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.h 15 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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466
  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. #ifndef GE_GRAPH_BUILD_MEMORY_BLOCK_MEM_ASSIGNER_H_
  17. #define GE_GRAPH_BUILD_MEMORY_BLOCK_MEM_ASSIGNER_H_
  18. #include <map>
  19. #include <string>
  20. #include <unordered_map>
  21. #include <unordered_set>
  22. #include <utility>
  23. #include <vector>
  24. #include <list>
  25. #include "common/ge_inner_error_codes.h"
  26. #include "common/types.h"
  27. #include "common/util.h"
  28. #include "graph/build/memory/mem_assigner.h"
  29. #include "graph/compute_graph.h"
  30. #include "graph/utils/graph_utils.h"
  31. namespace ge {
  32. const size_t kMaxLifeTime = 0xffffffff;
  33. using DependStreamLife = std::map<int64_t, std::map<int64_t, size_t>>;
  34. enum OpMemoryType { kOutput, kWorkspace };
  35. struct NodeTypeIndex {
  36. NodeTypeIndex(ge::NodePtr node, OpMemoryType mem_type, uint32_t index, bool ref_input = false)
  37. : node(std::move(node)), mem_type(mem_type), index(index), ref_input(ref_input) {}
  38. ge::NodePtr node = nullptr;
  39. OpMemoryType mem_type = kOutput;
  40. uint32_t index = 0;
  41. size_t life_time_end = kMaxLifeTime;
  42. bool ref_input = false;
  43. const string GetMemType() const {
  44. if (mem_type == kOutput) {
  45. return "output";
  46. } else if (mem_type == kWorkspace) {
  47. return "workspace";
  48. }
  49. return "unknown";
  50. }
  51. };
  52. class MemoryBlock {
  53. public:
  54. explicit MemoryBlock(size_t block_size, int64_t stream_id = 0, bool reuse_mem = true,
  55. int64_t memory_type = RT_MEMORY_HBM)
  56. : ref_count_(0),
  57. stream_id_(stream_id),
  58. deleted_block_(false),
  59. reuse_mem_(reuse_mem),
  60. same_stream_(true),
  61. input_index_(0),
  62. continuous_block_(false),
  63. first_continuous_block_(false),
  64. last_continuous_block_(false),
  65. is_zero_copy_(false),
  66. memory_type_(memory_type),
  67. block_size_(block_size),
  68. head_offset_(0),
  69. tail_offset_(0),
  70. child_offset_(0) {}
  71. MemoryBlock(const MemoryBlock &) = delete;
  72. MemoryBlock &operator=(const MemoryBlock &) = delete;
  73. ~MemoryBlock() {
  74. node_type_index_list_.clear();
  75. symbol_list_.clear();
  76. }
  77. void Init(size_t real_size, OpMemoryType type, const ge::NodePtr &node, uint32_t out_index, size_t no_align_size,
  78. int64_t stream_id) {
  79. real_size_list_.emplace_back(real_size);
  80. no_align_size_list_.emplace_back(no_align_size);
  81. node_type_index_list_.emplace_back(node, type, out_index, false);
  82. if (stream_id != stream_id_) {
  83. same_stream_ = false;
  84. }
  85. }
  86. size_t Size() const { return block_size_; }
  87. size_t AlignSize() const;
  88. void SetHeadOffset(size_t offset);
  89. void SetTailOffset(size_t offset);
  90. size_t HeadOffset() const { return head_offset_; }
  91. size_t TailOffset() const { return tail_offset_; }
  92. void AddNodeTypeIndex(const NodeTypeIndex &node_type_index, size_t real_size, size_t no_align_size) {
  93. node_type_index_list_.emplace_back(node_type_index);
  94. real_size_list_.emplace_back(real_size);
  95. no_align_size_list_.emplace_back(no_align_size);
  96. if ((node_type_index.node != nullptr) && (node_type_index.node->GetOpDesc() != nullptr)) {
  97. auto stream_id = node_type_index.node->GetOpDesc()->GetStreamId();
  98. if (stream_id != stream_id_) {
  99. same_stream_ = false;
  100. }
  101. }
  102. }
  103. void AddSymbol(const std::string &symbol) {
  104. symbol_list_.emplace_back(symbol);
  105. }
