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
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
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488
  1. /*
  2. * Copyright (c) Huawei Technologies Co., Ltd. 2020-2021. All rights reserved.
  3. * Description: rt_model.h
  4. * Create: 2020-01-01
  5. */
  6. #ifndef CCE_RUNTIME_RT_MODEL_H
  7. #define CCE_RUNTIME_RT_MODEL_H
  8. #include "base.h"
  9. #if defined(__cplusplus)
  10. extern "C" {
  11. #endif
  12. typedef enum tagModelTaskType {
  13. RT_MODEL_TASK_KERNEL = 0,
  14. RT_MODEL_TASK_EVENT_RECORD,
  15. RT_MODEL_TASK_EVENT_WAIT,
  16. RT_MODEL_TASK_FUSION_START,
  17. RT_MODEL_TASK_FUSION_END,
  18. RT_MODEL_TASK_KERNEL_EX,
  19. RT_MODEL_TASK_HCCL,
  20. RT_MODEL_TASK_STREAM_SWITCH,
  21. RT_MODEL_TASK_STREAM_ACTIVE,
  22. RT_MODEL_TASK_LABEL_SET,
  23. RT_MODEL_TASK_LABEL_SWITCH,
  24. RT_MODEL_TASK_LABEL_GOTO,
  25. RT_MODEL_TASK_PROFILER_TRACE,
  26. RT_MODEL_TASK_MEMCPY_ASYNC,
  27. RT_MODEL_TASK_NOTIFY_RECORD,
  28. RT_MODEL_TASK_NOTIFY_WAIT,
  29. RT_MODEL_TASK_REDUCE_ASYNC,
  30. RT_MODEL_TASK_RDMA_SEND,
  31. RT_MODEL_TASK_EVENT_RESET,
  32. RT_MODEL_TASK_MODEL_END_GRAPH,
  33. RT_MODEL_TASK_STREAM_SWITCH_N,
  34. RT_MODEL_TASK_RDMA_DB_SEND,
  35. RT_MODEL_TASK_MEMCPY_ADDR_ASYNC,
  36. RT_MODEL_TASK_STREAM_LABEL_SWITCH_BY_INDEX,
  37. RT_MODEL_TASK_STREAM_LABEL_GOTO,
  38. RT_MODEL_TASK_MODEL_EXIT,
  39. RT_MODEL_TASK_ALL_KERNEL,
  40. RT_MODEL_TASK_PROFILER_TRACE_EX,
  41. RT_MODEL_TASK_FFTS_TASK,
  42. RT_MODEL_TASK_FFTS_PLUS_TASK,
  43. } rtModelTaskType_t;
  44. typedef enum tagModelStreamType {
  45. RT_MODEL_HEAD_STREAM = 0,
  46. RT_MODEL_WAIT_ACTIVE_STREAM = 1
  47. } rtModelStreamType_t;
  48. typedef enum tagModelQueueFlag {
  49. RT_MODEL_INPUT_QUEUE = 0,
  50. RT_MODEL_OUTPUT_QUEUE = 1
  51. } rtModelQueueFlag_t;
  52. #define EXECUTOR_NONE (0x0U)
  53. #define EXECUTOR_TS (0x01U)
  54. #define EXECUTOR_AICPU (0x02U)
  55. /*
  56. * @ingroup rt_model
  57. * @brief debug flag for kernel exception dump
  58. */
  59. #define RT_DEBUG_FLAG_AICORE_OVERFLOW (0x1U << 0U)
  60. #define RT_DEBUG_FLAG_ATOMIC_ADD_OVERFLOW (0x1U << 1U)
  61. /**
  62. * @ingroup
  63. * @brief the type defination of aicpu model task command
  64. */
  65. typedef enum tagTsAicpuModelCmd {
  66. TS_AICPU_MODEL_LOAD = 1,
  67. TS_AICPU_MODEL_EXECUTE,
  68. TS_AICPU_MODEL_DESTROY,
  69. TS_AICPU_MODEL_ABORT,
  70. TS_AICPU_MODEL_RESERVED,
  71. } tsAicpuModelCmd;
  72. typedef struct tagAicpuTaskInfo {
  73. uint32_t taskID;
  74. uint32_t streamID;
  75. uint32_t kernelType;
  76. uint64_t kernelName;
  77. uint64_t kernelSo;
  78. uint64_t paraBase;
  79. uint32_t taskFlag;
  80. } rtAicpuTaskInfo_t;
  81. typedef struct tagModelStreamInfo {
  82. uint32_t streamID;
  83. uint32_t streamFlag;
  84. } rtModelStreamInfo_t;
  85. typedef struct tagModelQueueInfo {
  86. uint32_t queueID;
  87. uint32_t flag;
  88. } rtModelQueueInfo_t;
