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

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760
  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();
  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 SGEMM_P == sgemm_p
  60. BLASLONG sgemm_p = DEFAULT_GEMM_P;
  61. #else
  62. BLASLONG sgemm_p = SGEMM_P;
  63. #endif
  64. #if DGEMM_P == dgemm_p
  65. BLASLONG dgemm_p = DEFAULT_GEMM_P;
  66. #else
  67. BLASLONG dgemm_p = DGEMM_P;
  68. #endif
  69. #if CGEMM_P == cgemm_p
  70. BLASLONG cgemm_p = DEFAULT_GEMM_P;
  71. #else
  72. BLASLONG cgemm_p = CGEMM_P;
  73. #endif
  74. #if ZGEMM_P == zgemm_p
  75. BLASLONG zgemm_p = DEFAULT_GEMM_P;
  76. #else
  77. BLASLONG zgemm_p = ZGEMM_P;
  78. #endif
  79. #if SGEMM_Q == sgemm_q
  80. BLASLONG sgemm_q = DEFAULT_GEMM_Q;
  81. #else
  82. BLASLONG sgemm_q = SGEMM_Q;
  83. #endif
  84. #if DGEMM_Q == dgemm_q
  85. BLASLONG dgemm_q = DEFAULT_GEMM_Q;
  86. #else
  87. BLASLONG dgemm_q = DGEMM_Q;
  88. #endif
  89. #if CGEMM_Q == cgemm_q
  90. BLASLONG cgemm_q = DEFAULT_GEMM_Q;
  91. #else
  92. BLASLONG cgemm_q = CGEMM_Q;
  93. #endif
  94. #if ZGEMM_Q == zgemm_q
  95. BLASLONG zgemm_q = DEFAULT_GEMM_Q;
  96. #else
  97. BLASLONG zgemm_q = ZGEMM_Q;
  98. #endif
  99. #if SGEMM_R == sgemm_r
  100. BLASLONG sgemm_r = DEFAULT_GEMM_R;
  101. #else
  102. BLASLONG sgemm_r = SGEMM_R;
  103. #endif
  104. #if DGEMM_R == dgemm_r
  105. BLASLONG dgemm_r = DEFAULT_GEMM_R;
  106. #else
  107. BLASLONG dgemm_r = DGEMM_R;
  108. #endif
  109. #if CGEMM_R == cgemm_r
  110. BLASLONG cgemm_r = DEFAULT_GEMM_R;
  111. #else
  112. BLASLONG cgemm_r = CGEMM_R;
  113. #endif
  114. #if ZGEMM_R == zgemm_r
  115. BLASLONG zgemm_r = DEFAULT_GEMM_R;
  116. #else
  117. BLASLONG zgemm_r = ZGEMM_R;
  118. #endif
  119. #if defined(EXPRECISION) || defined(QUAD_PRECISION)
  120. #if QGEMM_P == qgemm_p
  121. BLASLONG qgemm_p = DEFAULT_GEMM_P;
  122. #else
  123. BLASLONG qgemm_p = QGEMM_P;
  124. #endif
  125. #if XGEMM_P == xgemm_p
  126. BLASLONG xgemm_p = DEFAULT_GEMM_P;
  127. #else
  128. BLASLONG xgemm_p = XGEMM_P;
  129. #endif
  130. #if QGEMM_Q == qgemm_q
  131. BLASLONG qgemm_q = DEFAULT_GEMM_Q;
  132. #else
  133. BLASLONG qgemm_q = QGEMM_Q;
  134. #endif
  135. #if XGEMM_Q == xgemm_q
  136. BLASLONG xgemm_q = DEFAULT_GEMM_Q;
  137. #else
  138. BLASLONG xgemm_q = XGEMM_Q;
  139. #endif
  140. #if QGEMM_R == qgemm_r
  141. BLASLONG qgemm_r = DEFAULT_GEMM_R;
  142. #else
  143. BLASLONG qgemm_r = QGEMM_R;
  144. #endif
  145. #if XGEMM_R == xgemm_r
  146. BLASLONG xgemm_r = DEFAULT_GEMM_R;
  147. #else
  148. BLASLONG xgemm_r = XGEMM_R;
  149. #endif
  150. #endif
  151. #if defined(ARCH_X86) || defined(ARCH_X86_64)
  152. int get_L2_size(void){
  153. int eax, ebx, ecx, edx;
  154. #if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER) || \
  155. defined(CORE_PRESCOTT) || defined(CORE_CORE2) || defined(PENRYN) || defined(DUNNINGTON) || \
  156. defined(CORE_NEHALEM) || defined(CORE_SANDYBRIDGE) || defined(ATOM) || defined(GENERIC) || \
