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 <ceed.h> 9 #include <ceed/backend.h> 10 #include <ceed/jit-tools.h> 11 #include <assert.h> 12 #include <stdbool.h> 13 #include <string.h> 14 #include <hip/hip_runtime.h> 15 16 #include "../hip/ceed-hip-common.h" 17 #include "../hip/ceed-hip-compile.h" 18 #include "ceed-hip-ref.h" 19 20 //------------------------------------------------------------------------------ 21 // Destroy operator 22 //------------------------------------------------------------------------------ 23 static int CeedOperatorDestroy_Hip(CeedOperator op) { 24 CeedOperator_Hip *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 diagonal 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 CeedCallHip(ceed, hipModuleUnload(impl->diag->module)); 54 CeedCallBackend(CeedFree(&impl->diag->h_emodein)); 55 CeedCallBackend(CeedFree(&impl->diag->h_emodeout)); 56 CeedCallHip(ceed, hipFree(impl->diag->d_emodein)); 57 CeedCallHip(ceed, hipFree(impl->diag->d_emodeout)); 58 CeedCallHip(ceed, hipFree(impl->diag->d_identity)); 59 CeedCallHip(ceed, hipFree(impl->diag->d_interpin)); 60 CeedCallHip(ceed, hipFree(impl->diag->d_interpout)); 61 CeedCallHip(ceed, hipFree(impl->diag->d_gradin)); 62 CeedCallHip(ceed, hipFree(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 CeedCallHip(ceed, hipModuleUnload(impl->asmb->module)); 73 CeedCallHip(ceed, hipFree(impl->asmb->d_B_in)); 74 CeedCallHip(ceed, hipFree(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_Hip(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, CEED_VECTOR_NONE, 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_Hip(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_Hip *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_Hip(qf, op, true, impl->evecs, impl->qvecsin, 0, numinputfields, Q, numelements)); 210 211 // Outfields 212 CeedCallBackend(CeedOperatorSetupFields_Hip(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_Hip(CeedInt numinputfields, CeedQFunctionField *qfinputfields, CeedOperatorField *opinputfields, 222 CeedVector invec, const bool skipactive, CeedScalar *edata[2 * CEED_FIELD_MAX], CeedOperator_Hip *impl, 223 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_Hip(CeedInt numelements, CeedQFunctionField *qfinputfields, CeedOperatorField *opinputfields, 262 CeedInt numinputfields, const bool skipactive, CeedScalar *edata[2 * CEED_FIELD_MAX], 263 CeedOperator_Hip *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_Hip(CeedInt numinputfields, CeedQFunctionField *qfinputfields, CeedOperatorField *opinputfields, 309 const bool skipactive, CeedScalar *edata[2 * CEED_FIELD_MAX], CeedOperator_Hip *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_Hip(CeedOperator op, CeedVector invec, CeedVector outvec, CeedRequest *request) { 337 CeedOperator_Hip *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]; 353 354 // Setup 355 CeedCallBackend(CeedOperatorSetup_Hip(op)); 356 357 // Input Evecs and Restriction 358 CeedCallBackend(CeedOperatorSetupInputs_Hip(numinputfields, qfinputfields, opinputfields, invec, false, edata, impl, request)); 359 360 // Input basis apply if needed 361 CeedCallBackend(CeedOperatorInputBasis_Hip(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_Hip(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_Hip(CeedOperator op, bool build_objects, CeedVector *assembled, CeedElemRestriction *rstr, 436 CeedRequest *request) { 437 CeedOperator_Hip *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_Hip(op)); 461 462 // Check for identity 463 bool identityqf; 464 CeedCallBackend(CeedQFunctionIsIdentity(qf, &identityqf)); 465 CeedCheck(!identityqf, ceed, CEED_ERROR_BACKEND, "Assembling identity QFunctions not supported"); 466 467 // Input Evecs and Restriction 468 CeedCallBackend(CeedOperatorSetupInputs_Hip(numinputfields, qfinputfields, opinputfields, NULL, true, edata, impl, request)); 469 470 // Count number of active input fields 471 if (!numactivein) { 472 for (CeedInt i = 0; i < numinputfields; i++) { 473 // Get input vector 474 CeedCallBackend(CeedOperatorFieldGetVector(opinputfields[i], &vec)); 475 // Check if active input 476 if (vec == CEED_VECTOR_ACTIVE) { 477 CeedCallBackend(CeedQFunctionFieldGetSize(qfinputfields[i], &size)); 478 CeedCallBackend(CeedVectorSetValue(impl->qvecsin[i], 0.0)); 479 CeedCallBackend(CeedVectorGetArray(impl->qvecsin[i], CEED_MEM_DEVICE, &tmp)); 480 CeedCallBackend(CeedRealloc(numactivein + size, &activein)); 481 for (CeedInt field = 0; field < size; field++) { 482 q_size = (CeedSize)Q * numelements; 483 CeedCallBackend(CeedVectorCreate(ceed, q_size, &activein[numactivein + field])); 484 CeedCallBackend(CeedVectorSetArray(activein[numactivein + field], CEED_MEM_DEVICE, CEED_USE_POINTER, &tmp[field * Q * numelements])); 485 } 486 numactivein += size; 487 CeedCallBackend(CeedVectorRestoreArray(impl->qvecsin[i], &tmp)); 488 } 489 } 490 impl->qfnumactivein = numactivein; 491 impl->qfactivein = activein; 492 } 493 494 // Count number of active output fields 495 if (!numactiveout) { 496 for (CeedInt i = 0; i < numoutputfields; i++) { 497 // Get output vector 498 CeedCallBackend(CeedOperatorFieldGetVector(opoutputfields[i], &vec)); 499 // Check if active output 500 if (vec == CEED_VECTOR_ACTIVE) { 501 CeedCallBackend(CeedQFunctionFieldGetSize(qfoutputfields[i], &size)); 502 numactiveout += size; 503 } 504 } 505 impl->qfnumactiveout = numactiveout; 506 } 507 508 // Check sizes 509 CeedCheck(numactivein > 0 && numactiveout > 0, ceed, CEED_ERROR_BACKEND, "Cannot assemble QFunction without active inputs and outputs"); 510 511 // Build objects if needed 512 if (build_objects) { 513 // Create output restriction 514 CeedInt strides[3] = {1, numelements * Q, Q}; /* *NOPAD* */ 515 CeedCallBackend(CeedElemRestrictionCreateStrided(ceedparent, numelements, Q, numactivein * numactiveout, 516 numactivein * numactiveout * numelements * Q, strides, rstr)); 517 // Create assembled vector 518 CeedSize l_size = (CeedSize)numelements * Q * numactivein * numactiveout; 519 CeedCallBackend(CeedVectorCreate(ceedparent, l_size, assembled)); 520 } 521 CeedCallBackend(CeedVectorSetValue(*assembled, 0.0)); 522 CeedCallBackend(CeedVectorGetArray(*assembled, CEED_MEM_DEVICE, &a)); 523 524 // Input basis apply 525 CeedCallBackend(CeedOperatorInputBasis_Hip(numelements, qfinputfields, opinputfields, numinputfields, true, edata, impl)); 526 527 // Assemble QFunction 528 for (CeedInt in = 0; in < numactivein; in++) { 529 // Set Inputs 530 CeedCallBackend(CeedVectorSetValue(activein[in], 1.0)); 531 if (numactivein > 1) { 532 CeedCallBackend(CeedVectorSetValue(activein[(in + numactivein - 1) % numactivein], 0.0)); 533 } 534 // Set Outputs 535 for (CeedInt out = 0; out < numoutputfields; out++) { 536 // Get output vector 537 CeedCallBackend(CeedOperatorFieldGetVector(opoutputfields[out], &vec)); 538 // Check if active output 539 if (vec == CEED_VECTOR_ACTIVE) { 540 CeedCallBackend(CeedVectorSetArray(impl->qvecsout[out], CEED_MEM_DEVICE, CEED_USE_POINTER, a)); 541 