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 13 kB

5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409
  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. #ifdef __cplusplus
  21. extern "C" {
  22. #endif // __cplusplus
  23. /*lint -e148*/
  24. /**
  25. * @ingroup rt_kernel
  26. * @brief shared memory data control
  27. */
  28. typedef struct tagRtSmData {
  29. uint64_t L2_mirror_addr; // preload or swap source address
  30. uint32_t L2_data_section_size; // every data size
  31. uint8_t L2_preload; // 1 - preload from mirrorAddr, 0 - no preload
  32. uint8_t modified; // 1 - data will be modified by kernel, 0 - no modified
  33. uint8_t priority; // data priority
  34. int8_t prev_L2_page_offset_base; // remap source section offset
  35. uint8_t L2_page_offset_base; // remap destination section offset
  36. uint8_t L2_load_to_ddr; // 1 - need load out, 0 - no need
  37. uint8_t reserved[2]; // reserved
  38. } rtSmData_t;
  39. /*lint -e148*/
  40. /**
  41. * @ingroup rt_kernel
  42. * @brief shared memory description
  43. */
  44. typedef struct tagRtSmCtrl {
  45. rtSmData_t data[8]; // data description
  46. uint64_t size; // max page Num
  47. uint8_t remap[64]; /* just using for static remap mode, default:0xFF
  48. * array index: virtual l2 page id, array value: physic l2 page id */
  49. uint8_t l2_in_main; // 0-DDR, 1-L2, default:0xF
  50. uint8_t reserved[3];
  51. } rtSmDesc_t;
  52. typedef rtSmDesc_t rtL2Ctrl_t;
  53. /**
  54. * @ingroup rt_kernel
  55. * @brief device binary type
  56. */
  57. typedef struct tagRtDevBinary {
  58. uint32_t magic; /**< magic number */
  59. uint32_t version; /**< version of binary */
  60. const void *data; /**< binary data */
  61. uint64_t length; /**< binary length */
  62. } rtDevBinary_t;
  63. /**
  64. * @ingroup rt_kernel
  65. * @brief function mode type
  66. */
  67. typedef enum {
  68. FUNC_MODE_NORMAL = 0,
  69. FUNC_MODE_PCTRACE_USERPROFILE_RECORDLOOP,
  70. FUNC_MODE_PCTRACE_USERPROFILE_SKIPLOOP,
  71. FUNC_MODE_PCTRACE_CYCLECNT_RECORDLOOP,
  72. FUNC_MODE_PCTRACE_CYCLECNT_SKIPLOOP,
  73. FUNC_MODE_BUTT
  74. } rtFuncModeType_t;
  75. /**
  76. * @ingroup rt_kernel
  77. * @brief kernel info
  78. */
  79. typedef struct rtKernelInfo {
  80. uint64_t task_offset; // kernel offset in module
  81. /* flowtable */
  82. void *arg; // launch kernel arg
  83. uint32_t arg_size;
  84. /* module */
  85. void *module_addr; // module::baseaddr_
  86. uint32_t module_size;
  87. } * rtKernelInfo_t;
  88. /**
  89. * @ingroup rt_KernelConfigDump
  90. * @brief device dump type
  91. */
  92. typedef enum tagRtDumpKind {
  93. RT_DATA_DUMP_KIND_INVALID = -1,
  94. RT_DATA_DUMP_KIND_DUMP = 0,
  95. RT_DATA_DUMP_KIND_RESERVED
  96. } rtDumpKind_t;
  97. /**
  98. * @ingroup rt_kernel
  99. * @brief report callback
  100. */
  101. typedef rtError_t (*rtKernelReportCallback)(rtStream_t stream, rtKernelInfo_t kernelInfo);
  102. /**
  103. * @ingroup rt_kernel
  104. * @brief magic number of plain binary for aicore
  105. */
  106. #define RT_DEV_BINARY_MAGIC_PLAIN 0xabceed50
  107. /**
  108. * @ingroup rt_kernel
  109. * @brief magic number of plain binary for aicpu
  110. */
  111. #define RT_DEV_BINARY_MAGIC_PLAIN_AICPU 0xabceed51
