matvecmult.c 6.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2010, 2011-2012, 2014 Université de Bordeaux
  4. * Copyright (C) 2010, 2011, 2012, 2013, 2014 Centre National de la Recherche Scientifique
  5. *
  6. * StarPU is free software; you can redistribute it and/or modify
  7. * it under the terms of the GNU Lesser General Public License as published by
  8. * the Free Software Foundation; either version 2.1 of the License, or (at
  9. * your option) any later version.
  10. *
  11. * StarPU is distributed in the hope that it will be useful, but
  12. * WITHOUT ANY WARRANTY; without even the implied warranty of
  13. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  14. *
  15. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  16. */
  17. #include <starpu.h>
  18. #include <math.h>
  19. #define FPRINTF(ofile, fmt, ...) do { if (!getenv("STARPU_SSILENT")) {fprintf(ofile, fmt, ## __VA_ARGS__); }} while(0)
  20. #ifdef STARPU_USE_OPENCL
  21. struct starpu_opencl_program opencl_code;
  22. void opencl_codelet(void *descr[], STARPU_ATTRIBUTE_UNUSED void *_args)
  23. {
  24. cl_kernel kernel;
  25. cl_command_queue queue;
  26. int id, devid, err, n;
  27. cl_mem matrix = (cl_mem)STARPU_MATRIX_GET_DEV_HANDLE(descr[0]);
  28. cl_mem vector = (cl_mem)STARPU_VECTOR_GET_DEV_HANDLE(descr[1]);
  29. cl_mem mult = (cl_mem)STARPU_VECTOR_GET_DEV_HANDLE(descr[2]);
  30. int nx = STARPU_MATRIX_GET_NX(descr[0]);
  31. int ny = STARPU_MATRIX_GET_NY(descr[0]);
  32. cl_event event;
  33. id = starpu_worker_get_id();
  34. devid = starpu_worker_get_devid(id);
  35. err = starpu_opencl_load_kernel(&kernel, &queue, &opencl_code, "matVecMult", devid);
  36. if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
  37. n=0;
  38. err = clSetKernelArg(kernel, n++, sizeof(matrix), &matrix);
  39. err |= clSetKernelArg(kernel, n++, sizeof(vector), &vector);
  40. err |= clSetKernelArg(kernel, n++, sizeof(nx), (void*)&nx);
  41. err |= clSetKernelArg(kernel, n++, sizeof(ny), (void*)&ny);
  42. err |= clSetKernelArg(kernel, n++, sizeof(mult), &mult);
  43. if (err) STARPU_OPENCL_REPORT_ERROR(err);
  44. {
  45. size_t global=nx*ny;
  46. err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, NULL, 0, NULL, NULL);
  47. if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
  48. }
  49. starpu_opencl_release_kernel(kernel);
  50. }
  51. #endif
  52. void fillArray(float* pfData, int iSize)
  53. {
  54. int i;
  55. const float fScale = 1.0f / (float)RAND_MAX;
  56. for (i = 0; i < iSize; ++i)
  57. {
  58. pfData[i] = fScale * rand();
  59. }
  60. }
  61. void printArray(float* pfData, int iSize)
  62. {
  63. int i;
  64. for (i = 0; i < iSize; ++i)
  65. {
  66. FPRINTF(stderr, "%f ", pfData[i]);
  67. }
  68. FPRINTF(stderr, "\n");
  69. }
  70. void matVecMult(const float *matrix, const float *vector, int width, int height, float *mult)
  71. {
  72. int i, j;
  73. for (i = 0; i < height; ++i)
  74. {
  75. double sum = 0;
  76. for (j = 0; j < width; ++j)
  77. {
  78. double a = matrix[i * width + j];
  79. double b = vector[j];
  80. sum += a * b;
  81. }
  82. mult[i] = (float)sum;
  83. }
  84. }
  85. int compareL2fe(const float* reference, const float* data, const unsigned int len, const float epsilon)
  86. {
  87. float error = 0;
  88. float ref = 0;
  89. unsigned int i;
  90. for(i = 0; i < len; ++i)
  91. {
  92. float diff = reference[i] - data[i];
  93. error += diff * diff;
  94. ref += reference[i] * reference[i];
  95. }
  96. float normRef = sqrtf(ref);
  97. if (fabs(ref) < 1e-7) return 1;
