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 #pragma once 8 9 #include <ceed.h> 10 #include <ceed/backend.h> 11 #include <ceed/jit-source/cuda/cuda-types.h> 12 #include <cublas_v2.h> 13 #include <cuda.h> 14 15 typedef struct { 16 CeedScalar *h_array; 17 CeedScalar *h_array_borrowed; 18 CeedScalar *h_array_owned; 19 CeedScalar *d_array; 20 CeedScalar *d_array_borrowed; 21 CeedScalar *d_array_owned; 22 } CeedVector_Cuda; 23 24 typedef struct { 25 CUmodule module; 26 CUfunction ApplyNoTranspose, ApplyTranspose; 27 CUfunction ApplyUnsignedNoTranspose, ApplyUnsignedTranspose; 28 CUfunction ApplyUnorientedNoTranspose, ApplyUnorientedTranspose; 29 CeedInt num_nodes; 30 const CeedInt *h_offsets; 31 const CeedInt *h_offsets_borrowed; 32 const CeedInt *h_offsets_owned; 33 const CeedInt *d_offsets; 34 const CeedInt *d_offsets_borrowed; 35 const CeedInt *d_offsets_owned; 36 const CeedInt *d_t_offsets; 37 const CeedInt *d_t_indices; 38 const CeedInt *d_l_vec_indices; 39 const bool *h_orients; 40 const bool *h_orients_borrowed; 41 const bool *h_orients_owned; 42 const bool *d_orients; 43 const bool *d_orients_borrowed; 44 const bool *d_orients_owned; 45 const CeedInt8 *h_curl_orients; 46 const CeedInt8 *h_curl_orients_borrowed; 47 const CeedInt8 *h_curl_orients_owned; 48 const CeedInt8 *d_curl_orients; 49 const CeedInt8 *d_curl_orients_borrowed; 50 const CeedInt8 *d_curl_orients_owned; 51 const CeedInt *h_offsets_at_points; 52 const CeedInt *h_offsets_at_points_borrowed; 53 const CeedInt *h_offsets_at_points_owned; 54 const CeedInt *d_offsets_at_points; 55 const CeedInt *d_offsets_at_points_borrowed; 56 const CeedInt *d_offsets_at_points_owned; 57 const CeedInt *h_points_per_elem; 58 const CeedInt *h_points_per_elem_borrowed; 59 const CeedInt *h_points_per_elem_owned; 60 const CeedInt *d_points_per_elem; 61 const CeedInt *d_points_per_elem_borrowed; 62 const CeedInt *d_points_per_elem_owned; 63 } CeedElemRestriction_Cuda; 64 65 typedef struct { 66 CUmodule module; 67 CUfunction Interp; 68 CUfunction Grad; 69 CUfunction Weight; 70 CUmodule moduleAtPoints; 71 CeedInt num_points; 72 CUfunction InterpAtPoints; 73 CUfunction GradAtPoints; 74 CeedScalar *d_interp_1d; 75 CeedScalar *d_grad_1d; 76 CeedScalar *d_q_weight_1d; 77 CeedScalar *d_chebyshev_interp_1d; 78 } CeedBasis_Cuda; 79 80 typedef struct { 81 CUmodule module; 82 CUfunction Interp; 83 CUfunction InterpTranspose; 84 CUfunction Deriv; 85 CUfunction DerivTranspose; 86 CUfunction Weight; 87 CeedScalar *d_interp; 88 CeedScalar *d_grad; 89 CeedScalar *d_div; 90 CeedScalar *d_curl; 91 CeedScalar *d_q_weight; 92 } CeedBasisNonTensor_Cuda; 93 94 typedef struct { 95 CUmodule module; 96 const char *qfunction_name; 97 const char *qfunction_source; 98 CUfunction QFunction; 99 Fields_Cuda fields; 100 void *d_c; 101 } CeedQFunction_Cuda; 102 103 typedef struct { 104 void *h_data; 105 void *h_data_borrowed; 106 void *h_data_owned; 107 void *d_data; 108 void *d_data_borrowed; 109 void *d_data_owned; 110 } CeedQFunctionContext_Cuda; 111 112 typedef struct { 113 CUmodule module, module_point_block; 114 CUfunction LinearDiagonal; 115 CUfunction LinearPointBlock; 116 CeedElemRestriction diag_rstr, point_block_diag_rstr; 117 CeedVector elem_diag, point_block_elem_diag; 118 CeedEvalMode *d_eval_modes_in, *d_eval_modes_out; 119 CeedScalar *d_identity, *d_interp_in, *d_grad_in, *d_div_in, *d_curl_in; 120 CeedScalar *d_interp_out, *d_grad_out, *d_div_out, *d_curl_out; 121 } CeedOperatorDiag_Cuda; 122 123 typedef struct { 124 CUmodule module; 125 CUfunction LinearAssemble; 126 CeedInt block_size_x, block_size_y, elems_per_block; 127 CeedScalar *d_B_in, *d_B_out; 128 } CeedOperatorAssemble_Cuda; 129 130 typedef struct { 131 bool *skip_rstr_in, *skip_rstr_out, *apply_add_basis_out; 132 uint64_t *input_states; // State tracking for passive inputs 133 CeedVector *e_vecs; // E-vectors, inputs followed by outputs 134 CeedVector *q_vecs_in; // Input Q-vectors needed to apply operator 135 CeedVector *q_vecs_out; // Output Q-vectors needed to apply operator 136 CeedInt num_inputs, num_outputs; 137 CeedInt num_active_in, num_active_out; 138 CeedInt max_num_points; 139 CeedVector *qf_active_in, point_coords_elem; 140 CeedOperatorDiag_Cuda *diag; 141 CeedOperatorAssemble_Cuda *asmb; 142 } CeedOperator_Cuda; 143 144 CEED_INTERN int CeedGetCublasHandle_Cuda(Ceed ceed, cublasHandle_t *handle); 145 146 CEED_INTERN int CeedVectorCreate_Cuda(CeedSize n, CeedVector vec); 147 148 CEED_INTERN int CeedElemRestrictionCreate_Cuda(CeedMemType mem_type, CeedCopyMode copy_mode, const CeedInt *offsets, const bool *orients, 149 const CeedInt8 *curl_orients, CeedElemRestriction r); 150 151 CEED_INTERN int CeedBasisCreateTensorH1_Cuda(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 152 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis); 153 CEED_INTERN int CeedBasisCreateH1_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 154 const CeedScalar *grad, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 155 CEED_INTERN int CeedBasisCreateHdiv_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 156 const CeedScalar *div, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 157 CEED_INTERN int CeedBasisCreateHcurl_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 158 const CeedScalar *curl, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 159 160 CEED_INTERN int CeedQFunctionCreate_Cuda(CeedQFunction qf); 161 162 CEED_INTERN int CeedQFunctionContextCreate_Cuda(CeedQFunctionContext ctx); 163 164 CEED_INTERN int CeedOperatorCreate_Cuda(CeedOperator op); 165 CEED_INTERN int CeedOperatorCreateAtPoints_Cuda(CeedOperator op); 166