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
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. * Licensed under the Apache License, Version 2.0 (the "License");
  4. * you may not use this file except in compliance with the License.
  5. * You may obtain a copy of the License at
  6. * http://www.apache.org/licenses/LICENSE-2.0
  7. * Unless required by applicable law or agreed to in writing, software
  8. * distributed under the License is distributed on an "AS IS" BASIS,
  9. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  10. * See the License for the specific language governing permissions and
  11. * limitations under the License.
  12. */
  13. #ifndef __CCE_RUNTIME_MODEL_H__
  14. #define __CCE_RUNTIME_MODEL_H__
  15. #include "base.h"
  16. #if defined(__cplusplus) && !defined(COMPILE_OMG_PACKAGE)
  17. extern "C" {
  18. #endif
  19. typedef enum tagModelTaskType {
  20. RT_MODEL_TASK_KERNEL = 0,
  21. RT_MODEL_TASK_EVENT_RECORD,
  22. RT_MODEL_TASK_EVENT_WAIT,
  23. RT_MODEL_TASK_FUSION_START,
  24. RT_MODEL_TASK_FUSION_END,
  25. RT_MODEL_TASK_KERNEL_EX,
  26. RT_MODEL_TASK_HCCL,
  27. RT_MODEL_TASK_STREAM_SWITCH,
  28. RT_MODEL_TASK_STREAM_ACTIVE,
  29. RT_MODEL_TASK_LABEL_SET,
  30. RT_MODEL_TASK_LABEL_SWITCH,
  31. RT_MODEL_TASK_LABEL_GOTO,
  32. RT_MODEL_TASK_PROFILER_TRACE,
  33. RT_MODEL_TASK_MEMCPY_ASYNC,
  34. RT_MODEL_TASK_NOTIFY_RECORD,
  35. RT_MODEL_TASK_NOTIFY_WAIT,
  36. RT_MODEL_TASK_REDUCE_ASYNC,
  37. RT_MODEL_TASK_RDMA_SEND,
  38. RT_MODEL_TASK_EVENT_RESET = 18,
  39. RT_MODEL_TASK_MODEL_END_GRAPH,
  40. RT_MODEL_TASK_STREAM_SWITCH_N,
  41. RT_MODEL_TASK_RDMA_DB_SEND,
  42. RT_MODEL_TASK_MEMCPY_ADDR_ASYNC,
  43. RT_MODEL_TASK_STREAM_LABEL_SWITCH_BY_INDEX,
  44. RT_MODEL_TASK_STREAM_LABEL_GOTO,
  45. RT_MODEL_TASK_MODEL_EXIT,
  46. RT_MODEL_TASK_ALL_KERNEL,
  47. } rtModelTaskType_t;
  48. typedef enum tagModelStreamType {
  49. RT_MODEL_HEAD_STREAM = 0,
  50. RT_MODEL_WAIT_ACTIVE_STREAM = 1
  51. } rtModelStreamType_t;
  52. typedef enum tagModelQueueFlag {
  53. RT_MODEL_INPUT_QUEUE = 0,
  54. RT_MODEL_OUTPUT_QUEUE = 1
  55. } rtModelQueueFlag_t;
  56. #define EXECUTOR_NONE ((uint32_t)0x0)
  57. #define EXECUTOR_TS ((uint32_t)0x01)
  58. #define EXECUTOR_AICPU ((uint32_t)0x02)
  59. /*
  60. * @ingroup rt_model
  61. * @brief debug flag for kernel exception dump
  62. */
  63. #define RT_DEBUG_FLAG_AICORE_OVERFLOW (0x1 << 0)
  64. #define RT_DEBUG_FLAG_ATOMIC_ADD_OVERFLOW (0x1 << 1)
  65. /**
  66. * @ingroup
  67. * @brief the type defination of aicpu model task command
  68. */
  69. typedef enum tagTsAicpuModelCmd {
  70. TS_AICPU_MODEL_LOAD = 1,
  71. TS_AICPU_MODEL_EXECUTE,
  72. TS_AICPU_MODEL_DESTROY,
  73. TS_AICPU_MODEL_ABORT,
  74. TS_AICPU_MODEL_RESERVED,
  75. } tsAicpuModelCmd;
  76. typedef struct tagAicpuTaskInfo {
  77. uint32_t taskID;
  78. uint32_t streamID;
  79. uint32_t kernelType;
  80. uint64_t kernelName;
  81. uint64_t kernelSo;
  82. uint64_t paraBase;
  83. uint32_t taskFlag;
