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.

parameter.c 19 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* All rights reserved. */
  4. /* */
  5. /* Redistribution and use in source and binary forms, with or */
  6. /* without modification, are permitted provided that the following */
  7. /* conditions are met: */
  8. /* */
  9. /* 1. Redistributions of source code must retain the above */
  10. /* copyright notice, this list of conditions and the following */
  11. /* disclaimer. */
  12. /* */
  13. /* 2. Redistributions in binary form must reproduce the above */
  14. /* copyright notice, this list of conditions and the following */
  15. /* disclaimer in the documentation and/or other materials */
  16. /* provided with the distribution. */
  17. /* */
  18. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  19. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  20. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  21. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  22. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  23. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  24. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  25. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  26. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  27. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  28. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  29. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  30. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  31. /* POSSIBILITY OF SUCH DAMAGE. */
  32. /* */
  33. /* The views and conclusions contained in the software and */
  34. /* documentation are those of the authors and should not be */
  35. /* interpreted as representing official policies, either expressed */
  36. /* or implied, of The University of Texas at Austin. */
  37. /*********************************************************************/
  38. #include <stdio.h>
  39. #include <string.h>
  40. #include "common.h"
  41. extern int openblas_block_factor(void);
  42. int get_L2_size(void);
  43. #define DEFAULT_GEMM_P 128
  44. #define DEFAULT_GEMM_Q 128
  45. #define DEFAULT_GEMM_R 128
  46. #define DEFAULT_GEMM_OFFSET_A 0
  47. #define DEFAULT_GEMM_OFFSET_B 0
  48. /* Global Parameter */
  49. #if GEMM_OFFSET_A == gemm_offset_a
  50. BLASLONG gemm_offset_a = DEFAULT_GEMM_OFFSET_A;
  51. #else
  52. BLASLONG gemm_offset_a = GEMM_OFFSET_A;
  53. #endif
  54. #if GEMM_OFFSET_B == gemm_offset_b
  55. BLASLONG gemm_offset_b = DEFAULT_GEMM_OFFSET_B;
  56. #else
  57. BLASLONG gemm_offset_b = GEMM_OFFSET_B;
  58. #endif
  59. #if SBGEMM_P == sbgemm_p
  60. BLASLONG sbgemm_p = DEFAULT_GEMM_P;
  61. #else
  62. BLASLONG sbgemm_p = SBGEMM_P;
  63. #endif
  64. #if SGEMM_P == sgemm_p
  65. BLASLONG sgemm_p = DEFAULT_GEMM_P;
  66. #else
  67. BLASLONG sgemm_p = SGEMM_P;
  68. #endif
  69. #if DGEMM_P == dgemm_p
  70. BLASLONG dgemm_p = DEFAULT_GEMM_P;
  71. #else
