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_8.c 12 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422
  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 *boffset;
  46. FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
  47. FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
  48. FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
  49. FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
  50. FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
  51. FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
  52. FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
  53. FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
  54. FLOAT ctemp33, ctemp34, ctemp35, ctemp36;
  55. FLOAT ctemp37, ctemp38, ctemp39, ctemp40;
  56. FLOAT ctemp41, ctemp42, ctemp43, ctemp44;
  57. FLOAT ctemp45, ctemp46, ctemp47, ctemp48;
  58. FLOAT ctemp49, ctemp50, ctemp51, ctemp52;
  59. FLOAT ctemp53, ctemp54, ctemp55, ctemp56;
  60. FLOAT ctemp57, ctemp58, ctemp59, ctemp60;
  61. FLOAT ctemp61, ctemp62, ctemp63, ctemp64;
  62. aoffset = a;
  63. boffset = b;
  64. j = (n >> 3);
  65. if (j > 0){
  66. do{
  67. aoffset1 = aoffset;
  68. aoffset2 = aoffset1 + lda;
  69. aoffset3 = aoffset2 + lda;
  70. aoffset4 = aoffset3 + lda;
  71. aoffset5 = aoffset4 + lda;
  72. aoffset6 = aoffset5 + lda;
  73. aoffset7 = aoffset6 + lda;
  74. aoffset8 = aoffset7 + lda;
  75. aoffset += 8 * lda;
  76. i = (m >> 3);
  77. if (i > 0){
  78. do{
  79. ctemp01 = *(aoffset1 + 0);
  80. ctemp02 = *(aoffset1 + 1);
  81. ctemp03 = *(aoffset1 + 2);
  82. ctemp04 = *(aoffset1 + 3);
  83. ctemp05 = *(aoffset1 + 4);
  84. ctemp06 = *(aoffset1 + 5);
  85. ctemp07 = *(aoffset1 + 6);
  86. ctemp08 = *(aoffset1 + 7);
  87. ctemp09 = *(aoffset2 + 0);
  88. ctemp10 = *(aoffset2 + 1);
  89. ctemp11 = *(aoffset2 + 2);
  90. ctemp12 = *(aoffset2 + 3);
  91. ctemp13 = *(aoffset2 + 4);
  92. ctemp14 = *(aoffset2 + 5);
  93. ctemp15 = *(aoffset2 + 6);
  94. ctemp16 = *(aoffset2 + 7);
  95. ctemp17 = *(aoffset3 + 0);
  96. ctemp18 = *(aoffset3 + 1);
  97. ctemp19 = *(aoffset3 + 2);
  98. ctemp20 = *(aoffset3 + 3);
  99. ctemp21 = *(aoffset3 + 4);
  100. ctemp22 = *(aoffset3 + 5);
  101. ctemp23 = *(aoffset3 + 6);
  102. ctemp24 = *(aoffset3 + 7);
  103. ctemp25 = *(aoffset4 + 0);
  104. ctemp26 = *(aoffset4 + 1);
  105. ctemp27 = *(aoffset4 + 2);
  106. ctemp28 = *(aoffset4 + 3);
  107. ctemp29 = *(aoffset4 + 4);
  108. ctemp30 = *(aoffset4 + 5);
  109. ctemp31 = *(aoffset4 + 6);
  110. ctemp32 = *(aoffset4 + 7);
  111. ctemp33 = *(aoffset5 + 0);
  112. ctemp34 = *(aoffset5 + 1);
  113. ctemp35 = *(aoffset5 + 2);
  114. ctemp36 = *(aoffset5 + 3);
  115. ctemp37 = *(aoffset5 + 4);
  116. ctemp38 = *(aoffset5 + 5);
  117. ctemp39 = *(aoffset5 + 6);
  118. ctemp40 = *(aoffset5 + 7);
  119. ctemp41 = *(aoffset6 + 0);
  120. ctemp42 = *(aoffset6 + 1);
  121. ctemp43 = *(aoffset6 + 2);
  122. ctemp44 = *(aoffset6 + 3);
  123. ctemp45 = *(aoffset6 + 4);
  124. ctemp46 = *(aoffset6 + 5);
  125. ctemp47 = *(aoffset6 + 6);
  126. ctemp48 = *(aoffset6 + 7);
  127. ctemp49 = *(aoffset7 + 0);
  128. ctemp50 = *(aoffset7 + 1);
  129. ctemp51 = *(aoffset7 + 2);
  130. ctemp52 = *(aoffset7 + 3);
  131. ctemp53 = *(aoffset7 + 4);
  132. ctemp54 = *(aoffset7 + 5);
  133. ctemp55 = *(aoffset7 + 6);
