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: 3597827c93
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '3597827c93'
${ noResults }
OpenBLAS/kernel/generic/trmmkernel_16x2.c

1154 lines
18 kB
Raw Blame History 

  1. #include "common.h"

  2. 

  3. int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc ,BLASLONG offset)

  4. {

  5. 

  6.    BLASLONG i,j,k;

  7.    FLOAT *C0,*C1,*ptrba,*ptrbb;

  8. 

  9.    FLOAT res0_0;

  10.    FLOAT res0_1;

  11.    FLOAT res0_2;

  12.    FLOAT res0_3;

  13.    FLOAT res0_4;

  14.    FLOAT res0_5;

  15.    FLOAT res0_6;

  16.    FLOAT res0_7;

  17. 

  18.    FLOAT res0_8;

  19.    FLOAT res0_9;

  20.    FLOAT res0_10;

  21.    FLOAT res0_11;

  22.    FLOAT res0_12;

  23.    FLOAT res0_13;

  24.    FLOAT res0_14;

  25.    FLOAT res0_15;

  26. 

  27.    FLOAT res1_0;

  28.    FLOAT res1_1;

  29.    FLOAT res1_2;

  30.    FLOAT res1_3;

  31.    FLOAT res1_4;

  32.    FLOAT res1_5;

  33.    FLOAT res1_6;

  34.    FLOAT res1_7;

  35. 

  36.    FLOAT res1_8;

  37.    FLOAT res1_9;

  38.    FLOAT res1_10;

  39.    FLOAT res1_11;

  40.    FLOAT res1_12;

  41.    FLOAT res1_13;

  42.    FLOAT res1_14;

  43.    FLOAT res1_15;

  44. 

  45.    FLOAT a0;

  46.    FLOAT a1;

  47. 

  48.    FLOAT b0;

  49.    FLOAT b1;

  50. 

  51.    BLASLONG off, temp;

  52. 

  53. #if !defined(LEFT)

  54.    off = -offset;

  55. #else

  56.    off = 0;

  57. #endif

  58. 

  59. 

  60. 

  61.    for (j=0; j<bn/2; j+=1)

  62.    {

  63.         C0 = C;

  64.         C1 = C0+ldc;

  65. 

  66. #if defined(TRMMKERNEL) && defined(LEFT)

  67. 	off = offset;

  68. #endif

  69. 

  70. 

  71.         ptrba = ba;

  72. 

  73. 

  74.         for (i=0; i<bm/16; i+=1)

  75.         {

  76. 

  77. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  78. 		ptrbb = bb;

  79. #else

  80. 		ptrba += off*16;

  81. 		ptrbb = bb + off*2;

  82. #endif

  83. 

  84. 		res0_0 = 0;

  85. 		res0_1 = 0;

  86. 		res0_2 = 0;

  87. 		res0_3 = 0;

  88. 		res0_4 = 0;

  89. 		res0_5 = 0;

  90. 		res0_6 = 0;

  91. 		res0_7 = 0;

  92. 

  93. 		res0_8  = 0;

  94. 		res0_9  = 0;

  95. 		res0_10 = 0;

  96. 		res0_11 = 0;

  97. 		res0_12 = 0;

  98. 		res0_13 = 0;

  99. 		res0_14 = 0;

  100. 		res0_15 = 0;

  101. 

  102. 		res1_0 = 0;

  103. 		res1_1 = 0;

  104. 		res1_2 = 0;

  105. 		res1_3 = 0;

  106. 		res1_4 = 0;

  107. 		res1_5 = 0;

  108. 		res1_6 = 0;

  109. 		res1_7 = 0;

  110. 

  111. 		res1_8  = 0;

  112. 		res1_9  = 0;

  113. 		res1_10 = 0;

  114. 		res1_11 = 0;

  115. 		res1_12 = 0;

  116. 		res1_13 = 0;

  117. 		res1_14 = 0;

  118. 		res1_15 = 0;

  119. 

  120. 

  121. 

  122. 

  123. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  124. 		temp = bk-off;

  125. #elif defined(LEFT)

  126. 		temp = off+16;	// number of values in A

  127. #else

  128. 		temp = off+2;	// number of values in B

  129. #endif

  130. 

  131. 		for (k=0; k<temp; k++)

  132.                 {

  133. 			b0 = ptrbb[0];

  134. 			b1 = ptrbb[1];

  135. 

  136. 			a0 = ptrba[0];

  137. 			res0_0 += a0*b0;

  138. 			res1_0 += a0*b1;

  139. 

  140. 			a1 = ptrba[1];

  141. 			res0_1 += a1*b0;

  142. 			res1_1 += a1*b1;

  143. 