  157. defined(PILEDRIVER) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR)
  158. cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
  159. return BITMASK(ecx, 16, 0xffff);
  160. #else
  161. int info[15];
  162. int i;
  163. cpuid(2, &eax, &ebx, &ecx, &edx);
  164. info[ 0] = BITMASK(eax, 8, 0xff);
  165. info[ 1] = BITMASK(eax, 16, 0xff);
  166. info[ 2] = BITMASK(eax, 24, 0xff);
  167. info[ 3] = BITMASK(ebx, 0, 0xff);
  168. info[ 4] = BITMASK(ebx, 8, 0xff);
  169. info[ 5] = BITMASK(ebx, 16, 0xff);
  170. info[ 6] = BITMASK(ebx, 24, 0xff);
  171. info[ 7] = BITMASK(ecx, 0, 0xff);
  172. info[ 8] = BITMASK(ecx, 8, 0xff);
  173. info[ 9] = BITMASK(ecx, 16, 0xff);
  174. info[10] = BITMASK(ecx, 24, 0xff);
  175. info[11] = BITMASK(edx, 0, 0xff);
  176. info[12] = BITMASK(edx, 8, 0xff);
  177. info[13] = BITMASK(edx, 16, 0xff);
  178. info[14] = BITMASK(edx, 24, 0xff);
  179. for (i = 0; i < 15; i++){
  180. switch (info[i]){
  181. case 0x3b :
  182. case 0x41 :
  183. case 0x79 :
  184. return 128;
  185. break;
  186. case 0x3c :
  187. case 0x42 :
  188. case 0x7a :
  189. case 0x7e :
  190. case 0x82 :
  191. return 256;
  192. break;
  193. case 0x43 :
  194. case 0x7b :
  195. case 0x7f :
  196. case 0x83 :
  197. case 0x86 :
  198. return 512;
  199. break;
  200. case 0x44 :
  201. case 0x78 :
  202. case 0x7c :
  203. case 0x84 :
  204. case 0x87 :
  205. return 1024;
  206. break;
  207. case 0x45 :
  208. case 0x7d :
  209. case 0x85 :
  210. return 2048;
  211. case 0x49 :
  212. return 4096;
  213. break;
  214. }
  215. }
  216. /* Never reached */
  217. return 0;
  218. #endif
  219. }
  220. void blas_set_parameter(void){
  221. int factor;
  222. #if defined(BULLDOZER) || defined(PILEDRIVER) || defined(SANDYBRIDGE) || defined(NEHALEM) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR)
  223. int size = 16;
  224. #else
  225. int size = get_L2_size();
  226. #endif
  227. #if defined(CORE_KATMAI) || defined(CORE_COPPERMINE) || defined(CORE_BANIAS)
  228. size >>= 7;
  229. #if defined(CORE_BANIAS) && (HAVE_HIT > 1)
  230. sgemm_p = 64 / HAVE_HIT * size;
  231. dgemm_p = 32 / HAVE_HIT * size;
  232. cgemm_p = 32 / HAVE_HIT * size;
  233. zgemm_p = 16 / HAVE_HIT * size;
  234. #ifdef EXPRECISION
  235. qgemm_p = 16 / HAVE_HIT * size;
  236. xgemm_p = 8 / HAVE_HIT * size;
  237. #endif
  238. #ifdef QUAD_PRECISION
  239. qgemm_p = 8 / HAVE_HIT * size;
  240. xgemm_p = 4 / HAVE_HIT * size;
  241. #endif
  242. #else
  243. sgemm_p = 64 * size;
  244. dgemm_p = 32 * size;
  245. cgemm_p = 32 * size;
  246. zgemm_p = 16 * size;
  247. #ifdef EXPRECISION
  248. qgemm_p = 16 * size;
  249. xgemm_p = 8 * size;
  250. #endif
  251. #ifdef QUAD_PRECISION
  252. qgemm_p = 8 * size;
  253. xgemm_p = 4 * size;
  254. #endif
  255. #endif
  256. #endif
  257. #if defined(CORE_NORTHWOOD)
  258. size >>= 7;
  259. #ifdef ALLOC_HUGETLB
  260. sgemm_p = 128 * size;
  261. dgemm_p = 64 * size;
  262. cgemm_p = 64 * size;
  263. zgemm_p = 32 * size;
  264. #ifdef EXPRECISION
