttrUtils::GetListInt(node_op_desc, TVM_ATTR_NAME_WORKSPACE_TYPE, tvm_workspace_memory_type);
  1222. vector<bool> workspace_reuse_flag;
  1223. GE_IF_BOOL_EXEC(!ge::AttrUtils::GetListBool(node_op_desc, kAttrNameWorkspaceReuseFlag, workspace_reuse_flag),
  1224. GELOGD("OP %s get workspace_reuse_flag attr failed", node_op_desc->GetName().c_str()));
  1225. GELOGI("Assign memory node[%s], size [temp:%zu, memory type size:%zu]", node_op_desc->GetName().c_str(),
  1226. temp.size(), tvm_workspace_memory_type.size());
  1227. if (has_tvm_workspace_mem_type_attr && (temp.size() != tvm_workspace_memory_type.size())) {
  1228. GELOGE(INTERNAL_ERROR, "fusion: node[%s], tvm workspace memory size error![v_temp:%zu, workspace:%zu]",
  1229. n->GetName().c_str(), temp.size(), tvm_workspace_memory_type.size());
  1230. return;
  1231. }
  1232. for (size_t i = 0; i < temp.size(); i++) {
  1233. // fusion: other type's size not means malloc HBM memory
  1234. bool workspace_skip_flag = false;
  1235. if (has_tvm_workspace_mem_type_attr && tvm_workspace_memory_type[i] == RT_MEMORY_L1) {
  1236. GELOGI(
  1237. "fusion: node[%s]workspace index[%d] is not hbm type, add to zero_memory_list, workspace memory type [%ld]",
  1238. node_op_desc->GetName().c_str(), i, tvm_workspace_memory_type[i]);
  1239. workspace_skip_flag = true;
  1240. }
  1241. if (temp[i] == 0 || workspace_skip_flag) {
  1242. zero_memory_list_.emplace_back(n, kWorkspace, static_cast<uint32_t>(i), false);
  1243. continue;
  1244. }
  1245. int64_t memory_type = RT_MEMORY_HBM;
  1246. if (!GetWorkSpaceMemoryType(n, i, memory_type)) {
  1247. GELOGW("Get workspace memory type failed.");
  1248. return;
  1249. }
  1250. MemoryBlock *mem_block = ApplyMemory(GetBlockSize(static_cast<size_t>(temp[i]), ranges),
  1251. static_cast<size_t>(temp[i]), static_cast<size_t>(temp[i]),
  1252. kWorkspace, n, static_cast<uint32_t>(i), workspace_reuse_flag,
  1253. is_op_reuse_mem_, false, memory_type);
  1254. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(mem_block == nullptr, continue, "failed to apply memory block.");
  1255. CheckWorkspaceReuse(workspace_reuse_flag, i, stream_id, mem_block, memory_type);
  1256. }
  1257. for (auto it = reusable_blocks_.begin(); it != reusable_blocks_.end(); ++it) {
  1258. ReleaseInputNodeOutMemory(node_out_blocks_, it->second[stream_id], n);
  1259. }
  1260. }
  1261. GELOGD("Assigned memory blocks:");
  1262. for (auto mem_block : memory_blocks_) {
  1263. GELOGD("%s", mem_block->String().c_str());
  1264. (void)mem_block; // Fix warning
  1265. }
  1266. bool merge_dynamic_batch = false;
  1267. GE_IF_BOOL_EXEC(!(ge_disable_reuse_mem_env_ == "1"), merge_dynamic_batch = MergeDynamicBatchBlocks());
  1268. GE_IF_BOOL_EXEC((!(ge_disable_reuse_mem_env_ == "1") && !merge_dynamic_batch), ReuseBlocksByLifeTime(ranges.size()));
  1269. AssignContinuousBlocks();
  1270. ResizeMemoryBlocks();
  1271. GELOGD("Memory blocks after resize:");
  1272. for (auto mem_block : memory_blocks_) {
  1273. GELOGD("%s", mem_block->String().c_str());
  1274. (void)mem_block; // Fix warning
  1275. }
  1276. }
  1277. void BlockMemAssigner::CheckWorkspaceReuse(const vector<bool> &workspace_reuse_flag, uint32_t index, int64_t stream_id,
