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 23 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
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
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701
  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_KERNEL_H
  17. #define CCE_RUNTIME_KERNEL_H
  18. #include "base.h"
  19. #include "stream.h"
  20. #if defined(__cplusplus)
  21. extern "C" {
  22. #endif
  23. /**
  24. * @ingroup rt_kernel
  25. * @brief shared memory data control
  26. */
  27. typedef struct tagRtSmData {
  28. uint64_t L2_mirror_addr; // preload or swap source address
  29. uint32_t L2_data_section_size; // every data size
  30. uint8_t L2_preload; // 1 - preload from mirrorAddr, 0 - no preload
  31. uint8_t modified; // 1 - data will be modified by kernel, 0 - no modified
  32. uint8_t priority; // data priority
  33. int8_t prev_L2_page_offset_base; // remap source section offset
  34. uint8_t L2_page_offset_base; // remap destination section offset
  35. uint8_t L2_load_to_ddr; // 1 - need load out, 0 - no need
  36. uint8_t reserved[2]; // reserved
  37. } rtSmData_t;
  38. /**
  39. * @ingroup rt_kernel
  40. * @brief shared memory description
  41. */
  42. typedef struct tagRtSmCtrl {
  43. rtSmData_t data[8]; // data description
  44. uint64_t size; // max page Num
  45. uint8_t remap[64]; /* just using for static remap mode, default:0xFF
  46. array index: virtual l2 page id, array value: physic l2 page id */
  47. uint8_t l2_in_main; // 0-DDR, 1-L2, default:0xFF
  48. uint8_t reserved[3];
  49. } rtSmDesc_t;
  50. typedef rtSmDesc_t rtL2Ctrl_t;
  51. /**
  52. * @ingroup rt_kernel
  53. * @brief device binary type
  54. */
  55. typedef struct tagRtDevBinary {
  56. uint32_t magic; // magic number
  57. uint32_t version; // version of binary
  58. const void *data; // binary data
  59. uint64_t length; // binary length
  60. } rtDevBinary_t;
  61. /**
  62. * @ingroup rt_kernel
  63. * @brief function mode type
  64. */
  65. #define ONLINE_PROF_MAX_PMU_NUM (8)
  66. typedef struct ProfilefDataInfo {
  67. const void *stubFunc;
  68. uint32_t blockDim;
  69. const void *args;
  70. uint32_t argsSize;
  71. rtSmDesc_t *smDesc;
  72. rtStream_t stream;
  73. uint64_t totalcycle;
  74. uint64_t ovcycle;
  75. uint64_t pmu_cnt[ONLINE_PROF_MAX_PMU_NUM];
  76. } rtProfDataInfo_t;
  77. /**
  78. * @ingroup rt_kernel
  79. * @brief function mode type
  80. */
  81. typedef enum {
  82. FUNC_MODE_NORMAL = 0,
  83. FUNC_MODE_PCTRACE_USERPROFILE_RECORDLOOP,
  84. FUNC_MODE_PCTRACE_USERPROFILE_SKIPLOOP,
  85. FUNC_MODE_PCTRACE_CYCLECNT_RECORDLOOP,
  86. FUNC_MODE_PCTRACE_CYCLECNT_SKIPLOOP,
  87. FUNC_MODE_BUTT
  88. } rtFuncModeType_t;
  89. /**
  90. * @ingroup rt_kernel
  91. * @brief kernel info
  92. */
  93. typedef struct rtKernelInfo {
  94. uint64_t task_offset; // kernel offset in module
  95. /* flowtable */
  96. void *arg; // launch kernel arg
  97. uint32_t arg_size;
  98. /* module */
  99. void *module_addr; // module::baseaddr_
  100. uint32_t module_size;
  101. } *rtKernelInfo_t;
  102. /**
  103. * @ingroup rt_kernel
  104. * @brief op name
  105. */
  106. typedef struct rtKernelLaunchNames {
  107. const char_t *soName; // defined for so name
  108. const char_t *kernelName; // defined for kernel type name
  109. const char_t *opName; // defined for operator name
  110. } rtKernelLaunchNames_t;
  111. /**
  112. * @ingroup rt_kernel
  113. * @brief args struct
  114. */
  115. typedef struct tagRtArgsWithTiling {
  116. void *args; // args host mem addr
  117. uint32_t argsSize; // input + output + tiling addr size + tiling data size
