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 70 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
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756
  1. /**
  2. * Copyright 2019-2020 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, bool &no_need_assign_memory) {
  505. if (n == nullptr || n->GetAllOutDataAnchors().size() <= 0) {
  506. return false;
  507. }
  508. if (static_cast<size_t>(out_index) < n->GetAllOutDataAnchors().size()) {
  509. auto out_anchor = n->GetOutDataAnchor(out_index);
  510. GE_IF_BOOL_EXEC(out_anchor == nullptr,
  511. GELOGE(FAILED, "Node[%s] output[%u] anchor is null.", n->GetName().c_str(), out_index);
  512. return false;);
  513. for (auto const &peer_in_anchor : out_anchor->GetPeerInDataAnchors()) {
  514. GE_IF_BOOL_EXEC(peer_in_anchor == nullptr,
  515. GELOGE(FAILED, "Node[%s] output[%u] peer_in_anchor 0 is null.", n->GetName().c_str(), out_index);
  516. return false;);
  517. auto peer_node = peer_in_anchor->GetOwnerNode();
  518. GE_IF_BOOL_EXEC(peer_node == nullptr,
  519. GELOGE(FAILED, "Node[%s] output[%u] node is null.", n->GetName().c_str(), out_index);
  520. return false;);
  521. // Get the continuous input type of the node, default is false
  522. bool is_input_continuous = false;
  523. auto peer_in_node_desc = peer_node->GetOpDesc();
  524. GE_IF_BOOL_EXEC(peer_in_node_desc == nullptr,
  525. GELOGE(FAILED, "Node[%s] output[%u] nodedesc is null.", n->GetName().c_str(), out_index);
  526. return false;);
  527. // If GetBool fail, is_input_continuous is false.
  528. (void)ge::AttrUtils::GetBool(peer_in_node_desc, ATTR_NAME_CONTINUOUS_INPUT, is_input_continuous);
  529. GE_IF_BOOL_EXEC(is_input_continuous && CheckIsZeroMemNodeType(peer_node->GetType()),
  530. GELOGI("Node[%s] output[%u] no_need_assign_memory.", n->GetName().c_str(), out_index);
  531. no_need_assign_memory = true;
  532. return false;);
  533. if (is_input_continuous) {
  534. if (n->GetOwnerComputeGraph() != nullptr) {
  535. string graph_name = n->GetOwnerComputeGraph()->GetName();
  536. GELOGI("%s name[%s] output[%u] node[%s] set input[%d] continuous, input size[%u].", graph_name.c_str(),
  537. n->GetName().c_str(), out_index, peer_in_node_desc->GetName().c_str(), peer_in_anchor->GetIdx(),
  538. peer_node->GetAllInDataAnchorsSize());
  539. // Only set attr one times.
  540. if (node_continuous_input_blocks_[peer_in_node_desc->GetName()].size() == 0) {
  541. (void)ge::AttrUtils::SetBool(peer_in_node_desc, ATTR_NAME_CONTINUOUS_INPUT_ALLOC, true);
  542. node_continuous_input_counts_[peer_in_node_desc->GetName()] = peer_node->GetAllInDataAnchorsSize();
  543. }
  544. peer_input_index = peer_in_anchor->GetIdx();
  545. peer_name = peer_in_node_desc->GetName();
  546. return true;
  547. }
  548. }
  549. }
  550. }
  551. return false;
  552. }
  553. ///
  554. /// @ingroup GE
  555. /// @brief Check pre_reuse flag & post_reuse glag for each symbol
  556. /// @return void
  557. ///
  558. void BlockMemAssigner::InitReuseFlag() {
  559. static const std::set<std::string> kPreReuseTypes = { ge::DATA_TYPE, ge::AIPP_DATA_TYPE, ge::ANN_DATA_TYPE,
  560. ge::NETOUTPUT, ge::PROPOSAL, ge::ZEROSLIKE,
  561. ge::CONSTANT, ge::CONSTANTOP };
  562. static const std::set<std::string> kPostReuseTypes = { ge::DATA_TYPE, ge::AIPP_DATA_TYPE, ge::ENTER, ge::REFENTER,
  563. ge::NEXTITERATION, ge::REFNEXTITERATION };
  564. for (const auto &pair : symbol_to_anchors_) {
  565. std::string symbol = pair.first;
  566. bool pre_reuse_flag = true;
  567. bool post_reuse_flag = true;
  568. // default memory type
  569. int64_t mem_type = RT_MEMORY_HBM;
  570. GetSymbolMemType(pair.second, mem_type);
  571. GELOGD("The memory type of symbol[%s] is [%ld]].", symbol.c_str(), mem_type);
  572. if (mem_type != RT_MEMORY_HBM) {
  573. UpdateOpTensorMemType(pair.second, mem_type);
  574. }
  575. // Only the memory with special requirements is processed. The HBM uses the default processing mode.
  576. if (mem_type != RT_MEMORY_HBM) {
  577. symbol_to_mem_type_[symbol] = mem_type;
  578. }
  579. for (const auto &node_index_io : pair.second) {
  580. if (node_index_io.io_type_ == kIn) {
  581. continue;
  582. }
  583. OutDataAnchorPtr out_anchor = node_index_io.node_->GetOutDataAnchor(node_index_io.index_);
  584. if (out_anchor == nullptr) {
  585. continue;
  586. }
  587. bool out_flg = false;
  588. if (node_index_io.node_->GetOutDataNodes().empty()) {
  589. out_flg = true;
  590. }
  591. for (const auto &in_anchor : out_anchor->GetPeerInDataAnchors()) {
  592. if (IsDirectOutputNode(in_anchor->GetOwnerNode(), in_anchor->GetIdx())) {
  593. out_flg = true;
  594. break;
  595. }
  596. }
  597. const std::string &type = out_anchor->GetOwnerNode()->GetType();
  598. pre_reuse_flag = pre_reuse_flag && !out_flg && (kPreReuseTypes.count(type) == 0);
  599. post_reuse_flag = post_reuse_flag && (kPostReuseTypes.count(type) == 0);
  600. if (!pre_reuse_flag && !post_reuse_flag) {
  601. break;
  602. }
  603. }
  604. pre_reuse_flag_[symbol] = pre_reuse_flag;
  605. post_reuse_flag_[symbol] = post_reuse_flag;
  606. }
  607. }
  608. ///
  609. /// @ingroup GE
  610. /// @brief get pre_reuse flag
  611. /// @param [in] node
  612. /// @param [in] out_index
  613. /// @return bool
  614. ///
  615. bool BlockMemAssigner::IsPreReuse(const NodePtr &node, uint32_t out_index) const {
  616. OutDataAnchorPtr out_data_anchor = nullptr;
  617. if (static_cast<size_t>(out_index) < node->GetAllOutDataAnchors().size()) {
  618. out_data_anchor = node->GetOutDataAnchor(out_index);
  619. }
  620. if (out_data_anchor == nullptr) {
  621. return false;
  622. }
  623. NodeIndexIO cur_node_index_io(out_data_anchor->GetOwnerNode(), out_data_anchor->GetIdx(), kOut);
  624. auto iter1 = anchor_to_symbol_.find(cur_node_index_io.ToString());
  625. if (iter1 == anchor_to_symbol_.end()) {
  626. return false;
  627. }
  628. const std::string &symbol = iter1->second;
  629. auto iter2 = pre_reuse_flag_.find(symbol);
  630. if (iter2 == pre_reuse_flag_.end()) {
  631. return false;
  632. }
  633. return iter2->second;
  634. }
  635. ///
  636. /// @ingroup GE
  637. /// @brief get post_reuse flag
  638. /// @param [in] mem_block
  639. /// @return bool
  640. ///
  641. bool BlockMemAssigner::IsPostReuse(const MemoryBlock *mem_block) const {
  642. if (mem_block == nullptr) {
  643. return false;
