vector_scal_opencl.texi 1.2 KB

1234567891011121314151617181920212223242526272829303132333435363738394041
  1. #include <starpu.h>
  2. #include <starpu_opencl.h>
  3. extern struct starpu_opencl_codelet codelet;
  4. void scal_opencl_func(void *buffers[], void *_args)
  5. @{
  6. float *factor = (float *)_args;
  7. int id, devid, err;
  8. cl_kernel kernel;
  9. cl_command_queue queue;
  10. /* length of the vector */
  11. unsigned n = STARPU_GET_VECTOR_NX(buffers[0]);
  12. /* local copy of the vector pointer */
  13. float *val = (float *)STARPU_GET_VECTOR_PTR(buffers[0]);
  14. id = starpu_worker_get_id();
  15. devid = starpu_worker_get_devid(id);
  16. err = starpu_opencl_load_kernel(&kernel, &queue, &codelet, "vector_mult_opencl", devid);
  17. if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
  18. err = 0;
  19. err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &val);
  20. err = clSetKernelArg(kernel, 1, sizeof(int), &n);
  21. err |= clSetKernelArg(kernel, 2, sizeof(float), (void*)factor);
  22. if (err) STARPU_OPENCL_REPORT_ERROR(err);
  23. @{
  24. size_t global=n;
  25. size_t local=n;
  26. err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 0, NULL, NULL);
  27. if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
  28. @}
  29. clFinish(queue);
  30. starpu_opencl_release_kernel(kernel);
  31. @}