1 // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3 // 4 // SPDX-License-Identifier: BSD-2-Clause 5 // 6 // This file is part of CEED: http://github.com/ceed 7 8 #include <assert.h> 9 #include <ceed/backend.h> 10 #include <ceed/ceed.h> 11 #include <ceed/jit-tools.h> 12 #include <cuda.h> 13 #include <cuda_runtime.h> 14 #include <stdbool.h> 15 #include <string.h> 16 17 #include "../cuda/ceed-cuda-compile.h" 18 #include "ceed-cuda-ref.h" 19 20 //------------------------------------------------------------------------------ 21 // Destroy operator 22 //------------------------------------------------------------------------------ 23 static int CeedOperatorDestroy_Cuda(CeedOperator op) { 24 CeedOperator_Cuda *impl; 25 CeedCallBackend(CeedOperatorGetData(op, &impl)); 26 27 // Apply data 28 for (CeedInt i = 0; i < impl->numein + impl->numeout; i++) { 29 CeedCallBackend(CeedVectorDestroy(&impl->evecs[i])); 30 } 31 CeedCallBackend(CeedFree(&impl->evecs)); 32 33 for (CeedInt i = 0; i < impl->numein; i++) { 34 CeedCallBackend(CeedVectorDestroy(&impl->qvecsin[i])); 35 } 36 CeedCallBackend(CeedFree(&impl->qvecsin)); 37 38 for (CeedInt i = 0; i < impl->numeout; i++) { 39 CeedCallBackend(CeedVectorDestroy(&impl->qvecsout[i])); 40 } 41 CeedCallBackend(CeedFree(&impl->qvecsout)); 42 43 // QFunction assembly data 44 for (CeedInt i = 0; i < impl->qfnumactivein; i++) { 45 CeedCallBackend(CeedVectorDestroy(&impl->qfactivein[i])); 46 } 47 CeedCallBackend(CeedFree(&impl->qfactivein)); 48 49 // Diag data 50 if (impl->diag) { 51 Ceed ceed; 52 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 53 CeedCallCuda(ceed, cuModuleUnload(impl->diag->module)); 54 CeedCallBackend(CeedFree(&impl->diag->h_emodein)); 55 CeedCallBackend(CeedFree(&impl->diag->h_emodeout)); 56 CeedCallCuda(ceed, cudaFree(impl->diag->d_emodein)); 57 CeedCallCuda(ceed, cudaFree(impl->diag->d_emodeout)); 58 CeedCallCuda(ceed, cudaFree(impl->diag->d_identity)); 59 CeedCallCuda(ceed, cudaFree(impl->diag->d_interpin)); 60 CeedCallCuda(ceed, cudaFree(impl->diag->d_interpout)); 61 CeedCallCuda(ceed, cudaFree(impl->diag->d_gradin)); 62 CeedCallCuda(ceed, cudaFree(impl->diag->d_gradout)); 63 CeedCallBackend(CeedElemRestrictionDestroy(&impl->diag->pbdiagrstr)); 64 CeedCallBackend(CeedVectorDestroy(&impl->diag->elemdiag)); 65 CeedCallBackend(CeedVectorDestroy(&impl->diag->pbelemdiag)); 66 } 67 CeedCallBackend(CeedFree(&impl->diag)); 68 69 if (impl->asmb) { 70 Ceed ceed; 71 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 72 CeedCallCuda(ceed, cuModuleUnload(impl->asmb->module)); 73 CeedCallCuda(ceed, cudaFree(impl->asmb->d_B_in)); 74 CeedCallCuda(ceed, cudaFree(impl->asmb->d_B_out)); 75 } 76 CeedCallBackend(CeedFree(&impl->asmb)); 77 78 CeedCallBackend(CeedFree(&impl)); 79 return CEED_ERROR_SUCCESS; 80 } 81 82 //------------------------------------------------------------------------------ 83 // Setup infields or outfields 84 //------------------------------------------------------------------------------ 85 static int CeedOperatorSetupFields_Cuda(CeedQFunction qf, CeedOperator op, bool isinput, CeedVector *evecs, CeedVector *qvecs, CeedInt starte, 86 CeedInt numfields, CeedInt Q, CeedInt numelements) { 87 CeedInt dim, size; 88 CeedSize q_size; 89 Ceed ceed; 90 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 91 CeedBasis basis; 92 CeedElemRestriction Erestrict; 93 CeedOperatorField *opfields; 94 CeedQFunctionField *qffields; 95 CeedVector fieldvec; 96 bool strided; 97 bool skiprestrict; 98 99 if (isinput) { 100 CeedCallBackend(CeedOperatorGetFields(op, NULL, &opfields, NULL, NULL)); 101 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qffields, NULL, NULL)); 102 } else { 103 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &opfields)); 104 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qffields)); 105 } 106 107 // Loop over fields 108 for (CeedInt i = 0; i < numfields; i++) { 109 CeedEvalMode emode; 110 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qffields[i], &emode)); 111 112 strided = false; 113 skiprestrict = false; 114 if (emode != CEED_EVAL_WEIGHT) { 115 CeedCallBackend(CeedOperatorFieldGetElemRestriction(opfields[i], &Erestrict)); 116 117 // Check whether this field can skip the element restriction: 118 // must be passive input, with emode NONE, and have a strided restriction with CEED_STRIDES_BACKEND. 119 120 // First, check whether the field is input or output: 121 if (isinput) { 122 // Check for passive input: 123 CeedCallBackend(CeedOperatorFieldGetVector(opfields[i], &fieldvec)); 124 if (fieldvec != CEED_VECTOR_ACTIVE) { 125 // Check emode 126 if (emode == CEED_EVAL_NONE) { 127 // Check for strided restriction 128 CeedCallBackend(CeedElemRestrictionIsStrided(Erestrict, &strided)); 129 if (strided) { 130 // Check if vector is already in preferred backend ordering 131 CeedCallBackend(CeedElemRestrictionHasBackendStrides(Erestrict, &skiprestrict)); 132 } 133 } 134 } 135 } 136 if (skiprestrict) { 137 // We do not need an E-Vector, but will use the input field vector's data directly in the operator application. 138 evecs[i + starte] = NULL; 139 } else { 140 CeedCallBackend(CeedElemRestrictionCreateVector(Erestrict, NULL, &evecs[i + starte])); 141 } 142 } 143 144 switch (emode) { 145 case CEED_EVAL_NONE: 146 CeedCallBackend(CeedQFunctionFieldGetSize(qffields[i], &size)); 147 q_size = (CeedSize)numelements * Q * size; 148 CeedCallBackend(CeedVectorCreate(ceed, q_size, &qvecs[i])); 149 break; 150 case CEED_EVAL_INTERP: 151 CeedCallBackend(CeedQFunctionFieldGetSize(qffields[i], &size)); 152 q_size = (CeedSize)numelements * Q * size; 153 CeedCallBackend(CeedVectorCreate(ceed, q_size, &qvecs[i])); 154 break; 155 case CEED_EVAL_GRAD: 156 CeedCallBackend(CeedOperatorFieldGetBasis(opfields[i], &basis)); 157 CeedCallBackend(CeedQFunctionFieldGetSize(qffields[i], &size)); 158 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 159 q_size = (CeedSize)numelements * Q * size; 160 CeedCallBackend(CeedVectorCreate(ceed, q_size, &qvecs[i])); 161 break; 162 case CEED_EVAL_WEIGHT: // Only on input fields 163 CeedCallBackend(CeedOperatorFieldGetBasis(opfields[i], &basis)); 164 q_size = (CeedSize)numelements * Q; 165 CeedCallBackend(CeedVectorCreate(ceed, q_size, &qvecs[i])); 166 CeedCallBackend(CeedBasisApply(basis, numelements, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, NULL, qvecs[i])); 167 break; 168 case CEED_EVAL_DIV: 169 break; // TODO: Not implemented 170 case CEED_EVAL_CURL: 171 break; // TODO: Not implemented 172 } 173 } 174 return CEED_ERROR_SUCCESS; 175 } 176 177 //------------------------------------------------------------------------------ 178 // CeedOperator needs to connect all the named fields (be they active or passive) to the named inputs and outputs of its CeedQFunction. 