#include <stdio.h>
#include <stdlib.h>

#include <CL/cl.h>


const char * program_src =
"#define __global __attribute__((address_space(1)))\n"
"__kernel\n"
"void pi(__global float * out) \n"
"{\n"
"  out[0] = 3.14159f;\n"
"}\n";


int main(int argc, char ** argv)
{
   cl_int error;

   cl_uint total_platforms;
   cl_platform_id platform_id;

   cl_uint total_gpu_devices;
   cl_device_id device_id;

   cl_context context;

   cl_command_queue command_queue;

   cl_program program;

   cl_kernel kernel;

   cl_mem out_buffer;
   float out_value = 0.0;
   size_t global_work_size = 1;

   error = clGetPlatformIDs(
                           1, /* Max number of platform IDs to return */
                           &platform_id, /* Pointer to platform_id */
                           &total_platforms); /* Total number of platforms
                                               * found on the system */

   if (error != CL_SUCCESS) {
      fprintf(stderr, "clGetPlatformIDs() failed: %s\n", clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "There are %u platforms.\n", total_platforms);



   error = clGetDeviceIDs(platform_id,
                          CL_DEVICE_TYPE_GPU,
                          1,
                          &device_id,
                          &total_gpu_devices);

   if (error != CL_SUCCESS) {
      fprintf(stderr, "clGetDeviceIDs() failed: %s\n", clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "There are %u GPU devices.\n", total_gpu_devices);

   context = clCreateContext(NULL, /* Properties */
                           1, /* Number of devices */
                           &device_id, /* Device pointer */
                           NULL, /* Callback for reporting errors */
                           NULL, /* User data to pass to error callback */
                           &error); /* Error code */

   if (error != CL_SUCCESS) {
      fprintf(stderr, "clCreateContext() failed: %s\n", clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "clCreateContext() succeeded.\n");

   command_queue = clCreateCommandQueue(context,
                                        device_id,
                                        0, /* Command queue properties */
                                        &error); /* Error code */

   if (error != CL_SUCCESS) {
      fprintf(stderr, "clCreateCommandQueue() failed: %s\n",
                      clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "clCreateCommandQueue() succeeded.\n");

   program = clCreateProgramWithSource(context,
                                       1, /* Number of strings */
                                       &program_src,
                                       NULL, /* String lengths, 0 means all the
                                              * strings are NULL terminated. */
                                       &error);

   if (error != CL_SUCCESS) {
      fprintf(stderr, "clCreateProgramWithSource() failed: %s\n",
                      clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "clCreateProgramWithSource() suceeded.\n");

   error = clBuildProgram(program,
                          1, /* Number of devices */
                          &device_id,
                          NULL, /* options */
                          NULL, /* callback function when compile is complete */
                          NULL); /* user data for callback */


   if (error != CL_SUCCESS) {
      char build_str[10000];
      error = clGetProgramBuildInfo(program,
                                    device_id,
                                    CL_PROGRAM_BUILD_LOG,
                                    10000, /* Size of output string */
                                    build_str, /* pointer to write the log to */
                                    NULL); /* Number of bytes written to the log */
      if (error != CL_SUCCESS) {
         fprintf(stderr, "clGetProgramBuildInfo() failed: %s\n",
                          clCheckErrorString(error));
      } else {
         fprintf(stderr, "Build Log: \n%s\n\n", build_str);
      }
      return EXIT_FAILURE;
   }

   fprintf(stderr, "clBuildProgram() suceeded.\n");

   kernel = clCreateKernel(program, "pi", &error);

   if (error != CL_SUCCESS) {
      fprintf(stderr, "clCreateKernel() failed: %s\n", clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "clCreateKernel() suceeded.\n");

   out_buffer = clCreateBuffer(context,
                               CL_MEM_WRITE_ONLY, /* Flags */
                               sizeof(float), /* Size of buffer */
                               NULL, /* Pointer to the data */
                               &error); /* error code */

   if (error != CL_SUCCESS) {
      fprintf(stderr, "clCreateBuffer() failed: %s\n", clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "clCreateBuffer() succeeded.\n");

   error = clSetKernelArg(kernel,
                          0, /* Arg index */
                          sizeof(cl_mem),
                          &out_buffer);

   if (error != CL_SUCCESS) {
      fprintf(stderr, "clSetKernelArg failed: %s\n", clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "clSetKernelArg() succeeded.\n");


   error = clEnqueueNDRangeKernel(command_queue,
                                  kernel,
                                  1, /* Number of dimensions */
                                  NULL, /* Global work offset */
                                  &global_work_size,
                                  &global_work_size, /* local work size */
                                  0, /* Events in wait list */
                                  NULL, /* Wait list */
                                  NULL); /* Event object for this event */

   if (error != CL_SUCCESS) {
      fprintf(stderr, "clEnqueueNDRangeKernel() failed: %s\n",
                      clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "clEnqueueNDRangeKernel() suceeded.\n");

   error = clEnqueueReadBuffer(command_queue,
                                out_buffer,
                                CL_TRUE, /* TRUE means it is a blocking read. */
                                0, /* Buffer offset to read from. */
                                sizeof(float), /* Bytes to read */
                                &out_value, /* Pointer to store the data */
                                0, /* Events in wait list */
                                NULL, /* Wait list */
                                NULL); /* Event object */


   if (error != CL_SUCCESS) {
      fprintf(stderr, "clEnqueueReadBuffer() failed: %s\n",
                      clCheckErrorString(error));
      return EXIT_FAILURE;
   }

   fprintf(stderr, "clEnqueueReadBuffer() suceeded.\n");

   fprintf(stderr, "pi = %f\n", out_value);
   return EXIT_SUCCESS;
}