  106. const std::vector<NodeTypeIndex> &NodeTypeIndexList() const { return node_type_index_list_; }
  107. const std::vector<std::string> &SymbolList() const { return symbol_list_; }
  108. const std::vector<size_t> &RealSizeList() const { return real_size_list_; }
  109. const std::vector<MemoryBlock *> &ChildBlockList() const { return child_blocks_; }
  110. const std::vector<size_t> &NoAlignSizeList() const { return no_align_size_list_; }
  111. void Resize();
  112. std::string String();
  113. bool IsSameBatchLabel();
  114. void AddContinuousLifeReuseBlock(MemoryBlock *block, DependStreamLife &total_node_depend_stream_life);
  115. void AddLifeReuseBlock(MemoryBlock *block, DependStreamLife &node_depend_stream_life);
  116. void SetLifeTimeEnd(size_t time);
  117. size_t GetLifeBegin();
  118. size_t GetLifeEnd();
  119. void AddDependLifeBegin(DependStreamLife &node_depend_stream_life);
  120. size_t GetDependLifeBegin(int64_t stream_id, DependStreamLife &node_depend_stream_life);
  121. int ref_count_;
  122. int64_t stream_id_;
  123. bool deleted_block_;
  124. bool reuse_mem_;
  125. bool same_stream_;
  126. uint32_t input_index_;
  127. bool continuous_block_;
  128. bool first_continuous_block_;
  129. bool last_continuous_block_;
  130. bool is_zero_copy_;
  131. std::map<int64_t, size_t> depend_stream_life_;
  132. int64_t memory_type_;
  133. std::string batch_label_;
  134. private:
  135. size_t block_size_;
  136. std::vector<size_t> real_size_list_;
  137. std::vector<size_t> no_align_size_list_;
  138. size_t head_offset_;
  139. size_t tail_offset_;
  140. size_t child_offset_;
  141. std::vector<NodeTypeIndex> node_type_index_list_;
  142. std::vector<std::string> symbol_list_;
  143. std::vector<MemoryBlock *> child_blocks_;
  144. };
  145. class BlockMemAssigner : public MemAssigner {
  146. public:
  147. BlockMemAssigner(ComputeGraphPtr compute_graph, const std::map<std::string, std::string> &anchor_to_symbol,
  148. const std::map<std::string, std::list<NodeIndexIO>> &symbol_to_anchors);
  149. BlockMemAssigner(const BlockMemAssigner &) = delete;
  150. BlockMemAssigner &operator=(const BlockMemAssigner &) = delete;
  151. ~BlockMemAssigner() override;
  152. Status Assign() override;
  153. size_t GetMemOffset() const { return mem_offset_; }
  154. size_t GetP2PMemOffset() const { return p2p_mem_offset_; }
  155. int64_t GetAtomicAddrCleanId() const { return atomic_addr_clean_id_; }
  156. std::vector<MemoryBlock *> GetMemoryBlocks() const { return memory_blocks_; }
  157. ///
  158. /// @ingroup domi
  159. /// @brief memory size fixed for reuse. get memory range
  160. /// @param [out] ranges return memory range
  161. /// @return Status result
  162. ///
  163. virtual Status GetMemoryRanges(std::vector<int64_t> &ranges) = 0;
  164. ///
  165. /// @ingroup domi
  166. /// @brief traverse all nodes' outputs and needed workspace mem, apply memory, consider reuse memory
  167. /// @param [in] ranges memory range provided
  168. /// @author
  169. ///
  170. void AssignMemoryWithReuse(std::vector<int64_t> &ranges);
  171. void SetOpMemOffset(bool is_zero_copy);
  172. protected:
  173. ///
  174. /// @ingroup domi
  175. /// @brief traverse all memory size, resize, and calculate offset
  176. /// @param [in&out] memory_blocks memory size, resize and calculate memory address after offset
  177. ///
  178. void ResizeMemoryBlocks();
  179. void GetOutAndWorkSpaceMem(std::vector<int64_t> &all_memory_size);
  180. void GetNodeWorkSpaceSize(const ge::NodePtr &node, std::vector<int64_t> &workspace_memory, int64_t &total_size);
  181. ///
  182. /// @ingroup GE
  183. /// @brief Determine whether it is the type of zero memory node.
  184. /// @param [in] node type.
