shl_thead_rvv.h 49 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732
  1. /*
  2. * Copyright (C) 2016-2022 T-Head Semiconductor Co., Ltd. All rights reserved.
  3. *
  4. * SPDX-License-Identifier: Apache-2.0
  5. *
  6. * Licensed under the Apache License, Version 2.0 (the License); you may
  7. * not use this file except in compliance with the License.
  8. * You may obtain a copy of the License at
  9. *
  10. * www.apache.org/licenses/LICENSE-2.0
  11. *
  12. * Unless required by applicable law or agreed to in writing, software
  13. * distributed under the License is distributed on an AS IS BASIS, WITHOUT
  14. * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  15. * See the License for the specific language governing permissions and
  16. * limitations under the License.
  17. */
  18. /* CSI-NN2 version 2.0.x */
  19. #ifndef INCLUDE_SHL_RVV_H_
  20. #define INCLUDE_SHL_RVV_H_
  21. #if __riscv_vector
  22. #include <riscv_vector.h>
  23. #if (__riscv_v == 1000000)
  24. #define RVV_1_0_0
  25. #elif (__riscv_v == 7000)
  26. #define RVV_0_7_1
  27. #endif
  28. #ifdef __riscv_xtheadvdot
  29. #define XTHEADVDOT
  30. #define SHL_USE_DOT_INT8 // default: support int8 dot
  31. // #define SHL_USE_DOT_INT4 // easter eggs
  32. #endif // __riscv_xtheadvdot
  33. #endif // __riscv_vector
  34. #include "csi_nn.h"
  35. #include "shl_gref.h"
  36. #include "shl_ref.h"
  37. #ifdef __cplusplus
  38. extern "C" {
  39. #endif
  40. /********************************** initialization ******************************/
  41. int shl_rvv_conv2d_init_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  42. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  43. struct csinn_conv2d_params *params);
  44. int shl_rvv_conv2d_init_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  45. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  46. struct csinn_conv2d_params *params);
  47. int shl_rvv_depthwise_conv2d_init_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  48. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  49. struct csinn_conv2d_params *params);
  50. int shl_rvv_depthwise_conv2d_init_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  51. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  52. struct csinn_conv2d_params *params);
  53. int shl_rvv_depthwise_conv2d_init_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  54. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  55. struct csinn_conv2d_params *params);
  56. int shl_rvv_depthwise_conv2d_init_int4(struct csinn_tensor *input, struct csinn_tensor *output,
  57. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  58. struct csinn_conv2d_params *params);
  59. int shl_rvv_avgpool2d_init_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  60. struct csinn_pool_params *params);
  61. int shl_rvv_avgpool2d_init_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  62. struct csinn_pool_params *params);
  63. int shl_rvv_avgpool2d_init_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  64. struct csinn_pool_params *params);
  65. int shl_rvv_global_avgpool2d_init(struct csinn_tensor *input, struct csinn_tensor *output,
  66. struct csinn_pool_params *params);
  67. int shl_rvv_maxpool2d_init_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  68. struct csinn_pool_params *params);
  69. int shl_rvv_maxpool2d_init_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  70. struct csinn_pool_params *params);
  71. int shl_rvv_maxpool2d_init_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  72. struct csinn_pool_params *params);
  73. int shl_rvv_global_maxpool2d_init(struct csinn_tensor *input, struct csinn_tensor *output,
  74. struct csinn_pool_params *params);
  75. int shl_rvv_fullyconnected_init(struct csinn_tensor *input, struct csinn_tensor *output,
  76. struct csinn_tensor *weights, struct csinn_tensor *bias,
  77. struct csinn_fc_params *params);
  78. int shl_rvv_data_convert_init(struct csinn_tensor *input, struct csinn_tensor *output,
  79. struct csinn_siso_params *params);
  80. /************************************ convolution *********************************/
  81. /*********************************** im2col + gemm ********************************/
  82. void shl_rvv_conv_im2col_gemm_reorder_kernel_fp32(struct csinn_tensor *kernel,
  83. struct csinn_conv2d_params *params);
  84. void shl_rvv_conv_im2col_gemm_reorder_kernel_fp16(struct csinn_tensor *kernel,
  85. struct csinn_conv2d_params *params);
  86. int shl_rvv_conv_im2col_gemm_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  87. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  88. struct csinn_conv2d_params *params);
  89. int shl_rvv_conv_im2col_gemm_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  90. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  91. struct csinn_conv2d_params *params);
  92. void shl_rvv_conv_im2col_gemm_reorder_kernel_packn_fp32(struct csinn_tensor *kernel,
  93. struct csinn_conv2d_params *params);
  94. void shl_rvv_conv_im2col_gemm_reorder_kernel_packn_fp16(struct csinn_tensor *kernel,
  95. struct csinn_conv2d_params *params);
