C++ ops_compute_transfer函数代码示例

//.........这里部分代码省略.........
  #ifdef OPS_GPU
  double *p_a0 = (double *)((char *)args[0].data_d + base0);
  #else
  double *p_a0 = (double *)((char *)args[0].data + base0);
  #endif

  #ifdef OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d] + OPS_sub_dat_list[args[1].dat->index]->d_im[d];
  #else //OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d];
  #endif //OPS_MPI
  int base1 = dat1 * 1 * 
    (start[0] * args[1].stencil->stride[0] - args[1].dat->base[0] - d_m[0]);
  base1 = base1+ dat1 *
    args[1].dat->size[0] *
    (start[1] * args[1].stencil->stride[1] - args[1].dat->base[1] - d_m[1]);
  base1 = base1+ dat1 *
    args[1].dat->size[0] *
    args[1].dat->size[1] *
    (start[2] * args[1].stencil->stride[2] - args[1].dat->base[2] - d_m[2]);
  #ifdef OPS_GPU
  double *p_a1 = (double *)((char *)args[1].data_d + base1);
  #else
  double *p_a1 = (double *)((char *)args[1].data + base1);
  #endif

  #ifdef OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[2].dat->d_m[d] + OPS_sub_dat_list[args[2].dat->index]->d_im[d];
  #else //OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[2].dat->d_m[d];
  #endif //OPS_MPI
  int base2 = dat2 * 1 * 
    (start[0] * args[2].stencil->stride[0] - args[2].dat->base[0] - d_m[0]);
  base2 = base2+ dat2 *
    args[2].dat->size[0] *
    (start[1] * args[2].stencil->stride[1] - args[2].dat->base[1] - d_m[1]);
  base2 = base2+ dat2 *
    args[2].dat->size[0] *
    args[2].dat->size[1] *
    (start[2] * args[2].stencil->stride[2] - args[2].dat->base[2] - d_m[2]);
  #ifdef OPS_GPU
  double *p_a2 = (double *)((char *)args[2].data_d + base2);
  #else
  double *p_a2 = (double *)((char *)args[2].data + base2);
  #endif

  #ifdef OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[3].dat->d_m[d] + OPS_sub_dat_list[args[3].dat->index]->d_im[d];
  #else //OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[3].dat->d_m[d];
  #endif //OPS_MPI
  int base3 = dat3 * 1 * 
    (start[0] * args[3].stencil->stride[0] - args[3].dat->base[0] - d_m[0]);
  base3 = base3+ dat3 *
    args[3].dat->size[0] *
    (start[1] * args[3].stencil->stride[1] - args[3].dat->base[1] - d_m[1]);
  base3 = base3+ dat3 *
    args[3].dat->size[0] *
    args[3].dat->size[1] *
    (start[2] * args[3].stencil->stride[2] - args[3].dat->base[2] - d_m[2]);
  #ifdef OPS_GPU
  double *p_a3 = (double *)((char *)args[3].data_d + base3);
  #else
  double *p_a3 = (double *)((char *)args[3].data + base3);
  #endif


  #ifdef OPS_GPU
  ops_H_D_exchanges_device(args, 4);
  #else
  ops_H_D_exchanges_host(args, 4);
  #endif
  ops_halo_exchanges(args,4,range);

  ops_timers_core(&c1,&t1);
  OPS_kernels[3].mpi_time += t1-t2;

  reset_field_kernel1_c_wrapper(
    p_a0,
    p_a1,
    p_a2,
    p_a3,
    x_size, y_size, z_size);

  ops_timers_core(&c2,&t2);
  OPS_kernels[3].time += t2-t1;
  #ifdef OPS_GPU
  ops_set_dirtybit_device(args, 4);
  #else
  ops_set_dirtybit_host(args, 4);
  #endif
  ops_set_halo_dirtybit3(&args[0],range);
  ops_set_halo_dirtybit3(&args[2],range);

  //Update kernel record
  OPS_kernels[3].transfer += ops_compute_transfer(dim, range, &arg0);
  OPS_kernels[3].transfer += ops_compute_transfer(dim, range, &arg1);
  OPS_kernels[3].transfer += ops_compute_transfer(dim, range, &arg2);
  OPS_kernels[3].transfer += ops_compute_transfer(dim, range, &arg3);
}

//.........这里部分代码省略.........
#else
  double *p_a2 = (double *)((char *)args[2].data + base2);
#endif

  int base3 = args[3].dat->base_offset +
              (OPS_soa ? args[3].dat->type_size : args[3].dat->elem_size) *
                  start[0] * args[3].stencil->stride[0];
  base3 = base3 +
          (OPS_soa ? args[3].dat->type_size : args[3].dat->elem_size) *
              args[3].dat->size[0] * start[1] * args[3].stencil->stride[1];
#ifdef OPS_GPU
  double *p_a3 = (double *)((char *)args[3].data_d + base3);
#else
  double *p_a3 = (double *)((char *)args[3].data + base3);
#endif

  int base4 = args[4].dat->base_offset +
              (OPS_soa ? args[4].dat->type_size : args[4].dat->elem_size) *
                  start[0] * args[4].stencil->stride[0];
  base4 = base4 +
          (OPS_soa ? args[4].dat->type_size : args[4].dat->elem_size) *
              args[4].dat->size[0] * start[1] * args[4].stencil->stride[1];
#ifdef OPS_GPU
  double *p_a4 = (double *)((char *)args[4].data_d + base4);
#else
  double *p_a4 = (double *)((char *)args[4].data + base4);
#endif

  int x_size = MAX(0, end[0] - start[0]);
  int y_size = MAX(0, end[1] - start[1]);

  // initialize global variable with the dimension of dats
  xdim0 = args[0].dat->size[0];
  xdim1 = args[1].dat->size[0];
  xdim2 = args[2].dat->size[0];
  xdim3 = args[3].dat->size[0];
  xdim4 = args[4].dat->size[0];
  if (xdim0 != xdim0_initialise_chunk_kernel_volume_h ||
      xdim1 != xdim1_initialise_chunk_kernel_volume_h ||
      xdim2 != xdim2_initialise_chunk_kernel_volume_h ||
      xdim3 != xdim3_initialise_chunk_kernel_volume_h ||
      xdim4 != xdim4_initialise_chunk_kernel_volume_h) {
    xdim0_initialise_chunk_kernel_volume = xdim0;
    xdim0_initialise_chunk_kernel_volume_h = xdim0;
    xdim1_initialise_chunk_kernel_volume = xdim1;
    xdim1_initialise_chunk_kernel_volume_h = xdim1;
    xdim2_initialise_chunk_kernel_volume = xdim2;
    xdim2_initialise_chunk_kernel_volume_h = xdim2;
    xdim3_initialise_chunk_kernel_volume = xdim3;
    xdim3_initialise_chunk_kernel_volume_h = xdim3;
    xdim4_initialise_chunk_kernel_volume = xdim4;
    xdim4_initialise_chunk_kernel_volume_h = xdim4;
  }

// Halo Exchanges

#ifdef OPS_GPU
  ops_H_D_exchanges_device(args, 5);
#else
  ops_H_D_exchanges_host(args, 5);
#endif
  ops_halo_exchanges(args, 5, range);