CeedCallBackend(CeedQFunctionFieldGetSize(qfoutputfields[out], &size)); 542 a += size * Q * numelements; // Advance the pointer by the size of the output 543 } 544 } 545 // Apply QFunction 546 CeedCallBackend(CeedQFunctionApply(qf, Q * numelements, impl->qvecsin, impl->qvecsout)); 547 } 548 549 // Un-set output Qvecs to prevent accidental overwrite of Assembled 550 for (CeedInt out = 0; out < numoutputfields; out++) { 551 // Get output vector 552 CeedCallBackend(CeedOperatorFieldGetVector(opoutputfields[out], &vec)); 553 // Check if active output 554 if (vec == CEED_VECTOR_ACTIVE) { 555 CeedCallBackend(CeedVectorTakeArray(impl->qvecsout[out], CEED_MEM_DEVICE, NULL)); 556 } 557 } 558 559 // Restore input arrays 560 CeedCallBackend(CeedOperatorRestoreInputs_Hip(numinputfields, qfinputfields, opinputfields, true, edata, impl)); 561 562 // Restore output 563 CeedCallBackend(CeedVectorRestoreArray(*assembled, &a)); 564 565 return CEED_ERROR_SUCCESS; 566 } 567 568 //------------------------------------------------------------------------------ 569 // Assemble Linear QFunction 570 //------------------------------------------------------------------------------ 571 static int CeedOperatorLinearAssembleQFunction_Hip(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 572 return CeedOperatorLinearAssembleQFunctionCore_Hip(op, true, assembled, rstr, request); 573 } 574 575 //------------------------------------------------------------------------------ 576 // Assemble Linear QFunction 577 //------------------------------------------------------------------------------ 578 static int CeedOperatorLinearAssembleQFunctionUpdate_Hip(CeedOperator op, CeedVector assembled, CeedElemRestriction rstr, CeedRequest *request) { 579 return CeedOperatorLinearAssembleQFunctionCore_Hip(op, false, &assembled, &rstr, request); 580 } 581 582 //------------------------------------------------------------------------------ 583 // Create point block restriction 584 //------------------------------------------------------------------------------ 585 static int CreatePBRestriction(CeedElemRestriction rstr, CeedElemRestriction *pbRstr) { 586 Ceed ceed; 587 CeedCallBackend(CeedElemRestrictionGetCeed(rstr, &ceed)); 588 const CeedInt *offsets; 589 CeedCallBackend(CeedElemRestrictionGetOffsets(rstr, CEED_MEM_HOST, &offsets)); 590 591 // Expand offsets 592 CeedInt nelem, ncomp, elemsize, compstride, *pbOffsets; 593 CeedSize l_size; 594 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr, &nelem)); 595 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr, &ncomp)); 596 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr, &elemsize)); 597 CeedCallBackend(CeedElemRestrictionGetCompStride(rstr, &compstride)); 598 CeedCallBackend(CeedElemRestrictionGetLVectorSize(rstr, &l_size)); 599 CeedInt shift = ncomp; 600 if (compstride != 1) shift *= ncomp; 601 CeedCallBackend(CeedCalloc(nelem * elemsize, &pbOffsets)); 602 for (CeedInt i = 0; i < nelem * elemsize; i++) { 603 pbOffsets[i] = offsets[i] * shift; 604 } 605 606 // Create new restriction 607 CeedCallBackend( 608 CeedElemRestrictionCreate(ceed, nelem, elemsize, ncomp * ncomp, 1, l_size * ncomp, CEED_MEM_HOST, CEED_OWN_POINTER, pbOffsets, pbRstr)); 609 610 // Cleanup 611 CeedCallBackend(CeedElemRestrictionRestoreOffsets(rstr, &offsets)); 612 613 return CEED_ERROR_SUCCESS; 614 } 615 616 //------------------------------------------------------------------------------ 617 // Assemble diagonal setup 618 //------------------------------------------------------------------------------ 