  265. qgemm_p = 32 * size;
  266. xgemm_p = 16 * size;
  267. #endif
  268. #ifdef QUAD_PRECISION
  269. qgemm_p = 16 * size;
  270. xgemm_p = 8 * size;
  271. #endif
  272. #else
  273. sgemm_p = 96 * size;
  274. dgemm_p = 48 * size;
  275. cgemm_p = 48 * size;
  276. zgemm_p = 24 * size;
  277. #ifdef EXPRECISION
  278. qgemm_p = 24 * size;
  279. xgemm_p = 12 * size;
  280. #endif
  281. #ifdef QUAD_PRECISION
  282. qgemm_p = 12 * size;
  283. xgemm_p = 6 * size;
  284. #endif
  285. #endif
  286. #endif
  287. #if defined(CORE_CORE2)
  288. size >>= 9;
  289. sgemm_p = 92 * size;
  290. dgemm_p = 46 * size;
  291. cgemm_p = 46 * size;
  292. zgemm_p = 23 * size;
  293. #ifdef EXPRECISION
  294. qgemm_p = 23 * size;
  295. xgemm_p = 11 * size;
  296. #endif
  297. #ifdef QUAD_PRECISION
  298. qgemm_p = 11 * size;
  299. xgemm_p = 5 * size;
  300. #endif
  301. #endif
  302. #if defined(PENRYN)
  303. size >>= 9;
  304. sgemm_p = 1024;
  305. dgemm_p = 512;
  306. cgemm_p = 512;
  307. zgemm_p = 256;
  308. #ifdef EXPRECISION
  309. qgemm_p = 256;
  310. xgemm_p = 128;
  311. #endif
  312. #ifdef QUAD_PRECISION
  313. qgemm_p = 21 * size + 4;
  314. xgemm_p = 10 * size + 2;
  315. #endif
  316. #endif
  317. #if defined(DUNNINGTON)
  318. size >>= 9;
  319. sgemm_p = 384;
  320. dgemm_p = 384;
  321. cgemm_p = 384;
  322. zgemm_p = 384;
  323. #ifdef EXPRECISION
  324. qgemm_p = 384;
  325. xgemm_p = 384;
  326. #endif
  327. #ifdef QUAD_PRECISION
  328. qgemm_p = 21 * size + 4;
  329. xgemm_p = 10 * size + 2;
  330. #endif
  331. #endif
  332. #if defined(NEHALEM)
  333. sgemm_p = 1024;
  334. dgemm_p = 512;
  335. cgemm_p = 512;
  336. zgemm_p = 256;
  337. #ifdef EXPRECISION
  338. qgemm_p = 256;
  339. xgemm_p = 128;
  340. #endif
  341. #endif
  342. #if defined(SANDYBRIDGE)
  343. sgemm_p = 1024;
  344. dgemm_p = 512;
  345. cgemm_p = 512;
  346. zgemm_p = 256;
  347. #ifdef EXPRECISION
  348. qgemm_p = 256;
  349. xgemm_p = 128;
  350. #endif
  351. #endif
  352. #if defined(CORE_PRESCOTT) || defined(GENERIC)
  353. size >>= 6;
  354. if (size > 16) size = 16;
  355. sgemm_p = 56 * size;
  356. dgemm_p = 28 * size;
  357. cgemm_p = 28 * size;
  358. zgemm_p = 14 * size;
  359. #ifdef EXPRECISION
  360. qgemm_p = 14 * size;
  361. xgemm_p = 7 * size;
  362. #endif
  363. #ifdef QUAD_PRECISION
  364. qgemm_p = 7 * size;
  365. xgemm_p = 3 * size;
  366. #endif
  367. #endif
  368. #if defined(CORE_OPTERON)
  369. sgemm_p = 224 + 14 * (size >> 5);
  370. dgemm_p = 112 + 14 * (size >> 6);
  371. cgemm_p = 116 + 14 * (size >> 6);
  372. zgemm_p = 58 + 14 * (size >> 7);
  373. #ifdef EXPRECISION
  374. qgemm_p = 58 + 14 * (size >> 7);
  375. xgemm_p = 29 + 14 * (size >> 8);
  376. #endif
  377. #ifdef QUAD_PRECISION
  378. qgemm_p = 29 + 14 * (size >> 8);
  379. xgemm_p = 15 + 14 * (size >> 9);
  380. #endif
  381. #endif
  382. #if defined(ATOM)
  383. size >>= 8;
  384. sgemm_p = 256;
  385. dgemm_p = 128;
  386. cgemm_p = 128;
  387. zgemm_p = 64;
  388. #ifdef EXPRECISION