#ifdef OPS_GPU
  ops_H_D_exchanges_device(args, 5);
#else
  ops_H_D_exchanges_host(args, 5);
#endif
  if (OPS_diags > 1) {
    ops_timers_core(&c2, &t2);
    OPS_kernels[14].mpi_time += t2 - t1;
  }

  initialise_chunk_kernel_volume_c_wrapper(p_a0, p_a1, p_a2, p_a3, p_a4, x_size,
                                           y_size);

  if (OPS_diags > 1) {
    ops_timers_core(&c1, &t1);
    OPS_kernels[14].time += t1 - t2;
  }
#ifdef OPS_GPU
  ops_set_dirtybit_device(args, 5);
#else
  ops_set_dirtybit_host(args, 5);
#endif
  ops_set_halo_dirtybit3(&args[0], range);
  ops_set_halo_dirtybit3(&args[2], range);
  ops_set_halo_dirtybit3(&args[4], range);

  if (OPS_diags > 1) {
    // Update kernel record
    ops_timers_core(&c2, &t2);
    OPS_kernels[14].mpi_time += t2 - t1;
    OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg0);
    OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg1);
    OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg2);
    OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg3);
    OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg4);
  }
}

//.........这里部分代码省略.........
    int start2 = start_i;

    // set up initial pointers
    int d_m[OPS_MAX_DIM];
#ifdef OPS_MPI
    for (int d = 0; d < dim; d++)
      d_m[d] =
          args[0].dat->d_m[d] + OPS_sub_dat_list[args[0].dat->index]->d_im[d];
#else
    for (int d = 0; d < dim; d++)
      d_m[d] = args[0].dat->d_m[d];
#endif
    int base0 = dat0 * 1 * (start0 * args[0].stencil->stride[0] -
                            args[0].dat->base[0] - d_m[0]);
    base0 = base0 +
            dat0 * args[0].dat->size[0] * (start1 * args[0].stencil->stride[1] -
                                           args[0].dat->base[1] - d_m[1]);
    base0 = base0 +
            dat0 * args[0].dat->size[0] * args[0].dat->size[1] *
                (start2 * args[0].stencil->stride[2] - args[0].dat->base[2] -
                 d_m[2]);
    p_a[0] = (char *)args[0].data + base0;

#ifdef OPS_MPI
    for (int d = 0; d < dim; d++)
      d_m[d] =
          args[1].dat->d_m[d] + OPS_sub_dat_list[args[1].dat->index]->d_im[d];
#else
    for (int d = 0; d < dim; d++)
      d_m[d] = args[1].dat->d_m[d];
#endif
    int base1 = dat1 * 1 * (start0 * args[1].stencil->stride[0] -
                            args[1].dat->base[0] - d_m[0]);
    base1 = base1 +
            dat1 * args[1].dat->size[0] * (start1 * args[1].stencil->stride[1] -
                                           args[1].dat->base[1] - d_m[1]);
    base1 = base1 +
            dat1 * args[1].dat->size[0] * args[1].dat->size[1] *
                (start2 * args[1].stencil->stride[2] - args[1].dat->base[2] -
                 d_m[2]);
    p_a[1] = (char *)args[1].data + base1;

    p_a[2] = (char *)args[2].data;

    for (int n_z = start_i; n_z < finish_i; n_z++) {
      for (int n_y = start[1]; n_y < end[1]; n_y++) {
        for (int n_x = start[0];
             n_x < start[0] + (end[0] - start[0]) / SIMD_VEC; n_x++) {
// call kernel function, passing in pointers to data -vectorised
#pragma simd
          for (int i = 0; i < SIMD_VEC; i++) {
            update_halo_kernel4_plus_2_b((double *)p_a[0] + i * 1 * 1,
                                         (double *)p_a[1] + i * 1 * 1,
                                         (int *)p_a[2]);
          }

          // shift pointers to data x direction
          p_a[0] = p_a[0] + (dat0 * off0_0) * SIMD_VEC;
          p_a[1] = p_a[1] + (dat1 * off1_0) * SIMD_VEC;
        }

        for (int n_x = start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC;
             n_x < end[0]; n_x++) {
          // call kernel function, passing in pointers to data - remainder
          update_halo_kernel4_plus_2_b((double *)p_a[0], (double *)p_a[1],
                                       (int *)p_a[2]);

          // shift pointers to data x direction
          p_a[0] = p_a[0] + (dat0 * off0_0);
          p_a[1] = p_a[1] + (dat1 * off1_0);
        }

        // shift pointers to data y direction
        p_a[0] = p_a[0] + (dat0 * off0_1);
        p_a[1] = p_a[1] + (dat1 * off1_1);
      }
      // shift pointers to data z direction
      p_a[0] = p_a[0] + (dat0 * off0_2);
      p_a[1] = p_a[1] + (dat1 * off1_2);
    }
  }

  if (OPS_diags > 1) {
    ops_timers_core(&c1, &t1);
    OPS_kernels[79].time += t1 - t2;
  }

  ops_set_dirtybit_host(args, 3);

  ops_set_halo_dirtybit3(&args[0], range);
  ops_set_halo_dirtybit3(&args[1], range);

  if (OPS_diags > 1) {
    // Update kernel record
    ops_timers_core(&c2, &t2);
    OPS_kernels[79].mpi_time += t2 - t1;
    OPS_kernels[79].transfer += ops_compute_transfer(dim, start, end, &arg0);
    OPS_kernels[79].transfer += ops_compute_transfer(dim, start, end, &arg1);
  }
}

//.........这里部分代码省略.........
  int nthreads = omp_get_max_threads( );
  #else
  int nthreads = 1;
  #endif
  xdim0 = args[0].dat->size[0]*args[0].dat->dim;

  ops_H_D_exchanges_host(args, 2);

  //Halo Exchanges
  ops_halo_exchanges(args,2,range);


  ops_timers_core(&c2,&t2);
  OPS_kernels[0].mpi_time += t2-t1;


  #pragma omp parallel for
  for ( int thr=0; thr<nthreads; thr++ ){

    int y_size = end[1]-start[1];
    char *p_a[2];

    int start_i = start[1] + ((y_size-1)/nthreads+1)*thr;
    int finish_i = start[1] + MIN(((y_size-1)/nthreads+1)*(thr+1),y_size);

    //get address per thread
    int start0 = start[0];
    int start1 = start_i;

    int arg_idx[2];
    #ifdef OPS_MPI
    arg_idx[0] = sb->decomp_disp[0]+start0;
    arg_idx[1] = sb->decomp_disp[1]+start1;
    #else //OPS_MPI
    arg_idx[0] = start0;
    arg_idx[1] = start1;
    #endif //OPS_MPI
    //set up initial pointers 
    int d_m[OPS_MAX_DIM];
    #ifdef OPS_MPI
    for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d] + OPS_sub_dat_list[args[0].dat->index]->d_im[d];
    #else //OPS_MPI
    for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d];
    #endif //OPS_MPI
    int base0 = dat0 * 1 * 
    (start0 * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]);
    base0 = base0+ dat0 *
      args[0].dat->size[0] *
      (start1 * args[0].stencil->stride[1] - args[0].dat->base[1] - d_m[1]);
    p_a[0] = (char *)args[0].data + base0;

    p_a[1] = (char *)arg_idx;


    for ( int n_y=start_i; n_y<finish_i; n_y++ ){
      for ( int n_x=start[0]; n_x<start[0]+(end[0]-start[0])/SIMD_VEC; n_x++ ){
        //call kernel function, passing in pointers to data -vectorised
        for ( int i=0; i<SIMD_VEC; i++ ){
          mblock_populate_kernel(  (double * )p_a[0]+ i*1, arg_idx );

          arg_idx[0]++;
        }

        //shift pointers to data x direction
        p_a[0]= p_a[0] + (dat0 * off0_0)*SIMD_VEC;
      }

      for ( int n_x=start[0]+((end[0]-start[0])/SIMD_VEC)*SIMD_VEC; n_x<end[0]; n_x++ ){
        //call kernel function, passing in pointers to data - remainder
        mblock_populate_kernel(  (double * )p_a[0], arg_idx );


        //shift pointers to data x direction
        p_a[0]= p_a[0] + (dat0 * off0_0);
        arg_idx[0]++;
      }