619 static inline int CeedOperatorAssembleDiagonalSetup_Hip(CeedOperator op, const bool pointBlock, CeedInt use_ceedsize_idx) { 620 Ceed ceed; 621 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 622 CeedQFunction qf; 623 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 624 CeedInt numinputfields, numoutputfields; 625 CeedCallBackend(CeedQFunctionGetNumArgs(qf, &numinputfields, &numoutputfields)); 626 627 // Determine active input basis 628 CeedOperatorField *opfields; 629 CeedQFunctionField *qffields; 630 CeedCallBackend(CeedOperatorGetFields(op, NULL, &opfields, NULL, NULL)); 631 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qffields, NULL, NULL)); 632 CeedInt numemodein = 0, ncomp = 0, dim = 1; 633 CeedEvalMode *emodein = NULL; 634 CeedBasis basisin = NULL; 635 CeedElemRestriction rstrin = NULL; 636 for (CeedInt i = 0; i < numinputfields; i++) { 637 CeedVector vec; 638 CeedCallBackend(CeedOperatorFieldGetVector(opfields[i], &vec)); 639 if (vec == CEED_VECTOR_ACTIVE) { 640 CeedElemRestriction rstr; 641 CeedCallBackend(CeedOperatorFieldGetBasis(opfields[i], &basisin)); 642 CeedCallBackend(CeedBasisGetNumComponents(basisin, &ncomp)); 643 CeedCallBackend(CeedBasisGetDimension(basisin, &dim)); 644 CeedCallBackend(CeedOperatorFieldGetElemRestriction(opfields[i], &rstr)); 645 CeedCheck(!rstrin || rstrin == rstr, ceed, CEED_ERROR_BACKEND, 646 "Backend does not implement multi-field non-composite operator diagonal assembly"); 647 rstrin = rstr; 648 CeedEvalMode emode; 649 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qffields[i], &emode)); 650 switch (emode) { 651 case CEED_EVAL_NONE: 652 case CEED_EVAL_INTERP: 653 CeedCallBackend(CeedRealloc(numemodein + 1, &emodein)); 654 emodein[numemodein] = emode; 655 numemodein += 1; 656 break; 657 case CEED_EVAL_GRAD: 658 CeedCallBackend(CeedRealloc(numemodein + dim, &emodein)); 659 for (CeedInt d = 0; d < dim; d++) emodein[numemodein + d] = emode; 660 numemodein += dim; 661 break; 662 case CEED_EVAL_WEIGHT: 663 case CEED_EVAL_DIV: 664 case CEED_EVAL_CURL: 665 break; // Caught by QF Assembly 666 } 667 } 668 } 669 670 // Determine active output basis 671 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &opfields)); 672 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qffields)); 673 CeedInt numemodeout = 0; 674 CeedEvalMode *emodeout = NULL; 675 CeedBasis basisout = NULL; 676 CeedElemRestriction rstrout = NULL; 677 for (CeedInt i = 0; i < numoutputfields; i++) { 678 CeedVector vec; 679 CeedCallBackend(CeedOperatorFieldGetVector(opfields[i], &vec)); 680 if (vec == CEED_VECTOR_ACTIVE) { 681 CeedElemRestriction rstr; 682 CeedCallBackend(CeedOperatorFieldGetBasis(opfields[i], &basisout)); 683 CeedCallBackend(CeedOperatorFieldGetElemRestriction(opfields[i], &rstr)); 684 CeedCheck(!rstrout || rstrout == rstr, ceed, CEED_ERROR_BACKEND, 685 "Backend does not implement multi-field non-composite operator diagonal assembly"); 686 rstrout = rstr; 687 CeedEvalMode emode; 688 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qffields[i], &emode)); 689 switch (emode) { 690 case CEED_EVAL_NONE: 691 case CEED_EVAL_INTERP: 692 CeedCallBackend(CeedRealloc(numemodeout + 1, &emodeout)); 693 emodeout[numemodeout] = emode; 694 numemodeout += 1; 695 break; 696 case CEED_EVAL_GRAD: 697 CeedCallBackend(CeedRealloc(numemodeout + dim, &emodeout)); 698 for (CeedInt d = 0; d < dim; d++) emodeout[numemodeout + d] = emode; 699 numemodeout += dim; 700 break; 701 case