OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
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.
Tree: bec9d9f63d
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'bec9d9f63d'
${ noResults }
OpenBLAS/kernel/x86_64/sgemm_small_kernel_nt_skyla...

536 lines
21 kB
Raw Blame History 

  1. /***************************************************************************

  2. Copyright (c) 2021, The OpenBLAS Project

  3. All rights reserved.

  4. Redistribution and use in source and binary forms, with or without

  5. modification, are permitted provided that the following conditions are

  6. met:

  7. 1. Redistributions of source code must retain the above copyright

  8. notice, this list of conditions and the following disclaimer.

  9. 2. Redistributions in binary form must reproduce the above copyright

  10. notice, this list of conditions and the following disclaimer in

  11. the documentation and/or other materials provided with the

  12. distribution.

  13. 3. Neither the name of the OpenBLAS project nor the names of

  14. its contributors may be used to endorse or promote products

  15. derived from this software without specific prior written permission.

  16. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"

  17. AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE

  18. IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE

  19. ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE

  20. LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL

  21. DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR

  22. SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER

  23. CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,

  24. OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE

  25. USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

  26. *****************************************************************************/

  27. 

  28. #include <immintrin.h>

  29. #include "common.h"

  30. #include <stdio.h>

  31. #include <memory.h>

  32. 

  33. #define DECLARE_RESULT_512(M, N) __m512 result##M##N = _mm512_setzero_ps()

  34. #define LOAD_A_512(M, N) __m512 Aval##M = _mm512_loadu_ps(&A[lda * k + i + (M*16)])

  35. #define MASK_LOAD_A_512(M, N) __m512 Aval##M = _mm512_maskz_loadu_ps(mask, &A[lda * k + i + (M*16)])

  36. #define BROADCAST_LOAD_B_512(M, N) __m512 Bval##N = _mm512_broadcastss_ps(_mm_load_ss(&B[ldb * k + j + N]))

  37. #define MATMUL_512(M, N) result##M##N = _mm512_fmadd_ps(Aval##M, Bval##N, result##M##N)

  38. 

  39. #define BROADCAST_LOAD_A_512(M, N) __m512 Aval##M = _mm512_broadcastss_ps(_mm_load_ss(&A[lda * k + i + M]))

  40. #define LOAD_B_512(M, N) __m512 Bval##N = _mm512_loadu_ps(&B[ldb * k + j + (N*16)])

  41. #define MASK_LOAD_B_512(M, N) __m512 Bval##N = _mm512_maskz_loadu_ps(mask, &B[ldb * k + j + (N*16)])

  42. #if defined(B0)

  43. #define STORE_512(M, N) result##M##N = _mm512_mul_ps(result##M##N, alpha_512); \

  44. 			_mm512_storeu_ps(&C[(j+N)*ldc + i + (M*16)], result##M##N)

  45. #define MASK_STORE_512(M, N) result##M##N = _mm512_mul_ps(result##M##N, alpha_512); \

  46. 			_mm512_mask_storeu_ps(&C[(j+N)*ldc + i + (M*16)], mask, result##M##N)

  47. #define SCATTER_STORE_512(M, N) result##M##N = _mm512_mul_ps(result##M##N, alpha_512); \

  48. 				_mm512_i32scatter_ps(&C[(j + N*16)*ldc + i + M], vindex_n, result##M##N, 4);

  49. #define MASK_SCATTER_STORE_512(M, N) result##M##N = _mm512_mul_ps(result##M##N, alpha_512); \

  50. 				_mm512_mask_i32scatter_ps(&C[(j + N*16)*ldc + i + M], mask, vindex_n, result##M##N, 4)

  51. #else

  52. #define STORE_512(M, N) \

  53. 	result##M##N = _mm512_mul_ps(result##M##N, alpha_512); \

  54. 	asm("vfmadd231ps (%1), %2, %0": "+v"(result##M##N):"r"(&C[(j+N)*ldc + i + (M*16)]), "v"(beta_512)); \

