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.

gemm_ncopy_16.c 12 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437
  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, FLOAT *b){
  41. BLASLONG i, j;
  42. FLOAT *aoffset;
  43. FLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
  44. FLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
  45. FLOAT *aoffset9, *aoffset10, *aoffset11, *aoffset12;
  46. FLOAT *aoffset13, *aoffset14, *aoffset15, *aoffset16;
  47. FLOAT *boffset;
  48. FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
  49. FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
  50. FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
  51. FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
  52. FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
  53. FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
  54. FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
  55. FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
  56. aoffset = a;
  57. boffset = b;
  58. j = (n >> 4);
  59. if (j > 0){
  60. do{
  61. aoffset1 = aoffset;
  62. aoffset2 = aoffset1 + lda;
  63. aoffset3 = aoffset2 + lda;
  64. aoffset4 = aoffset3 + lda;
  65. aoffset5 = aoffset4 + lda;
  66. aoffset6 = aoffset5 + lda;
  67. aoffset7 = aoffset6 + lda;
  68. aoffset8 = aoffset7 + lda;
  69. aoffset9 = aoffset8 + lda;
  70. aoffset10 = aoffset9 + lda;
  71. aoffset11 = aoffset10 + lda;
  72. aoffset12 = aoffset11 + lda;
  73. aoffset13 = aoffset12 + lda;
  74. aoffset14 = aoffset13 + lda;
  75. aoffset15 = aoffset14 + lda;
  76. aoffset16 = aoffset15 + lda;
  77. aoffset += 16 * lda;
  78. i = (m >> 1);
  79. if (i > 0){
  80. do{
  81. ctemp01 = *(aoffset1 + 0);
  82. ctemp02 = *(aoffset1 + 1);
  83. ctemp03 = *(aoffset2 + 0);
  84. ctemp04 = *(aoffset2 + 1);
  85. ctemp05 = *(aoffset3 + 0);
  86. ctemp06 = *(aoffset3 + 1);
  87. ctemp07 = *(aoffset4 + 0);
  88. ctemp08 = *(aoffset4 + 1);
  89. ctemp09 = *(aoffset5 + 0);
  90. ctemp10 = *(aoffset5 + 1);
  91. ctemp11 = *(aoffset6 + 0);
  92. ctemp12 = *(aoffset6 + 1);
  93. ctemp13 = *(aoffset7 + 0);
  94. ctemp14 = *(aoffset7 + 1);
  95. ctemp15 = *(aoffset8 + 0);
  96. ctemp16 = *(aoffset8 + 1);
  97. ctemp17 = *(aoffset9 + 0);
  98. ctemp18 = *(aoffset9 + 1);
  99. ctemp19 = *(aoffset10 + 0);
  100. ctemp20 = *(aoffset10 + 1);
  101. ctemp21 = *(aoffset11 + 0);
  102. ctemp22 = *(aoffset11 + 1);
  103. ctemp23 = *(aoffset12 + 0);
  104. ctemp24 = *(aoffset12 + 1);
  105. ctemp25 = *(aoffset13 + 0);
  106. ctemp26 = *(aoffset13 + 1);
  107. ctemp27 = *(aoffset14 + 0);
  108. ctemp28 = *(aoffset14 + 1);
  109. ctemp29 = *(aoffset15 + 0);
  110. ctemp30 = *(aoffset15 + 1);
  111. ctemp31 = *(aoffset16 + 0);
  112. ctemp32 = *(aoffset16 + 1);
  113. *(boffset + 0) = ctemp01;
  114. *(boffset + 1) = ctemp03;
  115. *(boffset + 2) = ctemp05;
  116. *(boffset + 3) = ctemp07;
  117. *(boffset + 4) = ctemp09;
  118. *(boffset + 5) = ctemp11;
  119. *(boffset + 6) = ctemp13;
  120. *(boffset + 7) = ctemp15;
  121. *(boffset + 8) = ctemp17;
  122. *(boffset + 9) = ctemp19;
  123. *(boffset + 10) = ctemp21;
  124. *(boffset + 11) = ctemp23;
  125. *(boffset + 12) = ctemp25;
  126. *(boffset + 13) = ctemp27;
  127. *(boffset + 14) = ctemp29;
  128. *(boffset + 15) = ctemp31;
  129. *(boffset + 16) = ctemp02;
  130. *(boffset + 17) = ctemp04;
  131. *(boffset + 18) = ctemp06;
