1 // Copyright (c) 2017-2025, 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 shared memory tensor product basis 10 #include <ceed/types.h> 11 12 #include "cuda-shared-basis-read-write-templates.h" 13 #include "cuda-shared-basis-tensor-templates.h" 14 15 //------------------------------------------------------------------------------ 16 // Interp kernels by dim 17 //------------------------------------------------------------------------------ 18 extern "C" __global__ void Interp(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *__restrict__ d_U, CeedScalar *__restrict__ d_V) { 19 extern __shared__ CeedScalar slice[]; 20 21 SharedData_Cuda data; 22 data.t_id_x = threadIdx.x; 23 data.t_id_y = threadIdx.y; 24 data.t_id_z = threadIdx.z; 25 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 26 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 27 28 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 29 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 30 31 // load interp_1d into shared memory 32 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 33 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 34 __syncthreads(); 35 36 // Apply basis element by element 37 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 38 if (BASIS_DIM == 1) { 39 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, d_U, r_U); 40 Interp1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 41 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, r_V, d_V); 42 } else if (BASIS_DIM == 2) { 43 ReadElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, d_U, r_U); 44 InterpTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 45 WriteElementStrided2d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, r_V, d_V); 46 } else if (BASIS_DIM == 3) { 47 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 48 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, d_U, r_U); 49 InterpTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 50 WriteElementStrided3d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 51 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, r_V, d_V); 52 } 53 } 54 } 55 56 extern "C" __global__ void InterpCollocated(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *__restrict__ d_U, 57 CeedScalar *__restrict__ d_V) { 58 extern __shared__ CeedScalar slice[]; 59 60 SharedData_Cuda data; 61 data.t_id_x = threadIdx.x; 62 data.t_id_y = threadIdx.y; 63 data.t_id_z = threadIdx.z; 64 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 65 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 66 67 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 68 69 // Apply basis element by element 70 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 71 if (BASIS_DIM == 1) { 72 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, d_U, r_U); 73 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, r_U, d_V); 74 } else if (BASIS_DIM == 2) { 75 ReadElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, d_U, r_U); 76 WriteElementStrided2d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, r_U, d_V); 77 } else if (BASIS_DIM == 3) { 78 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 79 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, d_U, r_U); 80 WriteElementStrided3d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 81 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, r_U, d_V); 82 } 83 } 84 } 85 86 extern "C" __global__ void InterpTranspose(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *__restrict__ d_U, 87 CeedScalar *__restrict__ d_V) { 88 extern __shared__ CeedScalar slice[]; 89 90 SharedData_Cuda data; 91 data.t_id_x = threadIdx.x; 92 data.t_id_y = threadIdx.y; 93 data.t_id_z = threadIdx.z; 94 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 95 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 96 97 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 98 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 99 100 // load interp_1d into shared memory 101 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 102 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 103 __syncthreads(); 104 105 // Apply basis element by element 106 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 107 if (BASIS_DIM == 1) { 108 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 109 InterpTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 110 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 111 } else if (BASIS_DIM == 2) { 112 ReadElementStrided2d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 113 InterpTransposeTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 114 WriteElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, r_V, d_V); 115 } else if (BASIS_DIM == 3) { 116 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 117 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 118 InterpTransposeTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 119 WriteElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 120 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 121 } 122 } 123 } 124 125 extern "C" __global__ void InterpCollocatedTranspose(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *__restrict__ d_U, 126 CeedScalar *__restrict__ d_V) { 127 extern __shared__ CeedScalar slice[]; 128 129 SharedData_Cuda data; 130 data.t_id_x = threadIdx.x; 131 data.t_id_y = threadIdx.y; 132 data.t_id_z = threadIdx.z; 133 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 134 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 135 136 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 137 138 // Apply basis element by element 139 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 140 if (BASIS_DIM == 1) { 141 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 142 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_U, d_V); 143 } else if (BASIS_DIM == 2) { 144 ReadElementStrided2d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 145 WriteElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, r_U, d_V); 146 } else if (BASIS_DIM == 3) { 147 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 148 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 149 WriteElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 150 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_U, d_V); 151 } 152 } 153 } 154 155 extern "C" __global__ void InterpTransposeAdd(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *__restrict__ d_U, 156 CeedScalar *__restrict__ d_V) { 157 extern __shared__ CeedScalar slice[]; 158 159 SharedData_Cuda data; 160 data.t_id_x = threadIdx.x; 161 data.t_id_y = threadIdx.y; 162 data.t_id_z = threadIdx.z; 163 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 164 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 165 166 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 167 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 168 169 // load interp_1d into shared memory 170 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 171 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 172 __syncthreads(); 173 174 // Apply basis element by element 175 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 176 if (BASIS_DIM == 1) { 177 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 178 InterpTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 179 SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 180 } else if (BASIS_DIM == 2) { 181 ReadElementStrided2d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 182 InterpTransposeTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 183 SumElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, r_V, d_V); 184 } else if (BASIS_DIM == 3) { 185 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 186 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 187 InterpTransposeTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 188 SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 189 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 190 } 191 } 192 } 193 194 extern "C" __global__ void InterpCollocatedTransposeAdd(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *__restrict__ d_U, 195 CeedScalar *__restrict__ d_V) { 196 extern __shared__ CeedScalar slice[]; 197 198 SharedData_Cuda data; 199 data.t_id_x = threadIdx.x; 200 data.t_id_y = threadIdx.y; 201 data.t_id_z = threadIdx.z; 202 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 203 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 204 205 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 206 207 // Apply basis element by element 208 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 209 if (BASIS_DIM == 1) { 210 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 211 SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_U, d_V); 212 } else if (BASIS_DIM == 2) { 213 ReadElementStrided2d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 214 SumElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, r_U, d_V); 215 } else if (BASIS_DIM == 3) { 216 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 217 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 218 SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 219 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_U, d_V); 220 } 221 } 222 } 223 224 //------------------------------------------------------------------------------ 225 // Grad kernel by dim 226 //------------------------------------------------------------------------------ 227 extern "C" __global__ void Grad(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const CeedScalar *__restrict__ d_U, 228 CeedScalar *__restrict__ d_V) { 229 extern __shared__ CeedScalar slice[]; 230 231 SharedData_Cuda data; 232 data.t_id_x = threadIdx.x; 233 data.t_id_y = threadIdx.y; 234 data.t_id_z = threadIdx.z; 235 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 236 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 237 238 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 239 CeedScalar r_V[BASIS_NUM_COMP * BASIS_DIM * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 240 241 // load interp_1d and grad_1d into shared memory 242 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 243 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 244 __shared__ CeedScalar s_G[BASIS_Q_1D * (BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D)]; 245 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); 246 __syncthreads(); 247 248 // Apply basis element by element 249 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 250 if (BASIS_DIM == 1) { 251 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, d_U, r_U); 252 Grad1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 253 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, r_V, d_V); 254 } else if (BASIS_DIM == 2) { 255 ReadElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, d_U, r_U); 256 GradTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 257 WriteElementStrided2d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, r_V, 258 d_V); 259 } else if (BASIS_DIM == 3) { 260 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 261 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, d_U, r_U); 262 if (BASIS_HAS_COLLOCATED_GRAD) GradTensorCollocated3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 263 else GradTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 264 