  96. int shl_rvv_conv_im2col_gemm_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  97. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  98. struct csinn_conv2d_params *params);
  99. int shl_rvv_conv_im2col_gemm_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  100. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  101. struct csinn_conv2d_params *params);
  102. void shl_rvv_conv_im2col_gemm_reorder_kernel_pack1ton_fp32(struct csinn_tensor *kernel,
  103. struct csinn_conv2d_params *params);
  104. void shl_rvv_conv_im2col_gemm_reorder_kernel_pack1ton_fp16(struct csinn_tensor *kernel,
  105. struct csinn_conv2d_params *params);
  106. int shl_rvv_conv_im2col_gemm_pack1ton_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  107. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  108. struct csinn_conv2d_params *params);
  109. int shl_rvv_conv_im2col_gemm_pack1ton_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  110. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  111. struct csinn_conv2d_params *params);
  112. void shl_rvv_conv_im2col_gemm_reorder_kernel_packnto1_fp32(struct csinn_tensor *kernel,
  113. struct csinn_conv2d_params *params);
  114. void shl_rvv_conv_im2col_gemm_reorder_kernel_packnto1_fp16(struct csinn_tensor *kernel,
  115. struct csinn_conv2d_params *params);
  116. int shl_rvv_conv_im2col_gemm_packnto1_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  117. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  118. struct csinn_conv2d_params *params);
  119. int shl_rvv_conv_im2col_gemm_packnto1_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  120. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  121. struct csinn_conv2d_params *params);
  122. /******************************** conv2d1x1s1 + gemm ******************************/
  123. void shl_rvv_conv1x1s1_gemm_reorder_kernel_fp32(struct csinn_tensor *kernel,
  124. struct csinn_conv2d_params *params);
  125. void shl_rvv_conv1x1s1_gemm_reorder_kernel_fp16(struct csinn_tensor *kernel,
  126. struct csinn_conv2d_params *params);
  127. int shl_rvv_conv1x1s1_gemm_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  128. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  129. struct csinn_conv2d_params *params);
  130. int shl_rvv_conv1x1s1_gemm_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  131. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  132. struct csinn_conv2d_params *params);
  133. void shl_rvv_conv1x1s1_gemm_reorder_kernel_packn_fp32(struct csinn_tensor *kernel,
  134. struct csinn_conv2d_params *params);
  135. void shl_rvv_conv1x1s1_gemm_reorder_kernel_packn_fp16(struct csinn_tensor *kernel,
  136. struct csinn_conv2d_params *params);
  137. int shl_rvv_conv1x1s1_gemm_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  138. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  139. struct csinn_conv2d_params *params);
  140. int shl_rvv_conv1x1s1_gemm_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  141. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  142. struct csinn_conv2d_params *params);
  143. void shl_rvv_conv1x1s1_gemm_reorder_kernel_pack1ton_fp32(struct csinn_tensor *kernel,
  144. struct csinn_conv2d_params *params);
  145. void shl_rvv_conv1x1s1_gemm_reorder_kernel_pack1ton_fp16(struct csinn_tensor *kernel,
  146. struct csinn_conv2d_params *params);
  147. int shl_rvv_conv1x1s1_gemm_pack1ton_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  148. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  149. struct csinn_conv2d_params *params);
  150. int shl_rvv_conv1x1s1_gemm_pack1ton_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  151. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  152. struct csinn_conv2d_params *params);
  153. void shl_rvv_conv1x1s1_gemm_reorder_kernel_packnto1_fp32(struct csinn_tensor *kernel,
  154. struct csinn_conv2d_params *params);
  155. void shl_rvv_conv1x1s1_gemm_reorder_kernel_packnto1_fp16(struct csinn_tensor *kernel,
  156. struct csinn_conv2d_params *params);
  157. int shl_rvv_conv1x1s1_gemm_packnto1_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  158. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  159. struct csinn_conv2d_params *params);
  160. int shl_rvv_conv1x1s1_gemm_packnto1_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  161. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  162. struct csinn_conv2d_params *params);
  163. /************************************* winograd ***********************************/
  164. void shl_rvv_wg_b6f3s1_trans_kernel_packn_fp32(struct csinn_tensor *src_kernel,
  165. struct csinn_tensor *dst_kernel);
  166. void shl_rvv_wg_b6f3s1_trans_kernel_packn_fp16(struct csinn_tensor *src_kernel,
  167. struct csinn_tensor *dst_kernel);
  168. int shl_rvv_wg_b6f3s1_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  169. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  170. struct csinn_conv2d_params *params);
  171. int shl_rvv_wg_b6f3s1_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  172. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  173. struct csinn_conv2d_params *params);
  174. void shl_rvv_wg_b4f3s1_trans_kernel_packn_fp32(struct csinn_tensor *src_kernel,