  72. BLASLONG dgemm_p = DGEMM_P;
  73. #endif
  74. #if CGEMM_P == cgemm_p
  75. BLASLONG cgemm_p = DEFAULT_GEMM_P;
  76. #else
  77. BLASLONG cgemm_p = CGEMM_P;
  78. #endif
  79. #if ZGEMM_P == zgemm_p
  80. BLASLONG zgemm_p = DEFAULT_GEMM_P;
  81. #else
  82. BLASLONG zgemm_p = ZGEMM_P;
  83. #endif
  84. #if SBGEMM_Q == sbgemm_q
  85. BLASLONG sbgemm_q = DEFAULT_GEMM_Q;
  86. #else
  87. BLASLONG sbgemm_q = SBGEMM_Q;
  88. #endif
  89. #if SGEMM_Q == sgemm_q
  90. BLASLONG sgemm_q = DEFAULT_GEMM_Q;
  91. #else
  92. BLASLONG sgemm_q = SGEMM_Q;
  93. #endif
  94. #if DGEMM_Q == dgemm_q
  95. BLASLONG dgemm_q = DEFAULT_GEMM_Q;
  96. #else
  97. BLASLONG dgemm_q = DGEMM_Q;
  98. #endif
  99. #if CGEMM_Q == cgemm_q
  100. BLASLONG cgemm_q = DEFAULT_GEMM_Q;
  101. #else
  102. BLASLONG cgemm_q = CGEMM_Q;
  103. #endif
  104. #if ZGEMM_Q == zgemm_q
  105. BLASLONG zgemm_q = DEFAULT_GEMM_Q;
  106. #else
  107. BLASLONG zgemm_q = ZGEMM_Q;
  108. #endif
  109. #if SBGEMM_R == sbgemm_r
  110. BLASLONG sbgemm_r = DEFAULT_GEMM_R;
  111. #else
  112. BLASLONG sbgemm_r = SBGEMM_R;
  113. #endif
  114. #if SGEMM_R == sgemm_r
  115. BLASLONG sgemm_r = DEFAULT_GEMM_R;
  116. #else
  117. BLASLONG sgemm_r = SGEMM_R;
  118. #endif
  119. #if DGEMM_R == dgemm_r
  120. BLASLONG dgemm_r = DEFAULT_GEMM_R;
  121. #else
  122. BLASLONG dgemm_r = DGEMM_R;
  123. #endif
  124. #if CGEMM_R == cgemm_r
  125. BLASLONG cgemm_r = DEFAULT_GEMM_R;
  126. #else
  127. BLASLONG cgemm_r = CGEMM_R;
  128. #endif
  129. #if ZGEMM_R == zgemm_r
  130. BLASLONG zgemm_r = DEFAULT_GEMM_R;
  131. #else
  132. BLASLONG zgemm_r = ZGEMM_R;
  133. #endif
  134. #if defined(EXPRECISION) || defined(QUAD_PRECISION)
  135. #if QGEMM_P == qgemm_p
  136. BLASLONG qgemm_p = DEFAULT_GEMM_P;
  137. #else
  138. BLASLONG qgemm_p = QGEMM_P;
  139. #endif
  140. #if XGEMM_P == xgemm_p
  141. BLASLONG xgemm_p = DEFAULT_GEMM_P;
  142. #else
  143. BLASLONG xgemm_p = XGEMM_P;
  144. #endif
  145. #if QGEMM_Q == qgemm_q
  146. BLASLONG qgemm_q = DEFAULT_GEMM_Q;
  147. #else
  148. BLASLONG qgemm_q = QGEMM_Q;
  149. #endif
  150. #if XGEMM_Q == xgemm_q
  151. BLASLONG xgemm_q = DEFAULT_GEMM_Q;
  152. #else
  153. BLASLONG xgemm_q = XGEMM_Q;
  154. #endif
  155. #if QGEMM_R == qgemm_r
  156. BLASLONG qgemm_r = DEFAULT_GEMM_R;
  157. #else
  158. BLASLONG qgemm_r = QGEMM_R;
  159. #endif
  160. #if XGEMM_R == xgemm_r
  161. BLASLONG xgemm_r = DEFAULT_GEMM_R;
  162. #else
  163. BLASLONG xgemm_r = XGEMM_R;
  164. #endif
  165. #endif
  166. #if defined(ARCH_X86) || defined(ARCH_X86_64)
  167. int get_L2_size(void){
  168. int eax, ebx, ecx, edx;
  169. #if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER) || \
  170. defined(CORE_PRESCOTT) || defined(CORE_CORE2) || defined(PENRYN) || defined(DUNNINGTON) || \
  171. defined(CORE_NEHALEM) || defined(CORE_SANDYBRIDGE) || defined(ATOM) || defined(GENERIC) || \
  172. defined(PILEDRIVER) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || \
  173. defined(ZEN) || defined(SKYLAKEX) || defined(COOPERLAKE) || defined(SAPPHIRERAPIDS)
  174. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  175. return BITMASK(ecx, 16, 0xffff);
  176. #else
  177. int info[15];
  178. int i;
  179. cpuid(2, &eax, &ebx, &ecx, &edx);
  180. info[ 0] = BITMASK(eax, 8, 0xff);
  181. info[ 1] = BITMASK(eax, 16, 0xff);
  182. info[ 2] = BITMASK(eax, 24, 0xff);
  183. info[ 3] = BITMASK(ebx, 0, 0xff);
  184. info[ 4] = BITMASK(ebx, 8, 0xff);
  185. info[ 5] = BITMASK(ebx, 16, 0xff);
  186. info[ 6] = BITMASK(ebx, 24, 0xff);
  187. info[ 7] = BITMASK(ecx, 0, 0xff);
  188. info[ 8] = BITMASK(ecx, 8, 0xff);
  189. info[ 9] = BITMASK(ecx, 16, 0xff);
  190. info[10] = BITMASK(ecx, 24, 0xff);
  191. info[11] = BITMASK(edx, 0, 0xff);
  192. info[12] = BITMASK(edx, 8, 0xff);
  193. info[13] = BITMASK(edx, 16, 0xff);
  194. info[14] = BITMASK(edx, 24, 0xff);
  195. for (i = 0; i < 15; i++){
  196. switch (info[i]){
  197. case 0x3b :
  198. case 0x41 :
  199. case 0x79 :
  200. return 128;
  201. break;
  202. case 0x3c :
  203. case 0x42 :
  204. case 0x7a :
  205. case 0x7e :
  206. case 0x82 :
  207. return 256;
  208. break;
  209. case 0x43 :
  210. case 0x7b :
  211. case 0x7f :
  212. case 0x83 :
  213. case 0x86 :
  214. return 512;
  215. break;
  216. case 0x44 :
  217. case 0x78 :
  218. case 0x7c :
  219. case 0x84 :
  220. case 0x87 :
  221. return 1024;
  222. break;
  223. case 0x45 :
  224. case 0x7d :
  225. case 0x85 :
  226. return 2048;
  227. case 0x49 :
  228. return 4096;
  229. break;
  230. }
  231. }
  232. /* Never reached */
  233. return 0;
  234. #endif
  235. }
  236. void blas_set_parameter(void){
  237. int factor;
  238. #if defined(BULLDOZER) || defined(PILEDRIVER) || defined(SANDYBRIDGE) || defined(NEHALEM) || \
  239. defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined(ZEN) || \
  240. defined(SKYLAKEX) || defined(COOPERLAKE) || defined(SAPPHIRERAPIDS)
  241. int size = 16;
  242. #else
  243. int size = get_L2_size();
  244. #endif
  245. #if defined(CORE_KATMAI) || defined(CORE_COPPERMINE) || defined(CORE_BANIAS)
  246. size >>= 7;
  247. #if defined(CORE_BANIAS) && (HAVE_HIT > 1)
  248. sgemm_p = 64 / HAVE_HIT * size;
  249. dgemm_p = 32 / HAVE_HIT * size;
  250. cgemm_p = 32 / HAVE_HIT * size;
  251. zgemm_p = 16 / HAVE_HIT * size;
  252. #ifdef EXPRECISION
  253. qgemm_p = 16 / HAVE_HIT * size;
  254. xgemm_p = 8 / HAVE_HIT * size;
  255. #endif
  256. #ifdef QUAD_PRECISION
  257. qgemm_p = 8 / HAVE_HIT * size;
  258. xgemm_p = 4 / HAVE_HIT * size;
  259. #endif
  260. #else
  261. sgemm_p = 64 * size;
  262. dgemm_p = 32 * size;
  263. cgemm_p = 32 * size;
  264. zgemm_p = 16 * size;
  265. #ifdef EXPRECISION
  266. qgemm_p = 16 * size;
  267. xgemm_p = 8 * size;
  268. #endif
  269. #ifdef QUAD_PRECISION
  270. qgemm_p = 8 * size;
  271. xgemm_p = 4 * size;
  272. #endif
  273. #endif
  274. #endif
  275. #if defined(CORE_NORTHWOOD)