  644. }
  645. for (const auto &symbol : mem_block->SymbolList()) {
  646. auto iter = post_reuse_flag_.find(symbol);
  647. if (iter == post_reuse_flag_.end()) {
  648. continue;
  649. }
  650. if (!iter->second) {
  651. return false;
  652. }
  653. }
  654. return true;
  655. }
  656. ///
  657. /// @ingroup GE
  658. /// @brief check if symbol of cur node_index_io has block
  659. /// @param [in] node_index_io
  660. /// @param [out] symbol
  661. /// @return bool
  662. ///
  663. bool BlockMemAssigner::IsSymbolExist(const NodeIndexIO &node_index_io, string &symbol) {
  664. auto iter = anchor_to_symbol_.find(node_index_io.ToString());
  665. if (iter == anchor_to_symbol_.end()) {
  666. return false;
  667. }
  668. symbol = iter->second;
  669. return symbol_blocks_.find(iter->second) != symbol_blocks_.end();
  670. }
  671. ///
  672. /// @ingroup GE
  673. /// @brief Print symbol
  674. /// @return void
  675. ///
  676. void BlockMemAssigner::PrintSymbolMap() {
  677. for (const auto &pair : symbol_to_anchors_) {
  678. GELOGD("symbol=%s, max_size=%zu, pre_reuse=%s, post_reuse=%s", pair.first.c_str(), symbol_size_[pair.first],
  679. pre_reuse_flag_[pair.first] ? "true" : "false", post_reuse_flag_[pair.first] ? "true" : "false");
  680. for (const auto &node_index_io : pair.second) {
  681. GELOGD("anchor:%s", node_index_io.ToString().c_str());
  682. }
  683. }
  684. }
  685. void BlockMemAssigner::GetSymbolMemType(std::list<NodeIndexIO> node_index_io_list, int64_t &memory_type) {
  686. memory_type = RT_MEMORY_HBM;
  687. vector<int64_t> memory_types;
  688. for (auto &node_index_io : node_index_io_list) {
  689. auto op_desc = node_index_io.node_->GetOpDesc();
  690. if (op_desc == nullptr) {
  691. GELOGW("Node[%s] op desc is null.", node_index_io.node_->GetName().c_str());
  692. return;
  693. }
  694. if (node_index_io.io_type_ == kIn) {
  695. vector<int64_t> input_memory_types;
  696. (void) ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_INPUT_MEM_TYPE_LIST, input_memory_types);
  697. if (!input_memory_types.empty() && node_index_io.index_ < input_memory_types.size()) {
  698. int64_t input_memory_type = input_memory_types[node_index_io.index_];
  699. GELOGD("Node[%s]: the memory type of input index [%u] is [%ld]].", op_desc->GetName().c_str(),
  700. node_index_io.index_, input_memory_type);
  701. memory_types.emplace_back(input_memory_type);
  702. }
  703. }
  704. if (node_index_io.io_type_ == kOut) {
  705. vector<int64_t> output_memory_types;
  706. (void) ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_OUTPUT_MEM_TYPE_LIST, output_memory_types);
  707. if (!output_memory_types.empty() && node_index_io.index_ < output_memory_types.size()) {
  708. int64_t output_memory_type = output_memory_types[node_index_io.index_];
  709. GELOGD("Node[%s]: the memory type of output index [%u] is [%ld]].", op_desc->GetName().c_str(),
  710. node_index_io.index_, output_memory_type);
  711. memory_types.emplace_back(output_memory_type);
  712. }
  713. }
  714. }
  715. // memory priority
  716. for (auto node_memory_type : memory_types) {
  717. if (node_memory_type > memory_type) {
  718. memory_type = node_memory_type;
  719. }
  720. }
  721. }
  722. void BlockMemAssigner::UpdateOpTensorMemType(std::list<NodeIndexIO> node_index_io_list, int64_t memory_type) {
  723. for (auto &node_index_io : node_index_io_list) {
  724. auto op_desc = node_index_io.node_->GetOpDesc();
  725. if (op_desc == nullptr) {
  726. GELOGW("Node[%s] op desc is null.", node_index_io.node_->GetName().c_str());
  727. return;
  728. }
  729. if (node_index_io.io_type_ == kIn) {
  730. auto input_desc = op_desc->GetInputDesc(node_index_io.index_);
  731. (void) AttrUtils::SetInt(input_desc, ATTR_NAME_TENSOR_MEM_TYPE, memory_type);
  732. }
  733. if (node_index_io.io_type_ == kOut) {
  734. auto output_desc = op_desc->GetOutputDesc(node_index_io.index_);
  735. (void) AttrUtils::SetInt(output_desc, ATTR_NAME_TENSOR_MEM_TYPE, memory_type);
  736. }
  737. }
  738. }
  739. bool BlockMemAssigner::IsContinuousOutput(const NodePtr &n) {
  740. if (n == nullptr) {
  741. GELOGE(FAILED, "Node is null.");
  742. return false;
  743. }
  744. // Get the continuous output type of the node, default is false
  745. bool is_output_continuous = false;
  746. auto node_desc = n->GetOpDesc();
  747. if (node_desc == nullptr) {
  748. GELOGE(FAILED, "Node[%s] nodedesc is null.", n->GetName().c_str());
  749. return false;
  750. }
  751. // If GetBool fail, is_output_continuous is false.
  752. (void)ge::AttrUtils::GetBool(node_desc, ATTR_NAME_CONTINUOUS_OUTPUT, is_output_continuous);
  753. if (is_output_continuous) {
  754. if (n->GetOwnerComputeGraph() != nullptr) {
  755. string graph_name = n->GetOwnerComputeGraph()->GetName();
  756. GELOGI("%s name[%s] set continuous, output size[%u].", graph_name.c_str(),
  757. n->GetName().c_str(), n->GetAllOutDataAnchorsSize());
  758. return true;
  759. }
  760. }
  761. return false;
  762. }
  763. bool BlockMemAssigner::IsZeroCopyBlock(const NodePtr &node, bool continuous) {
  764. if (NodeUtils::IsDynamicShape(node)) {
  765. return ((node->GetType() == DATA_TYPE) && !continuous) || (node->GetType() == NETOUTPUT);
  766. }
  767. if ((node->GetType() == DATA_TYPE) && !continuous) {
  768. return !node->GetOpDesc()->HasAttr(ATTR_NAME_PARENT_NODE_INDEX);
  769. }
  770. if (node->GetType() == NETOUTPUT) {
  771. const auto &owner = node->GetOwnerComputeGraph();
  772. return owner->GetParentGraph() == nullptr;
  773. }
  774. return false;
  775. }
  776. MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size, size_t no_align_size,
  777. OpMemoryType mem_type, const NodePtr &n, uint32_t out_index,
  778. const vector<bool> &workspace_reuse_flag, const bool is_op_reuse_mem,
  779. const bool continuous, int64_t memory_type) {
  780. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr, return nullptr, "Input parameter n is null.");
  781. auto node_op_desc = n->GetOpDesc();
  782. GE_IF_BOOL_EXEC(node_op_desc == nullptr, return nullptr);
  783. bool is_reuse_memory = false;
  784. string ge_disable_reuse_mem_env = "0";
  785. (void)ge::GetContext().GetOption(OPTION_EXEC_DISABLE_REUSED_MEMORY, ge_disable_reuse_mem_env);
  786. if (ge_disable_reuse_mem_env != "1") {
  787. bool reuse_mem_flag = (mem_type == kOutput) ? IsPreReuse(n, out_index) :