  1278. MemoryBlock *mem_block, int64_t memory_type) {
  1279. bool reuse_mem_flag =
  1280. ((workspace_reuse_flag.size() > index) && (workspace_reuse_flag[index] == false)) ? false : true;
  1281. if (reuse_mem_flag) {
  1282. stream_workspace_blocks_[memory_type][stream_id].emplace_back(mem_block);
  1283. }
  1284. }
  1285. void BlockMemAssigner::GetNodeWorkSpaceSize(const NodePtr &node, vector<int64_t> &workspace_memory) {
  1286. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node->GetOpDesc() == nullptr, return, "Op desc is null.");
  1287. vector<int64_t> workspace_byte_nums = node->GetOpDesc()->GetWorkspaceBytes();
  1288. GELOGD("node[%s] size:%zu", node->GetOpDesc()->GetName().c_str(), workspace_byte_nums.size());
  1289. for (int64_t byte_size : workspace_byte_nums) {
  1290. workspace_memory.emplace_back(byte_size);
  1291. GELOGD("push back size:%ld", byte_size);
  1292. }
  1293. }
  1294. // descending order
  1295. static bool CompareBlockMaxSize(MemoryBlock *left, MemoryBlock *right) {
  1296. if (left == nullptr || right == nullptr) {
  1297. return false;
  1298. }
  1299. auto left_max_size = std::max_element(left->RealSizeList().begin(), left->RealSizeList().end());
  1300. if (left_max_size != left->RealSizeList().end()) {
  1301. auto right_max_size = std::max_element(right->RealSizeList().begin(), right->RealSizeList().end());
  1302. if (right_max_size == right->RealSizeList().end() || (*left_max_size > *right_max_size)) {
  1303. return true;
  1304. }
  1305. }
  1306. return false;
  1307. }
  1308. void MergeBlocks(std::vector<MemoryBlock *> &dest, std::vector<MemoryBlock *> &src) {
  1309. for (size_t i = 0; i < dest.size(); ++i) {
  1310. if (i >= src.size()) {
  1311. return;
  1312. }
  1313. if (dest[i] != nullptr && src[i] != nullptr) {
  1314. if (!dest[i]->reuse_mem_ || !src[i]->reuse_mem_) {
  1315. GELOGD("Diff batch's workspace can't be reused, i: %zu, dest[i]: %s, stream: %ld, src[i]: %s, stream: %ld.",
  1316. i, dest[i]->String().c_str(), dest[i]->stream_id_, src[i]->String().c_str(), src[i]->stream_id_);
  1317. continue;
  1318. }
  1319. for (auto &symbol : src[i]->SymbolList()) {
  1320. dest[i]->AddSymbol(symbol);
  1321. }
  1322. for (size_t j = 0; j < src[i]->NodeTypeIndexList().size(); ++j) {
  1323. dest[i]->AddNodeTypeIndex(src[i]->NodeTypeIndexList()[j],
  1324. src[i]->RealSizeList()[j],
  1325. src[i]->NoAlignSizeList()[j]);
  1326. src[i]->deleted_block_ = true;
  1327. }
  1328. }
  1329. }
  1330. }
  1331. bool BlockMemAssigner::MergeDynamicBatchBlocks() {
  1332. bool merged = false;
  1333. std::map<std::string, std::vector<MemoryBlock *>> dynamic_batch_blocks;
  1334. for (auto block : memory_blocks_) {
  1335. if (block == nullptr) {
  1336. continue;
  1337. }
  1338. std::string batch_label;
  1339. if (block->IsSameLabel(batch_label)) {
  1340. dynamic_batch_blocks[batch_label].emplace_back(block);
  1341. }
  1342. }
  1343. auto it = dynamic_batch_blocks.begin();
  1344. auto it_max = it;
  1345. // find max block counts
  1346. for (; it != dynamic_batch_blocks.end(); ++it) {
  1347. if (it->second.size() > it_max->second.size()) {
  1348. it_max = it;
  1349. }