  55. 	_mm512_storeu_ps(&C[(j+N)*ldc + i + (M*16)], result##M##N)

  56. #define MASK_STORE_512(M, N) \

  57. 	result##M##N = _mm512_mul_ps(result##M##N, alpha_512); \

  58. 	asm("vfmadd231ps (%1), %2, %0 %{%3%}": "+v"(result##M##N):"r"(&C[(j+N)*ldc + i + (M*16)]), "v"(beta_512), "k"(mask)); \

  59. 	_mm512_mask_storeu_ps(&C[(j+N)*ldc + i + (M*16)], mask, result##M##N)

  60. #define SCATTER_STORE_512(M, N) result##M##N = _mm512_mul_ps(result##M##N, alpha_512); \

  61. 				__m512 tmp##M##N = _mm512_i32gather_ps(vindex_n, &C[(j + N*16)*ldc + i + M], 4); \

  62. 				result##M##N = _mm512_fmadd_ps(tmp##M##N, beta_512, result##M##N); \

  63. 				_mm512_i32scatter_ps(&C[(j + N*16)*ldc + i + M], vindex_n, result##M##N, 4);

  64. #define MASK_SCATTER_STORE_512(M, N) result##M##N = _mm512_mul_ps(result##M##N, alpha_512); \

  65. 				__m512 tmp##M##N = _mm512_mask_i32gather_ps(_mm512_setzero_ps(), mask, vindex_n, &C[(j + N*16)*ldc + i + M], 4); \

  66. 				result##M##N = _mm512_fmadd_ps(tmp##M##N, beta_512, result##M##N); \

  67. 				_mm512_mask_i32scatter_ps(&C[(j + N*16)*ldc + i + M], mask, vindex_n, result##M##N, 4);

  68. #endif

  69. 

  70. #if defined(B0)

  71. int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)

  72. #else

  73. int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha, FLOAT * B, BLASLONG ldb, FLOAT beta, FLOAT * C, BLASLONG ldc)

  74. #endif

  75. {

  76. 	// column major

  77. 	BLASLONG i, j, k;

  78. 

  79. 	BLASLONG m64 = M & ~63;

  80. 	BLASLONG m32 = M & ~31;

  81. 	BLASLONG m16 = M & ~15;

  82. 	BLASLONG m4 = M & ~3;

  83. 	BLASLONG m2 = M & ~1;

  84. 

  85. 	BLASLONG n64 = N & ~63;

  86. 	BLASLONG n32 = N & ~31;

  87. 	BLASLONG n8 = N & ~7;

  88. 	BLASLONG n6 = N - (N % 6);

  89. 	BLASLONG n4 = N & ~3;

  90. 	BLASLONG n2 = N & ~1;

  91. 

  92. 

  93. 	__m512 alpha_512 = _mm512_broadcastss_ps(_mm_load_ss(&alpha));

  94. #if !defined(B0)

  95. 	__m512 beta_512 = _mm512_broadcastss_ps(_mm_load_ss(&beta));

  96. #endif

  97. 

  98. 	for (i = 0; i < m64; i += 64) {

  99. 		for (j = 0; j < n6; j += 6) {

  100. 			DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0);

  101. 			DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1);

  102. 			DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2); DECLARE_RESULT_512(2, 2); DECLARE_RESULT_512(3, 2);

  103. 			DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3); DECLARE_RESULT_512(2, 3); DECLARE_RESULT_512(3, 3);

  104. 			DECLARE_RESULT_512(0, 4); DECLARE_RESULT_512(1, 4); DECLARE_RESULT_512(2, 4); DECLARE_RESULT_512(3, 4);

  105. 			DECLARE_RESULT_512(0, 5); DECLARE_RESULT_512(1, 5); DECLARE_RESULT_512(2, 5); DECLARE_RESULT_512(3, 5);

  106. 

  107. 			for (k = 0; k < K; k++) {

  108. 				LOAD_A_512(0, x); LOAD_A_512(1, x); LOAD_A_512(2, x); LOAD_A_512(3, x);

  109. 

