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.

rt_model.h 13 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498
  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. #ifndef CCE_RUNTIME_RT_MODEL_H
  17. #define CCE_RUNTIME_RT_MODEL_H
  18. #include "base.h"
  19. #if defined(__cplusplus)
  20. extern "C" {
  21. #endif
  22. typedef enum tagModelTaskType {
  23. RT_MODEL_TASK_KERNEL = 0,
  24. RT_MODEL_TASK_EVENT_RECORD,
  25. RT_MODEL_TASK_EVENT_WAIT,
  26. RT_MODEL_TASK_FUSION_START,
  27. RT_MODEL_TASK_FUSION_END,
  28. RT_MODEL_TASK_KERNEL_EX,
  29. RT_MODEL_TASK_HCCL,
  30. RT_MODEL_TASK_STREAM_SWITCH,
  31. RT_MODEL_TASK_STREAM_ACTIVE,
  32. RT_MODEL_TASK_LABEL_SET,
  33. RT_MODEL_TASK_LABEL_SWITCH,
  34. RT_MODEL_TASK_LABEL_GOTO,
  35. RT_MODEL_TASK_PROFILER_TRACE,
  36. RT_MODEL_TASK_MEMCPY_ASYNC,
  37. RT_MODEL_TASK_NOTIFY_RECORD,
  38. RT_MODEL_TASK_NOTIFY_WAIT,
  39. RT_MODEL_TASK_REDUCE_ASYNC,
  40. RT_MODEL_TASK_RDMA_SEND,
  41. RT_MODEL_TASK_EVENT_RESET,
  42. RT_MODEL_TASK_MODEL_END_GRAPH,
  43. RT_MODEL_TASK_STREAM_SWITCH_N,
  44. RT_MODEL_TASK_RDMA_DB_SEND,
  45. RT_MODEL_TASK_MEMCPY_ADDR_ASYNC,
  46. RT_MODEL_TASK_STREAM_LABEL_SWITCH_BY_INDEX,
  47. RT_MODEL_TASK_STREAM_LABEL_GOTO,
  48. RT_MODEL_TASK_MODEL_EXIT,
  49. RT_MODEL_TASK_ALL_KERNEL,
  50. RT_MODEL_TASK_PROFILER_TRACE_EX,
  51. RT_MODEL_TASK_FFTS_TASK,
  52. RT_MODEL_TASK_FFTS_PLUS_TASK,
  53. } rtModelTaskType_t;
  54. typedef enum tagModelStreamType {
  55. RT_MODEL_HEAD_STREAM = 0,
  56. RT_MODEL_WAIT_ACTIVE_STREAM = 1
  57. } rtModelStreamType_t;
  58. typedef enum tagModelQueueFlag {
  59. RT_MODEL_INPUT_QUEUE = 0,
  60. RT_MODEL_OUTPUT_QUEUE = 1
  61. } rtModelQueueFlag_t;
  62. #define EXECUTOR_NONE (0x0U)
  63. #define EXECUTOR_TS (0x01U)
  64. #define EXECUTOR_AICPU (0x02U)
  65. /*
  66. * @ingroup rt_model
  67. * @brief debug flag for kernel exception dump
  68. */
  69. #define RT_DEBUG_FLAG_AICORE_OVERFLOW (0x1U << 0U)
  70. #define RT_DEBUG_FLAG_ATOMIC_ADD_OVERFLOW (0x1U << 1U)
  71. /**
  72. * @ingroup
  73. * @brief the type defination of aicpu model task command
  74. */
  75. typedef enum tagTsAicpuModelCmd {
  76. TS_AICPU_MODEL_LOAD = 1,
  77. TS_AICPU_MODEL_EXECUTE,
  78. TS_AICPU_MODEL_DESTROY,
  79. TS_AICPU_MODEL_ABORT,
  80. TS_AICPU_MODEL_RESERVED,
  81. } tsAicpuModelCmd;
  82. typedef struct tagAicpuTaskInfo {
  83. uint32_t taskID;
  84. uint32_t streamID;
  85. uint32_t kernelType;
  86. uint64_t kernelName;
  87. uint64_t kernelSo;
  88. uint64_t paraBase;
  89. uint32_t taskFlag;
  90. } rtAicpuTaskInfo_t;
  91. typedef struct tagModelStreamInfo {
  92. uint32_t streamID;
  93. uint32_t streamFlag;
  94. } rtModelStreamInfo_t;
  95. typedef struct tagModelQueueInfo {
  96. uint32_t queueID;
  97. uint32_t flag;
  98. } rtModelQueueInfo_t;
  99. typedef struct tagAicpuModelInfo {
  100. uint32_t moduleID;
  101. uint32_t tsId;
  102. uint16_t streamInfoNum;
  103. uint16_t aicpuTaskNum;
