1 // SPDX-FileCopyrightText: Copyright (c) 2017-2024, HONEE contributors. 2 // SPDX-License-Identifier: Apache-2.0 OR BSD-2-Clause 3 4 #include "../../qfunctions/sgs_dd_training.h" 5 6 #include <differential_filter.h> 7 #include <navierstokes.h> 8 #include <petscdmplex.h> 9 #include <smartsim.h> 10 11 typedef struct { 12 DM dm_dd_training; 13 PetscInt num_comp_dd_inputs, write_data_interval, num_filter_widths; 14 PetscScalar filter_widths[16]; 15 OperatorApplyContext op_training_data_calc_ctx; 16 DiffFilterData diff_filter; 17 NodalProjectionData filtered_grad_velo_proj; 18 size_t training_data_array_dims[2]; 19 PetscBool overwrite_training_data; 20 } *SGS_DD_TrainingData; 21 22 #define SGS_DD_TRAIN_KEY "SGS Data Driven Training" 23 24 PetscErrorCode SGS_DD_TrainingDataDestroy(SGS_DD_TrainingData *sgs_dd_train) { 25 SGS_DD_TrainingData sgs_dd_train_ = *sgs_dd_train; 26 27 PetscFunctionBeginUser; 28 if (!sgs_dd_train_) PetscFunctionReturn(PETSC_SUCCESS); 29 PetscCall(OperatorApplyContextDestroy(sgs_dd_train_->op_training_data_calc_ctx)); 30 PetscCall(NodalProjectionDataDestroy(&sgs_dd_train_->filtered_grad_velo_proj)); 31 PetscCall(DMDestroy(&sgs_dd_train_->dm_dd_training)); 32 PetscCall(DifferentialFilterDataDestroy(&sgs_dd_train_->diff_filter)); 33 PetscCall(PetscFree(sgs_dd_train)); 34 PetscFunctionReturn(PETSC_SUCCESS); 35 } 36 37 typedef struct { 38 CeedElemRestriction elem_restr_grid_aniso; 39 CeedVector grid_aniso_ceed; 40 CeedQFunctionContext sgs_dd_train_qfctx; 41 } *SGS_DD_TrainingSetupData; 42 43 static PetscErrorCode SGS_DD_TrainingSetupDataDestroy(SGS_DD_TrainingSetupData sgs_dd_train_setup_data) { 44 Ceed ceed; 45 46 PetscFunctionBeginUser; 47 PetscCall(CeedElemRestrictionGetCeed(sgs_dd_train_setup_data->elem_restr_grid_aniso, &ceed)); 48 49 PetscCallCeed(ceed, CeedElemRestrictionDestroy(&sgs_dd_train_setup_data->elem_restr_grid_aniso)); 50 PetscCallCeed(ceed, CeedVectorDestroy(&sgs_dd_train_setup_data->grid_aniso_ceed)); 51 PetscCallCeed(ceed, CeedQFunctionContextDestroy(&sgs_dd_train_setup_data->sgs_dd_train_qfctx)); 52 PetscCall(PetscFree(sgs_dd_train_setup_data)); 53 PetscCheck(CeedDestroy(&ceed) == CEED_ERROR_SUCCESS, PETSC_COMM_SELF, PETSC_ERR_LIB, "Destroying Ceed object failed"); 54 PetscFunctionReturn(PETSC_SUCCESS); 55 } 56 57 // @brief Create DM for storing data-drive SGS model inputs 58 static PetscErrorCode SGS_DD_TrainingCreateDM(DM dm_source, DM *dm_dd_training, PetscInt degree, PetscInt q_extra, PetscInt *num_components) { 59 PetscSection section; 60 61 PetscFunctionBeginUser; 62 *num_components = 12; 63 64 PetscCall(DMClone(dm_source, dm_dd_training)); 65 PetscCall(DMSetMatrixPreallocateSkip(*dm_dd_training, PETSC_TRUE)); 66 PetscCall(PetscObjectSetName((PetscObject)*dm_dd_training, "Data-Driven SGS Training Data")); 67 68 PetscCall(DMSetupByOrder_FEM(PETSC_TRUE, PETSC_TRUE, degree, 1, q_extra, 1, num_components, *dm_dd_training)); 69 70 PetscCall(DMGetLocalSection(*dm_dd_training, §ion)); 71 PetscCall(PetscSectionSetFieldName(section, 0, "Data-Driven SGS Training Data")); 72 PetscCall(PetscSectionSetComponentName(section, 0, 0, "SGSInput1")); 73 PetscCall(PetscSectionSetComponentName(section, 0, 