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 8 #include <ceed-impl.h> 9 #include <ceed.h> 10 #include <ceed/backend.h> 11 #include <assert.h> 12 #include <math.h> 13 #include <stdbool.h> 14 #include <stdio.h> 15 #include <string.h> 16 17 /// @file 18 /// Implementation of CeedOperator preconditioning interfaces 19 20 /// ---------------------------------------------------------------------------- 21 /// CeedOperator Library Internal Preconditioning Functions 22 /// ---------------------------------------------------------------------------- 23 /// @addtogroup CeedOperatorDeveloper 24 /// @{ 25 26 /** 27 @brief Duplicate a `CeedQFunction` with a reference `Ceed` to fallback for advanced `CeedOperator` functionality 28 29 @param[in] fallback_ceed `Ceed` on which to create fallback `CeedQFunction` 30 @param[in] qf `CeedQFunction` to create fallback for 31 @param[out] qf_fallback Fallback `CeedQFunction` 32 33 @return An error code: 0 - success, otherwise - failure 34 35 @ref Developer 36 **/ 37 static int CeedQFunctionCreateFallback(Ceed fallback_ceed, CeedQFunction qf, CeedQFunction *qf_fallback) { 38 char *source_path_with_name = NULL; 39 CeedInt num_input_fields, num_output_fields; 40 CeedQFunctionField *input_fields, *output_fields; 41 42 // Check if NULL qf passed in 43 if (!qf) return CEED_ERROR_SUCCESS; 44 45 CeedDebug256(CeedQFunctionReturnCeed(qf), 1, "---------- CeedOperator Fallback ----------\n"); 46 CeedDebug(CeedQFunctionReturnCeed(qf), "Creating fallback CeedQFunction\n"); 47 48 if (qf->source_path) { 49 size_t path_len = strlen(qf->source_path), name_len = strlen(qf->kernel_name); 50 51 CeedCall(CeedCalloc(path_len + name_len + 2, &source_path_with_name)); 52 memcpy(source_path_with_name, qf->source_path, path_len); 53 memcpy(&source_path_with_name[path_len], ":", 1); 54 memcpy(&source_path_with_name[path_len + 1], qf->kernel_name, name_len); 55 } else if (qf->user_source) { 56 CeedCall(CeedStringAllocCopy(qf->user_source, &source_path_with_name)); 57 } else { 58 CeedCall(CeedCalloc(1, &source_path_with_name)); 59 } 60 61 { 62 CeedInt vec_length; 63 CeedQFunctionUser f; 64 65 CeedCall(CeedQFunctionGetVectorLength(qf, &vec_length)); 66 CeedCall(CeedQFunctionGetUserFunction(qf, &f)); 67 CeedCall(CeedQFunctionCreateInterior(fallback_ceed, vec_length, f, source_path_with_name, qf_fallback)); 68 } 69 { 70 CeedQFunctionContext ctx; 71 72 CeedCall(CeedQFunctionGetContext(qf, &ctx)); 73 CeedCall(CeedQFunctionSetContext(*qf_fallback, ctx)); 74 } 75 CeedCall(CeedQFunctionGetFields(qf, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 76 for (CeedInt i = 0; i < num_input_fields; i++) { 77 const char *field_name; 78 CeedInt size; 79 CeedEvalMode eval_mode; 80 81 CeedCall(CeedQFunctionFieldGetData(input_fields[i], &field_name, &size, &eval_mode)); 82 CeedCall(CeedQFunctionAddInput(*qf_fallback, field_name, size, eval_mode)); 83 } 84 for (CeedInt i = 0; i < num_output_fields; i++) { 85 const char *field_name; 86 CeedInt size; 87 CeedEvalMode eval_mode; 88 89 CeedCall(CeedQFunctionFieldGetData(output_fields[i], &field_name, &size, &eval_mode)); 90 CeedCall(CeedQFunctionAddOutput(*qf_fallback, field_name, size, eval_mode)); 91 } 92 CeedCall(CeedFree(&source_path_with_name)); 93 return CEED_ERROR_SUCCESS; 94 } 95 96 /** 97 @brief Duplicate a `CeedOperator` with a reference `Ceed` to fallback for advanced `CeedOperator` functionality 98 99 @param[in,out] op `CeedOperator` to create fallback for 100 101 @return An error code: 0 - success, otherwise - failure 102 103 @ref Developer 104 **/ 105 static int CeedOperatorCreateFallback(CeedOperator op) { 106 bool is_composite; 107 Ceed ceed, ceed_fallback; 108 CeedOperator op_fallback; 109 110 // Check not already created 111 if (op->op_fallback) return CEED_ERROR_SUCCESS; 112 113 // Fallback Ceed 114 CeedCall(CeedOperatorGetCeed(op, &ceed)); 115 CeedCall(CeedGetOperatorFallbackCeed(ceed, &ceed_fallback)); 116 CeedCall(CeedDestroy(&ceed)); 117 if (!ceed_fallback) return CEED_ERROR_SUCCESS; 118 119 CeedDebug256(CeedOperatorReturnCeed(op), 1, "---------- CeedOperator Fallback ----------\n"); 120 CeedDebug(CeedOperatorReturnCeed(op), "Creating fallback CeedOperator\n"); 121 122 // Clone Op 123 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 124 if (is_composite) { 125 CeedInt num_suboperators; 126 CeedOperator *sub_operators; 127 128 CeedCall(CeedCompositeOperatorCreate(ceed_fallback, &op_fallback)); 129 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 130 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 131 for (CeedInt i = 0; i < num_suboperators; i++) { 132 CeedOperator op_sub_fallback; 133 134 CeedCall(CeedOperatorGetFallback(sub_operators[i], &op_sub_fallback)); 135 CeedCall(CeedCompositeOperatorAddSub(op_fallback, op_sub_fallback)); 136 } 137 } else { 138 CeedInt num_input_fields, num_output_fields; 139 CeedQFunction qf_fallback = NULL, dqf_fallback = NULL, dqfT_fallback = NULL; 140 CeedOperatorField *input_fields, *output_fields; 141 142 CeedCall(CeedQFunctionCreateFallback(ceed_fallback, op->qf, &qf_fallback)); 143 CeedCall(CeedQFunctionCreateFallback(ceed_fallback, op->dqf, &dqf_fallback)); 144 CeedCall(CeedQFunctionCreateFallback(ceed_fallback, op->dqfT, &dqfT_fallback)); 145 CeedCall(CeedOperatorCreate(ceed_fallback, qf_fallback, dqf_fallback, dqfT_fallback, &op_fallback)); 146 CeedCall(CeedOperatorGetFields(op, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 147 for (CeedInt i = 0; i < num_input_fields; i++) { 148 const char *field_name; 149 CeedVector vec; 150 CeedElemRestriction rstr; 151 CeedBasis basis; 152 153 CeedCall(CeedOperatorFieldGetData(input_fields[i], &field_name, &rstr, &basis, &vec)); 154 CeedCall(CeedOperatorSetField(op_fallback, field_name, rstr, basis, vec)); 155 CeedCall(CeedVectorDestroy(&vec)); 156 CeedCall(CeedElemRestrictionDestroy(&rstr)); 157 CeedCall(CeedBasisDestroy(&basis)); 158 } 159 for (CeedInt i = 0; i < num_output_fields; i++) { 160 const char *field_name; 161 CeedVector vec; 162 CeedElemRestriction rstr; 163 CeedBasis basis; 164 165 CeedCall(CeedOperatorFieldGetData(output_fields[i], &field_name, &rstr, &basis, &vec)); 166 CeedCall(CeedOperatorSetField(op_fallback, field_name, rstr, basis, vec)); 167 CeedCall(CeedVectorDestroy(&vec)); 168 CeedCall(CeedElemRestrictionDestroy(&rstr)); 169 CeedCall(CeedBasisDestroy(&basis)); 170 } 171 { 172 CeedQFunctionAssemblyData data; 173 174 CeedCall(CeedOperatorGetQFunctionAssemblyData(op, &data)); 175 CeedCall(CeedQFunctionAssemblyDataReferenceCopy(data, &op_fallback->qf_assembled)); 176 } 177 // Cleanup 178 CeedCall(CeedQFunctionDestroy(&qf_fallback)); 179 CeedCall(CeedQFunctionDestroy(&dqf_fallback)); 180 CeedCall(CeedQFunctionDestroy(&dqfT_fallback)); 181 } 182 CeedCall(CeedOperatorSetName(op_fallback, op->name)); 183 CeedCall(CeedOperatorCheckReady(op_fallback)); 184 // Note: No ref-counting here so we don't get caught in a reference loop. 185 // The op holds the only reference to op_fallback and is responsible for deleting itself and op_fallback. 186 op->op_fallback = op_fallback; 187 op_fallback->op_fallback_parent = op; 188 CeedCall(CeedDestroy(&ceed_fallback)); 189 return CEED_ERROR_SUCCESS; 190 } 191 192 /** 193 @brief Core logic for assembling operator diagonal or point block diagonal 194 195 @param[in] op `CeedOperator` to assemble diagonal or point block diagonal 196 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 197 @param[in] is_point_block Boolean flag to assemble diagonal or point block diagonal 198 @param[out] assembled `CeedVector` to store assembled diagonal 199 200 @return An error code: 0 - success, otherwise - failure 201 202 @ref Developer 203 **/ 204 static inline int CeedSingleOperatorLinearAssembleAddDiagonal_Mesh(CeedOperator op, CeedRequest *request, const bool is_point_block, 205 CeedVector assembled) { 206 bool is_composite; 207 208 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 209 CeedCheck(!is_composite, CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, "Composite operator not supported"); 210 211 // Assemble QFunction 212 CeedInt layout_qf[3]; 213 const CeedScalar *assembled_qf_array; 214 CeedVector assembled_qf = NULL; 215 CeedElemRestriction assembled_elem_rstr = NULL; 216 217 CeedCall(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &assembled_elem_rstr, request)); 218 CeedCall(CeedElemRestrictionGetELayout(assembled_elem_rstr, layout_qf)); 219 CeedCall(CeedElemRestrictionDestroy(&assembled_elem_rstr)); 220 CeedCall(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_HOST, &assembled_qf_array)); 221 222 // Get assembly data 223 const CeedEvalMode **eval_modes_in, **eval_modes_out; 224 CeedInt num_active_bases_in, *num_eval_modes_in, num_active_bases_out, *num_eval_modes_out; 225 CeedSize **eval_mode_offsets_in, **eval_mode_offsets_out, num_output_components; 226 CeedBasis *active_bases_in, *active_bases_out; 227 CeedElemRestriction *active_elem_rstrs_in, *active_elem_rstrs_out; 228 CeedOperatorAssemblyData data; 229 230 CeedCall(CeedOperatorGetOperatorAssemblyData(op, &data)); 231 CeedCall(CeedOperatorAssemblyDataGetEvalModes(data, &num_active_bases_in, &num_eval_modes_in, &eval_modes_in, &eval_mode_offsets_in, 232 &num_active_bases_out, &num_eval_modes_out, &eval_modes_out, &eval_mode_offsets_out, 233 &num_output_components)); 234 CeedCall(CeedOperatorAssemblyDataGetBases(data, NULL, &active_bases_in, NULL, NULL, &active_bases_out, NULL)); 235 CeedCall(CeedOperatorAssemblyDataGetElemRestrictions(data, NULL, &active_elem_rstrs_in, NULL, &active_elem_rstrs_out)); 236 237 // Loop over all active bases (find matching input/output pairs) 238 for (CeedInt b = 0; b < CeedIntMin(num_active_bases_in, num_active_bases_out); b++) { 239 CeedInt b_in, b_out, num_elem, num_nodes, num_qpts, num_comp; 240 bool has_eval_none = false; 241 CeedScalar *elem_diag_array, *identity = NULL; 242 CeedVector elem_diag; 243 CeedElemRestriction diag_elem_rstr; 244 245 if (num_active_bases_in <= num_active_bases_out) { 246 b_in = b; 247 for (b_out = 0; b_out < num_active_bases_out; b_out++) { 248 if (active_bases_in[b_in] == active_bases_out[b_out]) { 249 break; 250 } 251 } 252 if (b_out == num_active_bases_out) { 253 continue; 254 } // No matching output basis found 255 } else { 256 b_out = b; 257 for (b_in = 0; b_in < num_active_bases_in; b_in++) { 258 if (active_bases_in[b_in] == active_bases_out[b_out]) { 259 break; 260 } 261 } 262 if (b_in == num_active_bases_in) { 263 continue; 264 } // No matching output basis found 265 } 266 CeedCheck(active_elem_rstrs_in[b_in] == active_elem_rstrs_out[b_out], CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, 267 "Cannot assemble operator diagonal with different input and output active element restrictions"); 268 269 // Assemble point block diagonal restriction, if needed 270 if (is_point_block) { 271 CeedCall(CeedOperatorCreateActivePointBlockRestriction(active_elem_rstrs_in[b_in], &diag_elem_rstr)); 272 } else { 273 CeedCall(CeedElemRestrictionCreateUnsignedCopy(active_elem_rstrs_in[b_in], &diag_elem_rstr)); 274 } 275 276 // Create diagonal vector 277 CeedCall(CeedElemRestrictionCreateVector(diag_elem_rstr, NULL, &elem_diag)); 278 279 // Assemble element operator diagonals 280 CeedCall(CeedVectorSetValue(elem_diag, 0.0)); 281 CeedCall(CeedVectorGetArray(elem_diag, CEED_MEM_HOST, &elem_diag_array)); 282 CeedCall(CeedElemRestrictionGetNumElements(diag_elem_rstr, &num_elem)); 283 CeedCall(CeedBasisGetNumNodes(active_bases_in[b_in], &num_nodes)); 284 CeedCall(CeedBasisGetNumComponents(active_bases_in[b_in], &num_comp)); 285 if (active_bases_in[b_in] == CEED_BASIS_NONE) num_qpts = num_nodes; 286 else CeedCall(CeedBasisGetNumQuadraturePoints(active_bases_in[b_in], &num_qpts)); 287 288 // Construct identity matrix for basis if required 289 for (CeedInt i = 0; i < num_eval_modes_in[b_in]; i++) { 290 has_eval_none = has_eval_none || (eval_modes_in[b_in][i] == CEED_EVAL_NONE); 291 } 292 for (CeedInt i = 0; i < num_eval_modes_out[b_out]; i++) { 293 has_eval_none = has_eval_none || (eval_modes_out[b_out][i] == CEED_EVAL_NONE); 294 } 295 if (has_eval_none) { 296 CeedCall(CeedCalloc(num_qpts * num_nodes, &identity)); 297 for (CeedInt i = 0; i < (num_nodes < num_qpts ? num_nodes : num_qpts); i++) identity[i * num_nodes + i] = 1.0; 298 } 299 300 // Compute the diagonal of B^T D B 301 // Each element 302 for (CeedSize e = 0; e < num_elem; e++) { 303 // Each basis eval mode pair 304 CeedInt d_out = 0, q_comp_out; 305 CeedEvalMode eval_mode_out_prev = CEED_EVAL_NONE; 306 307 for (CeedInt e_out = 0; e_out < num_eval_modes_out[b_out]; e_out++) { 308 CeedInt d_in = 0, q_comp_in; 309 const CeedScalar *B_t = NULL; 310 CeedEvalMode eval_mode_in_prev = CEED_EVAL_NONE; 311 312 CeedCall(CeedOperatorGetBasisPointer(active_bases_out[b_out], eval_modes_out[b_out][e_out], identity, &B_t)); 313 CeedCall(CeedBasisGetNumQuadratureComponents(active_bases_out[b_out], eval_modes_out[b_out][e_out], &q_comp_out)); 314 if (q_comp_out > 1) { 315 if (e_out == 0 || eval_modes_out[b_out][e_out] != eval_mode_out_prev) d_out = 0; 316 else B_t = &B_t[(++d_out) * num_qpts * num_nodes]; 317 } 318 eval_mode_out_prev = eval_modes_out[b_out][e_out]; 319 320 for (CeedInt e_in = 0; e_in < num_eval_modes_in[b_in]; e_in++) { 321 const CeedScalar *B = NULL; 322 323 CeedCall(CeedOperatorGetBasisPointer(active_bases_in[b_in], eval_modes_in[b_in][e_in], identity, &B)); 324 CeedCall(CeedBasisGetNumQuadratureComponents(active_bases_in[b_in], eval_modes_in[b_in][e_in], &q_comp_in)); 325 if (q_comp_in > 1) { 326 if (e_in == 0 || eval_modes_in[b_in][e_in] != eval_mode_in_prev) d_in = 0; 327 else B = &B[(++d_in) * num_qpts * num_nodes]; 328 } 329 eval_mode_in_prev = eval_modes_in[b_in][e_in]; 330 331 // Each component 332 for (CeedInt c_out = 0; c_out < num_comp; c_out++) { 333 // Each qpt/node pair 334 for (CeedInt q = 0; q < num_qpts; q++) { 335 if (is_point_block) { 336 // Point Block Diagonal 337 for (CeedInt c_in = 0; c_in < num_comp; c_in++) { 338 const CeedSize c_offset = 339 (eval_mode_offsets_in[b_in][e_in] + c_in) * num_output_components + eval_mode_offsets_out[b_out][e_out] + c_out; 340 const CeedScalar qf_value = assembled_qf_array[q * layout_qf[0] + c_offset * layout_qf[1] + e * layout_qf[2]]; 341 342 for (CeedInt n = 0; n < num_nodes; n++) { 343 elem_diag_array[((e * num_comp + c_out) * num_comp + c_in) * num_nodes + n] += 344 B_t[q * num_nodes + n] * qf_value * B[q * num_nodes + n]; 345 } 346 } 347 } else { 348 // Diagonal Only 349 const CeedInt c_offset = 350 (eval_mode_offsets_in[b_in][e_in] + c_out) * num_output_components + eval_mode_offsets_out[b_out][e_out] + c_out; 351 const CeedScalar qf_value = assembled_qf_array[q * layout_qf[0] + c_offset * layout_qf[1] + e * layout_qf[2]]; 352 353 for (CeedInt n = 0; n < num_nodes; n++) { 354 elem_diag_array[(e * num_comp + c_out) * num_nodes + n] += B_t[q * num_nodes + n] * qf_value * B[q * num_nodes + n]; 355 } 356 } 357 } 358 } 359 } 360 } 361 } 362 CeedCall(CeedVectorRestoreArray(elem_diag, &elem_diag_array)); 363 364 // Assemble local operator diagonal 365 CeedCall(CeedElemRestrictionApply(diag_elem_rstr, CEED_TRANSPOSE, elem_diag, assembled, request)); 366 367 // Cleanup 368 CeedCall(CeedElemRestrictionDestroy(&diag_elem_rstr)); 369 CeedCall(CeedVectorDestroy(&elem_diag)); 370 CeedCall(CeedFree(&identity)); 371 } 372 CeedCall(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 373 CeedCall(CeedVectorDestroy(&assembled_qf)); 374 return CEED_ERROR_SUCCESS; 375 } 376 377 /** 378 @brief Core logic for assembling operator diagonal or point block diagonal 379 380 @param[in] op `CeedOperator` to assemble diagonal or point block diagonal 381 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 382 @param[in] is_point_block Boolean flag to assemble diagonal or point block diagonal 383 @param[out] assembled `CeedVector` to store assembled diagonal 384 385 @return An error code: 0 - success, otherwise - failure 386 387 @ref Developer 388 **/ 389 static inline int CeedSingleOperatorLinearAssembleAddDiagonal(CeedOperator op, CeedRequest *request, const bool is_point_block, 390 CeedVector assembled) { 391 bool is_at_points; 392 393 CeedCall(CeedOperatorIsAtPoints(op, &is_at_points)); 394 CeedCheck(!is_at_points, CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, "AtPoints operator not supported"); 395 CeedCall(CeedSingleOperatorLinearAssembleAddDiagonal_Mesh(op, request, is_point_block, assembled)); 396 return CEED_ERROR_SUCCESS; 397 } 398 399 /** 400 @brief Core logic for assembling composite operator diagonal 401 402 @param[in] op `CeedOperator` to assemble point block diagonal 403 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 404 @param[in] is_point_block Boolean flag to assemble diagonal or point block diagonal 405 @param[out] assembled `CeedVector` to store assembled diagonal 406 407 @return An error code: 0 - success, otherwise - failure 408 409 @ref Developer 410 **/ 411 static inline int CeedCompositeOperatorLinearAssembleAddDiagonal(CeedOperator op, CeedRequest *request, const bool is_point_block, 412 CeedVector assembled) { 413 CeedInt num_sub; 414 CeedOperator *suboperators; 415 416 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_sub)); 417 CeedCall(CeedCompositeOperatorGetSubList(op, &suboperators)); 418 for (CeedInt i = 0; i < num_sub; i++) { 419 if (is_point_block) { 420 CeedCall(CeedOperatorLinearAssembleAddPointBlockDiagonal(suboperators[i], assembled, request)); 421 } else { 422 CeedCall(CeedOperatorLinearAssembleAddDiagonal(suboperators[i], assembled, request)); 423 } 424 } 425 return CEED_ERROR_SUCCESS; 426 } 427 428 /** 429 @brief Build nonzero pattern for non-composite CeedOperator`. 