  112. /**
  113. * @ingroup rt_kernel
  114. * @brief magic number of plain binary for aivector
  115. */
  116. #define RT_DEV_BINARY_MAGIC_PLAIN_AIVEC 0xabceed52
  117. /**
  118. * @ingroup rt_kernel
  119. * @brief magic number of elf binary for aicore
  120. */
  121. #define RT_DEV_BINARY_MAGIC_ELF 0x43554245
  122. /**
  123. * @ingroup rt_kernel
  124. * @brief magic number of elf binary for aicpu
  125. */
  126. #define RT_DEV_BINARY_MAGIC_ELF_AICPU 0x41415243
  127. /**
  128. * @ingroup rt_kernel
  129. * @brief magic number of elf binary for aivector
  130. */
  131. #define RT_DEV_BINARY_MAGIC_ELF_AIVEC 0x41415246
  132. /**
  133. * @ingroup rt_kernel_flags
  134. * @brief kernel op bit flags
  135. */
  136. #define RT_KERNEL_DEFAULT (0x00)
  137. #define RT_KERNEL_CONVERT (0x01)
  138. #define RT_KERNEL_DUMPFLAG (0x02)
  139. /**
  140. * @ingroup rt_kernel
  141. * @brief register device binary
  142. * @param [in] bin device binary description
  143. * @param [out] handle device binary handle
  144. * @return RT_ERROR_NONE for ok
  145. * @note:if this interface is changed, pls notify the compiler changing at the same time.
  146. */
  147. RTS_API rtError_t rtDevBinaryRegister(const rtDevBinary_t *bin, void **handle);
  148. /**
  149. * @ingroup rt_kernel
  150. * @brief register fast memeory device binary
  151. * @param [in] handle device binary handle
  152. * @return RT_ERROR_NONE for ok
  153. * @note:if this interface is changed, pls notify the compiler changing at the same time.
  154. */
  155. RTS_API rtError_t rtBinaryRegisterToFastMemory(void *handle);
  156. /**
  157. * @ingroup rt_kernel
  158. * @brief unregister device binary
  159. * @param [in] handle device binary handle
  160. * @return RT_ERROR_NONE for ok
  161. * @note:if this interface is changed, pls notify the compiler changing at the same time.
  162. */
  163. RTS_API rtError_t rtDevBinaryUnRegister(void *handle);
  164. /**
  165. * @ingroup rt_kernel
  166. * @brief register device binary metadata
  167. * @param [in] handle device binary description
  168. * @param [in] metadata device binary metadata
  169. * @return RT_ERROR_NONE for ok
  170. * @note:if this interface is changed, pls notify the compiler changing at the same time.
  171. */
  172. RTS_API rtError_t rtMetadataRegister(void *handle, const char *metadata);
  173. /**
  174. * @ingroup rt_kernel
  175. * @brief register device binary dependency
  176. * @param [in] mHandle master device binary description
  177. * @param [in] sHandle slave device binary description
  178. * @return RT_ERROR_NONE for ok
  179. * @note:if this interface is changed, pls notify the compiler changing at the same time.
  180. */
  181. RTS_API rtError_t rtDependencyRegister(void *mHandle, void *sHandle);
  182. /**
  183. * @ingroup rt_kernel
  184. * @brief register device function
  185. * @param [in] binHandle device binary handle
  186. * @param [in] stubFunc stub function
  187. * @param [in] stubName stub function name
  188. * @param [in] devFunc device function description. symbol name or address
  189. * offset, depending binary type.
  190. * @return RT_ERROR_NONE for ok
  191. * @note:if this interface is changed, pls notify the compiler changing at the same time.