  276. size >>= 7;
  277. #ifdef ALLOC_HUGETLB
  278. sgemm_p = 128 * size;
  279. dgemm_p = 64 * size;
  280. cgemm_p = 64 * size;
  281. zgemm_p = 32 * size;
  282. #ifdef EXPRECISION
  283. qgemm_p = 32 * size;
  284. xgemm_p = 16 * size;
  285. #endif
  286. #ifdef QUAD_PRECISION
  287. qgemm_p = 16 * size;
  288. xgemm_p = 8 * size;
  289. #endif
  290. #else
  291. sgemm_p = 96 * size;
  292. dgemm_p = 48 * size;
  293. cgemm_p = 48 * size;
  294. zgemm_p = 24 * size;
  295. #ifdef EXPRECISION
  296. qgemm_p = 24 * size;
  297. xgemm_p = 12 * size;
  298. #endif
  299. #ifdef QUAD_PRECISION
  300. qgemm_p = 12 * size;
  301. xgemm_p = 6 * size;
  302. #endif
  303. #endif
  304. #endif
  305. #if defined(CORE_CORE2)
  306. size >>= 9;
  307. sgemm_p = 92 * size;
  308. dgemm_p = 46 * size;
  309. cgemm_p = 46 * size;
  310. zgemm_p = 23 * size;
  311. #ifdef EXPRECISION
  312. qgemm_p = 23 * size;
  313. xgemm_p = 11 * size;
  314. #endif
  315. #ifdef QUAD_PRECISION
  316. qgemm_p = 11 * size;
  317. xgemm_p = 5 * size;
  318. #endif
  319. #endif
  320. #if defined(PENRYN)
  321. size >>= 9;
  322. sgemm_p = 1024;
  323. dgemm_p = 512;
  324. cgemm_p = 512;
  325. zgemm_p = 256;
  326. #ifdef EXPRECISION
  327. qgemm_p = 256;
  328. xgemm_p = 128;
  329. #endif
  330. #ifdef QUAD_PRECISION
  331. qgemm_p = 21 * size + 4;
  332. xgemm_p = 10 * size + 2;
  333. #endif
  334. #endif
  335. #if defined(DUNNINGTON)
  336. size >>= 9;
  337. sgemm_p = 384;
  338. dgemm_p = 384;
  339. cgemm_p = 384;
  340. zgemm_p = 384;
  341. #ifdef EXPRECISION
  342. qgemm_p = 384;
  343. xgemm_p = 384;
  344. #endif
  345. #ifdef QUAD_PRECISION
  346. qgemm_p = 21 * size + 4;
  347. xgemm_p = 10 * size + 2;
  348. #endif
  349. #endif
  350. #if defined(NEHALEM)
  351. sgemm_p = 1024;
  352. dgemm_p = 512;
  353. cgemm_p = 512;
  354. zgemm_p = 256;
  355. #ifdef EXPRECISION
  356. qgemm_p = 256;
  357. xgemm_p = 128;
  358. #endif
  359. #endif
  360. #if defined(SANDYBRIDGE)
  361. sgemm_p = 1024;
  362. dgemm_p = 512;
  363. cgemm_p = 512;
  364. zgemm_p = 256;
  365. #ifdef EXPRECISION
  366. qgemm_p = 256;
  367. xgemm_p = 128;
  368. #endif
  369. #endif
  370. #if defined(CORE_PRESCOTT) || defined(GENERIC)
  371. size >>= 6;
  372. if (size > 16) size = 16;
  373. sgemm_p = 56 * size;
  374. dgemm_p = 28 * size;
  375. cgemm_p = 28 * size;
  376. zgemm_p = 14 * size;
  377. #ifdef EXPRECISION
  378. qgemm_p = 14 * size;
  379. xgemm_p = 7 * size;
  380. #endif
  381. #ifdef QUAD_PRECISION
  382. qgemm_p = 7 * size;
  383. xgemm_p = 3 * size;
  384. #endif
  385. #endif
  386. #if defined(CORE_OPTERON)
  387. sgemm_p = 224 + 14 * (size >> 5);
  388. dgemm_p = 112 + 14 * (size >> 6);
  389. cgemm_p = 116 + 14 * (size >> 6);
  390. zgemm_p = 58 + 14 * (size >> 7);
  391. #ifdef EXPRECISION
  392. qgemm_p = 58 + 14 * (size >> 7);
  393. xgemm_p = 29 + 14 * (size >> 8);
  394. #endif
  395. #ifdef QUAD_PRECISION
