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.

kernel.h 25 kB

5 years ago
5 years ago
3 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
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
3 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741
  1. /*
  2. * Copyright (c) Huawei Technologies Co., Ltd. 2020-2021. All rights reserved.
  3. * Description: kernel.h
  4. * Create: 2020-01-01
  5. */
  6. #ifndef CCE_RUNTIME_KERNEL_H
  7. #define CCE_RUNTIME_KERNEL_H
  8. #include "base.h"
  9. #include "stream.h"
  10. #if defined(__cplusplus)
  11. extern "C" {
  12. #endif
  13. /**
  14. * @ingroup rt_kernel
  15. * @brief shared memory data control
  16. */
  17. typedef struct tagRtSmData {
  18. uint64_t L2_mirror_addr; // preload or swap source address
  19. uint32_t L2_data_section_size; // every data size
  20. uint8_t L2_preload; // 1 - preload from mirrorAddr, 0 - no preload
  21. uint8_t modified; // 1 - data will be modified by kernel, 0 - no modified
  22. uint8_t priority; // data priority
  23. int8_t prev_L2_page_offset_base; // remap source section offset
  24. uint8_t L2_page_offset_base; // remap destination section offset
  25. uint8_t L2_load_to_ddr; // 1 - need load out, 0 - no need
  26. uint8_t reserved[2]; // reserved
  27. } rtSmData_t;
  28. /**
  29. * @ingroup rt_kernel
  30. * @brief shared memory description
  31. */
  32. typedef struct tagRtSmCtrl {
  33. rtSmData_t data[8]; // data description
  34. uint64_t size; // max page Num
  35. uint8_t remap[64]; /* just using for static remap mode, default:0xFF
  36. array index: virtual l2 page id, array value: physic l2 page id */
  37. uint8_t l2_in_main; // 0-DDR, 1-L2, default:0xFF
  38. uint8_t reserved[3];
  39. } rtSmDesc_t;
  40. typedef rtSmDesc_t rtL2Ctrl_t;
  41. /**
  42. * @ingroup rt_kernel
  43. * @brief device binary type
  44. */
  45. typedef struct tagRtDevBinary {
  46. uint32_t magic; // magic number
  47. uint32_t version; // version of binary
  48. const void *data; // binary data
  49. uint64_t length; // binary length
  50. } rtDevBinary_t;
  51. /**
  52. * @ingroup rt_kernel
  53. * @brief function mode type
  54. */
  55. #define ONLINE_PROF_MAX_PMU_NUM (8)
  56. typedef struct ProfilefDataInfo {
  57. const void *stubFunc;
  58. uint32_t blockDim;
  59. const void *args;
  60. uint32_t argsSize;
  61. rtSmDesc_t *smDesc;
  62. rtStream_t stream;
  63. uint64_t totalcycle;
  64. uint64_t ovcycle;
  65. uint64_t pmu_cnt[ONLINE_PROF_MAX_PMU_NUM];
  66. } rtProfDataInfo_t;
  67. /**
  68. * @ingroup rt_kernel
  69. * @brief function mode type
  70. */
  71. typedef enum {
  72. FUNC_MODE_NORMAL = 0,
  73. FUNC_MODE_PCTRACE_USERPROFILE_RECORDLOOP,
  74. FUNC_MODE_PCTRACE_USERPROFILE_SKIPLOOP,
  75. FUNC_MODE_PCTRACE_CYCLECNT_RECORDLOOP,
  76. FUNC_MODE_PCTRACE_CYCLECNT_SKIPLOOP,
  77. FUNC_MODE_BUTT
  78. } rtFuncModeType_t;
  79. /**
  80. * @ingroup rt_kernel
  81. * @brief kernel info
  82. */
  83. typedef struct rtKernelInfo {
  84. uint64_t task_offset; // kernel offset in module
  85. /* flowtable */
  86. void *arg; // launch kernel arg
  87. uint32_t arg_size;
  88. /* module */
  89. void *module_addr; // module::baseaddr_
  90. uint32_t module_size;
  91. } *rtKernelInfo_t;
  92. /**
  93. * @ingroup rt_kernel
  94. * @brief op name
  95. */
  96. typedef struct rtKernelLaunchNames {
  97. const char_t *soName; // defined for so name
  98. const char_t *kernelName; // defined for kernel type name
  99. const char_t *opName; // defined for operator name
  100. } rtKernelLaunchNames_t;
  101. /**
  102. * @ingroup rt_kernel
  103. * @brief args struct
  104. */
  105. typedef struct tagRtArgsWithTiling {
  106. void *args; // args host mem addr
  107. uint32_t argsSize; // input + output + tiling addr size + tiling data size
  108. uint32_t argsSizeWithoutTiling; // input + output + tiling addr size
  109. uint16_t tilingAddrOffset; // tiling addr offset
  110. uint16_t tilingDataOffset; // tiling data offset
  111. uint16_t hostInputAddrOffset; // index of host_memory input in inputs_addrs list
  112. uint16_t hostInputDataOffset; // host_mem input data offset
  113. uint8_t hasHostMemInput; // has host_memory input data in args or not: 0 means no host_memory input data,
  114. // others means has host_memory input data.
