I'm absolutelly novice in OpenCL programming. For my app. (molecular simulaton) I wrote a kernel for calculate intermolecular potential of lennard-jones liquid. In this kernel I need to compute cumulative value of the potential of all particles with one:

__kernel void Molsim(__global const float* inmatrix, __global float* fi, const int c, const float r1, const float r2, const float r3, const float rc, const float epsilon, const float sigma, const float h1, const float h23)
{
   float fi0;
   float fi1;
   float d;

   unsigned int i = get_global_id(0); //number of particles (typically 2000)

   if(c!=i) {
      // potential before particle movement
      d=sqrt(pow((0.5*h1-fabs(0.5*h1-fabs(inmatrix[c*3]-inmatrix[i*3]))),2.0)+pow((0.5*h23-fabs(0.5*h23-fabs(inmatrix[c*3+1]-inmatrix[i*3+1]))),2.0)+pow((0.5*h23-fabs(0.5*h23-fabs(inmatrix[c*3+2]-inmatrix[i*3+2]))),2.0));
      if(d<rc) {
        fi0=4.0*epsilon*(pow(sigma/d,12.0)-pow(sigma/d,6.0));
      }
      else {
        fi0=0;
      }
      // potential after particle movement
      d=sqrt(pow((0.5*h1-fabs(0.5*h1-fabs(r1-inmatrix[i*3]))),2.0)+pow((0.5*h23-fabs(0.5*h23-fabs(r2-inmatrix[i*3+1]))),2.0)+pow((0.5*h23-fabs(0.5*h23-fabs(r3-inmatrix[i*3+2]))),2.0));
      if(d<rc) {
        fi1=4.0*epsilon*(pow(sigma/d,12.0)-pow(sigma/d,6.0));
      }
        else {
          fi1=0;
        }
      // cumulative difference of potentials
      fi[0]+=fi1-fi0;
      }
}         

My problem is in the line: fi[0]+=fi1-fi0;. In the one-element vector fi[0] are wrong results. I read something about sum reduction, but I do not know how to do it during the calculation.

Exist any simple solution of my problem?