  788. !((workspace_reuse_flag.size() > out_index) && !workspace_reuse_flag[out_index]);
  789. is_reuse_memory = !node_op_desc->HasAttr(kL2FusionDynamicConvergeOp) &&
  790. !node_op_desc->HasAttr(kOpNoReuseMem) && reuse_mem_flag && is_op_reuse_mem;
  791. auto stream_id = node_op_desc->GetStreamId();
  792. if (is_reuse_memory && !continuous && !reusable_blocks_[memory_type].empty()) {
  793. for (auto it = reusable_blocks_[memory_type][stream_id].begin();
  794. it != reusable_blocks_[memory_type][stream_id].end(); ++it) {
  795. MemoryBlock *reusable_block = *it;
  796. if (!IsPostReuse(reusable_block)) {
  797. reusable_block->reuse_mem_ = false;
  798. GELOGI("Unreusable block.");
  799. continue;
  800. }
  801. // A node can reuse blocks of the same stream and preorder streams
  802. if (CanReuseBySize(reusable_block_counts_, *reusable_block, block_size, real_size, continuous)) {
  803. reusable_block->AddNodeTypeIndex({n, mem_type, out_index, false}, real_size, no_align_size);
  804. if (mem_type == kOutput) {
  805. auto iter = anchor_to_symbol_.find(NodeIndexIO(n, out_index, kOut).ToString());
  806. if (iter != anchor_to_symbol_.end()) {
  807. reusable_block->AddSymbol(iter->second);
  808. }
  809. }
  810. reusable_block->continuous_block_ = continuous;
  811. reusable_block->ref_count_++;
  812. ReduceReusableBlockCount(*reusable_block, reusable_block_counts_);
  813. reusable_blocks_[memory_type][stream_id].erase(it);
  814. return reusable_block;
  815. }
  816. }
  817. }
  818. }
  819. auto block = new (std::nothrow) MemoryBlock(block_size, node_op_desc->GetStreamId(), is_reuse_memory, memory_type);
  820. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(block == nullptr, return nullptr, "new an object failed.");
  821. // Data and netoutput need zero copy block
  822. block->is_zero_copy_ = IsZeroCopyBlock(n, continuous);
  823. block->Init(real_size, mem_type, n, out_index, no_align_size);
  824. block->stream_id_ = node_op_desc->GetStreamId();
  825. block->ref_count_++;
  826. block->continuous_block_ = continuous;
  827. if (mem_type == kOutput) {
  828. auto iter = anchor_to_symbol_.find(NodeIndexIO(n, out_index, kOut).ToString());
  829. if (iter != anchor_to_symbol_.end()) {
  830. block->AddSymbol(iter->second);
  831. }
  832. }
  833. memory_blocks_.emplace_back(block);
  834. // cause memory_blocks_ may reduce when swap after,
  835. // create blocks_store_ to assure blocks deleted finally
  836. blocks_store_.emplace_back(block);
  837. return block;
  838. }
  839. MemoryBlock *BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<int64_t> &ranges,
  840. const bool is_op_reuse_mem) {
  841. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr, return nullptr, "input node is null.");
  842. auto node_op_desc = n->GetOpDesc();
  843. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node_op_desc == nullptr, return nullptr, "node_op_desc is null.");
  844. MemoryBlock *block = nullptr;
  845. int64_t total_size = 0;
  846. int64_t memory_type = RT_MEMORY_HBM;
  847. for (uint32_t index = 0; index < static_cast<uint32_t>(node_op_desc->GetOutputsSize()); index++) {
  848. auto output_op_desc = node_op_desc->GetOutputDescPtr(index);
  849. if (output_op_desc == nullptr) {
  850. return nullptr;
  851. }
  852. int64_t size = 0;
  853. if (ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS) {
  854. GELOGI("Get size failed");
  855. return nullptr;
  856. }
  857. size_t align_size = static_cast<size_t>(size);
  858. AlignMemOffset(align_size);
  859. total_size += align_size;
  860. // only apply total size in first block
  861. if (index != 0) {
  862. zero_memory_list_.emplace_back(n, kOutput, index);
  863. }
  864. if (index == 0) {
  865. NodeIndexIO node_index_io(n, index, kOut);
  866. auto iter = anchor_to_symbol_.find(node_index_io.ToString());
  867. if (iter != anchor_to_symbol_.end()) {
  868. string symbol = iter->second;
  869. if (symbol_to_mem_type_.find(symbol) != symbol_to_mem_type_.end()) {
  870. memory_type = symbol_to_mem_type_[symbol];
  871. GELOGD("Continuous out memory symbol is [%s], memory type is [%ld]", symbol.c_str(), memory_type);
  872. }
  873. }
  874. }
  875. }
  876. auto block_size = GetBlockSize(total_size, ranges);
  877. GELOGI("Node[%s] continuous out memory size[%ld] block size[%zu]", node_op_desc->GetName().c_str(),
  878. total_size, block_size);
  879. vector<bool> workspace_reuse_flag;
  880. block = ApplyMemory(block_size, total_size, total_size, kOutput, n, 0, workspace_reuse_flag, is_op_reuse_mem, true,
  881. memory_type);
  882. if (block != nullptr) {
  883. // hccl task need align header and tail
  884. block->first_continuous_block_ = true;
  885. block->last_continuous_block_ = true;
  886. }
  887. return block;
  888. }
  889. MemoryBlock *BlockMemAssigner::ApplyOutMemory(const NodePtr &n, uint32_t index, const vector<int64_t> &ranges,
  890. const bool is_op_reuse_mem, const bool continuous) {
  891. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr, return nullptr, "input node is null.");
  892. auto node_op_desc = n->GetOpDesc();
  893. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node_op_desc == nullptr, return nullptr, "node_op_desc is null.");
  894. MemoryBlock *block = nullptr;
  895. NodeIndexIO node_index_io(n, index, kOut);
  896. int64_t size = 0;
  897. auto output_op_desc = node_op_desc->GetOutputDescPtr(index);
  898. if (output_op_desc != nullptr) {
  899. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS, GELOGI("Get size failed"));
  900. }
  901. size_t no_align_size = 0;
  902. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(GetNoAlignSize(*node_op_desc, index, no_align_size) != SUCCESS,
  903. return nullptr, "Get no align size failed");
  904. std::string symbol;
  905. if (IsSymbolExist(node_index_io, symbol)) {
  906. block = symbol_blocks_[symbol];
  907. block->AddNodeTypeIndex({n, kOutput, index, true}, size, no_align_size);
  908. block->ref_count_++;
  909. } else {
  910. int64_t max_size = size;
  911. int64_t memory_type = RT_MEMORY_HBM;
  912. auto iter1 = anchor_to_symbol_.find(node_index_io.ToString());
  913. if (iter1 != anchor_to_symbol_.end()) {
  914. auto iter2 = symbol_size_.find(iter1->second);
  915. if (iter2 != symbol_size_.end()) {
  916. max_size = iter2->second;
  917. }
  918. auto iter3 = symbol_to_mem_type_.find(iter1->second);