  1350. std::sort(it->second.begin(), it->second.end(), CompareBlockMaxSize);
  1351. }
  1352. if (it_max != dynamic_batch_blocks.end()) {
  1353. GELOGD("MergeDynamicBatch %s block counts %zu", it_max->first.c_str(), it_max->second.size());
  1354. }
  1355. for (it = dynamic_batch_blocks.begin(); it != dynamic_batch_blocks.end(); ++it) {
  1356. if (it != it_max) {
  1357. GELOGD("MergeDynamicBatch from %s to %s", it->first.c_str(), it_max->first.c_str());
  1358. MergeBlocks(it_max->second, it->second);
  1359. merged = true;
  1360. }
  1361. }
  1362. return merged;
  1363. }
  1364. // asending order
  1365. static bool CompareBlockIndex(MemoryBlock *left, MemoryBlock *right) {
  1366. if (left == nullptr || right == nullptr) {
  1367. return false;
  1368. }
  1369. if (left->input_index_ < right->input_index_) {
  1370. return true;
  1371. }
  1372. return false;
  1373. }
  1374. ///
  1375. /// @ingroup domi
  1376. /// @brief order blocks by continuous input index
  1377. /// @param [in] blocks need be processed
  1378. /// @param [in] input blocks need continuous
  1379. /// @param [out] blocks after continuous order
  1380. /// @param [in/out] blocks ordered
  1381. /// @param [in] input or output
  1382. ///
  1383. void ReAssignContinuousBlocks(const std::vector<MemoryBlock *> &org_blocks,
  1384. const std::map<MemoryBlock *, uint32_t> block_map,
  1385. std::vector<MemoryBlock *> &dest_blocks, std::vector<MemoryBlock *> &continuous_blocks,
  1386. const std::string &type) {
  1387. for (auto &memory_block : org_blocks) {
  1388. if (memory_block == nullptr || memory_block->deleted_block_) {
  1389. continue;
  1390. }
  1391. if (block_map.find(memory_block) != block_map.end()) {
  1392. continue;
  1393. }
  1394. dest_blocks.emplace_back(memory_block);
  1395. }
  1396. // add continuous block
  1397. std::sort(continuous_blocks.begin(), continuous_blocks.end(), CompareBlockIndex);
  1398. size_t count = 0;
  1399. for (auto &memory_block : continuous_blocks) {
  1400. GE_IF_BOOL_EXEC(memory_block == nullptr, continue);
  1401. GELOGI("Block continuous %s index:%d", type.c_str(), memory_block->input_index_);
  1402. count++;
  1403. if (count == 1) {
  1404. memory_block->first_continuous_block_ = true;
  1405. }
  1406. if (count == continuous_blocks.size()) {
  1407. memory_block->last_continuous_block_ = true;
  1408. }
  1409. dest_blocks.emplace_back(memory_block);
  1410. }
  1411. }
  1412. void BlockMemAssigner::AssignContinuousBlocks() {
  1413. for (auto &block_map : node_continuous_input_blocks_) {
  1414. std::vector<MemoryBlock *> dest_memory_blocks;
  1415. std::map<MemoryBlock *, uint32_t> continuous_block_map;
  1416. std::vector<MemoryBlock *> continuous_blocks;
  1417. auto it = node_continuous_input_counts_.find(block_map.first);
  1418. GE_IF_BOOL_EXEC(it == node_continuous_input_counts_.end(), continue);
  1419. GELOGI("Node:%s continuous input block count:%zu input count:%u", block_map.first.c_str(), block_map.second.size(),
  1420. it->second);
  1421. GE_IF_BOOL_EXEC(it->second != block_map.second.size(), continue);
  1422. for (auto &it : block_map.second) {
  1423. if (it.second != nullptr) {
  1424. continuous_block_map[it.second] = it.first;
