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 curl-oriented element restriction kernels 10 #include <ceed/types.h> 11 12 //------------------------------------------------------------------------------ 13 // L-vector -> E-vector, curl-oriented 14 //------------------------------------------------------------------------------ 15 extern "C" __global__ void CurlOrientedNoTranspose(const CeedInt *__restrict__ indices, const CeedInt8 *__restrict__ curl_orients, 16 const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 17 for (CeedInt node = blockIdx.x * blockDim.x + threadIdx.x; node < RSTR_NUM_ELEM * RSTR_ELEM_SIZE; node += blockDim.x * gridDim.x) { 18 const CeedInt loc_node = node % RSTR_ELEM_SIZE; 19 const CeedInt elem = node / RSTR_ELEM_SIZE; 20 const CeedInt ind_dl = loc_node > 0 ? indices[node - 1] : 0; 21 const CeedInt ind_d = indices[node]; 22 const CeedInt ind_du = loc_node < (RSTR_ELEM_SIZE - 1) ? indices[node + 1] : 0; 23 const CeedInt8 curl_orient_dl = curl_orients[3 * node + 0]; 24 const CeedInt8 curl_orient_d = curl_orients[3 * node + 1]; 25 const CeedInt8 curl_orient_du = curl_orients[3 * node + 2]; 26 27 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) { 28 CeedScalar value = 0.0; 29 value += loc_node > 0 ? u[ind_dl + comp * RSTR_COMP_STRIDE] * curl_orient_dl : 0.0; 30 value += u[ind_d + comp * RSTR_COMP_STRIDE] * curl_orient_d; 31 value += loc_node < (RSTR_ELEM_SIZE - 1) ? u[ind_du + comp * RSTR_COMP_STRIDE] * curl_orient_du : 0.0; 32 v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = value; 33 } 34 } 35 } 36 37 //------------------------------------------------------------------------------ 38 // L-vector -> E-vector, unsigned curl-oriented 39 //------------------------------------------------------------------------------ 40 extern "C" __global__ void CurlOrientedUnsignedNoTranspose(const CeedInt *__restrict__ indices, const CeedInt8 *__restrict__ curl_orients, 41 const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 42 for (CeedInt node = blockIdx.x * blockDim.x + threadIdx.x; node < RSTR_NUM_ELEM * RSTR_ELEM_SIZE; node += blockDim.x * gridDim.x) { 43 const CeedInt loc_node = node % RSTR_ELEM_SIZE; 44 const CeedInt elem = node / RSTR_ELEM_SIZE; 45 const CeedInt ind_dl = loc_node > 0 ? indices[node - 1] : 0; 46 const CeedInt ind_d = indices[node]; 47 const CeedInt ind_du = loc_node < (RSTR_ELEM_SIZE - 1) ? indices[node + 1] : 0; 48 const CeedInt8 curl_orient_dl = abs(curl_orients[3 * node + 0]); 49 const CeedInt8 curl_orient_d = abs(curl_orients[3 * node + 1]); 50 const CeedInt8 curl_orient_du = abs(curl_orients[3 * node + 2]); 51 52 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) { 53 CeedScalar value = 0.0; 54 value += loc_node > 0 ? u[ind_dl + comp * RSTR_COMP_STRIDE] * curl_orient_dl : 0.0; 55 value += u[ind_d + comp * RSTR_COMP_STRIDE] * curl_orient_d; 56 value += loc_node < (RSTR_ELEM_SIZE - 1) ? u[ind_du + comp * RSTR_COMP_STRIDE] * curl_orient_du : 0.0; 57 v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = value; 58 } 59 } 60 } 61 62 //------------------------------------------------------------------------------ 63 // E-vector -> L-vector, curl-oriented 64 //------------------------------------------------------------------------------ 65 #if !USE_DETERMINISTIC 66 extern "C" __global__ void CurlOrientedTranspose(const CeedInt *__restrict__ indices, const CeedInt8 *__restrict__ curl_orients, 67 const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 68 for (CeedInt node = blockIdx.x * blockDim.x + threadIdx.x; node < RSTR_NUM_ELEM * RSTR_ELEM_SIZE; node += blockDim.x * gridDim.x) { 69 const CeedInt ind = indices[node]; 70 const CeedInt loc_node = node % RSTR_ELEM_SIZE; 71 const CeedInt elem = node / RSTR_ELEM_SIZE; 72 const CeedInt8 curl_orient_du = loc_node > 0 ? curl_orients[3 * node - 1] : 0.0; 73 const CeedInt8 curl_orient_d = curl_orients[3 * node + 1]; 74 const CeedInt8 curl_orient_dl = loc_node < (RSTR_ELEM_SIZE - 1) ? curl_orients[3 * node + 3] : 0.0; 75 76 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) { 77 CeedScalar value = 0.0; 78 value += loc_node > 0 ? u[loc_node - 1 + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_du : 0.0; 79 value += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_d; 80 value += 81 loc_node < (RSTR_ELEM_SIZE - 1) ? u[loc_node + 1 + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_dl : 0.0; 82 atomicAdd(&v[ind + comp * RSTR_COMP_STRIDE], value); 83 } 84 } 85 } 86 #else 87 extern "C" __global__ void CurlOrientedTranspose(const CeedInt *__restrict__ l_vec_indices, const CeedInt *__restrict__ t_indices, 88 const CeedInt *__restrict__ t_offsets, const CeedInt8 *__restrict__ curl_orients, 89 const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 90 CeedScalar value[RSTR_NUM_COMP]; 91 92 for (CeedInt i = blockIdx.x * blockDim.x + threadIdx.x; i < RSTR_NUM_NODES; i += blockDim.x * gridDim.x) { 93 const CeedInt ind = l_vec_indices[i]; 94 const CeedInt range_1 = t_offsets[i]; 95 const CeedInt range_N = t_offsets[i + 1]; 96 97 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) value[comp] = 0.0; 98 99 for (CeedInt j = range_1; j < range_N; j++) { 100 const CeedInt t_ind = t_indices[j]; 101 const CeedInt loc_node = t_ind % RSTR_ELEM_SIZE; 102 const CeedInt elem = t_ind / RSTR_ELEM_SIZE; 103 const CeedInt8 curl_orient_du = loc_node > 0 ? curl_orients[3 * t_ind - 1] : 0.0; 104 const CeedInt8 curl_orient_d = curl_orients[3 * t_ind + 1]; 105 const CeedInt8 curl_orient_dl = loc_node < (RSTR_ELEM_SIZE - 1) ? curl_orients[3 * t_ind + 3] : 0.0; 106 107 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) { 108 value[comp] += loc_node > 0 ? u[loc_node - 1 + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_du : 0.0; 109 value[comp] += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_d; 110 value[comp] += 111 loc_node < (RSTR_ELEM_SIZE - 1) ? u[loc_node + 1 + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_dl : 0.0; 112 } 113 } 114 115 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) v[ind + comp * RSTR_COMP_STRIDE] += value[comp]; 116 } 117 } 118 #endif 119 120 //------------------------------------------------------------------------------ 121 // E-vector -> L-vector, unsigned curl-oriented 122 //------------------------------------------------------------------------------ 123 #if !USE_DETERMINISTIC 124 extern "C" __global__ void CurlOrientedUnsignedTranspose(const CeedInt *__restrict__ indices, const CeedInt8 *__restrict__ curl_orients, 125 const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 126 for (CeedInt node = blockIdx.x * blockDim.x + threadIdx.x; node < RSTR_NUM_ELEM * RSTR_ELEM_SIZE; node += blockDim.x * gridDim.x) { 127 const CeedInt loc_node = node % RSTR_ELEM_SIZE; 128 const CeedInt elem = node / RSTR_ELEM_SIZE; 129 const CeedInt ind = indices[node]; 130 const CeedInt8 curl_orient_du = loc_node > 0 ? abs(curl_orients[3 * node - 1]) : 0.0; 131 const CeedInt8 curl_orient_d = abs(curl_orients[3 * node + 1]); 132 const CeedInt8 curl_orient_dl = loc_node < (RSTR_ELEM_SIZE - 1) ? abs(curl_orients[3 * node + 3]) : 0.0; 133 134 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) { 135 CeedScalar value = 0.0; 136 value += loc_node > 0 ? u[loc_node - 1 + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_du : 0.0; 137 value += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_d; 138 value += 139 loc_node < (RSTR_ELEM_SIZE - 1) ? u[loc_node + 1 + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_dl : 0.0; 140 atomicAdd(&v[ind + comp * RSTR_COMP_STRIDE], value); 141 } 142 } 143 } 144 #else 145 extern "C" __global__ void CurlOrientedUnsignedTranspose(const CeedInt *__restrict__ l_vec_indices, const CeedInt *__restrict__ t_indices, 146 const CeedInt *__restrict__ t_offsets, const CeedInt8 *__restrict__ curl_orients, 147 const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 148 CeedScalar value[RSTR_NUM_COMP]; 149 150 for (CeedInt i = blockIdx.x * blockDim.x + threadIdx.x; i < RSTR_NUM_NODES; i += blockDim.x * gridDim.x) { 151 const CeedInt ind = l_vec_indices[i]; 152 const CeedInt range_1 = t_offsets[i]; 153 const CeedInt range_N = t_offsets[i + 1]; 154 155 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) value[comp] = 0.0; 156 157 for (CeedInt j = range_1; j < range_N; j++) { 158 const CeedInt t_ind = t_indices[j]; 159 const CeedInt loc_node = t_ind % RSTR_ELEM_SIZE; 160 const CeedInt elem = t_ind / RSTR_ELEM_SIZE; 161 const CeedInt8 curl_orient_du = loc_node > 0 ? abs(curl_orients[3 * t_ind - 1]) : 0.0; 162 const CeedInt8 curl_orient_d = abs(curl_orients[3 * t_ind + 1]); 163 const CeedInt8 curl_orient_dl = loc_node < (RSTR_ELEM_SIZE - 1) ? abs(curl_orients[3 * t_ind + 3]) : 0.0; 164 165 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) { 166 value[comp] += loc_node > 0 ? u[loc_node - 1 + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_du : 0.0; 167 value[comp] += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_d; 168 value[comp] += 169 loc_node < (RSTR_ELEM_SIZE - 1) ? u[loc_node + 1 + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * curl_orient_dl : 0.0; 170 } 171 } 172 173 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) v[ind + comp * RSTR_COMP_STRIDE] += value[comp]; 174 } 175 } 176 #endif 177