1 // Copyright (c) 2017-2022, 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 #ifndef _ceed_cuda_ref_h 9 #define _ceed_cuda_ref_h 10 11 #include <ceed.h> 12 #include <ceed/backend.h> 13 #include <ceed/jit-source/cuda/cuda-types.h> 14 #include <cublas_v2.h> 15 #include <cuda.h> 16 17 typedef struct { 18 CeedScalar *h_array; 19 CeedScalar *h_array_borrowed; 20 CeedScalar *h_array_owned; 21 CeedScalar *d_array; 22 CeedScalar *d_array_borrowed; 23 CeedScalar *d_array_owned; 24 } CeedVector_Cuda; 25 26 typedef struct { 27 CUmodule module; 28 CUfunction StridedNoTranspose; 29 CUfunction StridedTranspose; 30 CUfunction OffsetNoTranspose; 31 CUfunction OffsetTranspose; 32 CUfunction OffsetTransposeDet; 33 CeedInt num_nodes; 34 CeedInt *h_ind; 35 CeedInt *h_ind_allocated; 36 CeedInt *d_ind; 37 CeedInt *d_ind_allocated; 38 CeedInt *d_t_offsets; 39 CeedInt *d_t_indices; 40 CeedInt *d_l_vec_indices; 41 } CeedElemRestriction_Cuda; 42 43 typedef struct { 44 CUmodule module; 45 CUfunction Interp; 46 CUfunction Grad; 47 CUfunction Weight; 48 CeedScalar *d_interp_1d; 49 CeedScalar *d_grad_1d; 50 CeedScalar *d_q_weight_1d; 51 } CeedBasis_Cuda; 52 53 typedef struct { 54 CUmodule module; 55 CUfunction Interp; 56 CUfunction Grad; 57 CUfunction Weight; 58 CeedScalar *d_interp; 59 CeedScalar *d_grad; 60 CeedScalar *d_q_weight; 61 } CeedBasisNonTensor_Cuda; 62 63 typedef struct { 64 CUmodule module; 65 char *qfunction_name; 66 char *qfunction_source; 67 CUfunction QFunction; 68 Fields_Cuda fields; 69 void *d_c; 70 } CeedQFunction_Cuda; 71 72 typedef struct { 73 void *h_data; 74 void *h_data_borrowed; 75 void *h_data_owned; 76 void *d_data; 77 void *d_data_borrowed; 78 void *d_data_owned; 79 } CeedQFunctionContext_Cuda; 80 81 typedef struct { 82 CUmodule module; 83 CUfunction linearDiagonal; 84 CUfunction linearPointBlock; 85 CeedBasis basisin, basisout; 86 CeedElemRestriction diagrstr, pbdiagrstr; 87 CeedVector elemdiag, pbelemdiag; 88 CeedInt numemodein, numemodeout, nnodes; 89 CeedEvalMode *h_emodein, *h_emodeout; 90 CeedEvalMode *d_emodein, *d_emodeout; 91 CeedScalar *d_identity, *d_interpin, *d_interpout, *d_gradin, *d_gradout; 92 } CeedOperatorDiag_Cuda; 93 94 typedef struct { 95 CUmodule module; 96 CUfunction linearAssemble; 97 CeedInt nelem, block_size_x, block_size_y, elemsPerBlock; 98 CeedScalar *d_B_in, *d_B_out; 99 } CeedOperatorAssemble_Cuda; 100 101 typedef struct { 102 CeedVector *evecs; // E-vectors, inputs followed by outputs 103 CeedVector *qvecsin; // Input Q-vectors needed to apply operator 104 CeedVector *qvecsout; // Output Q-vectors needed to apply operator 105 CeedInt numein; 106 CeedInt numeout; 107 CeedInt qfnumactivein, qfnumactiveout; 108 CeedVector *qfactivein; 109 CeedOperatorDiag_Cuda *diag; 110 CeedOperatorAssemble_Cuda *asmb; 111 } CeedOperator_Cuda; 112 113 CEED_INTERN int CeedGetCublasHandle_Cuda(Ceed ceed, cublasHandle_t *handle); 114 115 CEED_INTERN int CeedVectorCreate_Cuda(CeedSize n, CeedVector vec); 116 117 CEED_INTERN int CeedElemRestrictionCreate_Cuda(CeedMemType mem_type, CeedCopyMode copy_mode, const CeedInt *indices, const bool *orients, 118 const CeedInt8 *curl_orients, CeedElemRestriction r); 119 120 CEED_INTERN int CeedBasisCreateTensorH1_Cuda(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 121 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis); 122 123 CEED_INTERN int CeedBasisCreateH1_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 124 const CeedScalar *grad, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 125 126 CEED_INTERN int CeedQFunctionCreate_Cuda(CeedQFunction qf); 127 128 CEED_INTERN int CeedQFunctionContextCreate_Cuda(CeedQFunctionContext ctx); 129 130 CEED_INTERN int CeedOperatorCreate_Cuda(CeedOperator op); 131 132 #endif // _ceed_cuda_ref_h 133