1, "SGSInput2")); 74 PetscCall(PetscSectionSetComponentName(section, 0, 2, "SGSInput3")); 75 PetscCall(PetscSectionSetComponentName(section, 0, 3, "SGSInput4")); 76 PetscCall(PetscSectionSetComponentName(section, 0, 4, "SGSInput5")); 77 PetscCall(PetscSectionSetComponentName(section, 0, 5, "SGSInput6")); 78 PetscCall(PetscSectionSetComponentName(section, 0, 6, "FilteredSGSXX")); 79 PetscCall(PetscSectionSetComponentName(section, 0, 7, "FilteredSGSYY")); 80 PetscCall(PetscSectionSetComponentName(section, 0, 8, "FilteredSGSZZ")); 81 PetscCall(PetscSectionSetComponentName(section, 0, 9, "FilteredSGSYZ")); 82 PetscCall(PetscSectionSetComponentName(section, 0, 10, "FilteredSGSXZ")); 83 PetscCall(PetscSectionSetComponentName(section, 0, 11, "FilteredSGSXY")); 84 PetscFunctionReturn(PETSC_SUCCESS); 85 }; 86 87 // @brief Create CeedOperator to calculate training data for data-drive SGS model at nodes 88 static PetscErrorCode SetupTrainingDataCalculation(Ceed ceed, Honee honee, ProblemData problem, SGS_DD_TrainingSetupData sgs_dd_train_setup_data) { 89 SGS_DD_TrainingData sgs_dd_train; 90 CeedQFunction qf_sgs_dd_train; 91 CeedOperator op_sgs_dd_train; 92 CeedInt num_comp_grad_velo, num_comp_grid_aniso; 93 CeedVector inv_multiplicity, filtered_fields; 94 CeedElemRestriction elem_restr_inv_multiplicity, elem_restr_grad_velo, elem_restr_sgs_train; 95 DMLabel domain_label = NULL; 96 PetscInt label_value = 0, height = 0, dm_field = 0; 97 98 PetscFunctionBeginUser; 99 PetscCall(HoneeGetContainer(honee, SGS_DD_TRAIN_KEY, &sgs_dd_train)); 100 PetscCallCeed(ceed, CeedElemRestrictionGetNumComponents(sgs_dd_train_setup_data->elem_restr_grid_aniso, &num_comp_grid_aniso)); 101 102 PetscCall(DMPlexCeedElemRestrictionCreate(ceed, sgs_dd_train->dm_dd_training, domain_label, label_value, height, dm_field, &elem_restr_sgs_train)); 103 PetscCall(GetInverseMultiplicity(ceed, sgs_dd_train->dm_dd_training, domain_label, label_value, height, dm_field, PETSC_TRUE, 104 &elem_restr_inv_multiplicity, &inv_multiplicity)); 105 106 CeedElemRestriction elem_restr_filtered_state; 107 CeedInt num_comp_filtered_state; 108 { // -- Setup filtered velocity gradient projection 109 CeedBasis basis_filtered_state; 110 CeedOperatorField op_field; 111 PetscCallCeed(ceed, CeedOperatorGetFieldByName(sgs_dd_train->diff_filter->op_rhs_ctx->op, "v0", &op_field)); 112 PetscCallCeed(ceed, CeedOperatorFieldGetData(op_field, NULL, &elem_restr_filtered_state, &basis_filtered_state, NULL)); 113 PetscCallCeed(ceed, CeedElemRestrictionGetNumComponents(elem_restr_filtered_state, &num_comp_filtered_state)); 114 PetscCall(VelocityGradientProjectionSetup(ceed, honee, problem, STATEVAR_PRIMITIVE, elem_restr_filtered_state, basis_filtered_state, 115 &sgs_dd_train->filtered_grad_velo_proj)); 116 PetscCallCeed(ceed, CeedBasisDestroy(&basis_filtered_state)); 117 // Get velocity gradient information 118 PetscCallCeed(ceed, CeedOperatorGetFieldByName(sgs_dd_train->filtered_grad_velo_proj->l2_rhs_ctx->op, "velocity gradient", &op_field)); 119 PetscCallCeed(ceed, CeedOperatorFieldGetElemRestriction(op_field, &elem_restr_grad_velo)); 120 PetscCallCeed(ceed, CeedElemRestrictionGetNumComponents(elem_restr_grad_velo, &num_comp_grad_velo)); 