  132. *(boffset + 19) = ctemp08;
  133. *(boffset + 20) = ctemp10;
  134. *(boffset + 21) = ctemp12;
  135. *(boffset + 22) = ctemp14;
  136. *(boffset + 23) = ctemp16;
  137. *(boffset + 24) = ctemp18;
  138. *(boffset + 25) = ctemp20;
  139. *(boffset + 26) = ctemp22;
  140. *(boffset + 27) = ctemp24;
  141. *(boffset + 28) = ctemp26;
  142. *(boffset + 29) = ctemp28;
  143. *(boffset + 30) = ctemp30;
  144. *(boffset + 31) = ctemp32;
  145. aoffset1 += 2;
  146. aoffset2 += 2;
  147. aoffset3 += 2;
  148. aoffset4 += 2;
  149. aoffset5 += 2;
  150. aoffset6 += 2;
  151. aoffset7 += 2;
  152. aoffset8 += 2;
  153. aoffset9 += 2;
  154. aoffset10 += 2;
  155. aoffset11 += 2;
  156. aoffset12 += 2;
  157. aoffset13 += 2;
  158. aoffset14 += 2;
  159. aoffset15 += 2;
  160. aoffset16 += 2;
  161. boffset += 32;
  162. i --;
  163. }while(i > 0);
  164. }
  165. if (m & 1){
  166. ctemp01 = *(aoffset1 + 0);
  167. ctemp03 = *(aoffset2 + 0);
  168. ctemp05 = *(aoffset3 + 0);
  169. ctemp07 = *(aoffset4 + 0);
  170. ctemp09 = *(aoffset5 + 0);
  171. ctemp11 = *(aoffset6 + 0);
  172. ctemp13 = *(aoffset7 + 0);
  173. ctemp15 = *(aoffset8 + 0);
  174. ctemp17 = *(aoffset9 + 0);
  175. ctemp19 = *(aoffset10 + 0);
  176. ctemp21 = *(aoffset11 + 0);
  177. ctemp23 = *(aoffset12 + 0);
  178. ctemp25 = *(aoffset13 + 0);
  179. ctemp27 = *(aoffset14 + 0);
  180. ctemp29 = *(aoffset15 + 0);
  181. ctemp31 = *(aoffset16 + 0);
  182. *(boffset + 0) = ctemp01;
  183. *(boffset + 1) = ctemp03;
  184. *(boffset + 2) = ctemp05;
  185. *(boffset + 3) = ctemp07;
  186. *(boffset + 4) = ctemp09;
  187. *(boffset + 5) = ctemp11;
  188. *(boffset + 6) = ctemp13;
  189. *(boffset + 7) = ctemp15;
  190. *(boffset + 8) = ctemp17;
  191. *(boffset + 9) = ctemp19;
  192. *(boffset + 10) = ctemp21;
  193. *(boffset + 11) = ctemp23;
  194. *(boffset + 12) = ctemp25;
  195. *(boffset + 13) = ctemp27;
  196. *(boffset + 14) = ctemp29;
  197. *(boffset + 15) = ctemp31;
  198. boffset += 16;
  199. }
  200. j--;
  201. }while(j > 0);
  202. } /* end of if(j > 0) */
  203. if (n & 8){
  204. aoffset1 = aoffset;
  205. aoffset2 = aoffset1 + lda;
  206. aoffset3 = aoffset2 + lda;
  207. aoffset4 = aoffset3 + lda;
  208. aoffset5 = aoffset4 + lda;
  209. aoffset6 = aoffset5 + lda;
  210. aoffset7 = aoffset6 + lda;
  211. aoffset8 = aoffset7 + lda;
  212. aoffset += 8 * lda;
  213. i = (m >> 1);
  214. if (i > 0){
  215. do{
  216. ctemp01 = *(aoffset1 + 0);
  217. ctemp02 = *(aoffset1 + 1);
  218. ctemp03 = *(aoffset2 + 0);
  219. ctemp04 = *(aoffset2 + 1);
  220. ctemp05 = *(aoffset3 + 0);
  221. ctemp06 = *(aoffset3 + 1);
  222. ctemp07 = *(aoffset4 + 0);
  223. ctemp08 = *(aoffset4 + 1);
  224. ctemp09 = *(aoffset5 + 0);
  225. ctemp10 = *(aoffset5 + 1);
  226. ctemp11 = *(aoffset6 + 0);
  227. ctemp12 = *(aoffset6 + 1);
  228. ctemp13 = *(aoffset7 + 0);
  229. ctemp14 = *(aoffset7 + 1);
  230. ctemp15 = *(aoffset8 + 0);
  231. ctemp16 = *(aoffset8 + 1);
  232. *(boffset + 0) = ctemp01;
  233. *(boffset + 1) = ctemp03;
  234. *(boffset + 2) = ctemp05;
  235. *(boffset + 3) = ctemp07;
  236. *(boffset + 4) = ctemp09;
  237. *(boffset + 5) = ctemp11;
  238. *(boffset + 6) = ctemp13;
  239. *(boffset + 7) = ctemp15;
  240. *(boffset + 8) = ctemp02;
  241. *(boffset + 9) = ctemp04;
  242. *(boffset + 10) = ctemp06;
  243. *(boffset + 11) = ctemp08;