  118. uint32_t argsSizeWithoutTiling; // input + output + tiling addr size
  119. uint16_t tilingAddrOffset; // tiling addr offset
  120. uint16_t tilingDataOffset; // tiling data offset
  121. uint16_t hostInputAddrOffset; // index of host_memory input in inputs_addrs list
  122. uint16_t hostInputDataOffset; // host_mem input data offset
  123. bool hasHostMemInput; // has host_memory input data in args or not: ture or false
  124. uint8_t reserved[7];
  125. } rtArgsWithTiling_t;
  126. /**
  127. * @ingroup rt_KernelConfigDump
  128. * @brief device dump type
  129. */
  130. typedef enum tagRtDumpKind {
  131. RT_DATA_DUMP_KIND_INVALID = -1,
  132. RT_DATA_DUMP_KIND_DUMP = 0,
  133. RT_DATA_DUMP_KIND_RESERVED = 1,
  134. } rtDumpKind_t;
  135. /**
  136. * @ingroup rt_kernel
  137. * @brief report callback
  138. */
  139. typedef rtError_t (*rtKernelReportCallback)(rtStream_t stream, rtKernelInfo_t kernelInfo);
  140. /**
  141. * @ingroup rt_kernel
  142. * @brief stream report callback
  143. */
  144. typedef void (*rtCallback_t)(void *fnData);
  145. /**
  146. * @ingroup rt_kernel
  147. * @brief magic number of plain binary for aicore
  148. */
  149. #define RT_DEV_BINARY_MAGIC_PLAIN 0xabceed50U
  150. /**
  151. * @ingroup rt_kernel
  152. * @brief magic number of plain binary for aicpu
  153. */
  154. #define RT_DEV_BINARY_MAGIC_PLAIN_AICPU 0xabceed51U
  155. /**
  156. * @ingroup rt_kernel
  157. * @brief magic number of plain binary for aivector
  158. */
  159. #define RT_DEV_BINARY_MAGIC_PLAIN_AIVEC 0xabceed52U
  160. /**
  161. * @ingroup rt_kernel
  162. * @brief magic number of elf binary for aicore
  163. */
  164. #define RT_DEV_BINARY_MAGIC_ELF 0x43554245U
  165. /**
  166. * @ingroup rt_kernel
  167. * @brief magic number of elf binary for aicpu
  168. */
  169. #define RT_DEV_BINARY_MAGIC_ELF_AICPU 0x41415243U
  170. /**
  171. * @ingroup rt_kernel
  172. * @brief magic number of elf binary for aivector
  173. */
  174. #define RT_DEV_BINARY_MAGIC_ELF_AIVEC 0x41415246U
  175. /**
  176. * @ingroup rt_kernel
  177. * @brief magic number of elf binary for aicube
  178. */
  179. #define RT_DEV_BINARY_MAGIC_ELF_AICUBE 0x41494343U
  180. /**
  181. * @ingroup rt_kernel_flags
  182. * @brief kernel op bit flags
  183. */
  184. #define RT_KERNEL_DEFAULT (0x00U)
  185. #define RT_KERNEL_CONVERT (0x01U)
  186. #define RT_KERNEL_DUMPFLAG (0x02U)
  187. #define RT_FUSION_KERNEL_DUMPFLAG (0x04U)
  188. #define RT_KERNEL_CUSTOM_AICPU (0x08U)
  189. // STARS topic scheduler sqe : topic_type
  190. #define RT_KERNEL_DEVICE_FIRST (0x10U)
  191. #define RT_KERNEL_HOST_ONLY (0x20U)
  192. #define RT_KERNEL_HOST_FIRST (0x40U)
  193. /**
  194. * @ingroup rt_kernel
  195. * @brief kernel mode
  196. **/
  197. #define RT_DEFAULT_KERNEL_MODE (0x00U)
  198. #define RT_NORMAL_KERNEL_MODE (0x01U)
  199. #define RT_ALL_KERNEL_MODE (0x02U)
  200. /**
  201. * @ingroup rt_kernel
  202. * @brief kernel L1 Fusion Dump bit flags
  203. */
  204. #define RT_DDR_ADDR (0x0U)
  205. /**
  206. * @ingroup rt_kernel
  207. * @brief register device binary
  208. * @param [in] bin device binary description
  209. * @param [out] handle device binary handle
  210. * @return RT_ERROR_NONE for ok
  211. * @return RT_ERROR_INVALID_VALUE for error input
  212. */
  213. RTS_API rtError_t rtDevBinaryRegister(const rtDevBinary_t *bin, void **handle);
  214. /**
  215. * @ingroup rt_kernel
  216. * @brief register device binary with all kernel
  217. * @param [in] bin device binary description
  218. * @param [out] handle device binary handle
