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.

zgemm_tcopy_4.c 11 kB

9 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403
  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 *boffset, *boffset1, *boffset2, *boffset3;
  45. FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
  46. FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
  47. FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
  48. FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
  49. FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
  50. FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
  51. FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
  52. FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
  53. aoffset = a;
  54. boffset = b;
  55. lda *= 2;
  56. boffset2 = b + 2 * m * (n & ~3);
  57. boffset3 = b + 2 * m * (n & ~1);
  58. #if 0
  59. fprintf(stderr, "m = %d n = %d\n", m,n );
  60. #endif
  61. j = (m >> 2);
  62. if (j > 0){
  63. do{
  64. aoffset1 = aoffset;
  65. aoffset2 = aoffset1 + lda;
  66. aoffset3 = aoffset2 + lda;
  67. aoffset4 = aoffset3 + lda;
  68. aoffset += 4 * lda;
  69. boffset1 = boffset;
  70. boffset += 32;
  71. i = (n >> 2);
  72. if (i > 0){
  73. do{
  74. ctemp01 = *(aoffset1 + 0);
  75. ctemp02 = *(aoffset1 + 1);
  76. ctemp03 = *(aoffset1 + 2);
  77. ctemp04 = *(aoffset1 + 3);
  78. ctemp05 = *(aoffset1 + 4);
  79. ctemp06 = *(aoffset1 + 5);
  80. ctemp07 = *(aoffset1 + 6);
  81. ctemp08 = *(aoffset1 + 7);
  82. ctemp09 = *(aoffset2 + 0);
  83. ctemp10 = *(aoffset2 + 1);
  84. ctemp11 = *(aoffset2 + 2);
  85. ctemp12 = *(aoffset2 + 3);
  86. ctemp13 = *(aoffset2 + 4);
  87. ctemp14 = *(aoffset2 + 5);
  88. ctemp15 = *(aoffset2 + 6);
  89. ctemp16 = *(aoffset2 + 7);
  90. ctemp17 = *(aoffset3 + 0);
  91. ctemp18 = *(aoffset3 + 1);
  92. ctemp19 = *(aoffset3 + 2);
  93. ctemp20 = *(aoffset3 + 3);
  94. ctemp21 = *(aoffset3 + 4);
  95. ctemp22 = *(aoffset3 + 5);
  96. ctemp23 = *(aoffset3 + 6);
  97. ctemp24 = *(aoffset3 + 7);
  98. ctemp25 = *(aoffset4 + 0);
  99. ctemp26 = *(aoffset4 + 1);
  100. ctemp27 = *(aoffset4 + 2);
  101. ctemp28 = *(aoffset4 + 3);
  102. ctemp29 = *(aoffset4 + 4);
  103. ctemp30 = *(aoffset4 + 5);
  104. ctemp31 = *(aoffset4 + 6);
  105. ctemp32 = *(aoffset4 + 7);
  106. *(boffset1 + 0) = ctemp01;
  107. *(boffset1 + 1) = ctemp02;
  108. *(boffset1 + 2) = ctemp03;
  109. *(boffset1 + 3) = ctemp04;
  110. *(boffset1 + 4) = ctemp05;
  111. *(boffset1 + 5) = ctemp06;
  112. *(boffset1 + 6) = ctemp07;
  113. *(boffset1 + 7) = ctemp08;
  114. *(boffset1 + 8) = ctemp09;
  115. *(boffset1 + 9) = ctemp10;
  116. *(boffset1 + 10) = ctemp11;
  117. *(boffset1 + 11) = ctemp12;
  118. *(boffset1 + 12) = ctemp13;
  119. *(boffset1 + 13) = ctemp14;
  120. *(boffset1 + 14) = ctemp15;
  121. *(boffset1 + 15) = ctemp16;
  122. *(boffset1 + 16) = ctemp17;
  123. *(boffset1 + 17) = ctemp18;
  124. *(boffset1 + 18) = ctemp19;
  125. *(boffset1 + 19) = ctemp20;
  126. *(boffset1 + 20) = ctemp21;
  127. *(boffset1 + 21) = ctemp22;
  128. *(boffset1 + 22) = ctemp23;
  129. *(boffset1 + 23) = ctemp24;
