reproduce.cl

static inline int smin32(int x, int y)
{
    return x < y ? x : y;
}

static inline int sdiv_up32(int x, int y)
{
  return (x + y - 1) / y;
}

__kernel void mainzisegred_nonseg_169(int numD_115,
                                      int numBBins_116,
                                      __global double *datapoints_mem_171,
                                      __global double *binb_mem_172,
                                      __global int *mem_175)
{
  volatile __local bool local_failure;

  local_failure = false;
  barrier(CLK_LOCAL_MEM_FENCE);

  int x_acc_195;
  int chunk_sizze_196;

  chunk_sizze_196 = smin32(sdiv_up32(numD_115, get_local_size(0) *
                                     get_num_groups(0)),
                           sdiv_up32(numD_115 -
                                     get_global_id(0),
                                     get_global_size(0)));

  int x_124;
  int x_125;

  x_acc_195 = 0;

  for (int i_200 = 0; i_200 < chunk_sizze_196; i_200++) {
    int gtid_168 = get_global_id(0) + get_global_size(0) * i_200;
    int i_128 = gtid_168 / numD_115;
    int j_129 = gtid_168 % numD_115;
    bool x_130 = 0 <= i_128;
    bool y_131 = i_128 < numD_115;
    bool bounds_check_132 = x_130 && y_131;
    bool index_certs_133;

    if (!bounds_check_132) {
      local_failure = true;
      goto error_0;
    }

    double xOuter_134 = ((__global double *) datapoints_mem_171)[i_128];

    double xInner_139 = ((__global double *) datapoints_mem_171)[j_129];
    double dot_140 = xOuter_134 * xInner_139;
    bool res_141;
    int res_142;
    int res_143;
    bool loop_while_144;
    int min_145;
    int max_146;

    loop_while_144 = 1;
    min_145 = 0;
    max_146 = 20;
    while (loop_while_144) {
      int x_147 = min_145 + max_146;
      int k_148 = x_147 / 2;
      bool x_149 = 0 <= k_148;
      bool y_150 = k_148 < numBBins_116;
      bool bounds_check_151 = x_149 && y_150;
      bool index_certs_152;

      if (!bounds_check_151) {
        local_failure = true;
        goto error_0;
      }

      double y_153 = ((__global double *) binb_mem_172)[k_148];
      bool cond_154 = y_153 <= dot_140;
      int loopres_155;

      if (cond_154) {
        loopres_155 = min_145;
      } else {
        loopres_155 = k_148;
      }

      int loopres_156;

      if (cond_154) {
        loopres_156 = k_148;
      } else {
        loopres_156 = max_146;
      }

      int x_157 = 1 + loopres_155;
      bool loop_cond_158 = x_157 < loopres_156;
      bool loop_while_tmp_201 = loop_cond_158;
      int min_tmp_202 = loopres_155;
      int max_tmp_203 = loopres_156;

      loop_while_144 = loop_while_tmp_201;
      min_145 = min_tmp_202;
      max_146 = max_tmp_203;
    }
    res_141 = loop_while_144;
    res_142 = min_145;
    res_143 = max_146;
    x_124 = x_acc_195;
    x_125 = res_142;
    x_acc_195 = x_124 + x_125;
  }

  ((__global int *) mem_175)[get_local_id(0)] = 10 + x_acc_195;

 error_0:
  barrier(CLK_LOCAL_MEM_FENCE);
  if (local_failure)
    return;
  barrier(CLK_LOCAL_MEM_FENCE);
}