1  /* { dg-do run { target vect_simd_clones } } */
       2  /* { dg-additional-options "-msse2" { target sse2_runtime } } */
       3  /* { dg-additional-options "-mavx" { target avx_runtime } } */
       4  
       5  #include <stdlib.h>
       6  
       7  #define EPS 0.00001
       8  #define N 10000
       9  #define M 1024
      10  
      11  #pragma omp declare target
      12  float Q[N][N];
      13  #pragma omp declare simd uniform(i) linear(k) notinbranch
      14  float Pfun (const int i, const int k)
      15  {
      16    return Q[i][k] * Q[k][i];
      17  }
      18  #pragma omp end declare target
      19  
      20  void init ()
      21  {
      22    int i, j;
      23    for (i = 0; i < N; i++)
      24      for (j = 0; j < N; j++)
      25        Q[i][j] = 0.001 * i * j;
      26  }
      27  
      28  float accum_ref ()
      29  {
      30    int i, k;
      31    float tmp = 0.0;
      32  
      33    for (i = 0; i < N; i++)
      34      {
      35        float tmp1 = 0.0;
      36  
      37        for (k = 0; k < M; k++)
      38  	tmp1 += Pfun(i,k);
      39  
      40        tmp += tmp1;
      41      }
      42  
      43    return tmp;
      44  }
      45  
      46  float accum ()
      47  {
      48    int i, k;
      49    float tmp = 0.0;
      50  
      51    #pragma omp target map(tofrom:tmp)
      52      #pragma omp parallel for reduction(+:tmp)
      53        for (i = 0; i < N; i++)
      54  	{
      55  	  float tmp1 = 0.0;
      56  
      57  	  #pragma omp simd reduction(+:tmp1)
      58  	    for (k = 0; k < M; k++)
      59  	      tmp1 += Pfun(i,k);
      60  
      61  	  tmp += tmp1;
      62  	}
      63  
      64    return tmp;
      65  }
      66  
      67  void check (float a, float b)
      68  {
      69    float err = (b == 0.0) ? a : (a - b) / b;
      70    if (((err > 0) ? err : -err) > EPS)
      71      abort ();
      72  }
      73  
      74  int main ()
      75  {
      76    init ();
      77  
      78    #pragma omp target update to(Q)
      79  
      80    check (accum (), accum_ref ());
      81  
      82    return 0;
      83  }