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 tensor product basis 10 #include <ceed/types.h> 11 12 //------------------------------------------------------------------------------ 13 // Tensor Basis Kernels 14 //------------------------------------------------------------------------------ 15 16 //------------------------------------------------------------------------------ 17 // Interp 18 //------------------------------------------------------------------------------ 19 extern "C" __global__ void Interp(const CeedInt num_elem, const CeedInt is_transpose, const CeedScalar *__restrict__ interp_1d, 20 const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 21 const CeedInt i = threadIdx.x; 22 23 __shared__ CeedScalar s_mem[BASIS_Q_1D * BASIS_P_1D + 2 * BASIS_BUF_LEN]; 24 CeedScalar *s_interp_1d = s_mem; 25 CeedScalar *s_buffer_1 = s_mem + BASIS_Q_1D * BASIS_P_1D; 26 CeedScalar *s_buffer_2 = s_buffer_1 + BASIS_BUF_LEN; 27 for (CeedInt k = i; k < BASIS_Q_1D * BASIS_P_1D; k += blockDim.x) { 28 s_interp_1d[k] = interp_1d[k]; 29 } 30 31 const CeedInt P = is_transpose ? BASIS_Q_1D : BASIS_P_1D; 32 const CeedInt Q = is_transpose ? BASIS_P_1D : BASIS_Q_1D; 33 const CeedInt stride_0 = is_transpose ? 1 : BASIS_P_1D; 34 const CeedInt stride_1 = is_transpose ? BASIS_P_1D : 1; 35 const CeedInt u_stride = is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES; 36 const CeedInt v_stride = is_transpose ? BASIS_NUM_NODES : BASIS_NUM_QPTS; 37 const CeedInt u_comp_stride = num_elem * (is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES); 38 const CeedInt v_comp_stride = num_elem * (is_transpose ? BASIS_NUM_NODES : BASIS_NUM_QPTS); 39 const CeedInt u_size = is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES; 40 41 // Apply basis element by element 42 for (CeedInt elem = blockIdx.x; elem < num_elem; elem += gridDim.x) { 43 for (CeedInt comp = 0; comp < BASIS_NUM_COMP; comp++) { 44 const CeedScalar *cur_u = &u[elem * u_stride + comp * u_comp_stride]; 45 CeedScalar *cur_v = &v[elem * v_stride + comp * v_comp_stride]; 46 CeedInt pre = u_size; 47 CeedInt post = 1; 48 49 for (CeedInt d = 0; d < BASIS_DIM; d++) { 50 __syncthreads(); 51 // Update buffers used 52 pre /= P; 53 const CeedScalar *in = d == 0 ? cur_u : (d % 2 ? s_buffer_2 : s_buffer_1); 54 CeedScalar *out = d == BASIS_DIM - 1 ? cur_v : (d % 2 ? s_buffer_1 : s_buffer_2); 55 const CeedInt writeLen = pre * post * Q; 56 57 // Contract along middle index 58 for (CeedInt k = i; k < writeLen; k += blockDim.x) { 59 const CeedInt c = k % post; 60 const CeedInt j = (k / post) % Q; 61 const CeedInt a = k / (post * Q); 62 CeedScalar v_k = 0; 63 64 for (CeedInt b = 0; b < P; b++) v_k += s_interp_1d[j * stride_0 + b * stride_1] * in[(a * P + b) * post + c]; 65 if (is_transpose && d == BASIS_DIM - 1) out[k] += v_k; 66 else out[k] = v_k; 67 } 68 post *= Q; 69 } 70 } 71 } 72 } 73 74 //------------------------------------------------------------------------------ 75 // Grad 76 //------------------------------------------------------------------------------ 77 extern "C" __global__ void Grad(const CeedInt num_elem, const CeedInt is_transpose, const CeedScalar *__restrict__ interp_1d, 78 const CeedScalar *__restrict__ grad_1d, const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 79 const CeedInt i = threadIdx.x; 80 81 __shared__ CeedScalar s_mem[2 * (BASIS_Q_1D * BASIS_P_1D + BASIS_BUF_LEN)]; 82 CeedScalar *s_interp_1d = s_mem; 83 CeedScalar *s_grad_1d = s_interp_1d + BASIS_Q_1D * BASIS_P_1D; 84 CeedScalar *s_buffer_1 = s_grad_1d + BASIS_Q_1D * BASIS_P_1D; 85 CeedScalar *s_buffer_2 = s_buffer_1 + BASIS_BUF_LEN; 86 for (CeedInt k = i; k < BASIS_Q_1D * BASIS_P_1D; k += blockDim.x) { 87 s_interp_1d[k] = interp_1d[k]; 88 s_grad_1d[k] = grad_1d[k]; 89 } 90 91 const CeedInt P = is_transpose ? BASIS_Q_1D : BASIS_P_1D; 92 const CeedInt Q = is_transpose ? BASIS_P_1D : BASIS_Q_1D; 93 const CeedInt stride_0 = is_transpose ? 