  104. uint64_t streamInfoPtr;
  105. uint64_t aicpuTaskPtr;
  106. uint16_t queueSize;
  107. uint64_t queueInfoPtr;
  108. } rtAicpuModelInfo_t;
  109. typedef struct tagKernelTaskInfo {
  110. uint16_t blockDim;
  111. uint16_t argsCount;
  112. uint16_t argsSize;
  113. uint16_t reserved;
  114. char_t *stubFunc;
  115. uint8_t *smDesc;
  116. uint8_t *args;
  117. uint16_t *argsOffset;
  118. } rtKernelTaskInfo_t;
  119. typedef struct tagAllKernelTaskInfo {
  120. uint16_t blockDim;
  121. uint16_t argsCount;
  122. uint16_t argsSize;
  123. uint16_t reserved;
  124. void *devfunc;
  125. void *handle;
  126. uint8_t *smDesc;
  127. uint8_t *args;
  128. uint16_t *argsOffset;
  129. } rtAllKernelTaskInfo_t;
  130. typedef struct tagKernelTaskInfoEx {
  131. uint32_t flags;
  132. uint32_t argsSize;
  133. void *args;
  134. uint32_t reserved[6];
  135. } rtKernelTaskInfoEx_t;
  136. typedef struct tagEventTaskInfo {
  137. uint32_t eventID;
  138. uint32_t reserved[9];
  139. } rtEventTaskInfo_t;
  140. typedef struct tagStreamSwitchTaskInfo {
  141. int64_t value;
  142. uint64_t pValuePtr;
  143. uint32_t trueStreamID;
  144. uint32_t dataType;
  145. uint32_t reserved[4];
  146. } rtStreamSwitchTaskInfo_t;
  147. typedef struct tagStreamSwitchNTaskInfo {
  148. uint64_t pValuePtr;
  149. uint64_t pTrueStreamPtr;
  150. uint32_t size;
  151. uint32_t elementSize;
  152. uint32_t dataType;
  153. uint32_t reserved[3];
  154. } rtStreamSwitchNTaskInfo_t;
  155. typedef struct tagStreamActiveTaskInfo {
  156. uint32_t activeStreamID;
  157. uint32_t reserved[9];
  158. } rtStreamActiveTaskInfo_t;
  159. typedef struct tagSetTaskInfo {
  160. uint16_t labelId;
  161. uint32_t reserved[9];
  162. } rtLabelSetTaskInfo_t;
  163. typedef struct tagSwitchTaskInfo {
  164. uint32_t value;
  165. uint32_t reserved[9];
  166. } rtLabelSwitchTaskInfo_t;
  167. typedef struct tagLabelGotoTaskInfo {
  168. uint16_t labelId;
  169. uint32_t reserved[9];
  170. } rtLabelGotoTaskInfo_t;
  171. typedef struct tagProfilerTraceTaskInfo {
  172. uint64_t profilerTraceId;
  173. uint32_t notify : 8;
  174. uint32_t reserved_ : 24;
  175. uint32_t flags;
  176. uint32_t reserved[6];
  177. } rtProfilerTrace_t;
  178. typedef struct tagProfilerTraceExTaskInfo {
  179. uint64_t profilerTraceId;
  180. uint64_t modelId;
  181. uint16_t tagId;
  182. uint8_t reserved[22];
  183. } rtProfilerTraceEx_t;
  184. typedef struct tagrtMemcpyAsyncTaskInfo {
  185. void *dst;
  186. uint64_t destMax;
  187. void *src;
  188. uint64_t count;
  189. uint32_t kind;
  190. uint32_t reserved;
  191. } rtMemcpyAsyncTaskInfo_t;
  192. typedef struct tagrtNotifyTaskInfo {
  193. uint32_t notifyID;
  194. uint32_t reserved[9];
  195. } rtNotifyTaskInfo_t;
  196. typedef struct tagrtReduceAsyncTaskInfo {
  197. void *dst;
  198. uint64_t destMax;
  199. void *src;
  200. uint64_t count;
  201. uint32_t kind;
  202. uint32_t type;