  175. struct csinn_tensor *dst_kernel);
  176. void shl_rvv_wg_b4f3s1_trans_kernel_packn_fp16(struct csinn_tensor *src_kernel,
  177. struct csinn_tensor *dst_kernel);
  178. void shl_rvv_wg_b4f3s1_trans_kernel_packn_int8(struct csinn_tensor *src_kernel,
  179. struct csinn_tensor *dst_kernel);
  180. int shl_rvv_wg_b4f3s1_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  181. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  182. struct csinn_conv2d_params *params);
  183. int shl_rvv_wg_b4f3s1_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  184. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  185. struct csinn_conv2d_params *params);
  186. int shl_rvv_wg_b4f3s1_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  187. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  188. struct csinn_conv2d_params *params);
  189. /******************************* depthwise convolution ****************************/
  190. int shl_rvv_dwconv3x3s1_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  191. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  192. struct csinn_conv2d_params *params);
  193. int shl_rvv_dwconv3x3s2_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  194. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  195. struct csinn_conv2d_params *params);
  196. int shl_rvv_dwconv3x3s1_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  197. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  198. struct csinn_conv2d_params *params);
  199. int shl_rvv_dwconv3x3s2_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  200. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  201. struct csinn_conv2d_params *params);
  202. int shl_rvv_dwconv3x3s1_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  203. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  204. struct csinn_conv2d_params *params);
  205. int shl_rvv_dwconv3x3s2_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  206. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  207. struct csinn_conv2d_params *params);
  208. int shl_rvv_dwconv3x3s1_int4(struct csinn_tensor *input, struct csinn_tensor *output,
  209. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  210. struct csinn_conv2d_params *params);
  211. int shl_rvv_dwconv3x3s2_int4(struct csinn_tensor *input, struct csinn_tensor *output,
  212. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  213. struct csinn_conv2d_params *params);
  214. void shl_rvv_dwconv_reorder_kernel_packn_fp32(struct csinn_tensor *kernel,
  215. struct csinn_conv2d_params *params);
  216. void shl_rvv_dwconv_reorder_kernel_packn_fp16(struct csinn_tensor *kernel,
  217. struct csinn_conv2d_params *params);
  218. void shl_rvv_dwconv_reorder_kernel_packn_int8(struct csinn_tensor *kernel,
  219. struct csinn_conv2d_params *params);
  220. int shl_rvv_dwconv3x3s1_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  221. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  222. struct csinn_conv2d_params *params);
  223. int shl_rvv_dwconv3x3s2_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  224. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  225. struct csinn_conv2d_params *params);
  226. int shl_rvv_dwconv3x3s1_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  227. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  228. struct csinn_conv2d_params *params);
  229. int shl_rvv_dwconv3x3s2_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  230. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  231. struct csinn_conv2d_params *params);
  232. int shl_rvv_dwconv3x3s1_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  233. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  234. struct csinn_conv2d_params *params);
  235. int shl_rvv_dwconv3x3s2_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  236. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  237. struct csinn_conv2d_params *params);
  238. int shl_rvv_dwconv_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  239. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  240. struct csinn_conv2d_params *params);
  241. int shl_rvv_dwconv_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  242. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  243. struct csinn_conv2d_params *params);
  244. int shl_rvv_dwconv_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  245. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  246. struct csinn_conv2d_params *params);
  247. /*************************************** gemm *************************************/
  248. void shl_rvv_reorder_kernel_n8_fp32(float *a, float *sa, int m, int k, int ldx);
  249. void shl_rvv_reorder_input_z8_fp32(float *b, float *sb, int k, int n, int ldx);
  250. void shl_rvv_gemm_8x8_fp32(float *dst, const float *sa, const float *sb, float *bias, int m, int k,
  251. int n, int ldc);
  252. void shl_rvv256_reorder_input_z16_fp32(float *b, float *sb, int k, int n, int ldx);
  253. void shl_rvv256_gemm_8x16_fp32(float *dst, const float *sa, const float *sb, float *bias, int m,
  254. int k, int n, int ldc);
  255. void shl_rvv_reorder_kernel_n8_fp16(__fp16 *a, __fp16 *sa, int m, int k, int ldx);