  89. typedef struct tagAicpuModelInfo {
  90. uint32_t moduleID;
  91. uint32_t tsId;
  92. uint16_t streamInfoNum;
  93. uint16_t aicpuTaskNum;
  94. uint64_t streamInfoPtr;
  95. uint64_t aicpuTaskPtr;
  96. uint16_t queueSize;
  97. uint64_t queueInfoPtr;
  98. } rtAicpuModelInfo_t;
  99. typedef struct tagKernelTaskInfo {
  100. uint16_t blockDim;
  101. uint16_t argsCount;
  102. uint16_t argsSize;
  103. uint16_t reserved;
  104. char_t *stubFunc;
  105. uint8_t *smDesc;
  106. uint8_t *args;
  107. uint16_t *argsOffset;
  108. } rtKernelTaskInfo_t;
  109. typedef struct tagAllKernelTaskInfo {
  110. uint16_t blockDim;
  111. uint16_t argsCount;
  112. uint16_t argsSize;
  113. uint16_t reserved;
  114. void *devfunc;
  115. void *handle;
  116. uint8_t *smDesc;
  117. uint8_t *args;
  118. uint16_t *argsOffset;
  119. } rtAllKernelTaskInfo_t;
  120. typedef struct tagKernelTaskInfoEx {
  121. uint32_t flags;
  122. uint32_t argsSize;
  123. void *args;
  124. uint32_t reserved[6];
  125. } rtKernelTaskInfoEx_t;
  126. typedef struct tagEventTaskInfo {
  127. uint32_t eventID;
  128. uint32_t reserved[9];
  129. } rtEventTaskInfo_t;
  130. typedef struct tagStreamSwitchTaskInfo {
  131. int64_t value;
  132. uint64_t pValuePtr;
  133. uint32_t trueStreamID;
  134. uint32_t dataType;
  135. uint32_t reserved[4];
  136. } rtStreamSwitchTaskInfo_t;
  137. typedef struct tagStreamSwitchNTaskInfo {
  138. uint64_t pValuePtr;
  139. uint64_t pTrueStreamPtr;
  140. uint32_t size;
  141. uint32_t elementSize;
  142. uint32_t dataType;
  143. uint32_t reserved[3];
  144. } rtStreamSwitchNTaskInfo_t;
  145. typedef struct tagStreamActiveTaskInfo {
  146. uint32_t activeStreamID;
  147. uint32_t reserved[9];
  148. } rtStreamActiveTaskInfo_t;
  149. typedef struct tagSetTaskInfo {
  150. uint16_t labelId;
  151. uint32_t reserved[9];
  152. } rtLabelSetTaskInfo_t;
  153. typedef struct tagSwitchTaskInfo {
  154. uint32_t value;
  155. uint32_t reserved[9];
  156. } rtLabelSwitchTaskInfo_t;
  157. typedef struct tagLabelGotoTaskInfo {
  158. uint16_t labelId;
  159. uint32_t reserved[9];
  160. } rtLabelGotoTaskInfo_t;
  161. typedef struct tagProfilerTraceTaskInfo {
  162. uint64_t profilerTraceId;
  163. uint32_t notify : 8;
  164. uint32_t reserved_ : 24;
  165. uint32_t flags;
  166. uint32_t reserved[6];
  167. } rtProfilerTrace_t;
  168. typedef struct tagProfilerTraceExTaskInfo {
  169. uint64_t profilerTraceId;
  170. uint64_t modelId;
  171. uint16_t tagId;
  172. uint8_t reserved[22];
  173. } rtProfilerTraceEx_t;
  174. typedef struct tagrtMemcpyAsyncTaskInfo {
  175. void *dst;
  176. uint64_t destMax;
  177. void *src;
  178. uint64_t count;
  179. uint32_t kind;
  180. uint32_t reserved;
  181. } rtMemcpyAsyncTaskInfo_t;
  182. typedef struct tagrtNotifyTaskInfo {
  183. uint32_t notifyID;
  184. uint32_t reserved[9];
  185. } rtNotifyTaskInfo_t;
  186. typedef struct tagrtReduceAsyncTaskInfo {
  187. void *dst;