  219. * @return RT_ERROR_NONE for ok
  220. * @return RT_ERROR_INVALID_VALUE for error input
  221. */
  222. RTS_API rtError_t rtRegisterAllKernel(const rtDevBinary_t *bin, void **handle);
  223. /**
  224. * @ingroup rt_kernel
  225. * @brief register fast memeory device binary
  226. * @param [in] handle device binary handle
  227. * @return RT_ERROR_NONE for ok
  228. * @return RT_ERROR_INVALID_VALUE for error input
  229. */
  230. RTS_API rtError_t rtBinaryRegisterToFastMemory(void *handle);
  231. /**
  232. * @ingroup rt_kernel
  233. * @brief unregister device binary
  234. * @param [in] handle device binary handle
  235. * @return RT_ERROR_NONE for ok
  236. * @return RT_ERROR_INVALID_VALUE for error input
  237. */
  238. RTS_API rtError_t rtDevBinaryUnRegister(void *handle);
  239. /**
  240. * @ingroup rt_kernel
  241. * @brief register device binary metadata
  242. * @param [in] handle device binary description
  243. * @param [in] metadata device binary metadata
  244. * @return RT_ERROR_NONE for ok
  245. * @return RT_ERROR_INVALID_VALUE for error input
  246. */
  247. RTS_API rtError_t rtMetadataRegister(void *handle, const char_t *metadata);
  248. /**
  249. * @ingroup rt_kernel
  250. * @brief register device binary dependency
  251. * @param [in] mHandle master device binary description
  252. * @param [in] sHandle slave device binary description
  253. * @return RT_ERROR_NONE for ok
  254. * @return RT_ERROR_INVALID_VALUE for error input
  255. */
  256. RTS_API rtError_t rtDependencyRegister(void *mHandle, void *sHandle);
  257. /**
  258. * @ingroup rt_kernel
  259. * @brief register device function
  260. * @param [in] binHandle device binary handle
  261. * @param [in] stubFunc stub function
  262. * @param [in] stubName stub function name
  263. * @param [in] devFunc device function description. symbol name or address
  264. * offset, depending binary type.
  265. * @return RT_ERROR_NONE for ok
  266. * @return RT_ERROR_INVALID_VALUE for error input
  267. */
  268. RTS_API rtError_t rtFunctionRegister(void *binHandle, const void *stubFunc, const char_t *stubName, const void *devFunc,
  269. uint32_t funcMode);
  270. /**
  271. * @ingroup rt_kernel
  272. * @brief find stub function by name
  273. * @param [in] stubName stub function name
  274. * @param [out] stubFunc stub function
  275. * @return RT_ERROR_NONE for ok
  276. * @return RT_ERROR_INVALID_VALUE for error input
  277. */
  278. RTS_API rtError_t rtGetFunctionByName(const char_t *stubName, void **stubFunc);
  279. /**
  280. * @ingroup rt_kernel
  281. * @brief find addr by stub func
  282. * @param [in] stubFunc stub function
  283. * @param [out] addr
  284. * @return RT_ERROR_NONE for ok
  285. * @return RT_ERROR_INVALID_VALUE for error input
  286. */
  287. RTS_API rtError_t rtGetAddrByFun(const void *stubFunc, void **addr);
  288. /**
  289. * @ingroup rt_kernel
  290. * @brief query registered or not by stubName
  291. * @param [in] stubName stub function name
  292. * @return RT_ERROR_NONE for ok
  293. * @return RT_ERROR_INVALID_VALUE for error input
  294. */
  295. RTS_API rtError_t rtQueryFunctionRegistered(const char_t *stubName);
  296. /**
  297. * @ingroup rt_kernel
  298. * @brief config data dump
  299. * @param [in] dumpSizePerBlock dump size
  300. * @param [in] blockDim block dimentions
  301. * @param [in] dumpBaseAddr dump base address
  302. * @return RT_ERROR_NONE for ok
  303. * @return RT_ERROR_INVALID_VALUE for error input
  304. */
  305. RTS_API rtError_t rtKernelConfigDump(uint32_t kind, uint32_t dumpSizePerBlock, uint32_t blockDim, void **dumpBaseAddr,
