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 HIP offset element restriction kernels 10 11 #include <ceed.h> 12 13 //------------------------------------------------------------------------------ 14 // L-vector -> E-vector, standard (with offsets) 15 //------------------------------------------------------------------------------ 16 extern "C" __global__ void OffsetNoTranspose(const CeedInt *__restrict__ indices, 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 ind = indices[node]; 19 const CeedInt loc_node = node % RSTR_ELEM_SIZE; 20 const CeedInt elem = node / RSTR_ELEM_SIZE; 21 22 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) { 23 v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = u[ind + comp * RSTR_COMP_STRIDE]; 24 } 25 } 26 } 27 28 //------------------------------------------------------------------------------ 29 // E-vector -> L-vector, standard (with offsets) 30 //------------------------------------------------------------------------------ 31 #if !USE_DETERMINISTIC 32 extern "C" __global__ void OffsetTranspose(const CeedInt *__restrict__ indices, const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 33 for (CeedInt node = blockIdx.x * blockDim.x + threadIdx.x; node < RSTR_NUM_ELEM * RSTR_ELEM_SIZE; node += blockDim.x * gridDim.x) { 34 const CeedInt ind = indices[node]; 35 const CeedInt loc_node = node % RSTR_ELEM_SIZE; 36 const CeedInt elem = node / RSTR_ELEM_SIZE; 37 38 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) { 39 atomicAdd(&v[ind + comp * RSTR_COMP_STRIDE], u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE]); 40 } 41 } 42 } 43 #else 44 extern "C" __global__ void OffsetTranspose(const CeedInt *__restrict__ l_vec_indices, const CeedInt *__restrict__ t_indices, 45 const CeedInt *__restrict__ t_offsets, const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) { 46 CeedScalar value[RSTR_NUM_COMP]; 47 48 for (CeedInt i = blockIdx.x * blockDim.x + threadIdx.x; i < RSTR_NUM_NODES; i += blockDim.x * gridDim.x) { 49 const CeedInt ind = l_vec_indices[i]; 50 const CeedInt range_1 = t_offsets[i]; 51 const CeedInt range_N = t_offsets[i + 1]; 52 53 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) value[comp] = 0.0; 54 55 for (CeedInt j = range_1; j < range_N; j++) { 56 const CeedInt t_ind = t_indices[j]; 57 const CeedInt loc_node = t_ind % RSTR_ELEM_SIZE; 58 const CeedInt elem = t_ind / RSTR_ELEM_SIZE; 59 60 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) { 61 value[comp] += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE]; 62 } 63 } 64 65 for (CeedInt comp = 0; comp < RSTR_NUM_COMP; comp++) v[ind + comp * RSTR_COMP_STRIDE] += value[comp]; 66 } 67 } 68 #endif 69