  188. uint64_t destMax;
  189. void *src;
  190. uint64_t count;
  191. uint32_t kind;
  192. uint32_t type;
  193. } rtReduceAsyncTaskInfo_t;
  194. typedef struct tagrtRdmaSendTaskInfo {
  195. uint32_t index;
  196. uint32_t wqe_index;
  197. uint32_t reserved[8];
  198. } rtRdmaSendTaskInfo_t;
  199. typedef struct tagrtRdmaDbSendTaskInfo {
  200. uint64_t dbInfo;
  201. uint32_t dbIndex;
  202. uint32_t reserved[7]; // offset 7
  203. } rtRdmaDbSendTaskInfo_t;
  204. typedef struct tagrtModelEndGraphTaskInfo {
  205. uint32_t modelId;
  206. uint32_t executorFlag;
  207. uint32_t reserved[8];
  208. } rtModelEndGraphTaskInfo_t;
  209. typedef struct tagrtModelExitInfo {
  210. uint32_t modelId;
  211. uint32_t streamId;
  212. uint32_t reserved[8];
  213. } rtModelExitTaskInfo_t;
  214. typedef struct tagrtStreamLabelSwitchByIndexTask_t {
  215. uint64_t indexPtr;
  216. uint64_t labelInfoPtr;
  217. uint32_t max;
  218. uint8_t reserved[20];
  219. } rtStreamLabelSwitchByIndexTask_t;
  220. typedef struct tagrtStreamLabelGotoTask_t {
  221. uint16_t labelId;
  222. uint16_t modelId;
  223. uint8_t reserved[36];
  224. } rtStreamLabelGotoTask_t;
  225. typedef struct tagTaskInfo {
  226. uint32_t type;
  227. uint32_t streamID;
  228. union {
  229. rtKernelTaskInfoEx_t kernelTaskEx;
  230. rtKernelTaskInfo_t kernelTask;
  231. rtAllKernelTaskInfo_t allKernelTask;
  232. rtEventTaskInfo_t eventTask;
  233. rtStreamSwitchTaskInfo_t streamSwitchTask;
  234. rtStreamActiveTaskInfo_t streamActiveTask;
  235. rtLabelSetTaskInfo_t labelSetTask;
  236. rtLabelSwitchTaskInfo_t labelSwitchTask;
  237. rtLabelGotoTaskInfo_t labelGotoTask;
  238. rtProfilerTrace_t profilertraceTask;
  239. rtProfilerTraceEx_t profilertraceExTask;
  240. rtMemcpyAsyncTaskInfo_t memcpyAsyncTask;
  241. rtNotifyTaskInfo_t notifyTask;
  242. rtReduceAsyncTaskInfo_t reduceAsyncTask;
  243. rtRdmaSendTaskInfo_t rdmaSendTask;
  244. rtRdmaDbSendTaskInfo_t rdmaDbSendTask;
  245. rtModelEndGraphTaskInfo_t modelEndGraphTask;
  246. rtModelExitTaskInfo_t modelExitTask;
  247. rtStreamSwitchNTaskInfo_t streamSwitchNTask;
  248. rtStreamLabelSwitchByIndexTask_t streamLabelSwitchIndexTask;
  249. rtStreamLabelGotoTask_t streamLabelGotoTask;
  250. uint32_t reserved[10];
  251. } u;
  252. } rtTaskInfo_t;
  253. typedef struct tagNodeInfo_t {
  254. uint32_t nodeIdx;
  255. uint32_t reserved[1];
  256. } rtNodeInfo;
  257. typedef struct tagHwtsInfo_t {
  258. uint16_t taskId;
  259. uint16_t sqExeHead;
  260. uint16_t streamExeHead;
  261. uint16_t reserved[2];
  262. } rtHwtsInfo;
  263. typedef struct tagLabelDevInfo_t {
  264. uint16_t modelId;
  265. uint16_t streamId;
  266. uint16_t labelId;
  267. union {
  268. rtNodeInfo nodeInfo;
  269. rtHwtsInfo hwtsInfo;
  270. uint16_t reserved[5];
  271. }u;
  272. }rtLabelDevInfo;
  273. typedef rtError_t (*rtTaskGenCallback)(rtModel_t mdl, rtTaskInfo_t *taskInfo);
  274. /**