179 //------------------------------------------------------------------------------ 180 static int CeedOperatorSetup_Cuda(CeedOperator op) { 181 bool setupdone; 182 CeedCallBackend(CeedOperatorIsSetupDone(op, &setupdone)); 183 if (setupdone) return CEED_ERROR_SUCCESS; 184 Ceed ceed; 185 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 186 CeedOperator_Cuda *impl; 187 CeedCallBackend(CeedOperatorGetData(op, &impl)); 188 CeedQFunction qf; 189 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 190 CeedInt Q, numelements, numinputfields, numoutputfields; 191 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 192 CeedCallBackend(CeedOperatorGetNumElements(op, &numelements)); 193 CeedOperatorField *opinputfields, *opoutputfields; 194 CeedCallBackend(CeedOperatorGetFields(op, &numinputfields, &opinputfields, &numoutputfields, &opoutputfields)); 195 CeedQFunctionField *qfinputfields, *qfoutputfields; 196 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qfinputfields, NULL, &qfoutputfields)); 197 198 // Allocate 199 CeedCallBackend(CeedCalloc(numinputfields + numoutputfields, &impl->evecs)); 200 201 CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->qvecsin)); 202 CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->qvecsout)); 203 204 impl->numein = numinputfields; 205 impl->numeout = numoutputfields; 206 207 // Set up infield and outfield evecs and qvecs 208 // Infields 209 CeedCallBackend(CeedOperatorSetupFields_Cuda(qf, op, true, impl->evecs, impl->qvecsin, 0, numinputfields, Q, numelements)); 210 211 // Outfields 212 CeedCallBackend(CeedOperatorSetupFields_Cuda(qf, op, false, impl->evecs, impl->qvecsout, numinputfields, numoutputfields, Q, numelements)); 213 214 CeedCallBackend(CeedOperatorSetSetupDone(op)); 215 return CEED_ERROR_SUCCESS; 216 } 217 218 //------------------------------------------------------------------------------ 219 // Setup Operator Inputs 220 //------------------------------------------------------------------------------ 221 static inline int CeedOperatorSetupInputs_Cuda(CeedInt numinputfields, CeedQFunctionField *qfinputfields, CeedOperatorField *opinputfields, 222 CeedVector invec, const bool skipactive, CeedScalar *edata[2 * CEED_FIELD_MAX], 223 CeedOperator_Cuda *impl, CeedRequest *request) { 224 CeedEvalMode emode; 225 CeedVector vec; 226 CeedElemRestriction Erestrict; 227 228 for (CeedInt i = 0; i < numinputfields; i++) { 229 // Get input vector 230 CeedCallBackend(CeedOperatorFieldGetVector(opinputfields[i], &vec)); 231 if (vec == CEED_VECTOR_ACTIVE) { 232 if (skipactive) continue; 233 else vec = invec; 234 } 235 236 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode)); 237 if (emode == CEED_EVAL_WEIGHT) { // Skip 238 } else { 239 // Get input vector 240 CeedCallBackend(CeedOperatorFieldGetVector(opinputfields[i], &vec)); 241 // Get input element restriction 242 CeedCallBackend(CeedOperatorFieldGetElemRestriction(opinputfields[i], &Erestrict)); 243 if (vec == CEED_VECTOR_ACTIVE) vec = invec; 244 // Restrict, if necessary 245 if (!impl->evecs[i]) { 246 // No restriction for this field; read data directly from vec. 247 CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, (const CeedScalar **)&edata[i])); 248 } else { 249 CeedCallBackend(CeedElemRestrictionApply(Erestrict, CEED_NOTRANSPOSE, vec, impl->evecs[i], request)); 250 // Get evec 251 CeedCallBackend(CeedVectorGetArrayRead(impl->evecs[i], CEED_MEM_DEVICE, (const CeedScalar **)&edata[i])); 252 } 253 } 254 } 255 return CEED_ERROR_SUCCESS; 256 } 257 258 //------------------------------------------------------------------------------ 259 // Input Basis Action 260 //------------------------------------------------------------------------------ 261 static inline int CeedOperatorInputBasis_Cuda(CeedInt numelements, CeedQFunctionField *qfinputfields, CeedOperatorField *opinputfields, 262 CeedInt numinputfields, const bool skipactive, CeedScalar *edata[2 * CEED_FIELD_MAX], 263 CeedOperator_Cuda *impl) { 264 CeedInt elemsize, size; 265 CeedElemRestriction Erestrict; 266 CeedEvalMode emode; 267 CeedBasis basis; 268 269 for (CeedInt i = 0; i < numinputfields; i++) { 270 // Skip active input 271 if (skipactive) { 272 CeedVector vec; 273 CeedCallBackend(CeedOperatorFieldGetVector(opinputfields[i], &vec)); 274 if (vec == CEED_VECTOR_ACTIVE) continue; 275 } 276 // Get elemsize, emode, size 277 CeedCallBackend(CeedOperatorFieldGetElemRestriction(opinputfields[i], &Erestrict)); 278 CeedCallBackend(CeedElemRestrictionGetElementSize(Erestrict, &elemsize)); 279 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode)); 280 CeedCallBackend(CeedQFunctionFieldGetSize(qfinputfields[i], &size)); 281 // Basis action 282 switch (emode) { 283 case CEED_EVAL_NONE: 284 CeedCallBackend(CeedVectorSetArray(impl->qvecsin[i], CEED_MEM_DEVICE, CEED_USE_POINTER, edata[i])); 285 break; 286 case CEED_EVAL_INTERP: 287 CeedCallBackend(CeedOperatorFieldGetBasis(opinputfields[i], &basis)); 288 CeedCallBackend(CeedBasisApply(basis, numelements, CEED_NOTRANSPOSE, CEED_EVAL_INTERP, impl->evecs[i], impl->qvecsin[i])); 289 break; 290 case CEED_EVAL_GRAD: 291 CeedCallBackend(CeedOperatorFieldGetBasis(opinputfields[i], &basis)); 292 CeedCallBackend(CeedBasisApply(basis, numelements, CEED_NOTRANSPOSE, CEED_EVAL_GRAD, impl->evecs[i], impl->qvecsin[i])); 293 break; 294 case CEED_EVAL_WEIGHT: 295 break; // No action 296 case CEED_EVAL_DIV: 297 break; // TODO: Not implemented 298 case CEED_EVAL_CURL: 299 break; // TODO: Not implemented 300 } 301 } 302 return CEED_ERROR_SUCCESS; 303 } 304 305 //------------------------------------------------------------------------------ 306 // Restore Input Vectors 307 //------------------------------------------------------------------------------ 308 static inline int CeedOperatorRestoreInputs_Cuda(CeedInt