  144. 			a0 = ptrba[2];

  145. 			res0_2 += a0*b0;

  146. 			res1_2 += a0*b1;

  147. 

  148. 			a1 = ptrba[3];

  149. 			res0_3 += a1*b0;

  150. 			res1_3 += a1*b1;

  151. 

  152. 			a0 = ptrba[4];

  153. 			res0_4 += a0*b0;

  154. 			res1_4 += a0*b1;

  155. 

  156. 			a1 = ptrba[5];

  157. 			res0_5 += a1*b0;

  158. 			res1_5 += a1*b1;

  159. 

  160. 			a0 = ptrba[6];

  161. 			res0_6 += a0*b0;

  162. 			res1_6 += a0*b1;

  163. 

  164. 			a1 = ptrba[7];

  165. 			res0_7 += a1*b0;

  166. 			res1_7 += a1*b1;

  167. 

  168. 			a0 = ptrba[8];

  169. 			res0_8 += a0*b0;

  170. 			res1_8 += a0*b1;

  171. 

  172. 			a1 = ptrba[9];

  173. 			res0_9 += a1*b0;

  174. 			res1_9 += a1*b1;

  175. 

  176. 			a0 = ptrba[10];

  177. 			res0_10 += a0*b0;

  178. 			res1_10 += a0*b1;

  179. 

  180. 			a1 = ptrba[11];

  181. 			res0_11 += a1*b0;

  182. 			res1_11 += a1*b1;

  183. 

  184. 			a0 = ptrba[12];

  185. 			res0_12 += a0*b0;

  186. 			res1_12 += a0*b1;

  187. 

  188. 			a1 = ptrba[13];

  189. 			res0_13 += a1*b0;

  190. 			res1_13 += a1*b1;

  191. 

  192. 			a0 = ptrba[14];

  193. 			res0_14 += a0*b0;

  194. 			res1_14 += a0*b1;

  195. 

  196. 			a1 = ptrba[15];

  197. 			res0_15 += a1*b0;

  198. 			res1_15 += a1*b1;

  199. 

  200. 

  201. 			ptrba = ptrba+16;

  202. 			ptrbb = ptrbb+2;

  203.                 }

  204. 

  205. 		res0_0 *= alpha;

  206. 		res0_1 *= alpha;

  207. 		res0_2 *= alpha;

  208. 		res0_3 *= alpha;

  209. 		res0_4 *= alpha;

  210. 		res0_5 *= alpha;

  211. 		res0_6 *= alpha;

  212. 		res0_7 *= alpha;

  213. 

  214. 		res0_8  *= alpha;

  215. 		res0_9  *= alpha;

  216. 		res0_10 *= alpha;

  217. 		res0_11 *= alpha;

  218. 		res0_12 *= alpha;

  219. 		res0_13 *= alpha;

  220. 		res0_14 *= alpha;

  221. 		res0_15 *= alpha;

  222. 

  223. 		res1_0 *= alpha;

  224. 		res1_1 *= alpha;

  225. 		res1_2 *= alpha;

  226. 		res1_3 *= alpha;

  227. 		res1_4 *= alpha;

  228. 		res1_5 *= alpha;

  229. 		res1_6 *= alpha;

  230. 		res1_7 *= alpha;

  231. 

  232. 		res1_8  *= alpha;

  233. 		res1_9  *= alpha;

  234. 		res1_10 *= alpha;

  235. 		res1_11 *= alpha;

  236. 		res1_12 *= alpha;

  237. 		res1_13 *= alpha;

  238. 		res1_14 *= alpha;

  239. 		res1_15 *= alpha;

  240. 

  241. 		C0[0] = res0_0;

  242. 		C0[1] = res0_1;

  243. 		C0[2] = res0_2;

  244. 		C0[3] = res0_3;

  245. 		C0[4] = res0_4;

  246. 		C0[5] = res0_5;

  247. 		C0[6] = res0_6;

  248. 		C0[7] = res0_7;

  249. 

  250. 		C0[8]  = res0_8;

  251. 		C0[9]  = res0_9;

  252. 		C0[10] = res0_10;

  253. 		C0[11] = res0_11;

  254. 		C0[12] = res0_12;

  255. 		C0[13] = res0_13;

  256. 		C0[14] = res0_14;

  257. 		C0[15] = res0_15;

  258. 

  259. 		C1[0] = res1_0;

  260. 		C1[1] = res1_1;

  261. 		C1[2] = res1_2;

  262. 		C1[3] = res1_3;

  263. 		C1[4] = res1_4;

  264. 		C1[5] = res1_5;

  265. 		C1[6] = res1_6;

  266. 		C1[7] = res1_7;

  267. 