  275. * @ingroup rt_model
  276. * @brief set callback for generate model
  277. * @param [in] callBack callback function
  278. * @return RT_ERROR_NONE for ok
  279. * @return RT_ERROR_INVALID_VALUE for error input
  280. */
  281. RTS_API rtError_t rtSetTaskGenCallback(rtTaskGenCallback callback);
  282. /**
  283. * @ingroup rt_model
  284. * @brief create model instance
  285. * @param [out] mdl created model
  286. * @param [in] flag reserved
  287. * @return RT_ERROR_NONE for ok
  288. * @return RT_ERROR_INVALID_VALUE for error input
  289. */
  290. RTS_API rtError_t rtModelCreate(rtModel_t *mdl, uint32_t flag);
  291. /**
  292. * @ingroup rt_model
  293. * @brief destroy model instance
  294. * @param [in] mdl model to destroy
  295. * @return RT_ERROR_NONE for ok
  296. * @return RT_ERROR_INVALID_VALUE for error input
  297. */
  298. RTS_API rtError_t rtModelDestroy(rtModel_t mdl);
  299. /**
  300. * @ingroup rt_model
  301. * @brief bind model and stream instance
  302. * @param [in] mdl binded model
  303. * @param [in] stm binded stream
  304. * @param [in] flag reserved
  305. * @return RT_ERROR_NONE for ok
  306. * @return RT_ERROR_INVALID_VALUE for error input
  307. */
  308. RTS_API rtError_t rtModelBindStream(rtModel_t mdl, rtStream_t stm, uint32_t flag);
  309. /**
  310. * @ingroup rt_model
  311. * @brief unbind model and stream instance
  312. * @param [in] mdl unbinded model
  313. * @param [in] stm unbinded stream
  314. * @return RT_ERROR_NONE for ok
  315. * @return RT_ERROR_INVALID_VALUE for error input
  316. */
  317. RTS_API rtError_t rtModelUnbindStream(rtModel_t mdl, rtStream_t stm);
  318. /**
  319. * @ingroup rt_model
  320. * @brief tell runtime Model has been Loaded
  321. * @param [in] mdl model to execute
  322. * @return RT_ERROR_NONE for ok
  323. */
  324. RTS_API rtError_t rtModelLoadComplete(rtModel_t mdl);
  325. /**
  326. * @ingroup rt_model
  327. * @brief execute model instance
  328. * @param [in] mdl model to execute
  329. * @return RT_ERROR_NONE for ok
  330. * @return RT_ERROR_INVALID_VALUE for error input
  331. */
  332. RTS_API rtError_t rtModelExecute(rtModel_t mdl, rtStream_t stm, uint32_t flag);
  333. /**
  334. * @ingroup rt_model
  335. * @brief get model the last persist task id
  336. * @param [in] mdl model to execute
  337. * @param [out] taskId last task id of the model
  338. * @param [out] streamId last steam id of the model
  339. * @return RT_ERROR_NONE for ok
  340. * @return RT_ERROR_INVALID_VALUE for error input
  341. */
  342. RTS_API rtError_t rtModelGetTaskId(rtModel_t mdl, uint32_t *taskId, uint32_t *streamId);
  343. /**
  344. * @ingroup rt_model
  345. * @brief add a end graph task to stream
  346. * @param [in] mdl model to execute
  347. * @param [in] end graph stream
  348. * @return RT_ERROR_NONE for ok
  349. * @return RT_ERROR_INVALID_VALUE for error input
  350. */
  351. RTS_API rtError_t rtEndGraph(rtModel_t mdl, rtStream_t stm);