  203. } rtReduceAsyncTaskInfo_t;
  204. typedef struct tagrtRdmaSendTaskInfo {
  205. uint32_t index;
  206. uint32_t wqe_index;
  207. uint32_t reserved[8];
  208. } rtRdmaSendTaskInfo_t;
  209. typedef struct tagrtRdmaDbSendTaskInfo {
  210. uint64_t dbInfo;
  211. uint32_t dbIndex;
  212. uint32_t reserved[7]; // offset 7
  213. } rtRdmaDbSendTaskInfo_t;
  214. typedef struct tagrtModelEndGraphTaskInfo {
  215. uint32_t modelId;
  216. uint32_t executorFlag;
  217. uint32_t reserved[8];
  218. } rtModelEndGraphTaskInfo_t;
  219. typedef struct tagrtModelExitInfo {
  220. uint32_t modelId;
  221. uint32_t streamId;
  222. uint32_t reserved[8];
  223. } rtModelExitTaskInfo_t;
  224. typedef struct tagrtStreamLabelSwitchByIndexTask_t {
  225. uint64_t indexPtr;
  226. uint64_t labelInfoPtr;
  227. uint32_t max;
  228. uint8_t reserved[20];
  229. } rtStreamLabelSwitchByIndexTask_t;
  230. typedef struct tagrtStreamLabelGotoTask_t {
  231. uint16_t labelId;
  232. uint16_t modelId;
  233. uint8_t reserved[36];
  234. } rtStreamLabelGotoTask_t;
  235. typedef struct tagTaskInfo {
  236. uint32_t type;
  237. uint32_t streamID;
  238. union {
  239. rtKernelTaskInfoEx_t kernelTaskEx;
  240. rtKernelTaskInfo_t kernelTask;
  241. rtAllKernelTaskInfo_t allKernelTask;
  242. rtEventTaskInfo_t eventTask;
  243. rtStreamSwitchTaskInfo_t streamSwitchTask;
  244. rtStreamActiveTaskInfo_t streamActiveTask;
  245. rtLabelSetTaskInfo_t labelSetTask;
  246. rtLabelSwitchTaskInfo_t labelSwitchTask;
  247. rtLabelGotoTaskInfo_t labelGotoTask;
  248. rtProfilerTrace_t profilertraceTask;
  249. rtProfilerTraceEx_t profilertraceExTask;
  250. rtMemcpyAsyncTaskInfo_t memcpyAsyncTask;
  251. rtNotifyTaskInfo_t notifyTask;
  252. rtReduceAsyncTaskInfo_t reduceAsyncTask;
  253. rtRdmaSendTaskInfo_t rdmaSendTask;
  254. rtRdmaDbSendTaskInfo_t rdmaDbSendTask;
  255. rtModelEndGraphTaskInfo_t modelEndGraphTask;
  256. rtModelExitTaskInfo_t modelExitTask;
  257. rtStreamSwitchNTaskInfo_t streamSwitchNTask;
  258. rtStreamLabelSwitchByIndexTask_t streamLabelSwitchIndexTask;
  259. rtStreamLabelGotoTask_t streamLabelGotoTask;
  260. uint32_t reserved[10];
  261. } u;
  262. } rtTaskInfo_t;
  263. typedef struct tagNodeInfo_t {
  264. uint32_t nodeIdx;
  265. uint32_t reserved[1];
  266. } rtNodeInfo;
  267. typedef struct tagHwtsInfo_t {
  268. uint16_t taskId;
  269. uint16_t sqExeHead;
  270. uint16_t streamExeHead;
  271. uint16_t reserved[2];
  272. } rtHwtsInfo;
  273. typedef struct tagLabelDevInfo_t {
  274. uint16_t modelId;
  275. uint16_t streamId;
  276. uint16_t labelId;
  277. union {
  278. rtNodeInfo nodeInfo;
  279. rtHwtsInfo hwtsInfo;
  280. uint16_t reserved[5];
  281. }u;
  282. }rtLabelDevInfo;
  283. typedef rtError_t (*rtTaskGenCallback)(rtModel_t model, rtTaskInfo_t *taskInfo);
  284. /**
  285. * @ingroup rt_model
  286. * @brief set callback for generate model
  287. * @param [in] callBack callback function
