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

#include <CL/cl.h>

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

int main (int argc, char ** argv)
{

   struct cl_simple_context context;
   cl_mem src_buffer;
   cl_int error = CL_SUCCESS;
   unsigned num_items;
   int * src_data;
   int * out;
   unsigned src_size;
   unsigned i;
   int copy_host_ptr;
   cl_mem_flags mem_flags;
   size_t global_work_size = 1;

   if (argc != 3) {
      fprintf(stderr, "Usage: copy-host-ptr num_items copy_host_ptr(T/F)\n");
      return EXIT_FAILURE;
   }

   num_items = atoi(argv[1]);
   copy_host_ptr = atoi(argv[2]);
   src_size = num_items * sizeof(int);

   src_data = malloc(src_size * sizeof(int));
   out = malloc(src_size * sizeof(int));

   for (i = 0; i < num_items; i++) {
      src_data[i] = i;
   }

   if (!clSimpleSimpleInit(&context, "memcpy")) {
      return EXIT_FAILURE;
   }

   if (!clSimpleSetOutputBuffer(&context, src_size)) {
      return EXIT_FAILURE;
   }

   if (copy_host_ptr) {
      src_buffer = clCreateBuffer(context.cl_ctx,
                                  CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
                                  src_size, src_data, &error);
      if (error != CL_SUCCESS) {
         fprintf(stderr, "clCreateBuffer() failed.\n");
         return EXIT_FAILURE;
      }
   } else {
      if (!clSimpleCreateBuffer(&src_buffer, context.cl_ctx,
                                CL_MEM_READ_ONLY, src_size)) {
         return EXIT_FAILURE;
      }
      if (!clSimpleEnqueueWriteBuffer(context.command_queue, src_buffer,
                                      src_size, src_data)) {
         return EXIT_FAILURE;
      }
   }

   if (!clSimpleKernelSetArg(context.kernel, 1, sizeof(cl_mem), &src_buffer)
      || !clSimpleKernelSetArg(context.kernel, 2, sizeof(int), &num_items)) {
      return EXIT_FAILURE;
   }

   if (!clSimpleEnqueueNDRangeKernel(context.command_queue,
                                     context.kernel, 1, &global_work_size,
                                     &global_work_size)) {
      return EXIT_FAILURE;
   }

   if (!clSimpleReadOutput(&context, out, src_size)) {
      return EXIT_FAILURE;
   }

   for (i = 0; i < num_items; i++) {
      fprintf(stderr, "%4d ", out[i]);
      if (num_items % 10 == 9)
         fprintf(stderr, "\n");
   }
   fprintf(stderr, "\n");

   return EXIT_SUCCESS;

}