  134. ctemp56 = *(aoffset7 + 7);
  135. ctemp57 = *(aoffset8 + 0);
  136. ctemp58 = *(aoffset8 + 1);
  137. ctemp59 = *(aoffset8 + 2);
  138. ctemp60 = *(aoffset8 + 3);
  139. ctemp61 = *(aoffset8 + 4);
  140. ctemp62 = *(aoffset8 + 5);
  141. ctemp63 = *(aoffset8 + 6);
  142. ctemp64 = *(aoffset8 + 7);
  143. *(boffset + 0) = ctemp01;
  144. *(boffset + 1) = ctemp09;
  145. *(boffset + 2) = ctemp17;
  146. *(boffset + 3) = ctemp25;
  147. *(boffset + 4) = ctemp33;
  148. *(boffset + 5) = ctemp41;
  149. *(boffset + 6) = ctemp49;
  150. *(boffset + 7) = ctemp57;
  151. *(boffset + 8) = ctemp02;
  152. *(boffset + 9) = ctemp10;
  153. *(boffset + 10) = ctemp18;
  154. *(boffset + 11) = ctemp26;
  155. *(boffset + 12) = ctemp34;
  156. *(boffset + 13) = ctemp42;
  157. *(boffset + 14) = ctemp50;
  158. *(boffset + 15) = ctemp58;
  159. *(boffset + 16) = ctemp03;
  160. *(boffset + 17) = ctemp11;
  161. *(boffset + 18) = ctemp19;
  162. *(boffset + 19) = ctemp27;
  163. *(boffset + 20) = ctemp35;
  164. *(boffset + 21) = ctemp43;
  165. *(boffset + 22) = ctemp51;
  166. *(boffset + 23) = ctemp59;
  167. *(boffset + 24) = ctemp04;
  168. *(boffset + 25) = ctemp12;
  169. *(boffset + 26) = ctemp20;
  170. *(boffset + 27) = ctemp28;
  171. *(boffset + 28) = ctemp36;
  172. *(boffset + 29) = ctemp44;
  173. *(boffset + 30) = ctemp52;
  174. *(boffset + 31) = ctemp60;
  175. *(boffset + 32) = ctemp05;
  176. *(boffset + 33) = ctemp13;
  177. *(boffset + 34) = ctemp21;
  178. *(boffset + 35) = ctemp29;
  179. *(boffset + 36) = ctemp37;
  180. *(boffset + 37) = ctemp45;
  181. *(boffset + 38) = ctemp53;
  182. *(boffset + 39) = ctemp61;
  183. *(boffset + 40) = ctemp06;
  184. *(boffset + 41) = ctemp14;
  185. *(boffset + 42) = ctemp22;
  186. *(boffset + 43) = ctemp30;
  187. *(boffset + 44) = ctemp38;
  188. *(boffset + 45) = ctemp46;
  189. *(boffset + 46) = ctemp54;
  190. *(boffset + 47) = ctemp62;
  191. *(boffset + 48) = ctemp07;
  192. *(boffset + 49) = ctemp15;
  193. *(boffset + 50) = ctemp23;
  194. *(boffset + 51) = ctemp31;
  195. *(boffset + 52) = ctemp39;
  196. *(boffset + 53) = ctemp47;
  197. *(boffset + 54) = ctemp55;
  198. *(boffset + 55) = ctemp63;
  199. *(boffset + 56) = ctemp08;
  200. *(boffset + 57) = ctemp16;
  201. *(boffset + 58) = ctemp24;
  202. *(boffset + 59) = ctemp32;
  203. *(boffset + 60) = ctemp40;
  204. *(boffset + 61) = ctemp48;
  205. *(boffset + 62) = ctemp56;
  206. *(boffset + 63) = ctemp64;
  207. aoffset1 += 8;
  208. aoffset2 += 8;
  209. aoffset3 += 8;
  210. aoffset4 += 8;
  211. aoffset5 += 8;
  212. aoffset6 += 8;
  213. aoffset7 += 8;
  214. aoffset8 += 8;
  215. boffset += 64;
  216. i --;
  217. }while(i > 0);
  218. }
  219. i = (m & 7);
  220. if (i > 0){
  221. do{
  222. ctemp01 = *(aoffset1 + 0);
  223. ctemp09 = *(aoffset2 + 0);
  224. ctemp17 = *(aoffset3 + 0);
  225. ctemp25 = *(aoffset4 + 0);
  226. ctemp33 = *(aoffset5 + 0);
  227. ctemp41 = *(aoffset6 + 0);
  228. ctemp49 = *(aoffset7 + 0);
  229. ctemp57 = *(aoffset8 + 0);
  230. *(boffset + 0) = ctemp01;
  231. *(boffset + 1) = ctemp09;
  232. *(boffset + 2) = ctemp17;
  233. *(boffset + 3) = ctemp25;
  234. *(boffset + 4) = ctemp33;
  235. *(boffset + 5) = ctemp41;
  236. *(boffset + 6) = ctemp49;
  237. *(boffset + 7) = ctemp57;
  238. aoffset1 ++;