      //shift pointers to data y direction
      p_a[0]= p_a[0] + (dat0 * off0_1);
      #ifdef OPS_MPI
      arg_idx[0] = sb->decomp_disp[0]+start0;
      #else //OPS_MPI
      arg_idx[0] = start0;
      #endif //OPS_MPI
      arg_idx[1]++;
    }
  }

  ops_timers_core(&c1,&t1);
  OPS_kernels[0].time += t1-t2;

  ops_set_dirtybit_host(args, 2);

  ops_set_halo_dirtybit3(&args[0],range);

  //Update kernel record
  ops_timers_core(&c2,&t2);
  OPS_kernels[0].mpi_time += t2-t1;
  OPS_kernels[0].transfer += ops_compute_transfer(dim, range, &arg0);
}

//.........这里部分代码省略.........
  }
  #else
  for ( int n=0; n<2; n++ ){
    start[n] = range[2*n];end[n] = range[2*n+1];
  }
  #endif
  #ifdef OPS_DEBUG
  ops_register_args(args, "copy");
  #endif

  offs[0][0] = args[0].stencil->stride[0]*1;  //unit step in x dimension
  offs[0][1] = off2D(1, &start[0],
      &end[0],args[0].dat->size, args[0].stencil->stride) - offs[0][0];

  offs[1][0] = args[1].stencil->stride[0]*1;  //unit step in x dimension
  offs[1][1] = off2D(1, &start[0],
      &end[0],args[1].dat->size, args[1].stencil->stride) - offs[1][0];



  int off0_0 = offs[0][0];
  int off0_1 = offs[0][1];
  int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size);
  int off1_0 = offs[1][0];
  int off1_1 = offs[1][1];
  int dat1 = (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size);

  //set up initial pointers and exchange halos if necessary
  int base0 = args[0].dat->base_offset + (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size) * start[0] * args[0].stencil->stride[0];
  base0 = base0+ (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size) *
    args[0].dat->size[0] *
    start[1] * args[0].stencil->stride[1];
  p_a[0] = (char *)args[0].data + base0;

  int base1 = args[1].dat->base_offset + (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size) * start[0] * args[1].stencil->stride[0];
  base1 = base1+ (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size) *
    args[1].dat->size[0] *
    start[1] * args[1].stencil->stride[1];
  p_a[1] = (char *)args[1].data + base1;


  //initialize global variable with the dimension of dats
  xdim0 = args[0].dat->size[0];
  xdim1 = args[1].dat->size[0];

  //Halo Exchanges
  ops_H_D_exchanges_host(args, 2);
  ops_halo_exchanges(args,2,range);
  ops_H_D_exchanges_host(args, 2);

  if (OPS_diags > 1) {
    ops_timers_core(&c1,&t1);
    OPS_kernels[5].mpi_time += t1-t2;
  }

  int n_x;
  for ( int n_y=start[1]; n_y<end[1]; n_y++ ){
    #pragma novector
    for( n_x=start[0]; n_x<start[0]+((end[0]-start[0])/SIMD_VEC)*SIMD_VEC; n_x+=SIMD_VEC ) {
      //call kernel function, passing in pointers to data -vectorised
      #pragma simd
      for ( int i=0; i<SIMD_VEC; i++ ){
        copy(  (double *)p_a[0]+ i*1*1, (double *)p_a[1]+ i*1*1 );

      }

      //shift pointers to data x direction
      p_a[0]= p_a[0] + (dat0 * off0_0)*SIMD_VEC;
      p_a[1]= p_a[1] + (dat1 * off1_0)*SIMD_VEC;
    }

    for ( int n_x=start[0]+((end[0]-start[0])/SIMD_VEC)*SIMD_VEC; n_x<end[0]; n_x++ ){
      //call kernel function, passing in pointers to data - remainder
      copy(  (double *)p_a[0], (double *)p_a[1] );


      //shift pointers to data x direction
      p_a[0]= p_a[0] + (dat0 * off0_0);
      p_a[1]= p_a[1] + (dat1 * off1_0);
    }

    //shift pointers to data y direction
    p_a[0]= p_a[0] + (dat0 * off0_1);
    p_a[1]= p_a[1] + (dat1 * off1_1);
  }
  if (OPS_diags > 1) {
    ops_timers_core(&c2,&t2);
    OPS_kernels[5].time += t2-t1;
  }
  ops_set_dirtybit_host(args, 2);
  ops_set_halo_dirtybit3(&args[0],range);

  if (OPS_diags > 1) {
    //Update kernel record
    ops_timers_core(&c1,&t1);
    OPS_kernels[5].mpi_time += t1-t2;
    OPS_kernels[5].transfer += ops_compute_transfer(dim, start, end, &arg0);
    OPS_kernels[5].transfer += ops_compute_transfer(dim, start, end, &arg1);
  }
}

// host stub function
void ops_par_loop_flux_calc_kernelx_execute(ops_kernel_descriptor *desc) {
  ops_block block = desc->block;
  int dim = desc->dim;
  int *range = desc->range;
  ops_arg arg0 = desc->args[0];
  ops_arg arg1 = desc->args[1];
  ops_arg arg2 = desc->args[2];
  ops_arg arg3 = desc->args[3];

  // Timing
  double t1, t2, c1, c2;

  ops_arg args[4] = {arg0, arg1, arg2, arg3};

#ifdef CHECKPOINTING
  if (!ops_checkpointing_before(args, 4, range, 59))
    return;
#endif

  if (OPS_diags > 1) {
    OPS_kernels[59].count++;
    ops_timers_core(&c2, &t2);
  }

  // compute locally allocated range for the sub-block
  int start[2];
  int end[2];

  for (int n = 0; n < 2; n++) {
    start[n] = range[2 * n];
    end[n] = range[2 * n + 1];
  }

#ifdef OPS_DEBUG
  ops_register_args(args, "flux_calc_kernelx");
#endif

  // set up initial pointers and exchange halos if necessary
  int base0 = args[0].dat->base_offset;
  double *__restrict__ vol_flux_x = (double *)(args[0].data + base0);

  int base1 = args[1].dat->base_offset;
  const double *__restrict__ xarea = (double *)(args[1].data + base1);

  int base2 = args[2].dat->base_offset;
  const double *__restrict__ xvel0 = (double *)(args[2].data + base2);

  int base3 = args[3].dat->base_offset;
  const double *__restrict__ xvel1 = (double *)(args[3].data + base3);

  // initialize global variable with the dimension of dats
  int xdim0_flux_calc_kernelx = args[0].dat->size[0];
  int xdim1_flux_calc_kernelx = args[1].dat->size[0];
  int xdim2_flux_calc_kernelx = args[2].dat->size[0];
  int xdim3_flux_calc_kernelx = args[3].dat->size[0];

  if (OPS_diags > 1) {
    ops_timers_core(&c1, &t1);
    OPS_kernels[59].mpi_time += t1 - t2;
  }

#pragma omp parallel for
  for (int n_y = start[1]; n_y < end[1]; n_y++) {
#ifdef intel
#pragma loop_count(10000)
#pragma omp simd aligned(vol_flux_x, xarea, xvel0, xvel1)
#else
#pragma simd
#endif
    for (int n_x = start[0]; n_x < end[0]; n_x++) {

      vol_flux_x[OPS_ACC0(0, 0)] =
          0.25 * dt * (xarea[OPS_ACC1(0, 0)]) *
          ((xvel0[OPS_ACC2(0, 0)]) + (xvel0[OPS_ACC2(0, 1)]) +
           (xvel1[OPS_ACC3(0, 0)]) + (xvel1[OPS_ACC3(0, 1)]));
    }
  }
  if (OPS_diags > 1) {
    ops_timers_core(&c2, &t2);
    OPS_kernels[59].time += t2 - t1;
  }

  if (OPS_diags > 1) {
    // Update kernel record
    ops_timers_core(&c1, &t1);
    OPS_kernels[59].mpi_time += t1 - t2;
    OPS_kernels[59].transfer += ops_compute_transfer(dim, start, end, &arg0);
    OPS_kernels[59].transfer += ops_compute_transfer(dim, start, end, &arg1);
    OPS_kernels[59].transfer += ops_compute_transfer(dim, start, end, &arg2);
    OPS_kernels[59].transfer += ops_compute_transfer(dim, start, end, &arg3);
  }
}