WriteElementStrided3d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 265 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, r_V, d_V); 266 } 267 } 268 } 269 270 extern "C" __global__ void GradCollocated(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const CeedScalar *__restrict__ d_U, 271 CeedScalar *__restrict__ d_V) { 272 extern __shared__ CeedScalar slice[]; 273 274 SharedData_Cuda data; 275 data.t_id_x = threadIdx.x; 276 data.t_id_y = threadIdx.y; 277 data.t_id_z = threadIdx.z; 278 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 279 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 280 281 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 282 CeedScalar r_V[BASIS_NUM_COMP * BASIS_DIM * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 283 284 // load grad_1d into shared memory 285 __shared__ CeedScalar s_G[BASIS_Q_1D * (BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D)]; 286 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); 287 __syncthreads(); 288 289 // Apply basis element by element 290 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 291 if (BASIS_DIM == 1) { 292 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, d_U, r_U); 293 Grad1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, NULL, s_G, r_V); 294 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, r_V, d_V); 295 } else if (BASIS_DIM == 2) { 296 ReadElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, d_U, r_U); 297 GradTensorCollocatedNodes2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_G, r_V); 298 WriteElementStrided2d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, r_V, 299 d_V); 300 } else if (BASIS_DIM == 3) { 301 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 302 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, d_U, r_U); 303 GradTensorCollocatedNodes3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_G, r_V); 304 WriteElementStrided3d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 305 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, r_V, d_V); 306 } 307 } 308 } 309 310 extern "C" __global__ void GradTranspose(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const CeedScalar *__restrict__ d_U, 311 CeedScalar *__restrict__ d_V) { 312 extern __shared__ CeedScalar slice[]; 313 314 SharedData_Cuda data; 315 data.t_id_x = threadIdx.x; 316 data.t_id_y = threadIdx.y; 317 data.t_id_z = threadIdx.z; 318 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 319 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 320 321 CeedScalar r_U[BASIS_NUM_COMP * BASIS_DIM * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 322 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 323 324 // load interp_1d and grad_1d into shared memory 325 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 326 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 327 __shared__ CeedScalar s_G[BASIS_Q_1D * (BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D)]; 328 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); 329 __syncthreads(); 330 331 // Apply basis element by element 332 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 333 if (BASIS_DIM == 1) { 334 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 335 GradTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 336 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 337 } else if (BASIS_DIM == 2) { 338 ReadElementStrided2d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, d_U, 339 r_U); 340 GradTransposeTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 341 WriteElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, r_V, d_V); 342 } else if (BASIS_DIM == 3) { 343 ReadElementStrided3d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 344 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 345 if (BASIS_HAS_COLLOCATED_GRAD) GradTransposeTensorCollocated3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 346 else GradTransposeTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 347 WriteElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 348 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 349 } 350 } 351 } 352 353 extern "C" __global__ void GradCollocatedTranspose(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, 354 const CeedScalar *__restrict__ d_U, CeedScalar *__restrict__ d_V) { 355 extern __shared__ CeedScalar slice[]; 356 357 SharedData_Cuda data; 358 data.t_id_x = threadIdx.x; 359 data.t_id_y = threadIdx.y; 360 data.t_id_z = threadIdx.z; 361 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 362 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 363 364 CeedScalar r_U[BASIS_NUM_COMP * BASIS_DIM * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 365 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 366 367 // load grad_1d into shared memory 368 __shared__ CeedScalar s_G[BASIS_Q_1D * (BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D)]; 369 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); 370 __syncthreads(); 371 372 // Apply basis element by element 373 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 374 if (BASIS_DIM == 1) { 375 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 376 GradTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, NULL, s_G, r_V); 377 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 378 } else if (BASIS_DIM == 2) { 379 ReadElementStrided2d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, d_U, 380 r_U); 381 GradTransposeTensorCollocatedNodes2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_G, r_V); 382 WriteElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, r_V, d_V); 383 } else if (BASIS_DIM == 3) { 384 ReadElementStrided3d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 