121 } 122 123 CeedElemRestriction elem_restr_filtered_velo_prod; 124 CeedInt num_comp_filtered_velo_prod; 125 { // Get filtered velocity product information 126 CeedOperatorField op_field; 127 PetscCallCeed(ceed, CeedOperatorGetFieldByName(sgs_dd_train->diff_filter->op_rhs_ctx->op, "v1", &op_field)); 128 PetscCallCeed(ceed, CeedOperatorFieldGetElemRestriction(op_field, &elem_restr_filtered_velo_prod)); 129 PetscCallCeed(ceed, CeedElemRestrictionGetNumComponents(elem_restr_filtered_velo_prod, &num_comp_filtered_velo_prod)); 130 } 131 132 // -- Create operator for generating training data at nodes 133 // Differential Filter only provides filtered primitive variables 134 PetscCallCeed(ceed, CeedQFunctionCreateInterior(ceed, 1, ComputeSGS_DDAnisotropicTrainingDataNodal_Prim, 135 ComputeSGS_DDAnisotropicTrainingDataNodal_Prim_loc, &qf_sgs_dd_train)); 136 137 PetscCallCeed(ceed, CeedQFunctionSetContext(qf_sgs_dd_train, sgs_dd_train_setup_data->sgs_dd_train_qfctx)); 138 PetscCallCeed(ceed, CeedQFunctionAddInput(qf_sgs_dd_train, "q", num_comp_filtered_state, CEED_EVAL_NONE)); 139 PetscCallCeed(ceed, CeedQFunctionAddInput(qf_sgs_dd_train, "velocity product", num_comp_filtered_velo_prod, CEED_EVAL_NONE)); 140 PetscCallCeed(ceed, CeedQFunctionAddInput(qf_sgs_dd_train, "gradient velocity", num_comp_grad_velo, CEED_EVAL_NONE)); 141 PetscCallCeed(ceed, CeedQFunctionAddInput(qf_sgs_dd_train, "anisotropy tensor", num_comp_grid_aniso, CEED_EVAL_NONE)); 142 PetscCallCeed(ceed, CeedQFunctionAddInput(qf_sgs_dd_train, "inverse multiplicity", 1, CEED_EVAL_NONE)); 143 PetscCallCeed(ceed, CeedQFunctionAddOutput(qf_sgs_dd_train, "training data", sgs_dd_train->num_comp_dd_inputs, CEED_EVAL_NONE)); 144 145 PetscCallCeed(ceed, CeedElemRestrictionCreateVector(elem_restr_filtered_state, &filtered_fields, NULL)); 146 PetscCallCeed(ceed, CeedOperatorCreate(ceed, qf_sgs_dd_train, NULL, NULL, &op_sgs_dd_train)); 147 PetscCallCeed(ceed, CeedOperatorSetField(op_sgs_dd_train, "q", elem_restr_filtered_state, CEED_BASIS_NONE, filtered_fields)); 148 PetscCallCeed(ceed, CeedOperatorSetField(op_sgs_dd_train, "velocity product", elem_restr_filtered_velo_prod, CEED_BASIS_NONE, filtered_fields)); 149 PetscCallCeed(ceed, CeedOperatorSetField(op_sgs_dd_train, "gradient velocity", elem_restr_grad_velo, CEED_BASIS_NONE, CEED_VECTOR_ACTIVE)); 150 PetscCallCeed(ceed, CeedOperatorSetField(op_sgs_dd_train, "anisotropy tensor", sgs_dd_train_setup_data->elem_restr_grid_aniso, CEED_BASIS_NONE, 151 sgs_dd_train_setup_data->grid_aniso_ceed)); 152 PetscCallCeed(ceed, CeedOperatorSetField(op_sgs_dd_train, "inverse multiplicity", elem_restr_inv_multiplicity, CEED_BASIS_NONE, inv_multiplicity)); 153 PetscCallCeed(ceed, CeedOperatorSetField(op_sgs_dd_train, "training data", elem_restr_sgs_train, CEED_BASIS_NONE, CEED_VECTOR_ACTIVE)); 154 155 PetscCall(OperatorApplyContextCreate(sgs_dd_train->filtered_grad_velo_proj->dm, sgs_dd_train->dm_dd_training, ceed, op_sgs_dd_train, NULL, NULL, 156 NULL, NULL, &sgs_dd_train->op_training_data_calc_ctx)); 157 158 PetscCallCeed(ceed, CeedVectorDestroy(&inv_multiplicity)); 159 PetscCallCeed(ceed, CeedVectorDestroy(&filtered_fields)); 