1 // Copyright (c) 2017-2024, 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 non-tensor product basis templates 10 11 #include <ceed/types.h> 12 13 //------------------------------------------------------------------------------ 14 // Tensor contraction 15 //------------------------------------------------------------------------------ 16 template <int NUM_COMP, int Q_COMP, int P, int Q> 17 inline __device__ void Contract(const CeedInt elem, const CeedInt strides_elem_U, const CeedInt strides_elem_V, const CeedInt strides_comp_U, 18 const CeedInt strides_comp_V, const CeedInt strides_q_comp_V, const CeedScalar *__restrict__ d_B, 19 const CeedScalar *__restrict__ d_U, CeedScalar *__restrict__ d_V) { 20 const CeedInt t_id = threadIdx.x; 21 const CeedScalar *U; 22 CeedScalar r_V[Q_COMP]; 23 // TODO load B in shared memory if blockDim.z > 1? 24 25 for (CeedInt comp = 0; comp < NUM_COMP; comp++) { 26 // Run with Q threads 27 U = d_U + elem * strides_elem_U + comp * strides_comp_U; 28 for (CeedInt d = 0; d < Q_COMP; d++) r_V[d] = 0.0; 29 for (CeedInt i = 0; i < P; i++) { 30 const CeedScalar val = U[i]; 31 32 for (CeedInt d = 0; d < Q_COMP; d++) r_V[d] += d_B[i + t_id * P + d * P * Q] * val; 33 } 34 for (CeedInt d = 0; d < Q_COMP; d++) { 35 d_V[elem * strides_elem_V + comp * strides_comp_V + d * strides_q_comp_V + t_id] = r_V[d]; 36 } 37 } 38 } 39 40 //------------------------------------------------------------------------------ 41 // Tensor contraction transpose 42 //------------------------------------------------------------------------------ 43 template <int NUM_COMP, int Q_COMP, int P, int Q> 44 inline __device__ void ContractTranspose(const CeedInt elem, const CeedInt strides_elem_U, const CeedInt strides_elem_V, const CeedInt strides_comp_U, 45 const CeedInt strides_comp_V, const CeedInt strides_q_comp_U, const CeedScalar *__restrict__ d_B, 46 const CeedScalar *__restrict__ d_U, CeedScalar *__restrict__ d_V) { 47 const CeedInt t_id = threadIdx.x; 48 const CeedScalar *U; 49 CeedScalar r_V; 50 // TODO load B in shared memory if blockDim.z > 1? 51 52 for (CeedInt comp = 0; comp < NUM_COMP; comp++) { 53 // Run with P threads 54 r_V = 0.0; 55 for (CeedInt d = 0; d < Q_COMP; d++) { 56 U = &d_U[elem * strides_elem_U + comp * strides_comp_U + d * strides_q_comp_U]; 57 for (CeedInt i = 0; i < Q; i++) r_V += d_B[t_id + i * P + d * P * Q] * U[i]; 58 } 59 d_V[elem * strides_elem_V + comp * strides_comp_V + t_id] += r_V; 60 } 61 } 62