fmultiple_manual.c 7.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2015 Université Bordeaux
  4. *
  5. * StarPU is free software; you can redistribute it and/or modify
  6. * it under the terms of the GNU Lesser General Public License as published by
  7. * the Free Software Foundation; either version 2.1 of the License, or (at
  8. * your option) any later version.
  9. *
  10. * StarPU is distributed in the hope that it will be useful, but
  11. * WITHOUT ANY WARRANTY; without even the implied warranty of
  12. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  13. *
  14. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  15. */
  16. /*
  17. * This examplifies how to access the same matrix with different partitioned
  18. * views, doing the coherency by hand.
  19. * We first run a kernel on the whole matrix to fill it, then run a kernel on
  20. * each vertical slice to check the value and multiply it by two, then run a
  21. * kernel on each horizontal slice to do the same.
  22. */
  23. #include <starpu.h>
  24. #define NX 6
  25. #define NY 6
  26. #define PARTS 2
  27. #define FPRINTF(ofile, fmt, ...) do { if (!getenv("STARPU_SSILENT")) {fprintf(ofile, fmt, ## __VA_ARGS__); }} while(0)
  28. void matrix_fill(void *buffers[], void *cl_arg STARPU_ATTRIBUTE_UNUSED)
  29. {
  30. unsigned i, j;
  31. /* length of the matrix */
  32. unsigned nx = STARPU_MATRIX_GET_NX(buffers[0]);
  33. unsigned ny = STARPU_MATRIX_GET_NY(buffers[0]);
  34. unsigned ld = STARPU_MATRIX_GET_LD(buffers[0]);
  35. int *val = (int *)STARPU_MATRIX_GET_PTR(buffers[0]);
  36. for(j=0; j<ny ; j++)
  37. {
  38. for(i=0; i<nx ; i++)
  39. val[(j*ld)+i] = i+100*j;
  40. }
  41. }
  42. struct starpu_codelet cl_fill =
  43. {
  44. .cpu_funcs = {matrix_fill},
  45. .cpu_funcs_name = {"matrix_fill"},
  46. .nbuffers = 1,
  47. .modes = {STARPU_W},
  48. .name = "matrix_fill"
  49. };
  50. void fmultiple_check_scale(void *buffers[], void *cl_arg)
  51. {
  52. int start, factor;
  53. unsigned i, j;
  54. /* length of the matrix */
  55. unsigned nx = STARPU_MATRIX_GET_NX(buffers[0]);
  56. unsigned ny = STARPU_MATRIX_GET_NY(buffers[0]);
  57. unsigned ld = STARPU_MATRIX_GET_LD(buffers[0]);
  58. int *val = (int *)STARPU_MATRIX_GET_PTR(buffers[0]);
  59. starpu_codelet_unpack_args(cl_arg, &start, &factor);
  60. for(j=0; j<ny ; j++)
  61. {
  62. for(i=0; i<nx ; i++)
  63. {
  64. STARPU_ASSERT(val[(j*ld)+i] == start + factor*((int)(i+100*j)));
  65. val[(j*ld)+i] *= 2;
  66. }
  67. }
  68. }
  69. #ifdef STARPU_USE_CUDA
  70. extern void fmultiple_check_scale_cuda(void *buffers[], void *cl_arg);
  71. #endif
  72. struct starpu_codelet cl_check_scale =
  73. {
  74. #ifdef STARPU_USE_CUDA
  75. .cuda_funcs = {fmultiple_check_scale_cuda},
  76. .cuda_flags = {STARPU_CUDA_ASYNC},
  77. #else
  78. /* Only enable it on CPUs if we don't have a CUDA device, to force remote execution on the CUDA device */
  79. .cpu_funcs = {fmultiple_check_scale},
  80. .cpu_funcs_name = {"fmultiple_check_scale"},
  81. #endif
  82. .nbuffers = 1,
  83. .modes = {STARPU_RW},
  84. .name = "fmultiple_check_scale"
  85. };
  86. void empty(void *buffers[] STARPU_ATTRIBUTE_UNUSED, void *cl_arg STARPU_ATTRIBUTE_UNUSED)
  87. {
  88. /* This doesn't need to do anything, it's simply used to make coherency
  89. * between the two views, by simply running on the home node of the
  90. * data, thus getting back all data pieces there. */
  91. }
  92. struct starpu_codelet cl_switch =
  93. {
  94. .cpu_funcs = {empty},
  95. .nbuffers = STARPU_VARIABLE_NBUFFERS,
  96. .name = "switch"
  97. };
  98. int main(int argc, char **argv)
  99. {
  100. unsigned j, n=1;
  101. int matrix[NX][NY];
  102. int ret, i;
  103. /* We haven't taken care otherwise */
  104. STARPU_ASSERT((NX%PARTS) == 0);
  105. STARPU_ASSERT((NY%PARTS) == 0);
  106. starpu_data_handle_t handle;
  107. starpu_data_handle_t vert_handle[PARTS];
  108. starpu_data_handle_t horiz_handle[PARTS];
  109. ret = starpu_init(NULL);
  110. if (ret == -ENODEV)
  111. return 77;
  112. STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
  113. /* force to execute task on the home_node, here it is STARPU_MAIN_RAM */
  114. cl_switch.specific_nodes = 1;
  115. for (i = 0; i < STARPU_NMAXBUFS; i++)