numinputfields, CeedQFunctionField *qfinputfields, CeedOperatorField *opinputfields, 309 const bool skipactive, CeedScalar *edata[2 * CEED_FIELD_MAX], CeedOperator_Cuda *impl) { 310 CeedEvalMode emode; 311 CeedVector vec; 312 313 for (CeedInt i = 0; i < numinputfields; i++) { 314 // Skip active input 315 if (skipactive) { 316 CeedCallBackend(CeedOperatorFieldGetVector(opinputfields[i], &vec)); 317 if (vec == CEED_VECTOR_ACTIVE) continue; 318 } 319 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode)); 320 if (emode == CEED_EVAL_WEIGHT) { // Skip 321 } else { 322 if (!impl->evecs[i]) { // This was a skiprestrict case 323 CeedCallBackend(CeedOperatorFieldGetVector(opinputfields[i], &vec)); 324 CeedCallBackend(CeedVectorRestoreArrayRead(vec, (const CeedScalar **)&edata[i])); 325 } else { 326 CeedCallBackend(CeedVectorRestoreArrayRead(impl->evecs[i], (const CeedScalar **)&edata[i])); 327 } 328 } 329 } 330 return CEED_ERROR_SUCCESS; 331 } 332 333 //------------------------------------------------------------------------------ 334 // Apply and add to output 335 //------------------------------------------------------------------------------ 336 static int CeedOperatorApplyAdd_Cuda(CeedOperator op, CeedVector invec, CeedVector outvec, CeedRequest *request) { 337 CeedOperator_Cuda *impl; 338 CeedCallBackend(CeedOperatorGetData(op, &impl)); 339 CeedQFunction qf; 340 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 341 CeedInt Q, numelements, elemsize, numinputfields, numoutputfields, size; 342 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 343 CeedCallBackend(CeedOperatorGetNumElements(op, &numelements)); 344 CeedOperatorField *opinputfields, *opoutputfields; 345 CeedCallBackend(CeedOperatorGetFields(op, &numinputfields, &opinputfields, &numoutputfields, &opoutputfields)); 346 CeedQFunctionField *qfinputfields, *qfoutputfields; 347 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qfinputfields, NULL, &qfoutputfields)); 348 CeedEvalMode emode; 349 CeedVector vec; 350 CeedBasis basis; 351 CeedElemRestriction Erestrict; 352 CeedScalar *edata[2 * CEED_FIELD_MAX] = {0}; 353 354 // Setup 355 CeedCallBackend(CeedOperatorSetup_Cuda(op)); 356 357 // Input Evecs and Restriction 358 CeedCallBackend(CeedOperatorSetupInputs_Cuda(numinputfields, qfinputfields, opinputfields, invec, false, edata, impl, request)); 359 360 // Input basis apply if needed 361 CeedCallBackend(CeedOperatorInputBasis_Cuda(numelements, qfinputfields, opinputfields, numinputfields, false, edata, impl)); 362 363 // Output pointers, as necessary 364 for (CeedInt i = 0; i < numoutputfields; i++) { 365 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode)); 366 if (emode == CEED_EVAL_NONE) { 367 // Set the output Q-Vector to use the E-Vector data directly. 368 CeedCallBackend(CeedVectorGetArrayWrite(impl->evecs[i + impl->numein], CEED_MEM_DEVICE, &edata[i + numinputfields])); 369 CeedCallBackend(CeedVectorSetArray(impl->qvecsout[i], CEED_MEM_DEVICE, CEED_USE_POINTER, edata[i + numinputfields])); 370 } 371 } 372 373 // Q function 374 CeedCallBackend(CeedQFunctionApply(qf, numelements * Q, impl->qvecsin, impl->qvecsout)); 375 376 // Output basis apply if needed 377 for (CeedInt i = 0; i < numoutputfields; i++) { 378 // Get elemsize, emode, size 379 CeedCallBackend(CeedOperatorFieldGetElemRestriction(opoutputfields[i], &Erestrict)); 380 CeedCallBackend(CeedElemRestrictionGetElementSize(Erestrict, &elemsize)); 381 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode)); 382 CeedCallBackend(CeedQFunctionFieldGetSize(qfoutputfields[i], &size)); 383 // Basis action 384 switch (emode) { 385 case CEED_EVAL_NONE: 386 break; 387 case CEED_EVAL_INTERP: 388 CeedCallBackend(CeedOperatorFieldGetBasis(opoutputfields[i], &basis)); 389 CeedCallBackend(CeedBasisApply(basis, numelements, CEED_TRANSPOSE, CEED_EVAL_INTERP, impl->qvecsout[i], impl->evecs[i + impl->numein])); 390 break; 391 case CEED_EVAL_GRAD: 392 CeedCallBackend(CeedOperatorFieldGetBasis(opoutputfields[i], &basis)); 393 CeedCallBackend(CeedBasisApply(basis, numelements, CEED_TRANSPOSE, CEED_EVAL_GRAD, impl->qvecsout[i], impl->evecs[i + impl->numein])); 394 break; 395 // LCOV_EXCL_START 396 case CEED_EVAL_WEIGHT: { 397 Ceed ceed; 398 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 399 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 400 break; // Should not occur 401 } 402 case CEED_EVAL_DIV: 403 break; // TODO: Not implemented 404 case CEED_EVAL_CURL: 405 break; // TODO: Not implemented 406 // LCOV_EXCL_STOP 407 } 408 } 409 410 // Output restriction 411 for (CeedInt i = 0; i < numoutputfields; i++) { 412 // Restore evec 413 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode)); 414 if (emode == CEED_EVAL_NONE) { 415 CeedCallBackend(CeedVectorRestoreArray(impl->evecs[i + impl->numein], &edata[i + numinputfields])); 416 } 417 // Get output vector 418 CeedCallBackend(CeedOperatorFieldGetVector(opoutputfields[i], &vec)); 419 // Restrict 420 CeedCallBackend(CeedOperatorFieldGetElemRestriction(opoutputfields[i], &Erestrict)); 421 // Active 422 if (vec == CEED_VECTOR_ACTIVE) vec = outvec; 423 424 CeedCallBackend(CeedElemRestrictionApply(Erestrict, CEED_TRANSPOSE, impl->evecs[i + impl->numein], vec, request)); 425 } 426 427 // Restore input arrays 428 CeedCallBackend(CeedOperatorRestoreInputs_Cuda(numinputfields, qfinputfields, opinputfields, false, edata, impl)); 429 return CEED_ERROR_SUCCESS; 430 } 431 432 //------------------------------------------------------------------------------ 433 // Core code for assembling linear QFunction 434 //------------------------------------------------------------------------------ 435 static inline int CeedOperatorLinearAssembleQFunctionCore_Cuda(CeedOperator op, bool build_objects, CeedVector *assembled, CeedElemRestriction *rstr, 436 CeedRequest *request) { 437 CeedOperator_Cuda *impl; 438 CeedCallBackend(CeedOperatorGetData(op, &impl)); 439 CeedQFunction qf; 440 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 441 CeedInt Q, numelements, numinputfields, numoutputfields, size; 442 CeedSize q_size; 443 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 444 CeedCallBackend(CeedOperatorGetNumElements(op, &numelements)); 445 CeedOperatorField *opinputfields, *opoutputfields; 446 