  256. void shl_rvv_reorder_input_z16_fp16(__fp16 *b, __fp16 *sb, int k, int n, int ldx);
  257. void shl_rvv_gemm_8x16_fp16(__fp16 *dst, const __fp16 *sa, const __fp16 *sb, __fp16 *bias, int m,
  258. int k, int n, int ldc);
  259. void shl_rvv256_reorder_kernel_n16_fp16(__fp16 *a, __fp16 *sa, int m, int k, int ldx);
  260. void shl_rvv256_reorder_input_z16_fp16(__fp16 *b, __fp16 *sb, int k, int n, int ldx);
  261. void shl_rvv256_gemm_16x16_fp16(__fp16 *dst, const __fp16 *sa, const __fp16 *sb, __fp16 *bias,
  262. int m, int k, int n, int ldc);
  263. void shl_rvv_reorder_kernel_n8_int8(int8_t *a, int8_t *sa, int m, int k, int ldx);
  264. void shl_rvv_reorder_input_z8_int8(int8_t *b, int8_t *sb, int k, int n, int ldx);
  265. void shl_rvv_gemm_8x8_int32(int32_t *dst, const int8_t *sa, const int8_t *sb, int32_t *bias, int m,
  266. int k, int n, int ldc);
  267. void shl_rvv_gemm_8x8_int8(int8_t *dst, const int8_t *sa, const int8_t *sb, int32_t *bias, int m,
  268. int k, int n, int ldc, int32_t out_zp, int32_t *mult, int32_t *shift);
  269. void shl_rvv256_reorder_input_z16_int8(int8_t *b, int8_t *sb, int k, int n, int ldx);
  270. void shl_rvv256_gemm_8x16_int32(int32_t *dst, const int8_t *sa, const int8_t *sb, int32_t *bias,
  271. int m, int k, int n, int ldc);
  272. void shl_rvv_reorder_input_n8_int4(int8_t *a, int8_t *sa, int m, int k, int ldx);
  273. void shl_rvv_reorder_kernel_n8_int4(int8_t *b, int8_t *sb, int n, int k, int ldx);
  274. void shl_rvv_gemm_8x8_int4(int8_t *dst, const int8_t *sa, const int8_t *sb, int m, int k, int n,
  275. int ldc, int32_t *bias, int32_t out_zp, int32_t *mult, int32_t *shift);
  276. /************************************ gemm ncxhwx *********************************/
  277. void shl_rvv_reorder_kernel_packn_fp32(float *a, float *sa, int m, int k, int ldx);
  278. void shl_rvv_reorder_input_z8_packn_fp32(float *b, float *sb, int k, int n, int ldx);
  279. void shl_rvv_ncxhwx_gemm_8xpack2n_fp32(float *dst, const float *sa, const float *sb, float *bias,
  280. int m, int k, int n, int ldc);
  281. void shl_rvv_reorder_input_z12_packn_fp32(float *b, float *sb, int k, int n, int ldx);
  282. void shl_rvv_ncxhwx_gemm_12xpack2n_fp32(float *dst, const float *sa, const float *sb, float *bias,
  283. int m, int k, int n, int ldc);
  284. void shl_rvv_reorder_kernel_packn_fp16(__fp16 *a, __fp16 *sa, int m, int k, int ldx);
  285. void shl_rvv_reorder_input_z8_packn_fp16(__fp16 *b, __fp16 *sb, int k, int n, int ldx);
  286. void shl_rvv_ncxhwx_gemm_8xpack2n_fp16(__fp16 *dst, const __fp16 *sa, const __fp16 *sb,
  287. __fp16 *bias, int m, int k, int n, int ldc);
  288. void shl_rvv_reorder_input_z12_packn_fp16(__fp16 *b, __fp16 *sb, int k, int n, int ldx);
  289. void shl_rvv_ncxhwx_gemm_12xpack2n_fp16(__fp16 *dst, const __fp16 *sa, const __fp16 *sb,
  290. __fp16 *bias, int m, int k, int n, int ldc);
  291. void shl_rvv_reorder_input_z8_packn_int8(int8_t *b, int8_t *sb, int k, int n, int ldx);
  292. void shl_rvv_ncxhwx_gemm_8xpackn_int8(int8_t *dst, const int8_t *sa, const int8_t *sb,
  293. int32_t *bias, int m, int k, int n, int ldc, int32_t out_zp,
  294. int32_t *mult, int32_t *shift);
  295. void shl_rvv_reorder_input_z12_packn_int8(int8_t *b, int8_t *sb, int k, int n, int ldx);
  296. void shl_rvv_ncxhwx_gemm_12xpackn_int8(int8_t *dst, const int8_t *sa, const int8_t *sb,
  297. int32_t *bias, int m, int k, int n, int ldc, int32_t out_zp,
  298. int32_t *mult, int32_t *shift);
  299. void shl_rvv_reorder_input_z8_packn_int4(int8_t *b, int8_t *sb, int k, int n, int ldx);
  300. void shl_rvv_ncxhwx_gemm_8xpackn_int4(int8_t *dst, const int8_t *sa, const int8_t *sb,
  301. int32_t *bias, int m, int k, int n, int ldc, int32_t out_zp,
  302. int32_t *mult, int32_t *shift);
  303. void shl_rvv_reorder_input_z12_packn_int4(int8_t *b, int8_t *sb, int k, int n, int ldx);
  304. void shl_rvv_ncxhwx_gemm_12xpackn_int4(int8_t *dst, const int8_t *sa, const int8_t *sb,
  305. int32_t *bias, int m, int k, int n, int ldc, int32_t out_zp,
  306. int32_t *mult, int32_t *shift);
  307. void shl_rvv_reorder_input_z12_pack1ton_fp32(float *b, float *sb, int inc, int maxk, int n,
  308. int ldx);
  309. void shl_rvv_reorder_input_z12_pack1ton_fp16(__fp16 *b, __fp16 *sb, int inc, int maxk, int n,
  310. int ldx);
  311. void shl_rvv_reorder_input_z12_pack1ton_int8(int8_t *b, int8_t *sb, int inc, int maxk, int n,
  312. int ldx);
  313. /************************************ pooling *********************************/
  314. int shl_rvv_avgpool2x2s2_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  315. struct csinn_pool_params *params);
  316. int shl_rvv_avgpool2x2s2_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  317. struct csinn_pool_params *params);
  318. int shl_rvv_avgpool2x2s2_p1_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  319. struct csinn_pool_params *params);
  320. int shl_rvv_avgpool2x2s2_p1_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  321. struct csinn_pool_params *params);
  322. int shl_rvv_avgpool3x3s2_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  323. struct csinn_pool_params *params);
  324. int shl_rvv_avgpool3x3s2_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  325. struct csinn_pool_params *params);