  110. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  111. 				MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);

  112. 				MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1);

  113. 				BROADCAST_LOAD_B_512(x, 2); BROADCAST_LOAD_B_512(x, 3);

  114. 				MATMUL_512(0, 2); MATMUL_512(1, 2); MATMUL_512(2, 2); MATMUL_512(3, 2);

  115. 				MATMUL_512(0, 3); MATMUL_512(1, 3); MATMUL_512(2, 3); MATMUL_512(3, 3);

  116. 				BROADCAST_LOAD_B_512(x, 4); BROADCAST_LOAD_B_512(x, 5);

  117. 				MATMUL_512(0, 4); MATMUL_512(1, 4); MATMUL_512(2, 4); MATMUL_512(3, 4);

  118. 				MATMUL_512(0, 5); MATMUL_512(1, 5); MATMUL_512(2, 5); MATMUL_512(3, 5);

  119. 			}

  120. 			STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0);

  121. 			STORE_512(0, 1); STORE_512(1, 1); STORE_512(2, 1); STORE_512(3, 1);

  122. 			STORE_512(0, 2); STORE_512(1, 2); STORE_512(2, 2); STORE_512(3, 2);

  123. 			STORE_512(0, 3); STORE_512(1, 3); STORE_512(2, 3); STORE_512(3, 3);

  124. 			STORE_512(0, 4); STORE_512(1, 4); STORE_512(2, 4); STORE_512(3, 4);

  125. 			STORE_512(0, 5); STORE_512(1, 5); STORE_512(2, 5); STORE_512(3, 5);

  126. 		}

  127. 		for (; j < n2; j += 2) {

  128. 			DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0);

  129. 			DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1);

  130. 			for (k = 0; k < K; k++) {

  131. 				LOAD_A_512(0, x); LOAD_A_512(1, x); LOAD_A_512(2, x); LOAD_A_512(3, x);

  132. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  133. 				MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);

  134. 				MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1);

  135. 			}

  136. 			STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0);

  137. 			STORE_512(0, 1); STORE_512(1, 1); STORE_512(2, 1); STORE_512(3, 1);

  138. 		}

  139. 		for (; j < N; j++) {

  140. 			DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0);

  141. 			for (k = 0; k < K; k++) {

  142. 				LOAD_A_512(0, x); LOAD_A_512(1, x); LOAD_A_512(2, x); LOAD_A_512(3, x);

  143. 				BROADCAST_LOAD_B_512(x, 0);

  144. 				MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);

  145. 			}

  146. 			STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0);

  147. 		}

  148. 	}

  149. 	for (; i < m32; i += 32) {

  150. 		for (j = 0; j < n8; j += 8) {

  151. 			DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);

  152. 			DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1);

  153. 			DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2);

  154. 			DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3);

  155. 			DECLARE_RESULT_512(0, 4); DECLARE_RESULT_512(1, 4);

  156. 			DECLARE_RESULT_512(0, 5); DECLARE_RESULT_512(1, 5);

  157. 			DECLARE_RESULT_512(0, 6); DECLARE_RESULT_512(1, 6);

  158. 			DECLARE_RESULT_512(0, 7); DECLARE_RESULT_512(1, 7);

  159. 			for (k = 0; k < K; k++) {

  160. 				LOAD_A_512(0, x); LOAD_A_512(1, x);

  161. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  162. 				BROADCAST_LOAD_B_512(x, 2); BROADCAST_LOAD_B_512(x, 3);

  163. 				BROADCAST_LOAD_B_512(x, 4); BROADCAST_LOAD_B_512(x, 5);

  164. 				BROADCAST_LOAD_B_512(x, 6); BROADCAST_LOAD_B_512(x, 7);

  165. 