CeedCallBackend(CeedOperatorGetFields(op, &numinputfields, &opinputfields, &numoutputfields, &opoutputfields)); 447 CeedQFunctionField *qfinputfields, *qfoutputfields; 448 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qfinputfields, NULL, &qfoutputfields)); 449 CeedVector vec; 450 CeedInt numactivein = impl->qfnumactivein, numactiveout = impl->qfnumactiveout; 451 CeedVector *activein = impl->qfactivein; 452 CeedScalar *a, *tmp; 453 Ceed ceed, ceedparent; 454 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 455 CeedCallBackend(CeedGetOperatorFallbackParentCeed(ceed, &ceedparent)); 456 ceedparent = ceedparent ? ceedparent : ceed; 457 CeedScalar *edata[2 * CEED_FIELD_MAX]; 458 459 // Setup 460 CeedCallBackend(CeedOperatorSetup_Cuda(op)); 461 462 // Check for identity 463 bool identityqf; 464 CeedCallBackend(CeedQFunctionIsIdentity(qf, &identityqf)); 465 if (identityqf) { 466 // LCOV_EXCL_START 467 return CeedError(ceed, CEED_ERROR_BACKEND, "Assembling identity QFunctions not supported"); 468 // LCOV_EXCL_STOP 469 } 470 471 // Input Evecs and Restriction 472 CeedCallBackend(CeedOperatorSetupInputs_Cuda(numinputfields, qfinputfields, opinputfields, NULL, true, edata, impl, request)); 473 474 // Count number of active input fields 475 if (!numactivein) { 476 for (CeedInt i = 0; i < numinputfields; i++) { 477 // Get input vector 478 CeedCallBackend(CeedOperatorFieldGetVector(opinputfields[i], &vec)); 479 // Check if active input 480 if (vec == CEED_VECTOR_ACTIVE) { 481 CeedCallBackend(CeedQFunctionFieldGetSize(qfinputfields[i], &size)); 482 CeedCallBackend(CeedVectorSetValue(impl->qvecsin[i], 0.0)); 483 CeedCallBackend(CeedVectorGetArray(impl->qvecsin[i], CEED_MEM_DEVICE, &tmp)); 484 CeedCallBackend(CeedRealloc(numactivein + size, &activein)); 485 for (CeedInt field = 0; field < size; field++) { 486 q_size = (CeedSize)Q * numelements; 487 CeedCallBackend(CeedVectorCreate(ceed, q_size, &activein[numactivein + field])); 488 CeedCallBackend(CeedVectorSetArray(activein[numactivein + field], CEED_MEM_DEVICE, CEED_USE_POINTER, &tmp[field * Q * numelements])); 489 } 490 numactivein += size; 491 CeedCallBackend(CeedVectorRestoreArray(impl->qvecsin[i], &tmp)); 492 } 493 } 494 impl->qfnumactivein = numactivein; 495 impl->qfactivein = activein; 496 } 497 498 // Count number of active output fields 499 if (!numactiveout) { 500 for (CeedInt i = 0; i < numoutputfields; i++) { 501 // Get output vector 502 CeedCallBackend(CeedOperatorFieldGetVector(opoutputfields[i], &vec)); 503 // Check if active output 504 if (vec == CEED_VECTOR_ACTIVE) { 505 CeedCallBackend(CeedQFunctionFieldGetSize(qfoutputfields[i], &size)); 506 numactiveout += size; 507 } 508 } 509 impl->qfnumactiveout = numactiveout; 510 } 511 512 // Check sizes 513 if (!numactivein || !numactiveout) { 514 // LCOV_EXCL_START 515 return CeedError(ceed, CEED_ERROR_BACKEND, "Cannot assemble QFunction without active inputs and outputs"); 516 // LCOV_EXCL_STOP 517 } 518 519 // Build objects if needed 520 if (build_objects) { 521 // Create output restriction 522 CeedInt strides[3] = {1, numelements * Q, Q}; /* *NOPAD* */ 523 CeedCallBackend(CeedElemRestrictionCreateStrided(ceedparent, numelements, Q, numactivein * numactiveout, 524 numactivein * numactiveout * numelements * Q, strides, rstr)); 525 // Create assembled vector 526 CeedSize l_size = (CeedSize)numelements * Q * numactivein * numactiveout; 527 CeedCallBackend(CeedVectorCreate(ceedparent, l_size, assembled)); 528 } 529 CeedCallBackend(CeedVectorSetValue(*assembled, 0.0)); 530 CeedCallBackend(CeedVectorGetArray(*assembled, CEED_MEM_DEVICE, &a)); 531 532 // Input basis apply 533 CeedCallBackend(CeedOperatorInputBasis_Cuda(numelements, qfinputfields, opinputfields, numinputfields, true, edata, impl)); 534 535 // Assemble QFunction 536 for (CeedInt in = 0; in < numactivein; in++) { 537 // Set Inputs 538 CeedCallBackend(CeedVectorSetValue(activein[in], 1.0)); 539 if (numactivein > 1) { 540 CeedCallBackend(CeedVectorSetValue(activein[(in + numactivein - 1) % numactivein], 0.0)); 541 } 542 // Set Outputs 543 for (CeedInt out = 0; out < numoutputfields; out++) { 544 // Get output vector 545 CeedCallBackend(CeedOperatorFieldGetVector(opoutputfields[out], &vec)); 546 // Check if active output 547 if (vec == CEED_VECTOR_ACTIVE) { 548 CeedCallBackend(CeedVectorSetArray(impl->qvecsout[out], CEED_MEM_DEVICE, CEED_USE_POINTER, a)); 549 CeedCallBackend(CeedQFunctionFieldGetSize(qfoutputfields[out], &size)); 550 a += size * Q * numelements; // Advance the pointer by the size of the output 551 } 552 } 553 // Apply QFunction 554 CeedCallBackend(CeedQFunctionApply(qf, Q * numelements, impl->qvecsin, impl->qvecsout)); 555 } 556 557 // Un-set output Qvecs to prevent accidental overwrite of Assembled 558 for (CeedInt out = 0; out < numoutputfields; out++) { 559 // Get output vector 560 CeedCallBackend(CeedOperatorFieldGetVector(opoutputfields[out], &vec)); 561 // Check if active output 562 if (vec == CEED_VECTOR_ACTIVE) { 563 CeedCallBackend(CeedVectorTakeArray(impl->qvecsout[out], CEED_MEM_DEVICE, NULL)); 564 } 565 } 566 567 // Restore input arrays 568 CeedCallBackend(CeedOperatorRestoreInputs_Cuda(numinputfields, qfinputfields, opinputfields, true, edata, impl)); 569 570 // Restore output 571 CeedCallBackend(CeedVectorRestoreArray(*assembled, &a)); 572 573 return CEED_ERROR_SUCCESS; 574 } 575 576 //------------------------------------------------------------------------------ 577 // Assemble Linear QFunction 578 //------------------------------------------------------------------------------ 579 static int CeedOperatorLinearAssembleQFunction_Cuda(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 580 return CeedOperatorLinearAssembleQFunctionCore_Cuda(op, true, assembled, rstr, request); 581 } 582 583 //------------------------------------------------------------------------------ 584 // Update Assembled Linear QFunction 585 //------------------------------------------------------------------------------ 586 static int CeedOperatorLinearAssembleQFunctionUpdate_Cuda(CeedOperator op, CeedVector assembled, CeedElemRestriction rstr, CeedRequest *request) { 587 return CeedOperatorLinearAssembleQFunctionCore_Cuda(op, false, &assembled, &rstr, request); 588 } 589 590 //------------------------------------------------------------------------------ 