1 // Copyright (c) 2017-2018, Lawrence Livermore National Security, LLC. 2 // Produced at the Lawrence Livermore National Laboratory. LLNL-CODE-734707. 3 // All Rights reserved. See files LICENSE and NOTICE for details. 4 // 5 // This file is part of CEED, a collection of benchmarks, miniapps, software 6 // libraries and APIs for efficient high-order finite element and spectral 7 // element discretizations for exascale applications. For more information and 8 // source code availability see http://github.com/ceed. 9 // 10 // The CEED research is supported by the Exascale Computing Project 17-SC-20-SC, 11 // a collaborative effort of two U.S. Department of Energy organizations (Office 12 // of Science and the National Nuclear Security Administration) responsible for 13 // the planning and preparation of a capable exascale ecosystem, including 14 // software, applications, hardware, advanced system engineering and early 15 // testbed platforms, in support of the nation's exascale computing imperative. 16 17 #include <ceed-impl.h> 18 #include <string.h> 19 #include "ceed-ref.h" 20 21 // Contracts on the middle index 22 // NOTRANSPOSE: V_ajc = T_jb U_abc 23 // TRANSPOSE: V_ajc = T_bj U_abc 24 // If Add != 0, "=" is replaced by "+=" 25 static int CeedTensorContract_Ref(Ceed ceed, 26 CeedInt A, CeedInt B, CeedInt C, CeedInt J, 27 const CeedScalar *restrict t, CeedTransposeMode tmode, 28 const CeedInt Add, 29 const CeedScalar *restrict u, CeedScalar *restrict v) { 30 CeedInt tstride0 = B, tstride1 = 1; 31 if (tmode == CEED_TRANSPOSE) { 32 tstride0 = 1; tstride1 = J; 33 } 34 35 if (!Add) { 36 for (CeedInt q=0; q<A*J*C; q++) { 37 v[q] = (CeedScalar) 0.0; 38 } 39 } 40 41 for (CeedInt a=0; a<A; a++) { 42 for (CeedInt b=0; b<B; b++) { 43 for (CeedInt j=0; j<J; j++) { 44 CeedScalar tq = t[j*tstride0 + b*tstride1]; 45 for (CeedInt c=0; c<C; c++) { 46 v[(a*J+j)*C+c] += tq * u[(a*B+b)*C+c]; 47 } 48 } 49 } 50 } 51 return 0; 52 } 53 54 static int CeedBasisApply_Ref(CeedBasis basis, CeedTransposeMode tmode, 55 CeedEvalMode emode, 56 const CeedScalar *u, CeedScalar *v) { 57 int ierr; 58 const CeedInt dim = basis->dim; 59 const CeedInt ndof = basis->ndof; 60 const CeedInt nqpt = ndof*CeedPowInt(basis->Q1d, dim); 61 const CeedInt add = (tmode == CEED_TRANSPOSE); 62 63 if (tmode == CEED_TRANSPOSE) { 64 const CeedInt vsize = ndof*CeedPowInt(basis->P1d, dim); 65 for (CeedInt i = 0; i < vsize; i++) 66 v[i] = (CeedScalar) 0; 67 } 68 if (emode & CEED_EVAL_INTERP) { 69 CeedInt P = basis->P1d, Q = basis->Q1d; 70 if (tmode == CEED_TRANSPOSE) { 71 P = basis->Q1d; Q = basis->P1d; 72 } 73 CeedInt pre = ndof*CeedPowInt(P, dim-1), post = 1; 74 CeedScalar tmp[2][ndof*Q*CeedPowInt(P>Q?P:Q, dim-1)]; 75 for (CeedInt d=0; d<dim; d++) { 76 ierr = CeedTensorContract_Ref(basis->ceed, pre, P, post, Q, basis->interp1d, 77 tmode, add&&(d==dim-1), 78 d==0?u:tmp[d%2], d==dim-1?v:tmp[(d+1)%2]); 79 CeedChk(ierr); 80 pre /= P; 81 post *= Q; 82 } 83 if (tmode == CEED_NOTRANSPOSE) { 84 v += nqpt; 85 } else { 86 u += nqpt; 87 } 88 } 89 if (emode & CEED_EVAL_GRAD) { 90 CeedInt P = basis->P1d, Q = basis->Q1d; 91 // In CEED_NOTRANSPOSE mode: 92 // u is (P^dim x nc), column-major layout (nc = ndof) 93 // v is (Q^dim x nc x dim), column-major layout (nc = ndof) 94 // In CEED_TRANSPOSE mode, the sizes of u and v are switched. 95 if (tmode == CEED_TRANSPOSE) { 96 P = basis->Q1d, Q = basis->P1d; 97 } 98 CeedScalar tmp[2][ndof*Q*CeedPowInt(P>Q?P:Q, dim-1)]; 99 for (CeedInt p = 0; p < dim; p++) { 100 CeedInt pre = ndof*CeedPowInt(P, dim-1), post = 1; 101 for (CeedInt d=0; d<dim; d++) { 102 ierr = CeedTensorContract_Ref(basis->ceed, pre, P, post, Q, 103 (p==d)?basis->grad1d:basis->interp1d, 104 tmode, add&&(d==dim-1), 105 d==0?u:tmp[d%2], d==dim-1?v:tmp[(d+1)%2]); 106 CeedChk(ierr); 107 pre /= P; 108 post *= Q; 109 } 110 if (tmode == CEED_NOTRANSPOSE) { 111 v += nqpt; 112 } else { 113 u += nqpt; 114 } 115 } 116 } 117 if (emode & CEED_EVAL_WEIGHT) { 118 if (tmode == CEED_TRANSPOSE) 119 return CeedError(basis->ceed, 1, 120 "CEED_EVAL_WEIGHT incompatible with CEED_TRANSPOSE"); 121 CeedInt Q = basis->Q1d; 122 for (CeedInt d=0; d<dim; d++) { 123 CeedInt pre = CeedPowInt(Q, dim-d-1), post = CeedPowInt(Q, d); 124 for (CeedInt i=0; i<pre; i++) { 125 for (CeedInt j=0; j<Q; j++) { 126 for (CeedInt k=0; k<post; k++) { 127 v[(i*Q + j)*post + k] = basis->qweight1d[j] 128 * (d == 0 ? 1 : v[(i*Q + j)*post + k]); 129 } 130 } 131 } 132 } 133 } 134 return 0; 135 } 136 137 static int CeedBasisDestroy_Ref(CeedBasis basis) { 138 return 0; 139 } 140 141 int CeedBasisCreateTensorH1_Ref(Ceed ceed, CeedInt dim, CeedInt P1d, 142 CeedInt Q1d, const CeedScalar *interp1d, 143 const CeedScalar *grad1d, 144 const CeedScalar *qref1d, 145 const CeedScalar *qweight1d, 146 CeedBasis basis) { 147 basis->Apply = CeedBasisApply_Ref; 148 basis->Destroy = CeedBasisDestroy_Ref; 149 return 0; 150 } 151