  166. 				MATMUL_512(0, 0); MATMUL_512(1, 0);

  167. 				MATMUL_512(0, 1); MATMUL_512(1, 1);

  168. 				MATMUL_512(0, 2); MATMUL_512(1, 2);

  169. 				MATMUL_512(0, 3); MATMUL_512(1, 3);

  170. 				MATMUL_512(0, 4); MATMUL_512(1, 4);

  171. 				MATMUL_512(0, 5); MATMUL_512(1, 5);

  172. 				MATMUL_512(0, 6); MATMUL_512(1, 6);

  173. 				MATMUL_512(0, 7); MATMUL_512(1, 7);

  174. 			}

  175. 			STORE_512(0, 0); STORE_512(1, 0);

  176. 			STORE_512(0, 1); STORE_512(1, 1);

  177. 			STORE_512(0, 2); STORE_512(1, 2);

  178. 			STORE_512(0, 3); STORE_512(1, 3);

  179. 			STORE_512(0, 4); STORE_512(1, 4);

  180. 			STORE_512(0, 5); STORE_512(1, 5);

  181. 			STORE_512(0, 6); STORE_512(1, 6);

  182. 			STORE_512(0, 7); STORE_512(1, 7);

  183. 		}

  184. 		for (;j < n4; j += 4) {

  185. 			DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);

  186. 			DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1);

  187. 			DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2);

  188. 			DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3);

  189. 			for (k = 0; k < K; k++) {

  190. 				LOAD_A_512(0, x); LOAD_A_512(1, x);

  191. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  192. 				BROADCAST_LOAD_B_512(x, 2); BROADCAST_LOAD_B_512(x, 3);

  193. 

  194. 				MATMUL_512(0, 0); MATMUL_512(1, 0);

  195. 				MATMUL_512(0, 1); MATMUL_512(1, 1);

  196. 				MATMUL_512(0, 2); MATMUL_512(1, 2);

  197. 				MATMUL_512(0, 3); MATMUL_512(1, 3);

  198. 			}

  199. 			STORE_512(0, 0); STORE_512(1, 0);

  200. 			STORE_512(0, 1); STORE_512(1, 1);

  201. 			STORE_512(0, 2); STORE_512(1, 2);

  202. 			STORE_512(0, 3); STORE_512(1, 3);

  203. 		}

  204. 		for (; j < n2; j += 2) {

  205. 			DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);

  206. 			DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1);

  207. 			for (k = 0; k < K; k++) {

  208. 				LOAD_A_512(0, x); LOAD_A_512(1, x);

  209. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  210. 				MATMUL_512(0, 0); MATMUL_512(1, 0);

  211. 				MATMUL_512(0, 1); MATMUL_512(1, 1);

  212. 			}

  213. 			STORE_512(0, 0); STORE_512(1, 0);

  214. 			STORE_512(0, 1); STORE_512(1, 1);

  215. 		}

  216. 		for (; j < N; j++) {

  217. 			DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);

  218. 			for (k = 0; k < K; k++) {

  219. 				LOAD_A_512(0, x); LOAD_A_512(1, x);

  220. 				BROADCAST_LOAD_B_512(x, 0);

  221. 				MATMUL_512(0, 0); MATMUL_512(1, 0);

  222. 			}

  223. 			STORE_512(0, 0); STORE_512(1, 0);

  224. 		}

  225. 	}

  226. 	for (; i < m16; i += 16) {

  227. 		for (j = 0; j < n8; j += 8) {

  228. 			DECLARE_RESULT_512(0, 0);

  229. 			DECLARE_RESULT_512(0, 1);

  230. 			DECLARE_RESULT_512(0, 2);

  231. 			DECLARE_RESULT_512(0, 3);

  232. 			DECLARE_RESULT_512(0, 4);

  233. 			DECLARE_RESULT_512(0, 5);

  234. 			DECLARE_RESULT_512(0, 6);

  235. 			DECLARE_RESULT_512(0, 7);

  236. 			for (k = 0; k < K; k++) {

  237. 				LOAD_A_512(0, x);

  238. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  239. 				BROADCAST_LOAD_B_512(x, 2); BROADCAST_LOAD_B_512(x, 3);

  240. 				BROADCAST_LOAD_B_512(x, 4); BROADCAST_LOAD_B_512(x, 5);

  241. 				BROADCAST_LOAD_B_512(x, 6); BROADCAST_LOAD_B_512(x, 7);

  242. 