430 431 Users should generally use @ref CeedOperatorLinearAssembleSymbolic(). 432 433 @param[in] op `CeedOperator` to assemble nonzero pattern 434 @param[in] offset Offset for number of entries 435 @param[out] rows Row number for each entry 436 @param[out] cols Column number for each entry 437 438 @return An error code: 0 - success, otherwise - failure 439 440 @ref Developer 441 **/ 442 static int CeedSingleOperatorAssembleSymbolic(CeedOperator op, CeedInt offset, CeedInt *rows, CeedInt *cols) { 443 Ceed ceed; 444 bool is_composite; 445 CeedSize num_nodes_in, num_nodes_out, local_num_entries, count = 0; 446 CeedInt num_elem_in, elem_size_in, num_comp_in, layout_er_in[3]; 447 CeedInt num_elem_out, elem_size_out, num_comp_out, layout_er_out[3]; 448 CeedScalar *array; 449 const CeedScalar *elem_dof_a_in, *elem_dof_a_out; 450 CeedVector index_vec_in, index_vec_out, elem_dof_in, elem_dof_out; 451 CeedElemRestriction elem_rstr_in, elem_rstr_out, index_elem_rstr_in, index_elem_rstr_out; 452 453 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 454 CeedCheck(!is_composite, ceed, CEED_ERROR_UNSUPPORTED, "Composite operator not supported"); 455 CeedCall(CeedOperatorGetCeed(op, &ceed)); 456 457 CeedCall(CeedOperatorGetActiveVectorLengths(op, &num_nodes_in, &num_nodes_out)); 458 CeedCall(CeedOperatorGetActiveElemRestrictions(op, &elem_rstr_in, &elem_rstr_out)); 459 CeedCall(CeedElemRestrictionGetNumElements(elem_rstr_in, &num_elem_in)); 460 CeedCall(CeedElemRestrictionGetElementSize(elem_rstr_in, &elem_size_in)); 461 CeedCall(CeedElemRestrictionGetNumComponents(elem_rstr_in, &num_comp_in)); 462 CeedCall(CeedElemRestrictionGetELayout(elem_rstr_in, layout_er_in)); 463 464 // Determine elem_dof relation for input 465 CeedCall(CeedVectorCreate(ceed, num_nodes_in, &index_vec_in)); 466 CeedCall(CeedVectorGetArrayWrite(index_vec_in, CEED_MEM_HOST, &array)); 467 for (CeedSize i = 0; i < num_nodes_in; i++) array[i] = i; 468 CeedCall(CeedVectorRestoreArray(index_vec_in, &array)); 469 CeedCall(CeedVectorCreate(ceed, num_elem_in * elem_size_in * num_comp_in, &elem_dof_in)); 470 CeedCall(CeedVectorSetValue(elem_dof_in, 0.0)); 471 CeedCall(CeedElemRestrictionCreateUnorientedCopy(elem_rstr_in, &index_elem_rstr_in)); 472 CeedCall(CeedElemRestrictionApply(index_elem_rstr_in, CEED_NOTRANSPOSE, index_vec_in, elem_dof_in, CEED_REQUEST_IMMEDIATE)); 473 CeedCall(CeedVectorGetArrayRead(elem_dof_in, CEED_MEM_HOST, &elem_dof_a_in)); 474 CeedCall(CeedVectorDestroy(&index_vec_in)); 475 CeedCall(CeedElemRestrictionDestroy(&index_elem_rstr_in)); 476 477 if (elem_rstr_in != elem_rstr_out) { 478 CeedCall(CeedElemRestrictionGetNumElements(elem_rstr_out, &num_elem_out)); 479 CeedCheck(num_elem_in == num_elem_out, ceed, CEED_ERROR_UNSUPPORTED, 480 "Active input and output operator restrictions must have the same number of elements." 481 " Input has %" CeedInt_FMT " elements; output has %" CeedInt_FMT "elements.", 482 num_elem_in, num_elem_out); 483 CeedCall(CeedElemRestrictionGetElementSize(elem_rstr_out, &elem_size_out)); 484 CeedCall(CeedElemRestrictionGetNumComponents(elem_rstr_out, &num_comp_out)); 485 CeedCall(CeedElemRestrictionGetELayout(elem_rstr_out, layout_er_out)); 486 487 // Determine elem_dof relation for output 488 CeedCall(CeedVectorCreate(ceed, num_nodes_out, &index_vec_out)); 489 CeedCall(CeedVectorGetArrayWrite(index_vec_out, CEED_MEM_HOST, &array)); 490 for (CeedSize i = 0; i < num_nodes_out; i++) array[i] = i; 491 CeedCall(CeedVectorRestoreArray(index_vec_out, &array)); 492 CeedCall(CeedVectorCreate(ceed, num_elem_out * elem_size_out * num_comp_out, &elem_dof_out)); 493 CeedCall(CeedVectorSetValue(elem_dof_out, 0.0)); 494 CeedCall(CeedElemRestrictionCreateUnorientedCopy(elem_rstr_out, &index_elem_rstr_out)); 495 CeedCall(CeedElemRestrictionApply(index_elem_rstr_out, CEED_NOTRANSPOSE, index_vec_out, elem_dof_out, CEED_REQUEST_IMMEDIATE)); 496 CeedCall(CeedVectorGetArrayRead(elem_dof_out, CEED_MEM_HOST, &elem_dof_a_out)); 497 CeedCall(CeedVectorDestroy(&index_vec_out)); 498 CeedCall(CeedElemRestrictionDestroy(&index_elem_rstr_out)); 499 } else { 500 num_elem_out = num_elem_in; 501 elem_size_out = elem_size_in; 502 num_comp_out = num_comp_in; 503 layout_er_out[0] = layout_er_in[0]; 504 layout_er_out[1] = layout_er_in[1]; 505 layout_er_out[2] = layout_er_in[2]; 506 elem_dof_a_out = elem_dof_a_in; 507 } 508 local_num_entries = (CeedSize)elem_size_out * num_comp_out * elem_size_in * num_comp_in * num_elem_in; 509 510 // Determine i, j locations for element matrices 511 for (CeedInt e = 0; e < num_elem_in; e++) { 512 for (CeedInt comp_in = 0; comp_in < num_comp_in; comp_in++) { 513 for (CeedInt comp_out = 0; comp_out < num_comp_out; comp_out++) { 514 for (CeedInt i = 0; i < elem_size_out; i++) { 515 for (CeedInt j = 0; j < elem_size_in; j++) { 516 const CeedInt elem_dof_index_row = i * layout_er_out[0] + comp_out * layout_er_out[1] + e * layout_er_out[2]; 517 const CeedInt elem_dof_index_col = j * layout_er_in[0] + comp_in * layout_er_in[1] + e * layout_er_in[2]; 518 const CeedInt row = elem_dof_a_out[elem_dof_index_row]; 519 const CeedInt col = elem_dof_a_in[elem_dof_index_col]; 520 521 rows[offset + count] = row; 522 cols[offset + count] = col; 523 count++; 524 } 525 } 526 } 527 } 528 } 529 CeedCheck(count == local_num_entries, ceed, CEED_ERROR_MAJOR, "Error computing assembled entries"); 530 CeedCall(CeedVectorRestoreArrayRead(elem_dof_in, &elem_dof_a_in)); 531 CeedCall(CeedVectorDestroy(&elem_dof_in)); 532 if (elem_rstr_in != elem_rstr_out) { 533 CeedCall(CeedVectorRestoreArrayRead(elem_dof_out, &elem_dof_a_out)); 534 CeedCall(CeedVectorDestroy(&elem_dof_out)); 535 } 536 CeedCall(CeedElemRestrictionDestroy(&elem_rstr_in)); 537 CeedCall(CeedElemRestrictionDestroy(&elem_rstr_out)); 538 CeedCall(CeedDestroy(&ceed)); 539 return CEED_ERROR_SUCCESS; 540 } 541 542 /** 543 @brief Assemble nonzero entries for non-composite `CeedOperator`. 544 545 Users should generally use @ref CeedOperatorLinearAssemble(). 546 547 @param[in] op `CeedOperator` to assemble 548 @param[in] offset Offset for number of entries 549 @param[out] values Values to assemble into matrix 550 551 @return An error code: 0 - success, otherwise - failure 552 553 @ref Developer 554 **/ 555 static int CeedSingleOperatorAssemble(CeedOperator op, CeedInt offset, CeedVector values) { 556 bool is_composite; 557 558 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 559 CeedCheck(!is_composite, CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, "Composite operator not supported"); 560 561 // Early exit for empty operator 562 { 563 CeedInt num_elem = 0; 564 565 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 566 if (num_elem == 0) return CEED_ERROR_SUCCESS; 567 } 568 569 if (op->LinearAssembleSingle) { 570 // Backend version 571 CeedCall(op->LinearAssembleSingle(op, offset, values)); 572 return CEED_ERROR_SUCCESS; 573 } else { 574 // Operator fallback 575 CeedOperator op_fallback; 576 577 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 578 if (op_fallback) { 579 CeedCall(CeedSingleOperatorAssemble(op_fallback, offset, values)); 580 return CEED_ERROR_SUCCESS; 581 } 582 } 583 584 // Assemble QFunction 585 CeedInt layout_qf[3]; 586 const CeedScalar *assembled_qf_array; 587 CeedVector assembled_qf = NULL; 588 CeedElemRestriction assembled_elem_rstr = NULL; 589 590 CeedCall(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &assembled_elem_rstr, CEED_REQUEST_IMMEDIATE)); 591 CeedCall(CeedElemRestrictionGetELayout(assembled_elem_rstr, layout_qf)); 592 CeedCall(CeedElemRestrictionDestroy(&assembled_elem_rstr)); 593 CeedCall(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_HOST, &assembled_qf_array)); 594 595 // Get assembly data 596 CeedInt num_elem_in, elem_size_in, num_comp_in, num_qpts_in; 597 CeedInt num_elem_out, elem_size_out, num_comp_out, num_qpts_out; 598 CeedSize local_num_entries, count = 0; 599 const CeedEvalMode **eval_modes_in, **eval_modes_out; 600 CeedInt num_active_bases_in, *num_eval_modes_in, num_active_bases_out, *num_eval_modes_out; 601 CeedBasis *active_bases_in, *active_bases_out, basis_in, basis_out; 602 const CeedScalar **B_mats_in, **B_mats_out, *B_mat_in, *B_mat_out; 603 CeedElemRestriction elem_rstr_in, elem_rstr_out; 604 CeedRestrictionType elem_rstr_type_in, elem_rstr_type_out; 605 const bool *elem_rstr_orients_in = NULL, *elem_rstr_orients_out = NULL; 606 const CeedInt8 *elem_rstr_curl_orients_in = NULL, *elem_rstr_curl_orients_out = NULL; 607 CeedOperatorAssemblyData data; 608 609 CeedCall(CeedOperatorGetOperatorAssemblyData(op, &data)); 610 CeedCall(CeedOperatorAssemblyDataGetEvalModes(data, &num_active_bases_in, &num_eval_modes_in, &eval_modes_in, NULL, &num_active_bases_out, 611 &num_eval_modes_out, &eval_modes_out, NULL, NULL)); 612 613 CeedCheck(num_active_bases_in == 1 && num_active_bases_out == 1, CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, 614 "Cannot assemble operator with multiple active bases"); 615 CeedCheck(num_eval_modes_in[0] > 0 && num_eval_modes_out[0] > 0, CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, 616 "Cannot assemble operator without inputs/outputs"); 617 618 CeedCall(CeedOperatorAssemblyDataGetBases(data, NULL, &active_bases_in, &B_mats_in, NULL, &active_bases_out, &B_mats_out)); 619 CeedCall(CeedOperatorGetActiveElemRestrictions(op, &elem_rstr_in, &elem_rstr_out)); 620 basis_in = active_bases_in[0]; 621 basis_out = active_bases_out[0]; 622 B_mat_in = B_mats_in[0]; 623 B_mat_out = B_mats_out[0]; 624 625 CeedCall(CeedElemRestrictionGetNumElements(elem_rstr_in, &num_elem_in)); 626 CeedCall(CeedElemRestrictionGetElementSize(elem_rstr_in, &elem_size_in)); 627 CeedCall(CeedElemRestrictionGetNumComponents(elem_rstr_in, &num_comp_in)); 628 if (basis_in == CEED_BASIS_NONE) num_qpts_in = elem_size_in; 629 else CeedCall(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts_in)); 630 631 CeedCall(CeedElemRestrictionGetType(elem_rstr_in, &elem_rstr_type_in)); 632 if (elem_rstr_type_in == CEED_RESTRICTION_ORIENTED) { 633 CeedCall(CeedElemRestrictionGetOrientations(elem_rstr_in, CEED_MEM_HOST, &elem_rstr_orients_in)); 634 } else if (elem_rstr_type_in == CEED_RESTRICTION_CURL_ORIENTED) { 635 CeedCall(CeedElemRestrictionGetCurlOrientations(elem_rstr_in, CEED_MEM_HOST, &elem_rstr_curl_orients_in)); 636 } 637 638 if (elem_rstr_in != elem_rstr_out) { 639 CeedCall(CeedElemRestrictionGetNumElements(elem_rstr_out, &num_elem_out)); 640 CeedCheck(num_elem_in == num_elem_out, CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, 641 "Active input and output operator restrictions must have the same number of elements." 642 " Input has %" CeedInt_FMT " elements; output has %" CeedInt_FMT "elements.", 643 num_elem_in, num_elem_out); 644 CeedCall(CeedElemRestrictionGetElementSize(elem_rstr_out, &elem_size_out)); 645 CeedCall(CeedElemRestrictionGetNumComponents(elem_rstr_out, &num_comp_out)); 646 if (basis_out == CEED_BASIS_NONE) num_qpts_out = elem_size_out; 647 else CeedCall(CeedBasisGetNumQuadraturePoints(basis_out, &num_qpts_out)); 648 CeedCheck(num_qpts_in == num_qpts_out, CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, 649 "Active input and output bases must have the same number of quadrature points." 650 " Input has %" CeedInt_FMT " points; output has %" CeedInt_FMT "points.", 651 num_qpts_in, num_qpts_out); 652 653 CeedCall(CeedElemRestrictionGetType(elem_rstr_out, &elem_rstr_type_out)); 654 if (elem_rstr_type_out == CEED_RESTRICTION_ORIENTED) { 655 CeedCall(CeedElemRestrictionGetOrientations(elem_rstr_out, CEED_MEM_HOST, &elem_rstr_orients_out)); 656 } else if (elem_rstr_type_out == CEED_RESTRICTION_CURL_ORIENTED) { 657 CeedCall(CeedElemRestrictionGetCurlOrientations(elem_rstr_out, CEED_MEM_HOST, &elem_rstr_curl_orients_out)); 658 } 659 } else { 660 num_elem_out = num_elem_in; 661 elem_size_out = elem_size_in; 662 num_comp_out = num_comp_in; 663 num_qpts_out = num_qpts_in; 664 665 elem_rstr_orients_out = elem_rstr_orients_in; 666 elem_rstr_curl_orients_out = elem_rstr_curl_orients_in; 667 } 668 local_num_entries = (CeedSize)elem_size_out * num_comp_out * elem_size_in * num_comp_in * num_elem_in; 669 670 // Loop over elements and put in data structure 671 // We store B_mat_in, B_mat_out, BTD, elem_mat in row-major order 672 CeedTensorContract contract; 673 CeedScalar *vals, *BTD_mat = NULL, *elem_mat = NULL, *elem_mat_b = NULL; 674 675 CeedCall(CeedBasisGetTensorContract(basis_in, &contract)); 676 CeedCall(CeedCalloc(elem_size_out * num_qpts_in * num_eval_modes_in[0], &BTD_mat)); 677 CeedCall(CeedCalloc(elem_size_out * elem_size_in, &elem_mat)); 678 if (elem_rstr_curl_orients_in || elem_rstr_curl_orients_out) CeedCall(CeedCalloc(elem_size_out * elem_size_in, &elem_mat_b)); 679 680 CeedCall(CeedVectorGetArray(values, CEED_MEM_HOST, &vals)); 681 for (CeedSize e = 0; e < num_elem_in; e++) { 682 for (CeedInt comp_in = 0; comp_in < num_comp_in; comp_in++) { 683 for (CeedInt comp_out = 0; comp_out < num_comp_out; comp_out++) { 684 // Compute B^T*D 685 for (CeedSize n = 0; n < elem_size_out; n++) { 686 for (CeedSize q = 0; q < num_qpts_in; q++) { 687 for (CeedInt e_in = 0; e_in < num_eval_modes_in[0]; e_in++) { 688 const CeedSize btd_index = n * (num_qpts_in * num_eval_modes_in[0]) + q * num_eval_modes_in[0] + e_in; 689 CeedScalar sum = 0.0; 690 691 for (CeedInt e_out = 0; e_out < num_eval_modes_out[0]; e_out++) { 692 const CeedSize b_out_index = (q * num_eval_modes_out[0] + e_out) * elem_size_out + n; 693 const CeedSize eval_mode_index = ((e_in * num_comp_in + comp_in) * num_eval_modes_out[0] + e_out) * num_comp_out + comp_out; 694 const CeedSize qf_index = q * layout_qf[0] + eval_mode_index * layout_qf[1] + e * layout_qf[2]; 695 696 sum += B_mat_out[b_out_index] * assembled_qf_array[qf_index]; 697 } 698 BTD_mat[btd_index] = sum; 699 } 700 } 701 } 702 703 // Form element matrix itself (for each block component) 704 if (contract) { 705 CeedCall(CeedTensorContractApply(contract, 1, num_qpts_in * num_eval_modes_in[0], elem_size_in, elem_size_out, BTD_mat, CEED_NOTRANSPOSE, 706 false, B_mat_in, elem_mat)); 707 } else { 708 Ceed ceed; 709 710 CeedCall(CeedOperatorGetCeed(op, &ceed)); 711 CeedCall(CeedMatrixMatrixMultiply(ceed, BTD_mat, B_mat_in, elem_mat, elem_size_out, elem_size_in, num_qpts_in * num_eval_modes_in[0])); 712 CeedCall(CeedDestroy(&ceed)); 713 } 714 715 // Transform the element matrix if required 716 if (elem_rstr_orients_out) { 717 const bool *elem_orients = &elem_rstr_orients_out[e * elem_size_out]; 718 719 for (CeedInt i = 0; i < elem_size_out; i++) { 720 const double orient = elem_orients[i] ? -1.0 : 1.0; 721 722 for (CeedInt j = 0; j < elem_size_in; j++) { 723 elem_mat[i * elem_size_in + j] *= orient; 724 } 725 } 726 } else if (elem_rstr_curl_orients_out) { 727 const CeedInt8 *elem_curl_orients = &elem_rstr_curl_orients_out[e * 3 * elem_size_out]; 728 729 // T^T*(B^T*D*B) 730 memcpy(elem_mat_b, elem_mat, elem_size_out * elem_size_in * sizeof(CeedScalar)); 731 for (CeedInt i = 0; i < elem_size_out; i++) { 732 for (CeedInt j = 0; j < elem_size_in; j++) { 733 elem_mat[i * elem_size_in + j] = elem_mat_b[i * elem_size_in + j] * elem_curl_orients[3 * i + 1] + 734 (i > 0 ? elem_mat_b[(i - 1) * elem_size_in + j] * elem_curl_orients[3 * i - 1] : 0.0) + 735 (i < elem_size_out - 1 ? elem_mat_b[(i + 1) * elem_size_in + j] * elem_curl_orients[3 * i + 3] : 0.0); 736 } 737 } 738 } 739 if (elem_rstr_orients_in) { 740 const bool *elem_orients = &elem_rstr_orients_in[e * elem_size_in]; 741 742 for (CeedInt i = 0; i < elem_size_out; i++) { 743 for (CeedInt j = 0; j < elem_size_in; j++) { 744 elem_mat[i * elem_size_in + j] *= elem_orients[j] ? -1.0 : 1.0; 745 } 746 } 747 } else if (elem_rstr_curl_orients_in) { 748 const CeedInt8 *elem_curl_orients = &elem_rstr_curl_orients_in[e * 3 * elem_size_in]; 749 750 // (B^T*D*B)*T 751 memcpy(elem_mat_b, elem_mat, elem_size_out * elem_size_in * sizeof(CeedScalar)); 752 for (CeedInt i = 0; i < elem_size_out; i++) { 753 for (CeedInt j = 0; j < elem_size_in; j++) { 754 elem_mat[i * elem_size_in + j] = elem_mat_b[i * elem_size_in + j] * elem_curl_orients[3 * j + 1] + 755 (j > 0 ? elem_mat_b[i * elem_size_in + j - 1] * elem_curl_orients[3 * j - 1] : 0.0) + 756 (j < elem_size_in - 1 ? elem_mat_b[i * elem_size_in + j + 1] * elem_curl_orients[3 * j + 3] : 0.0); 757 } 758 } 759 } 760 761 // Put element matrix in coordinate data structure 762 for (CeedInt i = 0; i < elem_size_out; i++) { 763 for (CeedInt j = 0; j < elem_size_in; j++) { 764 vals[offset + count] = elem_mat[i * elem_size_in + j]; 765 count++; 766 } 767 } 768 } 769 } 770 } 771 CeedCheck(count == local_num_entries, CeedOperatorReturnCeed(op), CEED_ERROR_MAJOR, "Error computing entries"); 772 CeedCall(CeedVectorRestoreArray(values, &vals)); 773 774 // Cleanup 775 CeedCall(CeedFree(&BTD_mat)); 776 CeedCall(CeedFree(&elem_mat)); 777 CeedCall(CeedFree(&elem_mat_b)); 778 if (elem_rstr_type_in == CEED_RESTRICTION_ORIENTED) { 779 CeedCall(CeedElemRestrictionRestoreOrientations(elem_rstr_in, &elem_rstr_orients_in)); 780 } else if (elem_rstr_type_in == CEED_RESTRICTION_CURL_ORIENTED) { 781 CeedCall(CeedElemRestrictionRestoreCurlOrientations(elem_rstr_in, &elem_rstr_curl_orients_in)); 782 } 783 if (elem_rstr_in != elem_rstr_out) { 784 if (elem_rstr_type_out == CEED_RESTRICTION_ORIENTED) { 785 CeedCall(CeedElemRestrictionRestoreOrientations(elem_rstr_out, &elem_rstr_orients_out)); 786 } else if (elem_rstr_type_out == CEED_RESTRICTION_CURL_ORIENTED) { 787 CeedCall(CeedElemRestrictionRestoreCurlOrientations(elem_rstr_out, &elem_rstr_curl_orients_out)); 788 } 789 } 790 CeedCall(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 791 CeedCall(CeedVectorDestroy(&assembled_qf)); 792 CeedCall(CeedElemRestrictionDestroy(&elem_rstr_in)); 793 CeedCall(CeedElemRestrictionDestroy(&elem_rstr_out)); 794 return CEED_ERROR_SUCCESS; 795 } 796 797 /** 798 @brief Count number of entries for assembled `CeedOperator` 799 800 @param[in] op `CeedOperator` to assemble 801 @param[out] num_entries Number of entries in assembled representation 802 803 @return An error code: 0 - success, otherwise - failure 804 805 @ref Utility 806 **/ 807 static int CeedSingleOperatorAssemblyCountEntries(CeedOperator op, CeedSize *num_entries) { 808 bool is_composite; 809 CeedInt num_elem_in, elem_size_in, num_comp_in, num_elem_out, elem_size_out, num_comp_out; 810 CeedElemRestriction rstr_in, rstr_out; 811 812 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 813 CeedCheck(!is_composite, CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, "Composite operator not supported"); 814 815 CeedCall(CeedOperatorGetActiveElemRestrictions(op, &rstr_in, &rstr_out)); 816 CeedCall(CeedElemRestrictionGetNumElements(rstr_in, &num_elem_in)); 817 CeedCall(CeedElemRestrictionGetElementSize(rstr_in, &elem_size_in)); 818 CeedCall(CeedElemRestrictionGetNumComponents(rstr_in, &num_comp_in)); 819 if (rstr_in != rstr_out) { 820 CeedCall(CeedElemRestrictionGetNumElements(rstr_out, &num_elem_out)); 821 CeedCheck(num_elem_in == num_elem_out, CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, 822 "Active input and output operator restrictions must have the same number of elements." 823 " Input has %" CeedInt_FMT " elements; output has %" CeedInt_FMT "elements.", 824 num_elem_in, num_elem_out); 825 CeedCall(CeedElemRestrictionGetElementSize(rstr_out, &elem_size_out)); 826 CeedCall(CeedElemRestrictionGetNumComponents(rstr_out, &num_comp_out)); 827 } else { 828 num_elem_out = num_elem_in; 829 elem_size_out = elem_size_in; 830 num_comp_out = num_comp_in; 831 } 832 CeedCall(CeedElemRestrictionDestroy(&rstr_in)); 833 CeedCall(CeedElemRestrictionDestroy(&rstr_out)); 834 *num_entries = (CeedSize)elem_size_in * num_comp_in * elem_size_out * num_comp_out * num_elem_in; 835 return CEED_ERROR_SUCCESS; 836 } 837 838 /** 839 @brief Common code for creating a multigrid coarse `CeedOperator` and level transfer `CeedOperator` for a `CeedOperator` 840 841 @param[in] op_fine Fine grid `CeedOperator` 842 @param[in] p_mult_fine L-vector multiplicity in parallel gather/scatter, or `NULL` if not creating prolongation/restriction `CeedOperator` 843 @param[in] rstr_coarse Coarse grid `CeedElemRestriction` 844 @param[in] basis_coarse Coarse grid active vector `CeedBasis` 845 @param[in] basis_c_to_f `CeedBasis` for coarse to fine interpolation, or `NULL` if not creating prolongation/restriction operators 846 @param[out] op_coarse Coarse grid `CeedOperator` 847 @param[out] op_prolong Coarse to fine `CeedOperator`, or `NULL` 848 @param[out] op_restrict Fine to coarse `CeedOperator`, or `NULL` 849 850 @return An error code: 0 - success, otherwise - failure 851 852 @ref Developer 853 **/ 854 static int CeedSingleOperatorMultigridLevel(CeedOperator op_fine, CeedVector p_mult_fine, CeedElemRestriction rstr_coarse, CeedBasis basis_coarse, 855 CeedBasis basis_c_to_f, CeedOperator *op_coarse, CeedOperator *op_prolong, CeedOperator *op_restrict) { 856 bool is_composite; 857 Ceed ceed; 858 CeedInt num_comp, num_input_fields, num_output_fields; 859 CeedVector mult_vec = NULL; 860 CeedElemRestriction rstr_p_mult_fine = NULL, rstr_fine = NULL; 861 CeedOperatorField *input_fields, *output_fields; 862 863 CeedCall(CeedOperatorGetCeed(op_fine, &ceed)); 864 865 // Check for composite operator 866 CeedCall(CeedOperatorIsComposite(op_fine, &is_composite)); 867 CeedCheck(!is_composite, ceed, CEED_ERROR_UNSUPPORTED, "Automatic multigrid setup for composite operators not supported"); 868 869 // Coarse Grid 870 CeedCall(CeedOperatorCreate(ceed, op_fine->qf, op_fine->dqf, op_fine->dqfT, op_coarse)); 871 CeedCall(CeedOperatorGetFields(op_fine, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 872 // -- Clone input fields 873 for (CeedInt i = 0; i < num_input_fields; i++) { 874 const char *field_name; 875 CeedVector vec; 876 CeedElemRestriction rstr = NULL; 877 CeedBasis basis = NULL; 878 879 CeedCall(CeedOperatorFieldGetName(input_fields[i], &field_name)); 880 CeedCall(CeedOperatorFieldGetVector(input_fields[i], &vec)); 881 if (vec == CEED_VECTOR_ACTIVE) { 882 CeedCall(CeedElemRestrictionReferenceCopy(rstr_coarse, &rstr)); 883 CeedCall(CeedBasisReferenceCopy(basis_coarse, &basis)); 884 if (!rstr_fine) CeedCall(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr_fine)); 885 } else { 886 CeedCall(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr)); 887 CeedCall(CeedOperatorFieldGetBasis(input_fields[i], &basis)); 888 } 889 CeedCall(CeedOperatorSetField(*op_coarse, field_name, rstr, basis, vec)); 890 CeedCall(CeedVectorDestroy(&vec)); 891 CeedCall(CeedElemRestrictionDestroy(&rstr)); 892 CeedCall(CeedBasisDestroy(&basis)); 893 } 894 // -- Clone output fields 895 for (CeedInt i = 0; i < num_output_fields; i++) { 896 const char *field_name; 897 CeedVector vec; 898 CeedElemRestriction rstr = NULL; 899 CeedBasis basis = NULL; 900 901 CeedCall(CeedOperatorFieldGetName(output_fields[i], &field_name)); 902 CeedCall(CeedOperatorFieldGetVector(output_fields[i], &vec)); 903 if (vec == CEED_VECTOR_ACTIVE) { 904 CeedCall(CeedElemRestrictionReferenceCopy(rstr_coarse, &rstr)); 905 CeedCall(CeedBasisReferenceCopy(basis_coarse, &basis)); 906 if (!rstr_fine) CeedCall(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr_fine)); 907 } else { 908 CeedCall(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr)); 909 CeedCall(CeedOperatorFieldGetBasis(output_fields[i], &basis)); 910 } 911 CeedCall(CeedOperatorSetField(*op_coarse, field_name, rstr, basis, vec)); 912 CeedCall(CeedVectorDestroy(&vec)); 913 CeedCall(CeedElemRestrictionDestroy(&rstr)); 914 CeedCall(CeedBasisDestroy(&basis)); 915 } 916 // -- Clone QFunctionAssemblyData 917 { 918 CeedQFunctionAssemblyData fine_data; 919 920 CeedCall(CeedOperatorGetQFunctionAssemblyData(op_fine, &fine_data)); 921 CeedCall(CeedQFunctionAssemblyDataReferenceCopy(fine_data, &(*op_coarse)->qf_assembled)); 922 } 923 924 // Multiplicity vector 925 if (op_restrict || op_prolong) { 926 CeedVector mult_e_vec; 927 CeedRestrictionType rstr_type; 928 929 CeedCall(CeedElemRestrictionGetType(rstr_fine, &rstr_type)); 930 CeedCheck(rstr_type != CEED_RESTRICTION_CURL_ORIENTED, ceed, CEED_ERROR_UNSUPPORTED, 931 "Element restrictions created with CeedElemRestrictionCreateCurlOriented are not supported"); 932 CeedCheck(p_mult_fine, ceed, CEED_ERROR_INCOMPATIBLE, "Prolongation or restriction operator creation requires fine grid multiplicity vector"); 933 CeedCall(CeedElemRestrictionCreateUnsignedCopy(rstr_fine, &rstr_p_mult_fine)); 934 CeedCall(CeedElemRestrictionCreateVector(rstr_fine, &mult_vec, &mult_e_vec)); 935 CeedCall(CeedVectorSetValue(mult_e_vec, 0.0)); 936 CeedCall(CeedElemRestrictionApply(rstr_p_mult_fine, CEED_NOTRANSPOSE, p_mult_fine, mult_e_vec, CEED_REQUEST_IMMEDIATE)); 937 CeedCall(CeedVectorSetValue(mult_vec, 0.0)); 938 CeedCall(CeedElemRestrictionApply(rstr_p_mult_fine, CEED_TRANSPOSE, mult_e_vec, mult_vec, CEED_REQUEST_IMMEDIATE)); 939 CeedCall(CeedVectorDestroy(&mult_e_vec)); 940 CeedCall(CeedVectorReciprocal(mult_vec)); 941 } 942 943 // Clone name 944 bool has_name = op_fine->name; 945 size_t name_len = op_fine->name ? strlen(op_fine->name) : 0; 946 CeedCall(CeedOperatorSetName(*op_coarse, op_fine->name)); 947 948 // Check that coarse to fine basis is provided if prolong/restrict operators are requested 949 CeedCheck(basis_c_to_f || (!op_restrict && !op_prolong), ceed, CEED_ERROR_INCOMPATIBLE, 950 "Prolongation or restriction operator creation requires coarse-to-fine basis"); 951 952 // Restriction/Prolongation Operators 953 CeedCall(CeedBasisGetNumComponents(basis_coarse, &num_comp)); 954 955 // Restriction 956 if (op_restrict) { 957 CeedInt *num_comp_r_data; 958 CeedQFunctionContext ctx_r; 959 CeedQFunction qf_restrict; 960 961 CeedCall(CeedQFunctionCreateInteriorByName(ceed, "Scale", &qf_restrict)); 962 CeedCall(CeedCalloc(1, &num_comp_r_data)); 963 num_comp_r_data[0] = num_comp; 964 CeedCall(CeedQFunctionContextCreate(ceed, &ctx_r)); 965 CeedCall(CeedQFunctionContextSetData(ctx_r, CEED_MEM_HOST, CEED_OWN_POINTER, sizeof(*num_comp_r_data), num_comp_r_data)); 966 CeedCall(CeedQFunctionSetContext(qf_restrict, ctx_r)); 967 CeedCall(CeedQFunctionContextDestroy(&ctx_r)); 968 CeedCall(CeedQFunctionAddInput(qf_restrict, "input", num_comp, CEED_EVAL_NONE)); 969 CeedCall(CeedQFunctionAddInput(qf_restrict, "scale", num_comp, CEED_EVAL_NONE)); 970 CeedCall(CeedQFunctionAddOutput(qf_restrict, "output", num_comp, CEED_EVAL_INTERP)); 971 CeedCall(CeedQFunctionSetUserFlopsEstimate(qf_restrict, num_comp)); 972 973 CeedCall(CeedOperatorCreate(ceed, qf_restrict, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, op_restrict)); 974 CeedCall(CeedOperatorSetField(*op_restrict, "input", rstr_fine, CEED_BASIS_NONE, CEED_VECTOR_ACTIVE)); 975 CeedCall(CeedOperatorSetField(*op_restrict, "scale", rstr_p_mult_fine, CEED_BASIS_NONE, mult_vec)); 976 CeedCall(CeedOperatorSetField(*op_restrict, "output", rstr_coarse, basis_c_to_f, CEED_VECTOR_ACTIVE)); 977 978 // Set name 979 char *restriction_name; 980 981 CeedCall(CeedCalloc(17 + name_len, &restriction_name)); 982 sprintf(restriction_name, "restriction%s%s", has_name ? " for " : "", has_name ? op_fine->name : ""); 983 CeedCall(CeedOperatorSetName(*op_restrict, restriction_name)); 984 CeedCall(CeedFree(&restriction_name)); 985 986 // Check 987 CeedCall(CeedOperatorCheckReady(*op_restrict)); 988 989 // Cleanup 990 CeedCall(CeedQFunctionDestroy(&qf_restrict)); 991 } 992 993 // Prolongation 994 if (op_prolong) { 995 CeedInt *num_comp_p_data; 996 CeedQFunctionContext ctx_p; 997 CeedQFunction qf_prolong; 998 999 CeedCall(CeedQFunctionCreateInteriorByName(ceed, "Scale", &qf_prolong)); 1000 CeedCall(CeedCalloc(1, &num_comp_p_data)); 1001 num_comp_p_data[0] = num_comp; 1002 CeedCall(CeedQFunctionContextCreate(ceed, &ctx_p)); 1003 CeedCall(CeedQFunctionContextSetData(ctx_p, CEED_MEM_HOST, CEED_OWN_POINTER, sizeof(*num_comp_p_data), num_comp_p_data)); 1004 CeedCall(CeedQFunctionSetContext(qf_prolong, ctx_p)); 1005 CeedCall(CeedQFunctionContextDestroy(&ctx_p)); 1006 CeedCall(CeedQFunctionAddInput(qf_prolong, "input", num_comp, CEED_EVAL_INTERP)); 1007 CeedCall(CeedQFunctionAddInput(qf_prolong, "scale", num_comp, CEED_EVAL_NONE)); 1008 CeedCall(CeedQFunctionAddOutput(qf_prolong, "output", num_comp, CEED_EVAL_NONE)); 1009 CeedCall(CeedQFunctionSetUserFlopsEstimate(qf_prolong, num_comp)); 1010 1011 CeedCall(CeedOperatorCreate(ceed, qf_prolong, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, op_prolong)); 1012 CeedCall(CeedOperatorSetField(*op_prolong, "input", rstr_coarse, basis_c_to_f, CEED_VECTOR_ACTIVE)); 1013 CeedCall(CeedOperatorSetField(*op_prolong, "scale", rstr_p_mult_fine, CEED_BASIS_NONE, mult_vec)); 1014 CeedCall(CeedOperatorSetField(*op_prolong, "output", rstr_fine, CEED_BASIS_NONE, CEED_VECTOR_ACTIVE)); 1015 1016 // Set name 1017 char *prolongation_name; 1018 1019 CeedCall(CeedCalloc(18 + name_len, &prolongation_name)); 1020 sprintf(prolongation_name, "prolongation%s%s", has_name ? " for " : "", has_name ? op_fine->name : ""); 1021 CeedCall(CeedOperatorSetName(*op_prolong, prolongation_name)); 1022 CeedCall(CeedFree(&prolongation_name)); 1023 1024 // Check 1025 CeedCall(CeedOperatorCheckReady(*op_prolong)); 1026 1027 // Cleanup 1028 CeedCall(CeedQFunctionDestroy(&qf_prolong)); 1029 } 1030 1031 // Check 1032 CeedCall(CeedOperatorCheckReady(*op_coarse)); 1033 1034 // Cleanup 1035 CeedCall(CeedDestroy(&ceed)); 1036 CeedCall(CeedVectorDestroy(&mult_vec)); 1037 CeedCall(CeedElemRestrictionDestroy(&rstr_fine)); 1038 CeedCall(CeedElemRestrictionDestroy(&rstr_p_mult_fine)); 1039 CeedCall(CeedBasisDestroy(&basis_c_to_f)); 1040 return CEED_ERROR_SUCCESS; 1041 } 1042 1043 /** 1044 @brief Build 1D mass matrix and Laplacian with perturbation 1045 1046 @param[in] interp_1d Interpolation matrix in one dimension 1047 @param[in] grad_1d Gradient matrix in one dimension 1048 @param[in] q_weight_1d Quadrature weights in one dimension 1049 @param[in] P_1d Number of basis nodes in one dimension 1050 @param[in] Q_1d Number of quadrature points in one dimension 1051 @param[in] dim Dimension of basis 1052 @param[out] mass Assembled mass matrix in one dimension 1053 @param[out] laplace Assembled perturbed Laplacian in one dimension 1054 1055 @return An error code: 0 - success, otherwise - failure 1056 1057 @ref Developer 1058 **/ 1059 CeedPragmaOptimizeOff 1060 static int CeedBuildMassLaplace(const CeedScalar *interp_1d, const CeedScalar *grad_1d, const CeedScalar *q_weight_1d, CeedInt P_1d, CeedInt Q_1d, 1061 CeedInt dim, CeedScalar *mass, CeedScalar *laplace) { 1062 for (CeedInt i = 0; i < P_1d; i++) { 1063 for (CeedInt j = 0; j < P_1d; j++) { 1064 CeedScalar sum = 0.0; 1065 for (CeedInt k = 0; k < Q_1d; k++) sum += interp_1d[k * P_1d + i] * q_weight_1d[k] * interp_1d[k * P_1d + j]; 1066 mass[i + j * P_1d] = sum; 1067 } 1068 } 1069 // -- Laplacian 1070 for (CeedInt i = 0; i < P_1d; i++) { 1071 for (CeedInt j = 0; j < P_1d; j++) { 1072 CeedScalar sum = 0.0; 1073 1074 for (CeedInt k = 0; k < Q_1d; k++) sum += grad_1d[k * P_1d + i] * q_weight_1d[k] * grad_1d[k * P_1d + j]; 1075 laplace[i + j * P_1d] = sum; 1076 } 1077 } 1078 CeedScalar perturbation = dim > 2 ? 