  115. uint8_t isNoNeedH2DCopy; // is no need host to device copy: 0 means need H2D copy,
  116. // others means doesn't need H2D copy.
  117. uint8_t reserved[6];
  118. } rtArgsWithTiling_t;
  119. /**
  120. * @ingroup rt_KernelConfigDump
  121. * @brief device dump type
  122. */
  123. typedef enum tagRtDumpKind {
  124. RT_DATA_DUMP_KIND_INVALID = -1,
  125. RT_DATA_DUMP_KIND_DUMP = 0,
  126. RT_DATA_DUMP_KIND_RESERVED = 1,
  127. } rtDumpKind_t;
  128. /**
  129. * @ingroup rt_kernel
  130. * @brief report callback
  131. */
  132. typedef rtError_t (*rtKernelReportCallback)(rtStream_t stm, rtKernelInfo_t kernelInfo);
  133. /**
  134. * @ingroup rt_kernel
  135. * @brief stream report callback
  136. */
  137. typedef void (*rtCallback_t)(void *fnData);
  138. /**
  139. * @ingroup rt_kernel
  140. * @brief magic number of plain binary for aicore
  141. */
  142. #define RT_DEV_BINARY_MAGIC_PLAIN 0xabceed50U
  143. /**
  144. * @ingroup rt_kernel
  145. * @brief magic number of plain binary for aicpu
  146. */
  147. #define RT_DEV_BINARY_MAGIC_PLAIN_AICPU 0xabceed51U
  148. /**
  149. * @ingroup rt_kernel
  150. * @brief magic number of plain binary for aivector
  151. */
  152. #define RT_DEV_BINARY_MAGIC_PLAIN_AIVEC 0xabceed52U
  153. /**
  154. * @ingroup rt_kernel
  155. * @brief magic number of elf binary for aicore
  156. */
  157. #define RT_DEV_BINARY_MAGIC_ELF 0x43554245U
  158. /**
  159. * @ingroup rt_kernel
  160. * @brief magic number of elf binary for aicpu
  161. */
  162. #define RT_DEV_BINARY_MAGIC_ELF_AICPU 0x41415243U
  163. /**
  164. * @ingroup rt_kernel
  165. * @brief magic number of elf binary for aivector
  166. */
  167. #define RT_DEV_BINARY_MAGIC_ELF_AIVEC 0x41415246U
  168. /**
  169. * @ingroup rt_kernel
  170. * @brief magic number of elf binary for aicube
  171. */
  172. #define RT_DEV_BINARY_MAGIC_ELF_AICUBE 0x41494343U
  173. /**
  174. * @ingroup rt_kernel_flags
  175. * @brief kernel op bit flags
  176. */
  177. #define RT_KERNEL_DEFAULT (0x00U)
  178. #define RT_KERNEL_CONVERT (0x01U)
  179. #define RT_KERNEL_DUMPFLAG (0x02U)
  180. #define RT_FUSION_KERNEL_DUMPFLAG (0x04U)
  181. #define RT_KERNEL_CUSTOM_AICPU (0x08U)
  182. // STARS topic scheduler sqe : topic_type
  183. #define RT_KERNEL_DEVICE_FIRST (0x10U)
  184. #define RT_KERNEL_HOST_ONLY (0x20U)
  185. #define RT_KERNEL_HOST_FIRST (0x40U)
  186. /**
  187. * @ingroup rt_kernel
  188. * @brief kernel mode
  189. **/
  190. #define RT_DEFAULT_KERNEL_MODE (0x00U)
  191. #define RT_NORMAL_KERNEL_MODE (0x01U)
  192. #define RT_ALL_KERNEL_MODE (0x02U)
  193. /**
  194. * @ingroup rt_kernel
  195. * @brief SHAPE kernel type
  196. **/