  288. * @return RT_ERROR_NONE for ok
  289. * @return RT_ERROR_INVALID_VALUE for error input
  290. */
  291. RTS_API rtError_t rtSetTaskGenCallback(rtTaskGenCallback callback);
  292. /**
  293. * @ingroup rt_model
  294. * @brief create model instance
  295. * @param [out] model created model
  296. * @param [in] flag reserved
  297. * @return RT_ERROR_NONE for ok
  298. * @return RT_ERROR_INVALID_VALUE for error input
  299. */
  300. RTS_API rtError_t rtModelCreate(rtModel_t *model, uint32_t flag);
  301. /**
  302. * @ingroup rt_model
  303. * @brief destroy model instance
  304. * @param [in] model model to destroy
  305. * @return RT_ERROR_NONE for ok
  306. * @return RT_ERROR_INVALID_VALUE for error input
  307. */
  308. RTS_API rtError_t rtModelDestroy(rtModel_t model);
  309. /**
  310. * @ingroup rt_model
  311. * @brief bind model and stream instance
  312. * @param [in] model binded model
  313. * @param [in] stream binded stream
  314. * @param [in] flag reserved
  315. * @return RT_ERROR_NONE for ok
  316. * @return RT_ERROR_INVALID_VALUE for error input
  317. */
  318. RTS_API rtError_t rtModelBindStream(rtModel_t model, rtStream_t stream, uint32_t flag);
  319. /**
  320. * @ingroup rt_model
  321. * @brief unbind model and stream instance
  322. * @param [in] model unbinded model
  323. * @param [in] stream unbinded stream
  324. * @return RT_ERROR_NONE for ok
  325. * @return RT_ERROR_INVALID_VALUE for error input
  326. */
  327. RTS_API rtError_t rtModelUnbindStream(rtModel_t model, rtStream_t stream);
  328. /**
  329. * @ingroup rt_model
  330. * @brief tell runtime Model has been Loaded
  331. * @param [in] model model to execute
  332. * @return RT_ERROR_NONE for ok
  333. */
  334. RTS_API rtError_t rtModelLoadComplete(rtModel_t model);
  335. /**
  336. * @ingroup rt_model
  337. * @brief execute model instance
  338. * @param [in] model model to execute
  339. * @return RT_ERROR_NONE for ok
  340. * @return RT_ERROR_INVALID_VALUE for error input
  341. */
  342. RTS_API rtError_t rtModelExecute(rtModel_t model, rtStream_t stream, uint32_t flag);
  343. /**
  344. * @ingroup rt_model
  345. * @brief get model the last persist task id
  346. * @param [in] model model to execute
  347. * @param [out] taskId last task id of the model
  348. * @param [out] streamId last steam id of the model
  349. * @return RT_ERROR_NONE for ok
  350. * @return RT_ERROR_INVALID_VALUE for error input
  351. */
  352. RTS_API rtError_t rtModelGetTaskId(rtModel_t model, uint32_t *taskId, uint32_t *streamId);
  353. /**
  354. * @ingroup rt_model
  355. * @brief add a end graph task to stream
  356. * @param [in] model model to execute
  357. * @param [in] end graph stream
  358. * @return RT_ERROR_NONE for ok
  359. * @return RT_ERROR_INVALID_VALUE for error input
  360. */
  361. RTS_API rtError_t rtEndGraph(rtModel_t model, rtStream_t stream);
  362. /**
  363. * @ingroup rt_model