591 // Create point block restriction 592 //------------------------------------------------------------------------------ 593 static int CreatePBRestriction(CeedElemRestriction rstr, CeedElemRestriction *pbRstr) { 594 Ceed ceed; 595 CeedCallBackend(CeedElemRestrictionGetCeed(rstr, &ceed)); 596 const CeedInt *offsets; 597 CeedCallBackend(CeedElemRestrictionGetOffsets(rstr, CEED_MEM_HOST, &offsets)); 598 599 // Expand offsets 600 CeedInt nelem, ncomp, elemsize, compstride, *pbOffsets; 601 CeedSize l_size; 602 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr, &nelem)); 603 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr, &ncomp)); 604 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr, &elemsize)); 605 CeedCallBackend(CeedElemRestrictionGetCompStride(rstr, &compstride)); 606 CeedCallBackend(CeedElemRestrictionGetLVectorSize(rstr, &l_size)); 607 CeedInt shift = ncomp; 608 if (compstride != 1) shift *= ncomp; 609 CeedCallBackend(CeedCalloc(nelem * elemsize, &pbOffsets)); 610 for (CeedInt i = 0; i < nelem * elemsize; i++) { 611 pbOffsets[i] = offsets[i] * shift; 612 } 613 614 // Create new restriction 615 CeedCallBackend( 616 CeedElemRestrictionCreate(ceed, nelem, elemsize, ncomp * ncomp, 1, l_size * ncomp, CEED_MEM_HOST, CEED_OWN_POINTER, pbOffsets, pbRstr)); 617 618 // Cleanup 619 CeedCallBackend(CeedElemRestrictionRestoreOffsets(rstr, &offsets)); 620 621 return CEED_ERROR_SUCCESS; 622 } 623 624 //------------------------------------------------------------------------------ 625 // Assemble diagonal setup 626 //------------------------------------------------------------------------------ 627 static inline int CeedOperatorAssembleDiagonalSetup_Cuda(CeedOperator op, const bool pointBlock) { 628 Ceed ceed; 629 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 630 CeedQFunction qf; 631 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 632 CeedInt numinputfields, numoutputfields; 633 CeedCallBackend(CeedQFunctionGetNumArgs(qf, &numinputfields, &numoutputfields)); 634 635 // Determine active input basis 636 CeedOperatorField *opfields; 637 CeedQFunctionField *qffields; 638 CeedCallBackend(CeedOperatorGetFields(op, NULL, &opfields, NULL, NULL)); 639 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qffields, NULL, NULL)); 640 CeedInt numemodein = 0, ncomp = 0, dim = 1; 641 CeedEvalMode *emodein = NULL; 642 CeedBasis basisin = NULL; 643 CeedElemRestriction rstrin = NULL; 644 for (CeedInt i = 0; i < numinputfields; i++) { 645 CeedVector vec; 646 CeedCallBackend(CeedOperatorFieldGetVector(opfields[i], &vec)); 647 if (vec == CEED_VECTOR_ACTIVE) { 648 CeedElemRestriction rstr; 649 CeedCallBackend(CeedOperatorFieldGetBasis(opfields[i], &basisin)); 650 CeedCallBackend(CeedBasisGetNumComponents(basisin, &ncomp)); 651 CeedCallBackend(CeedBasisGetDimension(basisin, &dim)); 652 CeedCallBackend(CeedOperatorFieldGetElemRestriction(opfields[i], &rstr)); 653 if (rstrin && rstrin != rstr) { 654 // LCOV_EXCL_START 655 return CeedError(ceed, CEED_ERROR_BACKEND, "Backend does not implement multi-field non-composite operator diagonal assembly"); 656 // LCOV_EXCL_STOP 657 } 658 rstrin = rstr; 659 CeedEvalMode emode; 660 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qffields[i], &emode)); 661 switch (emode) { 662 case CEED_EVAL_NONE: 663 case CEED_EVAL_INTERP: 664 CeedCallBackend(CeedRealloc(numemodein + 1, &emodein)); 665 emodein[numemodein] = emode; 666 numemodein += 1; 667 break; 668 case CEED_EVAL_GRAD: 669 CeedCallBackend(CeedRealloc(numemodein + dim, &emodein)); 670 for (CeedInt d = 0; d < dim; d++) emodein[numemodein + d] = emode; 671 numemodein += dim; 672 break; 673 case CEED_EVAL_WEIGHT: 674 case CEED_EVAL_DIV: 675 case CEED_EVAL_CURL: 676 break; // Caught by QF Assembly 677 } 678 } 679 } 680 681 // Determine active output basis 682 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &opfields)); 683 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qffields)); 684 CeedInt numemodeout = 0; 685 CeedEvalMode *emodeout = NULL; 686 CeedBasis basisout = NULL; 687 CeedElemRestriction rstrout = NULL; 688 for (CeedInt i = 0; i < numoutputfields; i++) { 689 CeedVector vec; 690 CeedCallBackend(CeedOperatorFieldGetVector(opfields[i], &vec)); 691 if (vec == CEED_VECTOR_ACTIVE) { 692 CeedElemRestriction rstr; 693 CeedCallBackend(CeedOperatorFieldGetBasis(opfields[i], &basisout)); 694 CeedCallBackend(CeedOperatorFieldGetElemRestriction(opfields[i], &rstr)); 695 if (rstrout && rstrout != rstr) { 696 // LCOV_EXCL_START 697 return CeedError(ceed, CEED_ERROR_BACKEND, "Backend does not implement multi-field non-composite operator diagonal assembly"); 698 // LCOV_EXCL_STOP 699 } 700 rstrout = rstr; 701 CeedEvalMode emode; 702 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qffields[i], &emode)); 703 switch (emode) { 704 case CEED_EVAL_NONE: 705 case CEED_EVAL_INTERP: 706 CeedCallBackend(CeedRealloc(numemodeout + 1, &emodeout)); 707 emodeout[numemodeout] = emode; 708 numemodeout += 1; 709 break; 710 case CEED_EVAL_GRAD: 711 CeedCallBackend(CeedRealloc(numemodeout + dim, &emodeout)); 712 for (CeedInt d = 0; d < dim; d++) emodeout[numemodeout + d] = emode; 713 numemodeout += dim; 714 break; 715 case CEED_EVAL_WEIGHT: 716 case CEED_EVAL_DIV: 717 case CEED_EVAL_CURL: 718 break; // Caught by QF Assembly 719 } 720 } 721 } 722 723 // Operator data struct 724 CeedOperator_Cuda *impl; 725 CeedCallBackend(CeedOperatorGetData(op, &impl)); 726 CeedCallBackend(CeedCalloc(1, &impl->diag)); 727 CeedOperatorDiag_Cuda *diag = impl->diag; 728 diag->basisin = basisin; 729 diag->basisout = basisout; 730 diag->h_emodein = emodein; 731 diag->h_emodeout = emodeout; 732 diag->numemodein = numemodein; 733 diag->numemodeout = numemodeout; 734 735 // Assemble kernel 736 char *diagonal_kernel_path, *diagonal_kernel_source; 737 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-operator-assemble-diagonal.h", &diagonal_kernel_path)); 738 CeedDebug256(ceed, 2, "----- Loading Diagonal Assembly Kernel Source -----\n"); 739 CeedCallBackend(CeedLoadSourceToBuffer(ceed, diagonal_kernel_path, &diagonal_kernel_source)); 740 CeedDebug256(ceed, 2, "----- Loading Diagonal Assembly Source Complete! -----\n"); 741 CeedInt nnodes, nqpts; 742 CeedCallBackend(CeedBasisGetNumNodes(basisin, &nnodes)); 743 