CEED_EVAL_WEIGHT: 702 case CEED_EVAL_DIV: 703 case CEED_EVAL_CURL: 704 break; // Caught by QF Assembly 705 } 706 } 707 } 708 709 // Operator data struct 710 CeedOperator_Hip *impl; 711 CeedCallBackend(CeedOperatorGetData(op, &impl)); 712 CeedCallBackend(CeedCalloc(1, &impl->diag)); 713 CeedOperatorDiag_Hip *diag = impl->diag; 714 diag->basisin = basisin; 715 diag->basisout = basisout; 716 diag->h_emodein = emodein; 717 diag->h_emodeout = emodeout; 718 diag->numemodein = numemodein; 719 diag->numemodeout = numemodeout; 720 721 // Assemble kernel 722 723 char *diagonal_kernel_path, *diagonal_kernel_source; 724 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-ref-operator-assemble-diagonal.h", &diagonal_kernel_path)); 725 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Kernel Source -----\n"); 726 CeedCallBackend(CeedLoadSourceToBuffer(ceed, diagonal_kernel_path, &diagonal_kernel_source)); 727 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Source Complete! -----\n"); 728 CeedInt nnodes, nqpts; 729 CeedCallBackend(CeedBasisGetNumNodes(basisin, &nnodes)); 730 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basisin, &nqpts)); 731 diag->nnodes = nnodes; 732 CeedCallBackend(CeedCompile_Hip(ceed, diagonal_kernel_source, &diag->module, 6, "NUMEMODEIN", numemodein, "NUMEMODEOUT", numemodeout, "NNODES", 733 nnodes, "NQPTS", nqpts, "NCOMP", ncomp, "CEEDSIZE", use_ceedsize_idx)); 734 CeedCallBackend(CeedGetKernel_Hip(ceed, diag->module, "linearDiagonal", &diag->linearDiagonal)); 735 CeedCallBackend(CeedGetKernel_Hip(ceed, diag->module, "linearPointBlockDiagonal", &diag->linearPointBlock)); 736 CeedCallBackend(CeedFree(&diagonal_kernel_path)); 737 CeedCallBackend(CeedFree(&diagonal_kernel_source)); 738 739 // Basis matrices 740 const CeedInt qBytes = nqpts * sizeof(CeedScalar); 741 const CeedInt iBytes = qBytes * nnodes; 742 const CeedInt gBytes = qBytes * nnodes * dim; 743 const CeedInt eBytes = sizeof(CeedEvalMode); 744 const CeedScalar *interpin, *interpout, *gradin, *gradout; 745 746 // CEED_EVAL_NONE 747 CeedScalar *identity = NULL; 748 bool evalNone = false; 749 for (CeedInt i = 0; i < numemodein; i++) evalNone = evalNone || (emodein[i] == CEED_EVAL_NONE); 750 for (CeedInt i = 0; i < numemodeout; i++) evalNone = evalNone || (emodeout[i] == CEED_EVAL_NONE); 751 if (evalNone) { 752 CeedCallBackend(CeedCalloc(nqpts * nnodes, &identity)); 753 for (CeedInt i = 0; i < (nnodes < nqpts ? nnodes : nqpts); i++) identity[i * nnodes + i] = 1.0; 754 CeedCallHip(ceed, hipMalloc((void **)&diag->d_identity, iBytes)); 755 CeedCallHip(ceed, hipMemcpy(diag->d_identity, identity, iBytes, hipMemcpyHostToDevice)); 756 } 757 758 // CEED_EVAL_INTERP 759 CeedCallBackend(CeedBasisGetInterp(basisin, &interpin)); 760 CeedCallHip(ceed, hipMalloc((void **)&diag->d_interpin, iBytes)); 761 CeedCallHip(ceed, hipMemcpy(diag->d_interpin, interpin, iBytes, hipMemcpyHostToDevice)); 762 CeedCallBackend(CeedBasisGetInterp(basisout, &interpout)); 763 CeedCallHip(ceed, hipMalloc((void **)&diag->d_interpout, iBytes)); 764 CeedCallHip(ceed, hipMemcpy(diag->d_interpout, interpout, iBytes, hipMemcpyHostToDevice)); 765 766 // CEED_EVAL_GRAD 767 CeedCallBackend(CeedBasisGetGrad(basisin, &gradin)); 768 CeedCallHip(ceed, hipMalloc((void **)&diag->d_gradin, gBytes)); 769 CeedCallHip(ceed, hipMemcpy(diag->d_gradin, gradin, gBytes, hipMemcpyHostToDevice)); 770 CeedCallBackend(CeedBasisGetGrad(basisout, &gradout)); 771 CeedCallHip(ceed, hipMalloc((void **)&diag->d_gradout, gBytes)); 772 