  919. if (iter3 != symbol_to_mem_type_.end()) {
  920. memory_type = iter3->second;
  921. }
  922. }
  923. auto block_size = GetBlockSize(max_size, ranges);
  924. vector<bool> workspace_reuse_flag;
  925. block = ApplyMemory(block_size, size, no_align_size, kOutput, n, index,
  926. workspace_reuse_flag, is_op_reuse_mem, continuous, memory_type);
  927. }
  928. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(block == nullptr, return nullptr, "Block is nullptr.");
  929. int out_count_reuse_input = block->ref_count_;
  930. int out_count = 0;
  931. GE_IF_BOOL_EXEC(index >= n->GetAllOutDataAnchors().size(), GELOGE(FAILED, "index is out of range."); return nullptr);
  932. auto out_data_anchor = n->GetOutDataAnchor(index);
  933. GE_IF_BOOL_EXEC(out_data_anchor == nullptr, GELOGE(FAILED, "Out data anchor is nullptr."); return nullptr);
  934. for (const auto &in_anchor : out_data_anchor->GetPeerInDataAnchors()) {
  935. auto owner_node = in_anchor->GetOwnerNode();
  936. auto op_desc = owner_node->GetOpDesc();
  937. GE_IF_BOOL_EXEC(op_desc == nullptr, continue);
  938. Params *instance = Params::Instance();
  939. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(instance == nullptr, return nullptr, "Params instance is nullptr.");
  940. if (!((instance->GetTarget() == TARGET_TYPE_TINY) && (op_desc->GetType() == NETOUTPUT))) {
  941. out_count++;
  942. }
  943. }
  944. bool reuse_input = false;
  945. for (const auto &in_anchor : out_data_anchor->GetPeerInDataAnchors()) {
  946. auto owner_node = in_anchor->GetOwnerNode();
  947. GE_IF_BOOL_EXEC(owner_node == nullptr, continue);
  948. auto op_desc = owner_node->GetOpDesc();
  949. GE_IF_BOOL_EXEC(op_desc == nullptr, continue);
  950. for (uint32_t i = 0; i < static_cast<uint32_t>(op_desc->GetOutputsSize()); i++) {
  951. bool dst_reuse_input = false;
  952. uint32_t dst_reuse_input_index = 0;
  953. auto owner_node_op_desc = op_desc->GetOutputDescPtr(i);
  954. GE_IF_BOOL_EXEC(owner_node_op_desc == nullptr, continue);
  955. GE_IF_BOOL_EXEC(ge::TensorUtils::GetReuseInput(*owner_node_op_desc, dst_reuse_input) != SUCCESS,
  956. GELOGI("Get dst_reuse_input failed"));
  957. GE_IF_BOOL_EXEC(ge::TensorUtils::GetReuseInputIndex(*owner_node_op_desc, dst_reuse_input_index) != SUCCESS,
  958. GELOGI("Get dst_reuse_input_index failed"));
  959. if (dst_reuse_input && (dst_reuse_input_index == static_cast<uint32_t>(in_anchor->GetIdx()))) {
  960. block->AddNodeTypeIndex({owner_node, kOutput, i, true}, block->Size(), block->Size());
  961. out_count_reuse_input += 1;
  962. reuse_input = true;
  963. }
  964. }
  965. }
  966. block->ref_count_ = reuse_input ? out_count_reuse_input + out_count - 1 : out_count;
  967. return block;
  968. }
  969. bool IsOutputBlock(const ge::InDataAnchorPtr &in_data_anchor) {
  970. auto peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  971. GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, GELOGE(FAILED, "Peer out anchor is nullptr."); return false);
  972. auto src = peer_out_anchor->GetOwnerNode();
  973. int32_t index = peer_out_anchor->GetIdx();
  974. auto iter = GetLocalOmgContext().out_nodes_map.find(src->GetName());
  975. if (iter != GetLocalOmgContext().out_nodes_map.end()) {
  976. for (auto id : iter->second) {
  977. if (index == id) {
  978. return true;
  979. }
  980. }
  981. }
  982. return false;
  983. }
  984. // atomic out memory will be reassigned
  985. bool IsAtomicOutputMemory(const ge::NodePtr &node, uint32_t output_index, bool is_atomic,
  986. bool out_node_set_continuous_input) {
  987. auto op_desc = node->GetOpDesc();
  988. if (op_desc == nullptr) {
  989. return false;
  990. }
  991. vector<int64_t> atomic_output_index;
  992. // If GetListInt fail, atomic_output_index is empty.
  993. (void)ge::AttrUtils::GetListInt(op_desc, ATOMIC_ATTR_OUTPUT_INDEX, atomic_output_index);
  994. if (!out_node_set_continuous_input && is_atomic) {
  995. for (auto &index : atomic_output_index) {
  996. if (static_cast<uint32_t>(index) == output_index) {
  997. if (node->GetOwnerComputeGraph() != nullptr) {
  998. string graph_name = node->GetOwnerComputeGraph()->GetName();
  999. GELOGD("[IMAS]Atomic no assign %s name[%s] output[%d] streamid[%ld].", graph_name.c_str(),
  1000. op_desc->GetName().c_str(), index, op_desc->GetStreamId());
  1001. }
  1002. return true;
  1003. }
  1004. }
  1005. }
  1006. return false;
  1007. }
  1008. bool IsKnownSubgraphData(const NodePtr &node) {
  1009. if (NodeUtils::IsDynamicShape(node)) {
  1010. return false;
  1011. }
  1012. return node->GetOpDesc()->HasAttr(ATTR_NAME_PARENT_NODE_INDEX);
  1013. }
  1014. void BlockMemAssigner::ReleaseMemory(MemoryBlock *to_release, vector<MemoryBlock *> &reusable_memory) {
  1015. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(to_release == nullptr, return, "Input parameter to_release is null.");
  1016. GE_CHK_TRUE_EXEC_INFO(to_release->ref_count_ <= 0, return, "Release memory");
  1017. GE_CHK_TRUE_EXEC_INFO(!to_release->reuse_mem_, return, "doesn't reuse memory");
  1018. --to_release->ref_count_;
  1019. if (to_release->ref_count_ == 0) {
  1020. to_release->SetLifeTimeEnd(life_time_);
  1021. reusable_memory.emplace_back(to_release);
  1022. AddReusableBlockCount(*to_release, reusable_block_counts_);
  1023. }
  1024. }
  1025. void BlockMemAssigner::ReleaseMemorys(const vector<MemoryBlock *> &to_releases,
  1026. vector<MemoryBlock *> &reusable_memory) {
  1027. for (auto mem_block : to_releases) {
  1028. ReleaseMemory(mem_block, reusable_memory);
  1029. }
  1030. }
  1031. void BlockMemAssigner::ReleaseInputNodeOutMemory(const unordered_map<string, vector<MemoryBlock *>> &node_out_blocks,
  1032. vector<MemoryBlock *> &reusable_memory, NodePtr &node) {
  1033. for (const auto &in_anchor : node->GetAllInDataAnchors()) {
  1034. if ((in_anchor->GetPeerOutAnchor() == nullptr) ||
  1035. (in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetOpDesc() == nullptr) || (node->GetOpDesc() == nullptr)) {
  1036. return;
  1037. }
  1038. GE_IF_BOOL_EXEC(IsOutputBlock(in_anchor), continue);
  1039. auto node_name = in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetName();
  1040. GE_IF_BOOL_EXEC((in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetType() == CONSTANT) ||
  1041. (in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetType() == FASTRCNNPREDICTIONS) ||
