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 Ceed ceed; 41 CeedQFunctionField *input_fields, *output_fields; 42 43 // Check if NULL qf passed in 44 if (!qf) return CEED_ERROR_SUCCESS; 45 46 CeedCall(CeedQFunctionGetCeed(qf, &ceed)); 47 CeedDebug256(ceed, 1, "---------- CeedOperator Fallback ----------\n"); 48 CeedDebug(ceed, "Creating fallback CeedQFunction\n"); 49 50 if (qf->source_path) { 51 size_t path_len = strlen(qf->source_path), name_len = strlen(qf->kernel_name); 52 53 CeedCall(CeedCalloc(path_len + name_len + 2, &source_path_with_name)); 54 memcpy(source_path_with_name, qf->source_path, path_len); 55 memcpy(&source_path_with_name[path_len], ":", 1); 56 memcpy(&source_path_with_name[path_len + 1], qf->kernel_name, name_len); 57 } else if (qf->user_source) { 58 CeedCall(CeedStringAllocCopy(qf->user_source, &source_path_with_name)); 59 } else { 60 CeedCall(CeedCalloc(1, &source_path_with_name)); 61 } 62 63 { 64 CeedInt vec_length; 65 CeedQFunctionUser f; 66 67 CeedCall(CeedQFunctionGetVectorLength(qf, &vec_length)); 68 CeedCall(CeedQFunctionGetUserFunction(qf, &f)); 69 CeedCall(CeedQFunctionCreateInterior(fallback_ceed, vec_length, f, source_path_with_name, qf_fallback)); 70 } 71 { 72 CeedQFunctionContext ctx; 73 74 CeedCall(CeedQFunctionGetContext(qf, &ctx)); 75 CeedCall(CeedQFunctionSetContext(*qf_fallback, ctx)); 76 } 77 CeedCall(CeedQFunctionGetFields(qf, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 78 for (CeedInt i = 0; i < num_input_fields; i++) { 79 const char *field_name; 80 CeedInt size; 81 CeedEvalMode eval_mode; 82 83 CeedCall(CeedQFunctionFieldGetData(input_fields[i], &field_name, &size, &eval_mode)); 84 CeedCall(CeedQFunctionAddInput(*qf_fallback, field_name, size, eval_mode)); 85 } 86 for (CeedInt i = 0; i < num_output_fields; i++) { 87 const char *field_name; 88 CeedInt size; 89 CeedEvalMode eval_mode; 90 91 CeedCall(CeedQFunctionFieldGetData(output_fields[i], &field_name, &size, &eval_mode)); 92 CeedCall(CeedQFunctionAddOutput(*qf_fallback, field_name, size, eval_mode)); 93 } 94 CeedCall(CeedFree(&source_path_with_name)); 95 return CEED_ERROR_SUCCESS; 96 } 97 98 /** 99 @brief Duplicate a `CeedOperator` with a reference `Ceed` to fallback for advanced `CeedOperator` functionality 100 101 @param[in,out] op `CeedOperator` to create fallback for 102 103 @return An error code: 0 - success, otherwise - failure 104 105 @ref Developer 106 **/ 107 static int CeedOperatorCreateFallback(CeedOperator op) { 108 bool is_composite; 109 Ceed ceed, ceed_fallback; 110 CeedOperator op_fallback; 111 112 // Check not already created 113 if (op->op_fallback) return CEED_ERROR_SUCCESS; 114 115 // Fallback Ceed 116 CeedCall(CeedOperatorGetCeed(op, &ceed)); 117 CeedCall(CeedGetOperatorFallbackCeed(ceed, &ceed_fallback)); 118 if (!ceed_fallback) return CEED_ERROR_SUCCESS; 119 120 CeedDebug256(ceed, 1, "---------- CeedOperator Fallback ----------\n"); 121 CeedDebug(ceed, "Creating fallback CeedOperator\n"); 122 123 // Clone Op 124 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 125 if (is_composite) { 126 CeedInt num_suboperators; 127 CeedOperator *sub_operators; 128 129 CeedCall(CeedCompositeOperatorCreate(ceed_fallback, &op_fallback)); 130 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 131 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 132 for (CeedInt i = 0; i < num_suboperators; i++) { 133 CeedOperator op_sub_fallback; 134 135 CeedCall(CeedOperatorGetFallback(sub_operators[i], &op_sub_fallback)); 136 CeedCall(CeedCompositeOperatorAddSub(op_fallback, op_sub_fallback)); 137 } 138 } else { 139 CeedInt num_input_fields, num_output_fields; 140 CeedQFunction qf_fallback = NULL, dqf_fallback = NULL, dqfT_fallback = NULL; 141 CeedOperatorField *input_fields, *output_fields; 142 143 CeedCall(CeedQFunctionCreateFallback(ceed_fallback, op->qf, &qf_fallback)); 144 CeedCall(CeedQFunctionCreateFallback(ceed_fallback, op->dqf, &dqf_fallback)); 145 CeedCall(CeedQFunctionCreateFallback(ceed_fallback, op->dqfT, &dqfT_fallback)); 146 CeedCall(CeedOperatorCreate(ceed_fallback, qf_fallback, dqf_fallback, dqfT_fallback, &op_fallback)); 147 CeedCall(CeedOperatorGetFields(op, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 148 for (CeedInt i = 0; i < num_input_fields; i++) { 149 const char *field_name; 150 CeedVector vec; 151 CeedElemRestriction rstr; 152 CeedBasis basis; 153 154 CeedCall(CeedOperatorFieldGetData(input_fields[i], &field_name, &rstr, &basis, &vec)); 155 CeedCall(CeedOperatorSetField(op_fallback, field_name, rstr, basis, vec)); 156 CeedCall(CeedVectorDestroy(&vec)); 157 CeedCall(CeedElemRestrictionDestroy(&rstr)); 158 CeedCall(CeedBasisDestroy(&basis)); 159 } 160 for (CeedInt i = 0; i < num_output_fields; i++) { 161 const char *field_name; 162 CeedVector vec; 163 CeedElemRestriction rstr; 164 CeedBasis basis; 165 166 CeedCall(CeedOperatorFieldGetData(output_fields[i], &field_name, &rstr, &basis, &vec)); 167 CeedCall(CeedOperatorSetField(op_fallback, field_name, rstr, basis, vec)); 168 CeedCall(CeedVectorDestroy(&vec)); 169 CeedCall(CeedElemRestrictionDestroy(&rstr)); 170 CeedCall(CeedBasisDestroy(&basis)); 171 } 172 { 173 CeedQFunctionAssemblyData data; 174 175 CeedCall(CeedOperatorGetQFunctionAssemblyData(op, &data)); 176 CeedCall(CeedQFunctionAssemblyDataReferenceCopy(data, &op_fallback->qf_assembled)); 177 } 178 // Cleanup 179 CeedCall(CeedQFunctionDestroy(&qf_fallback)); 180 CeedCall(CeedQFunctionDestroy(&dqf_fallback)); 181 CeedCall(CeedQFunctionDestroy(&dqfT_fallback)); 182 } 183 CeedCall(CeedOperatorSetName(op_fallback, op->name)); 184 CeedCall(CeedOperatorCheckReady(op_fallback)); 185 // Note: No ref-counting here so we don't get caught in a reference loop. 186 // The op holds the only reference to op_fallback and is responsible for deleting itself and op_fallback. 187 op->op_fallback = op_fallback; 188 op_fallback->op_fallback_parent = op; 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 Ceed ceed; 207 bool is_composite; 208 209 CeedCall(CeedOperatorGetCeed(op, &ceed)); 210 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 211 CeedCheck(!is_composite, ceed, CEED_ERROR_UNSUPPORTED, "Composite operator not supported"); 212 213 // Assemble QFunction 214 CeedInt layout_qf[3]; 215 const CeedScalar *assembled_qf_array; 216 CeedVector assembled_qf = NULL; 217 CeedElemRestriction assembled_elem_rstr = NULL; 218 219 CeedCall(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &assembled_elem_rstr, request)); 220 CeedCall(CeedElemRestrictionGetELayout(assembled_elem_rstr, layout_qf)); 221 CeedCall(CeedElemRestrictionDestroy(&assembled_elem_rstr)); 222 CeedCall(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_HOST, &assembled_qf_array)); 223 224 // Get assembly data 225 const CeedEvalMode **eval_modes_in, **eval_modes_out; 226 CeedInt num_active_bases_in, *num_eval_modes_in, num_active_bases_out, *num_eval_modes_out; 227 CeedSize **eval_mode_offsets_in, **eval_mode_offsets_out, num_output_components; 228 CeedBasis *active_bases_in, *active_bases_out; 229 CeedElemRestriction *active_elem_rstrs_in, *active_elem_rstrs_out; 230 CeedOperatorAssemblyData data; 231 232 CeedCall(CeedOperatorGetOperatorAssemblyData(op, &data)); 233 CeedCall(CeedOperatorAssemblyDataGetEvalModes(data, &num_active_bases_in, &num_eval_modes_in, &eval_modes_in, &eval_mode_offsets_in, 234 &num_active_bases_out, &num_eval_modes_out, &eval_modes_out, &eval_mode_offsets_out, 235 &num_output_components)); 236 CeedCall(CeedOperatorAssemblyDataGetBases(data, NULL, &active_bases_in, NULL, NULL, &active_bases_out, NULL)); 237 CeedCall(CeedOperatorAssemblyDataGetElemRestrictions(data, NULL, &active_elem_rstrs_in, NULL, &active_elem_rstrs_out)); 238 239 // Loop over all active bases (find matching input/output pairs) 240 for (CeedInt b = 0; b < CeedIntMin(num_active_bases_in, num_active_bases_out); b++) { 241 CeedInt b_in, b_out, num_elem, num_nodes, num_qpts, num_comp; 242 bool has_eval_none = false; 243 CeedScalar *elem_diag_array, *identity = NULL; 244 CeedVector elem_diag; 245 CeedElemRestriction diag_elem_rstr; 246 247 if (num_active_bases_in <= num_active_bases_out) { 248 b_in = b; 249 for (b_out = 0; b_out < num_active_bases_out; b_out++) { 250 if (active_bases_in[b_in] == active_bases_out[b_out]) { 251 break; 252 } 253 } 254 if (b_out == num_active_bases_out) { 255 continue; 256 } // No matching output basis found 257 } else { 258 b_out = b; 259 for (b_in = 0; b_in < num_active_bases_in; b_in++) { 260 if (active_bases_in[b_in] == active_bases_out[b_out]) { 261 break; 262 } 263 } 264 if (b_in == num_active_bases_in) { 265 continue; 266 } // No matching output basis found 267 } 268 CeedCheck(active_elem_rstrs_in[b_in] == active_elem_rstrs_out[b_out], ceed, CEED_ERROR_UNSUPPORTED, 269 "Cannot assemble operator diagonal with different input and output active element restrictions"); 270 271 // Assemble point block diagonal restriction, if needed 272 if (is_point_block) { 273 CeedCall(CeedOperatorCreateActivePointBlockRestriction(active_elem_rstrs_in[b_in], &diag_elem_rstr)); 274 } else { 275 CeedCall(CeedElemRestrictionCreateUnsignedCopy(active_elem_rstrs_in[b_in], &diag_elem_rstr)); 276 } 277 278 // Create diagonal vector 279 CeedCall(CeedElemRestrictionCreateVector(diag_elem_rstr, NULL, &elem_diag)); 280 281 // Assemble element operator diagonals 282 CeedCall(CeedVectorSetValue(elem_diag, 0.0)); 283 CeedCall(CeedVectorGetArray(elem_diag, CEED_MEM_HOST, &elem_diag_array)); 284 CeedCall(CeedElemRestrictionGetNumElements(diag_elem_rstr, &num_elem)); 285 CeedCall(CeedBasisGetNumNodes(active_bases_in[b_in], &num_nodes)); 286 CeedCall(CeedBasisGetNumComponents(active_bases_in[b_in], &num_comp)); 287 if (active_bases_in[b_in] == CEED_BASIS_NONE) num_qpts = num_nodes; 288 else CeedCall(CeedBasisGetNumQuadraturePoints(active_bases_in[b_in], &num_qpts)); 289 290 // Construct identity matrix for basis if required 291 for (CeedInt i = 0; i < num_eval_modes_in[b_in]; i++) { 292 has_eval_none = has_eval_none || (eval_modes_in[b_in][i] == CEED_EVAL_NONE); 293 } 294 for (CeedInt i = 0; i < num_eval_modes_out[b_out]; i++) { 295 has_eval_none = has_eval_none || (eval_modes_out[b_out][i] == CEED_EVAL_NONE); 296 } 297 if (has_eval_none) { 298 CeedCall(CeedCalloc(num_qpts * num_nodes, &identity)); 299 for (CeedInt i = 0; i < (num_nodes < num_qpts ? num_nodes : num_qpts); i++) identity[i * num_nodes + i] = 1.0; 300 } 301 302 // Compute the diagonal of B^T D B 303 // Each element 304 for (CeedSize e = 0; e < num_elem; e++) { 305 // Each basis eval mode pair 306 CeedInt d_out = 0, q_comp_out; 307 CeedEvalMode eval_mode_out_prev = CEED_EVAL_NONE; 308 309 for (CeedInt e_out = 0; e_out < num_eval_modes_out[b_out]; e_out++) { 310 CeedInt d_in = 0, q_comp_in; 311 const CeedScalar *B_t = NULL; 312 CeedEvalMode eval_mode_in_prev = CEED_EVAL_NONE; 313 314 CeedCall(CeedOperatorGetBasisPointer(active_bases_out[b_out], eval_modes_out[b_out][e_out], identity, &B_t)); 315 CeedCall(CeedBasisGetNumQuadratureComponents(active_bases_out[b_out], eval_modes_out[b_out][e_out], &q_comp_out)); 316 if (q_comp_out > 1) { 317 if (e_out == 0 || eval_modes_out[b_out][e_out] != eval_mode_out_prev) d_out = 0; 318 else B_t = &B_t[(++d_out) * num_qpts * num_nodes]; 319 } 320 eval_mode_out_prev = eval_modes_out[b_out][e_out]; 321 322 for (CeedInt e_in = 0; e_in < num_eval_modes_in[b_in]; e_in++) { 323 const CeedScalar *B = NULL; 324 325 CeedCall(CeedOperatorGetBasisPointer(active_bases_in[b_in], eval_modes_in[b_in][e_in], identity, &B)); 326 CeedCall(CeedBasisGetNumQuadratureComponents(active_bases_in[b_in], eval_modes_in[b_in][e_in], &q_comp_in)); 327 if (q_comp_in > 1) { 328 if (e_in == 0 || eval_modes_in[b_in][e_in] != eval_mode_in_prev) d_in = 0; 329 else B = &B[(++d_in) * num_qpts * num_nodes]; 330 } 331 eval_mode_in_prev = eval_modes_in[b_in][e_in]; 332 333 // Each component 334 for (CeedInt c_out = 0; c_out < num_comp; c_out++) { 335 // Each qpt/node pair 336 for (CeedInt q = 0; q < num_qpts; q++) { 337 if (is_point_block) { 338 // Point Block Diagonal 339 for (CeedInt c_in = 0; c_in < num_comp; c_in++) { 340 const CeedSize c_offset = 341 (eval_mode_offsets_in[b_in][e_in] + c_in) * num_output_components + eval_mode_offsets_out[b_out][e_out] + c_out; 342 const CeedScalar qf_value = assembled_qf_array[q * layout_qf[0] + c_offset * layout_qf[1] + e * layout_qf[2]]; 343 344 for (CeedInt n = 0; n < num_nodes; n++) { 345 elem_diag_array[((e * num_comp + c_out) * num_comp + c_in) * num_nodes + n] += 346 B_t[q * num_nodes + n] * qf_value * B[q * num_nodes + n]; 347 } 348 } 349 } else { 350 // Diagonal Only 351 const CeedInt c_offset = 352 (eval_mode_offsets_in[b_in][e_in] + c_out) * num_output_components + eval_mode_offsets_out[b_out][e_out] + c_out; 353 const CeedScalar qf_value = assembled_qf_array[q * layout_qf[0] + c_offset * layout_qf[1] + e * layout_qf[2]]; 354 355 for (CeedInt n = 0; n < num_nodes; n++) { 356 elem_diag_array[(e * num_comp + c_out) * num_nodes + n] += B_t[q * num_nodes + n] * qf_value * B[q * num_nodes + n]; 357 } 358 } 359 } 360 } 361 } 362 } 363 } 364 CeedCall(CeedVectorRestoreArray(elem_diag, &elem_diag_array)); 365 366 // Assemble local operator diagonal 367 CeedCall(CeedElemRestrictionApply(diag_elem_rstr, CEED_TRANSPOSE, elem_diag, assembled, request)); 368 369 // Cleanup 370 CeedCall(CeedElemRestrictionDestroy(&diag_elem_rstr)); 371 CeedCall(CeedVectorDestroy(&elem_diag)); 372 CeedCall(CeedFree(&identity)); 373 } 374 CeedCall(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 375 CeedCall(CeedVectorDestroy(&assembled_qf)); 376 return CEED_ERROR_SUCCESS; 377 } 378 379 /** 380 @brief Core logic for assembling operator diagonal or point block diagonal 381 382 @param[in] op `CeedOperator` to assemble diagonal or point block diagonal 383 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 384 @param[in] is_point_block Boolean flag to assemble diagonal or point block diagonal 385 @param[out] assembled `CeedVector` to store assembled diagonal 386 387 @return An error code: 0 - success, otherwise - failure 388 389 @ref Developer 390 **/ 391 static inline int CeedSingleOperatorLinearAssembleAddDiagonal(CeedOperator op, CeedRequest *request, const bool is_point_block, 392 CeedVector assembled) { 393 Ceed ceed; 394 bool is_at_points; 395 396 CeedCall(CeedOperatorGetCeed(op, &ceed)); 397 CeedCall(CeedOperatorIsAtPoints(op, &is_at_points)); 398 CeedCheck(!is_at_points, ceed, CEED_ERROR_UNSUPPORTED, "AtPoints operator not supported"); 399 CeedCall(CeedSingleOperatorLinearAssembleAddDiagonal_Mesh(op, request, is_point_block, assembled)); 400 return CEED_ERROR_SUCCESS; 401 } 402 403 /** 404 @brief Core logic for assembling composite operator diagonal 405 406 @param[in] op `CeedOperator` to assemble point block diagonal 407 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 408 @param[in] is_point_block Boolean flag to assemble diagonal or point block diagonal 409 @param[out] assembled `CeedVector` to store assembled diagonal 410 411 @return An error code: 0 - success, otherwise - failure 412 413 @ref Developer 414 **/ 415 static inline int CeedCompositeOperatorLinearAssembleAddDiagonal(CeedOperator op, CeedRequest *request, const bool is_point_block, 416 CeedVector assembled) { 417 CeedInt num_sub; 418 CeedOperator *suboperators; 419 420 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_sub)); 421 CeedCall(CeedCompositeOperatorGetSubList(op, &suboperators)); 422 for (CeedInt i = 0; i < num_sub; i++) { 423 if (is_point_block) { 424 CeedCall(CeedOperatorLinearAssembleAddPointBlockDiagonal(suboperators[i], assembled, request)); 425 } else { 426 CeedCall(CeedOperatorLinearAssembleAddDiagonal(suboperators[i], assembled, request)); 427 } 428 } 429 return CEED_ERROR_SUCCESS; 430 } 431 432 /** 433 @brief Build nonzero pattern for non-composite CeedOperator`. 434 435 Users should generally use @ref CeedOperatorLinearAssembleSymbolic(). 436 437 @param[in] op `CeedOperator` to assemble nonzero pattern 438 @param[in] offset Offset for number of entries 439 @param[out] rows Row number for each entry 440 @param[out] cols Column number for each entry 441 442 @return An error code: 0 - success, otherwise - failure 443 444 @ref Developer 445 **/ 446 static int CeedSingleOperatorAssembleSymbolic(CeedOperator op, CeedInt offset, CeedInt *rows, CeedInt *cols) { 447 Ceed ceed; 448 bool is_composite; 449 CeedSize num_nodes_in, num_nodes_out, local_num_entries, count = 0; 450 CeedInt num_elem_in, elem_size_in, num_comp_in, layout_er_in[3]; 451 CeedInt num_elem_out, elem_size_out, num_comp_out, layout_er_out[3]; 452 CeedScalar *array; 453 const CeedScalar *elem_dof_a_in, *elem_dof_a_out; 454 CeedVector index_vec_in, index_vec_out, elem_dof_in, elem_dof_out; 455 CeedElemRestriction elem_rstr_in, elem_rstr_out, index_elem_rstr_in, index_elem_rstr_out; 456 457 CeedCall(CeedOperatorGetCeed(op, &ceed)); 458 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 459 CeedCheck(!is_composite, ceed, CEED_ERROR_UNSUPPORTED, "Composite operator not supported"); 460 461 CeedCall(CeedOperatorGetActiveVectorLengths(op, &num_nodes_in, &num_nodes_out)); 462 CeedCall(CeedOperatorGetActiveElemRestrictions(op, &elem_rstr_in, &elem_rstr_out)); 463 CeedCall(CeedElemRestrictionGetNumElements(elem_rstr_in, &num_elem_in)); 464 CeedCall(CeedElemRestrictionGetElementSize(elem_rstr_in, &elem_size_in)); 465 CeedCall(CeedElemRestrictionGetNumComponents(elem_rstr_in, &num_comp_in)); 466 CeedCall(CeedElemRestrictionGetELayout(elem_rstr_in, layout_er_in)); 467 468 // Determine elem_dof relation for input 469 CeedCall(CeedVectorCreate(ceed, num_nodes_in, &index_vec_in)); 470 CeedCall(CeedVectorGetArrayWrite(index_vec_in, CEED_MEM_HOST, &array)); 471 for (CeedSize i = 0; i < num_nodes_in; i++) array[i] = i; 472 CeedCall(CeedVectorRestoreArray(index_vec_in, &array)); 473 CeedCall(CeedVectorCreate(ceed, num_elem_in * elem_size_in * num_comp_in, &elem_dof_in)); 474 CeedCall(CeedVectorSetValue(elem_dof_in, 0.0)); 475 CeedCall(CeedElemRestrictionCreateUnorientedCopy(elem_rstr_in, &index_elem_rstr_in)); 476 CeedCall(CeedElemRestrictionApply(index_elem_rstr_in, CEED_NOTRANSPOSE, index_vec_in, elem_dof_in, CEED_REQUEST_IMMEDIATE)); 477 CeedCall(CeedVectorGetArrayRead(elem_dof_in, CEED_MEM_HOST, &elem_dof_a_in)); 478 CeedCall(CeedVectorDestroy(&index_vec_in)); 479 CeedCall(CeedElemRestrictionDestroy(&index_elem_rstr_in)); 480 481 if (elem_rstr_in != elem_rstr_out) { 482 CeedCall(CeedElemRestrictionGetNumElements(elem_rstr_out, &num_elem_out)); 483 CeedCheck(num_elem_in == num_elem_out, ceed, CEED_ERROR_UNSUPPORTED, 484 "Active input and output operator restrictions must have the same number of elements." 485 " Input has %" CeedInt_FMT " elements; output has %" CeedInt_FMT "elements.", 486 num_elem_in, num_elem_out); 487 CeedCall(CeedElemRestrictionGetElementSize(elem_rstr_out, &elem_size_out)); 488 CeedCall(CeedElemRestrictionGetNumComponents(elem_rstr_out, &num_comp_out)); 489 CeedCall(CeedElemRestrictionGetELayout(elem_rstr_out, layout_er_out)); 490 491 // Determine elem_dof relation for output 492 CeedCall(CeedVectorCreate(ceed, num_nodes_out, &index_vec_out)); 493 CeedCall(CeedVectorGetArrayWrite(index_vec_out, CEED_MEM_HOST, &array)); 494 for (CeedSize i = 0; i < num_nodes_out; i++) array[i] = i; 495 CeedCall(CeedVectorRestoreArray(index_vec_out, &array)); 496 CeedCall(CeedVectorCreate(ceed, num_elem_out * elem_size_out * num_comp_out, &elem_dof_out)); 497 CeedCall(CeedVectorSetValue(elem_dof_out, 0.0)); 498 CeedCall(CeedElemRestrictionCreateUnorientedCopy(elem_rstr_out, &index_elem_rstr_out)); 499 CeedCall(CeedElemRestrictionApply(index_elem_rstr_out, CEED_NOTRANSPOSE, index_vec_out, elem_dof_out, CEED_REQUEST_IMMEDIATE)); 500 CeedCall(CeedVectorGetArrayRead(elem_dof_out, CEED_MEM_HOST, &elem_dof_a_out)); 501 CeedCall(CeedVectorDestroy(&index_vec_out)); 502 CeedCall(CeedElemRestrictionDestroy(&index_elem_rstr_out)); 503 } else { 504 num_elem_out = num_elem_in; 505 elem_size_out = elem_size_in; 506 num_comp_out = num_comp_in; 507 layout_er_out[0] = layout_er_in[0]; 508 layout_er_out[1] = layout_er_in[1]; 509 layout_er_out[2] = layout_er_in[2]; 510 elem_dof_a_out = elem_dof_a_in; 511 } 512 local_num_entries = (CeedSize)elem_size_out * num_comp_out * elem_size_in * num_comp_in * num_elem_in; 513 514 // Determine i, j locations for element matrices 515 for (CeedInt e = 0; e < num_elem_in; e++) { 516 for (CeedInt comp_in = 0; comp_in < num_comp_in; comp_in++) { 517 for (CeedInt comp_out = 0; comp_out < num_comp_out; comp_out++) { 518 for (CeedInt i = 0; i < elem_size_out; i++) { 519 for (CeedInt j = 0; j < elem_size_in; j++) { 520 const CeedInt elem_dof_index_row = i * layout_er_out[0] + comp_out * layout_er_out[1] + e * layout_er_out[2]; 521 const CeedInt elem_dof_index_col = j * layout_er_in[0] + comp_in * layout_er_in[1] + e * layout_er_in[2]; 522 const CeedInt row = elem_dof_a_out[elem_dof_index_row]; 523 const CeedInt col = elem_dof_a_in[elem_dof_index_col]; 524 525 rows[offset + count] = row; 526 cols[offset + count] = col; 527 count++; 528 } 529 } 530 } 531 } 532 } 533 CeedCheck(count == local_num_entries, ceed, CEED_ERROR_MAJOR, "Error computing assembled entries"); 534 CeedCall(CeedVectorRestoreArrayRead(elem_dof_in, &elem_dof_a_in)); 535 CeedCall(CeedVectorDestroy(&elem_dof_in)); 536 if (elem_rstr_in != elem_rstr_out) { 537 CeedCall(CeedVectorRestoreArrayRead(elem_dof_out, &elem_dof_a_out)); 538 CeedCall(CeedVectorDestroy(&elem_dof_out)); 539 } 540 CeedCall(CeedElemRestrictionDestroy(&elem_rstr_in)); 541 CeedCall(CeedElemRestrictionDestroy(&elem_rstr_out)); 542 return CEED_ERROR_SUCCESS; 543 } 544 545 /** 546 @brief Assemble nonzero entries for non-composite `CeedOperator`. 547 548 Users should generally use @ref CeedOperatorLinearAssemble(). 