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_HCURL_HDIV_32_QF_H
       5              : #define PALACE_LIBCEED_HCURL_HDIV_32_QF_H
       6              : 
       7              : #include "../coeff/coeff_3_qf.h"
       8              : #include "utils_32_qf.h"
       9              : 
      10            0 : CEED_QFUNCTION(f_apply_hcurlhdiv_32)(void *__restrict__ ctx, CeedInt Q,
      11              :                                      const CeedScalar *const *in, CeedScalar *const *out)
      12              : {
      13            0 :   const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *adjJt = in[0] + 2 * Q, *u = in[1];
      14            0 :   CeedScalar *v = out[0];
      15              : 
      16            0 :   CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
      17              :   {
      18            0 :     const CeedScalar u_loc[2] = {u[i + Q * 0], u[i + Q * 1]};
      19              :     CeedScalar coeff[9], adjJt_loc[6], J_loc[6], v_loc[2];
      20            0 :     CoeffUnpack3((const CeedIntScalar *)ctx, (CeedInt)attr[i], coeff);
      21            0 :     MatUnpack32(adjJt + i, Q, adjJt_loc);
      22              :     AdjJt32(adjJt_loc, J_loc);
      23              :     MultAtBCx32(J_loc, coeff, adjJt_loc, u_loc, v_loc);
      24              : 
      25            0 :     v[i + Q * 0] = wdetJ[i] * v_loc[0];
      26            0 :     v[i + Q * 1] = wdetJ[i] * v_loc[1];
      27              :   }
      28            0 :   return 0;
      29              : }
      30              : 
      31            0 : CEED_QFUNCTION(f_apply_hdivhcurl_32)(void *__restrict__ ctx, CeedInt Q,
      32              :                                      const CeedScalar *const *in, CeedScalar *const *out)
      33              : {
      34            0 :   const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *adjJt = in[0] + 2 * Q, *u = in[1];
      35            0 :   CeedScalar *v = out[0];
      36              : 
      37            0 :   CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
      38              :   {
      39            0 :     const CeedScalar u_loc[2] = {u[i + Q * 0], u[i + Q * 1]};
      40              :     CeedScalar coeff[9], adjJt_loc[6], J_loc[6], v_loc[2];
      41            0 :     CoeffUnpack3((const CeedIntScalar *)ctx, (CeedInt)attr[i], coeff);
      42            0 :     MatUnpack32(adjJt + i, Q, adjJt_loc);
      43              :     AdjJt32(adjJt_loc, J_loc);
      44              :     MultAtBCx32(adjJt_loc, coeff, J_loc, u_loc, v_loc);
      45              : 
      46            0 :     v[i + Q * 0] = wdetJ[i] * v_loc[0];
      47            0 :     v[i + Q * 1] = wdetJ[i] * v_loc[1];
      48              :   }
      49            0 :   return 0;
      50              : }
      51              : 
      52              : #endif  // PALACE_LIBCEED_HCURL_HDIV_32_QF_H