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 HIP shared memory tensor product basis 10 #include <ceed/types.h> 11 12 #include "hip-shared-basis-read-write-templates.h" 13 #include "hip-shared-basis-tensor-templates.h" 14 15 //------------------------------------------------------------------------------ 16 // Interp kernel by dim 17 //------------------------------------------------------------------------------ 18 extern "C" __launch_bounds__(BASIS_INTERP_BLOCK_SIZE) __global__ 19 void Interp(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *__restrict__ d_U, CeedScalar *__restrict__ d_V) { 20 extern __shared__ CeedScalar slice[]; 21 22 SharedData_Hip data; 23 data.t_id_x = threadIdx.x; 24 data.t_id_y = threadIdx.y; 25 data.t_id_z = threadIdx.z; 26 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 27 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 28 29 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 30 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 31 32 // load interp_1d into shared memory 33 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 34 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 35 __syncthreads(); 36 37 // Apply basis element by element 38 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 39 if (BASIS_DIM == 1) { 40 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, d_U, r_U); 41 Interp1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 42 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, r_V, d_V); 43 } else if (BASIS_DIM == 2) { 44 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); 45 InterpTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 46 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); 47 } else if (BASIS_DIM == 3) { 48 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 49 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, d_U, r_U); 50 InterpTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 51 WriteElementStrided3d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 52 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, r_V, d_V); 53 } 54 } 55 } 56 57 extern "C" __launch_bounds__(BASIS_INTERP_BLOCK_SIZE) __global__ 58 void InterpTranspose(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *__restrict__ d_U, CeedScalar *__restrict__ d_V) { 59 extern __shared__ CeedScalar slice[]; 60 61 SharedData_Hip data; 62 data.t_id_x = threadIdx.x; 63 data.t_id_y = threadIdx.y; 64 data.t_id_z = threadIdx.z; 65 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 66 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 67 68 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 69 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 70 71 // load interp_1d into shared memory 72 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 73 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 74 __syncthreads(); 75 76 // Apply basis element by element 77 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 78 if (BASIS_DIM == 1) { 79 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 80 InterpTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 81 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 82 } else if (BASIS_DIM == 2) { 83 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); 84 InterpTransposeTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 85 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); 86 } else if (BASIS_DIM == 3) { 87 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 88 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 89 InterpTransposeTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 90 WriteElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 91 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 92 } 93 } 94 } 95 96 extern "C" __launch_bounds__(BASIS_INTERP_BLOCK_SIZE) __global__ 97 void InterpTransposeAdd(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *__restrict__ d_U, CeedScalar *__restrict__ d_V) { 98 extern __shared__ CeedScalar slice[]; 99 100 SharedData_Hip data; 101 data.t_id_x = threadIdx.x; 102 data.t_id_y = threadIdx.y; 103 data.t_id_z = threadIdx.z; 104 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 105 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 106 107 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 108 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 109 110 // load interp_1d into shared memory 111 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 112 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 113 __syncthreads(); 114 115 // Apply basis element by element 116 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 117 if (BASIS_DIM == 1) { 118 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 119 InterpTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 120 SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 121 } else if (BASIS_DIM == 2) { 122 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); 123 InterpTransposeTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 124 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); 125 } else if (BASIS_DIM == 3) { 126 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 127 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 128 InterpTransposeTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_V); 129 SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 130 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 131 } 132 } 133 } 134 135 //------------------------------------------------------------------------------ 136 // Grad kernel by dim 137 //------------------------------------------------------------------------------ 138 extern "C" __launch_bounds__(BASIS_GRAD_BLOCK_SIZE) __global__ void Grad(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, 139 const CeedScalar *__restrict__ d_U, CeedScalar *__restrict__ d_V) { 140 extern __shared__ CeedScalar slice[]; 141 142 SharedData_Hip data; 143 data.t_id_x = threadIdx.x; 144 data.t_id_y = threadIdx.y; 145 data.t_id_z = threadIdx.z; 146 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 147 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 148 149 CeedScalar r_U[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 150 CeedScalar r_V[BASIS_NUM_COMP * BASIS_DIM * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 151 152 // load interp_1d and grad_1d into shared memory 153 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 154 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 155 __shared__ CeedScalar s_G[BASIS_Q_1D * (BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D)]; 156 