549 550 @param[in] op `CeedOperator` to assemble 551 @param[in] offset Offset for number of entries 552 @param[out] values Values to assemble into matrix 553 554 @return An error code: 0 - success, otherwise - failure 555 556 @ref Developer 557 **/ 558 static int CeedSingleOperatorAssemble(CeedOperator op, CeedInt offset, CeedVector values) { 559 Ceed ceed; 560 bool is_composite; 561 562 CeedCall(CeedOperatorGetCeed(op, &ceed)); 563 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 564 CeedCheck(!is_composite, ceed, CEED_ERROR_UNSUPPORTED, "Composite operator not supported"); 565 566 // Early exit for empty operator 567 { 568 CeedInt num_elem = 0; 569 570 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 571 if (num_elem == 0) return CEED_ERROR_SUCCESS; 572 } 573 574 if (op->LinearAssembleSingle) { 575 // Backend version 576 CeedCall(op->LinearAssembleSingle(op, offset, values)); 577 return CEED_ERROR_SUCCESS; 578 } else { 579 // Operator fallback 580 CeedOperator op_fallback; 581 582 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 583 if (op_fallback) { 584 CeedCall(CeedSingleOperatorAssemble(op_fallback, offset, values)); 585 return CEED_ERROR_SUCCESS; 586 } 587 } 588 589 // Assemble QFunction 590 CeedInt layout_qf[3]; 591 const CeedScalar *assembled_qf_array; 592 CeedVector assembled_qf = NULL; 593 CeedElemRestriction assembled_elem_rstr = NULL; 594 595 CeedCall(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &assembled_elem_rstr, CEED_REQUEST_IMMEDIATE)); 596 CeedCall(CeedElemRestrictionGetELayout(assembled_elem_rstr, layout_qf)); 597 CeedCall(CeedElemRestrictionDestroy(&assembled_elem_rstr)); 598 CeedCall(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_HOST, &assembled_qf_array)); 599 600 // Get assembly data 601 CeedInt num_elem_in, elem_size_in, num_comp_in, num_qpts_in; 602 CeedInt num_elem_out, elem_size_out, num_comp_out, num_qpts_out; 603 CeedSize local_num_entries, count = 0; 604 const CeedEvalMode **eval_modes_in, **eval_modes_out; 605 CeedInt num_active_bases_in, *num_eval_modes_in, num_active_bases_out, *num_eval_modes_out; 606 CeedBasis *active_bases_in, *active_bases_out, basis_in, basis_out; 607 const CeedScalar **B_mats_in, **B_mats_out, *B_mat_in, *B_mat_out; 608 CeedElemRestriction elem_rstr_in, elem_rstr_out; 609 CeedRestrictionType elem_rstr_type_in, elem_rstr_type_out; 610 const bool *elem_rstr_orients_in = NULL, *elem_rstr_orients_out = NULL; 611 const CeedInt8 *elem_rstr_curl_orients_in = NULL, *elem_rstr_curl_orients_out = NULL; 612 CeedOperatorAssemblyData data; 613 614 CeedCall(CeedOperatorGetOperatorAssemblyData(op, &data)); 615 CeedCall(CeedOperatorAssemblyDataGetEvalModes(data, &num_active_bases_in, &num_eval_modes_in, &eval_modes_in, NULL, &num_active_bases_out, 616 &num_eval_modes_out, &eval_modes_out, NULL, NULL)); 617 618 CeedCheck(num_active_bases_in == 1 && num_active_bases_out == 1, ceed, CEED_ERROR_UNSUPPORTED, 619 "Cannot assemble operator with multiple active bases"); 620 CeedCheck(num_eval_modes_in[0] > 0 && num_eval_modes_out[0] > 0, ceed, CEED_ERROR_UNSUPPORTED, "Cannot assemble operator without inputs/outputs"); 621 622 CeedCall(CeedOperatorAssemblyDataGetBases(data, NULL, &active_bases_in, &B_mats_in, NULL, &active_bases_out, &B_mats_out)); 623 CeedCall(CeedOperatorGetActiveElemRestrictions(op, &elem_rstr_in, &elem_rstr_out)); 624 basis_in = active_bases_in[0]; 625 basis_out = active_bases_out[0]; 626 B_mat_in = B_mats_in[0]; 627 B_mat_out = B_mats_out[0]; 628 629 CeedCall(CeedElemRestrictionGetNumElements(elem_rstr_in, &num_elem_in)); 630 CeedCall(CeedElemRestrictionGetElementSize(elem_rstr_in, &elem_size_in)); 631 CeedCall(CeedElemRestrictionGetNumComponents(elem_rstr_in, &num_comp_in)); 632 if (basis_in == CEED_BASIS_NONE) num_qpts_in = elem_size_in; 633 else CeedCall(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts_in)); 634 635 CeedCall(CeedElemRestrictionGetType(elem_rstr_in, &elem_rstr_type_in)); 636 if (elem_rstr_type_in == CEED_RESTRICTION_ORIENTED) { 637 CeedCall(CeedElemRestrictionGetOrientations(elem_rstr_in, CEED_MEM_HOST, &elem_rstr_orients_in)); 638 } else if (elem_rstr_type_in == CEED_RESTRICTION_CURL_ORIENTED) { 639 CeedCall(CeedElemRestrictionGetCurlOrientations(elem_rstr_in, CEED_MEM_HOST, &elem_rstr_curl_orients_in)); 640 } 641 642 if (elem_rstr_in != elem_rstr_out) { 643 CeedCall(CeedElemRestrictionGetNumElements(elem_rstr_out, &num_elem_out)); 644 CeedCheck(num_elem_in == num_elem_out, ceed, CEED_ERROR_UNSUPPORTED, 645 "Active input and output operator restrictions must have the same number of elements." 646 " Input has %" CeedInt_FMT " elements; output has %" CeedInt_FMT "elements.", 647 num_elem_in, num_elem_out); 648 CeedCall(CeedElemRestrictionGetElementSize(elem_rstr_out, &elem_size_out)); 649 CeedCall(CeedElemRestrictionGetNumComponents(elem_rstr_out, &num_comp_out)); 650 if (basis_out == CEED_BASIS_NONE) num_qpts_out = elem_size_out; 651 else CeedCall(CeedBasisGetNumQuadraturePoints(basis_out, &num_qpts_out)); 652 CeedCheck(num_qpts_in == num_qpts_out, ceed, CEED_ERROR_UNSUPPORTED, 653 "Active input and output bases must have the same number of quadrature points." 654 " Input has %" CeedInt_FMT " points; output has %" CeedInt_FMT "points.", 655 num_qpts_in, num_qpts_out); 656 657 CeedCall(CeedElemRestrictionGetType(elem_rstr_out, &elem_rstr_type_out)); 658 if (elem_rstr_type_out == CEED_RESTRICTION_ORIENTED) { 659 CeedCall(CeedElemRestrictionGetOrientations(elem_rstr_out, CEED_MEM_HOST, &elem_rstr_orients_out)); 660 } else if (elem_rstr_type_out == CEED_RESTRICTION_CURL_ORIENTED) { 661 CeedCall(CeedElemRestrictionGetCurlOrientations(elem_rstr_out, CEED_MEM_HOST, &elem_rstr_curl_orients_out)); 662 } 663 } else { 664 num_elem_out = num_elem_in; 665 elem_size_out = elem_size_in; 666 num_comp_out = num_comp_in; 667 num_qpts_out = num_qpts_in; 668 669 elem_rstr_orients_out = elem_rstr_orients_in; 670 elem_rstr_curl_orients_out = elem_rstr_curl_orients_in; 671 } 672 local_num_entries = (CeedSize)elem_size_out * num_comp_out * elem_size_in * num_comp_in * num_elem_in; 673 674 // Loop over elements and put in data structure 675 // We store B_mat_in, B_mat_out, BTD, elem_mat in row-major order 676 CeedTensorContract contract; 677 CeedScalar *vals, *BTD_mat = NULL, *elem_mat = NULL, *elem_mat_b = NULL; 678 679 CeedCall(CeedBasisGetTensorContract(basis_in, &contract)); 680 CeedCall(CeedCalloc(elem_size_out * num_qpts_in * num_eval_modes_in[0], &BTD_mat)); 681 CeedCall(CeedCalloc(elem_size_out * elem_size_in, &elem_mat)); 682 if (elem_rstr_curl_orients_in || elem_rstr_curl_orients_out) CeedCall(CeedCalloc(elem_size_out * elem_size_in, &elem_mat_b)); 683 684 CeedCall(CeedVectorGetArray(values, CEED_MEM_HOST, &vals)); 685 for (CeedSize e = 0; e < num_elem_in; e++) { 686 for (CeedInt comp_in = 0; comp_in < num_comp_in; comp_in++) { 687 for (CeedInt comp_out = 0; comp_out < num_comp_out; comp_out++) { 688 // Compute B^T*D 689 for (CeedSize n = 0; n < elem_size_out; n++) { 690 for (CeedSize q = 0; q < num_qpts_in; q++) { 691 for (CeedInt e_in = 0; e_in < num_eval_modes_in[0]; e_in++) { 692 const CeedSize btd_index = n * (num_qpts_in * num_eval_modes_in[0]) + q * num_eval_modes_in[0] + e_in; 693 CeedScalar sum = 0.0; 694 695 for (CeedInt e_out = 0; e_out < num_eval_modes_out[0]; e_out++) { 696 const CeedSize b_out_index = (q * num_eval_modes_out[0] + e_out) * elem_size_out + n; 697 const CeedSize eval_mode_index = ((e_in * num_comp_in + comp_in) * num_eval_modes_out[0] + e_out) * num_comp_out + comp_out; 698 const CeedSize qf_index = q * layout_qf[0] + eval_mode_index * layout_qf[1] + e * layout_qf[2]; 699 700 sum += B_mat_out[b_out_index] * assembled_qf_array[qf_index]; 701 } 702 BTD_mat[btd_index] = sum; 703 } 704 } 705 } 706 707 // Form element matrix itself (for each block component) 708 if (contract) { 709 CeedCall(CeedTensorContractApply(contract, 1, num_qpts_in * num_eval_modes_in[0], elem_size_in, elem_size_out, BTD_mat, CEED_NOTRANSPOSE, 710 false, B_mat_in, elem_mat)); 711 } else { 712 CeedCall(CeedMatrixMatrixMultiply(ceed, BTD_mat, B_mat_in, elem_mat, elem_size_out, elem_size_in, num_qpts_in * num_eval_modes_in[0])); 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, ceed, 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 Ceed ceed; 811 CeedElemRestriction rstr_in, rstr_out; 812 813 CeedCall(CeedOperatorGetCeed(op, &ceed)); 814 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 815 CeedCheck(!is_composite, ceed, CEED_ERROR_UNSUPPORTED, "Composite operator not supported"); 816 817 CeedCall(CeedOperatorGetActiveElemRestrictions(op, &rstr_in, &rstr_out)); 818 CeedCall(CeedElemRestrictionGetNumElements(rstr_in, &num_elem_in)); 819 CeedCall(CeedElemRestrictionGetElementSize(rstr_in, &elem_size_in)); 820 CeedCall(CeedElemRestrictionGetNumComponents(rstr_in, &num_comp_in)); 821 if (rstr_in != rstr_out) { 822 CeedCall(CeedElemRestrictionGetNumElements(rstr_out, &num_elem_out)); 823 CeedCheck(num_elem_in == num_elem_out, ceed, CEED_ERROR_UNSUPPORTED, 824 "Active input and output operator restrictions must have the same number of elements." 825 " Input has %" CeedInt_FMT " elements; output has %" CeedInt_FMT "elements.", 826 num_elem_in, num_elem_out); 827 CeedCall(CeedElemRestrictionGetElementSize(rstr_out, &elem_size_out)); 828 CeedCall(CeedElemRestrictionGetNumComponents(rstr_out, &num_comp_out)); 829 } else { 830 num_elem_out = num_elem_in; 831 elem_size_out = elem_size_in; 832 num_comp_out = num_comp_in; 833 } 834 CeedCall(CeedElemRestrictionDestroy(&rstr_in)); 835 CeedCall(CeedElemRestrictionDestroy(&rstr_out)); 836 *num_entries = (CeedSize)elem_size_in * num_comp_in * elem_size_out * num_comp_out * num_elem_in; 837 return CEED_ERROR_SUCCESS; 838 } 839 840 /** 841 @brief Common code for creating a multigrid coarse `CeedOperator` and level transfer `CeedOperator` for a `CeedOperator` 842 843 @param[in] op_fine Fine grid `CeedOperator` 844 @param[in] p_mult_fine L-vector multiplicity in parallel gather/scatter, or `NULL` if not creating prolongation/restriction `CeedOperator` 845 @param[in] rstr_coarse Coarse grid `CeedElemRestriction` 846 @param[in] basis_coarse Coarse grid active vector `CeedBasis` 847 @param[in] basis_c_to_f `CeedBasis` for coarse to fine interpolation, or `NULL` if not creating prolongation/restriction operators 848 @param[out] op_coarse Coarse grid `CeedOperator` 849 @param[out] op_prolong Coarse to fine `CeedOperator`, or `NULL` 850 @param[out] op_restrict Fine to coarse `CeedOperator`, or `NULL` 851 852 @return An error code: 0 - success, otherwise - failure 853 854 @ref Developer 855 **/ 856 static int CeedSingleOperatorMultigridLevel(CeedOperator op_fine, CeedVector p_mult_fine, CeedElemRestriction rstr_coarse, CeedBasis basis_coarse, 857 CeedBasis basis_c_to_f, CeedOperator *op_coarse, CeedOperator *op_prolong, CeedOperator *op_restrict) { 858 bool is_composite; 859 Ceed ceed; 860 CeedInt num_comp, num_input_fields, num_output_fields; 861 CeedVector mult_vec = NULL; 862 CeedElemRestriction rstr_p_mult_fine = NULL, rstr_fine = NULL; 863 CeedOperatorField *input_fields, *output_fields; 864 865 CeedCall(CeedOperatorGetCeed(op_fine, &ceed)); 866 867 // Check for composite operator 868 CeedCall(CeedOperatorIsComposite(op_fine, &is_composite)); 869 CeedCheck(!is_composite, ceed, CEED_ERROR_UNSUPPORTED, "Automatic multigrid setup for composite operators not supported"); 870 871 // Coarse Grid 872 CeedCall(CeedOperatorCreate(ceed, op_fine->qf, op_fine->dqf, op_fine->dqfT, op_coarse)); 873 CeedCall(CeedOperatorGetFields(op_fine, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 874 // -- Clone input fields 875 for (CeedInt i = 0; i < num_input_fields; i++) { 876 const char *field_name; 877 CeedVector vec; 878 CeedElemRestriction rstr = NULL; 879 CeedBasis basis = NULL; 880 881 CeedCall(CeedOperatorFieldGetName(input_fields[i], &field_name)); 882 CeedCall(CeedOperatorFieldGetVector(input_fields[i], &vec)); 883 if (vec == CEED_VECTOR_ACTIVE) { 884 CeedCall(CeedElemRestrictionReferenceCopy(rstr_coarse, &rstr)); 885 CeedCall(CeedBasisReferenceCopy(basis_coarse, &basis)); 886 if (!rstr_fine) CeedCall(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr_fine)); 887 } else { 888 CeedCall(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr)); 889 CeedCall(CeedOperatorFieldGetBasis(input_fields[i], &basis)); 890 } 891 CeedCall(CeedOperatorSetField(*op_coarse, field_name, rstr, basis, vec)); 892 CeedCall(CeedVectorDestroy(&vec)); 893 CeedCall(CeedElemRestrictionDestroy(&rstr)); 894 CeedCall(CeedBasisDestroy(&basis)); 895 } 896 // -- Clone output fields 897 for (CeedInt i = 0; i < num_output_fields; i++) { 898 const char *field_name; 899 CeedVector vec; 900 CeedElemRestriction rstr = NULL; 901 CeedBasis basis = NULL; 902 903 CeedCall(CeedOperatorFieldGetName(output_fields[i], &field_name)); 904 CeedCall(CeedOperatorFieldGetVector(output_fields[i], &vec)); 905 if (vec == CEED_VECTOR_ACTIVE) { 906 CeedCall(CeedElemRestrictionReferenceCopy(rstr_coarse, &rstr)); 907 CeedCall(CeedBasisReferenceCopy(basis_coarse, &basis)); 908 if (!rstr_fine) CeedCall(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr_fine)); 909 } else { 910 CeedCall(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr)); 911 CeedCall(CeedOperatorFieldGetBasis(output_fields[i], &basis)); 912 } 913 CeedCall(CeedOperatorSetField(*op_coarse, field_name, rstr, basis, vec)); 914 CeedCall(CeedVectorDestroy(&vec)); 915 CeedCall(CeedElemRestrictionDestroy(&rstr)); 916 