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basisin, &nqpts)); 744 diag->nnodes = nnodes; 745 CeedCallCuda(ceed, CeedCompileCuda(ceed, diagonal_kernel_source, &diag->module, 5, "NUMEMODEIN", numemodein, "NUMEMODEOUT", numemodeout, "NNODES", 746 nnodes, "NQPTS", nqpts, "NCOMP", ncomp)); 747 CeedCallCuda(ceed, CeedGetKernelCuda(ceed, diag->module, "linearDiagonal", &diag->linearDiagonal)); 748 CeedCallCuda(ceed, CeedGetKernelCuda(ceed, diag->module, "linearPointBlockDiagonal", &diag->linearPointBlock)); 749 CeedCallBackend(CeedFree(&diagonal_kernel_path)); 750 CeedCallBackend(CeedFree(&diagonal_kernel_source)); 751 752 // Basis matrices 753 const CeedInt qBytes = nqpts * sizeof(CeedScalar); 754 const CeedInt iBytes = qBytes * nnodes; 755 const CeedInt gBytes = qBytes * nnodes * dim; 756 const CeedInt eBytes = sizeof(CeedEvalMode); 757 const CeedScalar *interpin, *interpout, *gradin, *gradout; 758 759 // CEED_EVAL_NONE 760 CeedScalar *identity = NULL; 761 bool evalNone = false; 762 for (CeedInt i = 0; i < numemodein; i++) evalNone = evalNone || (emodein[i] == CEED_EVAL_NONE); 763 for (CeedInt i = 0; i < numemodeout; i++) evalNone = evalNone || (emodeout[i] == CEED_EVAL_NONE); 764 if (evalNone) { 765 CeedCallBackend(CeedCalloc(nqpts * nnodes, &identity)); 766 for (CeedInt i = 0; i < (nnodes < nqpts ? nnodes : nqpts); i++) identity[i * nnodes + i] = 1.0; 767 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_identity, iBytes)); 768 CeedCallCuda(ceed, cudaMemcpy(diag->d_identity, identity, iBytes, cudaMemcpyHostToDevice)); 769 } 770 771 // CEED_EVAL_INTERP 772 CeedCallBackend(CeedBasisGetInterp(basisin, &interpin)); 773 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_interpin, iBytes)); 774 CeedCallCuda(ceed, cudaMemcpy(diag->d_interpin, interpin, iBytes, cudaMemcpyHostToDevice)); 775 CeedCallBackend(CeedBasisGetInterp(basisout, &interpout)); 776 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_interpout, iBytes)); 777 CeedCallCuda(ceed, cudaMemcpy(diag->d_interpout, interpout, iBytes, cudaMemcpyHostToDevice)); 778 779 // CEED_EVAL_GRAD 780 CeedCallBackend(CeedBasisGetGrad(basisin, &gradin)); 781 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_gradin, gBytes)); 782 CeedCallCuda(ceed, cudaMemcpy(diag->d_gradin, gradin, gBytes, cudaMemcpyHostToDevice)); 783 CeedCallBackend(CeedBasisGetGrad(basisout, &gradout)); 784 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_gradout, gBytes)); 785 CeedCallCuda(ceed, cudaMemcpy(diag->d_gradout, gradout, gBytes, cudaMemcpyHostToDevice)); 786 787 // Arrays of emodes 788 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_emodein, numemodein * eBytes)); 789 CeedCallCuda(ceed, cudaMemcpy(diag->d_emodein, emodein, numemodein * eBytes, cudaMemcpyHostToDevice)); 790 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_emodeout, numemodeout * eBytes)); 791 CeedCallCuda(ceed, cudaMemcpy(diag->d_emodeout, emodeout, numemodeout * eBytes, cudaMemcpyHostToDevice)); 792 793 // Restriction 794 diag->diagrstr = rstrout; 795 796 return CEED_ERROR_SUCCESS; 797 } 798 799 //------------------------------------------------------------------------------ 800 // Assemble diagonal common code 801 //------------------------------------------------------------------------------ 802 static inline int CeedOperatorAssembleDiagonalCore_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request, const bool pointBlock) { 803 Ceed ceed; 804 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 805 CeedOperator_Cuda *impl; 806 CeedCallBackend(CeedOperatorGetData(op, &impl)); 807 808 // Assemble QFunction 809 CeedVector assembledqf; 810 CeedElemRestriction rstr; 811 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembledqf, &rstr, request)); 812 CeedCallBackend(CeedElemRestrictionDestroy(&rstr)); 813 814 // Setup 815 if (!impl->diag) { 816 CeedCallBackend(CeedOperatorAssembleDiagonalSetup_Cuda(op, pointBlock)); 817 } 818 CeedOperatorDiag_Cuda *diag = impl->diag; 819 assert(diag != NULL); 820 821 // Restriction 822 if (pointBlock && !diag->pbdiagrstr) { 823 CeedElemRestriction pbdiagrstr; 824 CeedCallBackend(CreatePBRestriction(diag->diagrstr, &pbdiagrstr)); 825 diag->pbdiagrstr = pbdiagrstr; 826 } 827 CeedElemRestriction diagrstr = pointBlock ? diag->pbdiagrstr : diag->diagrstr; 828 829 // Create diagonal vector 830 CeedVector elemdiag = pointBlock ? diag->pbelemdiag : diag->elemdiag; 831 if (!elemdiag) { 832 CeedCallBackend(CeedElemRestrictionCreateVector(diagrstr, NULL, &elemdiag)); 833 if (pointBlock) diag->pbelemdiag = elemdiag; 834 else diag->elemdiag = elemdiag; 835 } 836 CeedCallBackend(CeedVectorSetValue(elemdiag, 0.0)); 837 838 // Assemble element operator diagonals 839 CeedScalar *elemdiagarray; 840 const CeedScalar *assembledqfarray; 841 CeedCallBackend(CeedVectorGetArray(elemdiag, CEED_MEM_DEVICE, &elemdiagarray)); 842 CeedCallBackend(CeedVectorGetArrayRead(assembledqf, CEED_MEM_DEVICE, &assembledqfarray)); 843 CeedInt nelem; 844 CeedCallBackend(CeedElemRestrictionGetNumElements(diagrstr, &nelem)); 845 846 // Compute the diagonal of B^T D B 847 int elemsPerBlock = 1; 848 int grid = nelem / elemsPerBlock + ((nelem / elemsPerBlock * elemsPerBlock < nelem) ? 1 : 0); 849 void *args[] = {(void *)&nelem, &diag->d_identity, &diag->d_interpin, &diag->d_gradin, &diag->d_interpout, 850 &diag->d_gradout, &diag->d_emodein, &diag->d_emodeout, &assembledqfarray, &elemdiagarray}; 851 if (pointBlock) { 852 CeedCallBackend(CeedRunKernelDimCuda(ceed, diag->linearPointBlock, grid, diag->nnodes, 1, elemsPerBlock, args)); 853 } else { 854 CeedCallBackend(CeedRunKernelDimCuda(ceed, diag->linearDiagonal, grid, diag->nnodes, 1, elemsPerBlock, args)); 855 } 856 857 // Restore arrays 858 CeedCallBackend(CeedVectorRestoreArray(elemdiag, &elemdiagarray)); 859 CeedCallBackend(CeedVectorRestoreArrayRead(assembledqf, &assembledqfarray)); 860 861 // Assemble local operator diagonal 862 CeedCallBackend(CeedElemRestrictionApply(diagrstr, CEED_TRANSPOSE, elemdiag, assembled, request)); 863 864 // Cleanup 865 CeedCallBackend(CeedVectorDestroy(&assembledqf)); 866 867 return CEED_ERROR_SUCCESS; 868 } 869 870 //------------------------------------------------------------------------------ 871 // Assemble Linear Diagonal 872 //------------------------------------------------------------------------------ 