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
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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497
  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_MODEL_H__
  17. #define __CCE_RUNTIME_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 = 18,
  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. } rtModelTaskType_t;
  53. typedef enum tagModelStreamType {
  54. RT_MODEL_HEAD_STREAM = 0,
  55. RT_MODEL_WAIT_ACTIVE_STREAM = 1
  56. } rtModelStreamType_t;
  57. typedef enum tagModelQueueFlag {
  58. RT_MODEL_INPUT_QUEUE = 0,
  59. RT_MODEL_OUTPUT_QUEUE = 1
  60. } rtModelQueueFlag_t;
  61. #define EXECUTOR_NONE ((uint32_t)0x0)
  62. #define EXECUTOR_TS ((uint32_t)0x01)
  63. #define EXECUTOR_AICPU ((uint32_t)0x02)
  64. /*
  65. * @ingroup rt_model
  66. * @brief debug flag for kernel exception dump
  67. */
  68. #define RT_DEBUG_FLAG_AICORE_OVERFLOW (0x1 << 0)
  69. #define RT_DEBUG_FLAG_ATOMIC_ADD_OVERFLOW (0x1 << 1)
  70. /**
  71. * @ingroup
  72. * @brief the type defination of aicpu model task command
  73. */
  74. typedef enum tagTsAicpuModelCmd {
  75. TS_AICPU_MODEL_LOAD = 1,
  76. TS_AICPU_MODEL_EXECUTE,
  77. TS_AICPU_MODEL_DESTROY,
  78. TS_AICPU_MODEL_ABORT,
  79. TS_AICPU_MODEL_RESERVED,
  80. } tsAicpuModelCmd;
  81. typedef struct tagAicpuTaskInfo {
  82. uint32_t taskID;
  83. uint32_t streamID;
  84. uint32_t kernelType;
  85. uint64_t kernelName;
  86. uint64_t kernelSo;
  87. uint64_t paraBase;
  88. uint32_t taskFlag;
  89. } rtAicpuTaskInfo_t;
  90. typedef struct tagModelStreamInfo {
  91. uint32_t streamID;
  92. uint32_t streamFlag;
  93. } rtModelStreamInfo_t;
  94. typedef struct tagModelQueueInfo {
  95. uint32_t queueID;
  96. uint32_t flag;
  97. } rtModelQueueInfo_t;
  98. typedef struct tagAicpuModelInfo {
  99. uint32_t moduleID;
  100. uint32_t tsId;
  101. uint16_t streamInfoNum;
  102. uint16_t aicpuTaskNum;
  103. uint64_t streamInfoPtr;
  104. uint64_t aicpuTaskPtr;
  105. uint16_t queueSize;
  106. uint64_t queueInfoPtr;
  107. } rtAicpuModelInfo_t;
  108. typedef struct tagKernelTaskInfo {
  109. uint16_t blockDim;
  110. uint16_t argsCount;
  111. uint16_t argsSize;
  112. uint16_t reserved;
  113. char *stubFunc;
  114. uint8_t *smDesc;
  115. uint8_t *args;
  116. uint16_t *argsOffset;
  117. } rtKernelTaskInfo_t;
  118. typedef struct tagAllKernelTaskInfo {
  119. uint16_t blockDim;
  120. uint16_t argsCount;
  121. uint16_t argsSize;
  122. uint16_t reserved;
  123. void *devfunc;
  124. void *handle;
  125. uint8_t *smDesc;
  126. uint8_t *args;
  127. uint16_t *argsOffset;
  128. } rtAllKernelTaskInfo_t;
  129. typedef struct tagKernelTaskInfoEx {
  130. uint32_t flags;
  131. uint32_t argsSize;
  132. void *args;
  133. uint32_t reserved[6];
  134. } rtKernelTaskInfoEx_t;
  135. typedef struct tagEventTaskInfo {
  136. uint32_t eventID;
  137. uint32_t reserved[9];
  138. } rtEventTaskInfo_t;
  139. typedef struct tagStreamSwitchTaskInfo {
  140. int64_t value;
  141. uint64_t pValuePtr;
  142. uint32_t trueStreamID;
  143. uint32_t dataType;
  144. uint32_t reserved[4];
  145. } rtStreamSwitchTaskInfo_t;
  146. typedef struct tagStreamSwitchNTaskInfo {
  147. uint64_t pValuePtr;
  148. uint64_t pTrueStreamPtr;
  149. uint32_t size;
  150. uint32_t elementSize;
  151. uint32_t dataType;
  152. uint32_t reserved[3];
  153. } rtStreamSwitchNTaskInfo_t;