  268. 		C1[8]  = res1_8;

  269. 		C1[9]  = res1_9;

  270. 		C1[10] = res1_10;

  271. 		C1[11] = res1_11;

  272. 		C1[12] = res1_12;

  273. 		C1[13] = res1_13;

  274. 		C1[14] = res1_14;

  275. 		C1[15] = res1_15;

  276. 

  277. 

  278. 

  279. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  280. 		temp = bk - off;

  281. #ifdef LEFT

  282. 		temp -= 16; // number of values in A

  283. #else

  284. 		temp -= 2; // number of values in B

  285. #endif

  286. 		ptrba += temp*16;

  287. 		ptrbb += temp*2;

  288. #endif

  289. 

  290. #ifdef LEFT

  291. 		off += 16; // number of values in A

  292. #endif

  293. 

  294. 		C0 = C0+16;

  295. 		C1 = C1+16;

  296. 	}

  297. 

  298. 

  299. 

  300. 

  301.         if ( bm & 8)

  302.         {

  303. 

  304. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  305. 		ptrbb = bb;

  306. #else

  307. 		ptrba += off*8;

  308. 		ptrbb = bb + off*2;

  309. #endif

  310. 

  311. 		res0_0 = 0;

  312. 		res0_1 = 0;

  313. 		res0_2 = 0;

  314. 		res0_3 = 0;

  315. 		res0_4 = 0;

  316. 		res0_5 = 0;

  317. 		res0_6 = 0;

  318. 		res0_7 = 0;

  319. 

  320. 		res1_0 = 0;

  321. 		res1_1 = 0;

  322. 		res1_2 = 0;

  323. 		res1_3 = 0;

  324. 		res1_4 = 0;

  325. 		res1_5 = 0;

  326. 		res1_6 = 0;

  327. 		res1_7 = 0;

  328. 

  329. 

  330. 

  331. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  332. 		temp = bk-off;

  333. #elif defined(LEFT)

  334. 		temp = off+8;	// number of values in A

  335. #else

  336. 		temp = off+2;	// number of values in B

  337. #endif

  338. 

  339. 		for (k=0; k<temp; k++)

  340.                 {

  341. 			b0 = ptrbb[0];

  342. 			b1 = ptrbb[1];

  343. 

  344. 			a0 = ptrba[0];

  345. 			res0_0 += a0*b0;

  346. 			res1_0 += a0*b1;

  347. 

  348. 			a1 = ptrba[1];

  349. 			res0_1 += a1*b0;

  350. 			res1_1 += a1*b1;

  351. 

  352. 			a0 = ptrba[2];

  353. 			res0_2 += a0*b0;

  354. 			res1_2 += a0*b1;

  355. 

  356. 			a1 = ptrba[3];

  357. 			res0_3 += a1*b0;

  358. 			res1_3 += a1*b1;

  359. 

  360. 			a0 = ptrba[4];

  361. 			res0_4 += a0*b0;

  362. 			res1_4 += a0*b1;

  363. 

  364. 			a1 = ptrba[5];

  365. 			res0_5 += a1*b0;

  366. 			res1_5 += a1*b1;

  367. 

  368. 			a0 = ptrba[6];

  369. 			res0_6 += a0*b0;

  370. 			res1_6 += a0*b1;

  371. 

  372. 			a1 = ptrba[7];

  373. 			res0_7 += a1*b0;

  374. 			res1_7 += a1*b1;

  375. 

  376. 			ptrba = ptrba+8;

  377. 			ptrbb = ptrbb+2;

  378.                 }

  379. 

  380. 		res0_0 *= alpha;

  381. 		res0_1 *= alpha;

  382. 		res0_2 *= alpha;

  383. 		res0_3 *= alpha;

  384. 		res0_4 *= alpha;

  385. 		res0_5 *= alpha;

  386. 		res0_6 *= alpha;

  387. 		res0_7 *= alpha;

  388. 

  389. 		res1_0 *= alpha;

  390. 		res1_1 *= alpha;

  391. 		res1_2 *= alpha;

  392. 		res1_3 *= alpha;

  393. 		res1_4 *= alpha;

  394. 		res1_5 *= alpha;

  395. 		res1_6 *= alpha;

  396. 		res1_7 *= alpha;

  397. 

  398. 		C0[0] = res0_0;

  399. 		C0[1] = res0_1;

  400. 		C0[2] = res0_2;

  401. 		C0[3] = res0_3;

  402. 		C0[4] = res0_4;

  403. 		C0[5] = res0_5;

  404. 		C0[6] = res0_6;

  405. 		C0[7] = res0_7;

  406. 