1e-6 : 1e-4; 1079 for (CeedInt i = 0; i < P_1d; i++) laplace[i + P_1d * i] += perturbation; 1080 return CEED_ERROR_SUCCESS; 1081 } 1082 CeedPragmaOptimizeOn 1083 1084 /// @} 1085 1086 /// ---------------------------------------------------------------------------- 1087 /// CeedOperator Backend API 1088 /// ---------------------------------------------------------------------------- 1089 /// @addtogroup CeedOperatorBackend 1090 /// @{ 1091 1092 /** 1093 @brief Select correct basis matrix pointer based on @ref CeedEvalMode 1094 1095 @param[in] basis `CeedBasis` from which to get the basis matrix 1096 @param[in] eval_mode Current basis evaluation mode 1097 @param[in] identity Pointer to identity matrix 1098 @param[out] basis_ptr `CeedBasis` pointer to set 1099 1100 @ref Backend 1101 **/ 1102 int CeedOperatorGetBasisPointer(CeedBasis basis, CeedEvalMode eval_mode, const CeedScalar *identity, const CeedScalar **basis_ptr) { 1103 switch (eval_mode) { 1104 case CEED_EVAL_NONE: 1105 *basis_ptr = identity; 1106 break; 1107 case CEED_EVAL_INTERP: 1108 CeedCall(CeedBasisGetInterp(basis, basis_ptr)); 1109 break; 1110 case CEED_EVAL_GRAD: 1111 CeedCall(CeedBasisGetGrad(basis, basis_ptr)); 1112 break; 1113 case CEED_EVAL_DIV: 1114 CeedCall(CeedBasisGetDiv(basis, basis_ptr)); 1115 break; 1116 case CEED_EVAL_CURL: 1117 CeedCall(CeedBasisGetCurl(basis, basis_ptr)); 1118 break; 1119 case CEED_EVAL_WEIGHT: 1120 break; // Caught by QF Assembly 1121 } 1122 assert(*basis_ptr != NULL); 1123 return CEED_ERROR_SUCCESS; 1124 } 1125 1126 /** 1127 @brief Create point block restriction for active `CeedOperatorField` 1128 1129 @param[in] rstr Original `CeedElemRestriction` for active field 1130 @param[out] point_block_rstr Address of the variable where the newly created `CeedElemRestriction` will be stored 1131 1132 @return An error code: 0 - success, otherwise - failure 1133 1134 @ref Backend 1135 **/ 1136 int CeedOperatorCreateActivePointBlockRestriction(CeedElemRestriction rstr, CeedElemRestriction *point_block_rstr) { 1137 Ceed ceed; 1138 CeedInt num_elem, num_comp, shift, elem_size, comp_stride, *point_block_offsets; 1139 CeedSize l_size; 1140 const CeedInt *offsets; 1141 1142 CeedCall(CeedElemRestrictionGetCeed(rstr, &ceed)); 1143 CeedCall(CeedElemRestrictionGetOffsets(rstr, CEED_MEM_HOST, &offsets)); 1144 1145 // Expand offsets 1146 CeedCall(CeedElemRestrictionGetNumElements(rstr, &num_elem)); 1147 CeedCall(CeedElemRestrictionGetNumComponents(rstr, &num_comp)); 1148 CeedCall(CeedElemRestrictionGetElementSize(rstr, &elem_size)); 1149 CeedCall(CeedElemRestrictionGetCompStride(rstr, &comp_stride)); 1150 CeedCall(CeedElemRestrictionGetLVectorSize(rstr, &l_size)); 1151 shift = num_comp; 1152 if (comp_stride != 1) shift *= num_comp; 1153 CeedCall(CeedCalloc(num_elem * elem_size, &point_block_offsets)); 1154 for (CeedInt i = 0; i < num_elem * elem_size; i++) { 1155 point_block_offsets[i] = offsets[i] * shift; 1156 } 1157 1158 // Create new restriction 1159 CeedCall(CeedElemRestrictionCreate(ceed, num_elem, elem_size, num_comp * num_comp, 1, l_size * num_comp, CEED_MEM_HOST, CEED_OWN_POINTER, 1160 point_block_offsets, point_block_rstr)); 1161 1162 // Cleanup 1163 CeedCall(CeedElemRestrictionRestoreOffsets(rstr, &offsets)); 1164 CeedCall(CeedDestroy(&ceed)); 1165 return CEED_ERROR_SUCCESS; 1166 } 1167 1168 /** 1169 @brief Get `CeedQFunctionAssemblyData` 1170 1171 @param[in] op `CeedOperator` to assemble 1172 @param[out] data `CeedQFunctionAssemblyData` 1173 1174 @return An error code: 0 - success, otherwise - failure 1175 1176 @ref Backend 1177 **/ 1178 int CeedOperatorGetQFunctionAssemblyData(CeedOperator op, CeedQFunctionAssemblyData *data) { 1179 if (!op->qf_assembled) { 1180 CeedQFunctionAssemblyData data; 1181 1182 CeedCall(CeedQFunctionAssemblyDataCreate(op->ceed, &data)); 1183 op->qf_assembled = data; 1184 } 1185 *data = op->qf_assembled; 1186 return CEED_ERROR_SUCCESS; 1187 } 1188 1189 /** 1190 @brief Create object holding `CeedQFunction` assembly data for `CeedOperator` 1191 1192 @param[in] ceed `Ceed` object used to create the `CeedQFunctionAssemblyData` 1193 @param[out] data Address of the variable where the newly created `CeedQFunctionAssemblyData` will be stored 1194 1195 @return An error code: 0 - success, otherwise - failure 1196 1197 @ref Backend 1198 **/ 1199 int CeedQFunctionAssemblyDataCreate(Ceed ceed, CeedQFunctionAssemblyData *data) { 1200 CeedCall(CeedCalloc(1, data)); 1201 (*data)->ref_count = 1; 1202 (*data)->ceed = ceed; 1203 CeedCall(CeedReference(ceed)); 1204 return CEED_ERROR_SUCCESS; 1205 } 1206 1207 /** 1208 @brief Increment the reference counter for a `CeedQFunctionAssemblyData` 1209 1210 @param[in,out] data `CeedQFunctionAssemblyData` to increment the reference counter 1211 1212 @return An error code: 0 - success, otherwise - failure 1213 1214 @ref Backend 1215 **/ 1216 int CeedQFunctionAssemblyDataReference(CeedQFunctionAssemblyData data) { 1217 data->ref_count++; 1218 return CEED_ERROR_SUCCESS; 1219 } 1220 1221 /** 1222 @brief Set re-use of `CeedQFunctionAssemblyData` 1223 1224 @param[in,out] data `CeedQFunctionAssemblyData` to mark for reuse 1225 @param[in] reuse_data Boolean flag indicating data re-use 1226 1227 @return An error code: 0 - success, otherwise - failure 1228 1229 @ref Backend 1230 **/ 1231 int CeedQFunctionAssemblyDataSetReuse(CeedQFunctionAssemblyData data, bool reuse_data) { 1232 data->reuse_data = reuse_data; 1233 data->needs_data_update = true; 1234 return CEED_ERROR_SUCCESS; 1235 } 1236 1237 /** 1238 @brief Mark `CeedQFunctionAssemblyData` as stale 1239 1240 @param[in,out] data `CeedQFunctionAssemblyData` to mark as stale 1241 @param[in] needs_data_update Boolean flag indicating if update is needed or completed 1242 1243 @return An error code: 0 - success, otherwise - failure 1244 1245 @ref Backend 1246 **/ 1247 int CeedQFunctionAssemblyDataSetUpdateNeeded(CeedQFunctionAssemblyData data, bool needs_data_update) { 1248 data->needs_data_update = needs_data_update; 1249 return CEED_ERROR_SUCCESS; 1250 } 1251 1252 /** 1253 @brief Determine if `CeedQFunctionAssemblyData` needs update 1254 1255 @param[in] data `CeedQFunctionAssemblyData` to mark as stale 1256 @param[out] is_update_needed Boolean flag indicating if re-assembly is required 1257 1258 @return An error code: 0 - success, otherwise - failure 1259 1260 @ref Backend 1261 **/ 1262 int CeedQFunctionAssemblyDataIsUpdateNeeded(CeedQFunctionAssemblyData data, bool *is_update_needed) { 1263 *is_update_needed = !data->reuse_data || data->needs_data_update; 1264 return CEED_ERROR_SUCCESS; 1265 } 1266 1267 /** 1268 @brief Copy the pointer to a `CeedQFunctionAssemblyData`. 1269 1270 Both pointers should be destroyed with @ref CeedQFunctionAssemblyDataDestroy(). 1271 1272 Note: If the value of ` *data_copy` passed to this function is non-`NULL` , then it is assumed that ` *data_copy` is a pointer to a `CeedQFunctionAssemblyData`. 1273 This `CeedQFunctionAssemblyData` will be destroyed if ` *data_copy` is the only reference to this `CeedQFunctionAssemblyData`. 1274 1275 @param[in] data `CeedQFunctionAssemblyData` to copy reference to 1276 @param[in,out] data_copy Variable to store copied reference 1277 1278 @return An error code: 0 - success, otherwise - failure 1279 1280 @ref Backend 1281 **/ 1282 int CeedQFunctionAssemblyDataReferenceCopy(CeedQFunctionAssemblyData data, CeedQFunctionAssemblyData *data_copy) { 1283 CeedCall(CeedQFunctionAssemblyDataReference(data)); 1284 CeedCall(CeedQFunctionAssemblyDataDestroy(data_copy)); 1285 *data_copy = data; 1286 return CEED_ERROR_SUCCESS; 1287 } 1288 1289 /** 1290 @brief Get setup status for internal objects for `CeedQFunctionAssemblyData` 1291 1292 @param[in] data `CeedQFunctionAssemblyData` to retrieve status 1293 @param[out] is_setup Boolean flag for setup status 1294 1295 @return An error code: 0 - success, otherwise - failure 1296 1297 @ref Backend 1298 **/ 1299 int CeedQFunctionAssemblyDataIsSetup(CeedQFunctionAssemblyData data, bool *is_setup) { 1300 *is_setup = data->is_setup; 1301 return CEED_ERROR_SUCCESS; 1302 } 1303 1304 /** 1305 @brief Set internal objects for `CeedQFunctionAssemblyData` 1306 1307 @param[in,out] data `CeedQFunctionAssemblyData` to set objects 1308 @param[in] vec `CeedVector` to store assembled `CeedQFunction` at quadrature points 1309 @param[in] rstr `CeedElemRestriction` for `CeedVector` containing assembled `CeedQFunction` 1310 1311 @return An error code: 0 - success, otherwise - failure 1312 1313 @ref Backend 1314 **/ 1315 int CeedQFunctionAssemblyDataSetObjects(CeedQFunctionAssemblyData data, CeedVector vec, CeedElemRestriction rstr) { 1316 CeedCall(CeedVectorReferenceCopy(vec, &data->vec)); 1317 CeedCall(CeedElemRestrictionReferenceCopy(rstr, &data->rstr)); 1318 1319 data->is_setup = true; 1320 return CEED_ERROR_SUCCESS; 1321 } 1322 1323 /** 1324 @brief Get internal objects for `CeedQFunctionAssemblyData` 1325 1326 @param[in,out] data `CeedQFunctionAssemblyData` to set objects 1327 @param[out] vec `CeedVector` to store assembled `CeedQFunction` at quadrature points 1328 @param[out] rstr `CeedElemRestriction` for `CeedVector` containing assembled `CeedQFunction` 1329 1330 @return An error code: 0 - success, otherwise - failure 1331 1332 @ref Backend 1333 **/ 1334 int CeedQFunctionAssemblyDataGetObjects(CeedQFunctionAssemblyData data, CeedVector *vec, CeedElemRestriction *rstr) { 1335 CeedCheck(data->is_setup, data->ceed, CEED_ERROR_INCOMPLETE, "Internal objects not set; must call CeedQFunctionAssemblyDataSetObjects first."); 1336 1337 CeedCall(CeedVectorReferenceCopy(data->vec, vec)); 1338 CeedCall(CeedElemRestrictionReferenceCopy(data->rstr, rstr)); 1339 return CEED_ERROR_SUCCESS; 1340 } 1341 1342 /** 1343 @brief Destroy `CeedQFunctionAssemblyData` 1344 1345 @param[in,out] data `CeedQFunctionAssemblyData` to destroy 1346 1347 @return An error code: 0 - success, otherwise - failure 1348 1349 @ref Backend 1350 **/ 1351 int CeedQFunctionAssemblyDataDestroy(CeedQFunctionAssemblyData *data) { 1352 if (!*data || --(*data)->ref_count > 0) { 1353 *data = NULL; 1354 return CEED_ERROR_SUCCESS; 1355 } 1356 CeedCall(CeedDestroy(&(*data)->ceed)); 1357 CeedCall(CeedVectorDestroy(&(*data)->vec)); 1358 CeedCall(CeedElemRestrictionDestroy(&(*data)->rstr)); 1359 1360 CeedCall(CeedFree(data)); 1361 return CEED_ERROR_SUCCESS; 1362 } 1363 1364 /** 1365 @brief Get `CeedOperatorAssemblyData` 1366 1367 @param[in] op `CeedOperator` to assemble 1368 @param[out] data `CeedOperatorAssemblyData` 1369 1370 @return An error code: 0 - success, otherwise - failure 1371 1372 @ref Backend 1373 **/ 1374 int CeedOperatorGetOperatorAssemblyData(CeedOperator op, CeedOperatorAssemblyData *data) { 1375 if (!op->op_assembled) { 1376 CeedOperatorAssemblyData data; 1377 1378 CeedCall(CeedOperatorAssemblyDataCreate(op->ceed, op, &data)); 1379 op->op_assembled = data; 1380 } 1381 *data = op->op_assembled; 1382 return CEED_ERROR_SUCCESS; 1383 } 1384 1385 /** 1386 @brief Create object holding `CeedOperator` assembly data. 1387 1388 The `CeedOperatorAssemblyData` holds an array with references to every active `CeedBasis` used in the `CeedOperator`. 1389 An array with references to the corresponding active `CeedElemRestriction` is also stored. 1390 For each active `CeedBasis, the `CeedOperatorAssemblyData` holds an array of all input and output @ref CeedEvalMode for this `CeedBasis`. 1391 The `CeedOperatorAssemblyData` holds an array of offsets for indexing into the assembled `CeedQFunction` arrays to the row representing each @ref CeedEvalMode. 1392 The number of input columns across all active bases for the assembled `CeedQFunction` is also stored. 1393 Lastly, the `CeedOperatorAssembly` data holds assembled matrices representing the full action of the `CeedBasis` for all @ref CeedEvalMode. 1394 1395 @param[in] ceed `Ceed` object used to create the `CeedOperatorAssemblyData` 1396 @param[in] op `CeedOperator` to be assembled 1397 @param[out] data Address of the variable where the newly created `CeedOperatorAssemblyData` will be stored 1398 1399 @return An error code: 0 - success, otherwise - failure 1400 1401 @ref Backend 1402 **/ 1403 int CeedOperatorAssemblyDataCreate(Ceed ceed, CeedOperator op, CeedOperatorAssemblyData *data) { 1404 CeedInt num_active_bases_in = 0, num_active_bases_out = 0, offset = 0; 1405 CeedInt num_input_fields, *num_eval_modes_in = NULL, num_output_fields, *num_eval_modes_out = NULL; 1406 CeedSize **eval_mode_offsets_in = NULL, **eval_mode_offsets_out = NULL; 1407 CeedEvalMode **eval_modes_in = NULL, **eval_modes_out = NULL; 1408 CeedQFunctionField *qf_fields; 1409 CeedQFunction qf; 1410 CeedOperatorField *op_fields; 1411 bool is_composite; 1412 1413 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 1414 CeedCheck(!is_composite, ceed, CEED_ERROR_INCOMPATIBLE, "Can only create CeedOperator assembly data for non-composite operators."); 1415 1416 // Allocate 1417 CeedCall(CeedCalloc(1, data)); 1418 (*data)->ceed = ceed; 1419 CeedCall(CeedReference(ceed)); 1420 1421 // Build OperatorAssembly data 1422 CeedCall(CeedOperatorGetQFunction(op, &qf)); 1423 1424 // Determine active input basis 1425 CeedCall(CeedQFunctionGetFields(qf, &num_input_fields, &qf_fields, NULL, NULL)); 1426 CeedCall(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 1427 for (CeedInt i = 0; i < num_input_fields; i++) { 1428 CeedVector vec; 1429 1430 CeedCall(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1431 if (vec == CEED_VECTOR_ACTIVE) { 1432 CeedInt index = -1, num_comp, q_comp; 1433 CeedEvalMode eval_mode; 1434 CeedBasis basis_in = NULL; 1435 1436 CeedCall(CeedOperatorFieldGetBasis(op_fields[i], &basis_in)); 1437 CeedCall(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1438 CeedCall(CeedBasisGetNumComponents(basis_in, &num_comp)); 1439 CeedCall(CeedBasisGetNumQuadratureComponents(basis_in, eval_mode, &q_comp)); 1440 for (CeedInt i = 0; i < num_active_bases_in; i++) { 1441 if ((*data)->active_bases_in[i] == basis_in) index = i; 1442 } 1443 if (index == -1) { 1444 CeedElemRestriction elem_rstr_in; 1445 1446 index = num_active_bases_in; 1447 CeedCall(CeedRealloc(num_active_bases_in + 1, &(*data)->active_bases_in)); 1448 (*data)->active_bases_in[num_active_bases_in] = NULL; 1449 CeedCall(CeedBasisReferenceCopy(basis_in, &(*data)->active_bases_in[num_active_bases_in])); 1450 CeedCall(CeedRealloc(num_active_bases_in + 1, &(*data)->active_elem_rstrs_in)); 1451 (*data)->active_elem_rstrs_in[num_active_bases_in] = NULL; 1452 CeedCall(CeedOperatorFieldGetElemRestriction(op_fields[i], &elem_rstr_in)); 1453 CeedCall(CeedElemRestrictionReferenceCopy(elem_rstr_in, &(*data)->active_elem_rstrs_in[num_active_bases_in])); 1454 CeedCall(CeedElemRestrictionDestroy(&elem_rstr_in)); 1455 CeedCall(CeedRealloc(num_active_bases_in + 1, &num_eval_modes_in)); 1456 num_eval_modes_in[index] = 0; 1457 CeedCall(CeedRealloc(num_active_bases_in + 1, &eval_modes_in)); 1458 eval_modes_in[index] = NULL; 1459 CeedCall(CeedRealloc(num_active_bases_in + 1, &eval_mode_offsets_in)); 1460 eval_mode_offsets_in[index] = NULL; 1461 CeedCall(CeedRealloc(num_active_bases_in + 1, &(*data)->assembled_bases_in)); 1462 (*data)->assembled_bases_in[index] = NULL; 1463 num_active_bases_in++; 1464 } 1465 if (eval_mode != CEED_EVAL_WEIGHT) { 1466 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF Assembly 1467 CeedCall(CeedRealloc(num_eval_modes_in[index] + q_comp, &eval_modes_in[index])); 1468 CeedCall(CeedRealloc(num_eval_modes_in[index] + q_comp, &eval_mode_offsets_in[index])); 1469 for (CeedInt d = 0; d < q_comp; d++) { 1470 eval_modes_in[index][num_eval_modes_in[index] + d] = eval_mode; 1471 eval_mode_offsets_in[index][num_eval_modes_in[index] + d] = offset; 1472 offset += num_comp; 1473 } 1474 num_eval_modes_in[index] += q_comp; 1475 } 1476 CeedCall(CeedBasisDestroy(&basis_in)); 1477 } 1478 CeedCall(CeedVectorDestroy(&vec)); 1479 } 1480 1481 // Determine active output basis 1482 CeedCall(CeedQFunctionGetFields(qf, NULL, NULL, &num_output_fields, &qf_fields)); 1483 CeedCall(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 1484 offset = 0; 1485 for (CeedInt i = 0; i < num_output_fields; i++) { 1486 CeedVector vec; 1487 1488 CeedCall(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1489 if (vec == CEED_VECTOR_ACTIVE) { 1490 CeedInt index = -1, num_comp, q_comp; 1491 CeedEvalMode eval_mode; 1492 CeedBasis basis_out = NULL; 1493 1494 CeedCall(CeedOperatorFieldGetBasis(op_fields[i], &basis_out)); 1495 CeedCall(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1496 CeedCall(CeedBasisGetNumComponents(basis_out, &num_comp)); 1497 CeedCall(CeedBasisGetNumQuadratureComponents(basis_out, eval_mode, &q_comp)); 1498 for (CeedInt i = 0; i < num_active_bases_out; i++) { 1499 if ((*data)->active_bases_out[i] == basis_out) index = i; 1500 } 1501 if (index == -1) { 1502 CeedElemRestriction elem_rstr_out; 1503 1504 index = num_active_bases_out; 1505 CeedCall(CeedRealloc(num_active_bases_out + 1, &(*data)->active_bases_out)); 1506 (*data)->active_bases_out[num_active_bases_out] = NULL; 1507 CeedCall(CeedBasisReferenceCopy(basis_out, &(*data)->active_bases_out[num_active_bases_out])); 1508 CeedCall(CeedRealloc(num_active_bases_out + 1, &(*data)->active_elem_rstrs_out)); 1509 (*data)->active_elem_rstrs_out[num_active_bases_out] = NULL; 1510 CeedCall(CeedOperatorFieldGetElemRestriction(op_fields[i], &elem_rstr_out)); 1511 CeedCall(CeedElemRestrictionReferenceCopy(elem_rstr_out, &(*data)->active_elem_rstrs_out[num_active_bases_out])); 1512 CeedCall(CeedElemRestrictionDestroy(&elem_rstr_out)); 1513 CeedCall(CeedRealloc(num_active_bases_out + 1, &num_eval_modes_out)); 1514 num_eval_modes_out[index] = 0; 1515 CeedCall(CeedRealloc(num_active_bases_out + 1, &eval_modes_out)); 1516 eval_modes_out[index] = NULL; 1517 CeedCall(CeedRealloc(num_active_bases_out + 1, &eval_mode_offsets_out)); 1518 eval_mode_offsets_out[index] = NULL; 1519 CeedCall(CeedRealloc(num_active_bases_out + 1, &(*data)->assembled_bases_out)); 1520 (*data)->assembled_bases_out[index] = NULL; 1521 num_active_bases_out++; 1522 } 1523 if (eval_mode != CEED_EVAL_WEIGHT) { 1524 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF Assembly 1525 CeedCall(CeedRealloc(num_eval_modes_out[index] + q_comp, &eval_modes_out[index])); 1526 CeedCall(CeedRealloc(num_eval_modes_out[index] + q_comp, &eval_mode_offsets_out[index])); 1527 for (CeedInt d = 0; d < q_comp; d++) { 1528 eval_modes_out[index][num_eval_modes_out[index] + d] = eval_mode; 1529 eval_mode_offsets_out[index][num_eval_modes_out[index] + d] = offset; 1530 offset += num_comp; 1531 } 1532 num_eval_modes_out[index] += q_comp; 1533 } 1534 CeedCall(CeedBasisDestroy(&basis_out)); 1535 } 1536 CeedCall(CeedVectorDestroy(&vec)); 1537 } 1538 (*data)->num_active_bases_in = num_active_bases_in; 1539 (*data)->num_eval_modes_in = num_eval_modes_in; 1540 (*data)->eval_modes_in = eval_modes_in; 1541 (*data)->eval_mode_offsets_in = eval_mode_offsets_in; 1542 (*data)->num_active_bases_out = num_active_bases_out; 1543 (*data)->num_eval_modes_out = num_eval_modes_out; 1544 (*data)->eval_modes_out = eval_modes_out; 1545 (*data)->eval_mode_offsets_out = eval_mode_offsets_out; 1546 (*data)->num_output_components = offset; 1547 return CEED_ERROR_SUCCESS; 1548 } 1549 1550 /** 1551 @brief Get `CeedOperator` @ref CeedEvalMode for assembly. 1552 1553 Note: See @ref CeedOperatorAssemblyDataCreate() for a full description of the data stored in this object. 1554 1555 @param[in] data `CeedOperatorAssemblyData` 1556 @param[out] num_active_bases_in Total number of active bases for input 1557 @param[out] num_eval_modes_in Pointer to hold array of numbers of input @ref CeedEvalMode, or `NULL`. 1558 `eval_modes_in[0]` holds an array of eval modes for the first active `CeedBasis`. 1559 @param[out] eval_modes_in Pointer to hold arrays of input @ref CeedEvalMode, or `NULL` 1560 @param[out] eval_mode_offsets_in Pointer to hold arrays of input offsets at each quadrature point 1561 @param[out] num_active_bases_out Total number of active bases for output 1562 @param[out] num_eval_modes_out Pointer to hold array of numbers of output @ref CeedEvalMode, or `NULL` 1563 @param[out] eval_modes_out Pointer to hold arrays of output @ref CeedEvalMode, or `NULL` 1564 @param[out] eval_mode_offsets_out Pointer to hold arrays of output offsets at each quadrature point 1565 @param[out] num_output_components The number of columns in the assembled `CeedQFunction` matrix for each quadrature point, including contributions of all active bases 1566 1567 @return An error code: 0 - success, otherwise - failure 1568 1569 @ref Backend 1570 **/ 1571 int CeedOperatorAssemblyDataGetEvalModes(CeedOperatorAssemblyData data, CeedInt *num_active_bases_in, CeedInt **num_eval_modes_in, 1572 const CeedEvalMode ***eval_modes_in, CeedSize ***eval_mode_offsets_in, CeedInt *num_active_bases_out, 1573 CeedInt **num_eval_modes_out, const CeedEvalMode ***eval_modes_out, CeedSize ***eval_mode_offsets_out, 1574 CeedSize *num_output_components) { 1575 if (num_active_bases_in) *num_active_bases_in = data->num_active_bases_in; 1576 if (num_eval_modes_in) *num_eval_modes_in = data->num_eval_modes_in; 1577 if (eval_modes_in) *eval_modes_in = (const CeedEvalMode **)data->eval_modes_in; 1578 if (eval_mode_offsets_in) *eval_mode_offsets_in = data->eval_mode_offsets_in; 1579 if (num_active_bases_out) *num_active_bases_out = data->num_active_bases_out; 1580 if (num_eval_modes_out) *num_eval_modes_out = data->num_eval_modes_out; 1581 if (eval_modes_out) *eval_modes_out = (const CeedEvalMode **)data->eval_modes_out; 1582 if (eval_mode_offsets_out) *eval_mode_offsets_out = data->eval_mode_offsets_out; 1583 if (num_output_components) *num_output_components = data->num_output_components; 1584 return CEED_ERROR_SUCCESS; 1585 } 1586 1587 /** 1588 @brief Get `CeedOperator` `CeedBasis` data for assembly. 1589 1590 Note: See @ref CeedOperatorAssemblyDataCreate() for a full description of the data stored in this object. 1591 1592 @param[in] data `CeedOperatorAssemblyData` 1593 @param[out] num_active_bases_in Number of active input bases, or `NULL` 1594 @param[out] active_bases_in Pointer to hold active input `CeedBasis`, or `NULL` 1595 @param[out] assembled_bases_in Pointer to hold assembled active input `B` , or `NULL` 1596 @param[out] num_active_bases_out Number of active output bases, or `NULL` 1597 @param[out] active_bases_out Pointer to hold active output `CeedBasis`, or `NULL` 1598 @param[out] assembled_bases_out Pointer to hold assembled active output `B` , or `NULL` 1599 1600 @return An error code: 0 - success, otherwise - failure 1601 1602 @ref Backend 1603 **/ 1604 int CeedOperatorAssemblyDataGetBases(CeedOperatorAssemblyData data, CeedInt *num_active_bases_in, CeedBasis **active_bases_in, 1605 const CeedScalar ***assembled_bases_in, CeedInt *num_active_bases_out, CeedBasis **active_bases_out, 1606 const CeedScalar ***assembled_bases_out) { 1607 // Assemble B_in, B_out if needed 1608 if (assembled_bases_in && !data->assembled_bases_in[0]) { 1609 CeedInt num_qpts; 1610 1611 if (data->active_bases_in[0] == CEED_BASIS_NONE) CeedCall(CeedElemRestrictionGetElementSize(data->active_elem_rstrs_in[0], &num_qpts)); 1612 else CeedCall(CeedBasisGetNumQuadraturePoints(data->active_bases_in[0], &num_qpts)); 1613 for (CeedInt b = 0; b < data->num_active_bases_in; b++) { 1614 bool has_eval_none = false; 1615 CeedInt num_nodes; 1616 CeedScalar *B_in = NULL, *identity = NULL; 1617 1618 CeedCall(CeedElemRestrictionGetElementSize(data->active_elem_rstrs_in[b], &num_nodes)); 1619 CeedCall(CeedCalloc(num_qpts * num_nodes * data->num_eval_modes_in[b], &B_in)); 1620 1621 for (CeedInt i = 0; i < data->num_eval_modes_in[b]; i++) { 1622 has_eval_none = has_eval_none || (data->eval_modes_in[b][i] == CEED_EVAL_NONE); 1623 } 1624 if (has_eval_none) { 1625 CeedCall(CeedCalloc(num_qpts * num_nodes, &identity)); 1626 for (CeedInt i = 0; i < (num_nodes < num_qpts ? num_nodes : num_qpts); i++) { 1627 identity[i * num_nodes + i] = 1.0; 1628 } 1629 } 1630 1631 for (CeedInt q = 0; q < num_qpts; q++) { 1632 for (CeedInt n = 0; n < num_nodes; n++) { 1633 CeedInt d_in = 0, q_comp_in; 1634 CeedEvalMode eval_mode_in_prev = CEED_EVAL_NONE; 1635 1636 for (CeedInt e_in = 0; e_in < data->num_eval_modes_in[b]; e_in++) { 1637 const CeedInt qq = data->num_eval_modes_in[b] * q; 1638 const CeedScalar *B = NULL; 1639 1640 CeedCall(CeedOperatorGetBasisPointer(data->active_bases_in[b], data->eval_modes_in[b][e_in], identity, &B)); 1641 CeedCall(CeedBasisGetNumQuadratureComponents(data->active_bases_in[b], data->eval_modes_in[b][e_in], &q_comp_in)); 1642 if (q_comp_in > 1) { 1643 if (e_in == 0 || data->eval_modes_in[b][e_in] != eval_mode_in_prev) d_in = 0; 1644 else B = &B[(++d_in) * num_qpts * num_nodes]; 1645 } 1646 eval_mode_in_prev = data->eval_modes_in[b][e_in]; 1647 B_in[(qq + e_in) * num_nodes + n] = B[q * num_nodes + n]; 1648 } 1649 } 1650 } 1651 if (identity) CeedCall(CeedFree(&identity)); 1652 data->assembled_bases_in[b] = B_in; 1653 } 1654 } 1655 1656 if (assembled_bases_out && !data->assembled_bases_out[0]) { 1657 CeedInt num_qpts; 1658 1659 if (data->active_bases_out[0] == CEED_BASIS_NONE) CeedCall(CeedElemRestrictionGetElementSize(data->active_elem_rstrs_out[0], &num_qpts)); 1660 else CeedCall(CeedBasisGetNumQuadraturePoints(data->active_bases_out[0], &num_qpts)); 1661 for (CeedInt b = 0; b < data->num_active_bases_out; b++) { 1662 bool has_eval_none = false; 1663 CeedInt num_nodes; 1664 CeedScalar *B_out = NULL, *identity = NULL; 1665 1666 CeedCall(CeedElemRestrictionGetElementSize(data->active_elem_rstrs_out[b], &num_nodes)); 1667 CeedCall(CeedCalloc(num_qpts * num_nodes * data->num_eval_modes_out[b], &B_out)); 1668 1669 for (CeedInt i = 0; i < data->num_eval_modes_out[b]; i++) { 1670 has_eval_none = has_eval_none || (data->eval_modes_out[b][i] == CEED_EVAL_NONE); 1671 } 1672 if (has_eval_none) { 1673 CeedCall(CeedCalloc(num_qpts * num_nodes, &identity)); 1674 for (CeedInt i = 0; i < (num_nodes < num_qpts ? num_nodes : num_qpts); i++) { 1675 identity[i * num_nodes + i] = 1.0; 1676 } 1677 } 1678 1679 for (CeedInt q = 0; q < num_qpts; q++) { 1680 for (CeedInt n = 0; n < num_nodes; n++) { 1681 CeedInt d_out = 0, q_comp_out; 1682 CeedEvalMode eval_mode_out_prev = CEED_EVAL_NONE; 1683 1684 for (CeedInt e_out = 0; e_out < data->num_eval_modes_out[b]; e_out++) { 1685 const CeedInt qq = data->num_eval_modes_out[b] * q; 1686 const CeedScalar *B = NULL; 1687 1688 CeedCall(CeedOperatorGetBasisPointer(data->active_bases_out[b], data->eval_modes_out[b][e_out], identity, &B)); 1689 CeedCall(CeedBasisGetNumQuadratureComponents(data->active_bases_out[b], data->eval_modes_out[b][e_out], &q_comp_out)); 1690 if (q_comp_out > 1) { 1691 if (e_out == 0 || data->eval_modes_out[b][e_out] != eval_mode_out_prev) d_out = 0; 1692 else B = &B[(++d_out) * num_qpts * num_nodes]; 1693 } 1694 eval_mode_out_prev = data->eval_modes_out[b][e_out]; 1695 B_out[(qq + e_out) * num_nodes + n] = B[q * num_nodes + n]; 1696 } 1697 } 1698 } 1699 if (identity) CeedCall(CeedFree(&identity)); 1700 data->assembled_bases_out[b] = B_out; 1701 } 1702 } 1703 1704 // Pass out assembled data 1705 if (num_active_bases_in) *num_active_bases_in = data->num_active_bases_in; 1706 if (active_bases_in) *active_bases_in = data->active_bases_in; 1707 if (assembled_bases_in) *assembled_bases_in = (const CeedScalar **)data->assembled_bases_in; 1708 if (num_active_bases_out) *num_active_bases_out = data->num_active_bases_out; 1709 if (active_bases_out) *active_bases_out = data->active_bases_out; 1710 if (assembled_bases_out) *assembled_bases_out = (const CeedScalar **)data->assembled_bases_out; 1711 return CEED_ERROR_SUCCESS; 1712 } 1713 1714 /** 1715 @brief Get `CeedOperator` `CeedBasis` data for assembly. 1716 1717 Note: See @ref CeedOperatorAssemblyDataCreate() for a full description of the data stored in this object. 1718 1719 @param[in] data `CeedOperatorAssemblyData` 1720 @param[out] num_active_elem_rstrs_in Number of active input element restrictions, or `NULL` 1721 @param[out] active_elem_rstrs_in Pointer to hold active input `CeedElemRestriction`, or `NULL` 1722 @param[out] num_active_elem_rstrs_out Number of active output element restrictions, or `NULL` 1723 @param[out] active_elem_rstrs_out Pointer to hold active output `CeedElemRestriction`, or `NULL` 1724 1725 @return An error code: 0 - success, otherwise - failure 1726 1727 @ref Backend 1728 **/ 1729 int CeedOperatorAssemblyDataGetElemRestrictions(CeedOperatorAssemblyData data, CeedInt *num_active_elem_rstrs_in, 1730 CeedElemRestriction **active_elem_rstrs_in, CeedInt *num_active_elem_rstrs_out, 1731 CeedElemRestriction **active_elem_rstrs_out) { 1732 if (num_active_elem_rstrs_in) *num_active_elem_rstrs_in = data->num_active_bases_in; 1733 if (active_elem_rstrs_in) *active_elem_rstrs_in = data->active_elem_rstrs_in; 1734 if (num_active_elem_rstrs_out) *num_active_elem_rstrs_out = data->num_active_bases_out; 1735 if (active_elem_rstrs_out) *active_elem_rstrs_out = data->active_elem_rstrs_out; 1736 return CEED_ERROR_SUCCESS; 1737 } 1738 1739 /** 1740 @brief Destroy `CeedOperatorAssemblyData` 1741 1742 @param[in,out] data `CeedOperatorAssemblyData` to destroy 1743 1744 @return An error code: 0 - success, otherwise - failure 1745 1746 @ref Backend 1747 **/ 1748 int CeedOperatorAssemblyDataDestroy(CeedOperatorAssemblyData *data) { 1749 if (!*data) { 1750 *data = NULL; 1751 return CEED_ERROR_SUCCESS; 1752 } 1753 CeedCall(CeedDestroy(&(*data)->ceed)); 1754 for (CeedInt b = 0; b < (*data)->num_active_bases_in; b++) { 1755 CeedCall(CeedBasisDestroy(&(*data)->active_bases_in[b])); 1756 CeedCall(CeedElemRestrictionDestroy(&(*data)->active_elem_rstrs_in[b])); 1757 CeedCall(CeedFree(&(*data)->eval_modes_in[b])); 1758 CeedCall(CeedFree(&(*data)->eval_mode_offsets_in[b])); 1759 CeedCall(CeedFree(&(*data)->assembled_bases_in[b])); 1760 } 1761 for (CeedInt b = 0; b < (*data)->num_active_bases_out; b++) { 1762 CeedCall(CeedBasisDestroy(&(*data)->active_bases_out[b])); 1763 CeedCall(CeedElemRestrictionDestroy(&(*data)->active_elem_rstrs_out[b])); 1764 CeedCall(CeedFree(&(*data)->eval_modes_out[b])); 1765 CeedCall(CeedFree(&(*data)->eval_mode_offsets_out[b])); 1766 CeedCall(CeedFree(&(*data)->assembled_bases_out[b])); 1767 } 1768 CeedCall(CeedFree(&(*data)->active_bases_in)); 1769 CeedCall(CeedFree(&(*data)->active_bases_out)); 1770 CeedCall(CeedFree(&(*data)->active_elem_rstrs_in)); 1771 CeedCall(CeedFree(&(*data)->active_elem_rstrs_out)); 1772 CeedCall(CeedFree(&(*data)->num_eval_modes_in)); 1773 CeedCall(CeedFree(&(*data)->num_eval_modes_out)); 1774 CeedCall(CeedFree(&(*data)->eval_modes_in)); 1775 CeedCall(CeedFree(&(*data)->eval_modes_out)); 1776 CeedCall(CeedFree(&(*data)->eval_mode_offsets_in)); 1777 CeedCall(CeedFree(&(*data)->eval_mode_offsets_out)); 1778 CeedCall(CeedFree(&(*data)->assembled_bases_in)); 1779 CeedCall(CeedFree(&(*data)->assembled_bases_out)); 1780 1781 CeedCall(CeedFree(data)); 1782 return CEED_ERROR_SUCCESS; 1783 } 1784 1785 /** 1786 @brief Retrieve fallback `CeedOperator` with a reference `Ceed` for advanced `CeedOperator` functionality 1787 1788 @param[in] op `CeedOperator` to retrieve fallback for 1789 @param[out] op_fallback Fallback `CeedOperator` 1790 1791 @return An error code: 0 - success, otherwise - failure 1792 1793 @ref Backend 1794 **/ 1795 int CeedOperatorGetFallback(CeedOperator op, CeedOperator *op_fallback) { 1796 // Create if needed 1797 if (!op->op_fallback) CeedCall(CeedOperatorCreateFallback(op)); 1798 if (op->op_fallback) { 1799 bool is_debug; 1800 Ceed ceed; 1801 1802 CeedCall(CeedOperatorGetCeed(op, &ceed)); 1803 CeedCall(CeedIsDebug(ceed, &is_debug)); 1804 if (is_debug) { 1805 Ceed ceed_fallback; 1806 const char *resource, *resource_fallback; 1807 1808 CeedCall(CeedGetOperatorFallbackCeed(ceed, &ceed_fallback)); 1809 CeedCall(CeedGetResource(ceed, &resource)); 1810 CeedCall(CeedGetResource(ceed_fallback, &resource_fallback)); 1811 1812 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "---------- CeedOperator Fallback ----------\n"); 1813 CeedDebug(ceed, "Falling back from %s operator at address %p to %s operator at address %p\n", resource, op, resource_fallback, op->op_fallback); 1814 CeedCall(CeedDestroy(&ceed_fallback)); 1815 } 1816 CeedCall(CeedDestroy(&ceed)); 1817 } 1818 *op_fallback = op->op_fallback; 1819 return CEED_ERROR_SUCCESS; 1820 } 1821 1822 /** 1823 @brief Get the parent `CeedOperator` for a fallback `CeedOperator` 1824 1825 @param[in] op `CeedOperator` context 1826 @param[out] parent Variable to store parent `CeedOperator` context 1827 1828 @return An error code: 0 - success, otherwise - failure 1829 1830 @ref Backend 1831 **/ 1832 int CeedOperatorGetFallbackParent(CeedOperator op, CeedOperator *parent) { 1833 *parent = op->op_fallback_parent ? op->op_fallback_parent : NULL; 1834 return CEED_ERROR_SUCCESS; 1835 } 1836 1837 /** 1838 @brief Get the `Ceed` context of the parent `CeedOperator` for a fallback `CeedOperator` 1839 1840 @param[in] op `CeedOperator` context 1841 @param[out] parent Variable to store parent `Ceed` context 1842 1843 @return An error code: 0 - success, otherwise - failure 1844 1845 @ref Backend 1846 **/ 1847 int CeedOperatorGetFallbackParentCeed(CeedOperator op, Ceed *parent) { 1848 *parent = NULL; 1849 if (op->op_fallback_parent) CeedCall(CeedReferenceCopy(op->op_fallback_parent->ceed, parent)); 1850 else CeedCall(CeedReferenceCopy(CeedOperatorReturnCeed(op), parent)); 1851 return CEED_ERROR_SUCCESS; 1852 } 1853 1854 /// @} 1855 1856 /// ---------------------------------------------------------------------------- 1857 /// CeedOperator Public API 1858 /// ---------------------------------------------------------------------------- 1859 /// @addtogroup CeedOperatorUser 1860 /// @{ 1861 1862 /** 1863 @brief Assemble a linear `CeedQFunction` associated with a `CeedOperator`. 