xlu.c 6.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2009, 2010-2011 Université de Bordeaux 1
  4. * Copyright (C) 2010 Mehdi Juhoor <mjuhoor@gmail.com>
  5. * Copyright (C) 2010, 2011, 2012 Centre National de la Recherche Scientifique
  6. *
  7. * StarPU is free software; you can redistribute it and/or modify
  8. * it under the terms of the GNU Lesser General Public License as published by
  9. * the Free Software Foundation; either version 2.1 of the License, or (at
  10. * your option) any later version.
  11. *
  12. * StarPU is distributed in the hope that it will be useful, but
  13. * WITHOUT ANY WARRANTY; without even the implied warranty of
  14. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  15. *
  16. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  17. */
  18. #include "xlu.h"
  19. #include "xlu_kernels.h"
  20. #define TAG11(k) ((starpu_tag_t)( (1ULL<<60) | (unsigned long long)(k)))
  21. #define TAG12(k,i) ((starpu_tag_t)(((2ULL<<60) | (((unsigned long long)(k))<<32) \
  22. | (unsigned long long)(i))))
  23. #define TAG21(k,j) ((starpu_tag_t)(((3ULL<<60) | (((unsigned long long)(k))<<32) \
  24. | (unsigned long long)(j))))
  25. #define TAG22(k,i,j) ((starpu_tag_t)(((4ULL<<60) | ((unsigned long long)(k)<<32) \
  26. | ((unsigned long long)(i)<<16) \
  27. | (unsigned long long)(j))))
  28. static unsigned no_prio = 0;
  29. /*
  30. * Construct the DAG
  31. */
  32. static struct starpu_task *create_task(starpu_tag_t id)
  33. {
  34. struct starpu_task *task = starpu_task_create();
  35. task->cl_arg = NULL;
  36. task->use_tag = 1;
  37. task->tag_id = id;
  38. return task;
  39. }
  40. static struct starpu_task *create_task_11(starpu_data_handle_t dataA, unsigned k)
  41. {
  42. /* printf("task 11 k = %d TAG = %llx\n", k, (TAG11(k))); */
  43. struct starpu_task *task = create_task(TAG11(k));
  44. task->cl = &cl11;
  45. /* which sub-data is manipulated ? */
  46. task->handles[0] = starpu_data_get_sub_data(dataA, 2, k, k);
  47. /* this is an important task */
  48. if (!no_prio)
  49. task->priority = STARPU_MAX_PRIO;
  50. /* enforce dependencies ... */
  51. if (k > 0)
  52. {
  53. starpu_tag_declare_deps(TAG11(k), 1, TAG22(k-1, k, k));
  54. }
  55. return task;
  56. }
  57. static void create_task_12(starpu_data_handle_t dataA, unsigned k, unsigned j)
  58. {
  59. /* printf("task 12 k,i = %d,%d TAG = %llx\n", k,i, TAG12(k,i)); */
  60. struct starpu_task *task = create_task(TAG12(k, j));
  61. task->cl = &cl12;
  62. /* which sub-data is manipulated ? */
  63. task->handles[0] = starpu_data_get_sub_data(dataA, 2, k, k);
  64. task->handles[1] = starpu_data_get_sub_data(dataA, 2, j, k);
  65. if (!no_prio && (j == k+1))
  66. {
  67. task->priority = STARPU_MAX_PRIO;
  68. }
  69. /* enforce dependencies ... */
  70. if (k > 0)
  71. {
  72. starpu_tag_declare_deps(TAG12(k, j), 2, TAG11(k), TAG22(k-1, k, j));
  73. }
  74. else
  75. {
  76. starpu_tag_declare_deps(TAG12(k, j), 1, TAG11(k));
  77. }
  78. starpu_task_submit(task);
  79. }
  80. static void create_task_21(starpu_data_handle_t dataA, unsigned k, unsigned i)
  81. {
  82. struct starpu_task *task = create_task(TAG21(k, i));
  83. task->cl = &cl21;
  84. /* which sub-data is manipulated ? */
  85. task->handles[0] = starpu_data_get_sub_data(dataA, 2, k, k);
  86. task->handles[1] = starpu_data_get_sub_data(dataA, 2, k, i);
  87. if (!no_prio && (i == k+1))
  88. {
  89. task->priority = STARPU_MAX_PRIO;
  90. }
  91. /* enforce dependencies ... */
  92. if (k > 0)
  93. {
  94. starpu_tag_declare_deps(TAG21(k, i), 2, TAG11(k), TAG22(k-1, i, k));
  95. }
  96. else
  97. {
  98. starpu_tag_declare_deps(TAG21(k, i), 1, TAG11(k));
