vector_scal_opencl.texi 1.3 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546
  1. /*
  2. * This example complements vector_scale.c: here we implement a OpenCL version.
  3. */
  4. #include <starpu.h>
  5. #include <starpu_opencl.h>
  6. extern struct starpu_opencl_codelet codelet;
  7. void scal_opencl_func(void *buffers[], void *_args)
  8. @{
  9. float *factor = (float *)_args;
  10. struct starpu_vector_interface_s *vector = (struct starpu_vector_interface_s *) buffers[0];
  11. int id, devid, err;
  12. cl_kernel kernel;
  13. cl_command_queue queue;
  14. /* length of the vector */
  15. unsigned n = STARPU_GET_VECTOR_NX(vector);
  16. /* local copy of the vector pointer */
  17. float *val = (float *)STARPU_GET_VECTOR_PTR(vector);
  18. id = starpu_worker_get_id();
  19. devid = starpu_worker_get_devid(id);
  20. err = starpu_opencl_load_kernel(&kernel, &queue, &codelet, "vector_mult_opencl", devid);
  21. if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
  22. err = 0;
  23. err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &val);
  24. err = clSetKernelArg(kernel, 1, sizeof(int), &n);
  25. err |= clSetKernelArg(kernel, 2, sizeof(float), (void*)factor);
  26. if (err) STARPU_OPENCL_REPORT_ERROR(err);
  27. @{
  28. size_t global=n;
  29. size_t local=n;
  30. err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 0, NULL, NULL);
  31. if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
  32. @}
  33. clFinish(queue);
  34. starpu_opencl_release_kernel(kernel);
  35. @}