  243. 				MATMUL_512(0, 0);

  244. 				MATMUL_512(0, 1);

  245. 				MATMUL_512(0, 2);

  246. 				MATMUL_512(0, 3);

  247. 				MATMUL_512(0, 4);

  248. 				MATMUL_512(0, 5);

  249. 				MATMUL_512(0, 6);

  250. 				MATMUL_512(0, 7);

  251. 			}

  252. 			STORE_512(0, 0);

  253. 			STORE_512(0, 1);

  254. 			STORE_512(0, 2);

  255. 			STORE_512(0, 3);

  256. 			STORE_512(0, 4);

  257. 			STORE_512(0, 5);

  258. 			STORE_512(0, 6);

  259. 			STORE_512(0, 7);

  260. 		}

  261. 		for (; j < n4; j += 4) {

  262. 			DECLARE_RESULT_512(0, 0);

  263. 			DECLARE_RESULT_512(0, 1);

  264. 			DECLARE_RESULT_512(0, 2);

  265. 			DECLARE_RESULT_512(0, 3);

  266. 			for (k = 0; k < K; k++) {

  267. 				LOAD_A_512(0, x);

  268. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  269. 				BROADCAST_LOAD_B_512(x, 2); BROADCAST_LOAD_B_512(x, 3);

  270. 

  271. 				MATMUL_512(0, 0);

  272. 				MATMUL_512(0, 1);

  273. 				MATMUL_512(0, 2);

  274. 				MATMUL_512(0, 3);

  275. 			}

  276. 			STORE_512(0, 0);

  277. 			STORE_512(0, 1);

  278. 			STORE_512(0, 2);

  279. 			STORE_512(0, 3);

  280. 		}

  281. 

  282. 		for (; j < n2; j += 2) {

  283. 			DECLARE_RESULT_512(0, 0);

  284. 			DECLARE_RESULT_512(0, 1);

  285. 			for (k = 0; k < K; k++) {

  286. 				LOAD_A_512(0, x);

  287. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  288. 				MATMUL_512(0, 0);

  289. 				MATMUL_512(0, 1);

  290. 			}

  291. 			STORE_512(0, 0);

  292. 			STORE_512(0, 1);

  293. 		}

  294. 		for (; j < N; j++) {

  295. 			DECLARE_RESULT_512(0, 0);

  296. 			for (k = 0; k < K; k++) {

  297. 				LOAD_A_512(0, x);

  298. 				BROADCAST_LOAD_B_512(x, 0);

  299. 				MATMUL_512(0, 0);

  300. 			}

  301. 			STORE_512(0, 0);

  302. 		}

  303. 	}

  304. 	int mm = M - i;

  305. 	if (mm >= 12) {

  306. 		register __mmask16 mask asm("k1") = (1UL << mm) - 1;

  307. 		for (j = 0; j < n8; j += 8) {

  308. 			DECLARE_RESULT_512(0, 0);

  309. 			DECLARE_RESULT_512(0, 1);

  310. 			DECLARE_RESULT_512(0, 2);

  311. 			DECLARE_RESULT_512(0, 3);

  312. 			DECLARE_RESULT_512(0, 4);

  313. 			DECLARE_RESULT_512(0, 5);

  314. 			DECLARE_RESULT_512(0, 6);

  315. 			DECLARE_RESULT_512(0, 7);

  316. 			for (k = 0; k < K; k++) {

  317. 				MASK_LOAD_A_512(0, x);

  318. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  319. 				BROADCAST_LOAD_B_512(x, 2); BROADCAST_LOAD_B_512(x, 3);

  320. 				BROADCAST_LOAD_B_512(x, 4); BROADCAST_LOAD_B_512(x, 5);

  321. 				BROADCAST_LOAD_B_512(x, 6); BROADCAST_LOAD_B_512(x, 7);

  322. 