  197. #define RT_STATIC_SHAPE_KERNEL (0x00U)
  198. #define RT_DYNAMIC_SHAPE_KERNEL (0x01U)
  199. /**
  200. * @ingroup rt_kernel
  201. * @brief kernel L1 Fusion Dump bit flags
  202. */
  203. #define RT_DDR_ADDR (0x0U)
  204. /**
  205. * @ingroup rt_kernel
  206. * @brief register device binary
  207. * @param [in] bin device binary description
  208. * @param [out] hdl device binary handle
  209. * @return RT_ERROR_NONE for ok
  210. * @return RT_ERROR_INVALID_VALUE for error input
  211. */
  212. RTS_API rtError_t rtDevBinaryRegister(const rtDevBinary_t *bin, void **hdl);
  213. /**
  214. * @ingroup rt_kernel
  215. * @brief register device binary with all kernel
  216. * @param [in] bin device binary description
  217. * @param [out] hdl device binary handle
  218. * @return RT_ERROR_NONE for ok
  219. * @return RT_ERROR_INVALID_VALUE for error input
  220. */
  221. RTS_API rtError_t rtRegisterAllKernel(const rtDevBinary_t *bin, void **hdl);
  222. /**
  223. * @ingroup rt_kernel
  224. * @brief register fast memeory device binary
  225. * @param [in] hdl device binary handle
  226. * @return RT_ERROR_NONE for ok
  227. * @return RT_ERROR_INVALID_VALUE for error input
  228. */
  229. RTS_API rtError_t rtBinaryRegisterToFastMemory(void *hdl);
  230. /**
  231. * @ingroup rt_kernel
  232. * @brief unregister device binary
  233. * @param [in] hdl device binary handle
  234. * @return RT_ERROR_NONE for ok
  235. * @return RT_ERROR_INVALID_VALUE for error input
  236. */
  237. RTS_API rtError_t rtDevBinaryUnRegister(void *hdl);
  238. /**
  239. * @ingroup rt_kernel
  240. * @brief register device binary metadata
  241. * @param [in] hdl device binary description
  242. * @param [in] metadata device binary metadata
  243. * @return RT_ERROR_NONE for ok
  244. * @return RT_ERROR_INVALID_VALUE for error input
  245. */
  246. RTS_API rtError_t rtMetadataRegister(void *hdl, const char_t *metadata);
  247. /**
  248. * @ingroup rt_kernel
  249. * @brief register device binary dependency
  250. * @param [in] mHandle master device binary description
  251. * @param [in] sHandle slave device binary description
  252. * @return RT_ERROR_NONE for ok
  253. * @return RT_ERROR_INVALID_VALUE for error input
  254. */
  255. RTS_API rtError_t rtDependencyRegister(void *mHandle, void *sHandle);
  256. /**
  257. * @ingroup rt_kernel
  258. * @brief register device function
  259. * @param [in] binHandle device binary handle
  260. * @param [in] stubFunc stub function
  261. * @param [in] stubName stub function name
  262. * @param [in] kernelInfoExt kernel Info extension. device function description or tiling key,
  263. * depending static shape or dynmaic shape.