CeedCall(CeedBasisDestroy(&basis)); 917 } 918 // -- Clone QFunctionAssemblyData 919 { 920 CeedQFunctionAssemblyData fine_data; 921 922 CeedCall(CeedOperatorGetQFunctionAssemblyData(op_fine, &fine_data)); 923 CeedCall(CeedQFunctionAssemblyDataReferenceCopy(fine_data, &(*op_coarse)->qf_assembled)); 924 } 925 926 // Multiplicity vector 927 if (op_restrict || op_prolong) { 928 CeedVector mult_e_vec; 929 CeedRestrictionType rstr_type; 930 931 CeedCall(CeedElemRestrictionGetType(rstr_fine, &rstr_type)); 932 CeedCheck(rstr_type != CEED_RESTRICTION_CURL_ORIENTED, ceed, CEED_ERROR_UNSUPPORTED, 933 "Element restrictions created with CeedElemRestrictionCreateCurlOriented are not supported"); 934 CeedCheck(p_mult_fine, ceed, CEED_ERROR_INCOMPATIBLE, "Prolongation or restriction operator creation requires fine grid multiplicity vector"); 935 CeedCall(CeedElemRestrictionCreateUnsignedCopy(rstr_fine, &rstr_p_mult_fine)); 936 CeedCall(CeedElemRestrictionCreateVector(rstr_fine, &mult_vec, &mult_e_vec)); 937 CeedCall(CeedVectorSetValue(mult_e_vec, 0.0)); 938 CeedCall(CeedElemRestrictionApply(rstr_p_mult_fine, CEED_NOTRANSPOSE, p_mult_fine, mult_e_vec, CEED_REQUEST_IMMEDIATE)); 939 CeedCall(CeedVectorSetValue(mult_vec, 0.0)); 940 CeedCall(CeedElemRestrictionApply(rstr_p_mult_fine, CEED_TRANSPOSE, mult_e_vec, mult_vec, CEED_REQUEST_IMMEDIATE)); 941 CeedCall(CeedVectorDestroy(&mult_e_vec)); 942 CeedCall(CeedVectorReciprocal(mult_vec)); 943 } 944 945 // Clone name 946 bool has_name = op_fine->name; 947 size_t name_len = op_fine->name ? strlen(op_fine->name) : 0; 948 CeedCall(CeedOperatorSetName(*op_coarse, op_fine->name)); 949 950 // Check that coarse to fine basis is provided if prolong/restrict operators are requested 951 CeedCheck(basis_c_to_f || (!op_restrict && !op_prolong), ceed, CEED_ERROR_INCOMPATIBLE, 952 "Prolongation or restriction operator creation requires coarse-to-fine basis"); 953 954 // Restriction/Prolongation Operators 955 CeedCall(CeedBasisGetNumComponents(basis_coarse, &num_comp)); 956 957 // Restriction 958 if (op_restrict) { 959 CeedInt *num_comp_r_data; 960 CeedQFunctionContext ctx_r; 961 CeedQFunction qf_restrict; 962 963 CeedCall(CeedQFunctionCreateInteriorByName(ceed, "Scale", &qf_restrict)); 964 CeedCall(CeedCalloc(1, &num_comp_r_data)); 965 num_comp_r_data[0] = num_comp; 966 CeedCall(CeedQFunctionContextCreate(ceed, &ctx_r)); 967 CeedCall(CeedQFunctionContextSetData(ctx_r, CEED_MEM_HOST, CEED_OWN_POINTER, sizeof(*num_comp_r_data), num_comp_r_data)); 968 CeedCall(CeedQFunctionSetContext(qf_restrict, ctx_r)); 969 CeedCall(CeedQFunctionContextDestroy(&ctx_r)); 970 CeedCall(CeedQFunctionAddInput(qf_restrict, "input", num_comp, CEED_EVAL_NONE)); 971 CeedCall(CeedQFunctionAddInput(qf_restrict, "scale", num_comp, CEED_EVAL_NONE)); 972 CeedCall(CeedQFunctionAddOutput(qf_restrict, "output", num_comp, CEED_EVAL_INTERP)); 973 CeedCall(CeedQFunctionSetUserFlopsEstimate(qf_restrict, num_comp)); 974 975 CeedCall(CeedOperatorCreate(ceed, qf_restrict, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, op_restrict)); 976 CeedCall(CeedOperatorSetField(*op_restrict, "input", rstr_fine, CEED_BASIS_NONE, CEED_VECTOR_ACTIVE)); 977 CeedCall(CeedOperatorSetField(*op_restrict, "scale", rstr_p_mult_fine, CEED_BASIS_NONE, mult_vec)); 978 CeedCall(CeedOperatorSetField(*op_restrict, "output", rstr_coarse, basis_c_to_f, CEED_VECTOR_ACTIVE)); 979 980 // Set name 981 char *restriction_name; 982 983 CeedCall(CeedCalloc(17 + name_len, &restriction_name)); 984 sprintf(restriction_name, "restriction%s%s", has_name ? " for " : "", has_name ? op_fine->name : ""); 985 CeedCall(CeedOperatorSetName(*op_restrict, restriction_name)); 986 CeedCall(CeedFree(&restriction_name)); 987 988 // Check 989 CeedCall(CeedOperatorCheckReady(*op_restrict)); 990 991 // Cleanup 992 CeedCall(CeedQFunctionDestroy(&qf_restrict)); 993 } 994 995 // Prolongation 996 if (op_prolong) { 997 CeedInt *num_comp_p_data; 998 CeedQFunctionContext ctx_p; 999 CeedQFunction qf_prolong; 1000 1001 CeedCall(CeedQFunctionCreateInteriorByName(ceed, "Scale", &qf_prolong)); 1002 CeedCall(CeedCalloc(1, &num_comp_p_data)); 1003 num_comp_p_data[0] = num_comp; 1004 CeedCall(CeedQFunctionContextCreate(ceed, &ctx_p)); 1005 CeedCall(CeedQFunctionContextSetData(ctx_p, CEED_MEM_HOST, CEED_OWN_POINTER, sizeof(*num_comp_p_data), num_comp_p_data)); 1006 CeedCall(CeedQFunctionSetContext(qf_prolong, ctx_p)); 1007 CeedCall(CeedQFunctionContextDestroy(&ctx_p)); 1008 CeedCall(CeedQFunctionAddInput(qf_prolong, "input", num_comp, CEED_EVAL_INTERP)); 1009 CeedCall(CeedQFunctionAddInput(qf_prolong, "scale", num_comp, CEED_EVAL_NONE)); 1010 CeedCall(CeedQFunctionAddOutput(qf_prolong, "output", num_comp, CEED_EVAL_NONE)); 1011 CeedCall(CeedQFunctionSetUserFlopsEstimate(qf_prolong, num_comp)); 1012 1013 CeedCall(CeedOperatorCreate(ceed, qf_prolong, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, op_prolong)); 1014 CeedCall(CeedOperatorSetField(*op_prolong, "input", rstr_coarse, basis_c_to_f, CEED_VECTOR_ACTIVE)); 1015 CeedCall(CeedOperatorSetField(*op_prolong, "scale", rstr_p_mult_fine, CEED_BASIS_NONE, mult_vec)); 1016 CeedCall(CeedOperatorSetField(*op_prolong, "output", rstr_fine, CEED_BASIS_NONE, CEED_VECTOR_ACTIVE)); 1017 1018 // Set name 1019 char *prolongation_name; 1020 1021 CeedCall(CeedCalloc(18 + name_len, &prolongation_name)); 1022 sprintf(prolongation_name, "prolongation%s%s", has_name ? " for " : "", has_name ? op_fine->name : ""); 1023 CeedCall(CeedOperatorSetName(*op_prolong, prolongation_name)); 1024 CeedCall(CeedFree(&prolongation_name)); 1025 1026 // Check 1027 CeedCall(CeedOperatorCheckReady(*op_prolong)); 1028 1029 // Cleanup 1030 CeedCall(CeedQFunctionDestroy(&qf_prolong)); 1031 } 1032 1033 // Check 1034 CeedCall(CeedOperatorCheckReady(*op_coarse)); 1035 1036 // Cleanup 1037 CeedCall(CeedVectorDestroy(&mult_vec)); 1038 CeedCall(CeedElemRestrictionDestroy(&rstr_fine)); 1039 CeedCall(CeedElemRestrictionDestroy(&rstr_p_mult_fine)); 1040 CeedCall(CeedBasisDestroy(&basis_c_to_f)); 1041 return CEED_ERROR_SUCCESS; 1042 } 1043 1044 /** 1045 @brief Build 1D mass matrix and Laplacian with perturbation 1046 1047 @param[in] interp_1d Interpolation matrix in one dimension 1048 @param[in] grad_1d Gradient matrix in one dimension 1049 @param[in] q_weight_1d Quadrature weights in one dimension 1050 @param[in] P_1d Number of basis nodes in one dimension 1051 @param[in] Q_1d Number of quadrature points in one dimension 1052 @param[in] dim Dimension of basis 1053 @param[out] mass Assembled mass matrix in one dimension 1054 @param[out] laplace Assembled perturbed Laplacian in one dimension 1055 1056 @return An error code: 0 - success, otherwise - failure 1057 1058 @ref Developer 1059 **/ 1060 CeedPragmaOptimizeOff 1061 static int CeedBuildMassLaplace(const CeedScalar *interp_1d, const CeedScalar *grad_1d, const CeedScalar *q_weight_1d, CeedInt P_1d, CeedInt Q_1d, 1062 CeedInt dim, CeedScalar *mass, CeedScalar *laplace) { 1063 for (CeedInt i = 0; i < P_1d; i++) { 1064 for (CeedInt j = 0; j < P_1d; j++) { 1065 CeedScalar sum = 0.0; 1066 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]; 1067 mass[i + j * P_1d] = sum; 1068 } 1069 } 1070 // -- Laplacian 1071 for (CeedInt i = 0; i < P_1d; i++) { 1072 for (CeedInt j = 0; j < P_1d; j++) { 1073 CeedScalar sum = 0.0; 1074 1075 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]; 1076 laplace[i + j * P_1d] = sum; 1077 } 1078 } 1079 CeedScalar perturbation = dim > 2 ? 1e-6 : 1e-4; 1080 for (CeedInt i = 0; i < P_1d; i++) laplace[i + P_1d * i] += perturbation; 1081 return CEED_ERROR_SUCCESS; 1082 } 1083 CeedPragmaOptimizeOn 1084 1085 /// @} 1086 1087 /// ---------------------------------------------------------------------------- 1088 /// CeedOperator Backend API 1089 /// ---------------------------------------------------------------------------- 1090 /// @addtogroup CeedOperatorBackend 1091 /// @{ 1092 1093 /** 1094 @brief Select correct basis matrix pointer based on @ref CeedEvalMode 1095 1096 @param[in] basis `CeedBasis` from which to get the basis matrix 1097 @param[in] eval_mode Current basis evaluation mode 1098 @param[in] identity Pointer to identity matrix 1099 @param[out] basis_ptr `CeedBasis` pointer to set 1100 1101 @ref Backend 1102 **/ 1103 int CeedOperatorGetBasisPointer(CeedBasis basis, CeedEvalMode eval_mode, const CeedScalar *identity, const CeedScalar **basis_ptr) { 1104 switch (eval_mode) { 1105 case CEED_EVAL_NONE: 1106 *basis_ptr = identity; 1107 break; 1108 case CEED_EVAL_INTERP: 1109 CeedCall(CeedBasisGetInterp(basis, basis_ptr)); 1110 break; 1111 case CEED_EVAL_GRAD: 1112 CeedCall(CeedBasisGetGrad(basis, basis_ptr)); 1113 break; 1114 case CEED_EVAL_DIV: 1115 CeedCall(CeedBasisGetDiv(basis, basis_ptr)); 1116 break; 1117 case CEED_EVAL_CURL: 1118 CeedCall(CeedBasisGetCurl(basis, basis_ptr)); 1119 break; 1120 case CEED_EVAL_WEIGHT: 1121 break; // Caught by QF Assembly 1122 } 1123 assert(*basis_ptr != NULL); 1124 return CEED_ERROR_SUCCESS; 1125 } 1126 1127 /** 1128 @brief Create point block restriction for active `CeedOperatorField` 1129 1130 @param[in] rstr Original `CeedElemRestriction` for active field 1131 @param[out] point_block_rstr Address of the variable where the newly created `CeedElemRestriction` will be stored 1132 1133 @return An error code: 0 - success, otherwise - failure 1134 1135 @ref Backend 1136 **/ 1137 int CeedOperatorCreateActivePointBlockRestriction(CeedElemRestriction rstr, CeedElemRestriction *point_block_rstr) { 1138 Ceed ceed; 1139 CeedInt num_elem, num_comp, shift, elem_size, comp_stride, *point_block_offsets; 1140 CeedSize l_size; 1141 const CeedInt *offsets; 1142 1143 CeedCall(CeedElemRestrictionGetCeed(rstr, &ceed)); 1144 CeedCall(CeedElemRestrictionGetOffsets(rstr, CEED_MEM_HOST, &offsets)); 1145 1146 // Expand offsets 1147 CeedCall(CeedElemRestrictionGetNumElements(rstr, &num_elem)); 1148 CeedCall(CeedElemRestrictionGetNumComponents(rstr, &num_comp)); 1149 CeedCall(CeedElemRestrictionGetElementSize(rstr, &elem_size)); 1150 CeedCall(CeedElemRestrictionGetCompStride(rstr, &comp_stride)); 1151 CeedCall(CeedElemRestrictionGetLVectorSize(rstr, &l_size)); 1152 shift = num_comp; 1153 if (comp_stride != 1) shift *= num_comp; 1154 CeedCall(CeedCalloc(num_elem * elem_size, &point_block_offsets)); 1155 for (CeedInt i = 0; i < num_elem * elem_size; i++) { 1156 point_block_offsets[i] = offsets[i] * shift; 1157 } 1158 1159 // Create new restriction 1160 CeedCall(CeedElemRestrictionCreate(ceed, num_elem, elem_size, num_comp * num_comp, 1, l_size * num_comp, CEED_MEM_HOST, CEED_OWN_POINTER, 1161 point_block_offsets, point_block_rstr)); 1162 1163 // Cleanup 1164 CeedCall(CeedElemRestrictionRestoreOffsets(rstr, &offsets)); 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 } 1815 } 1816 *op_fallback = op->op_fallback; 1817 return CEED_ERROR_SUCCESS; 1818 } 1819 1820 /** 1821 @brief Get the parent `CeedOperator` for a fallback `CeedOperator` 1822 1823 @param[in] op `CeedOperator` context 1824 @param[out] parent Variable to store parent `CeedOperator` context 1825 1826 @return An error code: 0 - success, otherwise - failure 1827 1828 @ref Backend 1829 **/ 1830 int CeedOperatorGetFallbackParent(CeedOperator op, CeedOperator *parent) { 1831 *parent = op->op_fallback_parent ? op->op_fallback_parent : NULL; 1832 return CEED_ERROR_SUCCESS; 1833 } 1834 1835 /** 1836 @brief Get the `Ceed` context of the parent `CeedOperator` for a fallback `CeedOperator` 1837 1838 @param[in] op `CeedOperator` context 1839 @param[out] parent Variable to store parent `Ceed` context 1840 1841 @return An error code: 0 - success, otherwise - failure 1842 1843 @ref Backend 1844 **/ 1845 int CeedOperatorGetFallbackParentCeed(CeedOperator op, Ceed *parent) { 1846 *parent = op->op_fallback_parent ? op->op_fallback_parent->ceed : op->ceed; 1847 return CEED_ERROR_SUCCESS; 1848 } 1849 1850 /// @} 1851 1852 /// ---------------------------------------------------------------------------- 1853 /// CeedOperator Public API 1854 /// ---------------------------------------------------------------------------- 1855 /// @addtogroup CeedOperatorUser 1856 /// @{ 1857 1858 /** 1859 @brief Assemble a linear `CeedQFunction` associated with a `CeedOperator`. 