1 // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3 // 4 // SPDX-License-Identifier: BSD-2-Clause 5 // 6 // This file is part of CEED: http://github.com/ceed 7 8 #include <ceed.h> 9 10 #if CEEDSIZE 11 typedef CeedSize IndexType; 12 #else 13 typedef CeedInt IndexType; 14 #endif 15 16 //------------------------------------------------------------------------------ 17 // Matrix assembly kernel for low-order elements (2D thread block) 18 //------------------------------------------------------------------------------ 19 extern "C" __launch_bounds__(BLOCK_SIZE) __global__ 20 void linearAssemble(const CeedScalar *B_in, const CeedScalar *B_out, const CeedScalar *__restrict__ qf_array, 21 CeedScalar *__restrict__ values_array) { 22 // This kernel assumes B_in and B_out have the same number of quadrature points and basis points. 23 // TODO: expand to more general cases 24 const int i = threadIdx.x; // The output row index of each B^TDB operation 25 const int l = threadIdx.y; // The output column index of each B^TDB operation 26 // such that we have (Bout^T)_ij D_jk Bin_kl = C_il 27 28 // Strides for final output ordering, determined by the reference (interface) implementation of the symbolic assembly, slowest --> fastest: element, 29 // comp_in, comp_out, node_row, node_col 30 const IndexType comp_out_stride = NNODES * NNODES; 31 const IndexType comp_in_stride = comp_out_stride * NCOMP; 32 const IndexType e_stride = comp_in_stride * NCOMP; 33 // Strides for QF array, slowest --> fastest: emode_in, comp_in, emode_out, comp_out, elem, qpt 34 const IndexType qe_stride = NQPTS; 35 const IndexType qcomp_out_stride = NELEM * qe_stride; 36 const IndexType qemode_out_stride = qcomp_out_stride * NCOMP; 37 const IndexType qcomp_in_stride = qemode_out_stride * NUMEMODEOUT; 38 const IndexType qemode_in_stride = qcomp_in_stride * NCOMP; 39 40 // Loop over each element (if necessary) 41 for (IndexType e = blockIdx.x * blockDim.z + threadIdx.z; e < NELEM; e += gridDim.x * blockDim.z) { 42 for (IndexType comp_in = 0; comp_in < NCOMP; comp_in++) { 43 for (IndexType comp_out = 0; comp_out < NCOMP; comp_out++) { 44 CeedScalar result = 0.0; 45 IndexType qf_index_comp = qcomp_in_stride * comp_in + qcomp_out_stride * comp_out + qe_stride * e; 46 for (IndexType emode_in = 0; emode_in < NUMEMODEIN; emode_in++) { 47 IndexType b_in_index = emode_in * NQPTS * NNODES; 48 for (IndexType emode_out = 0; emode_out < NUMEMODEOUT; emode_out++) { 49 IndexType b_out_index = emode_out * NQPTS * NNODES; 50 IndexType qf_index = qf_index_comp + qemode_out_stride * emode_out + qemode_in_stride * emode_in; 51 // Perform the B^T D B operation for this 'chunk' of D (the qf_array) 52 for (IndexType j = 0; j < NQPTS; j++) { 53 result += B_out[b_out_index + j * NNODES + i] * qf_array[qf_index + j] * B_in[b_in_index + j * NNODES + l]; 54 } 55 } // end of emode_out 56 } // end of emode_in 57 IndexType val_index = comp_in_stride * comp_in + comp_out_stride * comp_out + e_stride * e + NNODES * i + l; 58 values_array[val_index] = result; 59 } // end of out component 60 } // end of in component 61 } // end of element loop 62 } 63 64 //------------------------------------------------------------------------------ 65 // Fallback kernel for larger orders (1D thread block) 66 //------------------------------------------------------------------------------ 67 extern "C" __launch_bounds__(BLOCK_SIZE) __global__ 68 void linearAssembleFallback(const CeedScalar *B_in, const CeedScalar *B_out, const CeedScalar *__restrict__ qf_array, 69 CeedScalar *__restrict__ values_array) { 70 // This kernel assumes B_in and B_out have the same number of quadrature points and basis points. 71 // TODO: expand to more general cases 72 const int l = threadIdx.x; // The output column index of each B^TDB operation 73 // such that we have (Bout^T)_ij D_jk Bin_kl = C_il 74 75 // Strides for final output ordering, determined by the reference (interface) implementation of the symbolic assembly, slowest --> fastest: element, 76 // comp_in, comp_out, node_row, node_col 77 const IndexType comp_out_stride = NNODES * NNODES; 78 const IndexType comp_in_stride = comp_out_stride * NCOMP; 79 const IndexType e_stride = comp_in_stride * NCOMP; 80 // Strides for QF array, slowest --> fastest: emode_in, comp_in, emode_out, comp_out, elem, qpt 81 const IndexType qe_stride = NQPTS; 82 const IndexType qcomp_out_stride = NELEM * qe_stride; 83 const IndexType qemode_out_stride = qcomp_out_stride * NCOMP; 84 const IndexType qcomp_in_stride = qemode_out_stride * NUMEMODEOUT; 85 const IndexType qemode_in_stride = qcomp_in_stride * NCOMP; 86 87 // Loop over each element (if necessary) 88 for (IndexType e = blockIdx.x * blockDim.z + threadIdx.z; e < NELEM; e += gridDim.x * blockDim.z) { 89 for (IndexType comp_in = 0; comp_in < NCOMP; comp_in++) { 90 for (IndexType comp_out = 0; comp_out < NCOMP; comp_out++) { 91 for (IndexType i = 0; i < NNODES; i++) { 92 CeedScalar result = 0.0; 93 IndexType qf_index_comp = qcomp_in_stride * comp_in + qcomp_out_stride * comp_out + qe_stride * e; 94 for (IndexType emode_in = 0; emode_in < NUMEMODEIN; emode_in++) { 95 IndexType b_in_index = emode_in * NQPTS * NNODES; 96 for (IndexType emode_out = 0; emode_out < NUMEMODEOUT; emode_out++) { 97 IndexType b_out_index = emode_out * NQPTS * NNODES; 98 IndexType qf_index = qf_index_comp + qemode_out_stride * emode_out + qemode_in_stride * emode_in; 99 // Perform the B^T D B operation for this 'chunk' of D (the qf_array) 100 for (IndexType j = 0; j < NQPTS; j++) { 101 result += B_out[b_out_index + j * NNODES + i] * qf_array[qf_index + j] * B_in[b_in_index + j * NNODES + l]; 102 } 103 } // end of emode_out 104 } // end of emode_in 105 IndexType val_index = comp_in_stride * comp_in + comp_out_stride * comp_out + e_stride * e + NNODES * i + l; 106 values_array[val_index] = result; 107 } // end of loop over element node index, i 108 } // end of out component 109 } // end of in component 110 } // end of element loop 111 } 112 113 //------------------------------------------------------------------------------ 114