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 /// @file 9 /// Private header for frontend components of libCEED 10 #ifndef _ceed_impl_h 11 #define _ceed_impl_h 12 13 #include <ceed.h> 14 #include <ceed/backend.h> 15 #include <stdbool.h> 16 17 CEED_INTERN const char *CeedJitSourceRootDefault; 18 19 /** @defgroup CeedUser Public API for Ceed 20 @ingroup Ceed 21 */ 22 /** @defgroup CeedBackend Backend API for Ceed 23 @ingroup Ceed 24 */ 25 /** @defgroup CeedDeveloper Internal library functions for Ceed 26 @ingroup Ceed 27 */ 28 /** @defgroup CeedVectorUser Public API for CeedVector 29 @ingroup CeedVector 30 */ 31 /** @defgroup CeedVectorBackend Backend API for CeedVector 32 @ingroup CeedVector 33 */ 34 /** @defgroup CeedVectorDeveloper Internal library functions for CeedVector 35 @ingroup CeedVector 36 */ 37 /** @defgroup CeedElemRestrictionUser Public API for CeedElemRestriction 38 @ingroup CeedElemRestriction 39 */ 40 /** @defgroup CeedElemRestrictionBackend Backend API for CeedElemRestriction 41 @ingroup CeedElemRestriction 42 */ 43 /** @defgroup CeedElemRestrictionDeveloper Internal library functions for CeedElemRestriction 44 @ingroup CeedElemRestriction 45 */ 46 /** @defgroup CeedBasisUser Public API for CeedBasis 47 @ingroup CeedBasis 48 */ 49 /** @defgroup CeedBasisBackend Backend API for CeedBasis 50 @ingroup CeedBasis 51 */ 52 /** @defgroup CeedBasisDeveloper Internal library functions for CeedBasis 53 @ingroup CeedBasis 54 */ 55 /** @defgroup CeedQFunctionUser Public API for CeedQFunction 56 @ingroup CeedQFunction 57 */ 58 /** @defgroup CeedQFunctionBackend Backend API for CeedQFunction 59 @ingroup CeedQFunction 60 */ 61 /** @defgroup CeedQFunctionDeveloper Internal library functions for CeedQFunction 62 @ingroup CeedQFunction 63 */ 64 /** @defgroup CeedOperatorUser Public API for CeedOperator 65 @ingroup CeedOperator 66 */ 67 /** @defgroup CeedOperatorBackend Backend API for CeedOperator 68 @ingroup CeedOperator 69 */ 70 /** @defgroup CeedOperatorDeveloper Internal library functions for CeedOperator 71 @ingroup CeedOperator 72 */ 73 74 // Lookup table field for backend functions 75 typedef struct { 76 const char *func_name; 77 size_t offset; 78 } FOffset; 79 80 // Lookup table field for object delegates 81 typedef struct { 82 char *obj_name; 83 Ceed delegate; 84 } ObjDelegate; 85 86 struct Ceed_private { 87 const char *resource; 88 Ceed delegate; 89 Ceed parent; 90 ObjDelegate *obj_delegates; 91 int obj_delegate_count; 92 Ceed op_fallback_ceed, op_fallback_parent; 93 const char *op_fallback_resource; 94 char **jit_source_roots; 95 CeedInt num_jit_source_roots; 96 int (*Error)(Ceed, const char *, int, const char *, int, const char *, va_list *); 97 int (*SetStream)(Ceed, void *); 98 int (*GetPreferredMemType)(CeedMemType *); 99 int (*Destroy)(Ceed); 100 int (*VectorCreate)(CeedSize, CeedVector); 101 int (*ElemRestrictionCreate)(CeedMemType, CeedCopyMode, const CeedInt *, CeedElemRestriction); 102 int (*ElemRestrictionCreateOriented)(CeedMemType, CeedCopyMode, const CeedInt *, const bool *, CeedElemRestriction); 103 int (*ElemRestrictionCreateBlocked)(CeedMemType, CeedCopyMode, const CeedInt *, CeedElemRestriction); 