  264. * @return RT_ERROR_NONE for ok
  265. * @return RT_ERROR_INVALID_VALUE for error input
  266. */
  267. RTS_API rtError_t rtFunctionRegister(void *binHandle, const void *stubFunc, const char_t *stubName,
  268. const void *kernelInfoExt, uint32_t funcMode);
  269. /**
  270. * @ingroup rt_kernel
  271. * @brief find stub function by name
  272. * @param [in] stubName stub function name
  273. * @param [out] stubFunc stub function
  274. * @return RT_ERROR_NONE for ok
  275. * @return RT_ERROR_INVALID_VALUE for error input
  276. */
  277. RTS_API rtError_t rtGetFunctionByName(const char_t *stubName, void **stubFunc);
  278. /**
  279. * @ingroup rt_kernel
  280. * @brief find addr by stub func
  281. * @param [in] stubFunc stub function
  282. * @param [out] addr
  283. * @return RT_ERROR_NONE for ok
  284. * @return RT_ERROR_INVALID_VALUE for error input
  285. */
  286. RTS_API rtError_t rtGetAddrByFun(const void *stubFunc, void **addr);
  287. /**
  288. * @ingroup rt_kernel
  289. * @brief query registered or not by stubName
  290. * @param [in] stubName stub function name
  291. * @return RT_ERROR_NONE for ok
  292. * @return RT_ERROR_INVALID_VALUE for error input
  293. */
  294. RTS_API rtError_t rtQueryFunctionRegistered(const char_t *stubName);
  295. /**
  296. * @ingroup rt_kernel
  297. * @brief config data dump
  298. * @param [in] dumpSizePerBlock dump size
  299. * @param [in] blockDim block dimentions
  300. * @param [in] dumpBaseAddr dump base address
  301. * @return RT_ERROR_NONE for ok
  302. * @return RT_ERROR_INVALID_VALUE for error input
  303. */
  304. RTS_API rtError_t rtKernelConfigDump(uint32_t kind, uint32_t dumpSizePerBlock, uint32_t blockDim, void **dumpBaseAddr,
  305. rtStream_t stm);
  306. /**
  307. * @ingroup rt_kernel
  308. * @brief get kernel address and prefetchCnt
  309. * @param [in] hdl program for dynamic shape
  310. * @param [in] tilingKey tilingKey for dynamic shape
  311. * @param [in] stubFunc stubFunc for static shape
  312. * @param [in] flag flag for distinguishing between dynamic shape and static shape
  313. * @param [out] addr address of kernel function
  314. * @param [out] prefetchCnt prefetchCnt of kernel function
  315. * @return RT_ERROR_NONE for ok
  316. * @return RT_ERROR_INVALID_VALUE for error input
  317. */
  318. RTS_API rtError_t rtKernelGetAddrAndPrefCnt(void *hdl, const uint64_t tilingKey, const void * const stubFunc,
  319. const uint32_t flag, void **addr, uint32_t *prefetchCnt);
  320. /**
  321. * @ingroup rt_kernel
  322. * @brief launch kernel to device
  323. * @param [in] stubFunc stub function
  324. * @param [in] blockDim block dimentions
  325. * @param [in] args argments address for kernel function
  326. * @param [in] argsSize argements size
  327. * @param [in] smDesc shared memory description
  328. * @param [in] stm associated stream
  329. * @return RT_ERROR_NONE for ok
  330. * @return RT_ERROR_INVALID_VALUE for error input
  331. */
  332. RTS_API rtError_t rtKernelLaunch(const void *stubFunc, uint32_t blockDim, void *args, uint32_t argsSize,
  333. rtSmDesc_t *smDesc, rtStream_t stm);
  334. /**
  335. * @ingroup rt_kernel
  336. * @brief launch kernel with handle to device
  337. * @param [in] hdl program
  338. * @param [in] kernelInfoExt kernel Info extension. device function description or tiling key,
  339. * depending static shape or dynmaic shape.