  389. qgemm_p = 64;
  390. xgemm_p = 32;
  391. #endif
  392. #ifdef QUAD_PRECISION
  393. qgemm_p = 32;
  394. xgemm_p = 16;
  395. #endif
  396. #endif
  397. #if defined(CORE_BARCELONA) || defined(CORE_BOBCAT)
  398. size >>= 8;
  399. sgemm_p = 232 * size;
  400. dgemm_p = 116 * size;
  401. cgemm_p = 116 * size;
  402. zgemm_p = 58 * size;
  403. #ifdef EXPRECISION
  404. qgemm_p = 58 * size;
  405. xgemm_p = 26 * size;
  406. #endif
  407. #ifdef QUAD_PRECISION
  408. qgemm_p = 26 * size;
  409. xgemm_p = 13 * size;
  410. #endif
  411. #endif
  412. factor=openblas_block_factor();
  413. if (factor>0) {
  414. if (factor < 10) factor = 10;
  415. if (factor > 200) factor = 200;
  416. sgemm_p = ((long)((double)sgemm_p * (double)factor * 1.e-2)) & ~7L;
  417. dgemm_p = ((long)((double)dgemm_p * (double)factor * 1.e-2)) & ~7L;
  418. cgemm_p = ((long)((double)cgemm_p * (double)factor * 1.e-2)) & ~7L;
  419. zgemm_p = ((long)((double)zgemm_p * (double)factor * 1.e-2)) & ~7L;
  420. #ifdef EXPRECISION
  421. qgemm_p = ((long)((double)qgemm_p * (double)factor * 1.e-2)) & ~7L;
  422. xgemm_p = ((long)((double)xgemm_p * (double)factor * 1.e-2)) & ~7L;
  423. #endif
  424. }
  425. if (sgemm_p == 0) sgemm_p = 64;
  426. if (dgemm_p == 0) dgemm_p = 64;
  427. if (cgemm_p == 0) cgemm_p = 64;
  428. if (zgemm_p == 0) zgemm_p = 64;
  429. #ifdef EXPRECISION
  430. if (qgemm_p == 0) qgemm_p = 64;
  431. if (xgemm_p == 0) xgemm_p = 64;
  432. #endif
  433. #ifdef QUAD_PRECISION
  434. if (qgemm_p == 0) qgemm_p = 64;
  435. if (xgemm_p == 0) xgemm_p = 64;
  436. #endif
  437. sgemm_p = ((sgemm_p + SGEMM_UNROLL_M - 1)/SGEMM_UNROLL_M) * SGEMM_UNROLL_M;
  438. dgemm_p = ((dgemm_p + DGEMM_UNROLL_M - 1)/DGEMM_UNROLL_M) * DGEMM_UNROLL_M;
  439. cgemm_p = ((cgemm_p + CGEMM_UNROLL_M - 1)/CGEMM_UNROLL_M) * CGEMM_UNROLL_M;
  440. zgemm_p = ((zgemm_p + ZGEMM_UNROLL_M - 1)/ZGEMM_UNROLL_M) * ZGEMM_UNROLL_M;
  441. #ifdef QUAD_PRECISION
  442. qgemm_p = ((qgemm_p + QGEMM_UNROLL_M - 1)/QGEMM_UNROLL_M) * QGEMM_UNROLL_M;
  443. xgemm_p = ((xgemm_p + XGEMM_UNROLL_M - 1)/XGEMM_UNROLL_M) * XGEMM_UNROLL_M;
  444. #endif
  445. sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q * 4)) - 15) & ~15;
  446. dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q * 8)) - 15) & ~15;
  447. cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q * 8)) - 15) & ~15;
  448. zgemm_r = (((BUFFER_SIZE - ((ZGEMM_P * ZGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (ZGEMM_Q * 16)) - 15) & ~15;
  449. #if defined(EXPRECISION) || defined(QUAD_PRECISION)
  450. qgemm_r = (((BUFFER_SIZE - ((QGEMM_P * QGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (QGEMM_Q * 16)) - 15) & ~15;
  451. xgemm_r = (((BUFFER_SIZE - ((XGEMM_P * XGEMM_Q * 32 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (XGEMM_Q * 32)) - 15) & ~15;
  452. #endif
  453. #if 0
  454. fprintf(stderr, "SGEMM ... %3d, %3d, %3d\n", SGEMM_P, SGEMM_Q, SGEMM_R);
  455. fprintf(stderr, "DGEMM ... %3d, %3d, %3d\n", DGEMM_P, DGEMM_Q, DGEMM_R);