  396. qgemm_p = 29 + 14 * (size >> 8);
  397. xgemm_p = 15 + 14 * (size >> 9);
  398. #endif
  399. #endif
  400. #if defined(ATOM)
  401. size >>= 8;
  402. sgemm_p = 256;
  403. dgemm_p = 128;
  404. cgemm_p = 128;
  405. zgemm_p = 64;
  406. #ifdef EXPRECISION
  407. qgemm_p = 64;
  408. xgemm_p = 32;
  409. #endif
  410. #ifdef QUAD_PRECISION
  411. qgemm_p = 32;
  412. xgemm_p = 16;
  413. #endif
  414. #endif
  415. #if defined(CORE_BARCELONA) || defined(CORE_BOBCAT)
  416. size >>= 8;
  417. sgemm_p = 232 * size;
  418. dgemm_p = 116 * size;
  419. cgemm_p = 116 * size;
  420. zgemm_p = 58 * size;
  421. #ifdef EXPRECISION
  422. qgemm_p = 58 * size;
  423. xgemm_p = 26 * size;
  424. #endif
  425. #ifdef QUAD_PRECISION
  426. qgemm_p = 26 * size;
  427. xgemm_p = 13 * size;
  428. #endif
  429. #endif
  430. factor=openblas_block_factor();
  431. if (factor>0) {
  432. if (factor < 10) factor = 10;
  433. if (factor > 200) factor = 200;
  434. sgemm_p = ((long)((double)sgemm_p * (double)factor * 1.e-2)) & ~7L;
  435. dgemm_p = ((long)((double)dgemm_p * (double)factor * 1.e-2)) & ~7L;
  436. cgemm_p = ((long)((double)cgemm_p * (double)factor * 1.e-2)) & ~7L;
  437. zgemm_p = ((long)((double)zgemm_p * (double)factor * 1.e-2)) & ~7L;
  438. #ifdef EXPRECISION
  439. qgemm_p = ((long)((double)qgemm_p * (double)factor * 1.e-2)) & ~7L;
  440. xgemm_p = ((long)((double)xgemm_p * (double)factor * 1.e-2)) & ~7L;
  441. #endif
  442. }
  443. if (sgemm_p == 0) sgemm_p = 64;
  444. if (dgemm_p == 0) dgemm_p = 64;
  445. if (cgemm_p == 0) cgemm_p = 64;
  446. if (zgemm_p == 0) zgemm_p = 64;
  447. #ifdef EXPRECISION
  448. if (qgemm_p == 0) qgemm_p = 64;
  449. if (xgemm_p == 0) xgemm_p = 64;
  450. #endif
  451. #ifdef QUAD_PRECISION
  452. if (qgemm_p == 0) qgemm_p = 64;
  453. if (xgemm_p == 0) xgemm_p = 64;
  454. #endif
  455. sgemm_p = ((sgemm_p + SGEMM_UNROLL_M - 1)/SGEMM_UNROLL_M) * SGEMM_UNROLL_M;
  456. dgemm_p = ((dgemm_p + DGEMM_UNROLL_M - 1)/DGEMM_UNROLL_M) * DGEMM_UNROLL_M;
  457. cgemm_p = ((cgemm_p + CGEMM_UNROLL_M - 1)/CGEMM_UNROLL_M) * CGEMM_UNROLL_M;
  458. zgemm_p = ((zgemm_p + ZGEMM_UNROLL_M - 1)/ZGEMM_UNROLL_M) * ZGEMM_UNROLL_M;
  459. #ifdef QUAD_PRECISION
  460. qgemm_p = ((qgemm_p + QGEMM_UNROLL_M - 1)/QGEMM_UNROLL_M) * QGEMM_UNROLL_M;
  461. xgemm_p = ((xgemm_p + XGEMM_UNROLL_M - 1)/XGEMM_UNROLL_M) * XGEMM_UNROLL_M;
  462. #endif
  463. #ifdef BUILD_BFLOAT16
  464. sbgemm_r = (((BUFFER_SIZE - ((SBGEMM_P * SBGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SBGEMM_Q * 4)) - 15) & ~15;
  465. #endif
  466. sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q * 4)) - 15) & ~15;
  467. dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q * 8)) - 15) & ~15;
  468. cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q * 8)) - 15) & ~15;
  469. zgemm_r = (((BUFFER_SIZE - ((ZGEMM_P * ZGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (ZGEMM_Q * 16)) - 15) & ~15;
  470. #if defined(EXPRECISION) || defined(QUAD_PRECISION)