  1042. (in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetType() == CONSTANTOP),
  1043. continue);
  1044. auto it = node_out_blocks.find(node_name);
  1045. if (it == node_out_blocks.end()) {
  1046. continue;
  1047. }
  1048. for (auto block : it->second) {
  1049. const vector<NodeTypeIndex> &node_type_indexs = block->NodeTypeIndexList();
  1050. if (node_type_indexs.empty()) {
  1051. continue;
  1052. }
  1053. GELOGD("node_type_indexs: %d, %s", node_type_indexs.back().index,
  1054. node_type_indexs.back().node->GetName().c_str());
  1055. if ((node_type_indexs.back().node == in_anchor->GetPeerOutAnchor()->GetOwnerNode()) &&
  1056. (node_type_indexs.back().index == static_cast<uint32_t>(in_anchor->GetPeerOutAnchor()->GetIdx())) &&
  1057. (node->GetOpDesc()->GetStreamId() == block->stream_id_)) {
  1058. ReleaseMemory(block, reusable_memory);
  1059. if (block->ref_count_ == 0) {
  1060. SetLastUsedInputMemAttr(node, in_anchor->GetIdx());
  1061. }
  1062. }
  1063. }
  1064. }
  1065. }
  1066. void SplitStringByComma(const string &str, vector<string> &sub_str_vec) {
  1067. std::string tmp_string = str + ",";
  1068. std::string::size_type start_pos = 0;
  1069. std::string::size_type cur_pos = tmp_string.find(',', 0);
  1070. while (cur_pos != std::string::npos) {
  1071. std::string sub_str = tmp_string.substr(start_pos, cur_pos - start_pos);
  1072. if (!sub_str.empty()) {
  1073. vector<string>::iterator ret = std::find(sub_str_vec.begin(), sub_str_vec.end(), sub_str);
  1074. if (ret == sub_str_vec.end()) {
  1075. sub_str_vec.push_back(sub_str);
  1076. }
  1077. }
  1078. start_pos = cur_pos + 1;
  1079. cur_pos = tmp_string.find(',', start_pos);
  1080. }
  1081. }
  1082. void CheckAndGetOpReuseEnv(const string &env, vector<string> &env_vec, bool &op_reuse_env_valid) {
  1083. string env_str;
  1084. env_str = string(env);
  1085. if (env_str.size() > kReuseMaxCharNum) {
  1086. GELOGE(FAILED, "The OP_NO_REUSE_MEM has more than %d characters.", kReuseMaxCharNum);
  1087. return;
  1088. }
  1089. SplitStringByComma(env_str, env_vec);
  1090. if (env_vec.size() > kReuseMaxOpNum) {
  1091. GELOGE(FAILED, "The OP_NO_REUSE_MEM has more than %d nodes.", kReuseMaxOpNum);
  1092. return;
  1093. }
  1094. op_reuse_env_valid = true;
  1095. return;
  1096. }
  1097. Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector<int64_t> &ranges) {
  1098. auto op_desc = node->GetOpDesc();
  1099. int64_t stream_id = op_desc->GetStreamId();
  1100. vector<int64_t> memorys_type;
  1101. bool has_mem_type_attr = ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_OUTPUT_MEM_TYPE_LIST, memorys_type);
  1102. GELOGI("Assign memory node[%s], output size[%d], output memory type size[%d]", op_desc->GetName().c_str(),
  1103. op_desc->GetOutputsSize(), memorys_type.size());
  1104. if (has_mem_type_attr && (memorys_type.size() != op_desc->GetOutputsSize())) {
  1105. GELOGE(INTERNAL_ERROR, "fusion: node[%s], output memory size err[outputsize:%zu, memorysize:%zu]",
  1106. op_desc->GetName().c_str(), op_desc->GetOutputsSize(), memorys_type.size());
  1107. return INTERNAL_ERROR;
  1108. }
  1109. is_op_reuse_mem_ = true;
  1110. if (op_reuse_env_valid_ == true) {
  1111. vector<string>::iterator it_name =
  1112. std::find(op_no_reuse_mem_vec_.begin(), op_no_reuse_mem_vec_.end(), op_desc->GetName());
  1113. vector<string>::iterator it_type =
  1114. std::find(op_no_reuse_mem_vec_.begin(), op_no_reuse_mem_vec_.end(), op_desc->GetType());
  1115. GE_IF_BOOL_EXEC(it_name != op_no_reuse_mem_vec_.end() || it_type != op_no_reuse_mem_vec_.end(),
  1116. is_op_reuse_mem_ = false;);
  1117. }
  1118. bool is_atomic = false;
  1119. // If GetBool fail, is_atomic is false.
  1120. (void)ge::AttrUtils::GetBool(op_desc, ATOMIC_ATTR_IS_ATOMIC_NODE, is_atomic);
  1121. // Allocate memory for the current node and release node memory of the same size in the workspace
  1122. GE_IF_BOOL_EXEC(ge_disable_reuse_mem_env_ != "1",
  1123. for (auto iter = stream_workspace_blocks_.begin(); iter != stream_workspace_blocks_.end();
  1124. ++iter) { ReleaseMemorys(iter->second[stream_id], reusable_blocks_[iter->first][stream_id]); });
  1125. if (IsContinuousOutput(node)) {
  1126. (void)ApplyContinuousMemory(node, ranges, is_op_reuse_mem_);
  1127. return SUCCESS;
  1128. }
  1129. for (uint32_t i = 0; i < static_cast<uint32_t>(op_desc->GetOutputsSize()); i++) {
  1130. int64_t size = 0;
  1131. auto output_op_desc = op_desc->GetOutputDescPtr(i);
  1132. if (output_op_desc != nullptr) {
  1133. GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS, GELOGI("Get size failed"));
  1134. }
  1135. // fusion: other type's size not means malloc HBM memory
  1136. bool l1_flag = has_mem_type_attr && memorys_type[i] == RT_MEMORY_L1;
  1137. if (l1_flag) {
  1138. GELOGI("fusion: node[%s], output[%s], output memory type [%d]",
  1139. op_desc->GetName().c_str(), op_desc->GetOutputNameByIndex(i).c_str(), memorys_type[i]);
  1140. size = 0;
  1141. }
  1142. std::string peer_name;
  1143. uint32_t peer_input_index = 0;
  1144. bool out_node_set_continuous_input = false;
  1145. bool no_need_assign_memory = ((size == 0) || CheckIsZeroMemNodeType(node->GetType()));
  1146. if (!no_need_assign_memory) {
  1147. out_node_set_continuous_input =
  1148. IsOutNodeSetContinuousInput(node, i, peer_name, peer_input_index, no_need_assign_memory);
  1149. GE_IF_BOOL_EXEC(!no_need_assign_memory,
  1150. no_need_assign_memory = IsAtomicOutputMemory(node, i, is_atomic, out_node_set_continuous_input););
  1151. }
  1152. no_need_assign_memory = (no_need_assign_memory || IsKnownSubgraphData(node));
  1153. if (no_need_assign_memory) {
  1154. zero_memory_list_.emplace_back(node, kOutput, i, false);
  1155. continue;
  1156. }
  1157. // atomic can't be reused
  1158. bool need_change = is_op_reuse_mem_ && out_node_set_continuous_input && is_atomic;
  1159. if (need_change) {
  1160. is_op_reuse_mem_ = false;
  1161. }
  1162. MemoryBlock *mem_block = ApplyOutMemory(node, i, ranges, is_op_reuse_mem_, out_node_set_continuous_input);
  1163. if (mem_block != nullptr) {
  1164. node_out_blocks_[node->GetName()].emplace_back(mem_block);
  1165. if (out_node_set_continuous_input) {
  1166. node_continuous_input_blocks_[peer_name][peer_input_index] = mem_block;
  1167. }
  1168. NodeIndexIO node_index_io(node, i, kOut);
