#include <cmath>
#include <cstdio>
#include <cstdlib>

#include "reduction.h"
#include "util.h"

int main(int argc, char **argv) {
  size_t num_elements = 100000000;

  if (argc > 1) sscanf(argv[1], "%lu", &num_elements);

  printf("Number of elements: %lu\n", num_elements);

  double *A = (double *) malloc(num_elements * sizeof(double));
  rand_vec(A, num_elements);

  double cpu_st, cpu_en, gpu_st, gpu_en;
  cpu_st = get_current_time();
  double cpu_pi_estimate = reduction_cpu(A, num_elements);
  cpu_en = get_current_time();

  reduction_gpu_initialize(num_elements);
  gpu_st = get_current_time();
  double gpu_pi_estimate = reduction_gpu(A, num_elements);
  gpu_en = get_current_time();
  reduction_gpu_finalize();

  printf("Reduction result from CPU: %.16f\n", cpu_pi_estimate);
  printf("Reduction result from GPU: %.16f\n", gpu_pi_estimate);

  printf("CPU elapsed time: %.3f sec\n", cpu_en - cpu_st);
  printf("GPU elapsed time: %.3f sec\n", gpu_en - gpu_st);

  return 0;
}