  84. } rtAicpuTaskInfo_t;
  85. typedef struct tagModelStreamInfo {
  86. uint32_t streamID;
  87. uint32_t streamFlag;
  88. } rtModelStreamInfo_t;
  89. typedef struct tagModelQueueInfo {
  90. uint32_t queueID;
  91. uint32_t flag;
  92. } rtModelQueueInfo_t;
  93. typedef struct tagAicpuModelInfo {
  94. uint32_t moduleID;
  95. uint32_t tsId;
  96. uint16_t streamInfoNum;
  97. uint16_t aicpuTaskNum;
  98. uint64_t streamInfoPtr;
  99. uint64_t aicpuTaskPtr;
  100. uint16_t queueSize;
  101. uint64_t queueInfoPtr;
  102. } rtAicpuModelInfo_t;
  103. typedef struct tagKernelTaskInfo {
  104. uint16_t blockDim;
  105. uint16_t argsCount;
  106. uint16_t argsSize;
  107. uint16_t reserved;
  108. char *stubFunc;
  109. uint8_t *smDesc;
  110. uint8_t *args;
  111. uint16_t *argsOffset;
  112. } rtKernelTaskInfo_t;
  113. typedef struct tagAllKernelTaskInfo {
  114. uint16_t blockDim;
  115. uint16_t argsCount;
  116. uint16_t argsSize;
  117. uint16_t reserved;
  118. void *devfunc;
  119. void *handle;
  120. uint8_t *smDesc;
  121. uint8_t *args;
  122. uint16_t *argsOffset;
  123. } rtAllKernelTaskInfo_t;
  124. typedef struct tagKernelTaskInfoEx {
  125. uint32_t flags;
  126. uint32_t argsSize;
  127. void *args;
  128. uint32_t reserved[6];
  129. } rtKernelTaskInfoEx_t;
  130. typedef struct tagEventTaskInfo {
  131. uint32_t eventID;
  132. uint32_t reserved[9];
  133. } rtEventTaskInfo_t;
  134. typedef struct tagStreamSwitchTaskInfo {
  135. int64_t value;
  136. uint64_t pValuePtr;
  137. uint32_t trueStreamID;
  138. uint32_t dataType;
  139. uint32_t reserved[4];
  140. } rtStreamSwitchTaskInfo_t;
  141. typedef struct tagStreamSwitchNTaskInfo {
  142. uint64_t pValuePtr;
  143. uint64_t pTrueStreamPtr;
  144. uint32_t size;
  145. uint32_t elementSize;
  146. uint32_t dataType;
  147. uint32_t reserved[3];
  148. } rtStreamSwitchNTaskInfo_t;
  149. typedef struct tagStreamActiveTaskInfo {
  150. uint32_t activeStreamID;
  151. uint32_t reserved[9];
  152. } rtStreamActiveTaskInfo_t;
  153. typedef struct tagSetTaskInfo {
  154. uint16_t labelId;
  155. uint32_t reserved[9];
  156. } rtLabelSetTaskInfo_t;
  157. typedef struct tagSwitchTaskInfo {
  158. uint32_t value;
  159. uint32_t reserved[9];
  160. } rtLabelSwitchTaskInfo_t;
  161. typedef struct tagLabelGotoTaskInfo {
  162. uint16_t labelId;
  163. uint32_t reserved[9];
  164. } rtLabelGotoTaskInfo_t;
  165. typedef struct tagProfilerTraceTaskInfo {
  166. uint64_t profilerTraceId;
  167. uint32_t notify : 8;
  168. uint32_t reserved_ : 24;
  169. uint32_t flags;
  170. uint32_t reserved[6];
  171. } rtProfilerTrace_t;
  172. typedef struct tagrtMemcpyAsyncTaskInfo {
  173. void *dst;
  174. uint64_t destMax;
  175. void *src;
  176. uint64_t count;
  177. uint32_t kind;
  178. uint32_t reserved;
  179. } rtMemcpyAsyncTaskInfo_t;
  180. typedef struct tagrtNotifyTaskInfo {
  181. uint32_t notifyID;
  182. uint32_t reserved[9];
  183. } rtNotifyTaskInfo_t;