  192. */
  193. RTS_API rtError_t rtFunctionRegister(void *binHandle, const void *stubFunc, const char *stubName, const void *devFunc,
  194. uint32_t funcMode);
  195. /**
  196. * @ingroup rt_kernel
  197. * @brief find stub function by name
  198. * @param [in] stubName stub function name
  199. * @param [out] stubFunc stub function
  200. * @return RT_ERROR_NONE for ok
  201. */
  202. RTS_API rtError_t rtGetFunctionByName(const char *stubName, void **stubFunc);
  203. /**
  204. * @ingroup rt_kernel
  205. * @brief query registered or not by stubName
  206. * @param [in] stubName stub function name
  207. * @return RT_ERROR_NONE for ok
  208. */
  209. RTS_API rtError_t rtQueryFunctionRegistered(const char *stubName);
  210. /**
  211. * @ingroup rt_kernel
  212. * @brief config data dump
  213. * @param [in] dumpSizePerBlock dump size
  214. * @param [in] blockDim block dimentions
  215. * @param [in] dumpBaseAddr dump base address
  216. * @return RT_ERROR_NONE for ok, errno for failed
  217. */
  218. RTS_API rtError_t rtKernelConfigDump(uint32_t kind, uint32_t dumpSizePerBlock, uint32_t blockDim, void **dumpBaseAddr,
  219. rtStream_t stream_);
  220. /**
  221. * @ingroup rt_kernel
  222. * @brief launch kernel to device
  223. * @param [in] stubFunc stub function
  224. * @param [in] blockDim block dimentions
  225. * @param [in] args argments address for kernel function
  226. * @param [in] argsSize argements size
  227. * @param [in] smDesc shared memory description
  228. * @param [in] stream associated stream
  229. * @return RT_ERROR_NONE for ok, errno for failed
  230. */
  231. RTS_API rtError_t rtKernelLaunch(const void *stubFunc, uint32_t blockDim, void *args, uint32_t argsSize,
  232. rtSmDesc_t *smDesc, rtStream_t stream);
  233. /**
  234. * @ingroup rt_kernel
  235. * @brief launch kernel to device
  236. * @param [in] stubFunc stub function
  237. * @param [in] blockDim block dimentions
  238. * @param [in] args argments address for kernel function
  239. * @param [in] argsSize argements size
  240. * @param [in] smDesc shared memory description
  241. * @param [in] stream associated stream
  242. * @param [in] flag dump flag
  243. * @return RT_ERROR_NONE for ok, errno for failed
  244. */
  245. RTS_API rtError_t rtKernelLaunchWithFlag(const void *stubFunc, uint32_t blockDim, void *args, uint32_t argsSize,
  246. rtSmDesc_t *smDesc, rtStream_t stream, uint32_t flags);
  247. /**
  248. * @ingroup rt_kernel
  249. * @brief launch kernel to device
  250. * @param [in] args argments address for kernel function
  251. * @param [in] argsSize argements size
  252. * @param [in] flags launch flags
  253. * @param [in] stream associated stream
  254. * @return RT_ERROR_NONE for ok, errno for failed
  255. */
  256. RTS_API rtError_t rtKernelLaunchEx(void *args, uint32_t argsSize, uint32_t flags, rtStream_t stream);
  257. /**
  258. * @ingroup rt_kernel
  259. * @brief launch cpu kernel to device
  260. * @param [in] soName so name
  261. * @param [in] kernelName kernel name
  262. * @param [in] blockDim block dimentions
  263. * @param [in] args argments address for kernel function
  264. * @param [in] argsSize argments size
  265. * @param [in] smDesc shared memory description
  266. * @param [in] stream associated stream
  267. * @retval RT_ERROR_NONE for ok, errno for failed
  268. */
  269. RTS_API rtError_t rtCpuKernelLaunch(const void *soName, const void *kernelName, uint32_t blockDim, const void *args,
  270. uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stream);
  271. /**
  272. * @ingroup rt_kernel
  273. * @brief launch cpu kernel to device with dump identifier
  274. * @param [in] soName so name
  275. * @param [in] kernelName kernel name
  276. * @param [in] blockDim block dimentions
  277. * @param [in] args argments address for kernel function
  278. * @param [in] argsSize argments size
  279. * @param [in] smDesc shared memory description
  280. * @param [in] stream associated stream
  281. * @param [in] flag dump flag or others function flag
  282. * @retval RT_ERROR_NONE for ok, errno for failed
  283. */
  284. RTS_API rtError_t rtCpuKernelLaunchWithFlag(const void *soName, const void *kernelName, uint32_t blockDim,
  285. const void *args, uint32_t argsSize, rtSmDesc_t *smDesc,
  286. rtStream_t stream, uint32_t flags);
  287. /**
  288. * @ingroup rt_kernel
  289. * @brief load dump info to aicpu
  290. * @param [in] dumpInfo dump info
  291. * @param [in] length length of dump info
  292. * @return RT_ERROR_NONE for ok, errno for failed
  293. */
  294. RTS_API rtError_t rtDatadumpInfoLoad(const void *dumpInfo, uint32_t length);
  295. #ifndef __CLANG_CCE_RUNTIME_H__
  296. #define __CLANG_CCE_RUNTIME_H__
  297. /**
  298. * @ingroup rt_kernel
  299. * @brief configure call argment for next rtLaunch in current thread
  300. * @param [in] numBlocks block dimentions
  301. * @param [in] smDesc shared memory description
  302. * @param [in] stream associated stream
  303. * @return RT_ERROR_NONE for ok
  304. * @note:if this interface is changed, pls notify the compiler changing at the same time.