1860 1861 This returns a `CeedVector` containing a matrix at each quadrature point providing the action of the `CeedQFunction` associated with the `CeedOperator`. 1862 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. 1863 1864 Inputs and outputs are in the order provided by the user when adding `CeedOperator` fields. 1865 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]`. 1866 1867 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 1868 1869 @param[in] op `CeedOperator` to assemble `CeedQFunction` 1870 @param[out] assembled `CeedVector` to store assembled `CeedQFunction` at quadrature points 1871 @param[out] rstr `CeedElemRestriction` for `CeedVector` containing assembled `CeedQFunction` 1872 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 1873 1874 @return An error code: 0 - success, otherwise - failure 1875 1876 @ref User 1877 **/ 1878 int CeedOperatorLinearAssembleQFunction(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 1879 CeedCall(CeedOperatorCheckReady(op)); 1880 1881 if (op->LinearAssembleQFunction) { 1882 // Backend version 1883 CeedCall(op->LinearAssembleQFunction(op, assembled, rstr, request)); 1884 } else { 1885 // Operator fallback 1886 Ceed ceed; 1887 CeedOperator op_fallback; 1888 1889 CeedCall(CeedOperatorGetCeed(op, &ceed)); 1890 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 1891 if (op_fallback) CeedCall(CeedOperatorLinearAssembleQFunction(op_fallback, assembled, rstr, request)); 1892 else return CeedError(ceed, CEED_ERROR_UNSUPPORTED, "Backend does not support CeedOperatorLinearAssembleQFunction"); 1893 } 1894 return CEED_ERROR_SUCCESS; 1895 } 1896 1897 /** 1898 @brief Assemble `CeedQFunction` and store result internally. 1899 1900 Return copied references of stored data to the caller. 1901 Caller is responsible for ownership and destruction of the copied references. 1902 See also @ref CeedOperatorLinearAssembleQFunction(). 1903 1904 Note: If the value of `assembled` or `rstr` passed to this function are non-`NULL` , then it is assumed that they hold valid pointers. 1905 These objects will be destroyed if `*assembled` or `*rstr` is the only reference to the object. 1906 1907 @param[in] op `CeedOperator` to assemble `CeedQFunction` 1908 @param[out] assembled `CeedVector` to store assembled `CeedQFunction` at quadrature points 1909 @param[out] rstr `CeedElemRestriction` for `CeedVector` containing assembled `CeedQFunction` 1910 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 1911 1912 @return An error code: 0 - success, otherwise - failure 1913 1914 @ref User 1915 **/ 1916 int CeedOperatorLinearAssembleQFunctionBuildOrUpdate(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 1917 int (*LinearAssembleQFunctionUpdate)(CeedOperator, CeedVector, CeedElemRestriction, CeedRequest *) = NULL; 1918 CeedOperator op_assemble = NULL; 1919 CeedOperator op_fallback_parent = NULL; 1920 1921 CeedCall(CeedOperatorCheckReady(op)); 1922 1923 // Determine if fallback parent or operator has implementation 1924 CeedCall(CeedOperatorGetFallbackParent(op, &op_fallback_parent)); 1925 if (op_fallback_parent && op_fallback_parent->LinearAssembleQFunctionUpdate) { 1926 // -- Backend version for op fallback parent is faster, if it exists 1927 LinearAssembleQFunctionUpdate = op_fallback_parent->LinearAssembleQFunctionUpdate; 1928 op_assemble = op_fallback_parent; 1929 } else if (op->LinearAssembleQFunctionUpdate) { 1930 // -- Backend version for op 1931 LinearAssembleQFunctionUpdate = op->LinearAssembleQFunctionUpdate; 1932 op_assemble = op; 1933 } 1934 1935 // Assemble QFunction 1936 if (LinearAssembleQFunctionUpdate) { 1937 // Backend or fallback parent version 1938 CeedQFunctionAssemblyData data; 1939 bool data_is_setup; 1940 CeedVector assembled_vec = NULL; 1941 CeedElemRestriction assembled_rstr = NULL; 1942 1943 CeedCall(CeedOperatorGetQFunctionAssemblyData(op, &data)); 1944 CeedCall(CeedQFunctionAssemblyDataIsSetup(data, &data_is_setup)); 1945 if (data_is_setup) { 1946 bool update_needed; 1947 1948 CeedCall(CeedQFunctionAssemblyDataGetObjects(data, &assembled_vec, &assembled_rstr)); 1949 CeedCall(CeedQFunctionAssemblyDataIsUpdateNeeded(data, &update_needed)); 1950 if (update_needed) CeedCall(LinearAssembleQFunctionUpdate(op_assemble, assembled_vec, assembled_rstr, request)); 1951 } else { 1952 CeedCall(CeedOperatorLinearAssembleQFunction(op_assemble, &assembled_vec, &assembled_rstr, request)); 1953 CeedCall(CeedQFunctionAssemblyDataSetObjects(data, assembled_vec, assembled_rstr)); 1954 } 1955 CeedCall(CeedQFunctionAssemblyDataSetUpdateNeeded(data, false)); 1956 1957 // Copy reference from internally held copy 1958 CeedCall(CeedVectorReferenceCopy(assembled_vec, assembled)); 1959 CeedCall(CeedElemRestrictionReferenceCopy(assembled_rstr, rstr)); 1960 CeedCall(CeedVectorDestroy(&assembled_vec)); 1961 CeedCall(CeedElemRestrictionDestroy(&assembled_rstr)); 1962 } else { 1963 // Operator fallback 1964 Ceed ceed; 1965 CeedOperator op_fallback; 1966 1967 CeedCall(CeedOperatorGetCeed(op, &ceed)); 1968 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 1969 if (op_fallback) CeedCall(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op_fallback, assembled, rstr, request)); 1970 else return CeedError(ceed, 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 Ceed ceed; 1996 1997 CeedCall(CeedOperatorGetCeed(op, &ceed)); 1998 CeedCall(CeedOperatorCheckReady(op)); 1999 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2000 2001 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2002 CeedCheck(input_size == output_size, ceed, CEED_ERROR_DIMENSION, "Operator must be square"); 2003 2004 // Early exit for empty operator 2005 if (!is_composite) { 2006 CeedInt num_elem = 0; 2007 2008 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2009 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2010 } 2011 2012 if (op->LinearAssembleDiagonal) { 2013 // Backend version 2014 CeedCall(op->LinearAssembleDiagonal(op, assembled, request)); 2015 return CEED_ERROR_SUCCESS; 2016 } else if (op->LinearAssembleAddDiagonal) { 2017 // Backend version with zeroing first 2018 CeedCall(CeedVectorSetValue(assembled, 0.0)); 2019 CeedCall(op->LinearAssembleAddDiagonal(op, assembled, request)); 2020 return CEED_ERROR_SUCCESS; 2021 } else { 2022 // Operator fallback 2023 CeedOperator op_fallback; 2024 2025 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2026 if (op_fallback) { 2027 CeedCall(CeedOperatorLinearAssembleDiagonal(op_fallback, assembled, request)); 2028 return CEED_ERROR_SUCCESS; 2029 } 2030 } 2031 // Default interface implementation 2032 CeedCall(CeedVectorSetValue(assembled, 0.0)); 2033 CeedCall(CeedOperatorLinearAssembleAddDiagonal(op, assembled, request)); 2034 return CEED_ERROR_SUCCESS; 2035 } 2036 2037 /** 2038 @brief Assemble the diagonal of a square linear `CeedOperator`. 2039 2040 This sums into a `CeedVector` the diagonal of a linear `CeedOperator`. 2041 2042 Note: Currently only non-composite `CeedOperator` with a single field and composite `CeedOperator` with single field sub-operators are supported. 2043 2044 Note: Calling this function asserts that setup is complete and sets the CeedOperator as immutable. 2045 2046 @param[in] op `CeedOperator` to assemble `CeedQFunction` 2047 @param[out] assembled `CeedVector` to store assembled `CeedOperator` diagonal 2048 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 2049 2050 @return An error code: 0 - success, otherwise - failure 2051 2052 @ref User 2053 **/ 2054 int CeedOperatorLinearAssembleAddDiagonal(CeedOperator op, CeedVector assembled, CeedRequest *request) { 2055 bool is_composite; 2056 CeedSize input_size = 0, output_size = 0; 2057 Ceed ceed; 2058 2059 CeedCall(CeedOperatorGetCeed(op, &ceed)); 2060 CeedCall(CeedOperatorCheckReady(op)); 2061 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2062 2063 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2064 CeedCheck(input_size == output_size, ceed, CEED_ERROR_DIMENSION, "Operator must be square"); 2065 2066 // Early exit for empty operator 2067 if (!is_composite) { 2068 CeedInt num_elem = 0; 2069 2070 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2071 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2072 } 2073 2074 if (op->LinearAssembleAddDiagonal) { 2075 // Backend version 2076 CeedCall(op->LinearAssembleAddDiagonal(op, assembled, request)); 2077 return CEED_ERROR_SUCCESS; 2078 } else { 2079 // Operator fallback 2080 CeedOperator op_fallback; 2081 2082 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2083 if (op_fallback) { 2084 CeedCall(CeedOperatorLinearAssembleAddDiagonal(op_fallback, assembled, request)); 2085 return CEED_ERROR_SUCCESS; 2086 } 2087 } 2088 // Default interface implementation 2089 if (is_composite) { 2090 CeedCall(CeedCompositeOperatorLinearAssembleAddDiagonal(op, request, false, assembled)); 2091 } else { 2092 CeedCall(CeedSingleOperatorLinearAssembleAddDiagonal(op, request, false, assembled)); 2093 } 2094 return CEED_ERROR_SUCCESS; 2095 } 2096 2097 /** 2098 @brief Fully assemble the point-block diagonal pattern of a linear `CeedOperator`. 2099 2100 Expected to be used in conjunction with @ref CeedOperatorLinearAssemblePointBlockDiagonal(). 2101 2102 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)`. 2103 Note that the `(i, j)` pairs are unique. 2104 This function returns the number of entries and their `(i, j)` locations, while @ref CeedOperatorLinearAssemblePointBlockDiagonal() provides the values in the same ordering. 2105 2106 This will generally be slow unless your operator is low-order. 2107 2108 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2109 2110 @param[in] op `CeedOperator` to assemble 2111 @param[out] num_entries Number of entries in coordinate nonzero pattern 2112 @param[out] rows Row number for each entry 2113 @param[out] cols Column number for each entry 2114 2115 @ref User 2116 **/ 2117 int CeedOperatorLinearAssemblePointBlockDiagonalSymbolic(CeedOperator op, CeedSize *num_entries, CeedInt **rows, CeedInt **cols) { 2118 Ceed ceed; 2119 bool is_composite; 2120 CeedInt num_active_components, num_sub_operators; 2121 CeedOperator *sub_operators; 2122 2123 CeedCall(CeedOperatorGetCeed(op, &ceed)); 2124 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2125 2126 CeedSize input_size = 0, output_size = 0; 2127 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2128 CeedCheck(input_size == output_size, ceed, CEED_ERROR_DIMENSION, "Operator must be square"); 2129 2130 if (is_composite) { 2131 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_sub_operators)); 2132 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2133 } else { 2134 sub_operators = &op; 2135 num_sub_operators = 1; 2136 } 2137 2138 // Verify operator can be assembled correctly 2139 { 2140 CeedOperatorAssemblyData data; 2141 CeedInt num_active_elem_rstrs, comp_stride; 2142 CeedElemRestriction *active_elem_rstrs; 2143 2144 // Get initial values to check against 2145 CeedCall(CeedOperatorGetOperatorAssemblyData(sub_operators[0], &data)); 2146 CeedCall(CeedOperatorAssemblyDataGetElemRestrictions(data, &num_active_elem_rstrs, &active_elem_rstrs, NULL, NULL)); 2147 CeedCall(CeedElemRestrictionGetCompStride(active_elem_rstrs[0], &comp_stride)); 2148 CeedCall(CeedElemRestrictionGetNumComponents(active_elem_rstrs[0], &num_active_components)); 2149 2150 // Verify that all active element restrictions have same component stride and number of components 2151 for (CeedInt k = 0; k < num_sub_operators; k++) { 2152 CeedCall(CeedOperatorGetOperatorAssemblyData(sub_operators[k], &data)); 2153 CeedCall(CeedOperatorAssemblyDataGetElemRestrictions(data, &num_active_elem_rstrs, &active_elem_rstrs, NULL, NULL)); 2154 for (CeedInt i = 0; i < num_active_elem_rstrs; i++) { 2155 CeedInt comp_stride_sub, num_active_components_sub; 2156 2157 CeedCall(CeedElemRestrictionGetCompStride(active_elem_rstrs[i], &comp_stride_sub)); 2158 CeedCheck(comp_stride == comp_stride_sub, ceed, CEED_ERROR_DIMENSION, 2159 "Active element restrictions must have the same component stride: %d vs %d", comp_stride, comp_stride_sub); 2160 CeedCall(CeedElemRestrictionGetNumComponents(active_elem_rstrs[i], &num_active_components_sub)); 2161 CeedCheck(num_active_components == num_active_components_sub, ceed, CEED_ERROR_INCOMPATIBLE, 2162 "All suboperators must have the same number of output components." 