  184. typedef struct tagrtReduceAsyncTaskInfo {
  185. void *dst;
  186. uint64_t destMax;
  187. void *src;
  188. uint64_t count;
  189. uint32_t kind;
  190. uint32_t type;
  191. } rtReduceAsyncTaskInfo_t;
  192. typedef struct tagrtRdmaSendTaskInfo {
  193. uint32_t index;
  194. uint32_t wqe_index;
  195. uint32_t reserved[8];
  196. } rtRdmaSendTaskInfo_t;
  197. typedef struct tagrtRdmaDbSendTaskInfo {
  198. uint64_t dbInfo;
  199. uint32_t dbIndex;
  200. uint32_t reserved[7]; // offset 7
  201. } rtRdmaDbSendTaskInfo_t;
  202. typedef struct tagrtModelEndGraphTaskInfo {
  203. uint32_t modelId;
  204. uint32_t executorFlag;
  205. uint32_t reserved[8];
  206. } rtModelEndGraphTaskInfo_t;
  207. typedef struct tagrtModelExitInfo {
  208. uint32_t modelId;
  209. uint32_t streamId;
  210. uint32_t reserved[8];
  211. } rtModelExitTaskInfo_t;
  212. typedef struct tagrtStreamLabelSwitchByIndexTask_t {
  213. uint64_t indexPtr;
  214. uint64_t labelInfoPtr;
  215. uint32_t max;
  216. uint8_t reserved[20];
  217. } rtStreamLabelSwitchByIndexTask_t;
  218. typedef struct tagrtStreamLabelGotoTask_t {
  219. uint16_t labelId;
  220. uint16_t modelId;
  221. uint8_t reserved[36];
  222. } rtStreamLabelGotoTask_t;
  223. typedef struct tagTaskInfo {
  224. uint32_t type;
  225. uint32_t streamID;
  226. union {
  227. rtKernelTaskInfoEx_t kernelTaskEx;
  228. rtKernelTaskInfo_t kernelTask;
  229. rtAllKernelTaskInfo_t allKernelTask;
  230. rtEventTaskInfo_t eventTask;
  231. rtStreamSwitchTaskInfo_t streamSwitchTask;
  232. rtStreamActiveTaskInfo_t streamActiveTask;
  233. rtLabelSetTaskInfo_t labelSetTask;
  234. rtLabelSwitchTaskInfo_t labelSwitchTask;
  235. rtLabelGotoTaskInfo_t labelGotoTask;
  236. rtProfilerTrace_t profilertraceTask;
  237. rtMemcpyAsyncTaskInfo_t memcpyAsyncTask;
  238. rtNotifyTaskInfo_t notifyTask;
  239. rtReduceAsyncTaskInfo_t reduceAsyncTask;
  240. rtRdmaSendTaskInfo_t rdmaSendTask;
  241. rtRdmaDbSendTaskInfo_t rdmaDbSendTask;
  242. rtModelEndGraphTaskInfo_t modelEndGraphTask;
  243. rtModelExitTaskInfo_t modelExitTask;
  244. rtStreamSwitchNTaskInfo_t streamSwitchNTask;
  245. rtStreamLabelSwitchByIndexTask_t streamLabelSwitchIndexTask;
  246. rtStreamLabelGotoTask_t streamLabelGotoTask;
  247. uint32_t reserved[10];
  248. } u;
  249. } rtTaskInfo_t;
  250. typedef struct tagNodeInfo_t {
  251. uint32_t nodeIdx;
  252. uint32_t reserved[1];
  253. } rtNodeInfo;
  254. typedef struct tagHwtsInfo_t {
  255. uint16_t taskId;
  256. uint16_t sqExeHead;
  257. uint16_t streamExeHead;
  258. uint16_t reserved[2];
  259. } rtHwtsInfo;
  260. typedef struct tagLabelDevInfo_t {
  261. uint16_t modelId;
  262. uint16_t streamId;
  263. uint16_t labelId;
  264. union {
  265. rtNodeInfo nodeInfo;
  266. rtHwtsInfo hwtsInfo;
  267. uint16_t reserved[5];
  268. }u;
  269. }rtLabelDevInfo;
  270. typedef rtError_t (*rtTaskGenCallback)(rtModel_t model, rtTaskInfo_t *taskInfo);
  271. /**
  272. * @ingroup rt_model
  273. * @brief set callback for generate model