104 int (*BasisCreateTensorH1)(CeedInt, CeedInt, CeedInt, const CeedScalar *, const CeedScalar *, const CeedScalar *, const CeedScalar *, CeedBasis); 105 int (*BasisCreateH1)(CeedElemTopology, CeedInt, CeedInt, CeedInt, const CeedScalar *, const CeedScalar *, const CeedScalar *, const CeedScalar *, 106 CeedBasis); 107 int (*BasisCreateHdiv)(CeedElemTopology, CeedInt, CeedInt, CeedInt, const CeedScalar *, const CeedScalar *, const CeedScalar *, const CeedScalar *, 108 CeedBasis); 109 int (*BasisCreateHcurl)(CeedElemTopology, CeedInt, CeedInt, CeedInt, const CeedScalar *, const CeedScalar *, const CeedScalar *, const CeedScalar *, 110 CeedBasis); 111 int (*TensorContractCreate)(CeedBasis, CeedTensorContract); 112 int (*QFunctionCreate)(CeedQFunction); 113 int (*QFunctionContextCreate)(CeedQFunctionContext); 114 int (*OperatorCreate)(CeedOperator); 115 int (*CompositeOperatorCreate)(CeedOperator); 116 int ref_count; 117 void *data; 118 bool is_debug; 119 bool has_valid_op_fallback_resource; 120 bool is_deterministic; 121 char err_msg[CEED_MAX_RESOURCE_LEN]; 122 FOffset *f_offsets; 123 }; 124 125 struct CeedVector_private { 126 Ceed ceed; 127 int (*HasValidArray)(CeedVector, bool *); 128 int (*HasBorrowedArrayOfType)(CeedVector, CeedMemType, bool *); 129 int (*SetArray)(CeedVector, CeedMemType, CeedCopyMode, CeedScalar *); 130 int (*SetValue)(CeedVector, CeedScalar); 131 int (*SyncArray)(CeedVector, CeedMemType); 132 int (*TakeArray)(CeedVector, CeedMemType, CeedScalar **); 133 int (*GetArray)(CeedVector, CeedMemType, CeedScalar **); 134 int (*GetArrayRead)(CeedVector, CeedMemType, const CeedScalar **); 135 int (*GetArrayWrite)(CeedVector, CeedMemType, CeedScalar **); 136 int (*RestoreArray)(CeedVector); 137 int (*RestoreArrayRead)(CeedVector); 138 int (*Norm)(CeedVector, CeedNormType, CeedScalar *); 139 int (*Scale)(CeedVector, CeedScalar); 140 int (*AXPY)(CeedVector, CeedScalar, CeedVector); 141 int (*AXPBY)(CeedVector, CeedScalar, CeedScalar, CeedVector); 142 int (*PointwiseMult)(CeedVector, CeedVector, CeedVector); 143 int (*Reciprocal)(CeedVector); 144 int (*Destroy)(CeedVector); 145 int ref_count; 146 CeedSize length; 147 uint64_t state; 148 uint64_t num_readers; 149 void *data; 150 }; 151 152 struct CeedElemRestriction_private { 153 Ceed ceed; 154 CeedElemRestriction rstr_signed; 155 int (*Apply)(CeedElemRestriction, CeedTransposeMode, CeedVector, CeedVector, CeedRequest *); 156 int (*ApplyUnsigned)(CeedElemRestriction, CeedTransposeMode, CeedVector, CeedVector, CeedRequest *); 157 int (*ApplyBlock)(CeedElemRestriction, CeedInt, CeedTransposeMode, CeedVector, CeedVector, CeedRequest *); 158 int (*GetOffsets)(CeedElemRestriction, CeedMemType, const CeedInt **); 159 int (*Destroy)(CeedElemRestriction); 160 int ref_count; 161 CeedInt num_elem; /* number of elements */ 162 CeedInt elem_size; /* number of nodes per element */ 163 CeedInt num_comp; /* number of components */ 164 CeedInt comp_stride; /* Component stride for L-vector ordering */ 165 CeedSize l_size; /* size of the L-vector, can be used for checking for correct vector sizes */ 166 CeedInt blk_size; /* number of elements in a batch */ 167 CeedInt num_blk; /* number of blocks