1864 1865 This returns a `CeedVector` containing a matrix at each quadrature point providing the action of the `CeedQFunction` associated with the `CeedOperator`. 1866 The vector `assembled` is of shape `[num_elements, num_input_fields, num_output_fields, num_quad_points]` and contains column-major matrices representing the action of the `CeedQFunction` for a corresponding quadrature point on an element. 1867 1868 Inputs and outputs are in the order provided by the user when adding `CeedOperator` fields. 1869 For example, a `CeedQFunction` with inputs `u` and `gradu` and outputs `gradv` and `v` , provided in that order, would result in an assembled `CeedQFunction` that consists of `(1 + dim) x (dim + 1)` matrices at each quadrature point acting on the input ` [u, du_0, du_1]` and producing the output `[dv_0, dv_1, v]`. 1870 1871 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 1872 1873 @param[in] op `CeedOperator` to assemble `CeedQFunction` 1874 @param[out] assembled `CeedVector` to store assembled `CeedQFunction` at quadrature points 1875 @param[out] rstr `CeedElemRestriction` for `CeedVector` containing assembled `CeedQFunction` 1876 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 1877 1878 @return An error code: 0 - success, otherwise - failure 1879 1880 @ref User 1881 **/ 1882 int CeedOperatorLinearAssembleQFunction(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 1883 CeedCall(CeedOperatorCheckReady(op)); 1884 1885 if (op->LinearAssembleQFunction) { 1886 // Backend version 1887 CeedCall(op->LinearAssembleQFunction(op, assembled, rstr, request)); 1888 } else { 1889 // Operator fallback 1890 CeedOperator op_fallback; 1891 1892 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 1893 if (op_fallback) CeedCall(CeedOperatorLinearAssembleQFunction(op_fallback, assembled, rstr, request)); 1894 else return CeedError(CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, "Backend does not support CeedOperatorLinearAssembleQFunction"); 1895 } 1896 return CEED_ERROR_SUCCESS; 1897 } 1898 1899 /** 1900 @brief Assemble `CeedQFunction` and store result internally. 1901 1902 Return copied references of stored data to the caller. 1903 Caller is responsible for ownership and destruction of the copied references. 1904 See also @ref CeedOperatorLinearAssembleQFunction(). 1905 1906 Note: If the value of `assembled` or `rstr` passed to this function are non-`NULL` , then it is assumed that they hold valid pointers. 1907 These objects will be destroyed if `*assembled` or `*rstr` is the only reference to the object. 1908 1909 @param[in] op `CeedOperator` to assemble `CeedQFunction` 1910 @param[out] assembled `CeedVector` to store assembled `CeedQFunction` at quadrature points 1911 @param[out] rstr `CeedElemRestriction` for `CeedVector` containing assembled `CeedQFunction` 1912 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 1913 1914 @return An error code: 0 - success, otherwise - failure 1915 1916 @ref User 1917 **/ 1918 int CeedOperatorLinearAssembleQFunctionBuildOrUpdate(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 1919 int (*LinearAssembleQFunctionUpdate)(CeedOperator, CeedVector, CeedElemRestriction, CeedRequest *) = NULL; 1920 CeedOperator op_assemble = NULL; 1921 CeedOperator op_fallback_parent = NULL; 1922 1923 CeedCall(CeedOperatorCheckReady(op)); 1924 1925 // Determine if fallback parent or operator has implementation 1926 CeedCall(CeedOperatorGetFallbackParent(op, &op_fallback_parent)); 1927 if (op_fallback_parent && op_fallback_parent->LinearAssembleQFunctionUpdate) { 1928 // -- Backend version for op fallback parent is faster, if it exists 1929 LinearAssembleQFunctionUpdate = op_fallback_parent->LinearAssembleQFunctionUpdate; 1930 op_assemble = op_fallback_parent; 1931 } else if (op->LinearAssembleQFunctionUpdate) { 1932 // -- Backend version for op 1933 LinearAssembleQFunctionUpdate = op->LinearAssembleQFunctionUpdate; 1934 op_assemble = op; 1935 } 1936 1937 // Assemble QFunction 1938 if (LinearAssembleQFunctionUpdate) { 1939 // Backend or fallback parent version 1940 CeedQFunctionAssemblyData data; 1941 bool data_is_setup; 1942 CeedVector assembled_vec = NULL; 1943 CeedElemRestriction assembled_rstr = NULL; 1944 1945 CeedCall(CeedOperatorGetQFunctionAssemblyData(op, &data)); 1946 CeedCall(CeedQFunctionAssemblyDataIsSetup(data, &data_is_setup)); 1947 if (data_is_setup) { 1948 bool update_needed; 1949 1950 CeedCall(CeedQFunctionAssemblyDataGetObjects(data, &assembled_vec, &assembled_rstr)); 1951 CeedCall(CeedQFunctionAssemblyDataIsUpdateNeeded(data, &update_needed)); 1952 if (update_needed) CeedCall(LinearAssembleQFunctionUpdate(op_assemble, assembled_vec, assembled_rstr, request)); 1953 } else { 1954 CeedCall(CeedOperatorLinearAssembleQFunction(op_assemble, &assembled_vec, &assembled_rstr, request)); 1955 CeedCall(CeedQFunctionAssemblyDataSetObjects(data, assembled_vec, assembled_rstr)); 1956 } 1957 CeedCall(CeedQFunctionAssemblyDataSetUpdateNeeded(data, false)); 1958 1959 // Copy reference from internally held copy 1960 CeedCall(CeedVectorReferenceCopy(assembled_vec, assembled)); 1961 CeedCall(CeedElemRestrictionReferenceCopy(assembled_rstr, rstr)); 1962 CeedCall(CeedVectorDestroy(&assembled_vec)); 1963 CeedCall(CeedElemRestrictionDestroy(&assembled_rstr)); 1964 } else { 1965 // Operator fallback 1966 CeedOperator op_fallback; 1967 1968 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 1969 if (op_fallback) CeedCall(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op_fallback, assembled, rstr, request)); 1970 else return CeedError(CeedOperatorReturnCeed(op), CEED_ERROR_UNSUPPORTED, "Backend does not support CeedOperatorLinearAssembleQFunctionUpdate"); 1971 } 1972 return CEED_ERROR_SUCCESS; 1973 } 1974 1975 /** 1976 @brief Assemble the diagonal of a square linear `CeedOperator` 1977 1978 This overwrites a `CeedVector` with the diagonal of a linear `CeedOperator`. 1979 1980 Note: Currently only non-composite `CeedOperator` with a single field and composite `CeedOperator` with single field sub-operators are supported. 1981 1982 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 1983 1984 @param[in] op `CeedOperator` to assemble `CeedQFunction` 1985 @param[out] assembled `CeedVector` to store assembled `CeedOperator` diagonal 1986 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 1987 1988 @return An error code: 0 - success, otherwise - failure 1989 1990 @ref User 1991 **/ 1992 int CeedOperatorLinearAssembleDiagonal(CeedOperator op, CeedVector assembled, CeedRequest *request) { 1993 bool is_composite; 1994 CeedSize input_size = 0, output_size = 0; 1995 1996 CeedCall(CeedOperatorCheckReady(op)); 1997 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 1998 1999 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2000 CeedCheck(input_size == output_size, CeedOperatorReturnCeed(op), CEED_ERROR_DIMENSION, "Operator must be square"); 2001 2002 // Early exit for empty operator 2003 if (!is_composite) { 2004 CeedInt num_elem = 0; 2005 2006 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2007 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2008 } 2009 2010 if (op->LinearAssembleDiagonal) { 2011 // Backend version 2012 CeedCall(op->LinearAssembleDiagonal(op, assembled, request)); 2013 return CEED_ERROR_SUCCESS; 2014 } else if (op->LinearAssembleAddDiagonal) { 2015 // Backend version with zeroing first 2016 CeedCall(CeedVectorSetValue(assembled, 0.0)); 2017 CeedCall(op->LinearAssembleAddDiagonal(op, assembled, request)); 2018 return CEED_ERROR_SUCCESS; 2019 } else { 2020 // Operator fallback 2021 CeedOperator op_fallback; 2022 2023 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2024 if (op_fallback) { 2025 CeedCall(CeedOperatorLinearAssembleDiagonal(op_fallback, assembled, request)); 2026 return CEED_ERROR_SUCCESS; 2027 } 2028 } 2029 // Default interface implementation 2030 CeedCall(CeedVectorSetValue(assembled, 0.0)); 2031 CeedCall(CeedOperatorLinearAssembleAddDiagonal(op, assembled, request)); 2032 return CEED_ERROR_SUCCESS; 2033 } 2034 2035 /** 2036 @brief Assemble the diagonal of a square linear `CeedOperator`. 2037 2038 This sums into a `CeedVector` the diagonal of a linear `CeedOperator`. 2039 2040 Note: Currently only non-composite `CeedOperator` with a single field and composite `CeedOperator` with single field sub-operators are supported. 2041 2042 Note: Calling this function asserts that setup is complete and sets the CeedOperator as immutable. 2043 2044 @param[in] op `CeedOperator` to assemble `CeedQFunction` 2045 @param[out] assembled `CeedVector` to store assembled `CeedOperator` diagonal 2046 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 2047 2048 @return An error code: 0 - success, otherwise - failure 2049 2050 @ref User 2051 **/ 2052 int CeedOperatorLinearAssembleAddDiagonal(CeedOperator op, CeedVector assembled, CeedRequest *request) { 2053 bool is_composite; 2054 CeedSize input_size = 0, output_size = 0; 2055 2056 CeedCall(CeedOperatorCheckReady(op)); 2057 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2058 2059 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2060 CeedCheck(input_size == output_size, CeedOperatorReturnCeed(op), CEED_ERROR_DIMENSION, "Operator must be square"); 2061 2062 // Early exit for empty operator 2063 if (!is_composite) { 2064 CeedInt num_elem = 0; 2065 2066 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2067 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2068 } 2069 2070 if (op->LinearAssembleAddDiagonal) { 2071 // Backend version 2072 CeedCall(op->LinearAssembleAddDiagonal(op, assembled, request)); 2073 return CEED_ERROR_SUCCESS; 2074 } else { 2075 // Operator fallback 2076 CeedOperator op_fallback; 2077 2078 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2079 if (op_fallback) { 2080 CeedCall(CeedOperatorLinearAssembleAddDiagonal(op_fallback, assembled, request)); 2081 return CEED_ERROR_SUCCESS; 2082 } 2083 } 2084 // Default interface implementation 2085 if (is_composite) { 2086 CeedCall(CeedCompositeOperatorLinearAssembleAddDiagonal(op, request, false, assembled)); 2087 } else { 2088 CeedCall(CeedSingleOperatorLinearAssembleAddDiagonal(op, request, false, assembled)); 2089 } 2090 return CEED_ERROR_SUCCESS; 2091 } 2092 2093 /** 2094 @brief Fully assemble the point-block diagonal pattern of a linear `CeedOperator`. 2095 2096 Expected to be used in conjunction with @ref CeedOperatorLinearAssemblePointBlockDiagonal(). 2097 2098 The assembly routines use coordinate format, with `num_entries` tuples of the form `(i, j, value)` which indicate that value should be added to the matrix in entry `(i, j)`. 2099 Note that the `(i, j)` pairs are unique. 2100 This function returns the number of entries and their `(i, j)` locations, while @ref CeedOperatorLinearAssemblePointBlockDiagonal() provides the values in the same ordering. 2101 2102 This will generally be slow unless your operator is low-order. 2103 2104 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2105 2106 @param[in] op `CeedOperator` to assemble 2107 @param[out] num_entries Number of entries in coordinate nonzero pattern 2108 @param[out] rows Row number for each entry 2109 @param[out] cols Column number for each entry 2110 2111 @ref User 2112 **/ 2113 int CeedOperatorLinearAssemblePointBlockDiagonalSymbolic(CeedOperator op, CeedSize *num_entries, CeedInt **rows, CeedInt **cols) { 2114 bool is_composite; 2115 CeedInt num_active_components, num_sub_operators; 2116 CeedOperator *sub_operators; 2117 2118 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2119 2120 CeedSize input_size = 0, output_size = 0; 2121 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2122 CeedCheck(input_size == output_size, CeedOperatorReturnCeed(op), CEED_ERROR_DIMENSION, "Operator must be square"); 2123 2124 if (is_composite) { 2125 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_sub_operators)); 2126 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2127 } else { 2128 sub_operators = &op; 2129 num_sub_operators = 1; 2130 } 2131 2132 // Verify operator can be assembled correctly 2133 { 2134 CeedOperatorAssemblyData data; 2135 CeedInt num_active_elem_rstrs, comp_stride; 2136 CeedElemRestriction *active_elem_rstrs; 2137 2138 // Get initial values to check against 2139 CeedCall(CeedOperatorGetOperatorAssemblyData(sub_operators[0], &data)); 2140 CeedCall(CeedOperatorAssemblyDataGetElemRestrictions(data, &num_active_elem_rstrs, &active_elem_rstrs, NULL, NULL)); 2141 CeedCall(CeedElemRestrictionGetCompStride(active_elem_rstrs[0], &comp_stride)); 2142 CeedCall(CeedElemRestrictionGetNumComponents(active_elem_rstrs[0], &num_active_components)); 2143 2144 // Verify that all active element restrictions have same component stride and number of components 2145 for (CeedInt k = 0; k < num_sub_operators; k++) { 2146 CeedCall(CeedOperatorGetOperatorAssemblyData(sub_operators[k], &data)); 2147 CeedCall(CeedOperatorAssemblyDataGetElemRestrictions(data, &num_active_elem_rstrs, &active_elem_rstrs, NULL, NULL)); 2148 for (CeedInt i = 0; i < num_active_elem_rstrs; i++) { 2149 CeedInt comp_stride_sub, num_active_components_sub; 2150 2151 CeedCall(CeedElemRestrictionGetCompStride(active_elem_rstrs[i], &comp_stride_sub)); 2152 CeedCheck(comp_stride == comp_stride_sub, CeedOperatorReturnCeed(op), CEED_ERROR_DIMENSION, 2153 "Active element restrictions must have the same component stride: %d vs %d", comp_stride, comp_stride_sub); 2154 CeedCall(CeedElemRestrictionGetNumComponents(active_elem_rstrs[i], &num_active_components_sub)); 2155 CeedCheck(num_active_components == num_active_components_sub, CeedOperatorReturnCeed(op), CEED_ERROR_INCOMPATIBLE, 2156 "All suboperators must have the same number of output components." 2157 " Previous: %" CeedInt_FMT " Current: %" CeedInt_FMT, 2158 num_active_components, num_active_components_sub); 2159 } 2160 } 2161 } 2162 *num_entries = input_size * num_active_components; 2163 CeedCall(CeedCalloc(*num_entries, rows)); 2164 CeedCall(CeedCalloc(*num_entries, cols)); 2165 2166 for (CeedInt o = 0; o < num_sub_operators; o++) { 2167 CeedElemRestriction active_elem_rstr, point_block_active_elem_rstr; 2168 CeedInt comp_stride, num_elem, elem_size; 2169 const CeedInt *offsets, *point_block_offsets; 2170 2171 CeedCall(CeedOperatorGetActiveElemRestriction(sub_operators[o], &active_elem_rstr)); 2172 CeedCall(CeedElemRestrictionGetCompStride(active_elem_rstr, &comp_stride)); 2173 CeedCall(CeedElemRestrictionGetNumElements(active_elem_rstr, &num_elem)); 2174 CeedCall(CeedElemRestrictionGetElementSize(active_elem_rstr, &elem_size)); 2175 CeedCall(CeedElemRestrictionGetOffsets(active_elem_rstr, CEED_MEM_HOST, &offsets)); 2176 2177 CeedCall(CeedOperatorCreateActivePointBlockRestriction(active_elem_rstr, &point_block_active_elem_rstr)); 2178 CeedCall(CeedElemRestrictionGetOffsets(point_block_active_elem_rstr, CEED_MEM_HOST, &point_block_offsets)); 2179 2180 for (CeedSize i = 0; i < num_elem * elem_size; i++) { 2181 for (CeedInt c_out = 0; c_out < num_active_components; c_out++) { 2182 for (CeedInt c_in = 0; c_in < num_active_components; c_in++) { 2183 (*rows)[point_block_offsets[i] + c_out * num_active_components + c_in] = offsets[i] + c_out * comp_stride; 2184 (*cols)[point_block_offsets[i] + c_out * num_active_components + c_in] = offsets[i] + c_in * comp_stride; 2185 } 2186 } 2187 } 2188 2189 CeedCall(CeedElemRestrictionRestoreOffsets(active_elem_rstr, &offsets)); 2190 CeedCall(CeedElemRestrictionRestoreOffsets(point_block_active_elem_rstr, &point_block_offsets)); 2191 CeedCall(CeedElemRestrictionDestroy(&active_elem_rstr)); 2192 CeedCall(CeedElemRestrictionDestroy(&point_block_active_elem_rstr)); 2193 } 2194 return CEED_ERROR_SUCCESS; 2195 } 2196 2197 /** 2198 @brief Assemble the point block diagonal of a square linear `CeedOperator`. 