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

#include <CL/cl.h>

#include "cl_simple.h"
#include "cl_util.h"

#define GLOBAL_DIM_X 10
#define GLOBAL_DIM_Y 10

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

   cl_mem out_buffer;
   int out_data[GLOBAL_DIM_X * GLOBAL_DIM_Y];
   size_t global_work_size[2] = {GLOBAL_DIM_X, GLOBAL_DIM_Y};
   size_t local_work_size[2] = {5, 5};

   struct cl_simple_context context;

   clSimpleSimpleInit(&context, "global_id2d");

   /* XXX: Delete this to see a missing error path */
   out_buffer = clCreateBuffer(context.cl_ctx,
                               CL_MEM_WRITE_ONLY,
                               sizeof(out_data),
                               NULL, &error);

   assert(error == CL_SUCCESS);

   if (!clSimpleKernelSetArg(context.kernel, 0, sizeof(cl_mem), &out_buffer)) {
      return EXIT_FAILURE;
   }

   if (!clSimpleEnqueueNDRangeKernel(context.command_queue,
                                  context.kernel,
                                  2,
                                  global_work_size,
                                  local_work_size)) {
      return EXIT_FAILURE;
   }

   error = clEnqueueReadBuffer(context.command_queue,
                               out_buffer,
                               CL_TRUE,
                               0,
                               sizeof(out_data),
                               out_data,
                               0, NULL, NULL);

   assert(error == CL_SUCCESS);

   for (i = 0; i < GLOBAL_DIM_X; i++) {
      for (j = 0; j < GLOBAL_DIM_Y; j++) {
         fprintf(stderr, "%2u ", out_data[i * GLOBAL_DIM_Y + j]);
      }
      fprintf(stderr, "\n");
   }
}