#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
#include <civl-omp.cvl>

$input int THREAD_MAX;
$omp_team **parallelDataArr;
int $omp_numThreads = THREAD_MAX;

$input int CIVL_argc;
$input char  CIVL_argv[10][];
$scope $gen_root = $here;

float  a[100];
float  b[100];
float sum;
float dotprod(int _tid)
{
  int i;
  int tid;
  tid = _tid;
  $omp_team team = *parallelDataArr[_tid];
  $omp_shared *sharedArr = team->shared;
  void* b_local = sharedArr[1]->local;

  
  {
    $range r1 = 0 .. 100 - 1 # 1;
    $domain(1) loop_domain = ($domain){r1};
    $domain(1) my_iters = ($domain(1))($omp_arrive_loop(team, 0, ($domain)loop_domain, 2));
    float sum$omp_reduction = 0;
    $for(int i: my_iters)
    {
      float tmpRead0;
      $omp_read(sharedArr[1], &tmpRead0, &b_local[i]);
      float tmpRead1;
      $omp_read(sharedArr[0], &tmpRead1, &sharedArr[0]->local[i]);
      sum$omp_reduction = sum$omp_reduction + (tmpRead1 * tmpRead0);
      printf("  tid= %d i=%d\n", tid, i);
    }
    $omp_apply_assoc(sharedArr[2], 3, &sum$omp_reduction);
    $omp_barrier_and_flush(team);
  }
}
int main()
{
  int argc = CIVL_argc;
  char*  _argv[10];
  char** argv;
  for(int i = 0; i < 10; i++)
    _argv[i] = &CIVL_argv[i][0];
  argv = &_argv[0];
  int i;
  for(i = 0; i < 100; i++)
    a[i] = b[i] = 1.0 * i;
  sum = 0.0;
  {
    int _nthreads = 1 + $choose_int($omp_numThreads);
    $range thread_range = 0 .. _nthreads - 1;
    $domain(1) dom = ($domain){thread_range};
    $omp_gteam gteam = $omp_gteam_create($here, _nthreads);
    $omp_gshared sum_gshared = $omp_gshared_create(gteam, &sum);
    $omp_gshared a_gshared = $omp_gshared_create(gteam, &a);
    $omp_gshared b_gshared = $omp_gshared_create(gteam, &b);
    parallelDataArr = ($omp_team**)($malloc($gen_root, $omp_numThreads * sizeof($omp_team*)));
    $parfor(int _tid: dom)
    {
      $omp_team team = $omp_team_create($here, gteam, _tid);
      float sum_local;
      int sum_status;
      float  a_local[100];
      int  a_status[100];
      float  b_local[100];
      int  b_status[100];
      $omp_shared sum_shared = $omp_shared_create(team, sum_gshared, &sum_local, &sum_status);
      $omp_shared a_shared = $omp_shared_create(team, a_gshared, &a_local, &a_status);
      $omp_shared b_shared = $omp_shared_create(team, b_gshared, &b_local, &b_status);
      parallelDataArr[_tid] = &team;
      dotprod(_tid);
      $omp_barrier_and_flush(team);
      $omp_shared_destroy(sum_shared);
      $omp_shared_destroy(a_shared);
      $omp_shared_destroy(b_shared);
      $omp_team_destroy(team);
    }
    $omp_gshared_destroy(sum_gshared);
    $omp_gshared_destroy(a_gshared);
    $omp_gshared_destroy(b_gshared);
    $omp_gteam_destroy(gteam);
  }
  printf("Sum = %f\n", sum);
}