  99. }
  100. starpu_task_submit(task);
  101. }
  102. static void create_task_22(starpu_data_handle_t dataA, unsigned k, unsigned i, unsigned j)
  103. {
  104. /* printf("task 22 k,i,j = %d,%d,%d TAG = %llx\n", k,i,j, TAG22(k,i,j)); */
  105. struct starpu_task *task = create_task(TAG22(k, i, j));
  106. task->cl = &cl22;
  107. /* which sub-data is manipulated ? */
  108. task->handles[0] = starpu_data_get_sub_data(dataA, 2, k, i); /* produced by TAG21(k, i) */
  109. task->handles[1] = starpu_data_get_sub_data(dataA, 2, j, k); /* produced by TAG12(k, j) */
  110. task->handles[2] = starpu_data_get_sub_data(dataA, 2, j, i); /* produced by TAG22(k-1, i, j) */
  111. if (!no_prio && (i == k + 1) && (j == k +1) )
  112. {
  113. task->priority = STARPU_MAX_PRIO;
  114. }
  115. /* enforce dependencies ... */
  116. if (k > 0)
  117. {
  118. starpu_tag_declare_deps(TAG22(k, i, j), 3, TAG22(k-1, i, j), TAG12(k, j), TAG21(k, i));
  119. }
  120. else
  121. {
  122. starpu_tag_declare_deps(TAG22(k, i, j), 2, TAG12(k, j), TAG21(k, i));
  123. }
  124. starpu_task_submit(task);
  125. }
  126. /*
  127. * code to bootstrap the factorization
  128. */
  129. static void dw_codelet_facto_v3(starpu_data_handle_t dataA, unsigned nblocks)
  130. {
  131. struct timeval start;
  132. struct timeval end;
  133. struct starpu_task *entry_task = NULL;
  134. /* create all the DAG nodes */
  135. unsigned i,j,k;
  136. for (k = 0; k < nblocks; k++)
  137. {
  138. struct starpu_task *task = create_task_11(dataA, k);
  139. /* we defer the launch of the first task */
  140. if (k == 0)
  141. {
  142. entry_task = task;
  143. }
  144. else
  145. {
  146. starpu_task_submit(task);
  147. }
  148. for (i = k+1; i<nblocks; i++)
  149. {
  150. create_task_12(dataA, k, i);
  151. create_task_21(dataA, k, i);
  152. }
  153. for (i = k+1; i<nblocks; i++)
  154. {
  155. for (j = k+1; j<nblocks; j++)
  156. {
  157. create_task_22(dataA, k, i, j);
  158. }
  159. }
  160. }
  161. /* schedule the codelet */
  162. gettimeofday(&start, NULL);
  163. int ret = starpu_task_submit(entry_task);
  164. if (STARPU_UNLIKELY(ret == -ENODEV))
  165. {
  166. FPRINTF(stderr, "No worker may execute this task\n");
  167. exit(-1);
  168. }
  169. /* stall the application until the end of computations */
  170. starpu_tag_wait(TAG11(nblocks-1));
  171. gettimeofday(&end, NULL);
  172. double timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
  173. FPRINTF(stderr, "Computation took (in ms)\n");
  174. FPRINTF(stdout, "%2.2f\n", timing/1000);
  175. unsigned n = starpu_matrix_get_nx(dataA);
  176. double flop = (2.0f*n*n*n)/3.0f;
  177. FPRINTF(stderr, "Synthetic GFlops : %2.2f\n", (flop/timing/1000.0f));
  178. }
  179. void STARPU_LU(lu_decomposition)(TYPE *matA, unsigned size, unsigned ld, unsigned nblocks)
  180. {
  181. starpu_data_handle_t dataA;
  182. /* monitor and partition the A matrix into blocks :
  183. * one block is now determined by 2 unsigned (i,j) */
  184. starpu_matrix_data_register(&dataA, 0, (uintptr_t)matA, ld, size, size, sizeof(TYPE));
  185. /* We already enforce deps by hand */
  186. starpu_data_set_sequential_consistency_flag(dataA, 0);
  187. struct starpu_data_filter f =
  188. {
  189. .filter_func = starpu_vertical_block_filter_func,
  190. .nchildren = nblocks
  191. };
  192. struct starpu_data_filter f2 =
  193. {
  194. .filter_func = starpu_block_filter_func,
  195. .nchildren = nblocks
  196. };
  197. starpu_data_map_filters(dataA, 2, &f, &f2);
  198. dw_codelet_facto_v3(dataA, nblocks);
  199. /* gather all the data */
  200. starpu_data_unpartition(dataA, 0);
  201. starpu_data_unregister(dataA);
  202. }