  323. 				MATMUL_512(0, 0);

  324. 				MATMUL_512(0, 1);

  325. 				MATMUL_512(0, 2);

  326. 				MATMUL_512(0, 3);

  327. 				MATMUL_512(0, 4);

  328. 				MATMUL_512(0, 5);

  329. 				MATMUL_512(0, 6);

  330. 				MATMUL_512(0, 7);

  331. 			}

  332. 			MASK_STORE_512(0, 0);

  333. 			MASK_STORE_512(0, 1);

  334. 			MASK_STORE_512(0, 2);

  335. 			MASK_STORE_512(0, 3);

  336. 			MASK_STORE_512(0, 4);

  337. 			MASK_STORE_512(0, 5);

  338. 			MASK_STORE_512(0, 6);

  339. 			MASK_STORE_512(0, 7);

  340. 		}

  341. 		for (; j < n4; j += 4) {

  342. 			DECLARE_RESULT_512(0, 0);

  343. 			DECLARE_RESULT_512(0, 1);

  344. 			DECLARE_RESULT_512(0, 2);

  345. 			DECLARE_RESULT_512(0, 3);

  346. 			for (k = 0; k < K; k++) {

  347. 				MASK_LOAD_A_512(0, x);

  348. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  349. 				BROADCAST_LOAD_B_512(x, 2); BROADCAST_LOAD_B_512(x, 3);

  350. 

  351. 				MATMUL_512(0, 0);

  352. 				MATMUL_512(0, 1);

  353. 				MATMUL_512(0, 2);

  354. 				MATMUL_512(0, 3);

  355. 			}

  356. 			MASK_STORE_512(0, 0);

  357. 			MASK_STORE_512(0, 1);

  358. 			MASK_STORE_512(0, 2);

  359. 			MASK_STORE_512(0, 3);

  360. 		}

  361. 

  362. 		for (; j < n2; j += 2) {

  363. 			DECLARE_RESULT_512(0, 0);

  364. 			DECLARE_RESULT_512(0, 1);

  365. 			for (k = 0; k < K; k++) {

  366. 				MASK_LOAD_A_512(0, x);

  367. 				BROADCAST_LOAD_B_512(x, 0); BROADCAST_LOAD_B_512(x, 1);

  368. 				MATMUL_512(0, 0);

  369. 				MATMUL_512(0, 1);

  370. 			}

  371. 			MASK_STORE_512(0, 0);

  372. 			MASK_STORE_512(0, 1);

  373. 		}

  374. 		for (; j < N; j++) {

  375. 			DECLARE_RESULT_512(0, 0);

  376. 			for (k = 0; k < K; k++) {

  377. 				MASK_LOAD_A_512(0, x);

  378. 				BROADCAST_LOAD_B_512(x, 0);

  379. 				MATMUL_512(0, 0);

  380. 			}

  381. 			MASK_STORE_512(0, 0);

  382. 		}

  383. 	} else if (mm > 0) {

  384. 		int index_n[16];

  385. 		for (int ii = 0; ii < 16; ii++) {

  386. 			index_n[ii] = ii * ldc;

  387. 		}

  388. 		__m512i vindex_n = _mm512_loadu_si512(index_n);

  389. 		for (; i < m4; i += 4) {

  390. 			for (j = 0; j < n64; j += 64) {

  391. 				DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0);

  392. 				DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1);

  393. 				DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2); DECLARE_RESULT_512(2, 2); DECLARE_RESULT_512(3, 2);

  394. 				DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3); DECLARE_RESULT_512(2, 3); DECLARE_RESULT_512(3, 3);

  395. 				for (k = 0; k < K; k++) {

  396. 					BROADCAST_LOAD_A_512(0, x); BROADCAST_LOAD_A_512(1, x); BROADCAST_LOAD_A_512(2, x); BROADCAST_LOAD_A_512(3, x);