  1425. it.second->input_index_ = it.first;
  1426. continuous_blocks.emplace_back(it.second);
  1427. }
  1428. }
  1429. if (continuous_block_map.size() != continuous_blocks.size()) {
  1430. GELOGW("Node:%s continuous input map size:%zu vector size:%zu", block_map.first.c_str(),
  1431. continuous_block_map.size(), continuous_blocks.size());
  1432. continue;
  1433. }
  1434. ReAssignContinuousBlocks(memory_blocks_, continuous_block_map, dest_memory_blocks, continuous_blocks, "input");
  1435. memory_blocks_.swap(dest_memory_blocks);
  1436. }
  1437. }
  1438. void BlockMemAssigner::ReuseBlocksByLifeTime(size_t range_size) {
  1439. // 1 means block size is same so no need to do this
  1440. if (range_size <= 1) {
  1441. return;
  1442. }
  1443. for (size_t i = 0; i < memory_blocks_.size(); ++i) {
  1444. auto parent = memory_blocks_[i];
  1445. if (parent == nullptr || parent->deleted_block_ || parent->continuous_block_) {
  1446. continue;
  1447. }
  1448. if (parent->reuse_mem_ && !IsPostReuse(parent)) {
  1449. parent->reuse_mem_ = false;
  1450. }
  1451. for (size_t j = i + 1; j < memory_blocks_.size(); ++j) {
  1452. auto child = memory_blocks_[j];
  1453. if (child == nullptr) {
  1454. continue;
  1455. }
  1456. // If node is before atomic_addr_clean node, the continus memory can't be reused.
  1457. if (!parent->NodeTypeIndexList().empty() && child->continuous_block_) {
  1458. auto node = parent->NodeTypeIndexList()[0].node;
  1459. if (node == nullptr || node->GetOpDesc() == nullptr || (node->GetOpDesc()->GetId() < GetAtomicAddrCleanId())) {
  1460. continue;
  1461. }
  1462. }
  1463. parent->AddLifeReuseBlock(child, total_node_depend_stream_life_);
  1464. }
  1465. }
  1466. }
  1467. ///
  1468. /// @ingroup domi_omg
  1469. /// @brief traverse memory size, resize, calculate offset
  1470. /// @param [in&out] memory_blocks_ memory block, after calculating offset
  1471. ///
  1472. void BlockMemAssigner::ResizeMemoryBlocks() {
  1473. for (auto &memory_block : memory_blocks_) {
  1474. if (memory_block == nullptr || memory_block->deleted_block_ || memory_block->is_zero_copy_) {
  1475. continue;
  1476. }
  1477. if (memory_block->memory_type_ == RT_MEMORY_HBM) {
  1478. if (memory_block->first_continuous_block_) {
  1479. mem_offset_ += MEM_ALIGN_SIZE;
  1480. }
  1481. memory_block->Resize();
  1482. memory_block->SetHeadOffset(mem_offset_);
  1483. mem_offset_ += memory_block->Size();
  1484. memory_block->SetTailOffset(mem_offset_ - 1);
  1485. } else if (memory_block->memory_type_ == RT_MEMORY_P2P_DDR) {
  1486. if (memory_block->first_continuous_block_) {
  1487. p2p_mem_offset_ += MEM_ALIGN_SIZE;
  1488. }
  1489. memory_block->Resize();
  1490. memory_block->SetHeadOffset(p2p_mem_offset_);
  1491. p2p_mem_offset_ += memory_block->Size();
  1492. memory_block->SetTailOffset(p2p_mem_offset_ - 1);
  1493. }
  1494. }
  1495. GELOGI("mem_offset_ exclude zero_copy_memory is %zu, p2p_mem_offset_ exclude zero_copy_memory is %zu.",
  1496. mem_offset_, p2p_mem_offset_);
  1497. }
  1498. ///
  1499. /// @ingroup domi
  1500. /// @brief given NodeTypeIndex, set offset in Op's OpDef
  1501. /// @param [in&out] node_type_index <node, memory type, id>