2163 " Previous: %" CeedInt_FMT " Current: %" CeedInt_FMT, 2164 num_active_components, num_active_components_sub); 2165 } 2166 } 2167 } 2168 *num_entries = input_size * num_active_components; 2169 CeedCall(CeedCalloc(*num_entries, rows)); 2170 CeedCall(CeedCalloc(*num_entries, cols)); 2171 2172 for (CeedInt o = 0; o < num_sub_operators; o++) { 2173 CeedElemRestriction active_elem_rstr, point_block_active_elem_rstr; 2174 CeedInt comp_stride, num_elem, elem_size; 2175 const CeedInt *offsets, *point_block_offsets; 2176 2177 CeedCall(CeedOperatorGetActiveElemRestriction(sub_operators[o], &active_elem_rstr)); 2178 CeedCall(CeedElemRestrictionGetCompStride(active_elem_rstr, &comp_stride)); 2179 CeedCall(CeedElemRestrictionGetNumElements(active_elem_rstr, &num_elem)); 2180 CeedCall(CeedElemRestrictionGetElementSize(active_elem_rstr, &elem_size)); 2181 CeedCall(CeedElemRestrictionGetOffsets(active_elem_rstr, CEED_MEM_HOST, &offsets)); 2182 2183 CeedCall(CeedOperatorCreateActivePointBlockRestriction(active_elem_rstr, &point_block_active_elem_rstr)); 2184 CeedCall(CeedElemRestrictionGetOffsets(point_block_active_elem_rstr, CEED_MEM_HOST, &point_block_offsets)); 2185 2186 for (CeedSize i = 0; i < num_elem * elem_size; i++) { 2187 for (CeedInt c_out = 0; c_out < num_active_components; c_out++) { 2188 for (CeedInt c_in = 0; c_in < num_active_components; c_in++) { 2189 (*rows)[point_block_offsets[i] + c_out * num_active_components + c_in] = offsets[i] + c_out * comp_stride; 2190 (*cols)[point_block_offsets[i] + c_out * num_active_components + c_in] = offsets[i] + c_in * comp_stride; 2191 } 2192 } 2193 } 2194 2195 CeedCall(CeedElemRestrictionRestoreOffsets(active_elem_rstr, &offsets)); 2196 CeedCall(CeedElemRestrictionRestoreOffsets(point_block_active_elem_rstr, &point_block_offsets)); 2197 CeedCall(CeedElemRestrictionDestroy(&active_elem_rstr)); 2198 CeedCall(CeedElemRestrictionDestroy(&point_block_active_elem_rstr)); 2199 } 2200 return CEED_ERROR_SUCCESS; 2201 } 2202 2203 /** 2204 @brief Assemble the point block diagonal of a square linear `CeedOperator`. 2205 2206 This overwrites a `CeedVector` with the point block diagonal of a linear `CeedOperator`. 2207 2208 Note: Currently only non-composite `CeedOperator` with a single field and composite `CeedOperator` with single field sub-operators are supported. 2209 2210 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2211 2212 @param[in] op `CeedOperator` to assemble `CeedQFunction` 2213 @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. 2214 The dimensions of this vector are derived from the active vector for the `CeedOperator`. 2215 The array has shape `[nodes, component out, component in]`. 2216 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 2217 2218 @return An error code: 0 - success, otherwise - failure 2219 2220 @ref User 2221 **/ 2222 int CeedOperatorLinearAssemblePointBlockDiagonal(CeedOperator op, CeedVector assembled, CeedRequest *request) { 2223 bool is_composite; 2224 CeedSize input_size = 0, output_size = 0; 2225 Ceed ceed; 2226 2227 CeedCall(CeedOperatorGetCeed(op, &ceed)); 2228 CeedCall(CeedOperatorCheckReady(op)); 2229 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2230 2231 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2232 CeedCheck(input_size == output_size, ceed, CEED_ERROR_DIMENSION, "Operator must be square"); 2233 2234 // Early exit for empty operator 2235 if (!is_composite) { 2236 CeedInt num_elem = 0; 2237 2238 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2239 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2240 } 2241 2242 if (op->LinearAssemblePointBlockDiagonal) { 2243 // Backend version 2244 CeedCall(op->LinearAssemblePointBlockDiagonal(op, assembled, request)); 2245 return CEED_ERROR_SUCCESS; 2246 } else if (op->LinearAssembleAddPointBlockDiagonal) { 2247 // Backend version with zeroing first 2248 CeedCall(CeedVectorSetValue(assembled, 0.0)); 2249 CeedCall(CeedOperatorLinearAssembleAddPointBlockDiagonal(op, assembled, request)); 2250 return CEED_ERROR_SUCCESS; 2251 } else { 2252 // Operator fallback 2253 CeedOperator op_fallback; 2254 2255 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2256 if (op_fallback) { 2257 CeedCall(CeedOperatorLinearAssemblePointBlockDiagonal(op_fallback, assembled, request)); 2258 return CEED_ERROR_SUCCESS; 2259 } 2260 } 2261 // Default interface implementation 2262 CeedCall(CeedVectorSetValue(assembled, 0.0)); 2263 CeedCall(CeedOperatorLinearAssembleAddPointBlockDiagonal(op, assembled, request)); 2264 return CEED_ERROR_SUCCESS; 2265 } 2266 2267 /** 2268 @brief Assemble the point block diagonal of a square linear `CeedOperator`. 2269 2270 This sums into a `CeedVector` with the point block diagonal of a linear `CeedOperator`. 2271 2272 Note: Currently only non-composite `CeedOperator` with a single field and composite `CeedOperator` with single field sub-operators are supported. 2273 2274 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2275 2276 @param[in] op `CeedOperator` to assemble `CeedQFunction` 2277 @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. 2278 The dimensions of this vector are derived from the active vector for the `CeedOperator`. 2279 The array has shape `[nodes, component out, component in]`. 2280 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 2281 2282 @return An error code: 0 - success, otherwise - failure 2283 2284 @ref User 2285 **/ 2286 int CeedOperatorLinearAssembleAddPointBlockDiagonal(CeedOperator op, CeedVector assembled, CeedRequest *request) { 2287 bool is_composite; 2288 CeedSize input_size = 0, output_size = 0; 2289 Ceed ceed; 2290 2291 CeedCall(CeedOperatorGetCeed(op, &ceed)); 2292 CeedCall(CeedOperatorCheckReady(op)); 2293 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2294 2295 CeedCall(CeedOperatorGetActiveVectorLengths(op, &input_size, &output_size)); 2296 CeedCheck(input_size == output_size, ceed, CEED_ERROR_DIMENSION, "Operator must be square"); 2297 2298 // Early exit for empty operator 2299 if (!is_composite) { 2300 CeedInt num_elem = 0; 2301 2302 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2303 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2304 } 2305 2306 if (op->LinearAssembleAddPointBlockDiagonal) { 2307 // Backend version 2308 CeedCall(op->LinearAssembleAddPointBlockDiagonal(op, assembled, request)); 2309 return CEED_ERROR_SUCCESS; 2310 } else { 2311 // Operator fallback 2312 CeedOperator op_fallback; 2313 2314 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2315 if (op_fallback) { 2316 CeedCall(CeedOperatorLinearAssembleAddPointBlockDiagonal(op_fallback, assembled, request)); 2317 return CEED_ERROR_SUCCESS; 2318 } 2319 } 2320 // Default interface implementation 2321 if (is_composite) { 2322 CeedCall(CeedCompositeOperatorLinearAssembleAddDiagonal(op, request, true, assembled)); 2323 } else { 2324 CeedCall(CeedSingleOperatorLinearAssembleAddDiagonal(op, request, true, assembled)); 2325 } 2326 return CEED_ERROR_SUCCESS; 2327 } 2328 2329 /** 2330 @brief Fully assemble the nonzero pattern of a linear `CeedOperator`. 2331 2332 Expected to be used in conjunction with @ref CeedOperatorLinearAssemble(). 2333 2334 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)`. 2335 Note that the `(i, j)` pairs are not unique and may repeat. 2336 This function returns the number of entries and their `(i, j)` locations, while @ref CeedOperatorLinearAssemble() provides the values in the same ordering. 2337 2338 This will generally be slow unless your operator is low-order. 2339 2340 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2341 2342 @param[in] op `CeedOperator` to assemble 2343 @param[out] num_entries Number of entries in coordinate nonzero pattern 2344 @param[out] rows Row number for each entry 2345 @param[out] cols Column number for each entry 2346 2347 @ref User 2348 **/ 2349 int CeedOperatorLinearAssembleSymbolic(CeedOperator op, CeedSize *num_entries, CeedInt **rows, CeedInt **cols) { 2350 bool is_composite; 2351 CeedInt num_suboperators, offset = 0; 2352 CeedSize single_entries; 2353 CeedOperator *sub_operators; 2354 2355 CeedCall(CeedOperatorCheckReady(op)); 2356 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2357 2358 if (op->LinearAssembleSymbolic) { 2359 // Backend version 2360 CeedCall(op->LinearAssembleSymbolic(op, num_entries, rows, cols)); 2361 return CEED_ERROR_SUCCESS; 2362 } else { 2363 // Operator fallback 2364 CeedOperator op_fallback; 2365 2366 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2367 if (op_fallback) { 2368 CeedCall(CeedOperatorLinearAssembleSymbolic(op_fallback, num_entries, rows, cols)); 2369 return CEED_ERROR_SUCCESS; 2370 } 2371 } 2372 2373 // Default interface implementation 2374 2375 // Count entries and allocate rows, cols arrays 2376 *num_entries = 0; 2377 if (is_composite) { 2378 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 2379 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2380 for (CeedInt k = 0; k < num_suboperators; ++k) { 2381 CeedCall(CeedSingleOperatorAssemblyCountEntries(sub_operators[k], &single_entries)); 2382 *num_entries += single_entries; 2383 } 2384 } else { 2385 CeedCall(CeedSingleOperatorAssemblyCountEntries(op, &single_entries)); 2386 *num_entries += single_entries; 2387 } 2388 CeedCall(CeedCalloc(*num_entries, rows)); 2389 CeedCall(CeedCalloc(*num_entries, cols)); 2390 2391 // Assemble nonzero locations 2392 if (is_composite) { 2393 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 2394 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2395 for (CeedInt k = 0; k < num_suboperators; ++k) { 2396 CeedCall(CeedSingleOperatorAssembleSymbolic(sub_operators[k], offset, *rows, *cols)); 2397 CeedCall(CeedSingleOperatorAssemblyCountEntries(sub_operators[k], &single_entries)); 2398 offset += single_entries; 2399 } 2400 } else { 2401 CeedCall(CeedSingleOperatorAssembleSymbolic(op, offset, *rows, *cols)); 2402 } 2403 return CEED_ERROR_SUCCESS; 2404 } 2405 2406 /** 2407 @brief Fully assemble the nonzero entries of a linear operator. 2408 2409 Expected to be used in conjunction with @ref CeedOperatorLinearAssembleSymbolic(). 2410 2411 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)`. 2412 Note that the `(i, j)` pairs are not unique and may repeat. 2413 This function returns the values of the nonzero entries to be added, their `(i, j)` locations are provided by @ref CeedOperatorLinearAssembleSymbolic(). 2414 2415 This will generally be slow unless your operator is low-order. 