  397. 					LOAD_B_512(x, 0);

  398. 					LOAD_B_512(x, 1);

  399. 					LOAD_B_512(x, 2);

  400. 					LOAD_B_512(x, 3);

  401. 					MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);

  402. 					MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1);

  403. 					MATMUL_512(0, 2); MATMUL_512(1, 2); MATMUL_512(2, 2); MATMUL_512(3, 2);

  404. 					MATMUL_512(0, 3); MATMUL_512(1, 3); MATMUL_512(2, 3); MATMUL_512(3, 3);

  405. 				}

  406. 				SCATTER_STORE_512(0, 0); SCATTER_STORE_512(1, 0); SCATTER_STORE_512(2, 0); SCATTER_STORE_512(3, 0);

  407. 				SCATTER_STORE_512(0, 1); SCATTER_STORE_512(1, 1); SCATTER_STORE_512(2, 1); SCATTER_STORE_512(3, 1);

  408. 				SCATTER_STORE_512(0, 2); SCATTER_STORE_512(1, 2); SCATTER_STORE_512(2, 2); SCATTER_STORE_512(3, 2);

  409. 				SCATTER_STORE_512(0, 3); SCATTER_STORE_512(1, 3); SCATTER_STORE_512(2, 3); SCATTER_STORE_512(3, 3);

  410. 			}

  411. 			for (; j < n32; j += 32) {

  412. 				DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0);

  413. 				DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1);

  414. 				for (k = 0; k < K; k++) {

  415. 					BROADCAST_LOAD_A_512(0, x); BROADCAST_LOAD_A_512(1, x); BROADCAST_LOAD_A_512(2, x); BROADCAST_LOAD_A_512(3, x);

  416. 					LOAD_B_512(x, 0);

  417. 					LOAD_B_512(x, 1);

  418. 					MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);

  419. 					MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1);

  420. 				}

  421. 				SCATTER_STORE_512(0, 0); SCATTER_STORE_512(1, 0); SCATTER_STORE_512(2, 0); SCATTER_STORE_512(3, 0);

  422. 				SCATTER_STORE_512(0, 1); SCATTER_STORE_512(1, 1); SCATTER_STORE_512(2, 1); SCATTER_STORE_512(3, 1);

  423. 			}

  424. 			__mmask16 mask = 0xffff;

  425. 			for (; j < N; j += 16) {

  426. 				int remains = N - j;

  427. 				if (remains < 16) mask = (1UL << remains) - 1;

  428. 				DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0);

  429. 				for (k = 0; k < K; k++) {

  430. 					BROADCAST_LOAD_A_512(0, x); BROADCAST_LOAD_A_512(1, x); BROADCAST_LOAD_A_512(2, x); BROADCAST_LOAD_A_512(3, x);

  431. 					MASK_LOAD_B_512(x, 0);

  432. 					MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);

  433. 				}

  434. 				MASK_SCATTER_STORE_512(0, 0); MASK_SCATTER_STORE_512(1, 0); MASK_SCATTER_STORE_512(2, 0); MASK_SCATTER_STORE_512(3, 0);

