parallel_independent_heterogeneous_tasks.c 4.1 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2016 Bérangère Subervie
  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. #include <stdbool.h>
  17. #include <starpu.h>
  18. #include "../helper.h"
  19. /* Run a series of independent tasks with heterogeneous execution time */
  20. #define TIME 0.010
  21. #ifdef STARPU_QUICK_CHECK
  22. #define TASK_COEFFICIENT 20
  23. #define MARGIN 0.20
  24. #else
  25. #define TASK_COEFFICIENT 100
  26. #define MARGIN 0.10
  27. #endif
  28. #define TIME_CUDA_COEFFICIENT 10
  29. #define TIME_OPENCL_COEFFICIENT 5
  30. #define SECONDS_SCALE_COEFFICIENT_TIMING_NOW 1000000
  31. void wait_CPU(void *descr[] STARPU_ATTRIBUTE_UNUSED, void *_args)
  32. {
  33. starpu_sleep(TIME);
  34. }
  35. void wait_CUDA(void *descr[] STARPU_ATTRIBUTE_UNUSED, void *_args)
  36. {
  37. starpu_sleep(TIME/TIME_CUDA_COEFFICIENT);
  38. }
  39. void wait_OPENCL(void *descr[] STARPU_ATTRIBUTE_UNUSED, void *_args)
  40. {
  41. starpu_sleep(TIME/TIME_OPENCL_COEFFICIENT);
  42. }
  43. double cost_function(struct starpu_task *t, struct starpu_perfmodel_arch *a, unsigned i)
  44. {
  45. (void) t; (void) i;
  46. STARPU_ASSERT(a->ndevices == 1);
  47. if (a->devices[0].type == STARPU_CPU_WORKER)
  48. {
  49. STARPU_ASSERT(a->devices[0].ncores == 1);
  50. return TIME * 1000000;
  51. }
  52. else if (a->devices[0].type == STARPU_CUDA_WORKER)
  53. {
  54. return TIME/TIME_CUDA_COEFFICIENT * 1000000;
  55. }
  56. else if (a->devices[0].type == STARPU_OPENCL_WORKER)
  57. {
  58. return TIME/TIME_OPENCL_COEFFICIENT * 1000000;
  59. }
  60. STARPU_ASSERT(0);
  61. return 0.0;
  62. }
  63. static struct starpu_perfmodel perf_model =
  64. {
  65. .type = STARPU_PER_ARCH,
  66. .arch_cost_function = cost_function,
  67. };
  68. static struct starpu_codelet cl =
  69. {
  70. .cpu_funcs = { wait_CPU },
  71. .cuda_funcs = { wait_CUDA },
  72. .opencl_funcs = { wait_OPENCL },
  73. .cpu_funcs_name = { "wait_CPU" },
  74. .nbuffers = 0,
  75. .flags = STARPU_CODELET_SIMGRID_EXECUTE,
  76. .model = &perf_model,
  77. };
  78. int main(int argc, char *argv[])
  79. {
  80. int ret;
  81. ret = starpu_initialize(NULL, &argc, &argv);
  82. if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
  83. STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
  84. unsigned nb_tasks, nb_workers_CPU, nb_workers_CUDA, nb_workers_OPENCL, i;
  85. double begin_time, end_time, time_m, time_s, speed_up, expected_speed_up, percentage_expected_speed_up;
  86. bool check, check_sup;
  87. nb_workers_CPU = starpu_worker_get_count_by_type(STARPU_CPU_WORKER);
  88. nb_workers_CUDA = starpu_worker_get_count_by_type(STARPU_CUDA_WORKER);
  89. nb_workers_OPENCL = starpu_worker_get_count_by_type(STARPU_OPENCL_WORKER);
  90. nb_tasks = (nb_workers_CPU + nb_workers_CUDA + nb_workers_OPENCL)*TASK_COEFFICIENT;
  91. begin_time = starpu_timing_now();
  92. /*execution des tasks*/
  93. for (i=0; i<nb_tasks; i++)
  94. {
  95. starpu_task_insert(&cl,0);
  96. }
  97. starpu_task_wait_for_all();
  98. end_time = starpu_timing_now();
  99. /*on determine si le temps mesure est satisfaisant ou pas*/
  100. time_m = (end_time - begin_time)/SECONDS_SCALE_COEFFICIENT_TIMING_NOW; //pour ramener en secondes
  101. time_s = nb_tasks * TIME;
  102. speed_up = time_s/time_m;
  103. expected_speed_up = nb_workers_CPU + TIME_CUDA_COEFFICIENT*nb_workers_CUDA + TIME_OPENCL_COEFFICIENT*nb_workers_OPENCL;
  104. percentage_expected_speed_up = 100 * (speed_up/expected_speed_up);
  105. check = speed_up >= (1 - MARGIN) * expected_speed_up;
  106. check_sup = speed_up <= (1 + MARGIN) * expected_speed_up;
  107. printf("measured time = %f seconds\nsequential time = %f seconds\nspeed up = %f\nnumber of workers CPU = %u\nnumber of workers CUDA = %u\nnumber of workers OPENCL = %u\nnumber of tasks = %u\nexpected speed up = %f\npercentage of expected speed up = %.2f%%\n", time_m, time_s, speed_up, nb_workers_CPU, nb_workers_CUDA, nb_workers_OPENCL, nb_tasks, expected_speed_up, percentage_expected_speed_up);
  108. starpu_shutdown();
  109. //test reussi ou test echoue
  110. if (check && check_sup)
  111. {
  112. return EXIT_SUCCESS;
  113. }
  114. else
  115. {
  116. return EXIT_FAILURE;
  117. }
  118. }