2416 2417 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2418 2419 @param[in] op `CeedOperator` to assemble 2420 @param[out] values Values to assemble into matrix 2421 2422 @ref User 2423 **/ 2424 int CeedOperatorLinearAssemble(CeedOperator op, CeedVector values) { 2425 bool is_composite; 2426 CeedInt num_suboperators, offset = 0; 2427 CeedSize single_entries = 0; 2428 CeedOperator *sub_operators; 2429 2430 CeedCall(CeedOperatorCheckReady(op)); 2431 CeedCall(CeedOperatorIsComposite(op, &is_composite)); 2432 2433 // Early exit for empty operator 2434 if (!is_composite) { 2435 CeedInt num_elem = 0; 2436 2437 CeedCall(CeedOperatorGetNumElements(op, &num_elem)); 2438 if (num_elem == 0) return CEED_ERROR_SUCCESS; 2439 } 2440 2441 if (op->LinearAssemble) { 2442 // Backend version 2443 CeedCall(op->LinearAssemble(op, values)); 2444 return CEED_ERROR_SUCCESS; 2445 } else { 2446 // Operator fallback 2447 CeedOperator op_fallback; 2448 2449 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2450 if (op_fallback) { 2451 CeedCall(CeedOperatorLinearAssemble(op_fallback, values)); 2452 return CEED_ERROR_SUCCESS; 2453 } 2454 } 2455 2456 // Default interface implementation 2457 CeedCall(CeedVectorSetValue(values, 0.0)); 2458 if (is_composite) { 2459 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 2460 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2461 for (CeedInt k = 0; k < num_suboperators; k++) { 2462 CeedCall(CeedSingleOperatorAssemble(sub_operators[k], offset, values)); 2463 CeedCall(CeedSingleOperatorAssemblyCountEntries(sub_operators[k], &single_entries)); 2464 offset += single_entries; 2465 } 2466 } else { 2467 CeedCall(CeedSingleOperatorAssemble(op, offset, values)); 2468 } 2469 return CEED_ERROR_SUCCESS; 2470 } 2471 2472 /** 2473 @brief Get the multiplicity of nodes across sub-operators in a composite `CeedOperator`. 2474 2475 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2476 2477 @param[in] op Composite `CeedOperator` 2478 @param[in] num_skip_indices Number of sub-operators to skip 2479 @param[in] skip_indices Array of indices of sub-operators to skip 2480 @param[out] mult Vector to store multiplicity (of size `l_size` ) 2481 2482 @return An error code: 0 - success, otherwise - failure 2483 2484 @ref User 2485 **/ 2486 int CeedCompositeOperatorGetMultiplicity(CeedOperator op, CeedInt num_skip_indices, CeedInt *skip_indices, CeedVector mult) { 2487 Ceed ceed; 2488 CeedInt num_suboperators; 2489 CeedSize l_vec_len; 2490 CeedScalar *mult_array; 2491 CeedVector ones_l_vec; 2492 CeedElemRestriction elem_rstr, mult_elem_rstr; 2493 CeedOperator *sub_operators; 2494 2495 CeedCall(CeedOperatorCheckReady(op)); 2496 2497 CeedCall(CeedOperatorGetCeed(op, &ceed)); 2498 2499 // Zero mult vector 2500 CeedCall(CeedVectorSetValue(mult, 0.0)); 2501 2502 // Get suboperators 2503 CeedCall(CeedCompositeOperatorGetNumSub(op, &num_suboperators)); 2504 CeedCall(CeedCompositeOperatorGetSubList(op, &sub_operators)); 2505 if (num_suboperators == 0) return CEED_ERROR_SUCCESS; 2506 2507 // Work vector 2508 CeedCall(CeedVectorGetLength(mult, &l_vec_len)); 2509 CeedCall(CeedVectorCreate(ceed, l_vec_len, &ones_l_vec)); 2510 CeedCall(CeedVectorSetValue(ones_l_vec, 1.0)); 2511 CeedCall(CeedVectorGetArray(mult, CEED_MEM_HOST, &mult_array)); 2512 2513 // Compute multiplicity across suboperators 2514 for (CeedInt i = 0; i < num_suboperators; i++) { 2515 const CeedScalar *sub_mult_array; 2516 CeedVector sub_mult_l_vec, ones_e_vec; 2517 2518 // -- Check for suboperator to skip 2519 for (CeedInt j = 0; j < num_skip_indices; j++) { 2520 if (skip_indices[j] == i) continue; 2521 } 2522 2523 // -- Sub operator multiplicity 2524 CeedCall(CeedOperatorGetActiveElemRestriction(sub_operators[i], &elem_rstr)); 2525 CeedCall(CeedElemRestrictionCreateUnorientedCopy(elem_rstr, &mult_elem_rstr)); 2526 CeedCall(CeedElemRestrictionDestroy(&elem_rstr)); 2527 CeedCall(CeedElemRestrictionCreateVector(mult_elem_rstr, &sub_mult_l_vec, &ones_e_vec)); 2528 CeedCall(CeedVectorSetValue(sub_mult_l_vec, 0.0)); 2529 CeedCall(CeedElemRestrictionApply(mult_elem_rstr, CEED_NOTRANSPOSE, ones_l_vec, ones_e_vec, CEED_REQUEST_IMMEDIATE)); 2530 CeedCall(CeedElemRestrictionApply(mult_elem_rstr, CEED_TRANSPOSE, ones_e_vec, sub_mult_l_vec, CEED_REQUEST_IMMEDIATE)); 2531 CeedCall(CeedVectorGetArrayRead(sub_mult_l_vec, CEED_MEM_HOST, &sub_mult_array)); 2532 // ---- Flag every node present in the current suboperator 2533 for (CeedSize j = 0; j < l_vec_len; j++) { 2534 if (sub_mult_array[j] > 0.0) mult_array[j] += 1.0; 2535 } 2536 CeedCall(CeedVectorRestoreArrayRead(sub_mult_l_vec, &sub_mult_array)); 2537 CeedCall(CeedVectorDestroy(&sub_mult_l_vec)); 2538 CeedCall(CeedVectorDestroy(&ones_e_vec)); 2539 CeedCall(CeedElemRestrictionDestroy(&mult_elem_rstr)); 2540 } 2541 CeedCall(CeedVectorRestoreArray(mult, &mult_array)); 2542 CeedCall(CeedVectorDestroy(&ones_l_vec)); 2543 return CEED_ERROR_SUCCESS; 2544 } 2545 2546 /** 2547 @brief Create a multigrid coarse `CeedOperator` and level transfer `CeedOperator` for a `CeedOperator`, creating the prolongation basis from the fine and coarse grid interpolation. 2548 2549 Note: Calling this function asserts that setup is complete and sets all four `CeedOperator` as immutable. 2550 2551 @param[in] op_fine Fine grid `CeedOperator` 2552 @param[in] p_mult_fine L-vector multiplicity in parallel gather/scatter, or `NULL` if not creating prolongation/restriction `CeedOperator` 2553 @param[in] rstr_coarse Coarse grid `CeedElemRestriction` 2554 @param[in] basis_coarse Coarse grid active vector `CeedBasis` 2555 @param[out] op_coarse Coarse grid `CeedOperator` 2556 @param[out] op_prolong Coarse to fine `CeedOperator`, or `NULL` 2557 @param[out] op_restrict Fine to coarse `CeedOperator`, or `NULL` 2558 2559 @return An error code: 0 - success, otherwise - failure 2560 2561 @ref User 2562 **/ 2563 int CeedOperatorMultigridLevelCreate(CeedOperator op_fine, CeedVector p_mult_fine, CeedElemRestriction rstr_coarse, CeedBasis basis_coarse, 2564 CeedOperator *op_coarse, CeedOperator *op_prolong, CeedOperator *op_restrict) { 2565 CeedBasis basis_c_to_f = NULL; 2566 2567 CeedCall(CeedOperatorCheckReady(op_fine)); 2568 2569 // Build prolongation matrix, if required 2570 if (op_prolong || op_restrict) { 2571 CeedBasis basis_fine; 2572 2573 CeedCall(CeedOperatorGetActiveBasis(op_fine, &basis_fine)); 2574 CeedCall(CeedBasisCreateProjection(basis_coarse, basis_fine, &basis_c_to_f)); 2575 CeedCall(CeedBasisDestroy(&basis_fine)); 2576 } 2577 2578 // Core code 2579 CeedCall(CeedSingleOperatorMultigridLevel(op_fine, p_mult_fine, rstr_coarse, basis_coarse, basis_c_to_f, op_coarse, op_prolong, op_restrict)); 2580 return CEED_ERROR_SUCCESS; 2581 } 2582 2583 /** 2584 @brief Create a multigrid coarse `CeedOperator` and level transfer `CeedOperator` for a `CeedOperator` with a tensor basis for the active basis. 2585 2586 Note: Calling this function asserts that setup is complete and sets all four `CeedOperator` as immutable. 2587 2588 @param[in] op_fine Fine grid `CeedOperator` 2589 @param[in] p_mult_fine L-vector multiplicity in parallel gather/scatter, or `NULL` if not creating prolongation/restriction `CeedOperator` 2590 @param[in] rstr_coarse Coarse grid `CeedElemRestriction` 2591 @param[in] basis_coarse Coarse grid active vector `CeedBasis` 2592 @param[in] interp_c_to_f Matrix for coarse to fine interpolation, or `NULL` if not creating prolongation/restriction `CeedOperator` 2593 @param[out] op_coarse Coarse grid `CeedOperator` 2594 @param[out] op_prolong Coarse to fine `CeedOperator`, or `NULL` 2595 @param[out] op_restrict Fine to coarse `CeedOperator`, or `NULL` 2596 2597 @return An error code: 0 - success, otherwise - failure 2598 2599 @ref User 2600 **/ 2601 int CeedOperatorMultigridLevelCreateTensorH1(CeedOperator op_fine, CeedVector p_mult_fine, CeedElemRestriction rstr_coarse, CeedBasis basis_coarse, 2602 const CeedScalar *interp_c_to_f, CeedOperator *op_coarse, CeedOperator *op_prolong, 2603 CeedOperator *op_restrict) { 2604 Ceed ceed; 2605 CeedInt Q_f, Q_c; 2606 CeedBasis basis_fine, basis_c_to_f = NULL; 2607 2608 CeedCall(CeedOperatorCheckReady(op_fine)); 2609 CeedCall(CeedOperatorGetCeed(op_fine, &ceed)); 2610 2611 // Check for compatible quadrature spaces 2612 CeedCall(CeedOperatorGetActiveBasis(op_fine, &basis_fine)); 2613 CeedCall(CeedBasisGetNumQuadraturePoints(basis_fine, &Q_f)); 2614 CeedCall(CeedBasisGetNumQuadraturePoints(basis_coarse, &Q_c)); 2615 CeedCheck(Q_f == Q_c, ceed, CEED_ERROR_DIMENSION, 2616 "Bases must have compatible quadrature spaces." 2617 " Fine grid: %" CeedInt_FMT " points, Coarse grid: %" CeedInt_FMT " points", 2618 Q_f, Q_c); 2619 2620 // Create coarse to fine basis, if required 2621 if (op_prolong || op_restrict) { 2622 CeedInt dim, num_comp, num_nodes_c, P_1d_f, P_1d_c; 2623 CeedScalar *q_ref, *q_weight, *grad; 2624 2625 // Check if interpolation matrix is provided 2626 CeedCheck(interp_c_to_f, ceed, CEED_ERROR_INCOMPATIBLE, 2627 "Prolongation or restriction operator creation requires coarse-to-fine interpolation matrix"); 2628 CeedCall(CeedBasisGetDimension(basis_fine, &dim)); 2629 CeedCall(CeedBasisGetNumComponents(basis_fine, &num_comp)); 2630 CeedCall(CeedBasisGetNumNodes1D(basis_fine, &P_1d_f)); 2631 CeedCall(CeedBasisDestroy(&basis_fine)); 2632 CeedCall(CeedElemRestrictionGetElementSize(rstr_coarse, &num_nodes_c)); 2633 P_1d_c = dim == 1 ? num_nodes_c : dim == 2 ? sqrt(num_nodes_c) : cbrt(num_nodes_c); 2634 CeedCall(CeedCalloc(P_1d_f, &q_ref)); 2635 CeedCall(CeedCalloc(P_1d_f, &q_weight)); 2636 CeedCall(CeedCalloc(P_1d_f * P_1d_c * dim, &grad)); 2637 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)); 2638 CeedCall(CeedFree(&q_ref)); 2639 CeedCall(CeedFree(&q_weight)); 2640 CeedCall(CeedFree(&grad)); 2641 } 2642 2643 // Core code 2644 CeedCall(CeedSingleOperatorMultigridLevel(op_fine, p_mult_fine, rstr_coarse, basis_coarse, basis_c_to_f, op_coarse, op_prolong, op_restrict)); 2645 return CEED_ERROR_SUCCESS; 2646 } 2647 2648 /** 2649 @brief Create a multigrid coarse `CeedOperator` and level transfer `CeedOperator` for a `CeedOperator` with a non-tensor basis for the active vector 2650 2651 Note: Calling this function asserts that setup is complete and sets all four `CeedOperator` as immutable. 2652 2653 @param[in] op_fine Fine grid `CeedOperator` 2654 @param[in] p_mult_fine L-vector multiplicity in parallel gather/scatter, or `NULL` if not creating prolongation/restriction `CeedOperator` 2655 @param[in] rstr_coarse Coarse grid `CeedElemRestriction` 2656 @param[in] basis_coarse Coarse grid active vector `CeedBasis` 2657 @param[in] interp_c_to_f Matrix for coarse to fine interpolation, or `NULL` if not creating prolongation/restriction `CeedOperator` 2658 @param[out] op_coarse Coarse grid `CeedOperator` 2659 @param[out] op_prolong Coarse to fine `CeedOperator`, or `NULL` 2660 @param[out] op_restrict Fine to coarse `CeedOperator`, or `NULL` 2661 2662 @return An error code: 0 - success, otherwise - failure 2663 2664 @ref User 2665 **/ 2666 int CeedOperatorMultigridLevelCreateH1(CeedOperator op_fine, CeedVector p_mult_fine, CeedElemRestriction rstr_coarse, CeedBasis basis_coarse, 2667 const CeedScalar *interp_c_to_f, CeedOperator *op_coarse, CeedOperator *op_prolong, 2668 CeedOperator *op_restrict) { 2669 Ceed ceed; 2670 CeedInt Q_f, Q_c; 2671 CeedBasis basis_fine, basis_c_to_f = NULL; 2672 2673 CeedCall(CeedOperatorCheckReady(op_fine)); 2674 CeedCall(CeedOperatorGetCeed(op_fine, &ceed)); 2675 2676 // Check for compatible quadrature spaces 2677 CeedCall(CeedOperatorGetActiveBasis(op_fine, &basis_fine)); 2678 CeedCall(CeedBasisGetNumQuadraturePoints(basis_fine, &Q_f)); 2679 CeedCall(CeedBasisGetNumQuadraturePoints(basis_coarse, &Q_c)); 2680 CeedCheck(Q_f == Q_c, ceed, CEED_ERROR_DIMENSION, "Bases must have compatible quadrature spaces"); 2681 2682 // Coarse to fine basis 2683 if (op_prolong || op_restrict) { 2684 CeedInt dim, num_comp, num_nodes_c, num_nodes_f; 2685 CeedScalar *q_ref, *q_weight, *grad; 2686 CeedElemTopology topo; 2687 2688 // Check if interpolation matrix is provided 2689 CeedCheck(interp_c_to_f, ceed, CEED_ERROR_INCOMPATIBLE, 2690 "Prolongation or restriction operator creation requires coarse-to-fine interpolation matrix"); 2691 CeedCall(CeedBasisGetTopology(basis_fine, &topo)); 2692 CeedCall(CeedBasisGetDimension(basis_fine, &dim)); 2693 CeedCall(CeedBasisGetNumComponents(basis_fine, &num_comp)); 2694 CeedCall(CeedBasisGetNumNodes(basis_fine, &num_nodes_f)); 2695 CeedCall(CeedBasisDestroy(&basis_fine)); 2696 CeedCall(CeedElemRestrictionGetElementSize(rstr_coarse, &num_nodes_c)); 2697 CeedCall(CeedCalloc(num_nodes_f * dim, &q_ref)); 2698 CeedCall(CeedCalloc(num_nodes_f, &q_weight)); 2699 CeedCall(CeedCalloc(num_nodes_f * num_nodes_c * dim, &grad)); 2700 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)); 2701 CeedCall(CeedFree(&q_ref)); 2702 CeedCall(CeedFree(&q_weight)); 2703 CeedCall(CeedFree(&grad)); 2704 } 2705 2706 // Core code 2707 CeedCall(CeedSingleOperatorMultigridLevel(op_fine, p_mult_fine, rstr_coarse, basis_coarse, basis_c_to_f, op_coarse, op_prolong, op_restrict)); 2708 return CEED_ERROR_SUCCESS; 2709 } 2710 2711 /** 2712 @brief Build a FDM based approximate inverse for each element for a `CeedOperator`. 