  340. * @param [in] blockDim block dimentions
  341. * @param [in] args argments address for kernel function
  342. * @param [in] argsSize argements size
  343. * @param [in] smDesc shared memory description
  344. * @param [in] stm associated stream
  345. * @param [in] kernelInfo kernel info
  346. * @return RT_ERROR_NONE for ok
  347. * @return RT_ERROR_INVALID_VALUE for error input
  348. */
  349. RTS_API rtError_t rtKernelLaunchWithHandle(void *hdl, const void *kernelInfoExt, uint32_t blockDim,
  350. void *args, uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stream_,
  351. const void *kernelInfo);
  352. /**
  353. * @ingroup rt_kernel
  354. * @brief launch kernel to device
  355. * @param [in] stubFunc stub function
  356. * @param [in] blockDim block dimentions
  357. * @param [in] args argments address for kernel function
  358. * @param [in] argsSize argements size
  359. * @param [in] smDesc shared memory description
  360. * @param [in] stm associated stream
  361. * @param [in] flag dump flag
  362. * @return RT_ERROR_NONE for ok
  363. * @return RT_ERROR_INVALID_VALUE for error input
  364. */
  365. RTS_API rtError_t rtKernelLaunchWithFlag(const void *stubFunc, uint32_t blockDim, void *args, uint32_t argsSize,
  366. rtSmDesc_t *smDesc, rtStream_t stm, uint32_t flags);
  367. /**
  368. * @ingroup rt_kernel(abandoned)
  369. * @brief launch kernel to device
  370. * @param [in] args argments address for kernel function
  371. * @param [in] argsSize argements size
  372. * @param [in] flags launch flags
  373. * @param [in] stm associated stream
  374. * @return RT_ERROR_NONE for ok
  375. * @return RT_ERROR_INVALID_VALUE for error input
  376. */
  377. RTS_API rtError_t rtKernelLaunchEx(void *args, uint32_t argsSize, uint32_t flags, rtStream_t stm);
  378. /**
  379. * @ingroup rt_kernel(in use)
  380. * @brief launch kernel to device
  381. * @param [in] opName opkernel name
  382. * @param [in] args argments address for kernel function
  383. * @param [in] argsSize argements size
  384. * @param [in] flags launch flags
  385. * @param [in] stm associated stream
  386. * @return RT_ERROR_NONE for ok
  387. * @return RT_ERROR_INVALID_VALUE for error input
  388. */
  389. RTS_API rtError_t rtKernelLaunchFwk(const char_t *opName, void *args, uint32_t argsSize, uint32_t flags,
  390. rtStream_t rtStream);
  391. /**
  392. * @ingroup rt_kernel(abandoned)
  393. * @brief launch cpu kernel to device
  394. * @param [in] soName so name
  395. * @param [in] kernelName kernel name
  396. * @param [in] blockDim block dimentions
  397. * @param [in] args argments address for kernel function
  398. * @param [in] argsSize argments size
  399. * @param [in] smDesc shared memory description
  400. * @param [in] stm associated stream
  401. * @return RT_ERROR_NONE for ok
  402. * @return RT_ERROR_INVALID_VALUE for error input
  403. */
  404. RTS_API rtError_t rtCpuKernelLaunch(const void *soName, const void *kernelName, uint32_t blockDim, const void *args,
  405. uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stm);
  406. /**
  407. * @ingroup rt_kernel(in use)
  408. * @brief launch cpu kernel to device
  409. * @param [in] launchNames names for kernel launch
  410. * @param [in] blockDim block dimentions
  411. * @param [in] args argments address for kernel function
  412. * @param [in] argsSize argments size
  413. * @param [in] smDesc shared memory description
  414. * @param [in] stm associated stream
  415. * @return RT_ERROR_NONE for ok
  416. * @return RT_ERROR_INVALID_VALUE for error input
  417. */
  418. RTS_API rtError_t rtAicpuKernelLaunch(const rtKernelLaunchNames_t *launchNames,