  185. /// @return bool true: is zero memory node; false: is not zero memory node
  186. /// @author
  187. ///
  188. bool CheckIsZeroMemNodeType(const std::string &node_type) const;
  189. ///
  190. /// @ingroup GE
  191. /// @brief Check pre_reuse flag & post_reuse glag for each symbol
  192. /// @return void
  193. ///
  194. void InitReuseFlag();
  195. ///
  196. /// @ingroup GE
  197. /// @brief get pre_reuse flag
  198. /// @param [in] node
  199. /// @param [in] out_index
  200. /// @return bool
  201. ///
  202. bool IsPreReuse(const NodePtr &node, uint32_t out_index) const;
  203. ///
  204. /// @ingroup GE
  205. /// @brief get post_reuse flag
  206. /// @param [in] mem_block
  207. /// @return bool
  208. ///
  209. bool IsPostReuse(const MemoryBlock *mem_block) const;
  210. ///
  211. /// @ingroup GE
  212. /// @brief check if symbol of cur node_index_io has block
  213. /// @param [in] node_index_io
  214. /// @param [out] symbol
  215. /// @return bool
  216. ///
  217. bool IsSymbolExist(const NodeIndexIO &node_index_io, std::string &symbol);
  218. ///
  219. /// @ingroup GE
  220. /// @brief Print symbol
  221. /// @return void
  222. ///
  223. void PrintSymbolMap();
  224. ///
  225. /// @ingroup GE
  226. /// @brief Get the memory type corresponding to the current symbol.
  227. /// @param [in] node_index_io_list
  228. /// @param [out] memory_type
  229. /// @return void
  230. ///
  231. void GetSymbolMemType(std::list<NodeIndexIO> node_index_io_list, int64_t &memory_type);
  232. ///
  233. /// @ingroup GE
  234. /// @brief Update input tensor or output tensor of op to new memory type attr.
  235. /// @param [in] node_index_io_list
  236. /// @param [in] memory_type
  237. /// @return void
  238. ///
  239. void UpdateOpTensorMemType(std::list<NodeIndexIO> node_index_io_list, int64_t memory_type);
  240. size_t mem_offset_;
  241. size_t p2p_mem_offset_;
  242. ge::ComputeGraphPtr compute_graph_;
  243. std::vector<MemoryBlock *> memory_blocks_;
  244. std::vector<MemoryBlock *> blocks_store_;
  245. std::vector<NodeTypeIndex> zero_memory_list_;
  246. // ref mapping
  247. const std::map<std::string, std::list<NodeIndexIO>> &symbol_to_anchors_;
  248. const std::map<std::string, std::string> &anchor_to_symbol_;
  249. std::map<std::string, bool> pre_reuse_flag_;
  250. std::map<std::string, bool> post_reuse_flag_;
  251. std::map<std::string, size_t> symbol_size_;
  252. std::map<std::string, int64_t> symbol_to_mem_type_;
  253. private:
  254. ///
  255. /// @ingroup GE
  256. /// @brief Traversing the compute_graph_ to apply for output memory while considering reuse
  257. /// @param [in] n: node in compute_graph_
  258. /// @param [in] index: output node index
  259. /// @param [in] ranges: available memory specifications
  260. /// @param [in] is_op_reuse_mem: Whether the op reuses the memory, true: reuse; false: not reuse
  261. /// @param [in] continuous: Whether the op uses continuous memory
  262. /// @return MemoryBlock*
  263. /// @author
  264. ///
  265. MemoryBlock *ApplyOutMemory(const ge::NodePtr &n, uint32_t index, const std::vector<int64_t> &ranges,
  266. const bool is_op_reuse_mem, const bool continuous);
  267. Status AssignOutputMemoryWithReuse(const NodePtr &node, vector<int64_t> &ranges);
  268. ///
  269. /// @ingroup GE
  270. /// @brief Traversing the compute_graph_ to apply for memory while considering reuse
  271. /// @param [in] block_size applied memory block size
  272. /// @param [in] real_size actual memory size required
  273. /// @param [in] type output or workspace
  274. /// @param [in] n node in compute_graph_
  275. /// @param [in] out_index output node index
  276. /// @param [in] workspace_reuse_flag reuse flag for workspace
  277. /// @param [in] is_op_reuse_mem whether the op reuses memory
  278. /// @param [in] continuous whether the memory of op is continuous
  279. /// @param [in] memory_type device memory type
  280. /// @return MemoryBlock*
  281. /// @author
  282. ///
  283. MemoryBlock *ApplyMemory(size_t block_size, size_t real_size, size_t no_align_size, OpMemoryType mem_type,
  284. const ge::NodePtr &n, uint32_t out_index, const std::vector<bool> &workspace_reuse_flag,
  285. const bool is_op_reuse_mem, const bool continuous, int64_t memory_type);
  286. ///
  287. /// @ingroup GE