1 : BASIS_P_1D; 94 const CeedInt stride_1 = is_transpose ? BASIS_P_1D : 1; 95 const CeedInt u_stride = is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES; 96 const CeedInt v_stride = is_transpose ? BASIS_NUM_NODES : BASIS_NUM_QPTS; 97 const CeedInt u_comp_stride = num_elem * (is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES); 98 const CeedInt v_comp_stride = num_elem * (is_transpose ? BASIS_NUM_NODES : BASIS_NUM_QPTS); 99 const CeedInt u_dim_stride = is_transpose ? num_elem * BASIS_NUM_QPTS * BASIS_NUM_COMP : 0; 100 const CeedInt v_dim_stride = is_transpose ? 0 : num_elem * BASIS_NUM_QPTS * BASIS_NUM_COMP; 101 102 // Apply basis element by element 103 for (CeedInt elem = blockIdx.x; elem < num_elem; elem += gridDim.x) { 104 for (CeedInt comp = 0; comp < BASIS_NUM_COMP; comp++) { 105 // dim*dim contractions for grad 106 for (CeedInt dim_1 = 0; dim_1 < BASIS_DIM; dim_1++) { 107 CeedInt pre = is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES; 108 CeedInt post = 1; 109 const CeedScalar *cur_u = &u[elem * u_stride + dim_1 * u_dim_stride + comp * u_comp_stride]; 110 CeedScalar *cur_v = &v[elem * v_stride + dim_1 * v_dim_stride + comp * v_comp_stride]; 111 112 for (CeedInt dim_2 = 0; dim_2 < BASIS_DIM; dim_2++) { 113 __syncthreads(); 114 // Update buffers used 115 pre /= P; 116 const CeedScalar *op = dim_1 == dim_2 ? s_grad_1d : s_interp_1d; 117 const CeedScalar *in = dim_2 == 0 ? cur_u : (dim_2 % 2 ? s_buffer_2 : s_buffer_1); 118 CeedScalar *out = dim_2 == BASIS_DIM - 1 ? cur_v : (dim_2 % 2 ? s_buffer_1 : s_buffer_2); 119 const CeedInt writeLen = pre * post * Q; 120 121 // Contract along middle index 122 for (CeedInt k = i; k < writeLen; k += blockDim.x) { 123 const CeedInt c = k % post; 124 const CeedInt j = (k / post) % Q; 125 const CeedInt a = k / (post * Q); 126 CeedScalar v_k = 0; 127 128 for (CeedInt b = 0; b < P; b++) v_k += op[j * stride_0 + b * stride_1] * in[(a * P + b) * post + c]; 129 if (is_transpose && dim_2 == BASIS_DIM - 1) out[k] += v_k; 130 else out[k] = v_k; 131 } 132 post *= Q; 133 } 134 } 135 } 136 } 137 } 138 139 //------------------------------------------------------------------------------ 140 // 1D quadrature weights 141 //------------------------------------------------------------------------------ 142 __device__ void Weight1d(const CeedInt num_elem, const CeedScalar *q_weight_1d, CeedScalar *w) { 143 const CeedInt i = threadIdx.x; 144 145 if (i < BASIS_Q_1D) { 146 const size_t elem = blockIdx.x; 147 148 if (elem < num_elem) w[elem * BASIS_Q_1D + i] = q_weight_1d[i]; 149 } 150 } 151 152 //------------------------------------------------------------------------------ 153 // 2D quadrature weights 154 //------------------------------------------------------------------------------ 155 __device__ void Weight2d(const CeedInt num_elem, const CeedScalar *q_weight_1d, CeedScalar *w) { 156 const CeedInt i = threadIdx.x; 157 const CeedInt j = threadIdx.y; 158 159 if (i < BASIS_Q_1D && j < BASIS_Q_1D) { 160 const size_t elem = blockIdx.x; 161 162 if (elem < num_elem) { 163 const size_t ind = (elem * BASIS_Q_1D + j) * BASIS_Q_1D + i; 164 165 w[ind] = q_weight_1d[i] * q_weight_1d[j]; 166 } 167 } 168 } 169 170 //------------------------------------------------------------------------------ 171 // 3D quadrature weights 172 //------------------------------------------------------------------------------ 173 __device__ void Weight3d(const CeedInt num_elem, const CeedScalar *q_weight_1d, CeedScalar *w) { 174 const CeedInt i = threadIdx.x; 175 const CeedInt j = threadIdx.y; 176 177 if (i < BASIS_Q_1D && j < BASIS_Q_1D) { 178 const size_t elem = blockIdx.x; 179 180 if (elem < num_elem) { 181 for (CeedInt k = 0; k < BASIS_Q_1D; k++) { 182 const size_t ind = ((elem * BASIS_Q_1D + k) * BASIS_Q_1D + j) * BASIS_Q_1D + i; 183 184 w[ind] = q_weight_1d[i] * q_weight_1d[j] * q_weight_1d[k]; 185 } 186 } 187 } 188 } 189 190 //------------------------------------------------------------------------------ 191 // Quadrature weights 192 //------------------------------------------------------------------------------ 193 extern "C" __global__ void Weight(const CeedInt num_elem, const CeedScalar *__restrict__ q_weight_1d, CeedScalar *__restrict__ v) { 194 if (BASIS_DIM == 1) Weight1d(num_elem, q_weight_1d, v); 195 else if (BASIS_DIM == 2) Weight2d(num_elem, q_weight_1d, v); 196 else if (BASIS_DIM == 3) Weight3d(num_elem, q_weight_1d, v); 197 } 198