  306. rtStream_t stream);
  307. /**
  308. * @ingroup rt_kernel
  309. * @brief launch kernel to device
  310. * @param [in] stubFunc stub function
  311. * @param [in] blockDim block dimentions
  312. * @param [in] args argments address for kernel function
  313. * @param [in] argsSize argements size
  314. * @param [in] smDesc shared memory description
  315. * @param [in] stream associated stream
  316. * @return RT_ERROR_NONE for ok
  317. * @return RT_ERROR_INVALID_VALUE for error input
  318. */
  319. RTS_API rtError_t rtKernelLaunch(const void *stubFunc, uint32_t blockDim, void *args, uint32_t argsSize,
  320. rtSmDesc_t *smDesc, rtStream_t stream);
  321. /**
  322. * @ingroup rt_kernel
  323. * @brief launch kernel with handle to device
  324. * @param [in] handle program
  325. * @param [in] devFunc device function description.
  326. * @param [in] blockDim block dimentions
  327. * @param [in] args argments address for kernel function
  328. * @param [in] argsSize argements size
  329. * @param [in] smDesc shared memory description
  330. * @param [in] stream associated stream
  331. * @param [in] kernelInfo kernel info
  332. * @return RT_ERROR_NONE for ok
  333. * @return RT_ERROR_INVALID_VALUE for error input
  334. */
  335. RTS_API rtError_t rtKernelLaunchWithHandle(void *handle, const void *devFunc, uint32_t blockDim, void *args, uint32_t argsSize,
  336. rtSmDesc_t *smDesc, rtStream_t stream_, const void *kernelInfo);
  337. /**
  338. * @ingroup rt_kernel
  339. * @brief launch kernel to device
  340. * @param [in] stubFunc stub function
  341. * @param [in] blockDim block dimentions
  342. * @param [in] args argments address for kernel function
  343. * @param [in] argsSize argements size
  344. * @param [in] smDesc shared memory description
  345. * @param [in] stream associated stream
  346. * @param [in] flag dump flag
  347. * @return RT_ERROR_NONE for ok
  348. * @return RT_ERROR_INVALID_VALUE for error input
  349. */
  350. RTS_API rtError_t rtKernelLaunchWithFlag(const void *stubFunc, uint32_t blockDim, void *args, uint32_t argsSize,
  351. rtSmDesc_t *smDesc, rtStream_t stream, uint32_t flags);
  352. /**
  353. * @ingroup rt_kernel(abandoned)
  354. * @brief launch kernel to device
  355. * @param [in] args argments address for kernel function
  356. * @param [in] argsSize argements size
  357. * @param [in] flags launch flags
  358. * @param [in] stream associated stream
  359. * @return RT_ERROR_NONE for ok
  360. * @return RT_ERROR_INVALID_VALUE for error input
  361. */
  362. RTS_API rtError_t rtKernelLaunchEx(void *args, uint32_t argsSize, uint32_t flags, rtStream_t stream);
  363. /**
  364. * @ingroup rt_kernel(in use)
  365. * @brief launch kernel to device
  366. * @param [in] opName opkernel name
  367. * @param [in] args argments address for kernel function
  368. * @param [in] argsSize argements size
  369. * @param [in] flags launch flags
  370. * @param [in] stream associated stream
  371. * @return RT_ERROR_NONE for ok
  372. * @return RT_ERROR_INVALID_VALUE for error input
  373. */
  374. RTS_API rtError_t rtKernelLaunchFwk(const char_t *opName, void *args, uint32_t argsSize, uint32_t flags,
  375. rtStream_t rtStream);
  376. /**
  377. * @ingroup rt_kernel(abandoned)
  378. * @brief launch cpu kernel to device
  379. * @param [in] soName so name
  380. * @param [in] kernelName kernel name
  381. * @param [in] blockDim block dimentions
  382. * @param [in] args argments address for kernel function
  383. * @param [in] argsSize argments size
  384. * @param [in] smDesc shared memory description