160 PetscCallCeed(ceed, CeedElemRestrictionDestroy(&elem_restr_inv_multiplicity)); 161 PetscCallCeed(ceed, CeedElemRestrictionDestroy(&elem_restr_filtered_state)); 162 PetscCallCeed(ceed, CeedElemRestrictionDestroy(&elem_restr_grad_velo)); 163 PetscCallCeed(ceed, CeedElemRestrictionDestroy(&elem_restr_filtered_velo_prod)); 164 PetscCallCeed(ceed, CeedQFunctionDestroy(&qf_sgs_dd_train)); 165 PetscCallCeed(ceed, CeedOperatorDestroy(&op_sgs_dd_train)); 166 PetscFunctionReturn(PETSC_SUCCESS); 167 } 168 169 PetscErrorCode SGS_DD_TrainingSetup(Ceed ceed, Honee honee, ProblemData problem) { 170 SGS_DDTrainingContext sgsdd_train_ctx; 171 SGS_DD_TrainingSetupData sgs_dd_train_setup_data; 172 SGS_DD_TrainingData sgs_dd_train; 173 174 PetscFunctionBeginUser; 175 176 PetscCall(PetscNew(&sgsdd_train_ctx)); 177 PetscCall(PetscNew(&sgs_dd_train_setup_data)); 178 PetscCall(PetscNew(&sgs_dd_train)); 179 PetscCall(HoneeSetContainer(honee, SGS_DD_TRAIN_KEY, sgs_dd_train, (PetscCtxDestroyFn *)SGS_DD_TrainingDataDestroy)); 180 PetscCall(DifferentialFilterSetup(honee, &sgs_dd_train->diff_filter)); 181 182 sgs_dd_train->overwrite_training_data = PETSC_TRUE; 183 sgs_dd_train->write_data_interval = 1; 184 sgs_dd_train->num_filter_widths = sizeof(sgs_dd_train->filter_widths) / sizeof(sgs_dd_train->filter_widths[0]); 185 PetscOptionsBegin(honee->comm, NULL, "SGS Data-Driven Training Options", NULL); 186 PetscCall(PetscOptionsInt("-sgs_train_write_data_interval", "Number of timesteps between writing data into database", NULL, 187 sgs_dd_train->write_data_interval, &sgs_dd_train->write_data_interval, NULL)); 188 PetscCall(PetscOptionsBool("-sgs_train_overwrite_data", "Overwrite old training data in the database", NULL, sgs_dd_train->overwrite_training_data, 189 &sgs_dd_train->overwrite_training_data, NULL)); 190 PetscCall(PetscOptionsRealArray("-sgs_train_filter_width_scales", "Scales of each filter width put into training database", NULL, 191 sgs_dd_train->filter_widths, &sgs_dd_train->num_filter_widths, NULL)); 192 PetscOptionsEnd(); 193 194 // -- Create DM for storing training data 195 PetscCall(SGS_DD_TrainingCreateDM(honee->dm, &sgs_dd_train->dm_dd_training, honee->app_ctx->degree, honee->app_ctx->q_extra, 196 &sgs_dd_train->num_comp_dd_inputs)); 197 198 { // -- Create QFunction Context 199 NewtonianIdealGasContext gas; 200 PetscCallCeed(ceed, CeedQFunctionContextGetDataRead(problem->apply_vol_ifunction.qfctx, CEED_MEM_HOST, &gas)); 201 sgsdd_train_ctx->gas = *gas; 202 PetscCallCeed(ceed, CeedQFunctionContextRestoreDataRead(problem->apply_vol_ifunction.qfctx, &gas)); 203 PetscCallCeed(ceed, CeedQFunctionContextCreate(honee->ceed, &sgs_dd_train_setup_data->sgs_dd_train_qfctx)); 204 PetscCallCeed(ceed, CeedQFunctionContextSetData(sgs_dd_train_setup_data->sgs_dd_train_qfctx, CEED_MEM_HOST, CEED_USE_POINTER, 205 sizeof(*sgsdd_train_ctx), sgsdd_train_ctx)); 206 PetscCallCeed(ceed, CeedQFunctionContextSetDataDestroy(sgs_dd_train_setup_data->sgs_dd_train_qfctx, CEED_MEM_HOST, FreeContextPetsc)); 207 } 208 209 { // -- Send training data array info to SmartRedis database 210 PetscMPIInt rank, num_ranks; 211 SmartSimData smartsim; 212 PetscCall(HoneeGetSmartSimData(honee, &smartsim)); 213 PetscCallMPI(MPI_Comm_rank(honee->comm, &rank)); 214 PetscCallMPI(MPI_Comm_size(honee->comm, &num_ranks)); 215 216 { 217 PetscSection global_section; 218 PetscInt num_dofs, num_comps, local_min_max[2] = {0.