  326. int shl_rvv_avgpool3x3s2_p1_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  327. struct csinn_pool_params *params);
  328. int shl_rvv_avgpool3x3s2_p1_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  329. struct csinn_pool_params *params);
  330. int shl_rvv_avgpool3x3s1_p1_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  331. struct csinn_pool_params *params);
  332. int shl_rvv_avgpool3x3s1_p1_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  333. struct csinn_pool_params *params);
  334. int shl_rvv_maxpool2x2s2_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  335. struct csinn_pool_params *params);
  336. int shl_rvv_maxpool2x2s2_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  337. struct csinn_pool_params *params);
  338. int shl_rvv_maxpool2x2s2_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  339. struct csinn_pool_params *params);
  340. int shl_rvv_maxpool2x2s2_p1_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  341. struct csinn_pool_params *params);
  342. int shl_rvv_maxpool2x2s2_p1_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  343. struct csinn_pool_params *params);
  344. int shl_rvv_maxpool2x2s2_p1_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  345. struct csinn_pool_params *params);
  346. int shl_rvv_maxpool3x3s2_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  347. struct csinn_pool_params *params);
  348. int shl_rvv_maxpool3x3s2_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  349. struct csinn_pool_params *params);
  350. int shl_rvv_maxpool3x3s2_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  351. struct csinn_pool_params *params);
  352. int shl_rvv_maxpool3x3s2_p1_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  353. struct csinn_pool_params *params);
  354. int shl_rvv_maxpool3x3s2_p1_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  355. struct csinn_pool_params *params);
  356. int shl_rvv_maxpool3x3s2_p1_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  357. struct csinn_pool_params *params);
  358. int shl_rvv_maxpool3x3s1_p1_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  359. struct csinn_pool_params *params);
  360. int shl_rvv_maxpool3x3s1_p1_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  361. struct csinn_pool_params *params);
  362. int shl_rvv_maxpool3x3s1_p1_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  363. struct csinn_pool_params *params);
  364. int shl_rvv_global_avgpool2d_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  365. struct csinn_pool_params *params);
  366. int shl_rvv_global_avgpool2d_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  367. struct csinn_pool_params *params);
  368. int shl_rvv_global_maxpool2d_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  369. struct csinn_pool_params *params);
  370. int shl_rvv_global_maxpool2d_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  371. struct csinn_pool_params *params);
  372. int shl_rvv_maxpool2x2s2_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  373. struct csinn_pool_params *params);
  374. int shl_rvv_maxpool2x2s2_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  375. struct csinn_pool_params *params);
  376. int shl_rvv_maxpool3x3s2_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  377. struct csinn_pool_params *params);
  378. int shl_rvv_maxpool3x3s1_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  379. struct csinn_pool_params *params);
  380. int shl_rvv_maxpool3x3s2_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  381. struct csinn_pool_params *params);
  382. int shl_rvv_maxpool3x3s1_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  383. struct csinn_pool_params *params);
  384. int shl_rvv_maxpool2x2s2_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  385. struct csinn_pool_params *params);
  386. int shl_rvv_maxpool3x3s2_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  387. struct csinn_pool_params *params);
  388. int shl_rvv_maxpool3x3s1_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  389. struct csinn_pool_params *params);
  390. int shl_rvv_avgpool2x2s2_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  391. struct csinn_pool_params *params);
  392. int shl_rvv_avgpool2x2s2_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  393. struct csinn_pool_params *params);
  394. int shl_rvv_avgpool2x2s2_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  395. struct csinn_pool_params *params);
  396. int shl_rvv_avgpool3x3s2_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  397. struct csinn_pool_params *params);
  398. int shl_rvv_avgpool3x3s1_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  399. struct csinn_pool_params *params);
  400. int shl_rvv_avgpool3x3s2_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  401. struct csinn_pool_params *params);
  402. int shl_rvv_avgpool3x3s1_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  403. struct csinn_pool_params *params);
  404. int shl_rvv_avgpool3x3s2_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  405. struct csinn_pool_params *params);
