16e15d496SJeremy L Thompson /// @file 26e15d496SJeremy L Thompson /// Test FLOP estimation for composite mass matrix operator 36e15d496SJeremy L Thompson /// \test Test FLOP estimation for composite mass matrix operator 46e15d496SJeremy L Thompson #include <ceed.h> 56e15d496SJeremy L Thompson #include <math.h> 6*2b730f8bSJeremy L Thompson #include <stdlib.h> 7*2b730f8bSJeremy L Thompson 86e15d496SJeremy L Thompson #include "t320-basis.h" 96e15d496SJeremy L Thompson 106e15d496SJeremy L Thompson /* The mesh comprises of two rows of 3 quadralaterals followed by one row 116e15d496SJeremy L Thompson of 6 triangles: 126e15d496SJeremy L Thompson _ _ _ 136e15d496SJeremy L Thompson |_|_|_| 146e15d496SJeremy L Thompson |_|_|_| 156e15d496SJeremy L Thompson |/|/|/| 166e15d496SJeremy L Thompson 176e15d496SJeremy L Thompson */ 186e15d496SJeremy L Thompson 196e15d496SJeremy L Thompson int main(int argc, char **argv) { 206e15d496SJeremy L Thompson Ceed ceed; 219d36ca50SJeremy L Thompson CeedSize flop_estimate; 22*2b730f8bSJeremy L Thompson CeedElemRestriction elem_restr_x_tet, elem_restr_u_tet, elem_restr_qd_i_tet, elem_restr_x_hex, elem_restr_u_hex, elem_restr_qd_i_hex; 23*2b730f8bSJeremy L Thompson CeedBasis basis_x_tet, basis_u_tet, basis_x_hex, basis_u_hex; 246e15d496SJeremy L Thompson CeedQFunction qf_mass; 256e15d496SJeremy L Thompson CeedOperator op_mass_tet, op_mass_hex, op_mass; 266e15d496SJeremy L Thompson CeedVector q_data_tet, q_data_hex; 27*2b730f8bSJeremy L Thompson CeedInt num_elem_tet = 6, P_tet = 6, Q_tet = 4, num_elem_hex = 6, P_hex = 3, Q_hex = 4, dim = 2; 28*2b730f8bSJeremy L Thompson CeedInt n_x = 3, n_y = 3, n_x_tet = 3, n_y_tet = 1, n_x_hex = 3; 296e15d496SJeremy L Thompson CeedInt row, col, offset; 30*2b730f8bSJeremy L Thompson CeedInt num_dofs = (n_x * 2 + 1) * (n_y * 2 + 1), num_qpts_tet = num_elem_tet * Q_tet, num_qpts_hex = num_elem_hex * Q_hex * Q_hex; 31*2b730f8bSJeremy L Thompson CeedInt ind_x_tet[num_elem_tet * P_tet], ind_x_hex[num_elem_hex * P_hex * P_hex]; 326e15d496SJeremy L Thompson CeedScalar q_ref[dim * Q_tet], q_weight[Q_tet]; 336e15d496SJeremy L Thompson CeedScalar interp[P_tet * Q_tet], grad[dim * P_tet * Q_tet]; 346e15d496SJeremy L Thompson 356e15d496SJeremy L Thompson CeedInit(argv[1], &ceed); 366e15d496SJeremy L Thompson 376e15d496SJeremy L Thompson // Qdata Vectors 386e15d496SJeremy L Thompson CeedVectorCreate(ceed, num_qpts_tet, &q_data_tet); 396e15d496SJeremy L Thompson CeedVectorCreate(ceed, num_qpts_hex, &q_data_hex); 406e15d496SJeremy L Thompson 416e15d496SJeremy L Thompson // Set up Tet Elements 426e15d496SJeremy L Thompson for (CeedInt i = 0; i < num_elem_tet / 2; i++) { 436e15d496SJeremy L Thompson col = i % n_x_tet; 446e15d496SJeremy L Thompson row = i / n_x_tet; 456e15d496SJeremy L Thompson offset = col * 2 + row * (n_x_tet * 2 + 1) * 2; 466e15d496SJeremy L Thompson 476e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 0] = 2 + offset; 486e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 1] = 9 + offset; 496e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 2] = 16 + offset; 506e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 3] = 1 + offset; 516e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 4] = 8 + offset; 526e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 5] = 0 + offset; 536e15d496SJeremy L Thompson 546e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 6] = 14 + offset; 556e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 7] = 7 + offset; 566e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 8] = 0 + offset; 576e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 9] = 15 + offset; 586e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 10] = 8 + offset; 596e15d496SJeremy L Thompson ind_x_tet[i * 2 * P_tet + 11] = 16 + offset; 606e15d496SJeremy L Thompson } 616e15d496SJeremy L Thompson 626e15d496SJeremy L Thompson // -- Restrictions 63*2b730f8bSJeremy L Thompson CeedElemRestrictionCreate(ceed, num_elem_tet, P_tet, dim, num_dofs, dim * num_dofs, CEED_MEM_HOST, CEED_USE_POINTER, ind_x_tet, &elem_restr_x_tet); 646e15d496SJeremy L Thompson 65*2b730f8bSJeremy L Thompson CeedElemRestrictionCreate(ceed, num_elem_tet, P_tet, 1, 1, num_dofs, CEED_MEM_HOST, CEED_USE_POINTER, ind_x_tet, &elem_restr_u_tet); 666e15d496SJeremy L Thompson CeedInt strides_qd_tet[3] = {1, Q_tet, Q_tet}; 67*2b730f8bSJeremy L Thompson CeedElemRestrictionCreateStrided(ceed, num_elem_tet, Q_tet, 1, num_qpts_tet, strides_qd_tet, &elem_restr_qd_i_tet); 686e15d496SJeremy L Thompson 696e15d496SJeremy L Thompson // -- Bases 706e15d496SJeremy L Thompson buildmats(q_ref, q_weight, interp, grad); 71*2b730f8bSJeremy L Thompson CeedBasisCreateH1(ceed, CEED_TOPOLOGY_TRIANGLE, dim, P_tet, Q_tet, interp, grad, q_ref, q_weight, &basis_x_tet); 726e15d496SJeremy L Thompson 736e15d496SJeremy L Thompson buildmats(q_ref, q_weight, interp, grad); 74*2b730f8bSJeremy L Thompson CeedBasisCreateH1(ceed, CEED_TOPOLOGY_TRIANGLE, 1, P_tet, Q_tet, interp, grad, q_ref, q_weight, &basis_u_tet); 756e15d496SJeremy L Thompson 766e15d496SJeremy L Thompson // -- QFunction 776e15d496SJeremy L Thompson CeedQFunctionCreateInteriorByName(ceed, "MassApply", &qf_mass); 786e15d496SJeremy L Thompson 796e15d496SJeremy L Thompson // -- Operators 806e15d496SJeremy L Thompson // ---- Mass Tet 81*2b730f8bSJeremy L Thompson CeedOperatorCreate(ceed, qf_mass, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, &op_mass_tet); 82*2b730f8bSJeremy L Thompson CeedOperatorSetField(op_mass_tet, "u", elem_restr_u_tet, basis_u_tet, CEED_VECTOR_ACTIVE); 83*2b730f8bSJeremy L Thompson CeedOperatorSetField(op_mass_tet, "qdata", elem_restr_qd_i_tet, CEED_BASIS_COLLOCATED, q_data_tet); 84*2b730f8bSJeremy L Thompson CeedOperatorSetField(op_mass_tet, "v", elem_restr_u_tet, basis_u_tet, CEED_VECTOR_ACTIVE); 856e15d496SJeremy L Thompson 866e15d496SJeremy L Thompson // Set up Hex Elements 876e15d496SJeremy L Thompson for (CeedInt i = 0; i < num_elem_hex; i++) { 886e15d496SJeremy L Thompson col = i % n_x_hex; 896e15d496SJeremy L Thompson row = i / n_x_hex; 906e15d496SJeremy L Thompson offset = (n_x_tet * 2 + 1) * (n_y_tet * 2) * (1 + row) + col * 2; 91*2b730f8bSJeremy L Thompson for (CeedInt j = 0; j < P_hex; j++) { 92*2b730f8bSJeremy L Thompson for (CeedInt k = 0; k < P_hex; k++) ind_x_hex[P_hex * (P_hex * i + k) + j] = offset + k * (n_x_hex * 2 + 1) + j; 93*2b730f8bSJeremy L Thompson } 946e15d496SJeremy L Thompson } 956e15d496SJeremy L Thompson 966e15d496SJeremy L Thompson // -- Restrictions 97*2b730f8bSJeremy L Thompson CeedElemRestrictionCreate(ceed, num_elem_hex, P_hex * P_hex, dim, num_dofs, dim * num_dofs, CEED_MEM_HOST, CEED_USE_POINTER, ind_x_hex, 986e15d496SJeremy L Thompson &elem_restr_x_hex); 996e15d496SJeremy L Thompson 100*2b730f8bSJeremy L Thompson CeedElemRestrictionCreate(ceed, num_elem_hex, P_hex * P_hex, 1, 1, num_dofs, CEED_MEM_HOST, CEED_USE_POINTER, ind_x_hex, &elem_restr_u_hex); 1016e15d496SJeremy L Thompson CeedInt strides_qd_hex[3] = {1, Q_hex * Q_hex, Q_hex * Q_hex}; 102*2b730f8bSJeremy L Thompson CeedElemRestrictionCreateStrided(ceed, num_elem_hex, Q_hex * Q_hex, 1, num_qpts_hex, strides_qd_hex, &elem_restr_qd_i_hex); 1036e15d496SJeremy L Thompson 1046e15d496SJeremy L Thompson // -- Bases 105*2b730f8bSJeremy L Thompson CeedBasisCreateTensorH1Lagrange(ceed, dim, dim, P_hex, Q_hex, CEED_GAUSS, &basis_x_hex); 106*2b730f8bSJeremy L Thompson CeedBasisCreateTensorH1Lagrange(ceed, dim, 1, P_hex, Q_hex, CEED_GAUSS, &basis_u_hex); 1076e15d496SJeremy L Thompson 1086e15d496SJeremy L Thompson // -- Operators 109*2b730f8bSJeremy L Thompson CeedOperatorCreate(ceed, qf_mass, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, &op_mass_hex); 110*2b730f8bSJeremy L Thompson CeedOperatorSetField(op_mass_hex, "u", elem_restr_u_hex, basis_u_hex, CEED_VECTOR_ACTIVE); 111*2b730f8bSJeremy L Thompson CeedOperatorSetField(op_mass_hex, "qdata", elem_restr_qd_i_hex, CEED_BASIS_COLLOCATED, q_data_hex); 112*2b730f8bSJeremy L Thompson CeedOperatorSetField(op_mass_hex, "v", elem_restr_u_hex, basis_u_hex, CEED_VECTOR_ACTIVE); 1136e15d496SJeremy L Thompson 1146e15d496SJeremy L Thompson // Set up Composite Operator 1156e15d496SJeremy L Thompson // -- Create 1166e15d496SJeremy L Thompson CeedCompositeOperatorCreate(ceed, &op_mass); 1176e15d496SJeremy L Thompson // -- Add SubOperators 1186e15d496SJeremy L Thompson CeedCompositeOperatorAddSub(op_mass, op_mass_tet); 1196e15d496SJeremy L Thompson CeedCompositeOperatorAddSub(op_mass, op_mass_hex); 1206e15d496SJeremy L Thompson 1216e15d496SJeremy L Thompson // Estimate FLOPs 1226e15d496SJeremy L Thompson CeedQFunctionSetUserFlopsEstimate(qf_mass, 1); 1236e15d496SJeremy L Thompson CeedOperatorGetFlopsEstimate(op_mass, &flop_estimate); 1246e15d496SJeremy L Thompson 1256e15d496SJeremy L Thompson // Check output 126*2b730f8bSJeremy L Thompson if (flop_estimate != 3042) printf("Incorrect FLOP estimate computed, %ld != 3042\n", flop_estimate); 1276e15d496SJeremy L Thompson 1286e15d496SJeremy L Thompson // Cleanup 1296e15d496SJeremy L Thompson CeedQFunctionDestroy(&qf_mass); 1306e15d496SJeremy L Thompson CeedOperatorDestroy(&op_mass_tet); 1316e15d496SJeremy L Thompson CeedOperatorDestroy(&op_mass_hex); 1326e15d496SJeremy L Thompson CeedOperatorDestroy(&op_mass); 1336e15d496SJeremy L Thompson CeedElemRestrictionDestroy(&elem_restr_u_tet); 1346e15d496SJeremy L Thompson CeedElemRestrictionDestroy(&elem_restr_x_tet); 1356e15d496SJeremy L Thompson CeedElemRestrictionDestroy(&elem_restr_qd_i_tet); 1366e15d496SJeremy L Thompson CeedElemRestrictionDestroy(&elem_restr_u_hex); 1376e15d496SJeremy L Thompson CeedElemRestrictionDestroy(&elem_restr_x_hex); 1386e15d496SJeremy L Thompson CeedElemRestrictionDestroy(&elem_restr_qd_i_hex); 1396e15d496SJeremy L Thompson CeedBasisDestroy(&basis_u_tet); 1406e15d496SJeremy L Thompson CeedBasisDestroy(&basis_x_tet); 1416e15d496SJeremy L Thompson CeedBasisDestroy(&basis_u_hex); 1426e15d496SJeremy L Thompson CeedBasisDestroy(&basis_x_hex); 1436e15d496SJeremy L Thompson CeedVectorDestroy(&q_data_tet); 1446e15d496SJeremy L Thompson CeedVectorDestroy(&q_data_hex); 1456e15d496SJeremy L Thompson CeedDestroy(&ceed); 1466e15d496SJeremy L Thompson return 0; 1476e15d496SJeremy L Thompson } 148