2199 2200 This overwrites a `CeedVector` with the point block diagonal of a linear `CeedOperator`. 2201 2202 Note: Currently only non-composite `CeedOperator` with a single field and composite `CeedOperator` with single field sub-operators are supported. 2203 2204 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2205 2206 @param[in] op `CeedOperator` to assemble `CeedQFunction` 2207 @param[out] assembled `CeedVector` to store assembled `CeedOperator` point block diagonal, provided in row-major form with an `num_comp * num_comp` block at each node. 2208 The dimensions of this vector are derived from the active vector for the `CeedOperator`. 2209 The array has shape `[nodes, component out, component in]`. 2210 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 2211 2212 @return An error code: 0 - success, otherwise - failure 2213 2214 @ref User 2215 **/ 2216 int CeedOperatorLinearAssemblePointBlockDiagonal(CeedOperator op, CeedVector assembled, CeedRequest *request) { 2217 bool is_composite; 2218 CeedSize input_size = 0, output_size = 0; 2219 2220 CeedCall(CeedOperatorCheckReady(op)); 2221 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2222 2223 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2224 CeedCheck(input_size == output_size, CeedOperatorReturnCeed(op), CEED_ERROR_DIMENSION, "Operator must be square"); 2225 2226 // Early exit for empty operator 2227 if (!is_composite) { 2228 CeedInt num_elem = 0; 2229 2230 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2231 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2232 } 2233 2234 if (op->LinearAssemblePointBlockDiagonal) { 2235 // Backend version 2236 CeedCall(op->LinearAssemblePointBlockDiagonal(op, assembled, request)); 2237 return CEED_ERROR_SUCCESS; 2238 } else if (op->LinearAssembleAddPointBlockDiagonal) { 2239 // Backend version with zeroing first 2240 CeedCall(CeedVectorSetValue(assembled, 0.0)); 2241 CeedCall(CeedOperatorLinearAssembleAddPointBlockDiagonal(op, assembled, request)); 2242 return CEED_ERROR_SUCCESS; 2243 } else { 2244 // Operator fallback 2245 CeedOperator op_fallback; 2246 2247 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2248 if (op_fallback) { 2249 CeedCall(CeedOperatorLinearAssemblePointBlockDiagonal(op_fallback, assembled, request)); 2250 return CEED_ERROR_SUCCESS; 2251 } 2252 } 2253 // Default interface implementation 2254 CeedCall(CeedVectorSetValue(assembled, 0.0)); 2255 CeedCall(CeedOperatorLinearAssembleAddPointBlockDiagonal(op, assembled, request)); 2256 return CEED_ERROR_SUCCESS; 2257 } 2258 2259 /** 2260 @brief Assemble the point block diagonal of a square linear `CeedOperator`. 2261 2262 This sums into a `CeedVector` with the point block diagonal of a linear `CeedOperator`. 2263 2264 Note: Currently only non-composite `CeedOperator` with a single field and composite `CeedOperator` with single field sub-operators are supported. 2265 2266 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2267 2268 @param[in] op `CeedOperator` to assemble `CeedQFunction` 2269 @param[out] assembled `CeedVector` to store assembled CeedOperator point block diagonal, provided in row-major form with an `num_comp * num_comp` block at each node. 2270 The dimensions of this vector are derived from the active vector for the `CeedOperator`. 2271 The array has shape `[nodes, component out, component in]`. 2272 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 2273 2274 @return An error code: 0 - success, otherwise - failure 2275 2276 @ref User 2277 **/ 2278 int CeedOperatorLinearAssembleAddPointBlockDiagonal(CeedOperator op, CeedVector assembled, CeedRequest *request) { 2279 bool is_composite; 2280 CeedSize input_size = 0, output_size = 0; 2281 2282 CeedCall(CeedOperatorCheckReady(op)); 2283 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2284 2285 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2286 CeedCheck(input_size == output_size, CeedOperatorReturnCeed(op), CEED_ERROR_DIMENSION, "Operator must be square"); 2287 2288 // Early exit for empty operator 2289 if (!is_composite) { 2290 CeedInt num_elem = 0; 2291 2292 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2293 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2294 } 2295 2296 if (op->LinearAssembleAddPointBlockDiagonal) { 2297 // Backend version 2298 CeedCall(op->LinearAssembleAddPointBlockDiagonal(op, assembled, request)); 2299 return CEED_ERROR_SUCCESS; 2300 } else { 2301 // Operator fallback 2302 CeedOperator op_fallback; 2303 2304 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2305 if (op_fallback) { 2306 CeedCall(CeedOperatorLinearAssembleAddPointBlockDiagonal(op_fallback, assembled, request)); 2307 return CEED_ERROR_SUCCESS; 2308 } 2309 } 2310 // Default interface implementation 2311 if (is_composite) { 2312 CeedCall(CeedCompositeOperatorLinearAssembleAddDiagonal(op, request, true, assembled)); 2313 } else { 2314 CeedCall(CeedSingleOperatorLinearAssembleAddDiagonal(op, request, true, assembled)); 2315 } 2316 return CEED_ERROR_SUCCESS; 2317 } 2318 2319 /** 2320 @brief Fully assemble the nonzero pattern of a linear `CeedOperator`. 2321 2322 Expected to be used in conjunction with @ref CeedOperatorLinearAssemble(). 2323 2324 The assembly routines use coordinate format, with `num_entries` tuples of the form `(i, j, value)` which indicate that value should be added to the matrix in entry `(i, j)`. 2325 Note that the `(i, j)` pairs are not unique and may repeat. 2326 This function returns the number of entries and their `(i, j)` locations, while @ref CeedOperatorLinearAssemble() provides the values in the same ordering. 2327 2328 This will generally be slow unless your operator is low-order. 2329 2330 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2331 2332 @param[in] op `CeedOperator` to assemble 2333 @param[out] num_entries Number of entries in coordinate nonzero pattern 2334 @param[out] rows Row number for each entry 2335 @param[out] cols Column number for each entry 2336 2337 @ref User 2338 **/ 2339 int CeedOperatorLinearAssembleSymbolic(CeedOperator op, CeedSize *num_entries, CeedInt **rows, CeedInt **cols) { 2340 bool is_composite; 2341 CeedInt num_suboperators, offset = 0; 2342 CeedSize single_entries; 2343 CeedOperator *sub_operators; 2344 2345 CeedCall(CeedOperatorCheckReady(op)); 2346 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2347 2348 if (op->LinearAssembleSymbolic) { 2349 // Backend version 2350 CeedCall(op->LinearAssembleSymbolic(op, num_entries, rows, cols)); 2351 return CEED_ERROR_SUCCESS; 2352 } else { 2353 // Operator fallback 2354 CeedOperator op_fallback; 2355 2356 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2357 if (op_fallback) { 2358 CeedCall(CeedOperatorLinearAssembleSymbolic(op_fallback, num_entries, rows, cols)); 2359 return CEED_ERROR_SUCCESS; 2360 } 2361 } 2362 2363 // Default interface implementation 2364 2365 // Count entries and allocate rows, cols arrays 2366 *num_entries = 0; 2367 if (is_composite) { 2368 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 2369 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2370 for (CeedInt k = 0; k < num_suboperators; ++k) { 2371 CeedCall(CeedSingleOperatorAssemblyCountEntries(sub_operators[k], &single_entries)); 2372 *num_entries += single_entries; 2373 } 2374 } else { 2375 CeedCall(CeedSingleOperatorAssemblyCountEntries(op, &single_entries)); 2376 *num_entries += single_entries; 2377 } 2378 CeedCall(CeedCalloc(*num_entries, rows)); 2379 CeedCall(CeedCalloc(*num_entries, cols)); 2380 2381 // Assemble nonzero locations 2382 if (is_composite) { 2383 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 2384 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2385 for (CeedInt k = 0; k < num_suboperators; ++k) { 2386 CeedCall(CeedSingleOperatorAssembleSymbolic(sub_operators[k], offset, *rows, *cols)); 2387 CeedCall(CeedSingleOperatorAssemblyCountEntries(sub_operators[k], &single_entries)); 2388 offset += single_entries; 2389 } 2390 } else { 2391 CeedCall(CeedSingleOperatorAssembleSymbolic(op, offset, *rows, *cols)); 2392 } 2393 return CEED_ERROR_SUCCESS; 2394 } 2395 2396 /** 2397 @brief Fully assemble the nonzero entries of a linear operator. 2398 2399 Expected to be used in conjunction with @ref CeedOperatorLinearAssembleSymbolic(). 2400 2401 The assembly routines use coordinate format, with `num_entries` tuples of the form `(i, j, value)` which indicate that value should be added to the matrix in entry `(i, j)`. 2402 Note that the `(i, j)` pairs are not unique and may repeat. 2403 This function returns the values of the nonzero entries to be added, their `(i, j)` locations are provided by @ref CeedOperatorLinearAssembleSymbolic(). 2404 2405 This will generally be slow unless your operator is low-order. 2406 2407 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2408 2409 @param[in] op `CeedOperator` to assemble 2410 @param[out] values Values to assemble into matrix 2411 2412 @ref User 2413 **/ 2414 int CeedOperatorLinearAssemble(CeedOperator op, CeedVector values) { 2415 bool is_composite; 2416 CeedInt num_suboperators, offset = 0; 2417 CeedSize single_entries = 0; 2418 CeedOperator *sub_operators; 2419 2420 CeedCall(CeedOperatorCheckReady(op)); 2421 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2422 2423 // Early exit for empty operator 2424 if (!is_composite) { 2425 CeedInt num_elem = 0; 2426 2427 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2428 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2429 } 2430 2431 if (op->LinearAssemble) { 2432 // Backend version 2433 CeedCall(op->LinearAssemble(op, values)); 2434 return CEED_ERROR_SUCCESS; 2435 } else { 2436 // Operator fallback 2437 CeedOperator op_fallback; 2438 2439 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2440 if (op_fallback) { 2441 CeedCall(CeedOperatorLinearAssemble(op_fallback, values)); 2442 return CEED_ERROR_SUCCESS; 2443 } 2444 } 2445 2446 // Default interface implementation 2447 CeedCall(CeedVectorSetValue(values, 0.0)); 2448 if (is_composite) { 2449 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 2450 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2451 for (CeedInt k = 0; k < num_suboperators; k++) { 2452 CeedCall(CeedSingleOperatorAssemble(sub_operators[k], offset, values)); 2453 CeedCall(CeedSingleOperatorAssemblyCountEntries(sub_operators[k], &single_entries)); 2454 offset += single_entries; 2455 } 2456 } else { 2457 CeedCall(CeedSingleOperatorAssemble(op, offset, values)); 2458 } 2459 return CEED_ERROR_SUCCESS; 2460 } 2461 2462 /** 2463 @brief Get the multiplicity of nodes across sub-operators in a composite `CeedOperator`. 2464 2465 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2466 2467 @param[in] op Composite `CeedOperator` 2468 @param[in] num_skip_indices Number of sub-operators to skip 2469 @param[in] skip_indices Array of indices of sub-operators to skip 2470 @param[out] mult Vector to store multiplicity (of size `l_size` ) 2471 2472 @return An error code: 0 - success, otherwise - failure 2473 2474 @ref User 2475 **/ 2476 int CeedCompositeOperatorGetMultiplicity(CeedOperator op, CeedInt num_skip_indices, CeedInt *skip_indices, CeedVector mult) { 2477 Ceed ceed; 2478 CeedInt num_suboperators; 2479 CeedSize l_vec_len; 2480 CeedScalar *mult_array; 2481 CeedVector ones_l_vec; 2482 CeedElemRestriction elem_rstr, mult_elem_rstr; 2483 CeedOperator *sub_operators; 2484 2485 CeedCall(CeedOperatorCheckReady(op)); 2486 2487 // Zero mult vector 2488 CeedCall(CeedVectorSetValue(mult, 0.0)); 2489 2490 // Get suboperators 2491 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 2492 if (num_suboperators == 0) return CEED_ERROR_SUCCESS; 2493 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2494 2495 // Work vector 2496 CeedCall(CeedVectorGetLength(mult, &l_vec_len)); 2497 CeedCall(CeedOperatorGetCeed(op, &ceed)); 2498 CeedCall(CeedVectorCreate(ceed, l_vec_len, &ones_l_vec)); 2499 CeedCall(CeedDestroy(&ceed)); 2500 CeedCall(CeedVectorSetValue(ones_l_vec, 1.0)); 2501 CeedCall(CeedVectorGetArray(mult, CEED_MEM_HOST, &mult_array)); 2502 2503 // Compute multiplicity across suboperators 2504 for (CeedInt i = 0; i < num_suboperators; i++) { 2505 const CeedScalar *sub_mult_array; 2506 CeedVector sub_mult_l_vec, ones_e_vec; 2507 2508 // -- Check for suboperator to skip 2509 for (CeedInt j = 0; j < num_skip_indices; j++) { 2510 if (skip_indices[j] == i) continue; 2511 } 2512 2513 // -- Sub operator multiplicity 2514 CeedCall(CeedOperatorGetActiveElemRestriction(sub_operators[i], &elem_rstr)); 2515 CeedCall(CeedElemRestrictionCreateUnorientedCopy(elem_rstr, &mult_elem_rstr)); 2516 CeedCall(CeedElemRestrictionDestroy(&elem_rstr)); 2517 CeedCall(CeedElemRestrictionCreateVector(mult_elem_rstr, &sub_mult_l_vec, &ones_e_vec)); 2518 CeedCall(CeedVectorSetValue(sub_mult_l_vec, 0.0)); 2519 CeedCall(CeedElemRestrictionApply(mult_elem_rstr, CEED_NOTRANSPOSE, ones_l_vec, ones_e_vec, CEED_REQUEST_IMMEDIATE)); 2520 CeedCall(CeedElemRestrictionApply(mult_elem_rstr, CEED_TRANSPOSE, ones_e_vec, sub_mult_l_vec, CEED_REQUEST_IMMEDIATE)); 2521 CeedCall(CeedVectorGetArrayRead(sub_mult_l_vec, CEED_MEM_HOST, &sub_mult_array)); 2522 // ---- Flag every node present in the current suboperator 2523 for (CeedSize j = 0; j < l_vec_len; j++) { 2524 if (sub_mult_array[j] > 0.0) mult_array[j] += 1.0; 2525 } 2526 CeedCall(CeedVectorRestoreArrayRead(sub_mult_l_vec, &sub_mult_array)); 2527 CeedCall(CeedVectorDestroy(&sub_mult_l_vec)); 2528 CeedCall(CeedVectorDestroy(&ones_e_vec)); 2529 CeedCall(CeedElemRestrictionDestroy(&mult_elem_rstr)); 2530 } 2531 CeedCall(CeedVectorRestoreArray(mult, &mult_array)); 2532 CeedCall(CeedVectorDestroy(&ones_l_vec)); 2533 return CEED_ERROR_SUCCESS; 2534 } 2535 2536 /** 2537 @brief Create a multigrid coarse `CeedOperator` and level transfer `CeedOperator` for a `CeedOperator`, creating the prolongation basis from the fine and coarse grid interpolation. 2538 2539 Note: Calling this function asserts that setup is complete and sets all four `CeedOperator` as immutable. 2540 2541 @param[in] op_fine Fine grid `CeedOperator` 2542 @param[in] p_mult_fine L-vector multiplicity in parallel gather/scatter, or `NULL` if not creating prolongation/restriction `CeedOperator` 2543 @param[in] rstr_coarse Coarse grid `CeedElemRestriction` 2544 @param[in] basis_coarse Coarse grid active vector `CeedBasis` 2545 @param[out] op_coarse Coarse grid `CeedOperator` 2546 @param[out] op_prolong Coarse to fine `CeedOperator`, or `NULL` 2547 @param[out] op_restrict Fine to coarse `CeedOperator`, or `NULL` 2548 2549 @return An error code: 0 - success, otherwise - failure 2550 2551 @ref User 2552 **/ 2553 int CeedOperatorMultigridLevelCreate(CeedOperator op_fine, CeedVector p_mult_fine, CeedElemRestriction rstr_coarse, CeedBasis basis_coarse, 2554 CeedOperator *op_coarse, CeedOperator *op_prolong, CeedOperator *op_restrict) { 2555 CeedBasis basis_c_to_f = NULL; 2556 2557 CeedCall(CeedOperatorCheckReady(op_fine)); 2558 2559 // Build prolongation matrix, if required 2560 if (op_prolong || op_restrict) { 2561 CeedBasis basis_fine; 2562 2563 CeedCall(CeedOperatorGetActiveBasis(op_fine, &basis_fine)); 2564 CeedCall(CeedBasisCreateProjection(basis_coarse, basis_fine, &basis_c_to_f)); 2565 CeedCall(CeedBasisDestroy(&basis_fine)); 2566 } 2567 2568 // Core code 2569 CeedCall(CeedSingleOperatorMultigridLevel(op_fine, p_mult_fine, rstr_coarse, basis_coarse, basis_c_to_f, op_coarse, op_prolong, op_restrict)); 2570 return CEED_ERROR_SUCCESS; 2571 } 2572 2573 /** 2574 @brief Create a multigrid coarse `CeedOperator` and level transfer `CeedOperator` for a `CeedOperator` with a tensor basis for the active basis. 2575 2576 Note: Calling this function asserts that setup is complete and sets all four `CeedOperator` as immutable. 