  385. * @param [in] stream associated stream
  386. * @return RT_ERROR_NONE for ok
  387. * @return RT_ERROR_INVALID_VALUE for error input
  388. */
  389. RTS_API rtError_t rtCpuKernelLaunch(const void *soName, const void *kernelName, uint32_t blockDim, const void *args,
  390. uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stream);
  391. /**
  392. * @ingroup rt_kernel(in use)
  393. * @brief launch cpu kernel to device
  394. * @param [in] launchNames names for kernel launch
  395. * @param [in] blockDim block dimentions
  396. * @param [in] args argments address for kernel function
  397. * @param [in] argsSize argments size
  398. * @param [in] smDesc shared memory description
  399. * @param [in] stream associated stream
  400. * @return RT_ERROR_NONE for ok
  401. * @return RT_ERROR_INVALID_VALUE for error input
  402. */
  403. RTS_API rtError_t rtAicpuKernelLaunch(const rtKernelLaunchNames_t *launchNames,
  404. uint32_t blockDim, const void *args, uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stream);
  405. /**
  406. * @ingroup rt_kernel(abandoned)
  407. * @brief launch cpu kernel to device with dump identifier
  408. * @param [in] soName so name
  409. * @param [in] kernelName kernel name
  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] stream associated stream
  415. * @param [in] flag dump flag or others function flag
  416. * @return RT_ERROR_NONE for ok
  417. * @return RT_ERROR_INVALID_VALUE for error input
  418. */
  419. RTS_API rtError_t rtCpuKernelLaunchWithFlag(const void *soName, const void *kernelName, uint32_t blockDim,
  420. const void *args, uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stream,
  421. uint32_t flags);
  422. /**
  423. * @ingroup rt_kernel(in use)
  424. * @brief launch cpu kernel to device with dump identifier
  425. * @param [in] launchNames names for kernel launch
  426. * @param [in] blockDim block dimentions
  427. * @param [in] args argments address for kernel function
  428. * @param [in] argsSize argments size
  429. * @param [in] smDesc shared memory description
  430. * @param [in] stream associated stream
  431. * @param [in] flag dump flag or others function flag
  432. * @return RT_ERROR_NONE for ok
  433. * @return RT_ERROR_INVALID_VALUE for error input
  434. */
  435. RTS_API rtError_t rtAicpuKernelLaunchWithFlag(const rtKernelLaunchNames_t *launchNames, uint32_t blockDim,
  436. const void *args, uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stream, uint32_t flags);
  437. /**
  438. * @ingroup rt_kernel
  439. * @brief L1 fusion dump addr transfered to device
  440. * @param [in] model handle info
  441. * @param [in] addr ddr address of L1 Fusion Dump
  442. * @param [in] dumpSize memory size
  443. * @param [in] flag memory flag
  444. * @return RT_ERROR_NONE for ok
  445. * @return RT_ERROR_INVALID_VALUE for error input
  446. */
  447. RTS_API rtError_t rtDumpAddrSet(rtModel_t model, void *addr, uint32_t dumpSize, uint32_t flag);
  448. /**
  449. * @ingroup rt_kernel
  450. * @brief load dump info to aicpu
  451. * @param [in] dumpInfo dump info
  452. * @param [in] length length of dump info
  453. * @return RT_ERROR_NONE for ok
  454. * @return RT_ERROR_INVALID_VALUE for error input
  455. */
  456. RTS_API rtError_t rtDatadumpInfoLoad(const void *dumpInfo, uint32_t length);
  457. #ifndef __CLANG_CCE_RUNTIME_H__
  458. #define __CLANG_CCE_RUNTIME_H__
  459. /**
  460. * @ingroup rt_kernel
  461. * @brief configure call argment for next rtLaunch in current thread
  462. * @param [in] numBlocks block dimentions
  463. * @param [in] smDesc shared memory description