  407. 		C1[0] = res1_0;

  408. 		C1[1] = res1_1;

  409. 		C1[2] = res1_2;

  410. 		C1[3] = res1_3;

  411. 		C1[4] = res1_4;

  412. 		C1[5] = res1_5;

  413. 		C1[6] = res1_6;

  414. 		C1[7] = res1_7;

  415. 

  416. 

  417. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  418. 		temp = bk - off;

  419. #ifdef LEFT

  420. 		temp -= 8; // number of values in A

  421. #else

  422. 		temp -= 2; // number of values in B

  423. #endif

  424. 		ptrba += temp*8;

  425. 		ptrbb += temp*2;

  426. #endif

  427. 

  428. #ifdef LEFT

  429. 		off += 8; // number of values in A

  430. #endif

  431. 

  432. 		C0 = C0+8;

  433. 		C1 = C1+8;

  434. 	}

  435. 

  436. 	if ( bm & 4 )

  437. 	{

  438. 

  439. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  440. 		ptrbb = bb;

  441. #else

  442. 		ptrba += off*4;

  443. 		ptrbb = bb + off*2;

  444. #endif

  445. 

  446. 		res0_0 = 0;

  447. 		res0_1 = 0;

  448. 		res0_2 = 0;

  449. 		res0_3 = 0;

  450. 

  451. 		res1_0 = 0;

  452. 		res1_1 = 0;

  453. 		res1_2 = 0;

  454. 		res1_3 = 0;

  455. 

  456. 

  457. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  458. 		temp = bk-off;

  459. #elif defined(LEFT)

  460. 		temp = off+4;	// number of values in A

  461. #else

  462. 		temp = off+2;	// number of values in B

  463. #endif

  464. 

  465. 		for (k=0; k<temp; k++)

  466.                 {

  467. 			b0 = ptrbb[0];

  468. 			b1 = ptrbb[1];

  469. 

  470. 			a0 = ptrba[0];

  471. 			res0_0 += a0*b0;

  472. 			res1_0 += a0*b1;

  473. 

  474. 			a1 = ptrba[1];

  475. 			res0_1 += a1*b0;

  476. 			res1_1 += a1*b1;

  477. 

  478. 			a0 = ptrba[2];

  479. 			res0_2 += a0*b0;

  480. 			res1_2 += a0*b1;

  481. 

  482. 			a1 = ptrba[3];

  483. 			res0_3 += a1*b0;

  484. 			res1_3 += a1*b1;

  485. 

  486. 			ptrba = ptrba+4;

  487. 			ptrbb = ptrbb+2;

  488.                 }

  489. 

  490. 		res0_0 *= alpha;

  491. 		res0_1 *= alpha;

  492. 		res0_2 *= alpha;

  493. 		res0_3 *= alpha;

  494. 

  495. 		res1_0 *= alpha;

  496. 		res1_1 *= alpha;

  497. 		res1_2 *= alpha;

  498. 		res1_3 *= alpha;

  499. 

  500. 		C0[0] = res0_0;

  501. 		C0[1] = res0_1;

  502. 		C0[2] = res0_2;

  503. 		C0[3] = res0_3;

  504. 

  505. 		C1[0] = res1_0;

  506. 		C1[1] = res1_1;

  507. 		C1[2] = res1_2;

  508. 		C1[3] = res1_3;

  509. 

  510. 

  511. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  512. 		temp = bk - off;

  513. #ifdef LEFT

  514. 		temp -= 4; // number of values in A

  515. #else

  516. 		temp -= 2; // number of values in B

  517. #endif

  518. 		ptrba += temp*4;

  519. 		ptrbb += temp*2;

  520. #endif

  521. 

  522. #ifdef LEFT

  523. 		off += 4; // number of values in A

  524. #endif

  525. 

  526. 		C0 = C0+4;

  527. 		C1 = C1+4;

  528. 

  529. 	}

  530. 

  531. 	if ( bm & 2 )

  532. 	{

  533. 

  534. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  535. 		ptrbb = bb;

  536. #else

  537. 		ptrba += off*2;

  538. 		ptrbb = bb + off*2;

  539. #endif

  540. 

  541. 		res0_0 = 0;

  542. 		res0_1 = 0;

  543. 

  544. 		res1_0 = 0;

  545. 		res1_1 = 0;

  546. 

  547. 

  548. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  549. 		temp = bk-off;

  550. #elif defined(LEFT)

  551. 		temp = off+2;	// number of values in A

  552. #else

  553. 		temp = off+2;	// number of values in B

  554. #endif

  555. 

