rows_per_invocation;
  const uint first_col = gl_GlobalInvocationID.x*columns_per_invocation;
  const uint last_row_exclusive =
      min(first_row+rows_per_invocation, $input_data_0_h$);
  const uint last_column_exclusive =
      min(first_col+columns_per_invocation, $input_data_0_w$);
  vec4 value = vec4(0);
  for (uint h = first_row; h < last_row_exclusive; ++h) {
    for (uint w = first_col; w < last_column_exclusive; ++w) {
      value += $input_data_0[w, h, gid.z]$;
    }
  }
  highp vec4 subgroup_sum = subgroupAdd(value);
  if(subgroupElect()) {
    subgroup_sums[gl_SubgroupID] = subgroup_sum;
  }

  memoryBarrierShared();
  barrier();
  // Do the final reduction in the first subgroup.
  if(gl_SubgroupID == 0) {
    highp vec4 subtotal = vec4(0);
    if (gl_SubgroupInvocationID < gl_NumSubgroups) {
      subtotal = subgroup_sums[gl_SubgroupInvocationID];
    }
    highp vec4 grand_total = subgroupAdd(subtotal);
    if(subgroupElect()) {
      highp vec4 result = grand_total / $input_data_0_w$ / $input_data_0_h$;
      $output_data_0[0, 0, gid.z] = result$;
    }
  }