  406. int shl_rvv_avgpool3x3s1_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  407. struct csinn_pool_params *params);
  408. int shl_rvv_avgpool_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  409. struct csinn_pool_params *params);
  410. int shl_rvv_avgpool_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  411. struct csinn_pool_params *params);
  412. int shl_rvv_avgpool_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  413. struct csinn_pool_params *params);
  414. int shl_rvv_global_maxpool2d_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  415. struct csinn_pool_params *params);
  416. int shl_rvv_global_maxpool2d_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  417. struct csinn_pool_params *params);
  418. int shl_rvv_global_maxpool2d_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  419. struct csinn_pool_params *params);
  420. int shl_rvv_global_avgpool2d_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  421. struct csinn_pool_params *params);
  422. int shl_rvv_global_avgpool2d_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  423. struct csinn_pool_params *params);
  424. int shl_rvv_global_avgpool2d_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  425. struct csinn_pool_params *params);
  426. int shl_rvv_maxpool_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  427. struct csinn_pool_params *params);
  428. int shl_rvv_maxpool_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  429. struct csinn_pool_params *params);
  430. int shl_rvv_maxpool_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  431. struct csinn_pool_params *params);
  432. /************************************ fullyconnected *********************************/
  433. void shl_rvv_fc_gemv_transform_weight_fp32(struct csinn_tensor *weights);
  434. void shl_rvv_fc_gemv_transform_weight_fp16(struct csinn_tensor *weights);
  435. void shl_rvv_fc_gemv_transform_weight_int8(struct csinn_tensor *weights);
  436. int shl_rvv_fullyconnected_packn_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  437. struct csinn_tensor *weights, struct csinn_tensor *bias,
  438. struct csinn_fc_params *params);
  439. int shl_rvv_fullyconnected_packn_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  440. struct csinn_tensor *weights, struct csinn_tensor *bias,
  441. struct csinn_fc_params *params);
  442. int shl_rvv_fullyconnected_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  443. struct csinn_tensor *weights, struct csinn_tensor *bias,
  444. struct csinn_fc_params *params);
  445. /************************************ activation *********************************/
  446. int shl_rvv_relu_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  447. struct csinn_relu_params *params);
  448. int shl_rvv_relu_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  449. struct csinn_relu_params *params);
  450. int shl_rvv_relu_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  451. struct csinn_relu_params *params);
  452. int shl_rvv_relu6_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  453. struct csinn_relu_params *params);
  454. int shl_rvv_relu6_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  455. struct csinn_relu_params *params);
  456. int shl_rvv_relu6_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  457. struct csinn_relu_params *params);
  458. int shl_rvv_leaky_relu_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  459. struct csinn_relu_params *params);
  460. int shl_rvv_leaky_relu_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  461. struct csinn_relu_params *params);
  462. int shl_rvv_leaky_relu_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  463. struct csinn_relu_params *params);
  464. int shl_rvv_sigmoid_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  465. struct csinn_sigmoid_params *params);
  466. int shl_rvv_softmax_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  467. struct csinn_softmax_params *params);
  468. int shl_rvv_prelu_fp32(struct csinn_tensor *input, struct csinn_tensor *alpha,
  469. struct csinn_tensor *output, struct csinn_prelu_params *params);
  470. int shl_rvv_prelu_fp16(struct csinn_tensor *input, struct csinn_tensor *alpha,
  471. struct csinn_tensor *output, struct csinn_prelu_params *params);
  472. int shl_rvv_prelu_int8(struct csinn_tensor *input, struct csinn_tensor *alpha,
  473. struct csinn_tensor *output, struct csinn_prelu_params *params);
  474. /************************************ layout/memory transform *********************************/
  475. int shl_rvv_concat_fp32(struct csinn_tensor **input, struct csinn_tensor *output,
  476. struct csinn_concat_params *params);
  477. int shl_rvv_concat_fp16(struct csinn_tensor **input, struct csinn_tensor *output,
  478. struct csinn_concat_params *params);
  479. int shl_rvv_concat_int8(struct csinn_tensor **input, struct csinn_tensor *output,
  480. struct csinn_concat_params *params);
  481. int shl_rvv_reshape_fp32(struct csinn_tensor *input, struct csinn_tensor *output,
  482. struct csinn_reshape_params *params);
  483. int shl_rvv_reshape_fp16(struct csinn_tensor *input, struct csinn_tensor *output,
  484. struct csinn_reshape_params *params);