of elements */ 168 CeedInt *strides; /* strides between [nodes, components, elements] */ 169 CeedInt layout[3]; /* E-vector layout [nodes, components, elements] */ 170 uint64_t num_readers; /* number of instances of offset read only access */ 171 bool is_oriented; /* flag for oriented restriction */ 172 void *data; /* place for the backend to store any data */ 173 }; 174 175 struct CeedBasis_private { 176 Ceed ceed; 177 int (*Apply)(CeedBasis, CeedInt, CeedTransposeMode, CeedEvalMode, CeedVector, CeedVector); 178 int (*ApplyAtPoints)(CeedBasis, CeedInt, CeedTransposeMode, CeedEvalMode, CeedVector, CeedVector, CeedVector); 179 int (*Destroy)(CeedBasis); 180 int ref_count; 181 bool is_tensor_basis; /* flag for tensor basis */ 182 CeedInt dim; /* topological dimension */ 183 CeedElemTopology topo; /* element topology */ 184 CeedInt num_comp; /* number of field components (1 for scalar fields) */ 185 CeedInt P_1d; /* number of nodes in one dimension */ 186 CeedInt Q_1d; /* number of quadrature points in one dimension */ 187 CeedInt P; /* total number of nodes */ 188 CeedInt Q; /* total number of quadrature points */ 189 CeedFESpace fe_space; /* initialized in basis constructor with 1, 2, 3 for H^1, H(div), and H(curl) FE space */ 190 CeedTensorContract contract; /* tensor contraction object */ 191 CeedScalar *q_ref_1d; /* array of length Q1d holding the locations of quadrature points on the 1D reference element [-1, 1] */ 192 CeedScalar *q_weight_1d; /* array of length Q1d holding the quadrature weights on the reference element */ 193 CeedScalar *interp; /* row-major matrix of shape [Q, P] or [dim * Q, P] expressing the values of nodal basis functions or vector basis functions at 194 quadrature points */ 195 CeedScalar *interp_1d; /* row-major matrix of shape [Q1d, P1d] expressing the values of nodal basis functions at quadrature points */ 196 CeedScalar *grad; /* row-major matrix of shape [dim * Q, P] matrix expressing derivatives of nodal basis functions at quadrature points */ 197 CeedScalar *grad_1d; /* row-major matrix of shape [Q1d, P1d] matrix expressing derivatives of nodal basis functions at quadrature points */ 198 CeedScalar *div; /* row-major matrix of shape [Q, P] expressing the divergence of basis functions at quadrature points for H(div) discretizations */ 199 CeedScalar *curl; /* row-major matrix of shape [curl_dim * Q, P], curl_dim = 1 if dim < 3 else dim, expressing the curl of basis functions at 200 quadrature points for H(curl) discretizations */ 201 CeedVector vec_chebyshev; 202 CeedBasis basis_chebyshev; /* basis interpolating from nodes to Chebyshev polynomial coefficients */ 203 void *data; /* place for the backend to store any data */ 204 }; 205 206 struct CeedTensorContract_private { 207 Ceed ceed; 208 int (*Apply)(CeedTensorContract, CeedInt, CeedInt, CeedInt, CeedInt, const CeedScalar *restrict, CeedTransposeMode, const CeedInt, 209 const CeedScalar *restrict, CeedScalar *restrict); 210 int (*Destroy)(CeedTensorContract); 211 int ref_count; 212 void *data; 213 }; 214 215 struct CeedQFunctionField_private { 216 const char *field_name; 217 CeedInt size; 218 CeedEvalMode eval_mode; 219 }; 220 221 struct