  419. uint32_t blockDim, const void *args, uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stm);
  420. /**
  421. * @ingroup rt_kernel(abandoned)
  422. * @brief launch cpu kernel to device with dump identifier
  423. * @param [in] soName so name
  424. * @param [in] kernelName kernel name
  425. * @param [in] blockDim block dimentions
  426. * @param [in] args argments address for kernel function
  427. * @param [in] argsSize argments size
  428. * @param [in] smDesc shared memory description
  429. * @param [in] stm associated stream
  430. * @param [in] flag dump flag or others function flag
  431. * @return RT_ERROR_NONE for ok
  432. * @return RT_ERROR_INVALID_VALUE for error input
  433. */
  434. RTS_API rtError_t rtCpuKernelLaunchWithFlag(const void *soName, const void *kernelName, uint32_t blockDim,
  435. const void *args, uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stm,
  436. uint32_t flags);
  437. /**
  438. * @ingroup rt_kernel(in use)
  439. * @brief launch cpu kernel to device with dump identifier
  440. * @param [in] launchNames names for kernel launch
  441. * @param [in] blockDim block dimentions
  442. * @param [in] args argments address for kernel function
  443. * @param [in] argsSize argments size
  444. * @param [in] smDesc shared memory description
  445. * @param [in] stm associated stream
  446. * @param [in] flag dump flag or others function flag
  447. * @return RT_ERROR_NONE for ok
  448. * @return RT_ERROR_INVALID_VALUE for error input
  449. */
  450. RTS_API rtError_t rtAicpuKernelLaunchWithFlag(const rtKernelLaunchNames_t *launchNames, uint32_t blockDim,
  451. const void *args, uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stm, uint32_t flags);
  452. /**
  453. * @ingroup rt_kernel
  454. * @brief L1 fusion dump addr transfered to device
  455. * @param [in] mdl handle info
  456. * @param [in] addr ddr address of L1 Fusion Dump
  457. * @param [in] dumpSize memory size
  458. * @param [in] flag memory flag
  459. * @return RT_ERROR_NONE for ok
  460. * @return RT_ERROR_INVALID_VALUE for error input
  461. */
  462. RTS_API rtError_t rtDumpAddrSet(rtModel_t mdl, void *addr, uint32_t dumpSize, uint32_t flag);
  463. /**
  464. * @ingroup rt_kernel
  465. * @brief load dump info to aicpu
  466. * @param [in] dumpInfo dump info
  467. * @param [in] length length of dump info
  468. * @return RT_ERROR_NONE for ok
  469. * @return RT_ERROR_INVALID_VALUE for error input
  470. */
  471. RTS_API rtError_t rtDatadumpInfoLoad(const void *dumpInfo, uint32_t length);
  472. /**
  473. * @ingroup rt_kernel
  474. * @brief launch npu get float status task
  475. * @param [in] outputAddr pointer to op output addr
  476. * @param [in] outputSize op output size
  477. * @param [in] checkMode check mode
  478. * @param [in] stm associated stream
  479. * @return RT_ERROR_NONE for ok
  480. * @return RT_ERROR_INVALID_VALUE for error input
  481. */
  482. RTS_API rtError_t rtNpuGetFloatStatus(void *outputAddr, uint64_t outputSize, uint32_t checkMode, rtStream_t stm);
  483. /**
  484. * @ingroup rt_kernel
  485. * @brief launch npu clear float status task
  486. * @param [in] checkMode check mode
  487. * @param [in] stm associated stream
  488. * @return RT_ERROR_NONE for ok
  489. * @return RT_ERROR_INVALID_VALUE for error input
  490. */
  491. RTS_API rtError_t rtNpuClearFloatStatus(uint32_t checkMode, rtStream_t stm);
  492. #ifndef __CLANG_CCE_RUNTIME_H__
  493. #define __CLANG_CCE_RUNTIME_H__
  494. /**
  495. * @ingroup rt_kernel
  496. * @brief configure call argment for next rtLaunch in current thread