  485. int shl_rvv_reshape_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  486. struct csinn_reshape_params *params);
  487. /************************************ basic math *********************************/
  488. int shl_rvv_add_fp32(struct csinn_tensor *input0, struct csinn_tensor *input1,
  489. struct csinn_tensor *output, struct csinn_diso_params *params);
  490. int shl_rvv_add_fp16(struct csinn_tensor *input0, struct csinn_tensor *input1,
  491. struct csinn_tensor *output, struct csinn_diso_params *params);
  492. int shl_rvv_add_int8(struct csinn_tensor *input0, struct csinn_tensor *input1,
  493. struct csinn_tensor *output, struct csinn_diso_params *params);
  494. int shl_rvv_mul_fp32(struct csinn_tensor *input0, struct csinn_tensor *input1,
  495. struct csinn_tensor *output, struct csinn_diso_params *params);
  496. int shl_rvv_mul_fp16(struct csinn_tensor *input0, struct csinn_tensor *input1,
  497. struct csinn_tensor *output, struct csinn_diso_params *params);
  498. int shl_rvv_mul_int8(struct csinn_tensor *input0, struct csinn_tensor *input1,
  499. struct csinn_tensor *output, struct csinn_diso_params *params);
  500. int shl_rvv_sum_stride_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  501. struct csinn_reduce_params *params);
  502. /************************************ utils *********************************/
  503. void shl_rvv_pad_input_fp32(const float *input, float *input_padded, int inc, int inh, int inw,
  504. int padded_h, int padded_w, int pad_top, int pad_left);
  505. void shl_rvv_pad_input_fp16(const __fp16 *input, __fp16 *input_padded, int inc, int inh, int inw,
  506. int padded_h, int padded_w, int pad_top, int pad_left);
  507. void shl_rvv_pad_input_int8(const int8_t *input, int8_t *input_padded, int inc, int inh, int inw,
  508. int padded_h, int padded_w, int pad_top, int pad_left,
  509. int8_t pad_value);
  510. void shl_rvv_pad_input_packn_fp32(const float *input, float *input_padded, int inc, int inh,
  511. int inw, int padded_h, int padded_w, int pad_top, int pad_left);
  512. void shl_rvv_pad_input_packn_fp16(const __fp16 *input, __fp16 *input_padded, int inc, int inh,
  513. int inw, int padded_h, int padded_w, int pad_top, int pad_left);
  514. void shl_rvv_pad_input_packn_int8(const int8_t *input, int8_t *input_padded, int inc, int inh,
  515. int inw, int padded_h, int padded_w, int pad_top, int pad_left,
  516. int8_t pad_value);
  517. void shl_rvv_pad_input_pack1ton_fp32(const float *input, float *input_padded, int inc, int inh,
  518. int inw, int padded_h, int padded_w, int pad_top,
  519. int pad_left);
  520. void shl_rvv_pad_input_pack1ton_fp16(const __fp16 *input, __fp16 *input_padded, int inc, int inh,
  521. int inw, int padded_h, int padded_w, int pad_top,
  522. int pad_left);
  523. void shl_rvv_pad_input_pack1ton_int8(const int8_t *input, int8_t *input_padded, int inc, int inh,
  524. int inw, int padded_h, int padded_w, int pad_top, int pad_left,
  525. int8_t pad_value);
  526. void shl_rvv_reorder_input_pack1ton_fp32(const float *src, float *dst, int inc, int inh, int inw);
  527. void shl_rvv_reorder_input_pack1ton_fp16(const __fp16 *src, __fp16 *dst, int inc, int inh, int inw);
  528. void shl_rvv_reorder_input_pack1ton_int8(const int8_t *src, int8_t *dst, int inc, int inh, int inw);
  529. void shl_rvv_reorder_input_packnto1_fp32(const float *src, float *dst, int inc, int inh, int inw);
  530. void shl_rvv_reorder_input_packnto1_fp16(const __fp16 *src, __fp16 *dst, int inc, int inh, int inw);
  531. void shl_rvv_reorder_input_packnto1_int8(const int8_t *src, int8_t *dst, int inc, int inh, int inw);
  532. void shl_rvv_saturated_int8(int32_t *src, int8_t *dst, int32_t out_zp, int size);
  533. void shl_rvv_requantize(int32_t *src, int32_t multiplier, int32_t shift, int channel_size);
  534. void shl_rvv_pad_input_int4_trans_int8(const int8_t *input, int8_t *input_padded, int inc, int inh,
  535. int inw, int padded_h, int padded_w, int pad_top,
  536. int pad_left, int8_t pad_value);
  537. void shl_rvv_int4_to_int8(int8_t *src, int8_t *dst, int size);
  538. void shl_rvv_int8_to_int4(int8_t *src, int8_t *dst, int size);
  539. void shl_rvv_int4_trans_int8(int8_t *src, int8_t *dst, int size);
  540. void shl_rvv_saturated_int4(int32_t *src, int8_t *dst, int32_t out_zp, int size);
  541. int shl_rvv_data_convert_int8_to_int4(struct csinn_tensor *input, struct csinn_tensor *output,
  542. struct csinn_siso_params *params);
  543. int shl_rvv_data_convert_int4_to_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  544. struct csinn_siso_params *params);
  545. struct csinn_callback *shl_cb_map_rvv(int op, int dtype);
  546. void shl_rvv_reg_op(enum csinn_dtype_enum dtype, enum csinn_op_enum op_name, void *init, void *exec,
  547. void *est);
  548. int csrr_vl();
  549. int csrr_vlenb();
  550. enum avgpool_loc_enum {
  551. AVGPOOL_LEFT_TOP = 0,
  552. AVGPOOL_RIGHT_TOP,
  553. AVGPOOL_LEFT_BOTTOM,
  554. AVGPOOL_RIGHT_BOTTOM,
  555. AVGPOOL_LEFT,
  556. AVGPOOL_RIGHT,
  557. AVGPOOL_TOP,
  558. AVGPOOL_BOTTOM,
  559. AVGPOOL_CENTER,
  560. };
  561. int shl_rvv_avgpool_get_window_size(struct csinn_pool_params *params, int idx_h_start,