  1169. auto iter = anchor_to_symbol_.find(node_index_io.ToString());
  1170. if (iter == anchor_to_symbol_.end()) {
  1171. continue;
  1172. }
  1173. symbol_blocks_[iter->second] = mem_block;
  1174. }
  1175. }
  1176. return SUCCESS;
  1177. }
  1178. ///
  1179. /// @ingroup domi
  1180. /// @brief traverse all nodes outputs and workspace in need, apply memory block considering memory reuse
  1181. /// @param [in/out] ranges memory size provided
  1182. /// @return Status result
  1183. ///
  1184. void BlockMemAssigner::AssignMemoryWithReuse(vector<int64_t> &ranges) {
  1185. (void)ge::GetContext().GetOption(OPTION_EXEC_DISABLE_REUSED_MEMORY, ge_disable_reuse_mem_env_);
  1186. GEEVENT("Reuse memory %s", ge_disable_reuse_mem_env_ == "1" ? "close" : "open");
  1187. string op_no_reuse_mem_str;
  1188. const char *op_no_reuse_mem = std::getenv(OP_NO_REUSE_MEM);
  1189. GE_IF_BOOL_EXEC(op_no_reuse_mem != nullptr, op_no_reuse_mem_str = string(op_no_reuse_mem);
  1190. CheckAndGetOpReuseEnv(op_no_reuse_mem_str, op_no_reuse_mem_vec_, op_reuse_env_valid_););
  1191. for (NodePtr &n : compute_graph_->GetAllNodes()) {
  1192. auto node_op_desc = n->GetOpDesc();
  1193. GE_IF_BOOL_EXEC(node_op_desc == nullptr, continue);
  1194. life_time_ = node_op_desc->GetId();
  1195. int64_t stream_id = node_op_desc->GetStreamId();
  1196. if (AssignOutputMemoryWithReuse(n, ranges) != SUCCESS) {
  1197. return;
  1198. }
  1199. for (auto iter = stream_workspace_blocks_.begin(); iter != stream_workspace_blocks_.end(); ++iter) {
  1200. iter->second[stream_id].clear();
  1201. }
  1202. vector<int64_t> temp;
  1203. GetNodeWorkSpaceSize(n, temp);
  1204. vector<int64_t> workspace_bytes;
  1205. vector<int64_t> tvm_workspace_memory_type;
  1206. bool has_tvm_workspace_mem_type_attr =
  1207. ge::AttrUtils::GetListInt(node_op_desc, TVM_ATTR_NAME_WORKSPACE_TYPE, tvm_workspace_memory_type);
  1208. vector<bool> workspace_reuse_flag;
  1209. GE_IF_BOOL_EXEC(!ge::AttrUtils::GetListBool(node_op_desc, kAttrNameWorkspaceReuseFlag, workspace_reuse_flag),
  1210. GELOGD("OP %s get workspace_reuse_flag attr failed", node_op_desc->GetName().c_str()));
  1211. GELOGI("Assign memory node[%s], size [temp:%zu, memory type size:%zu]", node_op_desc->GetName().c_str(),
  1212. temp.size(), tvm_workspace_memory_type.size());
  1213. if (has_tvm_workspace_mem_type_attr && (temp.size() != tvm_workspace_memory_type.size())) {
  1214. GELOGE(INTERNAL_ERROR, "fusion: node[%s], tvm workspace memory size error![v_temp:%zu, workspace:%zu]",
  1215. n->GetName().c_str(), temp.size(), tvm_workspace_memory_type.size());
  1216. return;
  1217. }
  1218. for (size_t i = 0; i < temp.size(); i++) {
  1219. // fusion: other type's size not means malloc HBM memory
  1220. bool workspace_skip_flag = false;
  1221. if (has_tvm_workspace_mem_type_attr && tvm_workspace_memory_type[i] == RT_MEMORY_L1) {
  1222. GELOGI(
  1223. "fusion: node[%s]workspace index[%d] is not hbm type, add to zero_memory_list, workspace memory type [%ld]",
  1224. node_op_desc->GetName().c_str(), i, tvm_workspace_memory_type[i]);
  1225. workspace_skip_flag = true;
  1226. }
  1227. if (temp[i] == 0 || workspace_skip_flag) {
  1228. zero_memory_list_.emplace_back(n, kWorkspace, static_cast<uint32_t>(i), false);
  1229. continue;
  1230. }
  1231. int64_t memory_type = RT_MEMORY_HBM;
  1232. if (!GetWorkSpaceMemoryType(n, i, memory_type)) {
  1233. GELOGW("Get workspace memory type failed.");
  1234. return;
  1235. }
  1236. MemoryBlock *mem_block = ApplyMemory(GetBlockSize(static_cast<size_t>(temp[i]), ranges),
  1237. static_cast<size_t>(temp[i]), static_cast<size_t>(temp[i]),
  1238. kWorkspace, n, static_cast<uint32_t>(i), workspace_reuse_flag,
  1239. is_op_reuse_mem_, false, memory_type);
  1240. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(mem_block == nullptr, continue, "failed to apply memory block.");
  1241. CheckWorkspaceReuse(workspace_reuse_flag, i, stream_id, mem_block, memory_type);
  1242. }
  1243. for (auto it = reusable_blocks_.begin(); it != reusable_blocks_.end(); ++it) {
  1244. ReleaseInputNodeOutMemory(node_out_blocks_, it->second[stream_id], n);
  1245. }
  1246. }
  1247. GELOGD("Assigned memory blocks:");
  1248. for (auto mem_block : memory_blocks_) {
  1249. GELOGD("%s", mem_block->String().c_str());
  1250. (void)mem_block; // Fix warning
  1251. }
  1252. bool merge_dynamic_batch = false;
  1253. GE_IF_BOOL_EXEC(!(ge_disable_reuse_mem_env_ == "1"), merge_dynamic_batch = MergeDynamicBatchBlocks());
  1254. GE_IF_BOOL_EXEC((!(ge_disable_reuse_mem_env_ == "1") && !merge_dynamic_batch), ReuseBlocksByLifeTime(ranges.size()));
  1255. AssignContinuousBlocks();
  1256. ResizeMemoryBlocks();
  1257. GELOGD("Memory blocks after resize:");
  1258. for (auto mem_block : memory_blocks_) {
  1259. GELOGD("%s", mem_block->String().c_str());
  1260. (void)mem_block; // Fix warning
  1261. }
  1262. }
  1263. void BlockMemAssigner::CheckWorkspaceReuse(const vector<bool> &workspace_reuse_flag, uint32_t index, int64_t stream_id,
  1264. MemoryBlock *mem_block, int64_t memory_type) {
  1265. bool reuse_mem_flag =
  1266. ((workspace_reuse_flag.size() > index) && (workspace_reuse_flag[index] == false)) ? false : true;
  1267. if (reuse_mem_flag) {
  1268. stream_workspace_blocks_[memory_type][stream_id].emplace_back(mem_block);
  1269. }
  1270. }
  1271. void BlockMemAssigner::GetNodeWorkSpaceSize(const NodePtr &node, vector<int64_t> &workspace_memory) {
  1272. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node->GetOpDesc() == nullptr, return, "Op desc is null.");
  1273. vector<int64_t> workspace_byte_nums = node->GetOpDesc()->GetWorkspaceBytes();
  1274. GELOGD("node[%s] size:%zu", node->GetOpDesc()->GetName().c_str(), workspace_byte_nums.size());
  1275. for (int64_t byte_size : workspace_byte_nums) {
  1276. workspace_memory.emplace_back(byte_size);
  1277. GELOGD("push back size:%ld", byte_size);
  1278. }
  1279. }
  1280. // descending order
  1281. static bool CompareBlockMaxSize(MemoryBlock *left, MemoryBlock *right) {
  1282. if (left == nullptr || right == nullptr) {
  1283. return false;
  1284. }
  1285. auto left_max_size = std::max_element(left->RealSizeList().begin(), left->RealSizeList().end());
  1286. if (left_max_size != left->RealSizeList().end()) {
  1287. auto right_max_size = std::max_element(right->RealSizeList().begin(), right->RealSizeList().end());
