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.

ztrmm_lncopy_4.c 14 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664
  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 "common.h"
  40. int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, FLOAT *b){
  41. BLASLONG i, js;
  42. BLASLONG X;
  43. FLOAT data01, data02, data03, data04, data05, data06, data07, data08;
  44. FLOAT data09, data10, data11, data12, data13, data14, data15, data16;
  45. FLOAT data17, data18, data19, data20, data21, data22, data23, data24;
  46. FLOAT data25, data26, data27, data28, data29, data30, data31, data32;
  47. FLOAT *ao1, *ao2, *ao3, *ao4;
  48. lda += lda;
  49. js = (n >> 2);
  50. if (js > 0){
  51. do {
  52. X = posX;
  53. if (posX <= posY) {
  54. ao1 = a + posY * 2 + (posX + 0) * lda;
  55. ao2 = a + posY * 2 + (posX + 1) * lda;
  56. ao3 = a + posY * 2 + (posX + 2) * lda;
  57. ao4 = a + posY * 2 + (posX + 3) * lda;
  58. } else {
  59. ao1 = a + posX * 2 + (posY + 0) * lda;
  60. ao2 = a + posX * 2 + (posY + 1) * lda;
  61. ao3 = a + posX * 2 + (posY + 2) * lda;
  62. ao4 = a + posX * 2 + (posY + 3) * lda;
  63. }
  64. i = (m >> 2);
  65. if (i > 0) {
  66. do {
  67. if (X > posY) {
  68. data01 = *(ao1 + 0);
  69. data02 = *(ao1 + 1);
  70. data03 = *(ao1 + 2);
  71. data04 = *(ao1 + 3);
  72. data05 = *(ao1 + 4);
  73. data06 = *(ao1 + 5);
  74. data07 = *(ao1 + 6);
  75. data08 = *(ao1 + 7);
  76. data09 = *(ao2 + 0);
  77. data10 = *(ao2 + 1);
  78. data11 = *(ao2 + 2);
  79. data12 = *(ao2 + 3);
  80. data13 = *(ao2 + 4);
  81. data14 = *(ao2 + 5);
  82. data15 = *(ao2 + 6);
  83. data16 = *(ao2 + 7);
  84. data17 = *(ao3 + 0);
  85. data18 = *(ao3 + 1);
  86. data19 = *(ao3 + 2);
  87. data20 = *(ao3 + 3);
  88. data21 = *(ao3 + 4);
  89. data22 = *(ao3 + 5);
  90. data23 = *(ao3 + 6);
  91. data24 = *(ao3 + 7);
  92. data25 = *(ao4 + 0);
  93. data26 = *(ao4 + 1);
  94. data27 = *(ao4 + 2);
  95. data28 = *(ao4 + 3);
  96. data29 = *(ao4 + 4);
  97. data30 = *(ao4 + 5);
  98. data31 = *(ao4 + 6);
  99. data32 = *(ao4 + 7);
  100. b[ 0] = data01;
  101. b[ 1] = data02;
  102. b[ 2] = data09;
  103. b[ 3] = data10;
  104. b[ 4] = data17;
  105. b[ 5] = data18;
  106. b[ 6] = data25;
  107. b[ 7] = data26;
  108. b[ 8] = data03;
  109. b[ 9] = data04;
  110. b[10] = data11;
  111. b[11] = data12;
  112. b[12] = data19;
  113. b[13] = data20;
  114. b[14] = data27;
  115. b[15] = data28;
  116. b[16] = data05;
  117. b[17] = data06;
  118. b[18] = data13;
  119. b[19] = data14;
  120. b[20] = data21;
  121. b[21] = data22;
  122. b[22] = data29;
  123. b[23] = data30;
  124. b[24] = data07;
  125. b[25] = data08;
  126. b[26] = data15;
  127. b[27] = data16;
  128. b[28] = data23;
  129. b[29] = data24;
  130. b[30] = data31;
  131. b[31] = data32;
  132. ao1 += 8;
  133. ao2 += 8;
  134. ao3 += 8;
  135. ao4 += 8;
  136. b += 32;
  137. } else
  138. if (X < posY) {
  139. ao1 += 4 * lda;
  140. ao2 += 4 * lda;
  141. ao3 += 4 * lda;
  142. ao4 += 4 * lda;
  143. b += 32;
  144. } else {
  145. #ifdef UNIT
  146. data03 = *(ao1 + 2);
  147. data04 = *(ao1 + 3);