  239. aoffset2 ++;
  240. aoffset3 ++;
  241. aoffset4 ++;
  242. aoffset5 ++;
  243. aoffset6 ++;
  244. aoffset7 ++;
  245. aoffset8 ++;
  246. boffset += 8;
  247. i --;
  248. }while(i > 0);
  249. }
  250. j--;
  251. }while(j > 0);
  252. } /* end of if(j > 0) */
  253. if (n & 4){
  254. aoffset1 = aoffset;
  255. aoffset2 = aoffset1 + lda;
  256. aoffset3 = aoffset2 + lda;
  257. aoffset4 = aoffset3 + lda;
  258. aoffset += 4 * lda;
  259. i = (m >> 2);
  260. if (i > 0){
  261. do{
  262. ctemp01 = *(aoffset1 + 0);
  263. ctemp02 = *(aoffset1 + 1);
  264. ctemp03 = *(aoffset1 + 2);
  265. ctemp04 = *(aoffset1 + 3);
  266. ctemp05 = *(aoffset2 + 0);
  267. ctemp06 = *(aoffset2 + 1);
  268. ctemp07 = *(aoffset2 + 2);
  269. ctemp08 = *(aoffset2 + 3);
  270. ctemp09 = *(aoffset3 + 0);
  271. ctemp10 = *(aoffset3 + 1);
  272. ctemp11 = *(aoffset3 + 2);
  273. ctemp12 = *(aoffset3 + 3);
  274. ctemp13 = *(aoffset4 + 0);
  275. ctemp14 = *(aoffset4 + 1);
  276. ctemp15 = *(aoffset4 + 2);
  277. ctemp16 = *(aoffset4 + 3);
  278. *(boffset + 0) = ctemp01;
  279. *(boffset + 1) = ctemp05;
  280. *(boffset + 2) = ctemp09;
  281. *(boffset + 3) = ctemp13;
  282. *(boffset + 4) = ctemp02;
  283. *(boffset + 5) = ctemp06;
  284. *(boffset + 6) = ctemp10;
  285. *(boffset + 7) = ctemp14;
  286. *(boffset + 8) = ctemp03;
  287. *(boffset + 9) = ctemp07;
  288. *(boffset + 10) = ctemp11;
  289. *(boffset + 11) = ctemp15;
  290. *(boffset + 12) = ctemp04;
  291. *(boffset + 13) = ctemp08;
  292. *(boffset + 14) = ctemp12;
  293. *(boffset + 15) = ctemp16;
  294. aoffset1 += 4;
  295. aoffset2 += 4;
  296. aoffset3 += 4;
  297. aoffset4 += 4;
  298. boffset += 16;
  299. i --;
  300. }while(i > 0);
  301. }
  302. i = (m & 3);
  303. if (i > 0){
  304. do{
  305. ctemp01 = *(aoffset1 + 0);
  306. ctemp02 = *(aoffset2 + 0);
  307. ctemp03 = *(aoffset3 + 0);
  308. ctemp04 = *(aoffset4 + 0);
  309. *(boffset + 0) = ctemp01;
  310. *(boffset + 1) = ctemp02;
  311. *(boffset + 2) = ctemp03;
  312. *(boffset + 3) = ctemp04;
  313. aoffset1 ++;
  314. aoffset2 ++;
  315. aoffset3 ++;
  316. aoffset4 ++;
  317. boffset += 4;
  318. i --;
  319. }while(i > 0);
  320. }
  321. } /* end of if(j > 0) */
  322. if (n & 2){
  323. aoffset1 = aoffset;
  324. aoffset2 = aoffset1 + lda;
  325. aoffset += 2 * lda;
  326. i = (m >> 1);
  327. if (i > 0){
  328. do{
  329. ctemp01 = *(aoffset1 + 0);
  330. ctemp02 = *(aoffset1 + 1);
  331. ctemp03 = *(aoffset2 + 0);
  332. ctemp04 = *(aoffset2 + 1);
  333. *(boffset + 0) = ctemp01;
  334. *(boffset + 1) = ctemp03;
  335. *(boffset + 2) = ctemp02;
  336. *(boffset + 3) = ctemp04;
  337. aoffset1 += 2;
  338. aoffset2 += 2;
  339. boffset += 4;
  340. i --;
  341. }while(i > 0);
  342. }
  343. if (m & 1){
  344. ctemp01 = *(aoffset1 + 0);
  345. ctemp02 = *(aoffset2 + 0);
  346. *(boffset + 0) = ctemp01;
  347. *(boffset + 1) = ctemp02;
  348. aoffset1 ++;
  349. aoffset2 ++;
  350. boffset += 2;
  351. }
  352. } /* end of if(j > 0) */
  353. if (n & 1){
  354. aoffset1 = aoffset;
  355. i = m;
  356. if (i > 0){
  357. do{
  358. ctemp01 = *(aoffset1 + 0);
  359. *(boffset + 0) = ctemp01;
  360. aoffset1 ++;
  361. boffset ++;
  362. i --;
  363. }while(i > 0);
  364. }
  365. } /* end of if(j > 0) */
  366. return 0;
  367. }