  464. * @param [in] stream associated stream
  465. * @return RT_ERROR_NONE for ok
  466. * @return RT_ERROR_INVALID_VALUE for error input
  467. */
  468. #ifdef __cplusplus
  469. RTS_API rtError_t rtConfigureCall(uint32_t numBlocks, rtSmDesc_t *smDesc = nullptr, rtStream_t stream = nullptr);
  470. #else
  471. RTS_API rtError_t rtConfigureCall(uint32_t numBlocks, rtSmDesc_t *smDesc, rtStream_t stream);
  472. #endif
  473. #endif // __CLANG_CCE_RUNTIME_H__
  474. /**
  475. * @ingroup rt_kernel
  476. * @brief setup argment for next rtLaunch in current thread
  477. * @param [in] arg argment address for kernel function
  478. * @param [in] size argment size
  479. * @param [in] offset argment table offset
  480. * @return RT_ERROR_NONE for ok
  481. * @return RT_ERROR_INVALID_VALUE for error input
  482. */
  483. RTS_API rtError_t rtSetupArgument(const void *arg, uint32_t size, uint32_t offset);
  484. /**
  485. * @ingroup rt_kernel
  486. * @brief launch kernel to device with previous setting kernel argment
  487. * and call argment
  488. * @param [in] stubFunc stub function
  489. * @return RT_ERROR_NONE for ok
  490. * @return RT_ERROR_INVALID_VALUE for error input
  491. */
  492. RTS_API rtError_t rtLaunch(const void *stubFunc);
  493. /**
  494. * @ingroup rt_kernel
  495. * @brief implicitly transfered data to device.
  496. * lifecycle end after next kernel task finish
  497. * @param [in] ptr host memory
  498. * @param [in] size host memory size
  499. * @param [in] flag reserved. set to 0
  500. * @param [out] arg returned arg. used for next kernel's arg.
  501. * @return RT_ERROR_NONE for ok
  502. * @return RT_ERROR_INVALID_VALUE for error input
  503. */
  504. RTS_API rtError_t rtKernelConfigTransArg(const void *ptr, uint64_t size, uint32_t flag, void **arg);
  505. /**
  506. * @ingroup rt_kernel
  507. * @brief start fusion kernels.
  508. * @param [in] stream stream for fusion kernels
  509. * @return RT_ERROR_NONE for ok
  510. * @return RT_ERROR_INVALID_VALUE for error input
  511. */
  512. RTS_API rtError_t rtKernelFusionStart(rtStream_t stream);
  513. /**
  514. * @ingroup rt_kernel
  515. * @brief end fusion kernels.
  516. * @param [in] stream stream for fusion kernels
  517. * @return RT_ERROR_NONE for ok
  518. * @return RT_ERROR_INVALID_VALUE for error input
  519. */
  520. RTS_API rtError_t rtKernelFusionEnd(rtStream_t stream);
  521. /**
  522. * @ingroup rt_kernel
  523. * @brief set kernelinfo callback
  524. * @param [in] callback
  525. * @return RT_ERROR_NONE for ok
  526. * @return RT_ERROR_INVALID_VALUE for error input
  527. */
  528. RTS_API rtError_t rtSetKernelReportCallback(rtKernelReportCallback callBack);
  529. /**
  530. * @ingroup rt_kernel
  531. * @brief subscribe stream callback report.
  532. * @param [in] threadId thread id for stream
  533. * @param [in] stream stream for subscribe
  534. * @return RT_ERROR_NONE for ok
  535. * @return RT_ERROR_INVALID_VALUE for error input
  536. */
  537. RTS_API rtError_t rtSubscribeReport(uint64_t threadId, rtStream_t stream);
  538. /**
  539. * @ingroup rt_kernel
  540. * @brief add callback launch task in stream.
  541. * @param [in] callBackFunc app callback function
  542. * @param [in] fnData user data
  543. * @param [in] stream subscribed stream
  544. * @return RT_ERROR_NONE for ok
  545. * @return RT_ERROR_INVALID_VALUE for error input
  546. */
  547. RTS_API rtError_t rtCallbackLaunch(rtCallback_t callBackFunc, void *fnData, rtStream_t stream, bool isBlock);
  548. /**
  549. * @ingroup rt_kernel
  550. * @brief process callback report.