}, global_min_max[2] = {0.}; 219 220 PetscCall(DMGetGlobalSection(sgs_dd_train->dm_dd_training, &global_section)); 221 PetscCall(DMGetGlobalVectorInfo(sgs_dd_train->dm_dd_training, &num_dofs, NULL, NULL)); 222 PetscCall(PetscSectionGetFieldComponents(global_section, 0, &num_comps)); 223 local_min_max[0] = num_dofs; 224 PetscCall(PetscGlobalMinMaxInt(honee->comm, local_min_max, global_min_max)); 225 226 sgs_dd_train->training_data_array_dims[0] = global_min_max[0] / num_comps; 227 sgs_dd_train->training_data_array_dims[1] = num_comps; 228 } 229 230 if (rank % smartsim->collocated_database_num_ranks == 0) { 231 { // Communicate info on simulation size 232 const char tensor_name[] = "sizeInfo"; 233 size_t array_info_dim = 6; 234 PetscInt64 array_info[6] = {0}, num_features = 6; 235 236 array_info[0] = sgs_dd_train->training_data_array_dims[0]; 237 array_info[1] = sgs_dd_train->training_data_array_dims[1]; 238 array_info[2] = num_features; 239 array_info[3] = num_ranks; 240 array_info[4] = smartsim->collocated_database_num_ranks; 241 array_info[5] = rank; 242 243 PetscCall(PetscLogEventBegin(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 244 PetscCallSmartRedis( 245 put_tensor(smartsim->client, tensor_name, strlen(tensor_name), array_info, &array_info_dim, 1, SRTensorTypeInt64, SRMemLayoutContiguous)); 246 PetscCall(SmartRedisVerifyPutTensor(smartsim->client, tensor_name, strlen(tensor_name))); 247 PetscCall(PetscLogEventEnd(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 248 } 249 250 { // Send array that communicates if tensors are overwritten in database 251 const char tensor_name[] = "tensor-ow"; 252 PetscInt64 tensor_overwrite[2] = {sgs_dd_train->overwrite_training_data}; 253 size_t dim_2[1] = {2}; 254 255 PetscCall(PetscLogEventBegin(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 256 PetscCallSmartRedis( 257 put_tensor(smartsim->client, tensor_name, strlen(tensor_name), tensor_overwrite, dim_2, 1, SRTensorTypeInt64, SRMemLayoutContiguous)); 258 PetscCall(SmartRedisVerifyPutTensor(smartsim->client, tensor_name, strlen(tensor_name))); 259 PetscCall(PetscLogEventEnd(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 260 } 261 262 { // Communicate number of filter widths used 263 const char tensor_name[] = "num_filter_widths"; 264 PetscInt64 num_filter_widths = sgs_dd_train->num_filter_widths; 265 size_t dim_2 = 1; 266 267 PetscCall(PetscLogEventBegin(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 268 PetscCallSmartRedis( 269 put_tensor(smartsim->client, tensor_name, strlen(tensor_name), &num_filter_widths, &dim_2, 1, SRTensorTypeInt64, SRMemLayoutContiguous)); 270 PetscCall(SmartRedisVerifyPutTensor(smartsim->client, tensor_name, strlen(tensor_name))); 271 PetscCall(PetscLogEventEnd(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 272 } 273 } 274 } 275 276 // -- Compute and store anisotropy tensor 277 