  288. /// @brief check workspace_reuse_flag to judge if add workspace block wait reuse
  289. /// @param [in] workspace_reuse_flag mark out index if support resue
  290. /// @param [in] index out index
  291. /// @param [in] stream_id which stream op in
  292. /// @param [in] mem_block node workspace mem_block
  293. /// @param [in] memory_type workspace memory type
  294. /// @return void
  295. /// @author
  296. ///
  297. void CheckWorkspaceReuse(const vector<bool> &workspace_reuse_flag, uint32_t index, int64_t stream_id,
  298. MemoryBlock *mem_block, int64_t memory_type);
  299. ///
  300. /// @ingroup GE
  301. /// @brief Release memory block to reusable list
  302. /// @param [in] to_release memory block to be released
  303. /// @param [in] reusable_memory reusable list
  304. /// @return void
  305. /// @author
  306. ///
  307. void ReleaseMemory(MemoryBlock *to_release, vector<MemoryBlock *> &reusable_memory, bool same_stream = true);
  308. ///
  309. /// @ingroup GE
  310. /// @brief Release memory blocks to reusable list
  311. /// @param [in] to_releases memory blocks to be released
  312. /// @param [in] reusable_memory reusable list
  313. /// @return void
  314. /// @author
  315. ///
  316. void ReleaseMemorys(const vector<MemoryBlock *> &to_releases, vector<MemoryBlock *> &reusable_memory);
  317. ///
  318. /// @ingroup GE
  319. /// @brief Release memory block to reusable list
  320. /// @param [in] n node in compute_graph_
  321. /// @param [in] node_out_blocks output memory blocks for ops
  322. /// @param [in] reusable_memory reusable list
  323. /// @return void
  324. /// @author
  325. ///
  326. void ReleaseInputNodeOutMemory(const std::unordered_map<string, vector<MemoryBlock *>> &node_out_blocks,
  327. vector<MemoryBlock *> &reusable_memory, ge::NodePtr &n);
  328. ///
  329. /// @ingroup GE
  330. /// @brief Resize memory blocks for each batchs
  331. /// @return merge or not
  332. /// @author
  333. ///
  334. void ResizeDynamicBatchBlocks();
  335. void AssignContinuousBlocks();
  336. bool IsZeroCopyBlock(const NodePtr &node, bool continuous);
  337. bool IsOutNodeSetContinuousInput(const NodePtr &n, uint32_t out_index, std::string &peer_name,
  338. uint32_t &peer_input_index, bool &no_need_assign_memory, bool &reset_zero_copy_flag);
  339. ///
  340. /// @ingroup GE
  341. /// @|+++++++++block1++++++++| |+++++++++block1++++++++|
  342. /// @|+++++++++block1++++++++||++block2++| |+++++++++block1++++++++||++block2++|
  343. /// @ |++block2++||++block3++| ==> |++block3++| |++block2++|
  344. /// @ |++block3++| |++block3++|
  345. /// @return void
  346. /// @author
  347. ///
  348. void ReuseBlocksByLifeTime(size_t range_size);
  349. bool IsContinuousOutput(const NodePtr &n);
  350. bool GetWorkSpaceMemoryType(const NodePtr &node, size_t index, int64_t &memory_type);
  351. MemoryBlock *ApplyContinuousMemory(const NodePtr &n, const vector<int64_t> &ranges, const bool is_op_reuse_mem);
  352. std::unordered_map<int64_t, std::unordered_map<int64_t, std::vector<MemoryBlock *>>> reusable_blocks_;
  353. std::map<std::string, uint64_t> reusable_block_counts_;
  354. std::unordered_map<int64_t, std::unordered_map<int64_t, std::vector<MemoryBlock *>>> stream_workspace_blocks_;
  355. std::unordered_map<std::string, std::vector<MemoryBlock *>> node_out_blocks_;
  356. std::unordered_map<std::string, MemoryBlock *> symbol_blocks_;
  357. std::unordered_map<std::string, std::unordered_map<uint32_t, MemoryBlock *>> node_continuous_input_blocks_;
  358. std::unordered_map<std::string, uint32_t> node_continuous_input_counts_;
  359. // reuse memory
  360. vector<string> op_no_reuse_mem_vec_;
  361. bool op_reuse_env_valid_ = false;
  362. std::string ge_disable_reuse_mem_env_ = "0";
  363. bool is_op_reuse_mem_ = true;
  364. size_t life_time_;
  365. int64_t atomic_addr_clean_id_ = 0;
  366. size_t theory_min_memory_size_ = 0;
  367. size_t theory_memory_size_ = 0;
  368. std::string max_batch_label_;
  369. ///
  370. /// @ [stream1][nodeid]
  371. /// @[nodeid] [stream2][nodeid]
  372. /// @ [stream2][nodeid]
  373. ///
  374. DependStreamLife total_node_depend_stream_life_;
  375. };
  376. } // namespace ge
  377. #endif // GE_GRAPH_BUILD_MEMORY_BLOCK_MEM_ASSIGNER_H_

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