  116. cl_switch.nodes[i] = STARPU_MAIN_RAM;
  117. /* Declare the whole matrix to StarPU */
  118. starpu_matrix_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)matrix, NX, NX, NY, sizeof(matrix[0][0]));
  119. /* Also declare the vertical slices to StarPU */
  120. for (i = 0; i < PARTS; i++)
  121. {
  122. starpu_matrix_data_register(&vert_handle[i], STARPU_MAIN_RAM, (uintptr_t)&matrix[0][i*(NX/PARTS)], NX, NX/PARTS, NY, sizeof(matrix[0][0]));
  123. /* But make it invalid for now, we'll access data through the whole matrix first */
  124. starpu_data_invalidate(vert_handle[i]);
  125. }
  126. /* And the horizontal slices to StarPU */
  127. for (i = 0; i < PARTS; i++)
  128. {
  129. starpu_matrix_data_register(&horiz_handle[i], STARPU_MAIN_RAM, (uintptr_t)&matrix[i*(NY/PARTS)][0], NX, NX, NY/PARTS, sizeof(matrix[0][0]));
  130. starpu_data_invalidate(horiz_handle[i]);
  131. }
  132. /* Fill the matrix */
  133. ret = starpu_task_insert(&cl_fill, STARPU_W, handle, 0);
  134. if (ret == -ENODEV) goto enodev;
  135. STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  136. /* Now switch to vertical view of the matrix */
  137. struct starpu_data_descr vert_descr[PARTS];
  138. for (i = 0; i < PARTS; i++)
  139. {
  140. vert_descr[i].handle = vert_handle[i];
  141. vert_descr[i].mode = STARPU_W;
  142. }
  143. ret = starpu_task_insert(&cl_switch, STARPU_RW, handle, STARPU_DATA_MODE_ARRAY, vert_descr, PARTS, 0);
  144. if (ret == -ENODEV) goto enodev;
  145. STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  146. /* And make sure we don't accidentally access the matrix through the whole-matrix handle */
  147. starpu_data_invalidate_submit(handle);
  148. /* Check the values of the vertical slices */
  149. for (i = 0; i < PARTS; i++)
  150. {
  151. int factor = 1;
  152. int start = i*(NX/PARTS);
  153. ret = starpu_task_insert(&cl_check_scale,
  154. STARPU_RW, vert_handle[i],
  155. STARPU_VALUE, &start, sizeof(start),
  156. STARPU_VALUE, &factor, sizeof(factor),
  157. 0);
  158. if (ret == -ENODEV) goto enodev;
  159. STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  160. }
  161. /* Now switch back to total view of the matrix */
  162. for (i = 0; i < PARTS; i++)
  163. vert_descr[i].mode = STARPU_RW;
  164. ret = starpu_task_insert(&cl_switch, STARPU_DATA_MODE_ARRAY, vert_descr, PARTS, STARPU_W, handle, 0);
  165. if (ret == -ENODEV) goto enodev;
  166. STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  167. /* And make sure we don't accidentally access the matrix through the vertical slices */
  168. for (i = 0; i < PARTS; i++)
  169. starpu_data_invalidate_submit(vert_handle[i]);
  170. /* And switch to horizontal view of the matrix */
  171. struct starpu_data_descr horiz_descr[PARTS];
  172. for (i = 0; i < PARTS; i++)
  173. {
  174. horiz_descr[i].handle = horiz_handle[i];
  175. horiz_descr[i].mode = STARPU_W;
  176. }
  177. ret = starpu_task_insert(&cl_switch, STARPU_RW, handle, STARPU_DATA_MODE_ARRAY, horiz_descr, PARTS, 0);
  178. if (ret == -ENODEV) goto enodev;
  179. STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  180. /* And make sure we don't accidentally access the matrix through the whole-matrix handle */
  181. starpu_data_invalidate_submit(handle);
  182. /* Check the values of the horizontal slices */
  183. for (i = 0; i < PARTS; i++)
  184. {
  185. int factor = 2;
  186. int start = factor*100*i*(NY/PARTS);
  187. ret = starpu_task_insert(&cl_check_scale,
  188. STARPU_RW, horiz_handle[i],
  189. STARPU_VALUE, &start, sizeof(start),
  190. STARPU_VALUE, &factor, sizeof(factor),
  191. 0);
  192. if (ret == -ENODEV) goto enodev;
  193. STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
  194. }
  195. /*
  196. * Unregister data from StarPU and shutdown It does not really matter
  197. * which view is active at unregistration here, since all views cover
  198. * the whole matrix, so it will be completely updated in the main memory.
  199. */
  200. for (i = 0; i < PARTS; i++)
  201. {
  202. starpu_data_unregister(vert_handle[i]);
  203. starpu_data_unregister(horiz_handle[i]);
  204. }
  205. starpu_data_unregister(handle);
  206. starpu_shutdown();
  207. return ret;
  208. enodev:
  209. starpu_shutdown();
  210. return 77;
  211. }