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 CUfunction OrientedNoTranspose; 34 CUfunction OrientedTranspose; 35 CUfunction CurlOrientedNoTranspose; 36 CUfunction CurlOrientedTranspose; 37 CUfunction CurlOrientedUnsignedNoTranspose; 38 CUfunction CurlOrientedUnsignedTranspose; 39 CeedInt num_nodes; 40 CeedInt *h_ind; 41 CeedInt *h_ind_allocated; 42 CeedInt *d_ind; 43 CeedInt *d_ind_allocated; 44 CeedInt *d_t_offsets; 45 CeedInt *d_t_indices; 46 CeedInt *d_l_vec_indices; 47 bool *h_orients; 48 bool *h_orients_allocated; 49 bool *d_orients; 50 bool *d_orients_allocated; 51 CeedInt8 *h_curl_orients; 52 CeedInt8 *h_curl_orients_allocated; 53 CeedInt8 *d_curl_orients; 54 CeedInt8 *d_curl_orients_allocated; 55 } CeedElemRestriction_Cuda; 56 57 typedef struct { 58 CUmodule module; 59 CUfunction Interp; 60 CUfunction Grad; 61 CUfunction Weight; 62 CeedScalar *d_interp_1d; 63 CeedScalar *d_grad_1d; 64 CeedScalar *d_q_weight_1d; 65 } CeedBasis_Cuda; 66 67 typedef struct { 68 CUmodule module; 69 CUfunction Interp; 70 CUfunction InterpTranspose; 71 CUfunction Deriv; 72 CUfunction DerivTranspose; 73 CUfunction Weight; 74 CeedScalar *d_interp; 75 CeedScalar *d_grad; 76 CeedScalar *d_div; 77 CeedScalar *d_curl; 78 CeedScalar *d_q_weight; 79 } CeedBasisNonTensor_Cuda; 80 81 typedef struct { 82 CUmodule module; 83 char *qfunction_name; 84 char *qfunction_source; 85 CUfunction QFunction; 86 Fields_Cuda fields; 87 void *d_c; 88 } CeedQFunction_Cuda; 89 90 typedef struct { 91 void *h_data; 92 void *h_data_borrowed; 93 void *h_data_owned; 94 void *d_data; 95 void *d_data_borrowed; 96 void *d_data_owned; 97 } CeedQFunctionContext_Cuda; 98 99 typedef struct { 100 CUmodule module; 101 CUfunction LinearDiagonal; 102 CUfunction LinearPointBlock; 103 CeedElemRestriction diag_rstr, point_block_diag_rstr; 104 CeedVector elem_diag, point_block_elem_diag; 105 CeedEvalMode *d_eval_modes_in, *d_eval_modes_out; 106 CeedScalar *d_identity, *d_interp_in, *d_grad_in, *d_div_in, *d_curl_in; 107 CeedScalar *d_interp_out, *d_grad_out, *d_div_out, *d_curl_out; 108 } CeedOperatorDiag_Cuda; 109 110 typedef struct { 111 CUmodule module; 112 CUfunction LinearAssemble; 113 CeedInt block_size_x, block_size_y, elems_per_block; 114 CeedScalar *d_B_in, *d_B_out; 115 } CeedOperatorAssemble_Cuda; 116 117 typedef struct { 118 CeedVector *e_vecs; // E-vectors, inputs followed by outputs 119 CeedVector *q_vecs_in; // Input Q-vectors needed to apply operator 120 CeedVector *q_vecs_out; // Output Q-vectors needed to apply operator 121 CeedInt num_inputs, num_outputs; 122 CeedInt num_active_in, num_active_out; 123 CeedVector *qf_active_in; 124 CeedOperatorDiag_Cuda *diag; 125 CeedOperatorAssemble_Cuda *asmb; 126 } CeedOperator_Cuda; 127 128 CEED_INTERN int CeedGetCublasHandle_Cuda(Ceed ceed, cublasHandle_t *handle); 129 130 CEED_INTERN int CeedVectorCreate_Cuda(CeedSize n, CeedVector vec); 131 132 CEED_INTERN int CeedElemRestrictionCreate_Cuda(CeedMemType mem_type, CeedCopyMode copy_mode, const CeedInt *indices, const bool *orients, 133 const CeedInt8 *curl_orients, CeedElemRestriction r); 134 135 CEED_INTERN int CeedBasisCreateTensorH1_Cuda(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 136 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis); 137 CEED_INTERN int CeedBasisCreateH1_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 138 const CeedScalar *grad, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 139 CEED_INTERN int CeedBasisCreateHdiv_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 140 const CeedScalar *div, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 141 CEED_INTERN int CeedBasisCreateHcurl_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 142 const CeedScalar *curl, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 143 144 CEED_INTERN int CeedQFunctionCreate_Cuda(CeedQFunction qf); 145 146 CEED_INTERN int CeedQFunctionContextCreate_Cuda(CeedQFunctionContext ctx); 147 148 CEED_INTERN int CeedOperatorCreate_Cuda(CeedOperator op); 149 150 #endif // CEED_CUDA_REF_H 151