  154. typedef struct tagStreamActiveTaskInfo {
  155. uint32_t activeStreamID;
  156. uint32_t reserved[9];
  157. } rtStreamActiveTaskInfo_t;
  158. typedef struct tagSetTaskInfo {
  159. uint16_t labelId;
  160. uint32_t reserved[9];
  161. } rtLabelSetTaskInfo_t;
  162. typedef struct tagSwitchTaskInfo {
  163. uint32_t value;
  164. uint32_t reserved[9];
  165. } rtLabelSwitchTaskInfo_t;
  166. typedef struct tagLabelGotoTaskInfo {
  167. uint16_t labelId;
  168. uint32_t reserved[9];
  169. } rtLabelGotoTaskInfo_t;
  170. typedef struct tagProfilerTraceTaskInfo {
  171. uint64_t profilerTraceId;
  172. uint32_t notify : 8;
  173. uint32_t reserved_ : 24;
  174. uint32_t flags;
  175. uint32_t reserved[6];
  176. } rtProfilerTrace_t;
  177. typedef struct tagProfilerTraceExTaskInfo {
  178. uint64_t profilerTraceId;
  179. uint64_t modelId;
  180. uint16_t tagId;
  181. uint8_t reserved[22];
  182. } rtProfilerTraceEx_t;
  183. typedef struct tagrtMemcpyAsyncTaskInfo {
  184. void *dst;
  185. uint64_t destMax;
  186. void *src;
  187. uint64_t count;
  188. uint32_t kind;
  189. uint32_t reserved;
  190. } rtMemcpyAsyncTaskInfo_t;
  191. typedef struct tagrtNotifyTaskInfo {
  192. uint32_t notifyID;
  193. uint32_t reserved[9];
  194. } rtNotifyTaskInfo_t;
  195. typedef struct tagrtReduceAsyncTaskInfo {
  196. void *dst;
  197. uint64_t destMax;
  198. void *src;
  199. uint64_t count;
  200. uint32_t kind;
  201. uint32_t type;
  202. } rtReduceAsyncTaskInfo_t;
  203. typedef struct tagrtRdmaSendTaskInfo {
  204. uint32_t index;
  205. uint32_t wqe_index;
  206. uint32_t reserved[8];
  207. } rtRdmaSendTaskInfo_t;
  208. typedef struct tagrtRdmaDbSendTaskInfo {
  209. uint64_t dbInfo;
  210. uint32_t dbIndex;
  211. uint32_t reserved[7]; // offset 7
  212. } rtRdmaDbSendTaskInfo_t;
  213. typedef struct tagrtModelEndGraphTaskInfo {
  214. uint32_t modelId;
  215. uint32_t executorFlag;
  216. uint32_t reserved[8];
  217. } rtModelEndGraphTaskInfo_t;
  218. typedef struct tagrtModelExitInfo {
  219. uint32_t modelId;
  220. uint32_t streamId;
  221. uint32_t reserved[8];
  222. } rtModelExitTaskInfo_t;
  223. typedef struct tagrtStreamLabelSwitchByIndexTask_t {
  224. uint64_t indexPtr;
  225. uint64_t labelInfoPtr;
  226. uint32_t max;
  227. uint8_t reserved[20];
  228. } rtStreamLabelSwitchByIndexTask_t;
  229. typedef struct tagrtStreamLabelGotoTask_t {
  230. uint16_t labelId;
  231. uint16_t modelId;
  232. uint8_t reserved[36];
  233. } rtStreamLabelGotoTask_t;
  234. typedef struct tagTaskInfo {
  235. uint32_t type;
  236. uint32_t streamID;
  237. union {
  238. rtKernelTaskInfoEx_t kernelTaskEx;
  239. rtKernelTaskInfo_t kernelTask;
  240. rtAllKernelTaskInfo_t allKernelTask;
  241. rtEventTaskInfo_t eventTask;
  242. rtStreamSwitchTaskInfo_t streamSwitchTask;
  243. rtStreamActiveTaskInfo_t streamActiveTask;
  244. rtLabelSetTaskInfo_t labelSetTask;
  245. rtLabelSwitchTaskInfo_t labelSwitchTask;
  246. rtLabelGotoTaskInfo_t labelGotoTask;
  247. rtProfilerTrace_t profilertraceTask;
  248. rtProfilerTraceEx_t profilertraceExTask;
  249. rtMemcpyAsyncTaskInfo_t memcpyAsyncTask;
  250. rtNotifyTaskInfo_t notifyTask;
  251. rtReduceAsyncTaskInfo_t reduceAsyncTask;
  252. rtRdmaSendTaskInfo_t rdmaSendTask;
  253. rtRdmaDbSendTaskInfo_t rdmaDbSendTask;
  254. rtModelEndGraphTaskInfo_t modelEndGraphTask;
  255. rtModelExitTaskInfo_t modelExitTask;
