Line data    Source code

       1              : // Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
       2              : // SPDX-License-Identifier: Apache-2.0
       3              : 
       4              : #ifndef PALACE_LIBCEED_HCURLHDIV_ERROR_22_QF_H
       5              : #define PALACE_LIBCEED_HCURLHDIV_ERROR_22_QF_H
       6              : 
       7              : #include "../coeff/coeff_2_qf.h"
       8              : #include "utils_22_qf.h"
       9              : 
      10            0 : CEED_QFUNCTION(f_apply_hcurlhdiv_error_22)(void *__restrict__ ctx, CeedInt Q,
      11              :                                            const CeedScalar *const *in,
      12              :                                            CeedScalar *const *out)
      13              : {
      14            0 :   const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *adjJt = in[0] + 2 * Q, *u1 = in[1],
      15            0 :                    *u2 = in[2];
      16            0 :   CeedScalar *v = out[0];
      17              : 
      18            0 :   CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
      19              :   {
      20              :     CeedScalar adjJt_loc[4], v1_loc[2], v2_loc[2];
      21            0 :     MatUnpack22(adjJt + i, Q, adjJt_loc);
      22              :     {
      23            0 :       const CeedScalar u1_loc[2] = {u1[i + Q * 0], u1[i + Q * 1]};
      24              :       CeedScalar coeff[4];
      25            0 :       CoeffUnpack2((const CeedIntScalar *)ctx, (CeedInt)attr[i], coeff);
      26              :       MultBAx22(adjJt_loc, coeff, u1_loc, v1_loc);
      27              :     }
      28              :     {
      29            0 :       const CeedScalar u2_loc[2] = {u2[i + Q * 0], u2[i + Q * 1]};
      30              :       CeedScalar coeff[4], J_loc[4];
      31              :       CoeffUnpack2(CoeffPairSecond<2>((const CeedIntScalar *)ctx), (CeedInt)attr[i], coeff);
      32              :       AdjJt22(adjJt_loc, J_loc);
      33              :       MultBAx22(J_loc, coeff, u2_loc, v2_loc);
      34              :     }
      35            0 :     v2_loc[0] -= v1_loc[0];
      36            0 :     v2_loc[1] -= v1_loc[1];
      37            0 :     v[i] = wdetJ[i] * (v2_loc[0] * v2_loc[0] + v2_loc[1] * v2_loc[1]);
      38              :   }
      39            0 :   return 0;
      40              : }
      41              : 
      42              : CEED_QFUNCTION(f_apply_hdivhcurl_error_22)(void *__restrict__ ctx, CeedInt Q,
      43              :                                            const CeedScalar *const *in,
      44              :                                            CeedScalar *const *out)
      45              : {
      46              :   const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *adjJt = in[0] + 2 * Q, *u1 = in[1],
      47              :                    *u2 = in[2];
      48              :   CeedScalar *v = out[0];
      49              : 
      50              :   CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
      51              :   {
      52              :     CeedScalar adjJt_loc[4], v1_loc[2], v2_loc[2];
      53              :     MatUnpack22(adjJt + i, Q, adjJt_loc);
      54              :     {
      55              :       const CeedScalar u1_loc[2] = {u1[i + Q * 0], u1[i + Q * 1]};
      56              :       CeedScalar coeff[4], J_loc[4];
      57              :       CoeffUnpack2((const CeedIntScalar *)ctx, (CeedInt)attr[i], coeff);
      58              :       AdjJt22(adjJt_loc, J_loc);
      59              :       MultBAx22(J_loc, coeff, u1_loc, v1_loc);
      60              :     }
      61              :     {
      62              :       const CeedScalar u2_loc[2] = {u2[i + Q * 0], u2[i + Q * 1]};
      63              :       CeedScalar coeff[4];
      64              :       CoeffUnpack2(CoeffPairSecond<2>((const CeedIntScalar *)ctx), (CeedInt)attr[i], coeff);
      65              :       MultBAx22(adjJt_loc, coeff, u2_loc, v2_loc);
      66              :     }
      67              :     v2_loc[0] -= v1_loc[0];
      68              :     v2_loc[1] -= v1_loc[1];
      69              :     v[i] = wdetJ[i] * (v2_loc[0] * v2_loc[0] + v2_loc[1] * v2_loc[1]);
      70              :   }
      71              :   return 0;
      72              : }
      73              : 
      74              : #endif  // PALACE_LIBCEED_HCURLHDIV_ERROR_22_QF_H