  305. */
  306. #ifdef __cplusplus
  307. RTS_API rtError_t rtConfigureCall(uint32_t numBlocks, rtSmDesc_t *smDesc = nullptr, rtStream_t stream = nullptr);
  308. #else
  309. RTS_API rtError_t rtConfigureCall(uint32_t numBlocks, rtSmDesc_t *smDesc, rtStream_t stream);
  310. #endif
  311. #endif // __CLANG_CCE_RUNTIME_H__
  312. /**
  313. * @ingroup rt_kernel
  314. * @brief setup argment for next rtLaunch in current thread
  315. * @param [in] arg argment address for kernel function
  316. * @param [in] size argment size
  317. * @param [in] offset argment table offset
  318. * @return RT_ERROR_NONE for ok
  319. * @note:if this interface is changed, pls notify the compiler changing at the same time.
  320. */
  321. RTS_API rtError_t rtSetupArgument(const void *arg, uint32_t size, uint32_t offset);
  322. /**
  323. * @ingroup rt_kernel
  324. * @brief launch kernel to device with previous setting kernel argment
  325. * and call argment
  326. * @param [in] stubFunc stub function
  327. * @return RT_ERROR_NONE for ok
  328. * @note:if this interface is changed, pls notify the compiler changing at the same time.
  329. */
  330. RTS_API rtError_t rtLaunch(const void *stubFunc);
  331. /**
  332. * @ingroup rt_kernel
  333. * @brief implicitly transfered data to device.
  334. * lifecycle end after next kernel task finish
  335. * @param [in] ptr host memory
  336. * @param [in] size host memory size
  337. * @param [in] flag reserved. set to 0
  338. * @param [out] arg returned arg. used for next kernel's arg.
  339. * @return RT_ERROR_NONE for ok
  340. */
  341. RTS_API rtError_t rtKernelConfigTransArg(const void *ptr, uint64_t size, uint32_t flag, void **arg);
  342. /**
  343. * @ingroup rt_kernel
  344. * @brief start fusion kernels.
  345. * @param [in] stream stream for fusion kernels
  346. * @return RT_ERROR_NONE for ok, errno for failed
  347. */
  348. RTS_API rtError_t rtKernelFusionStart(rtStream_t stream);
  349. /**
  350. * @ingroup rt_kernel
  351. * @brief end fusion kernels.
  352. * @param [in] stream stream for fusion kernels
  353. * @return RT_ERROR_NONE for ok, errno for failed
  354. */
  355. RTS_API rtError_t rtKernelFusionEnd(rtStream_t stream);
  356. /**
  357. * @ingroup rt_kernel
  358. * @brief set kernelinfo callback
  359. * @param [in] callback
  360. * @return RT_ERROR_NONE for ok, errno for failed
  361. */
  362. RTS_API rtError_t rtSetKernelReportCallback(rtKernelReportCallback callBack);
  363. #ifdef __cplusplus
  364. }
  365. #endif
  366. #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