  130. *(boffset1 + 24) = ctemp25;
  131. *(boffset1 + 25) = ctemp26;
  132. *(boffset1 + 26) = ctemp27;
  133. *(boffset1 + 27) = ctemp28;
  134. *(boffset1 + 28) = ctemp29;
  135. *(boffset1 + 29) = ctemp30;
  136. *(boffset1 + 30) = ctemp31;
  137. *(boffset1 + 31) = ctemp32;
  138. aoffset1 += 8;
  139. aoffset2 += 8;
  140. aoffset3 += 8;
  141. aoffset4 += 8;
  142. boffset1 += m * 8;
  143. i --;
  144. }while(i > 0);
  145. }
  146. if (n & 2){
  147. ctemp01 = *(aoffset1 + 0);
  148. ctemp02 = *(aoffset1 + 1);
  149. ctemp03 = *(aoffset1 + 2);
  150. ctemp04 = *(aoffset1 + 3);
  151. ctemp05 = *(aoffset2 + 0);
  152. ctemp06 = *(aoffset2 + 1);
  153. ctemp07 = *(aoffset2 + 2);
  154. ctemp08 = *(aoffset2 + 3);
  155. ctemp09 = *(aoffset3 + 0);
  156. ctemp10 = *(aoffset3 + 1);
  157. ctemp11 = *(aoffset3 + 2);
  158. ctemp12 = *(aoffset3 + 3);
  159. ctemp13 = *(aoffset4 + 0);
  160. ctemp14 = *(aoffset4 + 1);
  161. ctemp15 = *(aoffset4 + 2);
  162. ctemp16 = *(aoffset4 + 3);
  163. *(boffset2 + 0) = ctemp01;
  164. *(boffset2 + 1) = ctemp02;
  165. *(boffset2 + 2) = ctemp03;
  166. *(boffset2 + 3) = ctemp04;
  167. *(boffset2 + 4) = ctemp05;
  168. *(boffset2 + 5) = ctemp06;
  169. *(boffset2 + 6) = ctemp07;
  170. *(boffset2 + 7) = ctemp08;
  171. *(boffset2 + 8) = ctemp09;
  172. *(boffset2 + 9) = ctemp10;
  173. *(boffset2 + 10) = ctemp11;
  174. *(boffset2 + 11) = ctemp12;
  175. *(boffset2 + 12) = ctemp13;
  176. *(boffset2 + 13) = ctemp14;
  177. *(boffset2 + 14) = ctemp15;
  178. *(boffset2 + 15) = ctemp16;
  179. aoffset1 += 4;
  180. aoffset2 += 4;
  181. aoffset3 += 4;
  182. aoffset4 += 4;
  183. boffset2 += 16;
  184. }
  185. if (n & 1){
  186. ctemp01 = *(aoffset1 + 0);
  187. ctemp02 = *(aoffset1 + 1);
  188. ctemp03 = *(aoffset2 + 0);
  189. ctemp04 = *(aoffset2 + 1);
  190. ctemp05 = *(aoffset3 + 0);
  191. ctemp06 = *(aoffset3 + 1);
  192. ctemp07 = *(aoffset4 + 0);
  193. ctemp08 = *(aoffset4 + 1);
  194. *(boffset3 + 0) = ctemp01;
  195. *(boffset3 + 1) = ctemp02;
  196. *(boffset3 + 2) = ctemp03;
  197. *(boffset3 + 3) = ctemp04;
  198. *(boffset3 + 4) = ctemp05;
  199. *(boffset3 + 5) = ctemp06;
  200. *(boffset3 + 6) = ctemp07;
  201. *(boffset3 + 7) = ctemp08;
  202. /* aoffset1 += 2;
  203. aoffset2 += 2;
  204. aoffset3 += 2;
  205. aoffset4 += 2; */
  206. boffset3 += 8;
  207. }
  208. j--;
  209. }while(j > 0);
  210. }
  211. if (m & 2){
  212. aoffset1 = aoffset;
  213. aoffset2 = aoffset1 + lda;
  214. aoffset += 2 * lda;
  215. boffset1 = boffset;
  216. boffset += 16;
  217. i = (n >> 2);
  218. if (i > 0){
  219. do{
  220. ctemp01 = *(aoffset1 + 0);
  221. ctemp02 = *(aoffset1 + 1);
  222. ctemp03 = *(aoffset1 + 2);
  223. ctemp04 = *(aoffset1 + 3);
  224. ctemp05 = *(aoffset1 + 4);
  225. ctemp06 = *(aoffset1 + 5);
  226. ctemp07 = *(aoffset1 + 6);
  227. ctemp08 = *(aoffset1 + 7);
  228. ctemp09 = *(aoffset2 + 0);
  229. ctemp10 = *(aoffset2 + 1);
  230. ctemp11 = *(aoffset2 + 2);
  231. ctemp12 = *(aoffset2 + 3);
  232. ctemp13 = *(aoffset2 + 4);
  233. ctemp14 = *(aoffset2 + 5);
