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 CeedScalar *d_interp_1d; 71 CeedScalar *d_grad_1d; 72 CeedScalar *d_q_weight_1d; 73 } CeedBasis_Cuda; 74 75 typedef struct { 76 CUmodule module; 77 CUfunction Interp; 78 CUfunction InterpTranspose; 79 CUfunction Deriv; 80 CUfunction DerivTranspose; 81 CUfunction Weight; 82 CeedScalar *d_interp; 83 CeedScalar *d_grad; 84 CeedScalar *d_div; 85 CeedScalar *d_curl; 86 CeedScalar *d_q_weight; 87 } CeedBasisNonTensor_Cuda; 88 89 typedef struct { 90 CUmodule module; 91 const char *qfunction_name; 92 const char *qfunction_source; 93 CUfunction QFunction; 94 Fields_Cuda fields; 95 void *d_c; 96 } CeedQFunction_Cuda; 97 98 typedef struct { 99 void *h_data; 100 void *h_data_borrowed; 101 void *h_data_owned; 102 void *d_data; 103 void *d_data_borrowed; 104 void *d_data_owned; 105 } CeedQFunctionContext_Cuda; 106 107 typedef struct { 108 CUmodule module, module_point_block; 109 CUfunction LinearDiagonal; 110 CUfunction LinearPointBlock; 111 CeedElemRestriction diag_rstr, point_block_diag_rstr; 112 CeedVector elem_diag, point_block_elem_diag; 113 CeedEvalMode *d_eval_modes_in, *d_eval_modes_out; 114 CeedScalar *d_identity, *d_interp_in, *d_grad_in, *d_div_in, *d_curl_in; 115 CeedScalar *d_interp_out, *d_grad_out, *d_div_out, *d_curl_out; 116 } CeedOperatorDiag_Cuda; 117 118 typedef struct { 119 CUmodule module; 120 CUfunction LinearAssemble; 121 CeedInt block_size_x, block_size_y, elems_per_block; 122 CeedScalar *d_B_in, *d_B_out; 123 } CeedOperatorAssemble_Cuda; 124 125 typedef struct { 126 CeedVector *e_vecs; // E-vectors, inputs followed by outputs 127 uint64_t *input_states; // State tracking for passive inputs 128 CeedVector *q_vecs_in; // Input Q-vectors needed to apply operator 129 CeedVector *q_vecs_out; // Output Q-vectors needed to apply operator 130 CeedInt num_inputs, num_outputs; 131 CeedInt num_active_in, num_active_out; 132 CeedVector *qf_active_in; 133 CeedOperatorDiag_Cuda *diag; 134 CeedOperatorAssemble_Cuda *asmb; 135 } CeedOperator_Cuda; 136 137 CEED_INTERN int CeedGetCublasHandle_Cuda(Ceed ceed, cublasHandle_t *handle); 138 139 CEED_INTERN int CeedVectorCreate_Cuda(CeedSize n, CeedVector vec); 140 141 CEED_INTERN int CeedElemRestrictionCreate_Cuda(CeedMemType mem_type, CeedCopyMode copy_mode, const CeedInt *offsets, const bool *orients, 142 const CeedInt8 *curl_orients, CeedElemRestriction r); 143 144 CEED_INTERN int CeedBasisCreateTensorH1_Cuda(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 145 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis); 146 CEED_INTERN int CeedBasisCreateH1_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 147 const CeedScalar *grad, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 148 CEED_INTERN int CeedBasisCreateHdiv_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 149 const CeedScalar *div, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 150 CEED_INTERN int CeedBasisCreateHcurl_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 151 const CeedScalar *curl, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 152 153 CEED_INTERN int CeedQFunctionCreate_Cuda(CeedQFunction qf); 154 155 CEED_INTERN int CeedQFunctionContextCreate_Cuda(CeedQFunctionContext ctx); 156 157 CEED_INTERN int CeedOperatorCreate_Cuda(CeedOperator op); 158