//.........这里部分代码省略.........
  double t1, t2, c1, c2;
  ops_arg args[2] = {arg0, arg1};

#ifdef CHECKPOINTING
  if (!ops_checkpointing_before(args, 2, range, 0))
    return;
#endif

  if (OPS_diags > 1) {
    ops_timing_realloc(0, "gridgen_kernel");
    OPS_kernels[0].count++;
    ops_timers_core(&c1, &t1);
  }

  // compute localy allocated range for the sub-block

  int start[1];
  int end[1];
#ifdef OPS_MPI
  sub_block_list sb = OPS_sub_block_list[block->index];
#endif // OPS_MPI

  int arg_idx[1];
  int arg_idx_base[1];
#ifdef OPS_MPI
  if (compute_ranges(args, 2, block, range, start, end, arg_idx) < 0)
    return;
#else // OPS_MPI
  for (int n = 0; n < 1; n++) {
    start[n] = range[2 * n];
    end[n] = range[2 * n + 1];
    arg_idx[n] = start[n];
  }
#endif
  for (int n = 0; n < 1; n++) {
    arg_idx_base[n] = arg_idx[n];
  }

  int dat0 = args[0].dat->elem_size;

  // set up initial pointers
  int base0 = args[0].dat->base_offset +
              (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size) *
                  start[0] * args[0].stencil->stride[0];
#ifdef OPS_GPU
  double *p_a0 = (double *)((char *)args[0].data_d + base0);
#else
  double *p_a0 = (double *)((char *)args[0].data + base0);
#endif

  int *p_a1 = NULL;

  int x_size = MAX(0, end[0] - start[0]);

  // initialize global variable with the dimension of dats
  xdim0 = args[0].dat->size[0];
  if (xdim0 != xdim0_gridgen_kernel_h) {
    xdim0_gridgen_kernel = xdim0;
    xdim0_gridgen_kernel_h = xdim0;
  }

// Halo Exchanges

#ifdef OPS_GPU
  ops_H_D_exchanges_device(args, 2);
#else
  ops_H_D_exchanges_host(args, 2);
#endif
  ops_halo_exchanges(args, 2, range);

#ifdef OPS_GPU
  ops_H_D_exchanges_device(args, 2);
#else
  ops_H_D_exchanges_host(args, 2);
#endif
  if (OPS_diags > 1) {
    ops_timers_core(&c2, &t2);
    OPS_kernels[0].mpi_time += t2 - t1;
  }

  gridgen_kernel_c_wrapper(p_a0, p_a1, arg_idx[0], x_size);

  if (OPS_diags > 1) {
    ops_timers_core(&c1, &t1);
    OPS_kernels[0].time += t1 - t2;
  }
#ifdef OPS_GPU
  ops_set_dirtybit_device(args, 2);
#else
  ops_set_dirtybit_host(args, 2);
#endif
  ops_set_halo_dirtybit3(&args[0], range);

  if (OPS_diags > 1) {
    // Update kernel record
    ops_timers_core(&c2, &t2);
    OPS_kernels[0].mpi_time += t2 - t1;
    OPS_kernels[0].transfer += ops_compute_transfer(dim, start, end, &arg0);
  }
}

// host stub function
void ops_par_loop_reset_field_kernel1_execute(ops_kernel_descriptor *desc) {
  ops_block block = desc->block;
  int dim = desc->dim;
  int *range = desc->range;
  ops_arg arg0 = desc->args[0];
  ops_arg arg1 = desc->args[1];
  ops_arg arg2 = desc->args[2];
  ops_arg arg3 = desc->args[3];

  // Timing
  double t1, t2, c1, c2;

  ops_arg args[4] = {arg0, arg1, arg2, arg3};

#ifdef CHECKPOINTING
  if (!ops_checkpointing_before(args, 4, range, 139))
    return;
#endif

  if (OPS_diags > 1) {
    OPS_kernels[139].count++;
    ops_timers_core(&c2, &t2);
  }

  // compute locally allocated range for the sub-block
  int start[3];
  int end[3];

  for (int n = 0; n < 3; n++) {
    start[n] = range[2 * n];
    end[n] = range[2 * n + 1];
  }

#ifdef OPS_DEBUG
  ops_register_args(args, "reset_field_kernel1");
#endif

  // set up initial pointers and exchange halos if necessary
  int base0 = args[0].dat->base_offset;
  double *__restrict__ density0 = (double *)(args[0].data + base0);

  int base1 = args[1].dat->base_offset;
  const double *__restrict__ density1 = (double *)(args[1].data + base1);

  int base2 = args[2].dat->base_offset;
  double *__restrict__ energy0 = (double *)(args[2].data + base2);

  int base3 = args[3].dat->base_offset;
  const double *__restrict__ energy1 = (double *)(args[3].data + base3);

  // initialize global variable with the dimension of dats
  int xdim0_reset_field_kernel1 = args[0].dat->size[0];
  int ydim0_reset_field_kernel1 = args[0].dat->size[1];
  int xdim1_reset_field_kernel1 = args[1].dat->size[0];
  int ydim1_reset_field_kernel1 = args[1].dat->size[1];
  int xdim2_reset_field_kernel1 = args[2].dat->size[0];
  int ydim2_reset_field_kernel1 = args[2].dat->size[1];
  int xdim3_reset_field_kernel1 = args[3].dat->size[0];
  int ydim3_reset_field_kernel1 = args[3].dat->size[1];

  if (OPS_diags > 1) {
    ops_timers_core(&c1, &t1);
    OPS_kernels[139].mpi_time += t1 - t2;
  }

#pragma omp parallel for collapse(2)
  for (int n_z = start[2]; n_z < end[2]; n_z++) {
    for (int n_y = start[1]; n_y < end[1]; n_y++) {
#ifdef intel
#pragma loop_count(10000)
#pragma omp simd aligned(density0, density1, energy0, energy1)
#else
#pragma simd
#endif
      for (int n_x = start[0]; n_x < end[0]; n_x++) {

        density0[OPS_ACC0(0, 0, 0)] = density1[OPS_ACC1(0, 0, 0)];
        energy0[OPS_ACC2(0, 0, 0)] = energy1[OPS_ACC3(0, 0, 0)];
      }
    }
  }
  if (OPS_diags > 1) {
    ops_timers_core(&c2, &t2);
    OPS_kernels[139].time += t2 - t1;
  }

  if (OPS_diags > 1) {
    // Update kernel record
    ops_timers_core(&c1, &t1);
    OPS_kernels[139].mpi_time += t1 - t2;
    OPS_kernels[139].transfer += ops_compute_transfer(dim, start, end, &arg0);
    OPS_kernels[139].transfer += ops_compute_transfer(dim, start, end, &arg1);
    OPS_kernels[139].transfer += ops_compute_transfer(dim, start, end, &arg2);
    OPS_kernels[139].transfer += ops_compute_transfer(dim, start, end, &arg3);
  }
}