  562. int idx_h_end, int idx_w_start, int idx_w_end,
  563. enum avgpool_loc_enum loc);
  564. #ifdef SHL_USE_DOT_INT8
  565. int shl_rvv_conv2d_init_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  566. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  567. struct csinn_conv2d_params *params);
  568. void shl_rvv_conv_im2col_gemm_reorder_kernel_int8(struct csinn_tensor *kernel,
  569. struct csinn_conv2d_params *params);
  570. int shl_rvv_conv_im2col_gemm_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  571. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  572. struct csinn_conv2d_params *params);
  573. void shl_rvv_conv_im2col_gemm_reorder_kernel_packn_int8(struct csinn_tensor *kernel,
  574. struct csinn_conv2d_params *params);
  575. int shl_rvv_conv_im2col_gemm_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  576. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  577. struct csinn_conv2d_params *params);
  578. void shl_rvv_conv_im2col_gemm_reorder_kernel_pack1ton_int8(struct csinn_tensor *kernel,
  579. struct csinn_conv2d_params *params);
  580. int shl_rvv_conv_im2col_gemm_pack1ton_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  581. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  582. struct csinn_conv2d_params *params);
  583. void shl_rvv_conv_im2col_gemm_reorder_kernel_packnto1_int8(struct csinn_tensor *kernel,
  584. struct csinn_conv2d_params *params);
  585. int shl_rvv_conv_im2col_gemm_packnto1_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  586. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  587. struct csinn_conv2d_params *params);
  588. void shl_rvv_conv1x1s1_gemm_reorder_kernel_int8(struct csinn_tensor *kernel,
  589. struct csinn_conv2d_params *params);
  590. int shl_rvv_conv1x1s1_gemm_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  591. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  592. struct csinn_conv2d_params *params);
  593. void shl_rvv_conv1x1s1_gemm_reorder_kernel_packn_int8(struct csinn_tensor *kernel,
  594. struct csinn_conv2d_params *params);
  595. int shl_rvv_conv1x1s1_gemm_packn_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  596. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  597. struct csinn_conv2d_params *params);
  598. void shl_rvv_conv1x1s1_gemm_reorder_kernel_pack1ton_int8(struct csinn_tensor *kernel,
  599. struct csinn_conv2d_params *params);
  600. int shl_rvv_conv1x1s1_gemm_pack1ton_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  601. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  602. struct csinn_conv2d_params *params);
  603. void shl_rvv_conv1x1s1_gemm_reorder_kernel_packnto1_int8(struct csinn_tensor *kernel,
  604. struct csinn_conv2d_params *params);
  605. int shl_rvv_conv1x1s1_gemm_packnto1_int8(struct csinn_tensor *input, struct csinn_tensor *output,
  606. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  607. struct csinn_conv2d_params *params);
  608. void shl_rvv_fc_gemv_transform_weight_int8_dot(struct csinn_tensor *weights);
  609. int shl_rvv_fullyconnected_packn_int8_dot(struct csinn_tensor *input, struct csinn_tensor *output,
  610. struct csinn_tensor *weights, struct csinn_tensor *bias,
  611. struct csinn_fc_params *params);
  612. #endif
  613. #ifdef SHL_USE_DOT_INT4
  614. int shl_rvv_conv2d_init_int4(struct csinn_tensor *input, struct csinn_tensor *output,
  615. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  616. struct csinn_conv2d_params *params);
  617. void shl_rvv_conv_im2col_gemm_reorder_kernel_int4(struct csinn_tensor *kernel,
  618. struct csinn_conv2d_params *params);
  619. int shl_rvv_conv_im2col_gemm_int4(struct csinn_tensor *input, struct csinn_tensor *output,
  620. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  621. struct csinn_conv2d_params *params);
  622. void shl_rvv_conv_im2col_gemm_reorder_kernel_packn_int4(struct csinn_tensor *kernel,
  623. struct csinn_conv2d_params *params);
  624. int shl_rvv_conv_im2col_gemm_packn_int4(struct csinn_tensor *input, struct csinn_tensor *output,
  625. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  626. struct csinn_conv2d_params *params);
  627. void shl_rvv_conv1x1s1_gemm_reorder_kernel_int4(struct csinn_tensor *kernel,
  628. struct csinn_conv2d_params *params);
  629. int shl_rvv_conv1x1s1_gemm_int4(struct csinn_tensor *input, struct csinn_tensor *output,
  630. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  631. struct csinn_conv2d_params *params);
  632. void shl_rvv_conv1x1s1_gemm_reorder_kernel_packn_int4(struct csinn_tensor *kernel,
  633. struct csinn_conv2d_params *params);
  634. int shl_rvv_conv1x1s1_gemm_packn_int4(struct csinn_tensor *input, struct csinn_tensor *output,
  635. struct csinn_tensor *kernel, struct csinn_tensor *bias,
  636. struct csinn_conv2d_params *params);
  637. void shl_rvv_fc_gemv_transform_weight_int4_dot(struct csinn_tensor *weights);
  638. int shl_rvv_fullyconnected_packn_int4_dot(struct csinn_tensor *input, struct csinn_tensor *output,
  639. struct csinn_tensor *weights, struct csinn_tensor *bias,
  640. struct csinn_fc_params *params);
  641. #endif
  642. #ifdef __cplusplus
  643. }
  644. #endif
  645. #endif // INCLUDE_SHL_RVV_H_