  1288. if (right_max_size == right->RealSizeList().end() || (*left_max_size > *right_max_size)) {
  1289. return true;
  1290. }
  1291. }
  1292. return false;
  1293. }
  1294. void MergeBlocks(std::vector<MemoryBlock *> &dest, std::vector<MemoryBlock *> &src) {
  1295. for (size_t i = 0; i < dest.size(); ++i) {
  1296. if (i >= src.size()) {
  1297. return;
  1298. }
  1299. if (dest[i] != nullptr && src[i] != nullptr) {
  1300. if (!dest[i]->reuse_mem_ || !src[i]->reuse_mem_) {
  1301. GELOGD("Diff batch's workspace can't be reused, i: %zu, dest[i]: %s, stream: %ld, src[i]: %s, stream: %ld.",
  1302. i, dest[i]->String().c_str(), dest[i]->stream_id_, src[i]->String().c_str(), src[i]->stream_id_);
  1303. continue;
  1304. }
  1305. for (auto &symbol : src[i]->SymbolList()) {
  1306. dest[i]->AddSymbol(symbol);
  1307. }
  1308. for (size_t j = 0; j < src[i]->NodeTypeIndexList().size(); ++j) {
  1309. dest[i]->AddNodeTypeIndex(src[i]->NodeTypeIndexList()[j],
  1310. src[i]->RealSizeList()[j],
  1311. src[i]->NoAlignSizeList()[j]);
  1312. src[i]->deleted_block_ = true;
  1313. }
  1314. }
  1315. }
  1316. }
  1317. bool BlockMemAssigner::MergeDynamicBatchBlocks() {
  1318. bool merged = false;
  1319. std::map<std::string, std::vector<MemoryBlock *>> dynamic_batch_blocks;
  1320. for (auto block : memory_blocks_) {
  1321. if (block == nullptr) {
  1322. continue;
  1323. }
  1324. std::string batch_label;
  1325. if (block->IsSameLabel(batch_label)) {
  1326. dynamic_batch_blocks[batch_label].emplace_back(block);
  1327. }
  1328. }
  1329. auto it = dynamic_batch_blocks.begin();
  1330. auto it_max = it;
  1331. // find max block counts
  1332. for (; it != dynamic_batch_blocks.end(); ++it) {
  1333. if (it->second.size() > it_max->second.size()) {
  1334. it_max = it;
  1335. }
  1336. std::sort(it->second.begin(), it->second.end(), CompareBlockMaxSize);
  1337. }
  1338. if (it_max != dynamic_batch_blocks.end()) {
  1339. GELOGD("MergeDynamicBatch %s block counts %zu", it_max->first.c_str(), it_max->second.size());
  1340. }
  1341. for (it = dynamic_batch_blocks.begin(); it != dynamic_batch_blocks.end(); ++it) {
  1342. if (it != it_max) {
  1343. GELOGD("MergeDynamicBatch from %s to %s", it->first.c_str(), it_max->first.c_str());
  1344. MergeBlocks(it_max->second, it->second);
  1345. merged = true;
  1346. }
  1347. }
  1348. return merged;
  1349. }
  1350. // asending order
  1351. static bool CompareBlockIndex(MemoryBlock *left, MemoryBlock *right) {
  1352. if (left == nullptr || right == nullptr) {
  1353. return false;
  1354. }
  1355. if (left->input_index_ < right->input_index_) {
  1356. return true;
  1357. }
  1358. return false;
  1359. }
  1360. ///
  1361. /// @ingroup domi
  1362. /// @brief order blocks by continuous input index
  1363. /// @param [in] blocks need be processed
  1364. /// @param [in] input blocks need continuous
  1365. /// @param [out] blocks after continuous order
  1366. /// @param [in/out] blocks ordered
  1367. /// @param [in] input or output
  1368. ///
  1369. void ReAssignContinuousBlocks(const std::vector<MemoryBlock *> &org_blocks,
  1370. const std::map<MemoryBlock *, uint32_t> block_map,
  1371. std::vector<MemoryBlock *> &dest_blocks, std::vector<MemoryBlock *> &continuous_blocks,
  1372. const std::string &type) {
  1373. for (auto &memory_block : org_blocks) {
  1374. if (memory_block == nullptr || memory_block->deleted_block_) {
  1375. continue;
  1376. }
  1377. if (block_map.find(memory_block) != block_map.end()) {
  1378. continue;
  1379. }
  1380. dest_blocks.emplace_back(memory_block);
  1381. }
  1382. // add continuous block
  1383. std::sort(continuous_blocks.begin(), continuous_blocks.end(), CompareBlockIndex);
  1384. size_t count = 0;
  1385. for (auto &memory_block : continuous_blocks) {
  1386. GE_IF_BOOL_EXEC(memory_block == nullptr, continue);
  1387. GELOGI("Block continuous %s index:%d", type.c_str(), memory_block->input_index_);
  1388. count++;
  1389. if (count == 1) {
  1390. memory_block->first_continuous_block_ = true;
  1391. }
  1392. if (count == continuous_blocks.size()) {
  1393. memory_block->last_continuous_block_ = true;
  1394. }
  1395. dest_blocks.emplace_back(memory_block);
  1396. }
  1397. }
  1398. void BlockMemAssigner::AssignContinuousBlocks() {
  1399. for (auto &block_map : node_continuous_input_blocks_) {
  1400. std::vector<MemoryBlock *> dest_memory_blocks;
  1401. std::map<MemoryBlock *, uint32_t> continuous_block_map;
  1402. std::vector<MemoryBlock *> continuous_blocks;
  1403. auto it = node_continuous_input_counts_.find(block_map.first);
  1404. GE_IF_BOOL_EXEC(it == node_continuous_input_counts_.end(), continue);
  1405. GELOGI("Node:%s continuous input block count:%zu input count:%u", block_map.first.c_str(), block_map.second.size(),
  1406. it->second);
  1407. GE_IF_BOOL_EXEC(it->second != block_map.second.size(), continue);
  1408. for (auto &it : block_map.second) {
  1409. if (it.second != nullptr) {
  1410. continuous_block_map[it.second] = it.first;
  1411. it.second->input_index_ = it.first;
  1412. continuous_blocks.emplace_back(it.second);
  1413. }
  1414. }
  1415. if (continuous_block_map.size() != continuous_blocks.size()) {
  1416. GELOGW("Node:%s continuous input map size:%zu vector size:%zu", block_map.first.c_str(),
  1417. continuous_block_map.size(), continuous_blocks.size());
  1418. continue;
  1419. }
  1420. ReAssignContinuousBlocks(memory_blocks_, continuous_block_map, dest_memory_blocks, continuous_blocks, "input");
  1421. memory_blocks_.swap(dest_memory_blocks);
  1422. }
  1423. }
  1424. void BlockMemAssigner::ReuseBlocksByLifeTime(size_t range_size) {
  1425. // 1 means block size is same so no need to do this
  1426. if (range_size <= 1) {
  1427. return;
  1428. }
  1429. for (size_t i = 0; i < memory_blocks_.size(); ++i) {
  1430. auto parent = memory_blocks_[i];
  1431. if (parent == nullptr || parent->deleted_block_ || parent->continuous_block_) {
  1432. continue;
  1433. }
  1434. if (parent->reuse_mem_ && !IsPostReuse(parent)) {
  1435. parent->reuse_mem_ = false;
  1436. }
  1437. for (size_t j = i + 1; j < memory_blocks_.size(); ++j) {
  1438. auto child = memory_blocks_[j];
  1439. if (child == nullptr) {
  1440. continue;
  1441. }
  1442. // If node is before atomic_addr_clean node, the continus memory can't be reused.