PetscCall(GridAnisotropyTensorProjectionSetupApply(ceed, honee, &sgs_dd_train_setup_data->elem_restr_grid_aniso, 278 &sgs_dd_train_setup_data->grid_aniso_ceed)); 279 280 // -- Create Nodal Evaluation Operator 281 PetscCall(SetupTrainingDataCalculation(ceed, honee, problem, sgs_dd_train_setup_data)); 282 283 PetscCall(SGS_DD_TrainingSetupDataDestroy(sgs_dd_train_setup_data)); 284 PetscFunctionReturn(PETSC_SUCCESS); 285 } 286 287 PetscErrorCode TSMonitor_SGS_DD_Training(TS ts, PetscInt step_num, PetscReal solution_time, Vec Q, void *ctx) { 288 Honee honee = (Honee)ctx; 289 Ceed ceed = honee->ceed; 290 SGS_DD_TrainingData sgs_dd_train; 291 SmartSimData smartsim; 292 Vec TrainingData; 293 PetscMPIInt rank; 294 295 PetscFunctionBeginUser; 296 PetscCall(HoneeGetSmartSimData(honee, &smartsim)); 297 PetscCall(HoneeGetContainer(honee, SGS_DD_TRAIN_KEY, &sgs_dd_train)); 298 PetscCallMPI(MPI_Comm_rank(honee->comm, &rank)); 299 300 if (step_num % sgs_dd_train->write_data_interval != 0) PetscFunctionReturn(PETSC_SUCCESS); 301 PetscCall(DMGetGlobalVector(sgs_dd_train->dm_dd_training, &TrainingData)); 302 303 for (PetscInt filter_index = 0; filter_index < sgs_dd_train->num_filter_widths; filter_index++) { 304 PetscCall(PetscLogEventBegin(HONEE_TrainDataCompute, 0, 0, 0, 0)); 305 { // -- Compute and assemble training data 306 Vec FilteredVelocityGradient, FilteredFields, FilteredFields_loc; 307 PetscMemType filtered_fields_mem_type; 308 CeedVector filtered_fields; 309 310 { // Set filter width for the current solve 311 double filter_width_scaling[3]; 312 CeedOperator op_mat; 313 Mat A_mat; 314 315 for (int j = 0; j < 3; j++) filter_width_scaling[j] = sgs_dd_train->filter_widths[filter_index]; 316 PetscCall(KSPGetOperators(sgs_dd_train->diff_filter->ksp, &A_mat, NULL)); 317 PetscCall(MatCeedGetCeedOperators(A_mat, &op_mat, NULL)); 318 PetscCall(CeedOperatorSetContextDouble(op_mat, sgs_dd_train->diff_filter->filter_width_scaling_label, filter_width_scaling)); 319 } 320 321 PetscCall(DMGetGlobalVector(sgs_dd_train->diff_filter->dm_filter, &FilteredFields)); 322 PetscCall(DMGetLocalVector(sgs_dd_train->diff_filter->dm_filter, &FilteredFields_loc)); 323 324 PetscCall(DifferentialFilterApply(honee, sgs_dd_train->diff_filter, solution_time, Q, FilteredFields)); 325 PetscCall(DMGlobalToLocal(sgs_dd_train->diff_filter->dm_filter, FilteredFields, INSERT_VALUES, FilteredFields_loc)); 326 327 PetscCall(DMGetGlobalVector(sgs_dd_train->filtered_grad_velo_proj->dm, &FilteredVelocityGradient)); 328 PetscCall(VelocityGradientProjectionApply(sgs_dd_train->filtered_grad_velo_proj, FilteredFields_loc, FilteredVelocityGradient)); 329 330 { 331 CeedOperatorField op_field; 332 333 PetscCallCeed(ceed, CeedOperatorGetFieldByName(sgs_dd_train->op_training_data_calc_ctx->op, "q", &op_field)); 334 PetscCallCeed(ceed, CeedOperatorFieldGetVector(op_field, &filtered_fields)); 335 } 336 337 PetscCall(VecPetscToCeed(FilteredFields_loc, &filtered_fields_mem_type, filtered_fields)); // filtered_fields is an implicit input 338 PetscCall(ApplyCeedOperatorGlobalToGlobal(FilteredVelocityGradient, TrainingData, sgs_dd_train->op_training_data_calc_ctx)); 339 PetscCall(VecCeedToPetsc(filtered_fields, filtered_fields_mem_type, FilteredFields_loc)); 