  471. qgemm_r = (((BUFFER_SIZE - ((QGEMM_P * QGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (QGEMM_Q * 16)) - 15) & ~15;
  472. xgemm_r = (((BUFFER_SIZE - ((XGEMM_P * XGEMM_Q * 32 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (XGEMM_Q * 32)) - 15) & ~15;
  473. #endif
  474. #if 0
  475. fprintf(stderr, "SGEMM ... %3d, %3d, %3d\n", SGEMM_P, SGEMM_Q, SGEMM_R);
  476. fprintf(stderr, "DGEMM ... %3d, %3d, %3d\n", DGEMM_P, DGEMM_Q, DGEMM_R);
  477. fprintf(stderr, "CGEMM ... %3d, %3d, %3d\n", CGEMM_P, CGEMM_Q, CGEMM_R);
  478. fprintf(stderr, "ZGEMM ... %3d, %3d, %3d\n", ZGEMM_P, ZGEMM_Q, ZGEMM_R);
  479. #endif
  480. return;
  481. }
  482. #if 0
  483. int get_current_cpu_info(void){
  484. int nlprocs, ncores, cmplegacy;
  485. int htt = 0;
  486. int apicid = 0;
  487. #if defined(CORE_PRESCOTT) || defined(CORE_OPTERON)
  488. int eax, ebx, ecx, edx;
  489. cpuid(1, &eax, &ebx, &ecx, &edx);
  490. nlprocs = BITMASK(ebx, 16, 0xff);
  491. apicid = BITMASK(ebx, 24, 0xff);
  492. htt = BITMASK(edx, 28, 0x01);
  493. #endif
  494. #if defined(CORE_PRESCOTT)
  495. cpuid(4, &eax, &ebx, &ecx, &edx);
  496. ncores = BITMASK(eax, 26, 0x3f);
  497. if (htt == 0) nlprocs = 0;
  498. #endif
  499. #if defined(CORE_OPTERON)
  500. cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
  501. ncores = BITMASK(ecx, 0, 0xff);
  502. cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
  503. cmplegacy = BITMASK(ecx, 1, 0x01);
  504. if (htt == 0) {
  505. nlprocs = 0;
  506. ncores = 0;
  507. cmplegacy = 0;
  508. }
  509. #endif
  510. ncores ++;
  511. fprintf(stderr, "APICID = %d Number of core = %d\n", apicid, ncores);
  512. return 0;
  513. }
  514. #endif
  515. #endif
  516. #if defined(ARCH_IA64)
  517. static inline BLASULONG cpuid(BLASULONG regnum){
  518. BLASULONG value;
  519. #ifndef __ECC
  520. asm ("mov %0=cpuid[%r1]" : "=r"(value) : "rO"(regnum));
  521. #else
  522. value = __getIndReg(_IA64_REG_INDR_CPUID, regnum);
  523. #endif
  524. return value;
  525. }
  526. #if 1
  527. void blas_set_parameter(void){
  528. BLASULONG cpuid3, size;
  529. cpuid3 = cpuid(3);
  530. size = BITMASK(cpuid3, 16, 0xff);
  531. sbgemm_p = 192 * (size + 1);
  532. sgemm_p = 192 * (size + 1);
  533. dgemm_p = 96 * (size + 1);
  534. cgemm_p = 96 * (size + 1);
  535. zgemm_p = 48 * (size + 1);
  536. #ifdef EXPRECISION
  537. qgemm_p = 64 * (size + 1);
  538. xgemm_p = 32 * (size + 1);
  539. #endif
  540. #ifdef QUAD_PRECISION
  541. qgemm_p = 32 * (size + 1);
  542. xgemm_p = 16 * (size + 1);
  543. #endif
  544. #ifdef BUILD_BFLOAT16
  545. sbgemm_r = (((BUFFER_SIZE - ((SBGEMM_P * SBGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SBGEMM_Q * 4)) - 15) & ~15;
  546. #endif
  547. sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q * 4)) - 15) & ~15;
  548. dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q * 8)) - 15) & ~15;
  549. cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q * 8)) - 15) & ~15;