CeedQFunction_private { 222 Ceed ceed; 223 int (*Apply)(CeedQFunction, CeedInt, CeedVector *, CeedVector *); 224 int (*SetCUDAUserFunction)(CeedQFunction, void *); 225 int (*SetHIPUserFunction)(CeedQFunction, void *); 226 int (*Destroy)(CeedQFunction); 227 int ref_count; 228 CeedInt vec_length; /* Number of quadrature points must be padded to a multiple of vec_length */ 229 CeedQFunctionField *input_fields; 230 CeedQFunctionField *output_fields; 231 CeedInt num_input_fields, num_output_fields; 232 CeedQFunctionUser function; 233 CeedInt user_flop_estimate; 234 const char *user_source; 235 const char *source_path; 236 const char *kernel_name; 237 const char *gallery_name; 238 bool is_gallery; 239 bool is_identity; 240 bool is_fortran; 241 bool is_immutable; 242 bool is_context_writable; 243 CeedQFunctionContext ctx; /* user context for function */ 244 void *data; /* place for the backend to store any data */ 245 }; 246 247 struct CeedQFunctionContext_private { 248 Ceed ceed; 249 int ref_count; 250 int (*HasValidData)(CeedQFunctionContext, bool *); 251 int (*HasBorrowedDataOfType)(CeedQFunctionContext, CeedMemType, bool *); 252 int (*SetData)(CeedQFunctionContext, CeedMemType, CeedCopyMode, void *); 253 int (*TakeData)(CeedQFunctionContext, CeedMemType, void *); 254 int (*GetData)(CeedQFunctionContext, CeedMemType, void *); 255 int (*GetDataRead)(CeedQFunctionContext, CeedMemType, void *); 256 int (*RestoreData)(CeedQFunctionContext); 257 int (*RestoreDataRead)(CeedQFunctionContext); 258 int (*DataDestroy)(CeedQFunctionContext); 259 int (*Destroy)(CeedQFunctionContext); 260 CeedQFunctionContextDataDestroyUser data_destroy_function; 261 CeedMemType data_destroy_mem_type; 262 CeedInt num_fields; 263 CeedInt max_fields; 264 CeedContextFieldLabel *field_labels; 265 uint64_t state; 266 uint64_t num_readers; 267 size_t ctx_size; 268 void *data; 269 }; 270 271 /// Struct to handle the context data to use the Fortran QFunction stub 272 /// @ingroup CeedQFunction 273 struct CeedFortranContext_private { 274 CeedQFunctionContext inner_ctx; 275 void (*f)(void *ctx, int *nq, const CeedScalar *u, const CeedScalar *u1, const CeedScalar *u2, const CeedScalar *u3, const CeedScalar *u4, 276 const CeedScalar *u5, const CeedScalar *u6, const CeedScalar *u7, const CeedScalar *u8, const CeedScalar *u9, const CeedScalar *u10, 277 const CeedScalar *u11, const CeedScalar *u12, const CeedScalar *u13, const CeedScalar *u14, const CeedScalar *u15, CeedScalar *v, 278 CeedScalar *v1, CeedScalar *v2, CeedScalar *v3, CeedScalar *v4, CeedScalar *v5, CeedScalar *v6, CeedScalar *v7, CeedScalar *v8, 279 CeedScalar *v9, CeedScalar *v10, CeedScalar *v11, CeedScalar *v12, CeedScalar *v13, CeedScalar *v14, CeedScalar *v15, int *err); 280 }; 281 typedef struct CeedFortranContext_private *CeedFortranContext; 282 283 struct CeedContextFieldLabel_private { 284 const char *name; 285 const char *description; 286 CeedContextFieldType type; 287 size_t size; 288 size_t num_values; 289 size_t offset; 290 CeedInt num_sub_labels; 291 CeedContextFieldLabel *sub_labels; 292 bool from_op; 293 }; 294 295 struct CeedOperatorField_private { 296 CeedElemRestriction elem_rstr; /* Restriction from