  148. data05 = *(ao1 + 4);
  149. data06 = *(ao1 + 5);
  150. data07 = *(ao1 + 6);
  151. data08 = *(ao1 + 7);
  152. data13 = *(ao2 + 4);
  153. data14 = *(ao2 + 5);
  154. data15 = *(ao2 + 6);
  155. data16 = *(ao2 + 7);
  156. data23 = *(ao3 + 6);
  157. data24 = *(ao3 + 7);
  158. b[ 0] = ONE;
  159. b[ 1] = ZERO;
  160. b[ 2] = ZERO;
  161. b[ 3] = ZERO;
  162. b[ 4] = ZERO;
  163. b[ 5] = ZERO;
  164. b[ 6] = ZERO;
  165. b[ 7] = ZERO;
  166. b[ 8] = data03;
  167. b[ 9] = data04;
  168. b[10] = ONE;
  169. b[11] = ZERO;
  170. b[12] = ZERO;
  171. b[13] = ZERO;
  172. b[14] = ZERO;
  173. b[15] = ZERO;
  174. b[16] = data05;
  175. b[17] = data06;
  176. b[18] = data13;
  177. b[19] = data14;
  178. b[20] = ONE;
  179. b[21] = ZERO;
  180. b[22] = ZERO;
  181. b[23] = ZERO;
  182. b[24] = data07;
  183. b[25] = data08;
  184. b[26] = data15;
  185. b[27] = data16;
  186. b[28] = data23;
  187. b[29] = data24;
  188. b[30] = ONE;
  189. b[31] = ZERO;
  190. #else
  191. data01 = *(ao1 + 0);
  192. data02 = *(ao1 + 1);
  193. data03 = *(ao1 + 2);
  194. data04 = *(ao1 + 3);
  195. data05 = *(ao1 + 4);
  196. data06 = *(ao1 + 5);
  197. data07 = *(ao1 + 6);
  198. data08 = *(ao1 + 7);
  199. data11 = *(ao2 + 2);
  200. data12 = *(ao2 + 3);
  201. data13 = *(ao2 + 4);
  202. data14 = *(ao2 + 5);
  203. data15 = *(ao2 + 6);
  204. data16 = *(ao2 + 7);
  205. data21 = *(ao3 + 4);
  206. data22 = *(ao3 + 5);
  207. data23 = *(ao3 + 6);
  208. data24 = *(ao3 + 7);
  209. data31 = *(ao4 + 6);
  210. data32 = *(ao4 + 7);
  211. b[ 0] = data01;
  212. b[ 1] = data02;
  213. b[ 2] = ZERO;
  214. b[ 3] = ZERO;
  215. b[ 4] = ZERO;
  216. b[ 5] = ZERO;
  217. b[ 6] = ZERO;
  218. b[ 7] = ZERO;
  219. b[ 8] = data03;
  220. b[ 9] = data04;
  221. b[10] = data11;
  222. b[11] = data12;
  223. b[12] = ZERO;
  224. b[13] = ZERO;
  225. b[14] = ZERO;
  226. b[15] = ZERO;
  227. b[16] = data05;
  228. b[17] = data06;
  229. b[18] = data13;
  230. b[19] = data14;
  231. b[20] = data21;
  232. b[21] = data22;
  233. b[22] = ZERO;
  234. b[23] = ZERO;
  235. b[24] = data07;
  236. b[25] = data08;
  237. b[26] = data15;
  238. b[27] = data16;
  239. b[28] = data23;
  240. b[29] = data24;
  241. b[30] = data31;
  242. b[31] = data32;
  243. #endif
  244. ao1 += 8;
  245. ao2 += 8;
  246. ao3 += 8;
  247. ao4 += 8;
  248. b += 32;
  249. }
  250. X += 4;
  251. i --;
  252. } while (i > 0);
  253. }
  254. i = (m & 3);
  255. if (i) {
  256. if (X > posY) {
  257. if (m & 2) {
  258. data01 = *(ao1 + 0);
  259. data02 = *(ao1 + 1);
  260. data03 = *(ao1 + 2);
  261. data04 = *(ao1 + 3);
  262. data09 = *(ao2 + 0);
  263. data10 = *(ao2 + 1);
  264. data11 = *(ao2 + 2);
  265. data12 = *(ao2 + 3);
  266. data17 = *(ao3 + 0);
  267. data18 = *(ao3 + 1);
  268. data19 = *(ao3 + 2);
  269. data20 = *(ao3 + 3);
  270. data25 = *(ao4 + 0);
  271. data26 = *(ao4 + 1);
  272. data27 = *(ao4 + 2);
  273. data28 = *(ao4 + 3);
  274. b[ 0] = data01;
  275. b[ 1] = data02;
  276. b[ 2] = data09;
  277. b[ 3] = data10;
  278. b[ 4] = data17;
  279. b[ 5] = data18;
  280. b[ 6] = data25;
  281. b[ 7] = data26;