  435. 			}

  436. 		}

  437. 		for (; i < m2; i += 2) {

  438. 			for (j = 0; j < n64; j += 64) {

  439. 				DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);

  440. 				DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1);

  441. 				DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2);

  442. 				DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3);

  443. 				for (k = 0; k < K; k++) {

  444. 					BROADCAST_LOAD_A_512(0, x); BROADCAST_LOAD_A_512(1, x);

  445. 					LOAD_B_512(x, 0);

  446. 					LOAD_B_512(x, 1);

  447. 					LOAD_B_512(x, 2);

  448. 					LOAD_B_512(x, 3);

  449. 					MATMUL_512(0, 0); MATMUL_512(1, 0);

  450. 					MATMUL_512(0, 1); MATMUL_512(1, 1);

  451. 					MATMUL_512(0, 2); MATMUL_512(1, 2);

  452. 					MATMUL_512(0, 3); MATMUL_512(1, 3);

  453. 				}

  454. 				SCATTER_STORE_512(0, 0); SCATTER_STORE_512(1, 0);

  455. 				SCATTER_STORE_512(0, 1); SCATTER_STORE_512(1, 1);

  456. 				SCATTER_STORE_512(0, 2); SCATTER_STORE_512(1, 2);

  457. 				SCATTER_STORE_512(0, 3); SCATTER_STORE_512(1, 3);

  458. 			}

  459. 			for (; j < n32; j += 32) {

  460. 				DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);

  461. 				DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1);

  462. 				for (k = 0; k < K; k++) {

  463. 					BROADCAST_LOAD_A_512(0, x); BROADCAST_LOAD_A_512(1, x);

  464. 					LOAD_B_512(x, 0);

  465. 					LOAD_B_512(x, 1);

  466. 					MATMUL_512(0, 0); MATMUL_512(1, 0);

  467. 					MATMUL_512(0, 1); MATMUL_512(1, 1);

  468. 				}

  469. 				SCATTER_STORE_512(0, 0); SCATTER_STORE_512(1, 0);

  470. 				SCATTER_STORE_512(0, 1); SCATTER_STORE_512(1, 1);

  471. 			}

  472. 			__mmask16 mask = 0xffff;

  473. 			for (; j < N; j += 16) {

  474. 				int remains = N - j;

  475. 				if (remains < 16) mask = (1UL << remains) - 1;

  476. 				DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);

  477. 				for (k = 0; k < K; k++) {

  478. 					BROADCAST_LOAD_A_512(0, x); BROADCAST_LOAD_A_512(1, x);

  479. 					MASK_LOAD_B_512(x, 0);

  480. 					MATMUL_512(0, 0); MATMUL_512(1, 0);

  481. 				}

  482. 				MASK_SCATTER_STORE_512(0, 0); MASK_SCATTER_STORE_512(1, 0);

  483. 			}

  484. 		}

  485. 		for (; i < M; i += 1) {

  486. 			for (j = 0; j < n64; j += 64) {

  487. 				DECLARE_RESULT_512(0, 0);

  488. 				DECLARE_RESULT_512(0, 1);

  489. 				DECLARE_RESULT_512(0, 2);

  490. 				DECLARE_RESULT_512(0, 3);

  491. 				for (k = 0; k < K; k++) {

  492. 					BROADCAST_LOAD_A_512(0, x);

  493. 					LOAD_B_512(x, 0);

  494. 					LOAD_B_512(x, 1);

  495. 					LOAD_B_512(x, 2);

  496. 					LOAD_B_512(x, 3);

  497. 					MATMUL_512(0, 0);

  498. 					MATMUL_512(0, 1);

  499. 					MATMUL_512(0, 2);

  500. 					MATMUL_512(0, 3);

  501. 				}

  502. 				SCATTER_STORE_512(0, 0);

  503. 				SCATTER_STORE_512(0, 1);

  504. 				SCATTER_STORE_512(0, 2);

  505. 				SCATTER_STORE_512(0, 3);

  506. 			}

  507. 			for (; j < n32; j += 32) {

  508. 				DECLARE_RESULT_512(0, 0);

  509. 				DECLARE_RESULT_512(0, 1);

  510. 				for (k = 0; k < K; k++) {

  511. 					BROADCAST_LOAD_A_512(0, x);

  512. 					LOAD_B_512(x, 0);

  513. 					LOAD_B_512(x, 1);

  514. 					MATMUL_512(0, 0);

  515. 					MATMUL_512(0, 1);

  516. 				}

  517. 				SCATTER_STORE_512(0, 0);

  518. 				SCATTER_STORE_512(0, 1);

  519. 			}

  520. 			__mmask16 mask = 0xffff;

  521. 			for (; j < N; j += 16) {

  522. 				int remains = N - j;

  523. 				if (remains < 16) mask = (1UL << remains) - 1;

  524. 				DECLARE_RESULT_512(0, 0);

  525. 				for (k = 0; k < K; k++) {

  526. 					BROADCAST_LOAD_A_512(0, x);

  527. 					MASK_LOAD_B_512(x, 0);

  528. 					MATMUL_512(0, 0);

  529. 				}

  530. 				MASK_SCATTER_STORE_512(0, 0);

  531. 			}

  532. 		}

  533. 	}

  534. 	return 0;

  535. }