2713 2714 This returns a `CeedOperator` and `CeedVector` to apply a Fast Diagonalization Method based approximate inverse. 2715 This function obtains the simultaneous diagonalization for the 1D mass and Laplacian operators, \f$M = V^T V, K = V^T S V\f$. 2716 The assembled `CeedQFunction` is used to modify the eigenvalues from simultaneous diagonalization and obtain an approximate inverse of the form \f$V^T \hat S V\f$. 2717 The `CeedOperator` must be linear and non-composite. 2718 The associated `CeedQFunction` must therefore also be linear. 2719 2720 Note: Calling this function asserts that setup is complete and sets the `CeedOperator` as immutable. 2721 2722 @param[in] op `CeedOperator` to create element inverses 2723 @param[out] fdm_inv `CeedOperator` to apply the action of a FDM based inverse for each element 2724 @param[in] request Address of @ref CeedRequest for non-blocking completion, else @ref CEED_REQUEST_IMMEDIATE 2725 2726 @return An error code: 0 - success, otherwise - failure 2727 2728 @ref User 2729 **/ 2730 int CeedOperatorCreateFDMElementInverse(CeedOperator op, CeedOperator *fdm_inv, CeedRequest *request) { 2731 Ceed ceed, ceed_parent; 2732 bool interp = false, grad = false, is_tensor_basis = true; 2733 CeedInt num_input_fields, P_1d, Q_1d, num_nodes, num_qpts, dim, num_comp = 1, num_elem = 1; 2734 CeedScalar *mass, *laplace, *x, *fdm_interp, *lambda, *elem_avg; 2735 const CeedScalar *interp_1d, *grad_1d, *q_weight_1d; 2736 CeedVector q_data; 2737 CeedElemRestriction rstr = NULL, rstr_qd_i; 2738 CeedBasis basis = NULL, fdm_basis; 2739 CeedQFunctionContext ctx_fdm; 2740 CeedQFunctionField *qf_fields; 2741 CeedQFunction qf, qf_fdm; 2742 CeedOperatorField *op_fields; 2743 2744 CeedCall(CeedOperatorCheckReady(op)); 2745 2746 if (op->CreateFDMElementInverse) { 2747 // Backend version 2748 CeedCall(op->CreateFDMElementInverse(op, fdm_inv, request)); 2749 return CEED_ERROR_SUCCESS; 2750 } else { 2751 // Operator fallback 2752 CeedOperator op_fallback; 2753 2754 CeedCall(CeedOperatorGetFallback(op, &op_fallback)); 2755 if (op_fallback) { 2756 CeedCall(CeedOperatorCreateFDMElementInverse(op_fallback, fdm_inv, request)); 2757 return CEED_ERROR_SUCCESS; 2758 } 2759 } 2760 2761 // Default interface implementation 2762 CeedCall(CeedOperatorGetCeed(op, &ceed)); 2763 CeedCall(CeedOperatorGetFallbackParentCeed(op, &ceed_parent)); 2764 CeedCall(CeedOperatorGetQFunction(op, &qf)); 2765 2766 // Determine active input basis 2767 CeedCall(CeedOperatorGetFields(op, &num_input_fields, &op_fields, NULL, NULL)); 2768 CeedCall(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 2769 for (CeedInt i = 0; i < num_input_fields; i++) { 2770 CeedVector vec; 2771 2772 CeedCall(CeedOperatorFieldGetVector(op_fields[i], &vec)); 2773 if (vec == CEED_VECTOR_ACTIVE) { 2774 CeedEvalMode eval_mode; 2775 2776 CeedCall(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 2777 interp = interp || eval_mode == CEED_EVAL_INTERP; 2778 grad = grad || eval_mode == CEED_EVAL_GRAD; 2779 if (!basis) CeedCall(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 2780 if (!rstr) CeedCall(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr)); 2781 } 2782 CeedCall(CeedVectorDestroy(&vec)); 2783 } 2784 CeedCheck(basis, ceed, CEED_ERROR_BACKEND, "No active field set"); 2785 CeedCall(CeedBasisGetNumNodes1D(basis, &P_1d)); 2786 CeedCall(CeedBasisGetNumNodes(basis, &num_nodes)); 2787 CeedCall(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 2788 CeedCall(CeedBasisGetNumQuadraturePoints(basis, &num_qpts)); 2789 CeedCall(CeedBasisGetDimension(basis, &dim)); 2790 CeedCall(CeedBasisGetNumComponents(basis, &num_comp)); 2791 CeedCall(CeedElemRestrictionGetNumElements(rstr, &num_elem)); 2792 2793 // Build and diagonalize 1D Mass and Laplacian 2794 CeedCall(CeedBasisIsTensor(basis, &is_tensor_basis)); 2795 CeedCheck(is_tensor_basis, ceed, CEED_ERROR_BACKEND, "FDMElementInverse only supported for tensor bases"); 2796 CeedCall(CeedCalloc(P_1d * P_1d, &mass)); 2797 CeedCall(CeedCalloc(P_1d * P_1d, &laplace)); 2798 CeedCall(CeedCalloc(P_1d * P_1d, &x)); 2799 CeedCall(CeedCalloc(P_1d * P_1d, &fdm_interp)); 2800 CeedCall(CeedCalloc(P_1d, &lambda)); 2801 // -- Build matrices 2802 CeedCall(CeedBasisGetInterp1D(basis, &interp_1d)); 2803 CeedCall(CeedBasisGetGrad1D(basis, &grad_1d)); 2804 CeedCall(CeedBasisGetQWeights(basis, &q_weight_1d)); 2805 CeedCall(CeedBuildMassLaplace(interp_1d, grad_1d, q_weight_1d, P_1d, Q_1d, dim, mass, laplace)); 2806 2807 // -- Diagonalize 2808 CeedCall(CeedSimultaneousDiagonalization(ceed, laplace, mass, x, lambda, P_1d)); 2809 CeedCall(CeedFree(&mass)); 2810 CeedCall(CeedFree(&laplace)); 2811 for (CeedInt i = 0; i < P_1d; i++) { 2812 for (CeedInt j = 0; j < P_1d; j++) fdm_interp[i + j * P_1d] = x[j + i * P_1d]; 2813 } 2814 CeedCall(CeedFree(&x)); 2815 2816 { 2817 CeedInt layout[3], num_modes = (interp ? 1 : 0) + (grad ? dim : 0); 2818 CeedScalar max_norm = 0; 2819 const CeedScalar *assembled_array, *q_weight_array; 2820 CeedVector assembled = NULL, q_weight; 2821 CeedElemRestriction rstr_qf = NULL; 2822 2823 // Assemble QFunction 2824 CeedCall(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled, &rstr_qf, request)); 2825 CeedCall(CeedElemRestrictionGetELayout(rstr_qf, layout)); 2826 CeedCall(CeedElemRestrictionDestroy(&rstr_qf)); 2827 CeedCall(CeedVectorNorm(assembled, CEED_NORM_MAX, &max_norm)); 2828 2829 // Calculate element averages 2830 CeedCall(CeedVectorCreate(ceed_parent, num_qpts, &q_weight)); 2831 CeedCall(CeedBasisApply(basis, 1, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, q_weight)); 2832 CeedCall(CeedVectorGetArrayRead(assembled, CEED_MEM_HOST, &assembled_array)); 2833 CeedCall(CeedVectorGetArrayRead(q_weight, CEED_MEM_HOST, &q_weight_array)); 2834 CeedCall(CeedCalloc(num_elem, &elem_avg)); 2835 const CeedScalar qf_value_bound = max_norm * 100 * CEED_EPSILON; 2836 2837 for (CeedInt e = 0; e < num_elem; e++) { 2838 CeedInt count = 0; 2839 2840 for (CeedInt q = 0; q < num_qpts; q++) { 2841 for (CeedInt i = 0; i < num_comp * num_comp * num_modes * num_modes; i++) { 2842 if (fabs(assembled_array[q * layout[0] + i * layout[1] + e * layout[2]]) > qf_value_bound) { 2843 elem_avg[e] += assembled_array[q * layout[0] + i * layout[1] + e * layout[2]] / q_weight_array[q]; 2844 count++; 2845 } 2846 } 2847 } 2848 if (count) { 2849 elem_avg[e] /= count; 2850 } else { 2851 elem_avg[e] = 1.0; 2852 } 2853 } 2854 CeedCall(CeedVectorRestoreArrayRead(assembled, &assembled_array)); 2855 CeedCall(CeedVectorDestroy(&assembled)); 2856 CeedCall(CeedVectorRestoreArrayRead(q_weight, &q_weight_array)); 2857 CeedCall(CeedVectorDestroy(&q_weight)); 2858 } 2859 2860 // Build FDM diagonal 2861 { 2862 CeedScalar *q_data_array, *fdm_diagonal; 2863 2864 CeedCall(CeedCalloc(num_comp * num_nodes, &fdm_diagonal)); 2865 const CeedScalar fdm_diagonal_bound = num_nodes * CEED_EPSILON; 2866 for (CeedInt c = 0; c < num_comp; c++) { 2867 for (CeedInt n = 0; n < num_nodes; n++) { 2868 if (interp) fdm_diagonal[c * num_nodes + n] = 1.0; 2869 if (grad) { 2870 for (CeedInt d = 0; d < dim; d++) { 2871 CeedInt i = (n / CeedIntPow(P_1d, d)) % P_1d; 2872 fdm_diagonal[c * num_nodes + n] += lambda[i]; 2873 } 2874 } 2875 if (fabs(fdm_diagonal[c * num_nodes + n]) < fdm_diagonal_bound) fdm_diagonal[c * num_nodes + n] = fdm_diagonal_bound; 2876 } 2877 } 2878 CeedCall(CeedVectorCreate(ceed_parent, num_elem * num_comp * num_nodes, &q_data)); 2879 CeedCall(CeedVectorSetValue(q_data, 0.0)); 2880 CeedCall(CeedVectorGetArrayWrite(q_data, CEED_MEM_HOST, &q_data_array)); 2881 for (CeedInt e = 0; e < num_elem; e++) { 2882 for (CeedInt c = 0; c < num_comp; c++) { 2883 for (CeedInt n = 0; n < num_nodes; n++) { 2884 q_data_array[(e * num_comp + c) * num_nodes + n] = 1. / (elem_avg[e] * fdm_diagonal[c * num_nodes + n]); 2885 } 2886 } 2887 } 2888 CeedCall(CeedFree(&elem_avg)); 2889 CeedCall(CeedFree(&fdm_diagonal)); 2890 CeedCall(CeedVectorRestoreArray(q_data, &q_data_array)); 2891 } 2892 2893 // Setup FDM operator 2894 // -- Basis 2895 { 2896 CeedScalar *grad_dummy, *q_ref_dummy, *q_weight_dummy; 2897 2898 CeedCall(CeedCalloc(P_1d * P_1d, &grad_dummy)); 2899 CeedCall(CeedCalloc(P_1d, &q_ref_dummy)); 2900 CeedCall(CeedCalloc(P_1d, &q_weight_dummy)); 2901 CeedCall(CeedBasisCreateTensorH1(ceed_parent, dim, num_comp, P_1d, P_1d, fdm_interp, grad_dummy, q_ref_dummy, q_weight_dummy, &fdm_basis)); 2902 CeedCall(CeedFree(&fdm_interp)); 2903 CeedCall(CeedFree(&grad_dummy)); 2904 CeedCall(CeedFree(&q_ref_dummy)); 2905 CeedCall(CeedFree(&q_weight_dummy)); 2906 CeedCall(CeedFree(&lambda)); 2907 } 2908 2909 // -- Restriction 2910 { 2911 CeedInt strides[3] = {1, num_nodes, num_nodes * num_comp}; 2912 CeedCall(CeedElemRestrictionCreateStrided(ceed_parent, num_elem, num_nodes, num_comp, 2913 (CeedSize)num_elem * (CeedSize)num_comp * (CeedSize)num_nodes, strides, &rstr_qd_i)); 2914 } 2915 2916 // -- QFunction 2917 CeedCall(CeedQFunctionCreateInteriorByName(ceed_parent, "Scale", &qf_fdm)); 2918 CeedCall(CeedQFunctionAddInput(qf_fdm, "input", num_comp, CEED_EVAL_INTERP)); 2919 CeedCall(CeedQFunctionAddInput(qf_fdm, "scale", num_comp, CEED_EVAL_NONE)); 2920 CeedCall(CeedQFunctionAddOutput(qf_fdm, "output", num_comp, CEED_EVAL_INTERP)); 2921 CeedCall(CeedQFunctionSetUserFlopsEstimate(qf_fdm, num_comp)); 2922 2923 // -- QFunction context 2924 { 2925 CeedInt *num_comp_data; 2926 2927 CeedCall(CeedCalloc(1, &num_comp_data)); 2928 num_comp_data[0] = num_comp; 2929 CeedCall(CeedQFunctionContextCreate(ceed, &ctx_fdm)); 2930 CeedCall(CeedQFunctionContextSetData(ctx_fdm, CEED_MEM_HOST, CEED_OWN_POINTER, sizeof(*num_comp_data), num_comp_data)); 2931 } 2932 CeedCall(CeedQFunctionSetContext(qf_fdm, ctx_fdm)); 2933 CeedCall(CeedQFunctionContextDestroy(&ctx_fdm)); 2934 2935 // -- Operator 2936 CeedCall(CeedOperatorCreate(ceed_parent, qf_fdm, NULL, NULL, fdm_inv)); 2937 CeedCall(CeedOperatorSetField(*fdm_inv, "input", rstr, fdm_basis, CEED_VECTOR_ACTIVE)); 2938 CeedCall(CeedOperatorSetField(*fdm_inv, "scale", rstr_qd_i, CEED_BASIS_NONE, q_data)); 2939 CeedCall(CeedOperatorSetField(*fdm_inv, "output", rstr, fdm_basis, CEED_VECTOR_ACTIVE)); 2940 2941 // Cleanup 2942 CeedCall(CeedVectorDestroy(&q_data)); 2943 CeedCall(CeedElemRestrictionDestroy(&rstr)); 2944 CeedCall(CeedElemRestrictionDestroy(&rstr_qd_i)); 2945 CeedCall(CeedBasisDestroy(&basis)); 2946 CeedCall(CeedBasisDestroy(&fdm_basis)); 2947 CeedCall(CeedQFunctionDestroy(&qf_fdm)); 2948 return CEED_ERROR_SUCCESS; 2949 } 2950 2951 /// @} 2952