  352. /**
  353. * @ingroup rt_model
  354. * @brief add a end graph task with flag to stream
  355. * @param [in] mdl model to execute
  356. * @param [in] end graph stream
  357. * @param [in] flags AICPU datadump
  358. * @return RT_ERROR_NONE for ok
  359. * @return RT_ERROR_INVALID_VALUE for error input
  360. */
  361. RTS_API rtError_t rtEndGraphEx(rtModel_t mdl, rtStream_t stm, uint32_t flags);
  362. /**
  363. * @ingroup rt_model
  364. * @brief add a end graph task to stream
  365. * @param [in] mdl model to execute
  366. * @param [in] flags EXECUTOR_TS | EXECUTOR_AICPU
  367. * @return RT_ERROR_NONE for ok
  368. * @return RT_ERROR_INVALID_VALUE for error input
  369. */
  370. RTS_API rtError_t rtModelExecutorSet(rtModel_t mdl, uint8_t flags);
  371. /**
  372. * @ingroup rt_model
  373. * @brief abort model
  374. * @param [in] mdl model to abort
  375. * @return RT_ERROR_NONE for ok
  376. * @return RT_ERROR_INVALID_VALUE for error input
  377. */
  378. RTS_API rtError_t rtModelAbort(rtModel_t mdl);
  379. /**
  380. * @ingroup rt_model
  381. * @brief end graph task to model default stream
  382. * @param [in] mdl model to execute
  383. * @param [in] end graph stream
  384. * @return RT_ERROR_NONE for ok
  385. * @return RT_ERROR_INVALID_VALUE for error input
  386. */
  387. RTS_API rtError_t rtModelExit(rtModel_t mdl, rtStream_t stm);
  388. /**
  389. * @ingroup rt_model
  390. * @brief bind queue
  391. * @param [in] mdl model to bind
  392. * @param [in] queueId queueId to bind
  393. * @param [in] flag
  394. * @return RT_ERROR_NONE for ok
  395. * @return RT_ERROR_INVALID_VALUE for error input
  396. */
  397. RTS_API rtError_t rtModelBindQueue(rtModel_t mdl, uint32_t queueId, rtModelQueueFlag_t flag);
  398. /**
  399. * @ingroup rt_model
  400. * @brief get model id
  401. * @param [in] mdl
  402. * @param [out] modelId model id
  403. * @return RT_ERROR_NONE for ok
  404. * @return RT_ERROR_INVALID_VALUE for error input
  405. */
  406. RTS_API rtError_t rtModelGetId(rtModel_t mdl, uint32_t *modelId);
  407. /*
  408. * @ingroup rt_model
  409. * @brief enable debug for dump overflow exception
  410. * @param [in] addr: ddr address of kernel exception dumpped
  411. * @param [in] mdl: model handle
  412. * @param [in] flag: debug flag
  413. * @return RT_ERROR_NONE for ok
  414. * @return RT_ERROR_INVALID_VALUE for error input
  415. */
  416. RTS_API rtError_t rtDebugRegister(rtModel_t mdl, uint32_t flag, const void *addr,
  417. uint32_t *streamId, uint32_t *taskId);
  418. /*
  419. * @ingroup rt_model
  420. * @brief disable debug for dump overflow exception
  421. * @param [in] mdl: model handle
  422. * @return RT_ERROR_NONE for ok
  423. * @return RT_ERROR_INVALID_VALUE for error input
  424. */
  425. RTS_API rtError_t rtDebugUnRegister(rtModel_t mdl);
  426. #if defined(__cplusplus)
  427. }
  428. #endif
  429. #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