  1443. if (!parent->NodeTypeIndexList().empty() && child->continuous_block_) {
  1444. auto node = parent->NodeTypeIndexList()[0].node;
  1445. if (node == nullptr || node->GetOpDesc() == nullptr || (node->GetOpDesc()->GetId() < GetAtomicAddrCleanId())) {
  1446. continue;
  1447. }
  1448. }
  1449. parent->AddLifeReuseBlock(child, total_node_depend_stream_life_);
  1450. }
  1451. }
  1452. }
  1453. ///
  1454. /// @ingroup domi_omg
  1455. /// @brief traverse memory size, resize, calculate offset
  1456. /// @param [in&out] memory_blocks_ memory block, after calculating offset
  1457. ///
  1458. void BlockMemAssigner::ResizeMemoryBlocks() {
  1459. for (auto &memory_block : memory_blocks_) {
  1460. if (memory_block == nullptr || memory_block->deleted_block_ || memory_block->is_zero_copy_) {
  1461. continue;
  1462. }
  1463. if (memory_block->memory_type_ == RT_MEMORY_HBM) {
  1464. if (memory_block->first_continuous_block_) {
  1465. mem_offset_ += MEM_ALIGN_SIZE;
  1466. }
  1467. memory_block->Resize();
  1468. memory_block->SetHeadOffset(mem_offset_);
  1469. mem_offset_ += memory_block->Size();
  1470. memory_block->SetTailOffset(mem_offset_ - 1);
  1471. } else {
  1472. if (memory_block->first_continuous_block_) {
  1473. p2p_mem_offset_ += MEM_ALIGN_SIZE;
  1474. }
  1475. memory_block->Resize();
  1476. memory_block->SetHeadOffset(p2p_mem_offset_);
  1477. p2p_mem_offset_ += memory_block->Size();
  1478. memory_block->SetTailOffset(p2p_mem_offset_ - 1);
  1479. }
  1480. }
  1481. GELOGI("mem_offset_ exclude zero_copy_memory is %zu, p2p_mem_offset_ exclude zero_copy_memory is %zu.",
  1482. mem_offset_, p2p_mem_offset_);
  1483. }
  1484. ///
  1485. /// @ingroup domi
  1486. /// @brief given NodeTypeIndex, set offset in Op's OpDef
  1487. /// @param [in&out] node_type_index <node, memory type, id>
  1488. /// @param [in] offset offset to be set
  1489. /// @param [in] size memory size
  1490. /// @param [in] real_size memory size in need
  1491. /// @return Status result
  1492. ///
  1493. void SetOffsetSize(const NodeTypeIndex &node_type, const MemoryBlock *block,
  1494. size_t real_size, size_t no_align_size, bool child_block) {
  1495. ge::OpDescPtr op_desc = node_type.node->GetOpDesc();
  1496. GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(op_desc == nullptr, return, "op_desc is null.");
  1497. string graph_name = node_type.node->GetOwnerComputeGraph()->GetName();
  1498. vector<int64_t> memorys_type;
  1499. int64_t offset = block->HeadOffset();
  1500. size_t end = node_type.life_time_end;
  1501. bool has_mem_type_attr = ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_OUTPUT_MEM_TYPE_LIST, memorys_type);
  1502. if (node_type.mem_type == kOutput) {
  1503. vector<int64_t> output_list = op_desc->GetOutputOffset();
  1504. for (auto i = static_cast<uint32_t>(output_list.size()); i < node_type.index + 1; i++) {
  1505. output_list.emplace_back(kInvalidOffset);
  1506. }
  1507. if (output_list.empty()) {
  1508. GELOGW("Empty output");
  1509. return;
  1510. }
  1511. static const set<string> kSetOffsetTypes = { DATA_TYPE, AIPP_DATA_TYPE, MULTISHAPE, NETOUTPUT };
  1512. if ((kSetOffsetTypes.count(op_desc->GetType()) > 0) && !IsKnownSubgraphData(node_type.node)) {
  1513. if ((output_list[node_type.index] == kInvalidOffset) || (output_list[node_type.index] < offset)) {
  1514. output_list.at(node_type.index) = offset;
  1515. }
  1516. } else {
  1517. // fusion: keep the original other type offset value from op_desc
  1518. bool set_out_offset = (!has_mem_type_attr) ||
  1519. (memorys_type.size() > node_type.index && memorys_type[node_type.index] != RT_MEMORY_L1);
  1520. if (set_out_offset) {
  1521. output_list.at(node_type.index) = offset;
  1522. }
  1523. }
  1524. op_desc->SetOutputOffset(output_list);
  1525. } else if (node_type.mem_type == kWorkspace) {
  1526. vector<int64_t> workspace_list;
  1527. workspace_list = op_desc->GetWorkspace();
  1528. for (auto i = static_cast<uint32_t>(workspace_list.size()); i < node_type.index + 1; i++) {
  1529. workspace_list.emplace_back(kInvalidOffset);
  1530. }
  1531. vector<int64_t> workspace_mem_type;
  1532. bool has_workspace_mem_type = ge::AttrUtils::GetListInt(op_desc, TVM_ATTR_NAME_WORKSPACE_TYPE, workspace_mem_type);
  1533. // fusion: keep the original other type offset value from op_desc
  1534. bool set_workspace_offset = (!has_workspace_mem_type) ||
  1535. (workspace_mem_type.size() > node_type.index && workspace_mem_type[node_type.index] != RT_MEMORY_L1);
  1536. if (set_workspace_offset) {
  1537. workspace_list.at(node_type.index) = offset;
  1538. }
  1539. op_desc->SetWorkspace(workspace_list);
  1540. }
  1541. GELOGI("[IMAS]Set %s name[%s] %s[%u] offset to [%ld] streamid[%ld] size[%zu] realsize[%zu]"
  1542. " noalignsize[%zu] life time begin[%zu] life time end[%zu] child[%d:%d:%d:%d] isref[%d].", graph_name.c_str(),
  1543. op_desc->GetName().c_str(), node_type.GetMemType().c_str(), node_type.index, offset, op_desc->GetStreamId(),
  1544. block->Size(), real_size, no_align_size, op_desc->GetId(), end, child_block, block->reuse_mem_,
  1545. block->continuous_block_, block->deleted_block_, node_type.ref_input);
  1546. }
  1547. void SetBlockOpMemOffset(MemoryBlock *block, bool child_block) {
  1548. if (block == nullptr) {
  1549. return;
  1550. }
  1551. size_t index = 0;
  1552. size_t real_size = 0;
  1553. size_t no_align_size = 0;
  1554. auto real_size_list_size = block->RealSizeList().size();
  1555. for (const NodeTypeIndex &node_type_index : block->NodeTypeIndexList()) {
  1556. if (index < real_size_list_size) {
  1557. real_size = block->RealSizeList()[index];
  1558. no_align_size = block->NoAlignSizeList()[index];
  1559. }
  1560. SetOffsetSize(node_type_index, block, real_size, no_align_size, child_block);
  1561. index++;
  1562. }
  1563. }
  1564. void BlockMemAssigner::SetOpMemOffset(bool is_zero_copy) {
  1565. for (MemoryBlock *memory_block : memory_blocks_) {
  1566. if (memory_block == nullptr || memory_block->deleted_block_) {
  1567. continue;
  1568. }
  1569. if ((is_zero_copy && !memory_block->is_zero_copy_) || (!is_zero_copy && memory_block->is_zero_copy_)) {
  1570. continue;
  1571. }
  1572. SetBlockOpMemOffset(memory_block, false);
  1573. for (MemoryBlock *child_block : memory_block->ChildBlockList()) {
  1574. SetBlockOpMemOffset(child_block, true);
  1575. }
  1576. }
  1577. if (!is_zero_copy) {
  1578. for (const NodeTypeIndex &node_type_index : zero_memory_list_) {
  1579. MemoryBlock block(0, 0);
  1580. SetOffsetSize(node_type_index, &block, 0, 0, false);
  1581. }
  1582. }
  1583. }
  1584. Status BlockMemAssigner::Assign() {
  1585. vector<int64_t> ranges;
  1586. if (GetMemoryRanges(ranges) != SUCCESS) {
  1587. GELOGE(FAILED, "GetMemoryRanges Fail!");
  1588. return FAILED;
  1589. }
  1590. GE_IF_BOOL_EXEC(ranges.empty(), return SUCCESS);
  1591. AssignMemoryWithReuse(ranges);
  1592. SetOpMemOffset(false);
  1593. return SUCCESS;
  1594. }
  1595. bool BlockMemAssigner::CheckIsZeroMemNodeType(const string &node_type) const {
  1596. return (node_type == VARIABLE) || (node_type == CONSTANT) || (node_type == MULTISHAPE) ||
  1597. (node_type == HCOMBROADCAST) || (node_type == CONSTANTOP) ||
  1598. (node_type == ASSIGNADD) || (node_type == ASSIGNSUB) || (node_type == ASSIGN) || (node_type == HVDWAIT) ||
  1599. (node_type == HVDCALLBACKBROADCAST);
  1600. }
  1601. bool BlockMemAssigner::GetWorkSpaceMemoryType(const NodePtr &node, size_t index, int64_t &memory_type) {
  1602. memory_type = RT_MEMORY_HBM;
  1603. vector<int64_t> workspace_memory_type;
  1604. auto op_desc = node->GetOpDesc();
  1605. bool has_workspace_mem_type_attr =
  1606. ge::AttrUtils::GetListInt(op_desc, TVM_ATTR_NAME_WORKSPACE_TYPE, workspace_memory_type);
  1607. if (has_workspace_mem_type_attr && (workspace_memory_type.size() <= index)) {
  1608. GELOGE(INTERNAL_ERROR, "node[%s], workspace_memory size error![index:%zu, workspace:%zu]",
  1609. node->GetName().c_str(), index, workspace_memory_type.size());
  1610. return false;
  1611. }
  1612. memory_type = has_workspace_mem_type_attr ? workspace_memory_type[index] : RT_MEMORY_HBM;
  1613. return true;
  1614. }
  1615. } // 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