  282. b[ 8] = data03;
  283. b[ 9] = data04;
  284. b[10] = data11;
  285. b[11] = data12;
  286. b[12] = data19;
  287. b[13] = data20;
  288. b[14] = data27;
  289. b[15] = data28;
  290. ao1 += 4;
  291. ao2 += 4;
  292. ao3 += 4;
  293. ao4 += 4;
  294. b += 16;
  295. }
  296. if (m & 1) {
  297. data01 = *(ao1 + 0);
  298. data02 = *(ao1 + 1);
  299. data09 = *(ao2 + 0);
  300. data10 = *(ao2 + 1);
  301. data17 = *(ao3 + 0);
  302. data18 = *(ao3 + 1);
  303. data25 = *(ao4 + 0);
  304. data26 = *(ao4 + 1);
  305. b[ 0] = data01;
  306. b[ 1] = data02;
  307. b[ 2] = data09;
  308. b[ 3] = data10;
  309. b[ 4] = data17;
  310. b[ 5] = data18;
  311. b[ 6] = data25;
  312. b[ 7] = data26;
  313. /* ao1 += 2;
  314. ao2 += 2;
  315. ao3 += 2;
  316. ao4 += 2; */
  317. b += 8;
  318. }
  319. } else
  320. if (X < posY) {
  321. if (m & 2) {
  322. /* ao1 += 2 * lda;
  323. ao2 += 2 * lda; */
  324. b += 16;
  325. }
  326. if (m & 1) {
  327. // ao1 += lda;
  328. b += 8;
  329. }
  330. } else {
  331. #ifdef UNIT
  332. if (i >= 2) {
  333. data03 = *(ao1 + 2);
  334. data04 = *(ao1 + 3);
  335. }
  336. if (i >= 3) {
  337. data05 = *(ao1 + 4);
  338. data06 = *(ao1 + 5);
  339. data13 = *(ao2 + 4);
  340. data14 = *(ao2 + 5);
  341. }
  342. b[ 0] = ONE;
  343. b[ 1] = ZERO;
  344. b[ 2] = ZERO;
  345. b[ 3] = ZERO;
  346. b[ 4] = ZERO;
  347. b[ 5] = ZERO;
  348. b[ 6] = ZERO;
  349. b[ 7] = ZERO;
  350. b += 8;
  351. if (i >= 2) {
  352. b[ 0] = data03;
  353. b[ 1] = data04;
  354. b[ 2] = ONE;
  355. b[ 3] = ZERO;
  356. b[ 4] = ZERO;
  357. b[ 5] = ZERO;
  358. b[ 6] = ZERO;
  359. b[ 7] = ZERO;
  360. b += 8;
  361. }
  362. if (i >= 3) {
  363. b[ 0] = data05;
  364. b[ 1] = data06;
  365. b[ 2] = data13;
  366. b[ 3] = data14;
  367. b[ 4] = ONE;
  368. b[ 5] = ZERO;
  369. b[ 6] = ZERO;
  370. b[ 7] = ZERO;
  371. b += 8;
  372. }
  373. #else
  374. data01 = *(ao1 + 0);
  375. data02 = *(ao1 + 1);
  376. if (i >= 2) {
  377. data03 = *(ao1 + 2);
  378. data04 = *(ao1 + 3);
  379. data11 = *(ao2 + 2);
  380. data12 = *(ao2 + 3);
  381. }
  382. if (i >= 3) {
  383. data05 = *(ao1 + 4);
  384. data06 = *(ao1 + 5);
  385. data13 = *(ao2 + 4);
  386. data14 = *(ao2 + 5);
  387. data21 = *(ao3 + 4);
  388. data22 = *(ao3 + 5);
  389. }
  390. b[ 0] = data01;
  391. b[ 1] = data02;
  392. b[ 2] = ZERO;
  393. b[ 3] = ZERO;
  394. b[ 4] = ZERO;
  395. b[ 5] = ZERO;
  396. b[ 6] = ZERO;
  397. b[ 7] = ZERO;
  398. b += 8;
  399. if (i >= 2) {
  400. b[ 0] = data03;
  401. b[ 1] = data04;
  402. b[ 2] = data11;
  403. b[ 3] = data12;
  404. b[ 4] = ZERO;
  405. b[ 5] = ZERO;
  406. b[ 6] = ZERO;
  407. b[ 7] = ZERO;
  408. b += 8;
  409. }
  410. if (i >= 3) {
  411. b[ 0] = data05;
  412. b[ 1] = data06;
  413. b[ 2] = data13;
  414. b[ 3] = data14;
  415. b[ 4] = data21;
  416. b[ 5] = data22;
  417. b[ 6] = ZERO;
  418. b[ 7] = ZERO;
  419. b += 8;
  420. }
  421. #endif
  422. }
  423. }
  424. posY += 4;
  425. js --;
  426. } while (js > 0);
  427. } /* End of main loop */
  428. if (n & 2){
  429. X = posX;
  430. if (posX <= posY) {