  556. 		for (k=0; k<temp; k++)

  557.                 {

  558. 			b0 = ptrbb[0];

  559. 			b1 = ptrbb[1];

  560. 

  561. 			a0 = ptrba[0];

  562. 			res0_0 += a0*b0;

  563. 			res1_0 += a0*b1;

  564. 

  565. 			a1 = ptrba[1];

  566. 			res0_1 += a1*b0;

  567. 			res1_1 += a1*b1;

  568. 

  569. 			ptrba = ptrba+2;

  570. 			ptrbb = ptrbb+2;

  571.                 }

  572. 

  573. 		res0_0 *= alpha;

  574. 		res0_1 *= alpha;

  575. 

  576. 		res1_0 *= alpha;

  577. 		res1_1 *= alpha;

  578. 

  579. 		C0[0] = res0_0;

  580. 		C0[1] = res0_1;

  581. 

  582. 		C1[0] = res1_0;

  583. 		C1[1] = res1_1;

  584. 

  585. 

  586. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  587. 		temp = bk - off;

  588. #ifdef LEFT

  589. 		temp -= 2; // number of values in A

  590. #else

  591. 		temp -= 2; // number of values in B

  592. #endif

  593. 		ptrba += temp*2;

  594. 		ptrbb += temp*2;

  595. #endif

  596. 

  597. #ifdef LEFT

  598. 		off += 2; // number of values in A

  599. #endif

  600. 

  601. 		C0 = C0+2;

  602. 		C1 = C1+2;

  603. 

  604. 	}

  605. 

  606. 	if ( bm & 1 )

  607. 	{

  608. 

  609. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  610. 		ptrbb = bb;

  611. #else

  612. 		ptrba += off*1;

  613. 		ptrbb = bb + off*2;

  614. #endif

  615. 

  616. 		res0_0 = 0;

  617. 

  618. 		res1_0 = 0;

  619. 

  620. 

  621. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  622. 		temp = bk-off;

  623. #elif defined(LEFT)

  624. 		temp = off+1;	// number of values in A

  625. #else

  626. 		temp = off+2;	// number of values in B

  627. #endif

  628. 

  629. 		for (k=0; k<temp; k++)

  630.                 {

  631. 			b0 = ptrbb[0];

  632. 			b1 = ptrbb[1];

  633. 

  634. 			a0 = ptrba[0];

  635. 			res0_0 += a0*b0;

  636. 			res1_0 += a0*b1;

  637. 

  638. 			ptrba = ptrba+1;

  639. 			ptrbb = ptrbb+2;

  640.                 }

  641. 

  642. 		res0_0 *= alpha;

  643. 

  644. 		res1_0 *= alpha;

  645. 

  646. 		C0[0] = res0_0;

  647. 

  648. 		C1[0] = res1_0;

  649. 

  650. 

  651. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  652. 		temp = bk - off;

  653. #ifdef LEFT

  654. 		temp -= 1; // number of values in A

  655. #else

  656. 		temp -= 2; // number of values in B

  657. #endif

  658. 		ptrba += temp*1;

  659. 		ptrbb += temp*2;

  660. #endif

  661. 

  662. #ifdef LEFT

  663. 		off += 1; // number of values in A

  664. #endif

  665. 

  666. 		C0 = C0+1;

  667. 		C1 = C1+1;

  668. 

  669. 	}

  670. 

  671. 

  672. #if defined(TRMMKERNEL) && !defined(LEFT)

  673. 		off += 2;

  674. #endif

  675. 

  676.         k = (bk<<1);

  677.         bb = bb+k;

  678.         i = (ldc<<1);

  679.         C = C+i;

  680.     }

  681. 

  682. 

  683. 

  684. 

  685. 

  686. 

  687. 

  688.    for (j=0; j<(bn&1); j+=1)

  689.    {

  690.         C0 = C;

  691. 

  692. #if defined(TRMMKERNEL) &&  defined(LEFT)

  693. 	off = offset;

  694. #endif

  695. 

  696.         ptrba = ba;

  697. 

  698. 

  699.         for (i=0; i<bm/16; i+=1)

  700.         {

  701. 

  702. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  703. 		ptrbb = bb;

  704. #else

  705. 		ptrba += off*16;

  706. 		ptrbb = bb + off*1;

  707. #endif

  708. 

  709. 		res0_0 = 0;

  710. 		res0_1 = 0;

  711. 		res0_2 = 0;

  712. 		res0_3 = 0;

  713. 		res0_4 = 0;

  714. 		res0_5 = 0;

  715. 		res0_6 = 0;

  716. 		res0_7 = 0;

  717. 