  497. * @param [in] numBlocks block dimentions
  498. * @param [in] smDesc shared memory description
  499. * @param [in] stm associated stream
  500. * @return RT_ERROR_NONE for ok
  501. * @return RT_ERROR_INVALID_VALUE for error input
  502. */
  503. #ifdef __cplusplus
  504. RTS_API rtError_t rtConfigureCall(uint32_t numBlocks, rtSmDesc_t *smDesc = nullptr, rtStream_t stm = nullptr);
  505. #else
  506. RTS_API rtError_t rtConfigureCall(uint32_t numBlocks, rtSmDesc_t *smDesc, rtStream_t stm);
  507. #endif
  508. #endif // __CLANG_CCE_RUNTIME_H__
  509. /**
  510. * @ingroup rt_kernel
  511. * @brief setup argment for next rtLaunch in current thread
  512. * @param [in] args argment address for kernel function
  513. * @param [in] size argment size
  514. * @param [in] offset argment table offset
  515. * @return RT_ERROR_NONE for ok
  516. * @return RT_ERROR_INVALID_VALUE for error input
  517. */
  518. RTS_API rtError_t rtSetupArgument(const void *args, uint32_t size, uint32_t offset);
  519. /**
  520. * @ingroup rt_kernel
  521. * @brief launch kernel to device with previous setting kernel argment
  522. * and call argment
  523. * @param [in] stubFunc stub function
  524. * @return RT_ERROR_NONE for ok
  525. * @return RT_ERROR_INVALID_VALUE for error input
  526. */
  527. RTS_API rtError_t rtLaunch(const void *stubFunc);
  528. /**
  529. * @ingroup rt_kernel
  530. * @brief implicitly transfered data to device.
  531. * lifecycle end after next kernel task finish
  532. * @param [in] ptr host memory
  533. * @param [in] size host memory size
  534. * @param [in] flag reserved. set to 0
  535. * @param [out] args returned arg. used for next kernel's arg.
  536. * @return RT_ERROR_NONE for ok
  537. * @return RT_ERROR_INVALID_VALUE for error input
  538. */
  539. RTS_API rtError_t rtKernelConfigTransArg(const void *ptr, uint64_t size, uint32_t flag, void **args);
  540. /**
  541. * @ingroup rt_kernel
  542. * @brief start fusion kernels.
  543. * @param [in] stm stream for fusion kernels
  544. * @return RT_ERROR_NONE for ok
  545. * @return RT_ERROR_INVALID_VALUE for error input
  546. */
  547. RTS_API rtError_t rtKernelFusionStart(rtStream_t stm);
  548. /**
  549. * @ingroup rt_kernel
  550. * @brief end fusion kernels.
  551. * @param [in] stm stream for fusion kernels
  552. * @return RT_ERROR_NONE for ok
  553. * @return RT_ERROR_INVALID_VALUE for error input
  554. */
  555. RTS_API rtError_t rtKernelFusionEnd(rtStream_t stm);
  556. /**
  557. * @ingroup rt_kernel
  558. * @brief set kernelinfo callback
  559. * @param [in] callback
  560. * @return RT_ERROR_NONE for ok
  561. * @return RT_ERROR_INVALID_VALUE for error input
  562. */
  563. RTS_API rtError_t rtSetKernelReportCallback(rtKernelReportCallback callBack);
  564. /**
  565. * @ingroup rt_kernel
  566. * @brief subscribe stream callback report.
  567. * @param [in] threadId thread id for stream
  568. * @param [in] stm stream for subscribe
  569. * @return RT_ERROR_NONE for ok
  570. * @return RT_ERROR_INVALID_VALUE for error input
  571. */
  572. RTS_API rtError_t rtSubscribeReport(uint64_t threadId, rtStream_t stm);
  573. /**
  574. * @ingroup rt_kernel
  575. * @brief add callback launch task in stream.
  576. * @param [in] callBackFunc app callback function
  577. * @param [in] fnData user data
  578. * @param [in] stm subscribed stream
  579. * @return RT_ERROR_NONE for ok
  580. * @return RT_ERROR_INVALID_VALUE for error input
  581. */
  582. RTS_API rtError_t rtCallbackLaunch(rtCallback_t callBackFunc, void *fnData, rtStream_t stm, bool isBlock);
  583. /**
  584. * @ingroup rt_kernel
  585. * @brief process callback report.
