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 #ifndef bp3_h 18 #define bp3_h 19 #include <ceed.h> 20 21 /// A structure used to pass additional data to f_build_diff and f_apply_diff 22 struct BuildContext { CeedInt dim, space_dim; }; 23 24 /// libCEED Q-function for building quadrature data for a diffusion operator 25 CEED_QFUNCTION(f_build_diff)(void *ctx, const CeedInt Q, 26 const CeedScalar *const *in, CeedScalar *const *out) { 27 BuildContext *bc = (BuildContext *)ctx; 28 // in[0] is Jacobians with shape [dim, nc=dim, Q] 29 // in[1] is quadrature weights, size (Q) 30 // 31 // At every quadrature point, compute w/det(J).adj(J).adj(J)^T and store 32 // the symmetric part of the result. 33 const CeedScalar *J = in[0], *w = in[1]; 34 CeedScalar *qdata = out[0]; 35 36 switch (bc->dim + 10*bc->space_dim) { 37 case 11: 38 // Quadrature Point Loop 39 CeedPragmaSIMD 40 for (CeedInt i=0; i<Q; i++) { 41 qdata[i] = w[i] / J[i]; 42 } 43 break; 44 case 22: 45 // Quadrature Point Loop 46 CeedPragmaSIMD 47 for (CeedInt i=0; i<Q; i++) { 48 // J: 0 2 qdata: 0 2 adj(J): J22 -J12 49 // 1 3 2 1 -J21 J11 50 const CeedScalar J11 = J[i+Q*0]; 51 const CeedScalar J21 = J[i+Q*1]; 52 const CeedScalar J12 = J[i+Q*2]; 53 const CeedScalar J22 = J[i+Q*3]; 54 const CeedScalar qw = w[i] / (J11*J22 - J21*J12); 55 qdata[i+Q*0] = qw * (J12*J12 + J22*J22); 56 qdata[i+Q*1] = qw * (J11*J11 + J21*J21); 57 qdata[i+Q*2] = - qw * (J11*J12 + J21*J22); 58 } 59 break; 60 case 33: 61 // Quadrature Point Loop 62 CeedPragmaSIMD 63 for (CeedInt i=0; i<Q; i++) { 64 // J: 0 3 6 qdata: 0 5 4 65 // 1 4 7 5 1 3 66 // 2 5 8 4 3 2 67 const CeedScalar J11 = J[i+Q*0]; 68 const CeedScalar J21 = J[i+Q*1]; 69 const CeedScalar J31 = J[i+Q*2]; 70 const CeedScalar J12 = J[i+Q*3]; 71 const CeedScalar J22 = J[i+Q*4]; 72 const CeedScalar J32 = J[i+Q*5]; 73 const CeedScalar J13 = J[i+Q*6]; 74 const CeedScalar J23 = J[i+Q*7]; 75 const CeedScalar J33 = J[i+Q*8]; 76 const CeedScalar A11 = J22*J33 - J23*J32; 77 const CeedScalar A12 = J13*J32 - J12*J33; 78 const CeedScalar A13 = J12*J23 - J13*J22; 79 const CeedScalar A21 = J23*J31 - J21*J33; 80 const CeedScalar A22 = J11*J33 - J13*J31; 81 const CeedScalar A23 = J13*J21 - J11*J23; 82 const CeedScalar A31 = J21*J32 - J22*J31; 83 const CeedScalar A32 = J12*J31 - J11*J32; 84 const CeedScalar A33 = J11*J22 - J12*J21; 85 const CeedScalar qw = w[i] / (J11*A11 + J21*A12 + J31*A13); 86 qdata[i+Q*0] = qw * (A11*A11 + A12*A12 + A13*A13); 87 qdata[i+Q*1] = qw * (A21*A21 + A22*A22 + A23*A23); 88 qdata[i+Q*2] = qw * (A31*A31 + A32*A32 + A33*A33); 89 qdata[i+Q*3] = qw * (A21*A31 + A22*A32 + A23*A33); 90 qdata[i+Q*4] = qw * (A11*A31 + A12*A32 + A13*A33); 91 qdata[i+Q*5] = qw * (A11*A21 + A12*A22 + A13*A23); 92 } 93 break; 94 } 95 return 0; 96 } 97 98 /// libCEED Q-function for applying a diff operator 99 CEED_QFUNCTION(f_apply_diff)(void *ctx, const CeedInt Q, 100 const CeedScalar *const *in, CeedScalar *const *out) { 101 BuildContext *bc = (BuildContext *)ctx; 102 // in[0], out[0] have shape [dim, nc=1, Q] 103 const CeedScalar *ug = in[0], *qdata = in[1]; 104 CeedScalar *vg = out[0]; 105 106 switch (bc->dim) { 107 case 1: 108 // Quadrature Point Loop 109 CeedPragmaSIMD 110 for (CeedInt i=0; i<Q; i++) { 111 vg[i] = ug[i] * qdata[i]; 112 } 113 break; 114 case 2: 115 // Quadrature Point Loop 116 CeedPragmaSIMD 117 for (CeedInt i=0; i<Q; i++) { 118 const CeedScalar ug0 = ug[i+Q*0]; 119 const CeedScalar ug1 = ug[i+Q*1]; 120 vg[i+Q*0] = qdata[i+Q*0]*ug0 + qdata[i+Q*2]*ug1; 121 vg[i+Q*1] = qdata[i+Q*2]*ug0 + qdata[i+Q*1]*ug1; 122 } 123 break; 124 case 3: 125 // Quadrature Point Loop 126 CeedPragmaSIMD 127 for (CeedInt i=0; i<Q; i++) { 128 const CeedScalar ug0 = ug[i+Q*0]; 129 const CeedScalar ug1 = ug[i+Q*1]; 130 const CeedScalar ug2 = ug[i+Q*2]; 131 vg[i+Q*0] = qdata[i+Q*0]*ug0 + qdata[i+Q*5]*ug1 + qdata[i+Q*4]*ug2; 132 vg[i+Q*1] = qdata[i+Q*5]*ug0 + qdata[i+Q*1]*ug1 + qdata[i+Q*3]*ug2; 133 vg[i+Q*2] = qdata[i+Q*4]*ug0 + qdata[i+Q*3]*ug1 + qdata[i+Q*2]*ug2; 134 } 135 break; 136 } 137 return 0; 138 } 139 140 #endif // bp3_h 141