  234. ctemp15 = *(aoffset2 + 6);
  235. ctemp16 = *(aoffset2 + 7);
  236. *(boffset1 + 0) = ctemp01;
  237. *(boffset1 + 1) = ctemp02;
  238. *(boffset1 + 2) = ctemp03;
  239. *(boffset1 + 3) = ctemp04;
  240. *(boffset1 + 4) = ctemp05;
  241. *(boffset1 + 5) = ctemp06;
  242. *(boffset1 + 6) = ctemp07;
  243. *(boffset1 + 7) = ctemp08;
  244. *(boffset1 + 8) = ctemp09;
  245. *(boffset1 + 9) = ctemp10;
  246. *(boffset1 + 10) = ctemp11;
  247. *(boffset1 + 11) = ctemp12;
  248. *(boffset1 + 12) = ctemp13;
  249. *(boffset1 + 13) = ctemp14;
  250. *(boffset1 + 14) = ctemp15;
  251. *(boffset1 + 15) = ctemp16;
  252. aoffset1 += 8;
  253. aoffset2 += 8;
  254. // aoffset3 += 8;
  255. // aoffset4 += 8;
  256. boffset1 += m * 8;
  257. i --;
  258. }while(i > 0);
  259. }
  260. if (n & 2){
  261. ctemp01 = *(aoffset1 + 0);
  262. ctemp02 = *(aoffset1 + 1);
  263. ctemp03 = *(aoffset1 + 2);
  264. ctemp04 = *(aoffset1 + 3);
  265. ctemp05 = *(aoffset2 + 0);
  266. ctemp06 = *(aoffset2 + 1);
  267. ctemp07 = *(aoffset2 + 2);
  268. ctemp08 = *(aoffset2 + 3);
  269. *(boffset2 + 0) = ctemp01;
  270. *(boffset2 + 1) = ctemp02;
  271. *(boffset2 + 2) = ctemp03;
  272. *(boffset2 + 3) = ctemp04;
  273. *(boffset2 + 4) = ctemp05;
  274. *(boffset2 + 5) = ctemp06;
  275. *(boffset2 + 6) = ctemp07;
  276. *(boffset2 + 7) = ctemp08;
  277. aoffset1 += 4;
  278. aoffset2 += 4;
  279. boffset2 += 8;
  280. }
  281. if (n & 1){
  282. ctemp01 = *(aoffset1 + 0);
  283. ctemp02 = *(aoffset1 + 1);
  284. ctemp03 = *(aoffset2 + 0);
  285. ctemp04 = *(aoffset2 + 1);
  286. *(boffset3 + 0) = ctemp01;
  287. *(boffset3 + 1) = ctemp02;
  288. *(boffset3 + 2) = ctemp03;
  289. *(boffset3 + 3) = ctemp04;
  290. /* aoffset1 += 2;
  291. aoffset2 += 2; */
  292. boffset3 += 4;
  293. }
  294. }
  295. if (m & 1){
  296. aoffset1 = aoffset;
  297. boffset1 = boffset;
  298. i = (n >> 2);
  299. if (i > 0){
  300. do{
  301. ctemp01 = *(aoffset1 + 0);
  302. ctemp02 = *(aoffset1 + 1);
  303. ctemp03 = *(aoffset1 + 2);
  304. ctemp04 = *(aoffset1 + 3);
  305. ctemp05 = *(aoffset1 + 4);
  306. ctemp06 = *(aoffset1 + 5);
  307. ctemp07 = *(aoffset1 + 6);
  308. ctemp08 = *(aoffset1 + 7);
  309. *(boffset1 + 0) = ctemp01;
  310. *(boffset1 + 1) = ctemp02;
  311. *(boffset1 + 2) = ctemp03;
  312. *(boffset1 + 3) = ctemp04;
  313. *(boffset1 + 4) = ctemp05;
  314. *(boffset1 + 5) = ctemp06;
  315. *(boffset1 + 6) = ctemp07;
  316. *(boffset1 + 7) = ctemp08;
  317. aoffset1 += 8;
  318. boffset1 += m * 8;
  319. i --;
  320. }while(i > 0);
  321. }
  322. if (n & 2){
  323. ctemp01 = *(aoffset1 + 0);
  324. ctemp02 = *(aoffset1 + 1);
  325. ctemp03 = *(aoffset1 + 2);
  326. ctemp04 = *(aoffset1 + 3);
  327. *(boffset2 + 0) = ctemp01;
  328. *(boffset2 + 1) = ctemp02;
  329. *(boffset2 + 2) = ctemp03;
  330. *(boffset2 + 3) = ctemp04;
  331. aoffset1 += 4;
  332. // boffset2 += 4;
  333. }
  334. if (n & 1){
  335. ctemp01 = *(aoffset1 + 0);
  336. ctemp02 = *(aoffset1 + 1);
  337. *(boffset3 + 0) = ctemp01;
  338. *(boffset3 + 1) = ctemp02;
  339. }
  340. }
  341. return 0;
  342. }