  274. * @param [in] callBack callback function
  275. * @return RT_ERROR_NONE for ok
  276. * @return RT_ERROR_INVALID_VALUE for error input
  277. */
  278. RTS_API rtError_t rtSetTaskGenCallback(rtTaskGenCallback callback);
  279. /**
  280. * @ingroup rt_model
  281. * @brief create model instance
  282. * @param [out] model created model
  283. * @param [in] flag reserved
  284. * @return RT_ERROR_NONE for ok
  285. * @return RT_ERROR_INVALID_VALUE for error input
  286. */
  287. RTS_API rtError_t rtModelCreate(rtModel_t *model, uint32_t flag);
  288. /**
  289. * @ingroup rt_model
  290. * @brief destroy model instance
  291. * @param [in] model model to destroy
  292. * @return RT_ERROR_NONE for ok
  293. * @return RT_ERROR_INVALID_VALUE for error input
  294. */
  295. RTS_API rtError_t rtModelDestroy(rtModel_t model);
  296. /**
  297. * @ingroup rt_model
  298. * @brief bind model and stream instance
  299. * @param [in] model binded model
  300. * @param [in] stream binded stream
  301. * @param [in] flag reserved
  302. * @return RT_ERROR_NONE for ok
  303. * @return RT_ERROR_INVALID_VALUE for error input
  304. */
  305. RTS_API rtError_t rtModelBindStream(rtModel_t model, rtStream_t stream, uint32_t flag);
  306. /**
  307. * @ingroup rt_model
  308. * @brief unbind model and stream instance
  309. * @param [in] model unbinded model
  310. * @param [in] stream unbinded stream
  311. * @return RT_ERROR_NONE for ok
  312. * @return RT_ERROR_INVALID_VALUE for error input
  313. */
  314. RTS_API rtError_t rtModelUnbindStream(rtModel_t model, rtStream_t stream);
  315. /**
  316. * @ingroup rt_model
  317. * @brief tell runtime Model has been Loaded
  318. * @param [in] model model to execute
  319. * @return RT_ERROR_NONE for ok
  320. */
  321. RTS_API rtError_t rtModelLoadComplete(rtModel_t model);
  322. /**
  323. * @ingroup rt_model
  324. * @brief execute model instance
  325. * @param [in] model model to execute
  326. * @return RT_ERROR_NONE for ok
  327. * @return RT_ERROR_INVALID_VALUE for error input
  328. */
  329. RTS_API rtError_t rtModelExecute(rtModel_t model, rtStream_t stream, uint32_t flag);
  330. /**
  331. * @ingroup rt_model
  332. * @brief get model the last persist task id
  333. * @param [in] model model to execute
  334. * @param [out] taskid last task id of the model
  335. * @param [out] streamid last steam id of the model
  336. * @return RT_ERROR_NONE for ok
  337. * @return RT_ERROR_INVALID_VALUE for error input
  338. */
  339. RTS_API rtError_t rtModelGetTaskId(rtModel_t model, uint32_t *taskid, uint32_t *streamid);
  340. /**
  341. * @ingroup rt_model
  342. * @brief add a end graph task to stream
  343. * @param [in] model model to execute
  344. * @param [in] end graph stream
  345. * @return RT_ERROR_NONE for ok
  346. * @return RT_ERROR_INVALID_VALUE for error input
  347. */
  348. RTS_API rtError_t rtEndGraph(rtModel_t model, rtStream_t stream);
  349. /**
  350. * @ingroup rt_model