L-vector */ 297 CeedBasis basis; /* Basis or CEED_BASIS_COLLOCATED for collocated fields */ 298 CeedVector vec; /* State vector for passive fields or CEED_VECTOR_NONE for no vector */ 299 const char *field_name; /* matching QFunction field name */ 300 }; 301 302 struct CeedQFunctionAssemblyData_private { 303 Ceed ceed; 304 int ref_count; 305 bool is_setup; 306 bool reuse_data; 307 bool needs_data_update; 308 CeedVector vec; 309 CeedElemRestriction rstr; 310 }; 311 312 struct CeedOperatorAssemblyData_private { 313 Ceed ceed; 314 CeedInt num_active_bases; 315 CeedBasis *active_bases; 316 CeedElemRestriction *active_elem_rstrs; 317 CeedInt *num_eval_modes_in, *num_eval_modes_out; 318 CeedEvalMode **eval_modes_in, **eval_modes_out; 319 CeedScalar **assembled_bases_in, **assembled_bases_out; 320 CeedSize **eval_mode_offsets_in, **eval_mode_offsets_out, num_output_components; 321 }; 322 323 struct CeedOperator_private { 324 Ceed ceed; 325 CeedOperator op_fallback, op_fallback_parent; 326 int ref_count; 327 int (*LinearAssembleQFunction)(CeedOperator, CeedVector *, CeedElemRestriction *, CeedRequest *); 328 int (*LinearAssembleQFunctionUpdate)(CeedOperator, CeedVector, CeedElemRestriction, CeedRequest *); 329 int (*LinearAssembleDiagonal)(CeedOperator, CeedVector, CeedRequest *); 330 int (*LinearAssembleAddDiagonal)(CeedOperator, CeedVector, CeedRequest *); 331 int (*LinearAssemblePointBlockDiagonal)(CeedOperator, CeedVector, CeedRequest *); 332 int (*LinearAssembleAddPointBlockDiagonal)(CeedOperator, CeedVector, CeedRequest *); 333 int (*LinearAssembleSymbolic)(CeedOperator, CeedSize *, CeedInt **, CeedInt **); 334 int (*LinearAssemble)(CeedOperator, CeedVector); 335 int (*LinearAssembleSingle)(CeedOperator, CeedInt, CeedVector); 336 int (*CreateFDMElementInverse)(CeedOperator, CeedOperator *, CeedRequest *); 337 int (*Apply)(CeedOperator, CeedVector, CeedVector, CeedRequest *); 338 int (*ApplyComposite)(CeedOperator, CeedVector, CeedVector, CeedRequest *); 339 int (*ApplyAdd)(CeedOperator, CeedVector, CeedVector, CeedRequest *); 340 int (*ApplyAddComposite)(CeedOperator, CeedVector, CeedVector, CeedRequest *); 341 int (*ApplyJacobian)(CeedOperator, CeedVector, CeedVector, CeedVector, CeedVector, CeedRequest *); 342 int (*Destroy)(CeedOperator); 343 CeedOperatorField *input_fields; 344 CeedOperatorField *output_fields; 345 CeedSize input_size, output_size; 346 CeedInt num_elem; /* Number of elements */ 347 CeedInt num_qpts; /* Number of quadrature points over all elements */ 348 CeedInt num_fields; /* Number of fields that have been set */ 349 CeedQFunction qf; 350 CeedQFunction dqf; 351 CeedQFunction dqfT; 352 const char *name; 353 bool is_immutable; 354 bool is_interface_setup; 355 bool is_backend_setup; 356 bool is_composite; 357 bool has_restriction; 358 CeedQFunctionAssemblyData qf_assembled; 359 CeedOperatorAssemblyData op_assembled; 360 CeedOperator *sub_operators; 361 CeedInt num_suboperators; 362 void *data; 363 CeedInt num_context_labels; 364 CeedInt max_context_labels; 365 CeedContextFieldLabel *context_labels; 366 }; 367 368 CEED_INTERN int CeedOperatorGetFallback(CeedOperator op, CeedOperator *op_fallback); 369 370 #endif 371