  718. 		res0_8  = 0;

  719. 		res0_9  = 0;

  720. 		res0_10 = 0;

  721. 		res0_11 = 0;

  722. 		res0_12 = 0;

  723. 		res0_13 = 0;

  724. 		res0_14 = 0;

  725. 		res0_15 = 0;

  726. 

  727. 

  728. 

  729. 

  730. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  731. 		temp = bk-off;

  732. #elif defined(LEFT)

  733. 		temp = off+16;	// number of values in A

  734. #else

  735. 		temp = off+1;	// number of values in B

  736. #endif

  737. 

  738. 		for (k=0; k<temp; k++)

  739.                 {

  740. 			b0 = ptrbb[0];

  741. 

  742. 			a0 = ptrba[0];

  743. 			res0_0 += a0*b0;

  744. 

  745. 			a1 = ptrba[1];

  746. 			res0_1 += a1*b0;

  747. 

  748. 			a0 = ptrba[2];

  749. 			res0_2 += a0*b0;

  750. 

  751. 			a1 = ptrba[3];

  752. 			res0_3 += a1*b0;

  753. 

  754. 			a0 = ptrba[4];

  755. 			res0_4 += a0*b0;

  756. 

  757. 			a1 = ptrba[5];

  758. 			res0_5 += a1*b0;

  759. 

  760. 			a0 = ptrba[6];

  761. 			res0_6 += a0*b0;

  762. 

  763. 			a1 = ptrba[7];

  764. 			res0_7 += a1*b0;

  765. 

  766. 			a0 = ptrba[8];

  767. 			res0_8 += a0*b0;

  768. 

  769. 			a1 = ptrba[9];

  770. 			res0_9 += a1*b0;

  771. 

  772. 			a0 = ptrba[10];

  773. 			res0_10 += a0*b0;

  774. 

  775. 			a1 = ptrba[11];

  776. 			res0_11 += a1*b0;

  777. 

  778. 			a0 = ptrba[12];

  779. 			res0_12 += a0*b0;

  780. 

  781. 			a1 = ptrba[13];

  782. 			res0_13 += a1*b0;

  783. 

  784. 			a0 = ptrba[14];

  785. 			res0_14 += a0*b0;

  786. 

  787. 			a1 = ptrba[15];

  788. 			res0_15 += a1*b0;

  789. 

  790. 

  791. 			ptrba = ptrba+16;

  792. 			ptrbb = ptrbb+1;

  793.                 }

  794. 

  795. 		res0_0 *= alpha;

  796. 		res0_1 *= alpha;

  797. 		res0_2 *= alpha;

  798. 		res0_3 *= alpha;

  799. 		res0_4 *= alpha;

  800. 		res0_5 *= alpha;

  801. 		res0_6 *= alpha;

  802. 		res0_7 *= alpha;

  803. 

  804. 		res0_8  *= alpha;

  805. 		res0_9  *= alpha;

  806. 		res0_10 *= alpha;

  807. 		res0_11 *= alpha;

  808. 		res0_12 *= alpha;

  809. 		res0_13 *= alpha;

  810. 		res0_14 *= alpha;

  811. 		res0_15 *= alpha;

  812. 

  813. 		C0[0] = res0_0;

  814. 		C0[1] = res0_1;

  815. 		C0[2] = res0_2;

  816. 		C0[3] = res0_3;

  817. 		C0[4] = res0_4;

  818. 		C0[5] = res0_5;

  819. 		C0[6] = res0_6;

  820. 		C0[7] = res0_7;

  821. 

  822. 		C0[8]  = res0_8;

  823. 		C0[9]  = res0_9;

  824. 		C0[10] = res0_10;

  825. 		C0[11] = res0_11;

  826. 		C0[12] = res0_12;

  827. 		C0[13] = res0_13;

  828. 		C0[14] = res0_14;

  829. 		C0[15] = res0_15;

  830. 

  831. 

  832. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  833. 		temp = bk - off;

  834. #ifdef LEFT

  835. 		temp -= 16; // number of values in A

  836. #else

  837. 		temp -= 1; // number of values in B

  838. #endif

  839. 		ptrba += temp*16;

  840. 		ptrbb += temp*1;

  841. #endif

  842. 

  843. #ifdef LEFT

  844. 		off += 16; // number of values in A

  845. #endif

  846. 

  847. 		C0 = C0+16;

  848. 	}

  849. 

  850. 

  851. 

  852. 

  853.         if ( bm & 8 )

  854.         {

  855. 