  244. *(boffset + 12) = ctemp10;
  245. *(boffset + 13) = ctemp12;
  246. *(boffset + 14) = ctemp14;
  247. *(boffset + 15) = ctemp16;
  248. aoffset1 += 2;
  249. aoffset2 += 2;
  250. aoffset3 += 2;
  251. aoffset4 += 2;
  252. aoffset5 += 2;
  253. aoffset6 += 2;
  254. aoffset7 += 2;
  255. aoffset8 += 2;
  256. boffset += 16;
  257. i --;
  258. }while(i > 0);
  259. }
  260. if (m & 1){
  261. ctemp01 = *(aoffset1 + 0);
  262. ctemp03 = *(aoffset2 + 0);
  263. ctemp05 = *(aoffset3 + 0);
  264. ctemp07 = *(aoffset4 + 0);
  265. ctemp09 = *(aoffset5 + 0);
  266. ctemp11 = *(aoffset6 + 0);
  267. ctemp13 = *(aoffset7 + 0);
  268. ctemp15 = *(aoffset8 + 0);
  269. *(boffset + 0) = ctemp01;
  270. *(boffset + 1) = ctemp03;
  271. *(boffset + 2) = ctemp05;
  272. *(boffset + 3) = ctemp07;
  273. *(boffset + 4) = ctemp09;
  274. *(boffset + 5) = ctemp11;
  275. *(boffset + 6) = ctemp13;
  276. *(boffset + 7) = ctemp15;
  277. boffset += 8;
  278. }
  279. }
  280. if (n & 4){
  281. aoffset1 = aoffset;
  282. aoffset2 = aoffset1 + lda;
  283. aoffset3 = aoffset2 + lda;
  284. aoffset4 = aoffset3 + lda;
  285. aoffset += 4 * lda;
  286. i = (m >> 1);
  287. if (i > 0){
  288. do{
  289. ctemp01 = *(aoffset1 + 0);
  290. ctemp02 = *(aoffset1 + 1);
  291. ctemp03 = *(aoffset2 + 0);
  292. ctemp04 = *(aoffset2 + 1);
  293. ctemp05 = *(aoffset3 + 0);
  294. ctemp06 = *(aoffset3 + 1);
  295. ctemp07 = *(aoffset4 + 0);
  296. ctemp08 = *(aoffset4 + 1);
  297. *(boffset + 0) = ctemp01;
  298. *(boffset + 1) = ctemp03;
  299. *(boffset + 2) = ctemp05;
  300. *(boffset + 3) = ctemp07;
  301. *(boffset + 4) = ctemp02;
  302. *(boffset + 5) = ctemp04;
  303. *(boffset + 6) = ctemp06;
  304. *(boffset + 7) = ctemp08;
  305. aoffset1 += 2;
  306. aoffset2 += 2;
  307. aoffset3 += 2;
  308. aoffset4 += 2;
  309. boffset += 8;
  310. i --;
  311. }while(i > 0);
  312. }
  313. if (m & 1){
  314. ctemp01 = *(aoffset1 + 0);
  315. ctemp03 = *(aoffset2 + 0);
  316. ctemp05 = *(aoffset3 + 0);
  317. ctemp07 = *(aoffset4 + 0);
  318. *(boffset + 0) = ctemp01;
  319. *(boffset + 1) = ctemp03;
  320. *(boffset + 2) = ctemp05;
  321. *(boffset + 3) = ctemp07;
  322. boffset += 4;
  323. }
  324. }
  325. if (n & 2){
  326. aoffset1 = aoffset;
  327. aoffset2 = aoffset1 + lda;
  328. aoffset += 2 * lda;
  329. i = (m >> 1);
  330. if (i > 0){
  331. do{
  332. ctemp01 = *(aoffset1 + 0);
  333. ctemp02 = *(aoffset1 + 1);
  334. ctemp03 = *(aoffset2 + 0);
  335. ctemp04 = *(aoffset2 + 1);
  336. *(boffset + 0) = ctemp01;
  337. *(boffset + 1) = ctemp03;
  338. *(boffset + 2) = ctemp02;
  339. *(boffset + 3) = ctemp04;
  340. aoffset1 += 2;
  341. aoffset2 += 2;
  342. boffset += 4;
  343. i --;
  344. }while(i > 0);
  345. }
  346. if (m & 1){
  347. ctemp01 = *(aoffset1 + 0);
  348. ctemp03 = *(aoffset2 + 0);
  349. *(boffset + 0) = ctemp01;
  350. *(boffset + 1) = ctemp03;
  351. boffset += 2;
  352. }
  353. }
  354. if (n & 1){
  355. aoffset1 = aoffset;
  356. i = (m >> 1);
  357. if (i > 0){
  358. do{
  359. ctemp01 = *(aoffset1 + 0);
  360. ctemp02 = *(aoffset1 + 1);
  361. *(boffset + 0) = ctemp01;
  362. *(boffset + 1) = ctemp02;
  363. aoffset1 += 2;
  364. boffset += 2;
  365. i --;
  366. }while(i > 0);
  367. }
  368. if (m & 1){
  369. ctemp01 = *(aoffset1 + 0);
  370. *(boffset + 0) = ctemp01;
  371. boffset += 1;
  372. }
  373. }
  374. return 0;
  375. }