1 // Copyright (c) 2017-2026, 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 /// @file 9 /// Internal header for CUDA operator full assembly 10 #include <ceed/types.h> 11 12 #if USE_CEEDSIZE 13 typedef CeedSize IndexType; 14 #else 15 typedef CeedInt IndexType; 16 #endif 17 18 //------------------------------------------------------------------------------ 19 // Matrix assembly kernel 20 //------------------------------------------------------------------------------ 21 extern "C" __launch_bounds__(BLOCK_SIZE) __global__ 22 void LinearAssemble(const CeedInt num_elem, const CeedScalar *B_in, const CeedScalar *B_out, const bool *orients_in, 23 const CeedInt8 *curl_orients_in, const bool *orients_out, const CeedInt8 *curl_orients_out, 24 const CeedScalar *__restrict__ qf_array, CeedScalar *__restrict__ values_array) { 25 extern __shared__ CeedScalar s_CT[]; 26 CeedScalar *s_C = &s_CT[NUM_NODES_OUT * NUM_NODES_IN]; 27 28 const int l = threadIdx.x; // The output column index of each B^T D B operation 29 // such that we have (Bout^T)_ij D_jk Bin_kl = C_il 30 31 // Strides for final output ordering, determined by the reference (interface) implementation of the symbolic assembly, slowest --> fastest: e, 32 // comp_in, comp_out, node_row, node_col 33 const IndexType comp_out_stride = NUM_NODES_OUT * NUM_NODES_IN; 34 const IndexType comp_in_stride = comp_out_stride * NUM_COMP_OUT; 35 const IndexType e_stride = comp_in_stride * NUM_COMP_IN; 36 37 // Strides for QF array, slowest --> fastest: e_in, comp_in, e_out, comp_out, e, q 38 const IndexType q_e_stride = NUM_QPTS; 39 const IndexType q_comp_out_stride = num_elem * q_e_stride; 40 const IndexType q_eval_mode_out_stride = q_comp_out_stride * NUM_COMP_OUT; 41 const IndexType q_comp_in_stride = q_eval_mode_out_stride * NUM_EVAL_MODES_OUT; 42 const IndexType q_eval_mode_in_stride = q_comp_in_stride * NUM_COMP_IN; 43 44 // Loop over each element (if necessary) 45 for (IndexType e = blockIdx.x * blockDim.z + threadIdx.z; e < num_elem; e += gridDim.x * blockDim.z) { 46 for (IndexType comp_in = 0; comp_in < NUM_COMP_IN; comp_in++) { 47 for (IndexType comp_out = 0; comp_out < NUM_COMP_OUT; comp_out++) { 48 for (IndexType i = threadIdx.y; i < NUM_NODES_OUT; i += BLOCK_SIZE_Y) { 49 CeedScalar result = 0.0; 50 IndexType qf_index_comp = q_comp_in_stride * comp_in + q_comp_out_stride * comp_out + q_e_stride * e; 51 52 for (IndexType e_in = 0; e_in < NUM_EVAL_MODES_IN; e_in++) { 53 IndexType b_in_index = e_in * NUM_QPTS * NUM_NODES_IN; 54 55 for (IndexType e_out = 0; e_out < NUM_EVAL_MODES_OUT; e_out++) { 56 IndexType b_out_index = e_out * NUM_QPTS * NUM_NODES_OUT; 57 IndexType qf_index = qf_index_comp + q_eval_mode_out_stride * e_out + q_eval_mode_in_stride * e_in; 58 59 // Perform the B^T D B operation for this 'chunk' of D (the qf_array) 60 for (IndexType j = 0; j < NUM_QPTS; j++) { 61 result += B_out[b_out_index + j * NUM_NODES_OUT + i] * qf_array[qf_index + j] * B_in[b_in_index + j * NUM_NODES_IN + l]; 62 } 63 } // end of out eval mode 64 } // end of in eval mode 65 if (orients_in) { 66 result *= orients_in[NUM_NODES_IN * e + l] ? -1.0 : 1.0; 67 } 68 if (orients_out) { 69 result *= orients_out[NUM_NODES_OUT * e + i] ? -1.0 : 1.0; 70 } 71 if (!curl_orients_in && !curl_orients_out) { 72 IndexType val_index = e_stride * e + comp_in_stride * comp_in + comp_out_stride * comp_out + NUM_NODES_IN * i + l; 73 74 values_array[val_index] = result; 75 } else if (curl_orients_in) { 76 s_C[NUM_NODES_IN * threadIdx.y + l] = result; 77 __syncthreads(); 78 s_CT[NUM_NODES_IN * i + l] = 79 (l > 0 ? s_C[NUM_NODES_IN * threadIdx.y + l - 1] * curl_orients_in[3 * NUM_NODES_IN * e + 3 * l - 1] : 0.0) + 80 s_C[NUM_NODES_IN * threadIdx.y + l] * curl_orients_in[3 * NUM_NODES_IN * e + 3 * l + 1] + 81 (l < (NUM_NODES_IN - 1) ? s_C[NUM_NODES_IN * threadIdx.y + l + 1] * curl_orients_in[3 * NUM_NODES_IN * e + 3 * l + 3] : 0.0); 82 } else { 83 s_CT[NUM_NODES_IN * i + l] = result; 84 } 85 } // end of loop over element node index, i 86 if (curl_orients_in || curl_orients_out) { 87 // Compute and store the final T^T (B^T D B T) using the fully computed C T product in shared memory 88 if (curl_orients_out) __syncthreads(); 89 for (IndexType i = threadIdx.y; i < NUM_NODES_OUT; i += BLOCK_SIZE_Y) { 90 IndexType val_index = e_stride * e + comp_in_stride * comp_in + comp_out_stride * comp_out + NUM_NODES_IN * i + l; 91 92 if (curl_orients_out) { 93 values_array[val_index] = 94 (i > 0 ? s_CT[NUM_NODES_IN * (i - 1) + l] * curl_orients_out[3 * NUM_NODES_OUT * e + 3 * i - 1] : 0.0) + 95 s_CT[NUM_NODES_IN * i + l] * curl_orients_out[3 * NUM_NODES_OUT * e + 3 * i + 1] + 96 (i < (NUM_NODES_OUT - 1) ? s_CT[NUM_NODES_IN * (i + 1) + l] * curl_orients_out[3 * NUM_NODES_OUT * e + 3 * i + 3] : 0.0); 97 } else { 98 values_array[val_index] = s_CT[NUM_NODES_IN * i + l]; 99 } 100 } 101 } 102 } // end of out component 103 } // end of in component 104 } // end of element loop 105 } 106