385 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 386 GradTransposeTensorCollocatedNodes3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_G, r_V); 387 WriteElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 388 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 389 } 390 } 391 } 392 393 extern "C" __global__ void GradTransposeAdd(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const CeedScalar *__restrict__ d_U, 394 CeedScalar *__restrict__ d_V) { 395 extern __shared__ CeedScalar slice[]; 396 397 SharedData_Cuda data; 398 data.t_id_x = threadIdx.x; 399 data.t_id_y = threadIdx.y; 400 data.t_id_z = threadIdx.z; 401 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 402 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 403 404 CeedScalar r_U[BASIS_NUM_COMP * BASIS_DIM * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 405 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 406 407 // load interp_1d and grad_1d into shared memory 408 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 409 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 410 __shared__ CeedScalar s_G[BASIS_Q_1D * (BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D)]; 411 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); 412 __syncthreads(); 413 414 // Apply basis element by element 415 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 416 if (BASIS_DIM == 1) { 417 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 418 GradTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 419 SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 420 } else if (BASIS_DIM == 2) { 421 ReadElementStrided2d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, d_U, 422 r_U); 423 GradTransposeTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 424 SumElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, r_V, d_V); 425 } else if (BASIS_DIM == 3) { 426 ReadElementStrided3d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 427 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 428 if (BASIS_HAS_COLLOCATED_GRAD) GradTransposeTensorCollocated3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 429 else GradTransposeTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 430 SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 431 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 432 } 433 } 434 } 435 436 extern "C" __global__ void GradCollocatedTransposeAdd(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, 437 const CeedScalar *__restrict__ d_U, CeedScalar *__restrict__ d_V) { 438 extern __shared__ CeedScalar slice[]; 439 440 SharedData_Cuda data; 441 data.t_id_x = threadIdx.x; 442 data.t_id_y = threadIdx.y; 443 data.t_id_z = threadIdx.z; 444 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 445 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 446 447 CeedScalar r_U[BASIS_NUM_COMP * BASIS_DIM * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 448 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 449 450 // load grad_1d into shared memory 451 __shared__ CeedScalar s_G[BASIS_Q_1D * (BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D)]; 452 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); 453 __syncthreads(); 454 455 // Apply basis element by element 456 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 457 if (BASIS_DIM == 1) { 458 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 459 GradTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, NULL, s_G, r_V); 460 SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 461 } else if (BASIS_DIM == 2) { 462 ReadElementStrided2d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, d_U, 463 r_U); 464 GradTransposeTensorCollocatedNodes2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_G, r_V); 465 SumElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem, BASIS_P_1D * BASIS_P_1D, r_V, d_V); 466 } else if (BASIS_DIM == 3) { 467 ReadElementStrided3d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 468 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 469 GradTransposeTensorCollocatedNodes3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_G, r_V); 470 SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 471 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 472 } 473 } 474 } 475 476 //------------------------------------------------------------------------------ 477 // Weight kernels by dim 478 //------------------------------------------------------------------------------ 479 extern "C" __global__ void Weight(const CeedInt num_elem, const CeedScalar *__restrict__ q_weight_1d, CeedScalar *__restrict__ d_W) { 480 extern __shared__ CeedScalar slice[]; 481 482 SharedData_Cuda data; 483 data.t_id_x = threadIdx.x; 484 data.t_id_y = threadIdx.y; 485 data.t_id_z = threadIdx.z; 486 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 487 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 488 489 CeedScalar r_W[BASIS_DIM > 2 ? BASIS_Q_1D : 1]; 490 491 // Apply basis element by element 492 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 493 if (BASIS_DIM == 1) { 494 Weight1d<BASIS_P_1D, BASIS_Q_1D>(data, q_weight_1d, r_W); 495 WriteElementStrided1d<1, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, r_W, d_W); 496 } else if (BASIS_DIM == 2) { 497 WeightTensor2d<BASIS_P_1D, BASIS_Q_1D>(data, q_weight_1d, r_W); 498 WriteElementStrided2d<1, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D, r_W, d_W); 499 } else if (BASIS_DIM == 3) { 500 WeightTensor3d<BASIS_P_1D, BASIS_Q_1D>(data, q_weight_1d, r_W); 501 WriteElementStrided3d<1, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, r_W, 502 d_W); 503 } 504 } 505 } 506