  456. fprintf(stderr, "CGEMM ... %3d, %3d, %3d\n", CGEMM_P, CGEMM_Q, CGEMM_R);
  457. fprintf(stderr, "ZGEMM ... %3d, %3d, %3d\n", ZGEMM_P, ZGEMM_Q, ZGEMM_R);
  458. #endif
  459. return;
  460. }
  461. #if 0
  462. int get_current_cpu_info(void){
  463. int nlprocs, ncores, cmplegacy;
  464. int htt = 0;
  465. int apicid = 0;
  466. #if defined(CORE_PRESCOTT) || defined(CORE_OPTERON)
  467. int eax, ebx, ecx, edx;
  468. cpuid(1, &eax, &ebx, &ecx, &edx);
  469. nlprocs = BITMASK(ebx, 16, 0xff);
  470. apicid = BITMASK(ebx, 24, 0xff);
  471. htt = BITMASK(edx, 28, 0x01);
  472. #endif
  473. #if defined(CORE_PRESCOTT)
  474. cpuid(4, &eax, &ebx, &ecx, &edx);
  475. ncores = BITMASK(eax, 26, 0x3f);
  476. if (htt == 0) nlprocs = 0;
  477. #endif
  478. #if defined(CORE_OPTERON)
  479. cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
  480. ncores = BITMASK(ecx, 0, 0xff);
  481. cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
  482. cmplegacy = BITMASK(ecx, 1, 0x01);
  483. if (htt == 0) {
  484. nlprocs = 0;
  485. ncores = 0;
  486. cmplegacy = 0;
  487. }
  488. #endif
  489. ncores ++;
  490. fprintf(stderr, "APICID = %d Number of core = %d\n", apicid, ncores);
  491. return 0;
  492. }
  493. #endif
  494. #endif
  495. #if defined(ARCH_IA64)
  496. static inline BLASULONG cpuid(BLASULONG regnum){
  497. BLASULONG value;
  498. #ifndef __ECC
  499. asm ("mov %0=cpuid[%r1]" : "=r"(value) : "rO"(regnum));
  500. #else
  501. value = __getIndReg(_IA64_REG_INDR_CPUID, regnum);
  502. #endif
  503. return value;
  504. }
  505. #if 1
  506. void blas_set_parameter(void){
  507. BLASULONG cpuid3, size;
  508. cpuid3 = cpuid(3);
  509. size = BITMASK(cpuid3, 16, 0xff);
  510. sgemm_p = 192 * (size + 1);
  511. dgemm_p = 96 * (size + 1);
  512. cgemm_p = 96 * (size + 1);
  513. zgemm_p = 48 * (size + 1);
  514. #ifdef EXPRECISION
  515. qgemm_p = 64 * (size + 1);
  516. xgemm_p = 32 * (size + 1);
  517. #endif
  518. #ifdef QUAD_PRECISION
  519. qgemm_p = 32 * (size + 1);
  520. xgemm_p = 16 * (size + 1);
  521. #endif
  522. sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q * 4)) - 15) & ~15;
  523. dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q * 8)) - 15) & ~15;
  524. cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q * 8)) - 15) & ~15;
  525. zgemm_r = (((BUFFER_SIZE - ((ZGEMM_P * ZGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (ZGEMM_Q * 16)) - 15) & ~15;
  526. #if defined(EXPRECISION) || defined(QUAD_PRECISION)
  527. qgemm_r = (((BUFFER_SIZE - ((QGEMM_P * QGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (QGEMM_Q * 16)) - 15) & ~15;
  528. xgemm_r = (((BUFFER_SIZE - ((XGEMM_P * XGEMM_Q * 32 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (XGEMM_Q * 32)) - 15) & ~15;
  529. #endif
  530. return;
  531. }
  532. #else
  533. #define IA64_SYS_NAME "/sys/devices/system/cpu/cpu0/cache/index3/size"
  534. #define IA64_PROC_NAME "/proc/pal/cpu0/cache_info"
  535. void blas_set_parameter(void){
  536. BLASULONG cpuid3;
  537. int size = 0;
  538. #if 1
