13d8e8822SJeremy L Thompson // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 23d8e8822SJeremy L Thompson // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 30d0321e0SJeremy L Thompson // 43d8e8822SJeremy L Thompson // SPDX-License-Identifier: BSD-2-Clause 50d0321e0SJeremy L Thompson // 63d8e8822SJeremy L Thompson // This file is part of CEED: http://github.com/ceed 70d0321e0SJeremy L Thompson 80d0321e0SJeremy L Thompson #ifndef _ceed_cuda_h 90d0321e0SJeremy L Thompson #define _ceed_cuda_h 100d0321e0SJeremy L Thompson 1149aac155SJeremy L Thompson #include <ceed.h> 120d0321e0SJeremy L Thompson #include <ceed/backend.h> 1349aac155SJeremy L Thompson #include <ceed/jit-source/cuda/cuda-types.h> 1449aac155SJeremy L Thompson #include <cublas_v2.h> 150d0321e0SJeremy L Thompson #include <cuda.h> 162b730f8bSJeremy L Thompson 170d0321e0SJeremy L Thompson typedef struct { 180d0321e0SJeremy L Thompson CeedScalar *h_array; 190d0321e0SJeremy L Thompson CeedScalar *h_array_borrowed; 200d0321e0SJeremy L Thompson CeedScalar *h_array_owned; 210d0321e0SJeremy L Thompson CeedScalar *d_array; 220d0321e0SJeremy L Thompson CeedScalar *d_array_borrowed; 230d0321e0SJeremy L Thompson CeedScalar *d_array_owned; 240d0321e0SJeremy L Thompson } CeedVector_Cuda; 250d0321e0SJeremy L Thompson 260d0321e0SJeremy L Thompson typedef struct { 270d0321e0SJeremy L Thompson CUmodule module; 28437930d1SJeremy L Thompson CUfunction StridedTranspose; 29437930d1SJeremy L Thompson CUfunction StridedNoTranspose; 30437930d1SJeremy L Thompson CUfunction OffsetTranspose; 31437930d1SJeremy L Thompson CUfunction OffsetNoTranspose; 32437930d1SJeremy L Thompson CeedInt num_nodes; 330d0321e0SJeremy L Thompson CeedInt *h_ind; 340d0321e0SJeremy L Thompson CeedInt *h_ind_allocated; 350d0321e0SJeremy L Thompson CeedInt *d_ind; 360d0321e0SJeremy L Thompson CeedInt *d_ind_allocated; 37437930d1SJeremy L Thompson CeedInt *d_t_offsets; 38437930d1SJeremy L Thompson CeedInt *d_t_indices; 39437930d1SJeremy L Thompson CeedInt *d_l_vec_indices; 400d0321e0SJeremy L Thompson } CeedElemRestriction_Cuda; 410d0321e0SJeremy L Thompson 42437930d1SJeremy L Thompson typedef struct { 43437930d1SJeremy L Thompson CUmodule module; 44437930d1SJeremy L Thompson CUfunction Interp; 45437930d1SJeremy L Thompson CUfunction Grad; 46437930d1SJeremy L Thompson CUfunction Weight; 47437930d1SJeremy L Thompson CeedScalar *d_interp_1d; 48437930d1SJeremy L Thompson CeedScalar *d_grad_1d; 49437930d1SJeremy L Thompson CeedScalar *d_q_weight_1d; 50437930d1SJeremy L Thompson } CeedBasis_Cuda; 51437930d1SJeremy L Thompson 52437930d1SJeremy L Thompson typedef struct { 53437930d1SJeremy L Thompson CUmodule module; 54437930d1SJeremy L Thompson CUfunction Interp; 55437930d1SJeremy L Thompson CUfunction Grad; 56437930d1SJeremy L Thompson CUfunction Weight; 57437930d1SJeremy L Thompson CeedScalar *d_interp; 58437930d1SJeremy L Thompson CeedScalar *d_grad; 59437930d1SJeremy L Thompson CeedScalar *d_q_weight; 60437930d1SJeremy L Thompson } CeedBasisNonTensor_Cuda; 61437930d1SJeremy L Thompson 620d0321e0SJeremy L Thompson typedef struct { 630d0321e0SJeremy L Thompson CUmodule module; 64437930d1SJeremy L Thompson char *qfunction_name; 65437930d1SJeremy L Thompson char *qfunction_source; 66437930d1SJeremy L Thompson CUfunction QFunction; 670d0321e0SJeremy L Thompson Fields_Cuda fields; 680d0321e0SJeremy L Thompson void *d_c; 690d0321e0SJeremy L Thompson } CeedQFunction_Cuda; 700d0321e0SJeremy L Thompson 710d0321e0SJeremy L Thompson typedef struct { 720d0321e0SJeremy L Thompson void *h_data; 730d0321e0SJeremy L Thompson void *h_data_borrowed; 740d0321e0SJeremy L Thompson void *h_data_owned; 750d0321e0SJeremy L Thompson void *d_data; 760d0321e0SJeremy L Thompson void *d_data_borrowed; 770d0321e0SJeremy L Thompson void *d_data_owned; 780d0321e0SJeremy L Thompson } CeedQFunctionContext_Cuda; 790d0321e0SJeremy L Thompson 800d0321e0SJeremy L Thompson typedef struct { 810d0321e0SJeremy L Thompson CUmodule module; 820d0321e0SJeremy L Thompson CUfunction linearDiagonal; 830d0321e0SJeremy L Thompson CUfunction linearPointBlock; 840d0321e0SJeremy L Thompson CeedBasis basisin, basisout; 850d0321e0SJeremy L Thompson CeedElemRestriction diagrstr, pbdiagrstr; 860d0321e0SJeremy L Thompson CeedVector elemdiag, pbelemdiag; 870d0321e0SJeremy L Thompson CeedInt numemodein, numemodeout, nnodes; 880d0321e0SJeremy L Thompson CeedEvalMode *h_emodein, *h_emodeout; 890d0321e0SJeremy L Thompson CeedEvalMode *d_emodein, *d_emodeout; 900d0321e0SJeremy L Thompson CeedScalar *d_identity, *d_interpin, *d_interpout, *d_gradin, *d_gradout; 910d0321e0SJeremy L Thompson } CeedOperatorDiag_Cuda; 920d0321e0SJeremy L Thompson 930d0321e0SJeremy L Thompson typedef struct { 94cc132f9aSnbeams CUmodule module; 95cc132f9aSnbeams CUfunction linearAssemble; 9659ad764aSnbeams CeedInt nelem, block_size_x, block_size_y, elemsPerBlock; 97cc132f9aSnbeams CeedScalar *d_B_in, *d_B_out; 98cc132f9aSnbeams } CeedOperatorAssemble_Cuda; 99cc132f9aSnbeams 100cc132f9aSnbeams typedef struct { 1010d0321e0SJeremy L Thompson CeedVector *evecs; // E-vectors, inputs followed by outputs 1020d0321e0SJeremy L Thompson CeedVector *qvecsin; // Input Q-vectors needed to apply operator 1030d0321e0SJeremy L Thompson CeedVector *qvecsout; // Output Q-vectors needed to apply operator 1040d0321e0SJeremy L Thompson CeedInt numein; 1050d0321e0SJeremy L Thompson CeedInt numeout; 1060d0321e0SJeremy L Thompson CeedInt qfnumactivein, qfnumactiveout; 1070d0321e0SJeremy L Thompson CeedVector *qfactivein; 1080d0321e0SJeremy L Thompson CeedOperatorDiag_Cuda *diag; 109cc132f9aSnbeams CeedOperatorAssemble_Cuda *asmb; 1100d0321e0SJeremy L Thompson } CeedOperator_Cuda; 1110d0321e0SJeremy L Thompson 112eb7e6cafSJeremy L Thompson CEED_INTERN int CeedGetCublasHandle_Cuda(Ceed ceed, cublasHandle_t *handle); 1130d0321e0SJeremy L Thompson 1141f9221feSJeremy L Thompson CEED_INTERN int CeedVectorCreate_Cuda(CeedSize n, CeedVector vec); 1150d0321e0SJeremy L Thompson 116*0305e208SSebastian Grimberg CEED_INTERN int CeedElemRestrictionCreate_Cuda(CeedMemType mem_type, CeedCopyMode copy_mode, const CeedInt *indices, const bool *orients, 117*0305e208SSebastian Grimberg const CeedInt *curl_orients, CeedElemRestriction r); 1180d0321e0SJeremy L Thompson 1192b730f8bSJeremy L Thompson CEED_INTERN int CeedBasisCreateTensorH1_Cuda(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 1206574a04fSJeremy L Thompson const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis); 1210d0321e0SJeremy L Thompson 12251475c7cSJeremy L Thompson CEED_INTERN int CeedBasisCreateH1_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 12351475c7cSJeremy L Thompson const CeedScalar *grad, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 1240d0321e0SJeremy L Thompson 1250d0321e0SJeremy L Thompson CEED_INTERN int CeedQFunctionCreate_Cuda(CeedQFunction qf); 1260d0321e0SJeremy L Thompson 1270d0321e0SJeremy L Thompson CEED_INTERN int CeedQFunctionContextCreate_Cuda(CeedQFunctionContext ctx); 1280d0321e0SJeremy L Thompson 1290d0321e0SJeremy L Thompson CEED_INTERN int CeedOperatorCreate_Cuda(CeedOperator op); 1300d0321e0SJeremy L Thompson 1310d0321e0SJeremy L Thompson #endif 132