  98. float normError = sqrtf(error);
  99. error = normError / normRef;
  100. return error < epsilon ? 0 : 1;
  101. }
  102. static struct starpu_perfmodel starpu_matvecmult_model =
  103. {
  104. .type = STARPU_HISTORY_BASED,
  105. .symbol = "matvecmult"
  106. };
  107. static struct starpu_codelet cl =
  108. {
  109. #ifdef STARPU_USE_OPENCL
  110. .opencl_funcs[0] = opencl_codelet,
  111. .opencl_flags = {STARPU_OPENCL_ASYNC},
  112. #endif
  113. .nbuffers = 3,
  114. .modes[0] = STARPU_R,
  115. .modes[1] = STARPU_R,
  116. .modes[2] = STARPU_RW,
  117. .model = &starpu_matvecmult_model
  118. };
  119. int main(int argc, char **argv)
  120. {
  121. struct starpu_conf conf;
  122. starpu_conf_init(&conf);
  123. conf.ncpus = 0;
  124. conf.ncuda = 0;
  125. conf.nopencl = 1;
  126. /* int width=1100; */
  127. /* int height=244021; */
  128. int width=20;
  129. int height=4;
  130. float *matrix, *vector, *mult;
  131. float *correctResult;
  132. unsigned int mem_size_matrix, mem_size_vector, mem_size_mult;
  133. starpu_data_handle_t matrix_handle, vector_handle, mult_handle;
  134. int ret, submit;
  135. ret = starpu_init(&conf);
  136. if (STARPU_UNLIKELY(ret == -ENODEV))
  137. {
  138. FPRINTF(stderr, "This application requires an OpenCL worker.\n");
  139. return 77;
  140. }
  141. mem_size_matrix = width * height * sizeof(float);
  142. matrix = (float*)malloc(mem_size_matrix);
  143. mem_size_vector = width * sizeof(float);
  144. vector = (float*)malloc(mem_size_vector);
  145. mem_size_mult = height * sizeof(float);
  146. mult = (float*)malloc(mem_size_mult);
  147. correctResult = (float*)malloc(mem_size_mult);
  148. assert(matrix);
  149. assert(vector);
  150. assert(mult);
  151. assert(correctResult);
  152. fillArray(matrix, width*height);
  153. fillArray(vector, width);
  154. fillArray(mult, height);
  155. matVecMult(matrix, vector, width, height, correctResult);
  156. starpu_matrix_data_register(&matrix_handle, STARPU_MAIN_RAM, (uintptr_t)matrix, width, width, height, sizeof(float));
  157. starpu_vector_data_register(&vector_handle, STARPU_MAIN_RAM, (uintptr_t)vector, width, sizeof(float));
  158. starpu_vector_data_register(&mult_handle, STARPU_MAIN_RAM, (uintptr_t)mult, height, sizeof(float));
  159. #ifdef STARPU_USE_OPENCL
  160. ret = starpu_opencl_load_opencl_from_file("examples/matvecmult/matvecmult_kernel.cl", &opencl_code, NULL);
  161. STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_load_opencl_from_file");
  162. #endif
  163. struct starpu_task *task = starpu_task_create();
  164. task->cl = &cl;
  165. task->callback_func = NULL;
  166. task->handles[0] = matrix_handle;
  167. task->handles[1] = vector_handle;
  168. task->handles[2] = mult_handle;
  169. submit = starpu_task_submit(task);
  170. if (STARPU_UNLIKELY(submit == -ENODEV))
  171. {
  172. FPRINTF(stderr, "No worker may execute this task. This application requires an OpenCL worker.\n");
  173. }
  174. else
  175. {
  176. starpu_task_wait_for_all();
  177. }
  178. starpu_data_unregister(matrix_handle);
  179. starpu_data_unregister(vector_handle);
  180. starpu_data_unregister(mult_handle);
  181. if (STARPU_LIKELY(submit != -ENODEV))
  182. {
  183. int res = compareL2fe(correctResult, mult, height, 1e-6f);
  184. FPRINTF(stdout, "TEST %s\n\n", (res == 0) ? "PASSED" : "FAILED !!!");
  185. }
  186. #if 0
  187. printArray(matrix, width*height);
  188. printArray(vector, width);
  189. printArray(mult, height);
  190. #endif
  191. starpu_shutdown();
  192. return (submit == -ENODEV) ? 77 : 0;
  193. }