//.........这里部分代码省略.........
                              (void *)&arg9.data_d));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 10, sizeof(cl_mem),
                              (void *)&arg10.data_d));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 11, sizeof(cl_mem),
                              (void *)&arg11.data_d));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 12, sizeof(cl_mem),
                              (void *)&arg12.data_d));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 13, sizeof(cl_mem),
                              (void *)&arg13.data_d));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 14, sizeof(cl_mem),
                              (void *)&arg14.data_d));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 15, sizeof(cl_mem),
                              (void *)&arg15.data_d));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 16, sizeof(cl_mem),
                              (void *)&arg16.data_d));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 17,
                              sizeof(cl_double), (void *)&dt));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 18, sizeof(cl_int),
                              (void *)&base0));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 19, sizeof(cl_int),
                              (void *)&base1));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 20, sizeof(cl_int),
                              (void *)&base2));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 21, sizeof(cl_int),
                              (void *)&base3));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 22, sizeof(cl_int),
                              (void *)&base4));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 23, sizeof(cl_int),
                              (void *)&base5));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 24, sizeof(cl_int),
                              (void *)&base6));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 25, sizeof(cl_int),
                              (void *)&base7));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 26, sizeof(cl_int),
                              (void *)&base8));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 27, sizeof(cl_int),
                              (void *)&base9));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 28, sizeof(cl_int),
                              (void *)&base10));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 29, sizeof(cl_int),
                              (void *)&base11));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 30, sizeof(cl_int),
                              (void *)&base12));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 31, sizeof(cl_int),
                              (void *)&base13));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 32, sizeof(cl_int),
                              (void *)&base14));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 33, sizeof(cl_int),
                              (void *)&base15));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 34, sizeof(cl_int),
                              (void *)&base16));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 35, sizeof(cl_int),
                              (void *)&x_size));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 36, sizeof(cl_int),
                              (void *)&y_size));
    clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 37, sizeof(cl_int),
                              (void *)&z_size));

    // call/enque opencl kernel wrapper function
    clSafeCall(clEnqueueNDRangeKernel(
        OPS_opencl_core.command_queue, OPS_opencl_core.kernel[103], 3, NULL,
        globalWorkSize, localWorkSize, 0, NULL, NULL));
  }
  if (OPS_diags > 1) {
    clSafeCall(clFinish(OPS_opencl_core.command_queue));
  }

  if (OPS_diags > 1) {
    ops_timers_core(&c1, &t1);
    OPS_kernels[103].time += t1 - t2;
  }

  ops_set_dirtybit_device(args, 17);
  ops_set_halo_dirtybit3(&args[6], range);
  ops_set_halo_dirtybit3(&args[10], range);
  ops_set_halo_dirtybit3(&args[13], range);

  if (OPS_diags > 1) {
    // Update kernel record
    ops_timers_core(&c2, &t2);
    OPS_kernels[103].mpi_time += t2 - t1;
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg0);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg1);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg2);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg3);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg4);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg5);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg6);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg7);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg8);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg9);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg10);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg11);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg12);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg13);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg14);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg15);
    OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg16);
  }
}

//.........这里部分代码省略.........
  int base1 = 1 * 
  (start[0] * args[1].stencil->stride[0] - args[1].dat->base[0] - d_m[0]);
  base1 = base1 + args[1].dat->size[0] *
  (start[1] * args[1].stencil->stride[1] - args[1].dat->base[1] - d_m[1]);

  #ifdef OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[2].dat->d_m[d] + OPS_sub_dat_list[args[2].dat->index]->d_im[d];
  #else //OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[2].dat->d_m[d];
  #endif //OPS_MPI
  int base2 = 1 * 
  (start[0] * args[2].stencil->stride[0] - args[2].dat->base[0] - d_m[0]);
  base2 = base2 + args[2].dat->size[0] *
  (start[1] * args[2].stencil->stride[1] - args[2].dat->base[1] - d_m[1]);

  #ifdef OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[3].dat->d_m[d] + OPS_sub_dat_list[args[3].dat->index]->d_im[d];
  #else //OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[3].dat->d_m[d];
  #endif //OPS_MPI
  int base3 = 1 * 
  (start[0] * args[3].stencil->stride[0] - args[3].dat->base[0] - d_m[0]);
  base3 = base3 + args[3].dat->size[0] *
  (start[1] * args[3].stencil->stride[1] - args[3].dat->base[1] - d_m[1]);

  #ifdef OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[4].dat->d_m[d] + OPS_sub_dat_list[args[4].dat->index]->d_im[d];
  #else //OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[4].dat->d_m[d];
  #endif //OPS_MPI
  int base4 = 1 * 
  (start[0] * args[4].stencil->stride[0] - args[4].dat->base[0] - d_m[0]);
  base4 = base4 + args[4].dat->size[0] *
  (start[1] * args[4].stencil->stride[1] - args[4].dat->base[1] - d_m[1]);

  #ifdef OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[5].dat->d_m[d] + OPS_sub_dat_list[args[5].dat->index]->d_im[d];
  #else //OPS_MPI
  for (int d = 0; d < dim; d++) d_m[d] = args[5].dat->d_m[d];
  #endif //OPS_MPI
  int base5 = 1 * 
  (start[0] * args[5].stencil->stride[0] - args[5].dat->base[0] - d_m[0]);
  base5 = base5 + args[5].dat->size[0] *
  (start[1] * args[5].stencil->stride[1] - args[5].dat->base[1] - d_m[1]);


  ops_H_D_exchanges_device(args, 7);
  ops_halo_exchanges(args,7,range);
  ops_H_D_exchanges_device(args, 7);

  ops_timers_core(&c1,&t1);
  OPS_kernels[30].mpi_time += t1-t2;

  int nthread = OPS_block_size_x*OPS_block_size_y;


  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 0, sizeof(cl_mem), (void*) &arg0.data_d ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 1, sizeof(cl_mem), (void*) &arg1.data_d ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 2, sizeof(cl_mem), (void*) &arg2.data_d ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 3, sizeof(cl_mem), (void*) &arg3.data_d ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 4, sizeof(cl_mem), (void*) &arg4.data_d ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 5, sizeof(cl_mem), (void*) &arg5.data_d ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 6, sizeof(cl_mem), (void*) &arg6.data_d ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 7, nthread*sizeof(double), NULL));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 8, sizeof(cl_int), (void*) &r_bytes6 ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 9, sizeof(cl_int), (void*) &base0 ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 10, sizeof(cl_int), (void*) &base1 ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 11, sizeof(cl_int), (void*) &base2 ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 12, sizeof(cl_int), (void*) &base3 ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 13, sizeof(cl_int), (void*) &base4 ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 14, sizeof(cl_int), (void*) &base5 ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 15, sizeof(cl_int), (void*) &x_size ));
  clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 16, sizeof(cl_int), (void*) &y_size ));

  //call/enque opencl kernel wrapper function
  clSafeCall( clEnqueueNDRangeKernel(OPS_opencl_core.command_queue, OPS_opencl_core.kernel[30], 3, NULL, globalWorkSize, localWorkSize, 0, NULL, NULL) );
  if (OPS_diags>1) {
    clSafeCall( clFinish(OPS_opencl_core.command_queue) );
  }

  mvReductArraysToHost(reduct_bytes);
  for ( int b=0; b<maxblocks; b++ ){
    for ( int d=0; d<12; d++ ){
      arg6h[d] = arg6h[d] + ((double *)arg6.data)[d+b*12];
    }
  }
  arg6.data = (char *)arg6h;

  ops_set_dirtybit_device(args, 7);

  //Update kernel record
  ops_timers_core(&c2,&t2);
  OPS_kernels[30].time += t2-t1;
  OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg0);
  OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg1);
  OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg2);
  OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg3);
  OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg4);
  OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg5);
}

//.........这里部分代码省略.........
    d_m[d] = args[2].dat->d_m[d];
#endif
  int base2 = dat2 * 1 * (start[0] * args[2].stencil->stride[0] -
                          args[2].dat->base[0] - d_m[0]);
  base2 = base2 +
          dat2 * args[2].dat->size[0] * (start[1] * args[2].stencil->stride[1] -
                                         args[2].dat->base[1] - d_m[1]);
  double *p_a2 = (double *)((char *)args[2].data + base2);