  256. rtStreamSwitchNTaskInfo_t streamSwitchNTask;
  257. rtStreamLabelSwitchByIndexTask_t streamLabelSwitchIndexTask;
  258. rtStreamLabelGotoTask_t streamLabelGotoTask;
  259. uint32_t reserved[10];
  260. } u;
  261. } rtTaskInfo_t;
  262. typedef struct tagNodeInfo_t {
  263. uint32_t nodeIdx;
  264. uint32_t reserved[1];
  265. } rtNodeInfo;
  266. typedef struct tagHwtsInfo_t {
  267. uint16_t taskId;
  268. uint16_t sqExeHead;
  269. uint16_t streamExeHead;
  270. uint16_t reserved[2];
  271. } rtHwtsInfo;
  272. typedef struct tagLabelDevInfo_t {
  273. uint16_t modelId;
  274. uint16_t streamId;
  275. uint16_t labelId;
  276. union {
  277. rtNodeInfo nodeInfo;
  278. rtHwtsInfo hwtsInfo;
  279. uint16_t reserved[5];
  280. }u;
  281. }rtLabelDevInfo;
  282. typedef rtError_t (*rtTaskGenCallback)(rtModel_t model, rtTaskInfo_t *taskInfo);
  283. /**
  284. * @ingroup rt_model
  285. * @brief set callback for generate model
  286. * @param [in] callBack callback function
  287. * @return RT_ERROR_NONE for ok
  288. * @return RT_ERROR_INVALID_VALUE for error input
  289. */
  290. RTS_API rtError_t rtSetTaskGenCallback(rtTaskGenCallback callback);
  291. /**
  292. * @ingroup rt_model
  293. * @brief create model instance
  294. * @param [out] model created model
  295. * @param [in] flag reserved
  296. * @return RT_ERROR_NONE for ok
  297. * @return RT_ERROR_INVALID_VALUE for error input
  298. */
  299. RTS_API rtError_t rtModelCreate(rtModel_t *model, uint32_t flag);
  300. /**
  301. * @ingroup rt_model
  302. * @brief destroy model instance
  303. * @param [in] model model to destroy
  304. * @return RT_ERROR_NONE for ok
  305. * @return RT_ERROR_INVALID_VALUE for error input
  306. */
  307. RTS_API rtError_t rtModelDestroy(rtModel_t model);
  308. /**
  309. * @ingroup rt_model
  310. * @brief bind model and stream instance
  311. * @param [in] model binded model
  312. * @param [in] stream binded stream
  313. * @param [in] flag reserved
  314. * @return RT_ERROR_NONE for ok
  315. * @return RT_ERROR_INVALID_VALUE for error input
  316. */
  317. RTS_API rtError_t rtModelBindStream(rtModel_t model, rtStream_t stream, uint32_t flag);
  318. /**
  319. * @ingroup rt_model
  320. * @brief unbind model and stream instance
  321. * @param [in] model unbinded model
  322. * @param [in] stream unbinded stream
  323. * @return RT_ERROR_NONE for ok
  324. * @return RT_ERROR_INVALID_VALUE for error input
  325. */
  326. RTS_API rtError_t rtModelUnbindStream(rtModel_t model, rtStream_t stream);
  327. /**
  328. * @ingroup rt_model
  329. * @brief tell runtime Model has been Loaded
  330. * @param [in] model model to execute
  331. * @return RT_ERROR_NONE for ok
  332. */
  333. RTS_API rtError_t rtModelLoadComplete(rtModel_t model);
  334. /**
  335. * @ingroup rt_model
  336. * @brief execute model instance
  337. * @param [in] model model to execute
  338. * @return RT_ERROR_NONE for ok
  339. * @return RT_ERROR_INVALID_VALUE for error input
  340. */
  341. RTS_API rtError_t rtModelExecute(rtModel_t model, rtStream_t stream, uint32_t flag);
  342. /**
  343. * @ingroup rt_model
  344. * @brief get model the last persist task id
  345. * @param [in] model model to execute
  346. * @param [out] taskid last task id of the model
  347. * @param [out] streamid last steam id of the model
  348. * @return RT_ERROR_NONE for ok
  349. * @return RT_ERROR_INVALID_VALUE for error input
  350. */