CeedCallHip(ceed, hipMemcpy(diag->d_gradout, gradout, gBytes, hipMemcpyHostToDevice)); 773 774 // Arrays of emodes 775 CeedCallHip(ceed, hipMalloc((void **)&diag->d_emodein, numemodein * eBytes)); 776 CeedCallHip(ceed, hipMemcpy(diag->d_emodein, emodein, numemodein * eBytes, hipMemcpyHostToDevice)); 777 CeedCallHip(ceed, hipMalloc((void **)&diag->d_emodeout, numemodeout * eBytes)); 778 CeedCallHip(ceed, hipMemcpy(diag->d_emodeout, emodeout, numemodeout * eBytes, hipMemcpyHostToDevice)); 779 780 // Restriction 781 diag->diagrstr = rstrout; 782 783 return CEED_ERROR_SUCCESS; 784 } 785 786 //------------------------------------------------------------------------------ 787 // Assemble diagonal common code 788 //------------------------------------------------------------------------------ 789 static inline int CeedOperatorAssembleDiagonalCore_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request, const bool pointBlock) { 790 Ceed ceed; 791 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 792 CeedOperator_Hip *impl; 793 CeedCallBackend(CeedOperatorGetData(op, &impl)); 794 795 // Assemble QFunction 796 CeedVector assembledqf = NULL; 797 CeedElemRestriction rstr = NULL; 798 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembledqf, &rstr, request)); 799 CeedCallBackend(CeedElemRestrictionDestroy(&rstr)); 800 801 CeedSize assembled_length = 0, assembledqf_length = 0; 802 CeedCallBackend(CeedVectorGetLength(assembled, &assembled_length)); 803 CeedCallBackend(CeedVectorGetLength(assembledqf, &assembledqf_length)); 804 CeedInt use_ceedsize_idx = 0; 805 if ((assembled_length > INT_MAX) || (assembledqf_length > INT_MAX)) use_ceedsize_idx = 1; 806 807 // Setup 808 if (!impl->diag) CeedCallBackend(CeedOperatorAssembleDiagonalSetup_Hip(op, pointBlock, use_ceedsize_idx)); 809 CeedOperatorDiag_Hip *diag = impl->diag; 810 assert(diag != NULL); 811 812 // Restriction 813 if (pointBlock && !diag->pbdiagrstr) { 814 CeedElemRestriction pbdiagrstr; 815 CeedCallBackend(CreatePBRestriction(diag->diagrstr, &pbdiagrstr)); 816 diag->pbdiagrstr = pbdiagrstr; 817 } 818 CeedElemRestriction diagrstr = pointBlock ? diag->pbdiagrstr : diag->diagrstr; 819 820 // Create diagonal vector 821 CeedVector elemdiag = pointBlock ? diag->pbelemdiag : diag->elemdiag; 822 if (!elemdiag) { 823 // Element diagonal vector 824 CeedCallBackend(CeedElemRestrictionCreateVector(diagrstr, NULL, &elemdiag)); 825 if (pointBlock) diag->pbelemdiag = elemdiag; 826 else diag->elemdiag = elemdiag; 827 } 828 CeedCallBackend(CeedVectorSetValue(elemdiag, 0.0)); 829 830 // Assemble element operator diagonals 831 CeedScalar *elemdiagarray; 832 const CeedScalar *assembledqfarray; 833 CeedCallBackend(CeedVectorGetArray(elemdiag, CEED_MEM_DEVICE, &elemdiagarray)); 834 CeedCallBackend(CeedVectorGetArrayRead(assembledqf, CEED_MEM_DEVICE, &assembledqfarray)); 835 CeedInt nelem; 836 CeedCallBackend(CeedElemRestrictionGetNumElements(diagrstr, &nelem)); 837 838 // Compute the diagonal of B^T D B 839 int elemsPerBlock = 1; 840 int grid = nelem / elemsPerBlock + ((nelem / elemsPerBlock * elemsPerBlock < nelem) ? 