873 static int CeedOperatorLinearAssembleAddDiagonal_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request) { 874 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Cuda(op, assembled, request, false)); 875 return CEED_ERROR_SUCCESS; 876 } 877 878 //------------------------------------------------------------------------------ 879 // Assemble Linear Point Block Diagonal 880 //------------------------------------------------------------------------------ 881 static int CeedOperatorLinearAssembleAddPointBlockDiagonal_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request) { 882 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Cuda(op, assembled, request, true)); 883 return CEED_ERROR_SUCCESS; 884 } 885 886 //------------------------------------------------------------------------------ 887 // Single operator assembly setup 888 //------------------------------------------------------------------------------ 889 static int CeedSingleOperatorAssembleSetup_Cuda(CeedOperator op) { 890 Ceed ceed; 891 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 892 CeedOperator_Cuda *impl; 893 CeedCallBackend(CeedOperatorGetData(op, &impl)); 894 895 // Get intput and output fields 896 CeedInt num_input_fields, num_output_fields; 897 CeedOperatorField *input_fields; 898 CeedOperatorField *output_fields; 899 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 900 901 // Determine active input basis eval mode 902 CeedQFunction qf; 903 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 904 CeedQFunctionField *qf_fields; 905 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 906 // Note that the kernel will treat each dimension of a gradient action separately; 907 // i.e., when an active input has a CEED_EVAL_GRAD mode, num_emode_in will increment by dim. 908 // However, for the purposes of loading the B matrices, it will be treated as one mode, and we will load/copy the entire gradient matrix at once, so 909 // num_B_in_mats_to_load will be incremented by 1. 910 CeedInt num_emode_in = 0, dim = 1, num_B_in_mats_to_load = 0, size_B_in = 0; 911 CeedEvalMode *eval_mode_in = NULL; // will be of size num_B_in_mats_load 912 CeedBasis basis_in = NULL; 913 CeedInt nqpts = 0, esize = 0; 914 CeedElemRestriction rstr_in = NULL; 915 for (CeedInt i = 0; i < num_input_fields; i++) { 916 CeedVector vec; 917 CeedCallBackend(CeedOperatorFieldGetVector(input_fields[i], &vec)); 918 if (vec == CEED_VECTOR_ACTIVE) { 919 CeedCallBackend(CeedOperatorFieldGetBasis(input_fields[i], &basis_in)); 920 CeedCallBackend(CeedBasisGetDimension(basis_in, &dim)); 921 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &nqpts)); 922 CeedCallBackend(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr_in)); 923 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_in, &esize)); 924 CeedEvalMode eval_mode; 925 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 926 if (eval_mode != CEED_EVAL_NONE) { 927 CeedCallBackend(CeedRealloc(num_B_in_mats_to_load + 1, &eval_mode_in)); 928 eval_mode_in[num_B_in_mats_to_load] = eval_mode; 929 num_B_in_mats_to_load += 1; 930 if (eval_mode == CEED_EVAL_GRAD) { 931 num_emode_in += dim; 932 size_B_in += dim * esize * nqpts; 933 } else { 934 num_emode_in += 1; 935 size_B_in += esize * nqpts; 936 } 937 } 938 } 939 } 940 941 // Determine active output basis; basis_out and rstr_out only used if same as input, TODO 942 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 943 CeedInt num_emode_out = 0, num_B_out_mats_to_load = 0, size_B_out = 0; 944 CeedEvalMode *eval_mode_out = NULL; 945 CeedBasis basis_out = NULL; 946 CeedElemRestriction rstr_out = NULL; 947 for (CeedInt i = 0; i < num_output_fields; i++) { 948 CeedVector vec; 949 CeedCallBackend(CeedOperatorFieldGetVector(output_fields[i], &vec)); 950 if (vec == CEED_VECTOR_ACTIVE) { 951 CeedCallBackend(CeedOperatorFieldGetBasis(output_fields[i], &basis_out)); 952 CeedCallBackend(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr_out)); 953 if (rstr_out && rstr_out != rstr_in) { 954 // LCOV_EXCL_START 955 return CeedError(ceed, CEED_ERROR_BACKEND, "Backend does not implement multi-field non-composite operator assembly"); 956 // LCOV_EXCL_STOP 957 } 958 CeedEvalMode eval_mode; 959 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 960 if (eval_mode != CEED_EVAL_NONE) { 961 CeedCallBackend(CeedRealloc(num_B_out_mats_to_load + 1, &eval_mode_out)); 962 eval_mode_out[num_B_out_mats_to_load] = eval_mode; 963 num_B_out_mats_to_load += 1; 964 if (eval_mode == CEED_EVAL_GRAD) { 965 num_emode_out += dim; 966 size_B_out += dim * esize * nqpts; 967 } else { 968 num_emode_out += 1; 969 size_B_out += esize * nqpts; 970 } 971 } 972 } 973 } 974 975 if (num_emode_in == 0 || num_emode_out == 0) { 976 // LCOV_EXCL_START 977 return CeedError(ceed, CEED_ERROR_UNSUPPORTED, "Cannot assemble operator without inputs/outputs"); 978 // LCOV_EXCL_STOP 979 } 980 981 CeedInt nelem, ncomp; 982 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr_in, &nelem)); 983 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr_in, &ncomp)); 984 985 CeedCallBackend(CeedCalloc(1, &impl->asmb)); 986 CeedOperatorAssemble_Cuda *asmb = impl->asmb; 987 asmb->nelem = nelem; 988 989 // Compile kernels 990 int elemsPerBlock = 1; 991 asmb->elemsPerBlock = elemsPerBlock; 992 CeedInt block_size = esize * esize * elemsPerBlock; 993 Ceed_Cuda *cuda_data; 994 CeedCallBackend(CeedGetData(ceed, &cuda_data)); 995 char *assembly_kernel_path, *assembly_kernel_source; 996 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-operator-assemble.h", &assembly_kernel_path)); 997 CeedDebug256(ceed, 2, "----- Loading Assembly Kernel Source -----\n"); 998 CeedCallBackend(CeedLoadSourceToBuffer(ceed, assembly_kernel_path, &assembly_kernel_source)); 999 CeedDebug256(ceed, 2, "----- Loading Assembly Source Complete! -----\n"); 1000 bool fallback = block_size > cuda_data->device_prop.maxThreadsPerBlock; 1001 if (fallback) { 1002 // Use fallback kernel with 1D threadblock 1003 block_size = esize * elemsPerBlock; 1004 asmb->block_size_x = esize; 1005 asmb->block_size_y = 1; 1006 } else { // Use kernel with 2D threadblock 1007 asmb->block_size_x = esize; 1008 asmb->block_size_y = esize; 1009 } 1010 CeedCallCuda(ceed, CeedCompileCuda(ceed, assembly_kernel_source, &asmb->module, 