  1502. /// @param [in] offset offset to be set
  1503. /// @param [in] size memory size
  1504. /// @param [in] real_size memory size in need
  1505. /// @return Status result
  1506. ///
  1507. void SetOffsetSize(const NodeTypeIndex &node_type, const MemoryBlock *block,
  1508. size_t real_size, size_t no_align_size, bool child_block) {
  1509. ge::OpDescPtr op_desc = node_type.node->GetOpDesc();
  1510. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(op_desc == nullptr, return, "op_desc is null.");
  1511. string graph_name = node_type.node->GetOwnerComputeGraph()->GetName();
  1512. vector<int64_t> memorys_type;
  1513. int64_t offset = block->HeadOffset();
  1514. size_t end = node_type.life_time_end;
  1515. bool has_mem_type_attr = ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_OUTPUT_MEM_TYPE_LIST, memorys_type);
  1516. if (node_type.mem_type == kOutput) {
  1517. vector<int64_t> output_list = op_desc->GetOutputOffset();
  1518. for (auto i = static_cast<uint32_t>(output_list.size()); i < node_type.index + 1; i++) {
  1519. output_list.emplace_back(kInvalidOffset);
  1520. }
  1521. if (output_list.empty()) {
  1522. GELOGW("Empty output");
  1523. return;
  1524. }
  1525. static const set<string> kSetOffsetTypes = { DATA_TYPE, AIPP_DATA_TYPE, MULTISHAPE, NETOUTPUT };
  1526. if ((kSetOffsetTypes.count(op_desc->GetType()) > 0) && !IsKnownSubgraphData(node_type.node)) {
  1527. if ((output_list[node_type.index] == kInvalidOffset) || (output_list[node_type.index] < offset)) {
  1528. output_list.at(node_type.index) = offset;
  1529. }
  1530. } else {
  1531. // fusion: keep the original other type offset value from op_desc
  1532. bool set_out_offset = (!has_mem_type_attr) ||
  1533. (memorys_type.size() > node_type.index && memorys_type[node_type.index] != RT_MEMORY_L1);
  1534. if (set_out_offset) {
  1535. output_list.at(node_type.index) = offset;
  1536. }
  1537. }
  1538. op_desc->SetOutputOffset(output_list);
  1539. } else if (node_type.mem_type == kWorkspace) {
  1540. vector<int64_t> workspace_list;
  1541. workspace_list = op_desc->GetWorkspace();
  1542. for (auto i = static_cast<uint32_t>(workspace_list.size()); i < node_type.index + 1; i++) {
  1543. workspace_list.emplace_back(kInvalidOffset);
  1544. }
  1545. vector<int64_t> workspace_mem_type;
  1546. bool has_workspace_mem_type = ge::AttrUtils::GetListInt(op_desc, TVM_ATTR_NAME_WORKSPACE_TYPE, workspace_mem_type);
  1547. // fusion: keep the original other type offset value from op_desc
  1548. bool set_workspace_offset = (!has_workspace_mem_type) ||
  1549. (workspace_mem_type.size() > node_type.index && workspace_mem_type[node_type.index] != RT_MEMORY_L1);
  1550. if (set_workspace_offset) {
  1551. workspace_list.at(node_type.index) = offset;
  1552. }
  1553. op_desc->SetWorkspace(workspace_list);
  1554. }
  1555. GELOGI("[IMAS]Set %s name[%s] %s[%u] offset to [%ld] streamid[%ld] size[%zu] realsize[%zu]"
  1556. " noalignsize[%zu] life time begin[%zu] life time end[%zu] child[%d:%d:%d:%d] isref[%d].", graph_name.c_str(),
  1557. op_desc->GetName().c_str(), node_type.GetMemType().c_str(), node_type.index, offset, op_desc->GetStreamId(),