1 : 0); 841 void *args[] = {(void *)&nelem, &diag->d_identity, &diag->d_interpin, &diag->d_gradin, &diag->d_interpout, 842 &diag->d_gradout, &diag->d_emodein, &diag->d_emodeout, &assembledqfarray, &elemdiagarray}; 843 if (pointBlock) { 844 CeedCallBackend(CeedRunKernelDim_Hip(ceed, diag->linearPointBlock, grid, diag->nnodes, 1, elemsPerBlock, args)); 845 } else { 846 CeedCallBackend(CeedRunKernelDim_Hip(ceed, diag->linearDiagonal, grid, diag->nnodes, 1, elemsPerBlock, args)); 847 } 848 849 // Restore arrays 850 CeedCallBackend(CeedVectorRestoreArray(elemdiag, &elemdiagarray)); 851 CeedCallBackend(CeedVectorRestoreArrayRead(assembledqf, &assembledqfarray)); 852 853 // Assemble local operator diagonal 854 CeedCallBackend(CeedElemRestrictionApply(diagrstr, CEED_TRANSPOSE, elemdiag, assembled, request)); 855 856 // Cleanup 857 CeedCallBackend(CeedVectorDestroy(&assembledqf)); 858 859 return CEED_ERROR_SUCCESS; 860 } 861 862 //------------------------------------------------------------------------------ 863 // Assemble Linear Diagonal 864 //------------------------------------------------------------------------------ 865 static int CeedOperatorLinearAssembleAddDiagonal_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request) { 866 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Hip(op, assembled, request, false)); 867 return CEED_ERROR_SUCCESS; 868 } 869 870 //------------------------------------------------------------------------------ 871 // Assemble Linear Point Block Diagonal 872 //------------------------------------------------------------------------------ 873 static int CeedOperatorLinearAssembleAddPointBlockDiagonal_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request) { 874 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Hip(op, assembled, request, true)); 875 return CEED_ERROR_SUCCESS; 876 } 877 878 //------------------------------------------------------------------------------ 879 // Single operator assembly setup 880 //------------------------------------------------------------------------------ 881 static int CeedSingleOperatorAssembleSetup_Hip(CeedOperator op, CeedInt use_ceedsize_idx) { 882 Ceed ceed; 883 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 884 CeedOperator_Hip *impl; 885 CeedCallBackend(CeedOperatorGetData(op, &impl)); 886 887 // Get intput and output fields 888 CeedInt num_input_fields, num_output_fields; 889 CeedOperatorField *input_fields; 890 CeedOperatorField *output_fields; 891 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 892 893 // Determine active input basis eval mode 894 CeedQFunction qf; 895 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 896 CeedQFunctionField *qf_fields; 897 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 898 // Note that the kernel will treat each dimension of a gradient action separately; 899 // i.e., when an active input has a CEED_EVAL_GRAD mode, num_emode_in will increment by dim. 900 // 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 901 // num_B_in_mats_to_load will be incremented by 1. 902 CeedInt num_emode_in = 0, dim = 1, num_B_in_mats_to_load = 0, size_B_in = 0; 903 CeedEvalMode *eval_mode_in = NULL; // will be of size num_B_in_mats_load 904 CeedBasis basis_in = NULL; 905 CeedInt nqpts = 0, esize = 0; 906 CeedElemRestriction rstr_in = NULL; 907 for (CeedInt i = 0; i < num_input_fields; i++) { 908 CeedVector vec; 909 CeedCallBackend(CeedOperatorFieldGetVector(input_fields[i], &vec)); 910 if (vec == CEED_VECTOR_ACTIVE) { 911 CeedCallBackend(CeedOperatorFieldGetBasis(input_fields[i], &basis_in)); 912 CeedCallBackend(CeedBasisGetDimension(basis_in, &dim)); 913 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &nqpts)); 914 CeedCallBackend(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr_in)); 915 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_in, &esize)); 916 CeedEvalMode eval_mode; 917 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 