  550. zgemm_r = (((BUFFER_SIZE - ((ZGEMM_P * ZGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (ZGEMM_Q * 16)) - 15) & ~15;
  551. #if defined(EXPRECISION) || defined(QUAD_PRECISION)
  552. qgemm_r = (((BUFFER_SIZE - ((QGEMM_P * QGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (QGEMM_Q * 16)) - 15) & ~15;
  553. xgemm_r = (((BUFFER_SIZE - ((XGEMM_P * XGEMM_Q * 32 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (XGEMM_Q * 32)) - 15) & ~15;
  554. #endif
  555. return;
  556. }
  557. #else
  558. #define IA64_SYS_NAME "/sys/devices/system/cpu/cpu0/cache/index3/size"
  559. #define IA64_PROC_NAME "/proc/pal/cpu0/cache_info"
  560. void blas_set_parameter(void){
  561. BLASULONG cpuid3;
  562. int size = 0;
  563. #if 1
  564. char buffer[128];
  565. FILE *infile;
  566. if ((infile = fopen(IA64_SYS_NAME, "r")) != NULL) {
  567. fgets(buffer, sizeof(buffer), infile);
  568. fclose(infile);
  569. size = atoi(buffer) / 1536;
  570. }
  571. if (size <= 0) {
  572. if ((infile = fopen(IA64_PROC_NAME, "r")) != NULL) {
  573. while(fgets(buffer, sizeof(buffer), infile) != NULL) {
  574. if ((!strncmp("Data/Instruction Cache level 3", buffer, 30))) break;
  575. }
  576. fgets(buffer, sizeof(buffer), infile);
  577. fclose(infile);
  578. *strstr(buffer, "bytes") = (char)NULL;
  579. size = atoi(strchr(buffer, ':') + 1) / 1572864;
  580. }
  581. }
  582. #endif
  583. /* The last resort */
  584. if (size <= 0) {
  585. cpuid3 = cpuid(3);
  586. size = BITMASK(cpuid3, 16, 0xff) + 1;
  587. }
  588. sgemm_p = 320 * size;
  589. dgemm_p = 160 * size;
  590. cgemm_p = 160 * size;
  591. zgemm_p = 80 * size;
  592. #ifdef EXPRECISION
  593. qgemm_p = 80 * size;
  594. xgemm_p = 40 * size;
  595. #endif
  596. sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q * 4)) - 15) & ~15;
  597. dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q * 8)) - 15) & ~15;
  598. cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q * 8)) - 15) & ~15;
  599. zgemm_r = (((BUFFER_SIZE - ((ZGEMM_P * ZGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (ZGEMM_Q * 16)) - 15) & ~15;
  600. #ifdef EXPRECISION
  601. qgemm_r = (((BUFFER_SIZE - ((QGEMM_P * QGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (QGEMM_Q * 16)) - 15) & ~15;
  602. xgemm_r = (((BUFFER_SIZE - ((XGEMM_P * XGEMM_Q * 32 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (XGEMM_Q * 32)) - 15) & ~15;
  603. #endif
  604. return;
  605. }
  606. #endif
  607. #endif
  608. #if defined(ARCH_MIPS64)
  609. void blas_set_parameter(void){
  610. #if defined(LOONGSON3R3) || defined(LOONGSON3R4)
  611. #ifdef SMP
  612. if(blas_num_threads == 1){
  613. #endif
  614. //single thread
  615. dgemm_r = 1024;
  616. #ifdef SMP
  617. }else{
  618. //multi thread
  619. dgemm_r = 200;
  620. }
  621. #endif
  622. #endif
  623. }
  624. #endif
  625. #if defined(ARCH_ARM64)
  626. void blas_set_parameter(void)
  627. {
  628. }
  629. #endif