  351. * @brief add a end graph task with flag to stream
  352. * @param [in] model model to execute
  353. * @param [in] end graph stream
  354. * @param [in] flags AICPU datadump
  355. * @return RT_ERROR_NONE for ok
  356. * @return RT_ERROR_INVALID_VALUE for error input
  357. */
  358. RTS_API rtError_t rtEndGraphEx(rtModel_t model, rtStream_t stream, uint32_t flags);
  359. /**
  360. * @ingroup rt_model
  361. * @brief add a end graph task to stream
  362. * @param [in] model model to execute
  363. * @param [in] flags EXECUTOR_TS | EXECUTOR_AICPU
  364. * @return RT_ERROR_NONE for ok
  365. * @return RT_ERROR_INVALID_VALUE for error input
  366. */
  367. RTS_API rtError_t rtModelExecutorSet(rtModel_t model, uint8_t flags);
  368. /**
  369. * @ingroup rt_model
  370. * @brief abort model
  371. * @param [in] model model to abort
  372. * @return RT_ERROR_NONE for ok
  373. * @return RT_ERROR_INVALID_VALUE for error input
  374. */
  375. RTS_API rtError_t rtModelAbort(rtModel_t model);
  376. /**
  377. * @ingroup rt_model
  378. * @brief end graph task to model default stream
  379. * @param [in] model model to execute
  380. * @param [in] end graph stream
  381. * @return RT_ERROR_NONE for ok
  382. * @return RT_ERROR_INVALID_VALUE for error input
  383. */
  384. RTS_API rtError_t rtModelExit(rtModel_t model, rtStream_t stream);
  385. /**
  386. * @ingroup rt_model
  387. * @brief bind queue
  388. * @param [in] model model to bind
  389. * @param [in] queueId queueId to bind
  390. * @param [in] flag
  391. * @return RT_ERROR_NONE for ok
  392. * @return RT_ERROR_INVALID_VALUE for error input
  393. */
  394. RTS_API rtError_t rtModelBindQueue(rtModel_t model, uint32_t queueId, rtModelQueueFlag_t flag);
  395. /**
  396. * @ingroup rt_model
  397. * @brief get model id
  398. * @param [in] model
  399. * @param [out] modelId model id
  400. * @return RT_ERROR_NONE for ok
  401. * @return RT_ERROR_INVALID_VALUE for error input
  402. */
  403. RTS_API rtError_t rtModelGetId(rtModel_t model, uint32_t *modelId);
  404. /*
  405. * @ingroup rt_model
  406. * @brief enable debug for dump overflow exception
  407. * @param [in] addr: ddr address of kernel exception dumpped
  408. * @param [in] model: model handle
  409. * @param [in] flag: debug flag
  410. * @return RT_ERROR_NONE for ok
  411. * @return RT_ERROR_INVALID_VALUE for error input
  412. */
  413. RTS_API rtError_t rtDebugRegister(rtModel_t model, uint32_t flag, const void *addr,
  414. uint32_t *streamId, uint32_t *taskId);
  415. /*
  416. * @ingroup rt_model
  417. * @brief disable debug for dump overflow exception
  418. * @param [in] model: model handle
  419. * @return RT_ERROR_NONE for ok
  420. * @return RT_ERROR_INVALID_VALUE for error input
  421. */
  422. RTS_API rtError_t rtDebugUnRegister(rtModel_t model);
  423. #if defined(__cplusplus) && !defined(COMPILE_OMG_PACKAGE)
  424. }
  425. #endif
  426. #endif // __CCE_RUNTIME_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