  431. ao1 = a + posY * 2 + (posX + 0) * lda;
  432. ao2 = a + posY * 2 + (posX + 1) * lda;
  433. } else {
  434. ao1 = a + posX * 2 + (posY + 0) * lda;
  435. ao2 = a + posX * 2 + (posY + 1) * lda;
  436. }
  437. i = (m >> 1);
  438. if (i > 0) {
  439. do {
  440. if (X > posY) {
  441. data01 = *(ao1 + 0);
  442. data02 = *(ao1 + 1);
  443. data03 = *(ao1 + 2);
  444. data04 = *(ao1 + 3);
  445. data09 = *(ao2 + 0);
  446. data10 = *(ao2 + 1);
  447. data11 = *(ao2 + 2);
  448. data12 = *(ao2 + 3);
  449. b[ 0] = data01;
  450. b[ 1] = data02;
  451. b[ 2] = data09;
  452. b[ 3] = data10;
  453. b[ 4] = data03;
  454. b[ 5] = data04;
  455. b[ 6] = data11;
  456. b[ 7] = data12;
  457. ao1 += 4;
  458. ao2 += 4;
  459. b += 8;
  460. } else
  461. if (X < posY) {
  462. ao1 += 2 * lda;
  463. ao2 += 2 * lda;
  464. b += 8;
  465. } else {
  466. #ifdef UNIT
  467. data03 = *(ao1 + 2);
  468. data04 = *(ao1 + 3);
  469. b[ 0] = ONE;
  470. b[ 1] = ZERO;
  471. b[ 2] = ZERO;
  472. b[ 3] = ZERO;
  473. b[ 4] = data03;
  474. b[ 5] = data04;
  475. b[ 6] = ONE;
  476. b[ 7] = ZERO;
  477. #else
  478. data01 = *(ao1 + 0);
  479. data02 = *(ao1 + 1);
  480. data03 = *(ao1 + 2);
  481. data04 = *(ao1 + 3);
  482. data11 = *(ao2 + 2);
  483. data12 = *(ao2 + 3);
  484. b[ 0] = data01;
  485. b[ 1] = data02;
  486. b[ 2] = ZERO;
  487. b[ 3] = ZERO;
  488. b[ 4] = data03;
  489. b[ 5] = data04;
  490. b[ 6] = data11;
  491. b[ 7] = data12;
  492. #endif
  493. ao1 += 4;
  494. ao2 += 4;
  495. b += 8;
  496. }
  497. X += 2;
  498. i --;
  499. } while (i > 0);
  500. }
  501. i = (m & 1);
  502. if (i) {
  503. if (X > posY) {
  504. data01 = *(ao1 + 0);
  505. data02 = *(ao1 + 1);
  506. data09 = *(ao2 + 0);
  507. data10 = *(ao2 + 1);
  508. b[ 0] = data01;
  509. b[ 1] = data02;
  510. b[ 2] = data09;
  511. b[ 3] = data10;
  512. /* ao1 += 2;
  513. ao2 += 2; */
  514. b += 4;
  515. } else
  516. if (X < posY) {
  517. // ao1 += lda;
  518. b += 4;
  519. } else {
  520. #ifdef UNIT
  521. b[ 0] = ONE;
  522. b[ 1] = ZERO;
  523. b[ 2] = ZERO;
  524. b[ 3] = ZERO;
  525. #else
  526. data01 = *(ao1 + 0);
  527. data02 = *(ao1 + 1);
  528. b[ 0] = data01;
  529. b[ 1] = data02;
  530. b[ 2] = ZERO;
  531. b[ 3] = ZERO;
  532. #endif
  533. b += 2;
  534. }
  535. }
  536. posY += 2;
  537. }
  538. if (n & 1){
  539. X = posX;
  540. if (posX <= posY) {
  541. ao1 = a + posY * 2 + (posX + 0) * lda;
  542. } else {
  543. ao1 = a + posX * 2 + (posY + 0) * lda;
  544. }
  545. i = m;
  546. if (i > 0) {
  547. do {
  548. if (X > posY) {
  549. data01 = *(ao1 + 0);
  550. data02 = *(ao1 + 1);
  551. b[ 0] = data01;
  552. b[ 1] = data02;
  553. ao1 += 2;
  554. b += 2;
  555. } else
  556. if (X < posY) {
  557. ao1 += lda;
  558. b += 2;
  559. } else {
  560. #ifdef UNIT
  561. b[ 0] = ONE;
  562. b[ 1] = ZERO;
  563. #else
  564. data01 = *(ao1 + 0);
  565. data02 = *(ao1 + 1);
  566. b[ 0] = data01;
  567. b[ 1] = data02;
  568. #endif
  569. ao1 += lda;
  570. b += 2;
  571. }
  572. X ++;
  573. i --;
  574. } while (i > 0);
  575. }
  576. // posY += 1;
  577. }
  578. return 0;
  579. }