7, "NELEM", nelem, "NUMEMODEIN", num_emode_in, "NUMEMODEOUT", 1011 num_emode_out, "NQPTS", nqpts, "NNODES", esize, "BLOCK_SIZE", block_size, "NCOMP", ncomp)); 1012 CeedCallCuda(ceed, CeedGetKernelCuda(ceed, asmb->module, fallback ? "linearAssembleFallback" : "linearAssemble", &asmb->linearAssemble)); 1013 CeedCallBackend(CeedFree(&assembly_kernel_path)); 1014 CeedCallBackend(CeedFree(&assembly_kernel_source)); 1015 1016 // Build 'full' B matrices (not 1D arrays used for tensor-product matrices) 1017 const CeedScalar *interp_in, *grad_in; 1018 CeedCallBackend(CeedBasisGetInterp(basis_in, &interp_in)); 1019 CeedCallBackend(CeedBasisGetGrad(basis_in, &grad_in)); 1020 1021 // Load into B_in, in order that they will be used in eval_mode 1022 const CeedInt inBytes = size_B_in * sizeof(CeedScalar); 1023 CeedInt mat_start = 0; 1024 CeedCallCuda(ceed, cudaMalloc((void **)&asmb->d_B_in, inBytes)); 1025 for (int i = 0; i < num_B_in_mats_to_load; i++) { 1026 CeedEvalMode eval_mode = eval_mode_in[i]; 1027 if (eval_mode == CEED_EVAL_INTERP) { 1028 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_in[mat_start], interp_in, esize * nqpts * sizeof(CeedScalar), cudaMemcpyHostToDevice)); 1029 mat_start += esize * nqpts; 1030 } else if (eval_mode == CEED_EVAL_GRAD) { 1031 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_in[mat_start], grad_in, dim * esize * nqpts * sizeof(CeedScalar), cudaMemcpyHostToDevice)); 1032 mat_start += dim * esize * nqpts; 1033 } 1034 } 1035 1036 const CeedScalar *interp_out, *grad_out; 1037 // Note that this function currently assumes 1 basis, so this should always be true 1038 // for now 1039 if (basis_out == basis_in) { 1040 interp_out = interp_in; 1041 grad_out = grad_in; 1042 } else { 1043 CeedCallBackend(CeedBasisGetInterp(basis_out, &interp_out)); 1044 CeedCallBackend(CeedBasisGetGrad(basis_out, &grad_out)); 1045 } 1046 1047 // Load into B_out, in order that they will be used in eval_mode 1048 const CeedInt outBytes = size_B_out * sizeof(CeedScalar); 1049 mat_start = 0; 1050 CeedCallCuda(ceed, cudaMalloc((void **)&asmb->d_B_out, outBytes)); 1051 for (int i = 0; i < num_B_out_mats_to_load; i++) { 1052 CeedEvalMode eval_mode = eval_mode_out[i]; 1053 if (eval_mode == CEED_EVAL_INTERP) { 1054 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_out[mat_start], interp_out, esize * nqpts * sizeof(CeedScalar), cudaMemcpyHostToDevice)); 1055 mat_start += esize * nqpts; 1056 } else if (eval_mode == CEED_EVAL_GRAD) { 1057 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_out[mat_start], grad_out, dim * esize * nqpts * sizeof(CeedScalar), cudaMemcpyHostToDevice)); 1058 mat_start += dim * esize * nqpts; 1059 } 1060 } 1061 return CEED_ERROR_SUCCESS; 1062 } 1063 1064 //------------------------------------------------------------------------------ 1065 // Assemble matrix data for COO matrix of assembled operator. 1066 // The sparsity pattern is set by CeedOperatorLinearAssembleSymbolic. 1067 // 1068 // Note that this (and other assembly routines) currently assume only one active input restriction/basis per operator (could have multiple basis eval 1069 // modes). 1070 // TODO: allow multiple active input restrictions/basis objects 1071 //------------------------------------------------------------------------------ 1072 static int CeedSingleOperatorAssemble_Cuda(CeedOperator op, CeedInt offset, CeedVector values) { 1073 Ceed ceed; 1074 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1075 CeedOperator_Cuda *impl; 1076 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1077 1078 // Setup 1079 if (!impl->asmb) { 1080 CeedCallBackend(CeedSingleOperatorAssembleSetup_Cuda(op)); 1081 assert(impl->asmb != NULL); 1082 } 1083 1084 // Assemble QFunction 1085 CeedVector assembled_qf; 1086 CeedElemRestriction rstr_q; 1087 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &rstr_q, CEED_REQUEST_IMMEDIATE)); 1088 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_q)); 1089 CeedScalar *values_array; 1090 CeedCallBackend(CeedVectorGetArray(values, CEED_MEM_DEVICE, &values_array)); 1091 values_array += offset; 1092 const CeedScalar *qf_array; 1093 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &qf_array)); 1094 1095 // Compute B^T D B 1096 const CeedInt nelem = impl->asmb->nelem; 1097 const CeedInt elemsPerBlock = impl->asmb->elemsPerBlock; 1098 const CeedInt grid = nelem / elemsPerBlock + ((nelem / elemsPerBlock * elemsPerBlock < nelem) ? 1 : 0); 1099 void *args[] = {&impl->asmb->d_B_in, &impl->asmb->d_B_out, &qf_array, &values_array}; 1100 CeedCallBackend( 1101 CeedRunKernelDimCuda(ceed, impl->asmb->linearAssemble, grid, impl->asmb->block_size_x, impl->asmb->block_size_y, elemsPerBlock, args)); 1102 1103 // Restore arrays 1104 CeedCallBackend(CeedVectorRestoreArray(values, &values_array)); 1105 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &qf_array)); 1106 1107 // Cleanup 1108 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 1109 1110 return CEED_ERROR_SUCCESS; 1111 } 1112 1113 //------------------------------------------------------------------------------ 1114 // Create operator 1115 //------------------------------------------------------------------------------ 1116 int CeedOperatorCreate_Cuda(CeedOperator op) { 1117 Ceed ceed; 1118 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1119 CeedOperator_Cuda *impl; 1120 1121 CeedCallBackend(CeedCalloc(1, &impl)); 1122 CeedCallBackend(CeedOperatorSetData(op, impl)); 1123 1124 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunction_Cuda)); 1125 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunctionUpdate", CeedOperatorLinearAssembleQFunctionUpdate_Cuda)); 1126 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddDiagonal", CeedOperatorLinearAssembleAddDiagonal_Cuda)); 1127 CeedCallBackend( 1128 CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddPointBlockDiagonal", CeedOperatorLinearAssembleAddPointBlockDiagonal_Cuda)); 1129 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleSingle", CeedSingleOperatorAssemble_Cuda)); 1130 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Cuda)); 1131 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Cuda)); 1132 return CEED_ERROR_SUCCESS; 1133 } 1134 1135 //------------------------------------------------------------------------------ 1136