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); 157 __syncthreads(); 158 159 // Apply basis element by element 160 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 161 if (BASIS_DIM == 1) { 162 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, d_U, r_U); 163 Grad1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 164 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, r_V, d_V); 165 } else if (BASIS_DIM == 2) { 166 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); 167 GradTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 168 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, 169 d_V); 170 } else if (BASIS_DIM == 3) { 171 ReadElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 172 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, d_U, r_U); 173 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); 174 else GradTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 175 WriteElementStrided3d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 176 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, r_V, d_V); 177 } 178 } 179 } 180 181 extern "C" __launch_bounds__(BASIS_GRAD_BLOCK_SIZE) __global__ 182 void GradTranspose(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const CeedScalar *__restrict__ d_U, 183 CeedScalar *__restrict__ d_V) { 184 extern __shared__ CeedScalar slice[]; 185 186 SharedData_Hip data; 187 data.t_id_x = threadIdx.x; 188 data.t_id_y = threadIdx.y; 189 data.t_id_z = threadIdx.z; 190 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 191 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 192 193 CeedScalar r_U[BASIS_NUM_COMP * BASIS_DIM * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 194 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 195 196 // load interp_1d and grad_1d into shared memory 197 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 198 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 199 __shared__ CeedScalar s_G[BASIS_Q_1D * (BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D)]; 200 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); 201 __syncthreads(); 202 203 // Apply basis element by element 204 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 205 if (BASIS_DIM == 1) { 206 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 207 GradTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 208 WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 209 } else if (BASIS_DIM == 2) { 210 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, 211 r_U); 212 GradTransposeTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 213 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); 214 } else if (BASIS_DIM == 3) { 215 ReadElementStrided3d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 216 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 217 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); 218 else GradTransposeTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 219 WriteElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 220 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 221 } 222 } 223 } 224 225 extern "C" __launch_bounds__(BASIS_GRAD_BLOCK_SIZE) __global__ 226 void GradTransposeAdd(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const CeedScalar *__restrict__ d_U, 227 CeedScalar *__restrict__ d_V) { 228 extern __shared__ CeedScalar slice[]; 229 230 SharedData_Hip data; 231 data.t_id_x = threadIdx.x; 232 data.t_id_y = threadIdx.y; 233 data.t_id_z = threadIdx.z; 234 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 235 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 236 237 CeedScalar r_U[BASIS_NUM_COMP * BASIS_DIM * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; 238 CeedScalar r_V[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_P_1D : 1)]; 239 240 // load interp_1d and grad_1d into shared memory 241 __shared__ CeedScalar s_B[BASIS_P_1D * BASIS_Q_1D]; 242 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); 243 __shared__ CeedScalar s_G[BASIS_Q_1D * (BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D)]; 244 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); 245 __syncthreads(); 246 247 // Apply basis element by element 248 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 249 if (BASIS_DIM == 1) { 250 ReadElementStrided1d<BASIS_NUM_COMP, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, d_U, r_U); 251 GradTranspose1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 252 SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, r_V, d_V); 253 } else if (BASIS_DIM == 2) { 254 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, 255 r_U); 256 GradTransposeTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 257 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); 258 } else if (BASIS_DIM == 3) { 259 ReadElementStrided3d<BASIS_NUM_COMP * BASIS_DIM, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D * num_elem, 260 BASIS_Q_1D * BASIS_Q_1D * BASIS_Q_1D, d_U, r_U); 261 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); 262 else GradTransposeTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, s_G, r_V); 263 SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1D * num_elem, 264 BASIS_P_1D * BASIS_P_1D * BASIS_P_1D, r_V, d_V); 265 } 266 } 267 } 268 269 //------------------------------------------------------------------------------ 270 // Weight kernels by dim 271 //------------------------------------------------------------------------------ 272 extern "C" __launch_bounds__(BASIS_WEIGHT_BLOCK_SIZE) __global__ 273 void Weight(const CeedInt num_elem, const CeedScalar *__restrict__ q_weight_1d, CeedScalar *__restrict__ d_W) { 274 extern __shared__ CeedScalar slice[]; 275 276 SharedData_Hip data; 277 data.t_id_x = threadIdx.x; 278 data.t_id_y = threadIdx.y; 279 data.t_id_z = threadIdx.z; 280 data.t_id = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x; 281 data.slice = slice + data.t_id_z * BASIS_T_1D * (BASIS_DIM > 1 ? BASIS_T_1D : 1); 282 283 CeedScalar r_W[BASIS_DIM > 2 ? BASIS_Q_1D : 1]; 284 285 for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * blockDim.z) { 286 if (BASIS_DIM == 1) { 287 Weight1d<BASIS_Q_1D>(data, q_weight_1d, r_W); 288 WriteElementStrided1d<1, BASIS_Q_1D>(data, elem, 1, BASIS_Q_1D * num_elem, BASIS_Q_1D, r_W, d_W); 289 } else if (BASIS_DIM == 2) { 290 WeightTensor2d<BASIS_Q_1D>(data, q_weight_1d, r_W); 291 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); 292 } else if (BASIS_DIM == 3) { 293 WeightTensor3d<BASIS_Q_1D>(data, q_weight_1d, r_W); 294 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, 295 d_W); 296 } 297 } 298 } 299