  351. RTS_API rtError_t rtModelGetTaskId(rtModel_t model, uint32_t *taskid, uint32_t *streamid);
  352. /**
  353. * @ingroup rt_model
  354. * @brief add a end graph task to stream
  355. * @param [in] model model to execute
  356. * @param [in] end graph stream
  357. * @return RT_ERROR_NONE for ok
  358. * @return RT_ERROR_INVALID_VALUE for error input
  359. */
  360. RTS_API rtError_t rtEndGraph(rtModel_t model, rtStream_t stream);
  361. /**
  362. * @ingroup rt_model
  363. * @brief add a end graph task with flag to stream
  364. * @param [in] model model to execute
  365. * @param [in] end graph stream
  366. * @param [in] flags AICPU datadump
  367. * @return RT_ERROR_NONE for ok
  368. * @return RT_ERROR_INVALID_VALUE for error input
  369. */
  370. RTS_API rtError_t rtEndGraphEx(rtModel_t model, rtStream_t stream, uint32_t flags);
  371. /**
  372. * @ingroup rt_model
  373. * @brief add a end graph task to stream
  374. * @param [in] model model to execute
  375. * @param [in] flags EXECUTOR_TS | EXECUTOR_AICPU
  376. * @return RT_ERROR_NONE for ok
  377. * @return RT_ERROR_INVALID_VALUE for error input
  378. */
  379. RTS_API rtError_t rtModelExecutorSet(rtModel_t model, uint8_t flags);
  380. /**
  381. * @ingroup rt_model
  382. * @brief abort model
  383. * @param [in] model model to abort
  384. * @return RT_ERROR_NONE for ok
  385. * @return RT_ERROR_INVALID_VALUE for error input
  386. */
  387. RTS_API rtError_t rtModelAbort(rtModel_t model);
  388. /**
  389. * @ingroup rt_model
  390. * @brief end graph task to model default stream
  391. * @param [in] model model to execute
  392. * @param [in] end graph stream
  393. * @return RT_ERROR_NONE for ok
  394. * @return RT_ERROR_INVALID_VALUE for error input
  395. */
  396. RTS_API rtError_t rtModelExit(rtModel_t model, rtStream_t stream);
  397. /**
  398. * @ingroup rt_model
  399. * @brief bind queue
  400. * @param [in] model model to bind
  401. * @param [in] queueId queueId to bind
  402. * @param [in] flag
  403. * @return RT_ERROR_NONE for ok
  404. * @return RT_ERROR_INVALID_VALUE for error input
  405. */
  406. RTS_API rtError_t rtModelBindQueue(rtModel_t model, uint32_t queueId, rtModelQueueFlag_t flag);
  407. /**
  408. * @ingroup rt_model
  409. * @brief get model id
  410. * @param [in] model
  411. * @param [out] modelId model id
  412. * @return RT_ERROR_NONE for ok
  413. * @return RT_ERROR_INVALID_VALUE for error input
  414. */
  415. RTS_API rtError_t rtModelGetId(rtModel_t model, uint32_t *modelId);
  416. /*
  417. * @ingroup rt_model
  418. * @brief enable debug for dump overflow exception
  419. * @param [in] addr: ddr address of kernel exception dumpped
  420. * @param [in] model: model handle
  421. * @param [in] flag: debug flag
  422. * @return RT_ERROR_NONE for ok
  423. * @return RT_ERROR_INVALID_VALUE for error input
  424. */
  425. RTS_API rtError_t rtDebugRegister(rtModel_t model, uint32_t flag, const void *addr,
  426. uint32_t *streamId, uint32_t *taskId);
  427. /*
  428. * @ingroup rt_model
  429. * @brief disable debug for dump overflow exception
  430. * @param [in] model: model handle
  431. * @return RT_ERROR_NONE for ok
  432. * @return RT_ERROR_INVALID_VALUE for error input
  433. */
  434. RTS_API rtError_t rtDebugUnRegister(rtModel_t model);
  435. #if defined(__cplusplus)
  436. }
  437. #endif
  438. #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