#ifdef OPS_MPI
  for (int d = 0; d < dim; d++)
    d_m[d] =
        args[3].dat->d_m[d] + OPS_sub_dat_list[args[3].dat->index]->d_im[d];
#else
  for (int d = 0; d < dim; d++)
    d_m[d] = args[3].dat->d_m[d];
#endif
  int base3 = dat3 * 1 * (start[0] * args[3].stencil->stride[0] -
                          args[3].dat->base[0] - d_m[0]);
  base3 = base3 +
          dat3 * args[3].dat->size[0] * (start[1] * args[3].stencil->stride[1] -
                                         args[3].dat->base[1] - d_m[1]);
  double *p_a3 = (double *)((char *)args[3].data + base3);

#ifdef OPS_MPI
  for (int d = 0; d < dim; d++)
    d_m[d] =
        args[4].dat->d_m[d] + OPS_sub_dat_list[args[4].dat->index]->d_im[d];
#else
  for (int d = 0; d < dim; d++)
    d_m[d] = args[4].dat->d_m[d];
#endif
  int base4 = dat4 * 1 * (start[0] * args[4].stencil->stride[0] -
                          args[4].dat->base[0] - d_m[0]);
  base4 = base4 +
          dat4 * args[4].dat->size[0] * (start[1] * args[4].stencil->stride[1] -
                                         args[4].dat->base[1] - d_m[1]);
  double *p_a4 = (double *)((char *)args[4].data + base4);

#ifdef OPS_MPI
  for (int d = 0; d < dim; d++)
    d_m[d] =
        args[5].dat->d_m[d] + OPS_sub_dat_list[args[5].dat->index]->d_im[d];
#else
  for (int d = 0; d < dim; d++)
    d_m[d] = args[5].dat->d_m[d];
#endif
  int base5 = dat5 * 1 * (start[0] * args[5].stencil->stride[0] -
                          args[5].dat->base[0] - d_m[0]);
  base5 = base5 +
          dat5 * args[5].dat->size[0] * (start[1] * args[5].stencil->stride[1] -
                                         args[5].dat->base[1] - d_m[1]);
  double *p_a5 = (double *)((char *)args[5].data + base5);

#ifdef OPS_MPI
  for (int d = 0; d < dim; d++)
    d_m[d] =
        args[6].dat->d_m[d] + OPS_sub_dat_list[args[6].dat->index]->d_im[d];
#else
  for (int d = 0; d < dim; d++)
    d_m[d] = args[6].dat->d_m[d];
#endif
  int base6 = dat6 * 1 * (start[0] * args[6].stencil->stride[0] -
                          args[6].dat->base[0] - d_m[0]);
  base6 = base6 +
          dat6 * args[6].dat->size[0] * (start[1] * args[6].stencil->stride[1] -
                                         args[6].dat->base[1] - d_m[1]);
  double *p_a6 = (double *)((char *)args[6].data + base6);

  int *p_a7 = arg7h;

  ops_H_D_exchanges_host(args, 8);
  ops_halo_exchanges(args, 8, range);

  ops_timers_core(&c1, &t1);
  OPS_kernels[16].mpi_time += t1 - t2;

  update_halo_kernel1_r1_c_wrapper(p_a0, p_a1, p_a2, p_a3, p_a4, p_a5, p_a6,
                                   p_a7, x_size, y_size);

  ops_timers_core(&c2, &t2);
  OPS_kernels[16].time += t2 - t1;
  ops_set_dirtybit_host(args, 8);
  ops_set_halo_dirtybit3(&args[0], range);
  ops_set_halo_dirtybit3(&args[1], range);
  ops_set_halo_dirtybit3(&args[2], range);
  ops_set_halo_dirtybit3(&args[3], range);
  ops_set_halo_dirtybit3(&args[4], range);
  ops_set_halo_dirtybit3(&args[5], range);
  ops_set_halo_dirtybit3(&args[6], range);

  // Update kernel record
  OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg0);
  OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg1);
  OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg2);
  OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg3);
  OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg4);
  OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg5);
  OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg6);
}

//.........这里部分代码省略.........
          (OPS_soa ? args[6].dat->type_size : args[6].dat->elem_size) *
              args[6].dat->size[0] * start[1] * args[6].stencil->stride[1];
#ifdef OPS_GPU
  double *p_a6 = (double *)((char *)args[6].data_d + base6);
#else
  double *p_a6 = (double *)((char *)args[6].data + base6);
#endif

#ifdef OPS_GPU
  int *p_a7 = (int *)args[7].data_d;
#else
  int *p_a7 = arg7h;
#endif

  int x_size = MAX(0, end[0] - start[0]);
  int y_size = MAX(0, end[1] - start[1]);

  // initialize global variable with the dimension of dats
  xdim0 = args[0].dat->size[0];
  xdim1 = args[1].dat->size[0];
  xdim2 = args[2].dat->size[0];
  xdim3 = args[3].dat->size[0];
  xdim4 = args[4].dat->size[0];
  xdim5 = args[5].dat->size[0];
  xdim6 = args[6].dat->size[0];
  if (xdim0 != xdim0_update_halo_kernel1_b1_h ||
      xdim1 != xdim1_update_halo_kernel1_b1_h ||
      xdim2 != xdim2_update_halo_kernel1_b1_h ||
      xdim3 != xdim3_update_halo_kernel1_b1_h ||
      xdim4 != xdim4_update_halo_kernel1_b1_h ||
      xdim5 != xdim5_update_halo_kernel1_b1_h ||
      xdim6 != xdim6_update_halo_kernel1_b1_h) {
    xdim0_update_halo_kernel1_b1 = xdim0;
    xdim0_update_halo_kernel1_b1_h = xdim0;
    xdim1_update_halo_kernel1_b1 = xdim1;
    xdim1_update_halo_kernel1_b1_h = xdim1;
    xdim2_update_halo_kernel1_b1 = xdim2;
    xdim2_update_halo_kernel1_b1_h = xdim2;
    xdim3_update_halo_kernel1_b1 = xdim3;
    xdim3_update_halo_kernel1_b1_h = xdim3;
    xdim4_update_halo_kernel1_b1 = xdim4;
    xdim4_update_halo_kernel1_b1_h = xdim4;
    xdim5_update_halo_kernel1_b1 = xdim5;
    xdim5_update_halo_kernel1_b1_h = xdim5;
    xdim6_update_halo_kernel1_b1 = xdim6;
    xdim6_update_halo_kernel1_b1_h = xdim6;
  }

// Halo Exchanges

#ifdef OPS_GPU
  ops_H_D_exchanges_device(args, 8);
#else
  ops_H_D_exchanges_host(args, 8);
#endif
  ops_halo_exchanges(args, 8, range);

#ifdef OPS_GPU
  ops_H_D_exchanges_device(args, 8);
#else
  ops_H_D_exchanges_host(args, 8);
#endif
  if (OPS_diags > 1) {
    ops_timers_core(&c2, &t2);
    OPS_kernels[10].mpi_time += t2 - t1;
  }

  update_halo_kernel1_b1_c_wrapper(p_a0, p_a1, p_a2, p_a3, p_a4, p_a5, p_a6,
                                   p_a7, x_size, y_size);

  if (OPS_diags > 1) {
    ops_timers_core(&c1, &t1);
    OPS_kernels[10].time += t1 - t2;
  }
#ifdef OPS_GPU
  ops_set_dirtybit_device(args, 8);
#else
  ops_set_dirtybit_host(args, 8);
#endif
  ops_set_halo_dirtybit3(&args[0], range);
  ops_set_halo_dirtybit3(&args[1], range);
  ops_set_halo_dirtybit3(&args[2], range);
  ops_set_halo_dirtybit3(&args[3], range);
  ops_set_halo_dirtybit3(&args[4], range);
  ops_set_halo_dirtybit3(&args[5], range);
  ops_set_halo_dirtybit3(&args[6], range);

  if (OPS_diags > 1) {
    // Update kernel record
    ops_timers_core(&c2, &t2);
    OPS_kernels[10].mpi_time += t2 - t1;
    OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg0);
    OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg1);
    OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg2);
    OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg3);
    OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg4);
    OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg5);
    OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg6);
  }
}