  364. * @brief add a end graph task with flag to stream
  365. * @param [in] model model to execute
  366. * @param [in] end graph stream
  367. * @param [in] flags AICPU datadump
  368. * @return RT_ERROR_NONE for ok
  369. * @return RT_ERROR_INVALID_VALUE for error input
  370. */
  371. RTS_API rtError_t rtEndGraphEx(rtModel_t model, rtStream_t stream, uint32_t flags);
  372. /**
  373. * @ingroup rt_model
  374. * @brief add a end graph task to stream
  375. * @param [in] model model to execute
  376. * @param [in] flags EXECUTOR_TS | EXECUTOR_AICPU
  377. * @return RT_ERROR_NONE for ok
  378. * @return RT_ERROR_INVALID_VALUE for error input
  379. */
  380. RTS_API rtError_t rtModelExecutorSet(rtModel_t model, uint8_t flags);
  381. /**
  382. * @ingroup rt_model
  383. * @brief abort model
  384. * @param [in] model model to abort
  385. * @return RT_ERROR_NONE for ok
  386. * @return RT_ERROR_INVALID_VALUE for error input
  387. */
  388. RTS_API rtError_t rtModelAbort(rtModel_t model);
  389. /**
  390. * @ingroup rt_model
  391. * @brief end graph task to model default stream
  392. * @param [in] model model to execute
  393. * @param [in] end graph stream
  394. * @return RT_ERROR_NONE for ok
  395. * @return RT_ERROR_INVALID_VALUE for error input
  396. */
  397. RTS_API rtError_t rtModelExit(rtModel_t model, rtStream_t stream);
  398. /**
  399. * @ingroup rt_model
  400. * @brief bind queue
  401. * @param [in] model model to bind
  402. * @param [in] queueId queueId to bind
  403. * @param [in] flag
  404. * @return RT_ERROR_NONE for ok
  405. * @return RT_ERROR_INVALID_VALUE for error input
  406. */
  407. RTS_API rtError_t rtModelBindQueue(rtModel_t model, uint32_t queueId, rtModelQueueFlag_t flag);
  408. /**
  409. * @ingroup rt_model
  410. * @brief get model id
  411. * @param [in] model
  412. * @param [out] modelId model id
  413. * @return RT_ERROR_NONE for ok
  414. * @return RT_ERROR_INVALID_VALUE for error input
  415. */
  416. RTS_API rtError_t rtModelGetId(rtModel_t model, uint32_t *modelId);
  417. /*
  418. * @ingroup rt_model
  419. * @brief enable debug for dump overflow exception
  420. * @param [in] addr: ddr address of kernel exception dumpped
  421. * @param [in] model: model handle
  422. * @param [in] flag: debug flag
  423. * @return RT_ERROR_NONE for ok
  424. * @return RT_ERROR_INVALID_VALUE for error input
  425. */
  426. RTS_API rtError_t rtDebugRegister(rtModel_t model, uint32_t flag, const void *addr,
  427. uint32_t *streamId, uint32_t *taskId);
  428. /*
  429. * @ingroup rt_model
  430. * @brief disable debug for dump overflow exception
  431. * @param [in] model: model handle
  432. * @return RT_ERROR_NONE for ok
  433. * @return RT_ERROR_INVALID_VALUE for error input
  434. */
  435. RTS_API rtError_t rtDebugUnRegister(rtModel_t model);
  436. #if defined(__cplusplus)
  437. }
  438. #endif
  439. #endif // CCE_RUNTIME_RT_MODEL_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