  539. char buffer[128];
  540. FILE *infile;
  541. if ((infile = fopen(IA64_SYS_NAME, "r")) != NULL) {
  542. fgets(buffer, sizeof(buffer), infile);
  543. fclose(infile);
  544. size = atoi(buffer) / 1536;
  545. }
  546. if (size <= 0) {
  547. if ((infile = fopen(IA64_PROC_NAME, "r")) != NULL) {
  548. while(fgets(buffer, sizeof(buffer), infile) != NULL) {
  549. if ((!strncmp("Data/Instruction Cache level 3", buffer, 30))) break;
  550. }
  551. fgets(buffer, sizeof(buffer), infile);
  552. fclose(infile);
  553. *strstr(buffer, "bytes") = (char)NULL;
  554. size = atoi(strchr(buffer, ':') + 1) / 1572864;
  555. }
  556. }
  557. #endif
  558. /* The last resort */
  559. if (size <= 0) {
  560. cpuid3 = cpuid(3);
  561. size = BITMASK(cpuid3, 16, 0xff) + 1;
  562. }
  563. sgemm_p = 320 * size;
  564. dgemm_p = 160 * size;
  565. cgemm_p = 160 * size;
  566. zgemm_p = 80 * size;
  567. #ifdef EXPRECISION
  568. qgemm_p = 80 * size;
  569. xgemm_p = 40 * size;
  570. #endif
  571. sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q * 4)) - 15) & ~15;
  572. dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q * 8)) - 15) & ~15;
  573. cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q * 8)) - 15) & ~15;
  574. zgemm_r = (((BUFFER_SIZE - ((ZGEMM_P * ZGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (ZGEMM_Q * 16)) - 15) & ~15;
  575. #ifdef EXPRECISION
  576. qgemm_r = (((BUFFER_SIZE - ((QGEMM_P * QGEMM_Q * 16 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (QGEMM_Q * 16)) - 15) & ~15;
  577. xgemm_r = (((BUFFER_SIZE - ((XGEMM_P * XGEMM_Q * 32 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (XGEMM_Q * 32)) - 15) & ~15;
  578. #endif
  579. return;
  580. }
  581. #endif
  582. #endif
  583. #if defined(ARCH_MIPS64)
  584. void blas_set_parameter(void){
  585. #if defined(LOONGSON3A)
  586. #ifdef SMP
  587. if(blas_num_threads == 1){
  588. #endif
  589. //single thread
  590. dgemm_r = 1024;
  591. #ifdef SMP
  592. }else{
  593. //multi thread
  594. dgemm_r = 200;
  595. }
  596. #endif
  597. #endif
  598. #if defined(LOONGSON3B)
  599. #ifdef SMP
  600. if(blas_num_threads == 1 || blas_num_threads == 2){
  601. #endif
  602. //single thread
  603. dgemm_r = 640;
  604. #ifdef SMP
  605. }else{
  606. //multi thread
  607. dgemm_r = 160;
  608. }
  609. #endif
  610. #endif
  611. }
  612. #endif
  613. #if defined(ARCH_ARM64)
  614. #if defined(VULCAN) || defined(THUNDERX2T99)
  615. unsigned long dgemm_prefetch_size_a;
  616. unsigned long dgemm_prefetch_size_b;
  617. unsigned long dgemm_prefetch_size_c;
  618. #endif
  619. void blas_set_parameter(void)
  620. {
  621. #if defined(VULCAN) || defined(THUNDERX2T99)
  622. dgemm_p = 160;
  623. dgemm_q = 128;
  624. dgemm_r = 4096;
  625. sgemm_p = 128;
  626. sgemm_q = 352;
  627. sgemm_r = 4096;
  628. cgemm_p = 128;
  629. cgemm_q = 224;
  630. cgemm_r = 4096;
  631. dgemm_prefetch_size_a = 3584;
  632. dgemm_prefetch_size_b = 512;
  633. dgemm_prefetch_size_c = 128;
  634. #endif
  635. }
  636. #endif