2577 2578 @param[in] op_fine Fine grid `CeedOperator` 2579 @param[in] p_mult_fine L-vector multiplicity in parallel gather/scatter, or `NULL` if not creating prolongation/restriction `CeedOperator` 2580 @param[in] rstr_coarse Coarse grid `CeedElemRestriction` 2581 @param[in] basis_coarse Coarse grid active vector `CeedBasis` 2582 @param[in] interp_c_to_f Matrix for coarse to fine interpolation, or `NULL` if not creating prolongation/restriction `CeedOperator` 2583 @param[out] op_coarse Coarse grid `CeedOperator` 2584 @param[out] op_prolong Coarse to fine `CeedOperator`, or `NULL` 2585 @param[out] op_restrict Fine to coarse `CeedOperator`, or `NULL` 2586 2587 @return An error code: 0 - success, otherwise - failure 2588 2589 @ref User 2590 **/ 2591 int CeedOperatorMultigridLevelCreateTensorH1(CeedOperator op_fine, CeedVector p_mult_fine, CeedElemRestriction rstr_coarse, CeedBasis basis_coarse, 2592 const CeedScalar *interp_c_to_f, CeedOperator *op_coarse, CeedOperator *op_prolong, 2593 CeedOperator *op_restrict) { 2594 Ceed ceed; 2595 CeedInt Q_f, Q_c; 2596 CeedBasis basis_fine, basis_c_to_f = NULL; 2597 2598 CeedCall(CeedOperatorCheckReady(op_fine)); 2599 CeedCall(CeedOperatorGetCeed(op_fine, &ceed)); 2600 2601 // Check for compatible quadrature spaces 2602 CeedCall(CeedOperatorGetActiveBasis(op_fine, &basis_fine)); 2603 CeedCall(CeedBasisGetNumQuadraturePoints(basis_fine, &Q_f)); 2604 CeedCall(CeedBasisGetNumQuadraturePoints(basis_coarse, &Q_c)); 2605 CeedCheck(Q_f == Q_c, ceed, CEED_ERROR_DIMENSION, 2606 "Bases must have compatible quadrature spaces." 2607 " Fine grid: %" CeedInt_FMT " points, Coarse grid: %" CeedInt_FMT " points", 2608 Q_f, Q_c); 2609 2610 // Create coarse to fine basis, if required 2611 if (op_prolong || op_restrict) { 2612 CeedInt dim, num_comp, num_nodes_c, P_1d_f, P_1d_c; 2613 CeedScalar *q_ref, *q_weight, *grad; 2614 2615 // Check if interpolation matrix is provided 2616 CeedCheck(interp_c_to_f, ceed, CEED_ERROR_INCOMPATIBLE, 2617 "Prolongation or restriction operator creation requires coarse-to-fine interpolation matrix"); 2618 CeedCall(CeedBasisGetDimension(basis_fine, &dim)); 2619 CeedCall(CeedBasisGetNumComponents(basis_fine, &num_comp)); 2620 CeedCall(CeedBasisGetNumNodes1D(basis_fine, &P_1d_f)); 2621 CeedCall(CeedBasisDestroy(&basis_fine)); 2622 CeedCall(CeedElemRestrictionGetElementSize(rstr_coarse, &num_nodes_c)); 2623 P_1d_c = dim == 1 ? num_nodes_c : dim == 2 ? sqrt(num_nodes_c) : cbrt(num_nodes_c); 2624 CeedCall(CeedCalloc(P_1d_f, &q_ref)); 2625 CeedCall(CeedCalloc(P_1d_f, &q_weight)); 2626 CeedCall(CeedCalloc(P_1d_f * P_1d_c * dim, &grad)); 2627 CeedCall(CeedBasisCreateTensorH1(ceed, dim, num_comp, P_1d_c, P_1d_f, interp_c_to_f, grad, q_ref, q_weight, &basis_c_to_f)); 2628 CeedCall(CeedFree(&q_ref)); 2629 CeedCall(CeedFree(&q_weight)); 2630 CeedCall(CeedFree(&grad)); 2631 } 2632 2633 // Core code 2634 CeedCall(CeedSingleOperatorMultigridLevel(op_fine, p_mult_fine, rstr_coarse, basis_coarse, basis_c_to_f, op_coarse, op_prolong, op_restrict)); 2635 CeedCall(CeedDestroy(&ceed)); 2636 return CEED_ERROR_SUCCESS; 2637 } 2638 2639 /** 2640 @brief Create a multigrid coarse `CeedOperator` and level transfer `CeedOperator` for a `CeedOperator` with a non-tensor basis for the active vector 2641 2642 Note: Calling this function asserts that setup is complete and sets all four `CeedOperator` as immutable. 2643 2644 @param[in] op_fine Fine grid `CeedOperator` 2645 @param[in] p_mult_fine L-vector multiplicity in parallel gather/scatter, or `NULL` if not creating prolongation/restriction `CeedOperator` 2646 @param[in] rstr_coarse Coarse grid `CeedElemRestriction` 2647 @param[in] basis_coarse Coarse grid active vector `CeedBasis` 2648 @param[in] interp_c_to_f Matrix for coarse to fine interpolation, or `NULL` if not creating prolongation/restriction `CeedOperator` 2649 @param[out] op_coarse Coarse grid `CeedOperator` 2650 @param[out] op_prolong Coarse to fine `CeedOperator`, or `NULL` 2651 @param[out] op_restrict Fine to coarse `CeedOperator`, or `NULL` 2652 2653 @return An error code: 0 - success, otherwise - failure 2654 2655 @ref User 2656 **/ 2657 int CeedOperatorMultigridLevelCreateH1(CeedOperator op_fine, CeedVector p_mult_fine, CeedElemRestriction rstr_coarse, CeedBasis basis_coarse, 2658 const CeedScalar *interp_c_to_f, CeedOperator *op_coarse, CeedOperator *op_prolong, 2659 CeedOperator *op_restrict) { 2660 Ceed ceed; 2661 CeedInt Q_f, Q_c; 2662 CeedBasis basis_fine, basis_c_to_f = NULL; 2663 2664 CeedCall(CeedOperatorCheckReady(op_fine)); 2665 CeedCall(CeedOperatorGetCeed(op_fine, &ceed)); 2666 2667 // Check for compatible quadrature spaces 2668 CeedCall(CeedOperatorGetActiveBasis(op_fine, &basis_fine)); 2669 CeedCall(CeedBasisGetNumQuadraturePoints(basis_fine, &Q_f)); 2670 CeedCall(CeedBasisGetNumQuadraturePoints(basis_coarse, &Q_c)); 2671 CeedCheck(Q_f == Q_c, ceed, CEED_ERROR_DIMENSION, "Bases must have compatible quadrature spaces"); 2672 2673 // Coarse to fine basis 2674 if (op_prolong || op_restrict) { 2675 CeedInt dim, num_comp, num_nodes_c, num_nodes_f; 2676 CeedScalar *q_ref, *q_weight, *grad; 2677 CeedElemTopology topo; 2678 2679 // Check if interpolation matrix is provided 2680 CeedCheck(interp_c_to_f, ceed, CEED_ERROR_INCOMPATIBLE, 2681 "Prolongation or restriction operator creation requires coarse-to-fine interpolation matrix"); 2682 CeedCall(CeedBasisGetTopology(basis_fine, &topo)); 2683 CeedCall(CeedBasisGetDimension(basis_fine, &dim)); 2684 CeedCall(CeedBasisGetNumComponents(basis_fine, &num_comp)); 2685 CeedCall(CeedBasisGetNumNodes(basis_fine, &num_nodes_f)); 2686 CeedCall(CeedBasisDestroy(&basis_fine)); 2687 CeedCall(CeedElemRestrictionGetElementSize(rstr_coarse, &num_nodes_c)); 2688 CeedCall(CeedCalloc(num_nodes_f * dim, &q_ref)); 2689 CeedCall(CeedCalloc(num_nodes_f, &q_weight)); 2690 CeedCall(CeedCalloc(num_nodes_f * num_nodes_c * dim, &grad)); 2691 CeedCall(CeedBasisCreateH1(ceed, topo, num_comp, num_nodes_c, num_nodes_f, interp_c_to_f, grad, q_ref, q_weight, &basis_c_to_f)); 2692 CeedCall(CeedFree(&q_ref)); 2693 CeedCall(CeedFree(&q_weight)); 2694 CeedCall(CeedFree(&grad)); 2695 } 2696 2697 // Core code 2698 CeedCall(CeedSingleOperatorMultigridLevel(op_fine, p_mult_fine, rstr_coarse, basis_coarse, basis_c_to_f, op_coarse, op_prolong, op_restrict)); 2699 CeedCall(CeedDestroy(&ceed)); 2700 return CEED_ERROR_SUCCESS; 2701 } 2702 2703 /** 2704 @brief Build a FDM based approximate inverse for each element for a `CeedOperator`. 2705 2706 This returns a `CeedOperator` and `CeedVector` to apply a Fast Diagonalization Method based approximate inverse. 2707 This function obtains the simultaneous diagonalization for the 1D mass and Laplacian operators, \f$M = V^T V, K = V^T S V\f$. 2708 The assembled `CeedQFunction` is used to modify the eigenvalues from simultaneous diagonalization and obtain an approximate inverse of the form \f$V^T \hat S V\f$. 2709 The `CeedOperator` must be linear and non-composite. 2710 The associated `CeedQFunction` must therefore also be linear. 2711 2712 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2713 2714 @param[in] op `CeedOperator` to create element inverses 2715 @param[out] fdm_inv `CeedOperator` to apply the action of a FDM based inverse for each element 2716 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 2717 2718 @return An error code: 0 - success, otherwise - failure 2719 2720 @ref User 2721 **/ 2722 int CeedOperatorCreateFDMElementInverse(CeedOperator op, CeedOperator *fdm_inv, CeedRequest *request) { 2723 Ceed ceed, ceed_parent; 2724 bool interp = false, grad = false, is_tensor_basis = true; 2725 CeedInt num_input_fields, P_1d, Q_1d, num_nodes, num_qpts, dim, num_comp = 1, num_elem = 1; 2726 CeedScalar *mass, *laplace, *x, *fdm_interp, *lambda, *elem_avg; 2727 const CeedScalar *interp_1d, *grad_1d, *q_weight_1d; 2728 CeedVector q_data; 2729 CeedElemRestriction rstr = NULL, rstr_qd_i; 2730 CeedBasis basis = NULL, fdm_basis; 2731 CeedQFunctionContext ctx_fdm; 2732 CeedQFunctionField *qf_fields; 2733 CeedQFunction qf, qf_fdm; 2734 CeedOperatorField *op_fields; 2735 2736 CeedCall(CeedOperatorCheckReady(op)); 2737 2738 if (op->CreateFDMElementInverse) { 2739 // Backend version 2740 CeedCall(op->CreateFDMElementInverse(op, fdm_inv, request)); 2741 return CEED_ERROR_SUCCESS; 2742 } else { 2743 // Operator fallback 2744 CeedOperator op_fallback; 2745 2746 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2747 if (op_fallback) { 2748 CeedCall(CeedOperatorCreateFDMElementInverse(op_fallback, fdm_inv, request)); 2749 return CEED_ERROR_SUCCESS; 2750 } 2751 } 2752 2753 // Default interface implementation 2754 CeedCall(CeedOperatorGetCeed(op, &ceed)); 2755 CeedCall(CeedOperatorGetFallbackParentCeed(op, &ceed_parent)); 2756 CeedCall(CeedOperatorGetQFunction(op, &qf)); 2757 2758 // Determine active input basis 2759 CeedCall(CeedOperatorGetFields(op, &num_input_fields, &op_fields, NULL, NULL)); 2760 CeedCall(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 2761 for (CeedInt i = 0; i < num_input_fields; i++) { 2762 CeedVector vec; 2763 2764 CeedCall(CeedOperatorFieldGetVector(op_fields[i], &vec)); 2765 if (vec == CEED_VECTOR_ACTIVE) { 2766 CeedEvalMode eval_mode; 2767 2768 CeedCall(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 2769 interp = interp || eval_mode == CEED_EVAL_INTERP; 2770 grad = grad || eval_mode == CEED_EVAL_GRAD; 2771 if (!basis) CeedCall(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 2772 if (!rstr) CeedCall(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr)); 2773 } 2774 CeedCall(CeedVectorDestroy(&vec)); 2775 } 2776 CeedCheck(basis, ceed, CEED_ERROR_BACKEND, "No active field set"); 2777 CeedCall(CeedBasisGetNumNodes1D(basis, &P_1d)); 2778 CeedCall(CeedBasisGetNumNodes(basis, &num_nodes)); 2779 CeedCall(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 2780 CeedCall(CeedBasisGetNumQuadraturePoints(basis, &num_qpts)); 2781 CeedCall(CeedBasisGetDimension(basis, &dim)); 2782 CeedCall(CeedBasisGetNumComponents(basis, &num_comp)); 2783 CeedCall(CeedElemRestrictionGetNumElements(rstr, &num_elem)); 2784 2785 // Build and diagonalize 1D Mass and Laplacian 2786 CeedCall(CeedBasisIsTensor(basis, &is_tensor_basis)); 2787 CeedCheck(is_tensor_basis, ceed, CEED_ERROR_BACKEND, "FDMElementInverse only supported for tensor bases"); 2788 CeedCall(CeedCalloc(P_1d * P_1d, &mass)); 2789 CeedCall(CeedCalloc(P_1d * P_1d, &laplace)); 2790 CeedCall(CeedCalloc(P_1d * P_1d, &x)); 2791 CeedCall(CeedCalloc(P_1d * P_1d, &fdm_interp)); 2792 CeedCall(CeedCalloc(P_1d, &lambda)); 2793 // -- Build matrices 2794 CeedCall(CeedBasisGetInterp1D(basis, &interp_1d)); 2795 CeedCall(CeedBasisGetGrad1D(basis, &grad_1d)); 2796 CeedCall(CeedBasisGetQWeights(basis, &q_weight_1d)); 2797 CeedCall(CeedBuildMassLaplace(interp_1d, grad_1d, q_weight_1d, P_1d, Q_1d, dim, mass, laplace)); 2798 2799 // -- Diagonalize 2800 CeedCall(CeedSimultaneousDiagonalization(ceed, laplace, mass, x, lambda, P_1d)); 2801 CeedCall(CeedFree(&mass)); 2802 CeedCall(CeedFree(&laplace)); 2803 for (CeedInt i = 0; i < P_1d; i++) { 2804 for (CeedInt j = 0; j < P_1d; j++) fdm_interp[i + j * P_1d] = x[j + i * P_1d]; 2805 } 2806 CeedCall(CeedFree(&x)); 2807 2808 { 2809 CeedInt layout[3], num_modes = (interp ? 1 : 0) + (grad ? dim : 0); 2810 CeedScalar max_norm = 0; 2811 const CeedScalar *assembled_array, *q_weight_array; 2812 CeedVector assembled = NULL, q_weight; 2813 CeedElemRestriction rstr_qf = NULL; 2814 2815 // Assemble QFunction 2816 CeedCall(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled, &rstr_qf, request)); 2817 CeedCall(CeedElemRestrictionGetELayout(rstr_qf, layout)); 2818 CeedCall(CeedElemRestrictionDestroy(&rstr_qf)); 2819 CeedCall(CeedVectorNorm(assembled, CEED_NORM_MAX, &max_norm)); 2820 2821 // Calculate element averages 2822 CeedCall(CeedVectorCreate(ceed_parent, num_qpts, &q_weight)); 2823 CeedCall(CeedBasisApply(basis, 1, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, q_weight)); 2824 CeedCall(CeedVectorGetArrayRead(assembled, CEED_MEM_HOST, &assembled_array)); 2825 CeedCall(CeedVectorGetArrayRead(q_weight, CEED_MEM_HOST, &q_weight_array)); 2826 CeedCall(CeedCalloc(num_elem, &elem_avg)); 2827 const CeedScalar qf_value_bound = max_norm * 100 * CEED_EPSILON; 2828 2829 for (CeedInt e = 0; e < num_elem; e++) { 2830 CeedInt count = 0; 2831 2832 for (CeedInt q = 0; q < num_qpts; q++) { 2833 for (CeedInt i = 0; i < num_comp * num_comp * num_modes * num_modes; i++) { 2834 if (fabs(assembled_array[q * layout[0] + i * layout[1] + e * layout[2]]) > qf_value_bound) { 2835 elem_avg[e] += assembled_array[q * layout[0] + i * layout[1] + e * layout[2]] / q_weight_array[q]; 2836 count++; 2837 } 2838 } 2839 } 2840 if (count) { 2841 elem_avg[e] /= count; 2842 } else { 2843 elem_avg[e] = 1.0; 2844 } 2845 } 2846 CeedCall(CeedVectorRestoreArrayRead(assembled, &assembled_array)); 2847 CeedCall(CeedVectorDestroy(&assembled)); 2848 CeedCall(CeedVectorRestoreArrayRead(q_weight, &q_weight_array)); 2849 CeedCall(CeedVectorDestroy(&q_weight)); 2850 } 2851 2852 // Build FDM diagonal 2853 { 2854 CeedScalar *q_data_array, *fdm_diagonal; 2855 2856 CeedCall(CeedCalloc(num_comp * num_nodes, &fdm_diagonal)); 2857 const CeedScalar fdm_diagonal_bound = num_nodes * CEED_EPSILON; 2858 for (CeedInt c = 0; c < num_comp; c++) { 2859 for (CeedInt n = 0; n < num_nodes; n++) { 2860 if (interp) fdm_diagonal[c * num_nodes + n] = 1.0; 2861 if (grad) { 2862 for (CeedInt d = 0; d < dim; d++) { 2863 CeedInt i = (n / CeedIntPow(P_1d, d)) % P_1d; 2864 fdm_diagonal[c * num_nodes + n] += lambda[i]; 2865 } 2866 } 2867 if (fabs(fdm_diagonal[c * num_nodes + n]) < fdm_diagonal_bound) fdm_diagonal[c * num_nodes + n] = fdm_diagonal_bound; 2868 } 2869 } 2870 CeedCall(CeedVectorCreate(ceed_parent, num_elem * num_comp * num_nodes, &q_data)); 2871 CeedCall(CeedVectorSetValue(q_data, 0.0)); 2872 CeedCall(CeedVectorGetArrayWrite(q_data, CEED_MEM_HOST, &q_data_array)); 2873 for (CeedInt e = 0; e < num_elem; e++) { 2874 for (CeedInt c = 0; c < num_comp; c++) { 2875 for (CeedInt n = 0; n < num_nodes; n++) { 2876 q_data_array[(e * num_comp + c) * num_nodes + n] = 1. / (elem_avg[e] * fdm_diagonal[c * num_nodes + n]); 2877 } 2878 } 2879 } 2880 CeedCall(CeedFree(&elem_avg)); 2881 CeedCall(CeedFree(&fdm_diagonal)); 2882 CeedCall(CeedVectorRestoreArray(q_data, &q_data_array)); 2883 } 2884 2885 // Setup FDM operator 2886 // -- Basis 2887 { 2888 CeedScalar *grad_dummy, *q_ref_dummy, *q_weight_dummy; 2889 2890 CeedCall(CeedCalloc(P_1d * P_1d, &grad_dummy)); 2891 CeedCall(CeedCalloc(P_1d, &q_ref_dummy)); 2892 CeedCall(CeedCalloc(P_1d, &q_weight_dummy)); 2893 CeedCall(CeedBasisCreateTensorH1(ceed_parent, dim, num_comp, P_1d, P_1d, fdm_interp, grad_dummy, q_ref_dummy, q_weight_dummy, &fdm_basis)); 2894 CeedCall(CeedFree(&fdm_interp)); 2895 CeedCall(CeedFree(&grad_dummy)); 2896 CeedCall(CeedFree(&q_ref_dummy)); 2897 CeedCall(CeedFree(&q_weight_dummy)); 2898 CeedCall(CeedFree(&lambda)); 2899 } 2900 2901 // -- Restriction 2902 { 2903 CeedInt strides[3] = {1, num_nodes, num_nodes * num_comp}; 2904 CeedCall(CeedElemRestrictionCreateStrided(ceed_parent, num_elem, num_nodes, num_comp, 2905 (CeedSize)num_elem * (CeedSize)num_comp * (CeedSize)num_nodes, strides, &rstr_qd_i)); 2906 } 2907 2908 // -- QFunction 2909 CeedCall(CeedQFunctionCreateInteriorByName(ceed_parent, "Scale", &qf_fdm)); 2910 CeedCall(CeedQFunctionAddInput(qf_fdm, "input", num_comp, CEED_EVAL_INTERP)); 2911 CeedCall(CeedQFunctionAddInput(qf_fdm, "scale", num_comp, CEED_EVAL_NONE)); 2912 CeedCall(CeedQFunctionAddOutput(qf_fdm, "output", num_comp, CEED_EVAL_INTERP)); 2913 CeedCall(CeedQFunctionSetUserFlopsEstimate(qf_fdm, num_comp)); 2914 2915 // -- QFunction context 2916 { 2917 CeedInt *num_comp_data; 2918 2919 CeedCall(CeedCalloc(1, &num_comp_data)); 2920 num_comp_data[0] = num_comp; 2921 CeedCall(CeedQFunctionContextCreate(ceed, &ctx_fdm)); 2922 CeedCall(CeedQFunctionContextSetData(ctx_fdm, CEED_MEM_HOST, CEED_OWN_POINTER, sizeof(*num_comp_data), num_comp_data)); 2923 } 2924 CeedCall(CeedQFunctionSetContext(qf_fdm, ctx_fdm)); 2925 CeedCall(CeedQFunctionContextDestroy(&ctx_fdm)); 2926 2927 // -- Operator 2928 CeedCall(CeedOperatorCreate(ceed_parent, qf_fdm, NULL, NULL, fdm_inv)); 2929 CeedCall(CeedOperatorSetField(*fdm_inv, "input", rstr, fdm_basis, CEED_VECTOR_ACTIVE)); 2930 CeedCall(CeedOperatorSetField(*fdm_inv, "scale", rstr_qd_i, CEED_BASIS_NONE, q_data)); 2931 CeedCall(CeedOperatorSetField(*fdm_inv, "output", rstr, fdm_basis, CEED_VECTOR_ACTIVE)); 2932 2933 // Cleanup 2934 CeedCall(CeedDestroy(&ceed)); 2935 CeedCall(CeedDestroy(&ceed_parent)); 2936 CeedCall(CeedVectorDestroy(&q_data)); 2937 CeedCall(CeedElemRestrictionDestroy(&rstr)); 2938 CeedCall(CeedElemRestrictionDestroy(&rstr_qd_i)); 2939 CeedCall(CeedBasisDestroy(&basis)); 2940 CeedCall(CeedBasisDestroy(&fdm_basis)); 2941 CeedCall(CeedQFunctionDestroy(&qf_fdm)); 2942 return CEED_ERROR_SUCCESS; 2943 } 2944 2945 /// @} 2946