  586. * @param [in] timeout if timeout=-1, while(1); else timeout
  587. * @return RT_ERROR_NONE for ok
  588. * @return RT_ERROR_INVALID_VALUE for error input
  589. */
  590. RTS_API rtError_t rtProcessReport(int32_t timeout);
  591. /**
  592. * @ingroup rt_kernel
  593. * @brief unsubscribe callback report.
  594. * @param [in] threadId thread id for stream
  595. * @param [in] stm stream for subscribe
  596. * @return RT_ERROR_NONE for ok
  597. * @return RT_ERROR_INVALID_VALUE for error input
  598. */
  599. RTS_API rtError_t rtUnSubscribeReport(uint64_t threadId, rtStream_t stm);
  600. /**
  601. * @ingroup profiling_base
  602. * @brief start online prof.
  603. * @return RT_ERROR_NONE for ok
  604. * @return RT_ERROR_INVALID_VALUE for error input
  605. */
  606. RTS_API rtError_t rtStartOnlineProf(rtStream_t stm, uint32_t sampleNum);
  607. /**
  608. * @ingroup profiling_base
  609. * @brief stop online prof.
  610. * @return RT_ERROR_NONE for ok
  611. * @return RT_ERROR_INVALID_VALUE for error input
  612. */
  613. RTS_API rtError_t rtStopOnlineProf(rtStream_t stm);
  614. /**
  615. * @ingroup profiling_base
  616. * @brief get online prof.
  617. * @return RT_ERROR_NONE for ok
  618. * @return RT_ERROR_INVALID_VALUE for error input
  619. */
  620. RTS_API rtError_t rtGetOnlineProfData(rtStream_t stm, rtProfDataInfo_t *pProfData, uint32_t profDataNum);
  621. /**
  622. * @ingroup profiling_base
  623. * @brief start mdc profiler.
  624. * @return RT_ERROR_NONE for ok
  625. * @return RT_ERROR_INVALID_VALUE for error input
  626. */
  627. RTS_API rtError_t rtStartMDCProfiler(void **addr, uint32_t length);
  628. /**
  629. * @ingroup profiling_base
  630. * @brief stop mdc profiler.
  631. * @return RT_ERROR_NONE for ok
  632. * @return RT_ERROR_INVALID_VALUE for error input
  633. */
  634. RTS_API rtError_t rtStopMDCProfiler(void *addr);
  635. /**
  636. * @ingroup rt_kernel
  637. * @brief launch kernel with tiling data to device
  638. * @param [in] stubFunc stub function
  639. * @param [in] blockDim block dimentions
  640. * @param [in] argsInfo argments info address for kernel function
  641. * @param [in] smDesc shared memory description
  642. * @param [in] stm associated stream
  643. * @return RT_ERROR_NONE for ok
  644. * @return RT_ERROR_INVALID_VALUE for error input
  645. */
  646. RTS_API rtError_t rtKernelLaunchWithTiling(const void *stubFunc, uint32_t blockDim,
  647. rtArgsWithTiling_t *argsInfo, rtSmDesc_t *smDesc, rtStream_t stm);
  648. /**
  649. * @ingroup rt_kernel
  650. * @brief launch kernel with handle and tiling data to device
  651. * @param [in] hdl program
  652. * @param [in] kernelInfoExt kernel Info extension. device function description or tiling key,
  653. * depending static shape or dynmaic shape.
  654. * @param [in] blockDim block dimentions
  655. * @param [in] argsInfo argments info address for kernel function
  656. * @param [in] smDesc shared memory description
  657. * @param [in] stm associated stream
  658. * @param [in] kernelInfo kernel info
  659. * @return RT_ERROR_NONE for ok
  660. * @return RT_ERROR_INVALID_VALUE for error input
  661. */
  662. RTS_API rtError_t rtKernelLaunchWithHandleAndTiling(void *hdl, const void *kernelInfoExt, uint32_t blockDim,
  663. rtArgsWithTiling_t *argsInfo, rtSmDesc_t *smDesc, rtStream_t stm, const void* kernelInfo);
  664. #if defined(__cplusplus)
  665. }
  666. #endif
  667. #endif // CCE_RUNTIME_KERNEL_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