  1558. block->Size(), real_size, no_align_size, op_desc->GetId(), end, child_block, block->reuse_mem_,
  1559. block->continuous_block_, block->deleted_block_, node_type.ref_input);
  1560. }
  1561. void SetBlockOpMemOffset(MemoryBlock *block, bool child_block) {
  1562. if (block == nullptr) {
  1563. return;
  1564. }
  1565. size_t index = 0;
  1566. size_t real_size = 0;
  1567. size_t no_align_size = 0;
  1568. auto real_size_list_size = block->RealSizeList().size();
  1569. for (const NodeTypeIndex &node_type_index : block->NodeTypeIndexList()) {
  1570. if (index < real_size_list_size) {
  1571. real_size = block->RealSizeList()[index];
  1572. no_align_size = block->NoAlignSizeList()[index];
  1573. }
  1574. SetOffsetSize(node_type_index, block, real_size, no_align_size, child_block);
  1575. index++;
  1576. }
  1577. }
  1578. void BlockMemAssigner::SetOpMemOffset(bool is_zero_copy) {
  1579. for (MemoryBlock *memory_block : memory_blocks_) {
  1580. if (memory_block == nullptr || memory_block->deleted_block_) {
  1581. continue;
  1582. }
  1583. if ((is_zero_copy && !memory_block->is_zero_copy_) || (!is_zero_copy && memory_block->is_zero_copy_)) {
  1584. continue;
  1585. }
  1586. SetBlockOpMemOffset(memory_block, false);
  1587. for (MemoryBlock *child_block : memory_block->ChildBlockList()) {
  1588. SetBlockOpMemOffset(child_block, true);
  1589. }
  1590. }
  1591. if (!is_zero_copy) {
  1592. for (const NodeTypeIndex &node_type_index : zero_memory_list_) {
  1593. MemoryBlock block(0, 0);
  1594. SetOffsetSize(node_type_index, &block, 0, 0, false);
  1595. }
  1596. }
  1597. }
  1598. Status BlockMemAssigner::Assign() {
  1599. vector<int64_t> ranges;
  1600. if (GetMemoryRanges(ranges) != SUCCESS) {
  1601. GELOGE(FAILED, "GetMemoryRanges Fail!");
  1602. return FAILED;
  1603. }
  1604. GE_IF_BOOL_EXEC(ranges.empty(), return SUCCESS);
  1605. AssignMemoryWithReuse(ranges);
  1606. SetOpMemOffset(false);
  1607. return SUCCESS;
  1608. }
  1609. bool BlockMemAssigner::CheckIsZeroMemNodeType(const string &node_type) const {
  1610. return (node_type == VARIABLE) || (node_type == CONSTANT) || (node_type == MULTISHAPE) ||
  1611. (node_type == HCOMBROADCAST) || (node_type == CONSTANTOP) ||
  1612. (node_type == ASSIGNADD) || (node_type == ASSIGNSUB) || (node_type == ASSIGN) || (node_type == HVDWAIT) ||
  1613. (node_type == HVDCALLBACKBROADCAST);
  1614. }
  1615. bool BlockMemAssigner::GetWorkSpaceMemoryType(const NodePtr &node, size_t index, int64_t &memory_type) {
  1616. memory_type = RT_MEMORY_HBM;
  1617. vector<int64_t> workspace_memory_type;
  1618. auto op_desc = node->GetOpDesc();
  1619. bool has_workspace_mem_type_attr =
  1620. ge::AttrUtils::GetListInt(op_desc, TVM_ATTR_NAME_WORKSPACE_TYPE, workspace_memory_type);
  1621. if (has_workspace_mem_type_attr && (workspace_memory_type.size() <= index)) {
  1622. GELOGE(INTERNAL_ERROR, "node[%s], workspace_memory size error![index:%zu, workspace:%zu]",
  1623. node->GetName().c_str(), index, workspace_memory_type.size());
  1624. return false;
  1625. }
  1626. memory_type = has_workspace_mem_type_attr ? workspace_memory_type[index] : RT_MEMORY_HBM;
  1627. return true;
  1628. }
  1629. } // 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