  551. * @param [in] timeout if timeout=-1, while(1); else timeout
  552. * @return RT_ERROR_NONE for ok
  553. * @return RT_ERROR_INVALID_VALUE for error input
  554. */
  555. RTS_API rtError_t rtProcessReport(int32_t timeout);
  556. /**
  557. * @ingroup rt_kernel
  558. * @brief unsubscribe callback report.
  559. * @param [in] threadId thread id for stream
  560. * @param [in] stream stream for subscribe
  561. * @return RT_ERROR_NONE for ok
  562. * @return RT_ERROR_INVALID_VALUE for error input
  563. */
  564. RTS_API rtError_t rtUnSubscribeReport(uint64_t threadId, rtStream_t stream);
  565. /**
  566. * @ingroup profiling_base
  567. * @brief start online prof.
  568. * @return RT_ERROR_NONE for ok
  569. * @return RT_ERROR_INVALID_VALUE for error input
  570. */
  571. RTS_API rtError_t rtStartOnlineProf(rtStream_t stream, uint32_t sampleNum);
  572. /**
  573. * @ingroup profiling_base
  574. * @brief stop online prof.
  575. * @return RT_ERROR_NONE for ok
  576. * @return RT_ERROR_INVALID_VALUE for error input
  577. */
  578. RTS_API rtError_t rtStopOnlineProf(rtStream_t stream);
  579. /**
  580. * @ingroup profiling_base
  581. * @brief get online prof.
  582. * @return RT_ERROR_NONE for ok
  583. * @return RT_ERROR_INVALID_VALUE for error input
  584. */
  585. RTS_API rtError_t rtGetOnlineProfData(rtStream_t stream, rtProfDataInfo_t *pProfData, uint32_t profDataNum);
  586. /**
  587. * @ingroup profiling_base
  588. * @brief start mdc profiler.
  589. * @return RT_ERROR_NONE for ok
  590. * @return RT_ERROR_INVALID_VALUE for error input
  591. */
  592. RTS_API rtError_t rtStartMDCProfiler(void **addr, uint32_t length);
  593. /**
  594. * @ingroup profiling_base
  595. * @brief stop mdc profiler.
  596. * @return RT_ERROR_NONE for ok
  597. * @return RT_ERROR_INVALID_VALUE for error input
  598. */
  599. RTS_API rtError_t rtStopMDCProfiler(void *addr);
  600. /**
  601. * @ingroup rt_kernel
  602. * @brief launch kernel with tiling data to device
  603. * @param [in] stubFunc stub function
  604. * @param [in] blockDim block dimentions
  605. * @param [in] argsInfo argments info address for kernel function
  606. * @param [in] smDesc shared memory description
  607. * @param [in] stream associated stream
  608. * @return RT_ERROR_NONE for ok
  609. * @return RT_ERROR_INVALID_VALUE for error input
  610. */
  611. RTS_API rtError_t rtKernelLaunchWithTiling(const void *stubFunc, uint32_t blockDim,
  612. rtArgsWithTiling_t *argsInfo, rtSmDesc_t *smDesc, rtStream_t stream);
  613. /**
  614. * @ingroup rt_kernel
  615. * @brief launch kernel with handle and tiling data to device
  616. * @param [in] handle program
  617. * @param [in] devFunc device function description.
  618. * @param [in] blockDim block dimentions
  619. * @param [in] argsInfo argments info address for kernel function
  620. * @param [in] smDesc shared memory description
  621. * @param [in] stream associated stream
  622. * @param [in] kernelInfo kernel info
  623. * @return RT_ERROR_NONE for ok
  624. * @return RT_ERROR_INVALID_VALUE for error input
  625. */
  626. RTS_API rtError_t rtKernelLaunchWithHandleAndTiling(void *handle, const void *devFunc, uint32_t blockDim,
  627. rtArgsWithTiling_t *argsInfo, rtSmDesc_t *smDesc, rtStream_t stream, const void* kernelInfo);
  628. #if defined(__cplusplus)
  629. }
  630. #endif
  631. #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