//.........这里部分代码省略.........
                 xarea[OPS_ACC5(0, -1, 0)] * (pressure[OPS_ACC6(0, -1, 0)] -
                                              pressure[OPS_ACC6(-1, -1, 0)]) +
                 xarea[OPS_ACC5(0, 0, -1)] * (pressure[OPS_ACC6(0, 0, -1)] -
                                              pressure[OPS_ACC6(-1, 0, -1)]) +
                 xarea[OPS_ACC5(0, -1, -1)] * (pressure[OPS_ACC6(0, -1, -1)] -
                                               pressure[OPS_ACC6(-1, -1, -1)]));

        yvel1[OPS_ACC8(0, 0, 0)] =
            yvel0[OPS_ACC7(0, 0, 0)] -
            stepbymass[OPS_ACC2(0, 0, 0)] *
                (yarea[OPS_ACC9(0, 0, 0)] * (pressure[OPS_ACC6(0, 0, 0)] -
                                             pressure[OPS_ACC6(0, -1, 0)]) +
                 yarea[OPS_ACC9(-1, 0, 0)] * (pressure[OPS_ACC6(-1, 0, 0)] -
                                              pressure[OPS_ACC6(-1, -1, 0)]) +
                 yarea[OPS_ACC9(0, 0, -1)] * (pressure[OPS_ACC6(0, 0, -1)] -
                                              pressure[OPS_ACC6(0, -1, -1)]) +
                 yarea[OPS_ACC9(-1, 0, -1)] * (pressure[OPS_ACC6(-1, 0, -1)] -
                                               pressure[OPS_ACC6(-1, -1, -1)]));

        zvel1[OPS_ACC12(0, 0, 0)] =
            zvel0[OPS_ACC11(0, 0, 0)] -
            stepbymass[OPS_ACC2(0, 0, 0)] *
                (zarea[OPS_ACC13(0, 0, 0)] * (pressure[OPS_ACC6(0, 0, 0)] -
                                              pressure[OPS_ACC6(0, 0, -1)]) +
                 zarea[OPS_ACC13(0, -1, 0)] * (pressure[OPS_ACC6(0, -1, 0)] -
                                               pressure[OPS_ACC6(0, -1, -1)]) +
                 zarea[OPS_ACC13(-1, 0, 0)] * (pressure[OPS_ACC6(-1, 0, 0)] -
                                               pressure[OPS_ACC6(-1, 0, -1)]) +
                 zarea[OPS_ACC13(-1, -1, 0)] *
                     (pressure[OPS_ACC6(-1, -1, 0)] -
                      pressure[OPS_ACC6(-1, -1, -1)]));

        xvel1[OPS_ACC4(0, 0, 0)] =
            xvel1[OPS_ACC4(0, 0, 0)] -
            stepbymass[OPS_ACC2(0, 0, 0)] *
                (xarea[OPS_ACC5(0, 0, 0)] * (viscosity[OPS_ACC10(0, 0, 0)] -
                                             viscosity[OPS_ACC10(-1, 0, 0)]) +
                 xarea[OPS_ACC5(0, -1, 0)] * (viscosity[OPS_ACC10(0, -1, 0)] -
                                              viscosity[OPS_ACC10(-1, -1, 0)]) +
                 xarea[OPS_ACC5(0, 0, -1)] * (viscosity[OPS_ACC10(0, 0, -1)] -
                                              viscosity[OPS_ACC10(-1, 0, -1)]) +
                 xarea[OPS_ACC5(0, -1, -1)] *
                     (viscosity[OPS_ACC10(0, -1, -1)] -
                      viscosity[OPS_ACC10(-1, -1, -1)]));

        yvel1[OPS_ACC8(0, 0, 0)] =
            yvel1[OPS_ACC8(0, 0, 0)] -
            stepbymass[OPS_ACC2(0, 0, 0)] *
                (yarea[OPS_ACC9(0, 0, 0)] * (viscosity[OPS_ACC10(0, 0, 0)] -
                                             viscosity[OPS_ACC10(0, -1, 0)]) +
                 yarea[OPS_ACC9(-1, 0, 0)] * (viscosity[OPS_ACC10(-1, 0, 0)] -
                                              viscosity[OPS_ACC10(-1, -1, 0)]) +
                 yarea[OPS_ACC9(0, 0, -1)] * (viscosity[OPS_ACC10(0, 0, -1)] -
                                              viscosity[OPS_ACC10(0, -1, -1)]) +
                 yarea[OPS_ACC9(-1, 0, -1)] *
                     (viscosity[OPS_ACC10(-1, 0, -1)] -
                      viscosity[OPS_ACC10(-1, -1, -1)]));

        zvel1[OPS_ACC12(0, 0, 0)] =
            zvel1[OPS_ACC12(0, 0, 0)] -
            stepbymass[OPS_ACC2(0, 0, 0)] *
                (zarea[OPS_ACC13(0, 0, 0)] * (viscosity[OPS_ACC10(0, 0, 0)] -
                                              viscosity[OPS_ACC10(0, 0, -1)]) +
                 zarea[OPS_ACC13(0, -1, 0)] *
                     (viscosity[OPS_ACC10(0, -1, 0)] -
                      viscosity[OPS_ACC10(0, -1, -1)]) +
                 zarea[OPS_ACC13(-1, 0, 0)] *
                     (viscosity[OPS_ACC10(-1, 0, 0)] -
                      viscosity[OPS_ACC10(-1, 0, -1)]) +
                 zarea[OPS_ACC13(-1, -1, 0)] *
                     (viscosity[OPS_ACC10(-1, -1, 0)] -
                      viscosity[OPS_ACC10(-1, -1, -1)]));
      }
    }
  }
  if (OPS_diags > 1) {
    ops_timers_core(&c2, &t2);
    OPS_kernels[105].time += t2 - t1;
  }

  if (OPS_diags > 1) {
    // Update kernel record
    ops_timers_core(&c1, &t1);
    OPS_kernels[105].mpi_time += t1 - t2;
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg0);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg1);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg2);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg3);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg4);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg5);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg6);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg7);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg8);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg9);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg10);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg11);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg12);
    OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg13);
  }
}

// host stub function
void ops_par_loop_update_halo_kernel5_minus_4_back_execute(
    ops_kernel_descriptor *desc) {
  ops_block block = desc->block;
  int dim = desc->dim;
  int *range = desc->range;
  ops_arg arg0 = desc->args[0];
  ops_arg arg1 = desc->args[1];
  ops_arg arg2 = desc->args[2];

  // Timing
  double t1, t2, c1, c2;

  ops_arg args[3] = {arg0, arg1, arg2};

#ifdef CHECKPOINTING
  if (!ops_checkpointing_before(args, 3, range, 92))
    return;
#endif

  if (OPS_diags > 1) {
    OPS_kernels[92].count++;
    ops_timers_core(&c2, &t2);
  }

  // compute locally allocated range for the sub-block
  int start[3];
  int end[3];

  for (int n = 0; n < 3; n++) {
    start[n] = range[2 * n];
    end[n] = range[2 * n + 1];
  }