  856. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  857. 		ptrbb = bb;

  858. #else

  859. 		ptrba += off*8;

  860. 		ptrbb = bb + off*1;

  861. #endif

  862. 

  863. 		res0_0 = 0;

  864. 		res0_1 = 0;

  865. 		res0_2 = 0;

  866. 		res0_3 = 0;

  867. 		res0_4 = 0;

  868. 		res0_5 = 0;

  869. 		res0_6 = 0;

  870. 		res0_7 = 0;

  871. 

  872. 

  873. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  874. 		temp = bk-off;

  875. #elif defined(LEFT)

  876. 		temp = off+8;	// number of values in A

  877. #else

  878. 		temp = off+1;	// number of values in B

  879. #endif

  880. 

  881. 		for (k=0; k<temp; k++)

  882.                 {

  883. 			b0 = ptrbb[0];

  884. 

  885. 			a0 = ptrba[0];

  886. 			res0_0 += a0*b0;

  887. 

  888. 			a1 = ptrba[1];

  889. 			res0_1 += a1*b0;

  890. 

  891. 			a0 = ptrba[2];

  892. 			res0_2 += a0*b0;

  893. 

  894. 			a1 = ptrba[3];

  895. 			res0_3 += a1*b0;

  896. 

  897. 			a0 = ptrba[4];

  898. 			res0_4 += a0*b0;

  899. 

  900. 			a1 = ptrba[5];

  901. 			res0_5 += a1*b0;

  902. 

  903. 			a0 = ptrba[6];

  904. 			res0_6 += a0*b0;

  905. 

  906. 			a1 = ptrba[7];

  907. 			res0_7 += a1*b0;

  908. 

  909. 			ptrba = ptrba+8;

  910. 			ptrbb = ptrbb+1;

  911.                 }

  912. 

  913. 		res0_0 *= alpha;

  914. 		res0_1 *= alpha;

  915. 		res0_2 *= alpha;

  916. 		res0_3 *= alpha;

  917. 		res0_4 *= alpha;

  918. 		res0_5 *= alpha;

  919. 		res0_6 *= alpha;

  920. 		res0_7 *= alpha;

  921. 

  922. 		C0[0] = res0_0;

  923. 		C0[1] = res0_1;

  924. 		C0[2] = res0_2;

  925. 		C0[3] = res0_3;

  926. 		C0[4] = res0_4;

  927. 		C0[5] = res0_5;

  928. 		C0[6] = res0_6;

  929. 		C0[7] = res0_7;

  930. 

  931. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  932. 		temp = bk - off;

  933. #ifdef LEFT

  934. 		temp -= 8; // number of values in A

  935. #else

  936. 		temp -= 1; // number of values in B

  937. #endif

  938. 		ptrba += temp*8;

  939. 		ptrbb += temp*1;

  940. #endif

  941. 

  942. #ifdef LEFT

  943. 		off += 8; // number of values in A

  944. #endif

  945. 

  946. 		C0 = C0+8;

  947. 	}

  948. 

  949. 	if ( bm & 4 )

  950. 	{

  951. 

  952. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  953. 		ptrbb = bb;

  954. #else

  955. 		ptrba += off*4;

  956. 		ptrbb = bb + off*1;

  957. #endif

  958. 

  959. 		res0_0 = 0;

  960. 		res0_1 = 0;

  961. 		res0_2 = 0;

  962. 		res0_3 = 0;

  963. 

  964. 

  965. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  966. 		temp = bk-off;

  967. #elif defined(LEFT)

  968. 		temp = off+4;	// number of values in A

  969. #else

  970. 		temp = off+1;	// number of values in B

  971. #endif

  972. 

  973. 		for (k=0; k<temp; k++)

  974.                 {

  975. 			b0 = ptrbb[0];

  976. 

  977. 			a0 = ptrba[0];

  978. 			res0_0 += a0*b0;

  979. 

  980. 			a1 = ptrba[1];

  981. 			res0_1 += a1*b0;

  982. 

  983. 			a0 = ptrba[2];

  984. 			res0_2 += a0*b0;

  985. 

  986. 			a1 = ptrba[3];

  987. 			res0_3 += a1*b0;

  988. 

  989. 			ptrba = ptrba+4;

  990. 			ptrbb = ptrbb+1;

  991.                 }

  992. 

  993. 		res0_0 *= alpha;

  994. 		res0_1 *= alpha;

  995. 		res0_2 *= alpha;

  996. 		res0_3 *= alpha;

  997. 