340 341 PetscCall(DMRestoreGlobalVector(sgs_dd_train->filtered_grad_velo_proj->dm, &FilteredVelocityGradient)); 342 PetscCall(DMRestoreGlobalVector(sgs_dd_train->diff_filter->dm_filter, &FilteredFields)); 343 PetscCall(DMRestoreLocalVector(sgs_dd_train->diff_filter->dm_filter, &FilteredFields_loc)); 344 PetscCallCeed(ceed, CeedVectorDestroy(&filtered_fields)); 345 } 346 PetscCall(PetscLogEventEnd(HONEE_TrainDataCompute, 0, 0, 0, 0)); 347 348 { // -- Send training data to SmartSim 349 char array_key[PETSC_MAX_PATH_LEN]; 350 size_t array_key_len; 351 352 if (sgs_dd_train->overwrite_training_data) { 353 PetscCall(PetscSNPrintf(array_key, sizeof array_key, "%s.%" PetscInt_FMT, smartsim->rank_id_name, filter_index)); 354 } else { 355 PetscCall(PetscSNPrintf(array_key, sizeof array_key, "%s.%" PetscInt_FMT "%" PetscInt_FMT, smartsim->rank_id_name, step_num, filter_index)); 356 } 357 PetscCall(PetscStrlen(array_key, &array_key_len)); 358 359 { 360 const PetscScalar *training_data; 361 PetscCall(VecGetArrayRead(TrainingData, &training_data)); 362 PetscCall(PetscLogEventBegin(HONEE_SmartRedis_Train, 0, 0, 0, 0)); 363 PetscCallSmartRedis(put_tensor(smartsim->client, array_key, array_key_len, (void *)training_data, sgs_dd_train->training_data_array_dims, 2, 364 SRTensorTypeDouble, SRMemLayoutContiguous)); 365 PetscCall(PetscLogEventEnd(HONEE_SmartRedis_Train, 0, 0, 0, 0)); 366 PetscCall(VecRestoreArrayRead(TrainingData, &training_data)); 367 } 368 } 369 } 370 371 if (rank % smartsim->collocated_database_num_ranks == 0) { 372 const char tensor_name[] = "step"; 373 size_t dim_2[1] = {2}; 374 PetscInt64 step_array[2] = {step_num, step_num}; 375 376 PetscCall(PetscLogEventBegin(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 377 PetscCallSmartRedis( 378 put_tensor(smartsim->client, tensor_name, strlen(tensor_name), step_array, dim_2, 1, SRTensorTypeInt64, SRMemLayoutContiguous)); 379 PetscCall(PetscLogEventEnd(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 380 } 381 382 PetscCall(DMRestoreGlobalVector(sgs_dd_train->dm_dd_training, &TrainingData)); 383 PetscFunctionReturn(PETSC_SUCCESS); 384 } 385 386 PetscErrorCode TSPostStep_SGS_DD_Training(TS ts) { 387 Honee honee; 388 const char check_run_key[] = "check-run"; 389 PetscReal check_run[2] = {1}; 390 const size_t check_run_dims[1] = {2}; 391 size_t check_run_key_size; 392 SmartSimData smartsim; 393 394 PetscFunctionBeginUser; 395 PetscCall(PetscStrlen(check_run_key, &check_run_key_size)); 396 PetscCall(TSGetApplicationContext(ts, &honee)); 397 PetscCall(HoneeGetSmartSimData(honee, &smartsim)); 398 399 PetscCall(PetscLogEventBegin(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 400 PetscCallSmartRedis( 401 unpack_tensor(smartsim->client, check_run_key, check_run_key_size, check_run, check_run_dims, 1, SRTensorTypeDouble, SRMemLayoutContiguous)); 402 PetscCall(PetscLogEventEnd(HONEE_SmartRedis_Meta, 0, 0, 0, 0)); 403 if (check_run[0] == 0) { 404 PetscCall(PetscPrintf(honee->comm, "-- Simulation stopped by 'check-run' tensor in Redis database\n")); 405 PetscCall(TSSetConvergedReason(ts, TS_CONVERGED_USER)); 406 } 407 408 PetscFunctionReturn(PETSC_SUCCESS); 409 } 410