#ifdef OPS_DEBUG
  ops_register_args(args, "update_halo_kernel5_minus_4_back");
#endif

  // set up initial pointers and exchange halos if necessary
  int base0 = args[0].dat->base_offset;
  double *__restrict__ vol_flux_z = (double *)(args[0].data + base0);

  int base1 = args[1].dat->base_offset;
  double *__restrict__ mass_flux_z = (double *)(args[1].data + base1);

  const int *__restrict__ fields = (int *)args[2].data;

  // initialize global variable with the dimension of dats
  int xdim0_update_halo_kernel5_minus_4_back = args[0].dat->size[0];
  int ydim0_update_halo_kernel5_minus_4_back = args[0].dat->size[1];
  int xdim1_update_halo_kernel5_minus_4_back = args[1].dat->size[0];
  int ydim1_update_halo_kernel5_minus_4_back = args[1].dat->size[1];

  if (OPS_diags > 1) {
    ops_timers_core(&c1, &t1);
    OPS_kernels[92].mpi_time += t1 - t2;
  }

#pragma omp parallel for collapse(2)
  for (int n_z = start[2]; n_z < end[2]; n_z++) {
    for (int n_y = start[1]; n_y < end[1]; n_y++) {
#ifdef intel
#pragma loop_count(10000)
#pragma omp simd aligned(vol_flux_z, mass_flux_z)
#else
#pragma simd
#endif
      for (int n_x = start[0]; n_x < end[0]; n_x++) {

        if (fields[FIELD_VOL_FLUX_Z] == 1)
          vol_flux_z[OPS_ACC0(0, 0, 0)] = -vol_flux_z[OPS_ACC0(0, 0, 4)];
        if (fields[FIELD_MASS_FLUX_Z] == 1)
          mass_flux_z[OPS_ACC1(0, 0, 0)] = -mass_flux_z[OPS_ACC1(0, 0, 4)];
      }
    }
  }
  if (OPS_diags > 1) {
    ops_timers_core(&c2, &t2);
    OPS_kernels[92].time += t2 - t1;
  }

  if (OPS_diags > 1) {
    // Update kernel record
    ops_timers_core(&c1, &t1);
    OPS_kernels[92].mpi_time += t1 - t2;
    OPS_kernels[92].transfer += ops_compute_transfer(dim, start, end, &arg0);
    OPS_kernels[92].transfer += ops_compute_transfer(dim, start, end, &arg1);
  }
}

//.........这里部分代码省略.........

        }

        //shift pointers to data x direction
        p_a[0]= p_a[0] + (dat0 * off0_0)*SIMD_VEC;
        p_a[1]= p_a[1] + (dat1 * off1_0)*SIMD_VEC;
        p_a[2]= p_a[2] + (dat2 * off2_0)*SIMD_VEC;
        p_a[3]= p_a[3] + (dat3 * off3_0)*SIMD_VEC;
        p_a[4]= p_a[4] + (dat4 * off4_0)*SIMD_VEC;
        p_a[5]= p_a[5] + (dat5 * off5_0)*SIMD_VEC;
        p_a[6]= p_a[6] + (dat6 * off6_0)*SIMD_VEC;
        p_a[7]= p_a[7] + (dat7 * off7_0)*SIMD_VEC;
        p_a[8]= p_a[8] + (dat8 * off8_0)*SIMD_VEC;
        p_a[9]= p_a[9] + (dat9 * off9_0)*SIMD_VEC;
        p_a[10]= p_a[10] + (dat10 * off10_0)*SIMD_VEC;
        p_a[11]= p_a[11] + (dat11 * off11_0)*SIMD_VEC;
        p_a[12]= p_a[12] + (dat12 * off12_0)*SIMD_VEC;
        p_a[13]= p_a[13] + (dat13 * off13_0)*SIMD_VEC;
      }

      for ( int n_x=start[0]+((end[0]-start[0])/SIMD_VEC)*SIMD_VEC; n_x<end[0]; n_x++ ){
        //call kernel function, passing in pointers to data - remainder
        PdV_kernel_predict(  (double *)p_a[0], (double *)p_a[1], (double *)p_a[2],
           (double *)p_a[3], (double *)p_a[4], (double *)p_a[5], (double *)p_a[6],
           (double *)p_a[7], (double *)p_a[8], (double *)p_a[9], (double *)p_a[10],
           (double *)p_a[11], (double *)p_a[12], (double *)p_a[13] );


        //shift pointers to data x direction
        p_a[0]= p_a[0] + (dat0 * off0_0);
        p_a[1]= p_a[1] + (dat1 * off1_0);
        p_a[2]= p_a[2] + (dat2 * off2_0);
        p_a[3]= p_a[3] + (dat3 * off3_0);
        p_a[4]= p_a[4] + (dat4 * off4_0);
        p_a[5]= p_a[5] + (dat5 * off5_0);
        p_a[6]= p_a[6] + (dat6 * off6_0);
        p_a[7]= p_a[7] + (dat7 * off7_0);
        p_a[8]= p_a[8] + (dat8 * off8_0);
        p_a[9]= p_a[9] + (dat9 * off9_0);
        p_a[10]= p_a[10] + (dat10 * off10_0);
        p_a[11]= p_a[11] + (dat11 * off11_0);
        p_a[12]= p_a[12] + (dat12 * off12_0);
        p_a[13]= p_a[13] + (dat13 * off13_0);
      }

      //shift pointers to data y direction
      p_a[0]= p_a[0] + (dat0 * off0_1);
      p_a[1]= p_a[1] + (dat1 * off1_1);
      p_a[2]= p_a[2] + (dat2 * off2_1);
      p_a[3]= p_a[3] + (dat3 * off3_1);
      p_a[4]= p_a[4] + (dat4 * off4_1);
      p_a[5]= p_a[5] + (dat5 * off5_1);
      p_a[6]= p_a[6] + (dat6 * off6_1);
      p_a[7]= p_a[7] + (dat7 * off7_1);
      p_a[8]= p_a[8] + (dat8 * off8_1);
      p_a[9]= p_a[9] + (dat9 * off9_1);
      p_a[10]= p_a[10] + (dat10 * off10_1);
      p_a[11]= p_a[11] + (dat11 * off11_1);
      p_a[12]= p_a[12] + (dat12 * off12_1);
      p_a[13]= p_a[13] + (dat13 * off13_1);
    }
    //shift pointers to data z direction
    p_a[0]= p_a[0] + (dat0 * off0_2);
    p_a[1]= p_a[1] + (dat1 * off1_2);
    p_a[2]= p_a[2] + (dat2 * off2_2);
    p_a[3]= p_a[3] + (dat3 * off3_2);
    p_a[4]= p_a[4] + (dat4 * off4_2);
    p_a[5]= p_a[5] + (dat5 * off5_2);
    p_a[6]= p_a[6] + (dat6 * off6_2);
    p_a[7]= p_a[7] + (dat7 * off7_2);
    p_a[8]= p_a[8] + (dat8 * off8_2);
    p_a[9]= p_a[9] + (dat9 * off9_2);
    p_a[10]= p_a[10] + (dat10 * off10_2);
    p_a[11]= p_a[11] + (dat11 * off11_2);
    p_a[12]= p_a[12] + (dat12 * off12_2);
    p_a[13]= p_a[13] + (dat13 * off13_2);
  }
  ops_timers_core(&c2,&t2);
  OPS_kernels[5].time += t2-t1;
  ops_set_dirtybit_host(args, 14);
  ops_set_halo_dirtybit3(&args[4],range);
  ops_set_halo_dirtybit3(&args[8],range);
  ops_set_halo_dirtybit3(&args[11],range);

  //Update kernel record
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg0);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg1);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg2);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg3);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg4);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg5);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg6);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg7);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg8);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg9);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg10);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg11);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg12);
  OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg13);
}