  998. 		C0[0] = res0_0;

  999. 		C0[1] = res0_1;

  1000. 		C0[2] = res0_2;

  1001. 		C0[3] = res0_3;

  1002. 

  1003. 

  1004. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1005. 		temp = bk - off;

  1006. #ifdef LEFT

  1007. 		temp -= 4; // number of values in A

  1008. #else

  1009. 		temp -= 1; // number of values in B

  1010. #endif

  1011. 		ptrba += temp*4;

  1012. 		ptrbb += temp*1;

  1013. #endif

  1014. 

  1015. #ifdef LEFT

  1016. 		off += 4; // number of values in A

  1017. #endif

  1018. 

  1019. 		C0 = C0+4;

  1020. 

  1021. 	}

  1022. 

  1023. 	if ( bm & 2 )

  1024. 	{

  1025. 

  1026. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1027. 		ptrbb = bb;

  1028. #else

  1029. 		ptrba += off*2;

  1030. 		ptrbb = bb + off*1;

  1031. #endif

  1032. 

  1033. 		res0_0 = 0;

  1034. 		res0_1 = 0;

  1035. 

  1036. 

  1037. 

  1038. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  1039. 		temp = bk-off;

  1040. #elif defined(LEFT)

  1041. 		temp = off+2;	// number of values in A

  1042. #else

  1043. 		temp = off+1;	// number of values in B

  1044. #endif

  1045. 

  1046. 		for (k=0; k<temp; k++)

  1047.                 {

  1048. 			b0 = ptrbb[0];

  1049. 

  1050. 			a0 = ptrba[0];

  1051. 			res0_0 += a0*b0;

  1052. 

  1053. 			a1 = ptrba[1];

  1054. 			res0_1 += a1*b0;

  1055. 

  1056. 			ptrba = ptrba+2;

  1057. 			ptrbb = ptrbb+1;

  1058.                 }

  1059. 

  1060. 		res0_0 *= alpha;

  1061. 		res0_1 *= alpha;

  1062. 

  1063. 		C0[0] = res0_0;

  1064. 		C0[1] = res0_1;

  1065. 

  1066. 

  1067. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1068. 		temp = bk - off;

  1069. #ifdef LEFT

  1070. 		temp -= 2; // number of values in A

  1071. #else

  1072. 		temp -= 1; // number of values in B

  1073. #endif

  1074. 		ptrba += temp*2;

  1075. 		ptrbb += temp*1;

  1076. #endif

  1077. 

  1078. #ifdef LEFT

  1079. 		off += 2; // number of values in A

  1080. #endif

  1081. 

  1082. 		C0 = C0+2;

  1083. 

  1084. 	}

  1085. 

  1086. 	if ( bm & 1 )

  1087. 	{

  1088. 

  1089. #if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1090. 		ptrbb = bb;

  1091. #else

  1092. 		ptrba += off*1;

  1093. 		ptrbb = bb + off*1;

  1094. #endif

  1095. 

  1096. 		res0_0 = 0;

  1097. 

  1098. 

  1099. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))

  1100. 		temp = bk-off;

  1101. #elif defined(LEFT)

  1102. 		temp = off+1;	// number of values in A

  1103. #else

  1104. 		temp = off+1;	// number of values in B

  1105. #endif

  1106. 

  1107. 		for (k=0; k<temp; k++)

  1108.                 {

  1109. 			b0 = ptrbb[0];

  1110. 

  1111. 			a0 = ptrba[0];

  1112. 			res0_0 += a0*b0;

  1113. 

  1114. 			ptrba = ptrba+1;

  1115. 			ptrbb = ptrbb+1;

  1116.                 }

  1117. 

  1118. 		res0_0 *= alpha;

  1119. 

  1120. 		C0[0] = res0_0;

  1121. 

  1122. 

  1123. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))

  1124. 		temp = bk - off;

  1125. #ifdef LEFT

  1126. 		temp -= 1; // number of values in A

  1127. #else

  1128. 		temp -= 1; // number of values in B

  1129. #endif

  1130. 		ptrba += temp*1;

  1131. 		ptrbb += temp*1;

  1132. #endif

  1133. 

  1134. #ifdef LEFT

  1135. 		off += 1; // number of values in A

  1136. #endif

  1137. 

  1138. 		C0 = C0+1;

  1139. 

  1140. 	}

  1141. 

  1142. 

  1143. 

  1144. #if defined(TRMMKERNEL) && !defined(LEFT)

  1145. 		off += 1;

  1146. #endif

  1147. 

  1148.         k = (bk<<0);

  1149.         bb = bb+k;

  1150.         C = C+ldc;

  1151.    }

  1152.    return 0;

  1153. }