918 if (eval_mode != CEED_EVAL_NONE) { 919 CeedCallBackend(CeedRealloc(num_B_in_mats_to_load + 1, &eval_mode_in)); 920 eval_mode_in[num_B_in_mats_to_load] = eval_mode; 921 num_B_in_mats_to_load += 1; 922 if (eval_mode == CEED_EVAL_GRAD) { 923 num_emode_in += dim; 924 size_B_in += dim * esize * nqpts; 925 } else { 926 num_emode_in += 1; 927 size_B_in += esize * nqpts; 928 } 929 } 930 } 931 } 932 933 // Determine active output basis; basis_out and rstr_out only used if same as input, TODO 934 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 935 CeedInt num_emode_out = 0, num_B_out_mats_to_load = 0, size_B_out = 0; 936 CeedEvalMode *eval_mode_out = NULL; 937 CeedBasis basis_out = NULL; 938 CeedElemRestriction rstr_out = NULL; 939 for (CeedInt i = 0; i < num_output_fields; i++) { 940 CeedVector vec; 941 CeedCallBackend(CeedOperatorFieldGetVector(output_fields[i], &vec)); 942 if (vec == CEED_VECTOR_ACTIVE) { 943 CeedCallBackend(CeedOperatorFieldGetBasis(output_fields[i], &basis_out)); 944 CeedCallBackend(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr_out)); 945 CeedCheck(!rstr_out || rstr_out == rstr_in, ceed, CEED_ERROR_BACKEND, "Backend does not implement multi-field non-composite operator assembly"); 946 CeedEvalMode eval_mode; 947 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 948 if (eval_mode != CEED_EVAL_NONE) { 949 CeedCallBackend(CeedRealloc(num_B_out_mats_to_load + 1, &eval_mode_out)); 950 eval_mode_out[num_B_out_mats_to_load] = eval_mode; 951 num_B_out_mats_to_load += 1; 952 if (eval_mode == CEED_EVAL_GRAD) { 953 num_emode_out += dim; 954 size_B_out += dim * esize * nqpts; 955 } else { 956 num_emode_out += 1; 957 size_B_out += esize * nqpts; 958 } 959 } 960 } 961 } 962 963 CeedCheck(num_emode_in > 0 && num_emode_out > 0, ceed, CEED_ERROR_UNSUPPORTED, "Cannot assemble operator without inputs/outputs"); 964 965 CeedInt nelem, ncomp; 966 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr_in, &nelem)); 967 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr_in, &ncomp)); 968 969 CeedCallBackend(CeedCalloc(1, &impl->asmb)); 970 CeedOperatorAssemble_Hip *asmb = impl->asmb; 971 asmb->nelem = nelem; 972 973 // Compile kernels 974 int elemsPerBlock = 1; 975 asmb->elemsPerBlock = elemsPerBlock; 976 CeedInt block_size = esize * esize * elemsPerBlock; 977 char *assembly_kernel_path, *assembly_kernel_source; 978 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-ref-operator-assemble.h", &assembly_kernel_path)); 979 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Kernel Source -----\n"); 980 CeedCallBackend(CeedLoadSourceToBuffer(ceed, assembly_kernel_path, &assembly_kernel_source)); 981 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Source Complete! -----\n"); 982 bool fallback = block_size > 1024; 983 if (fallback) { // Use fallback kernel with 1D threadblock 984 block_size = esize * elemsPerBlock; 985 asmb->block_size_x = esize; 986 asmb->block_size_y = 1; 987 } else { // Use kernel with 2D threadblock 988 asmb->block_size_x = esize; 989 asmb->block_size_y = esize; 990 } 991 CeedCallBackend(CeedCompile_Hip(ceed, assembly_kernel_source, &asmb->module, 8, "NELEM", nelem, "NUMEMODEIN", num_emode_in, "NUMEMODEOUT", 992 num_emode_out, "NQPTS", nqpts, "NNODES", esize, "BLOCK_SIZE", block_size, "NCOMP", ncomp, "CEEDSIZE", 993 use_ceedsize_idx)); 994 CeedCallBackend(CeedGetKernel_Hip(ceed, asmb->module, fallback ? "linearAssembleFallback" : "linearAssemble", &asmb->linearAssemble)); 995 CeedCallBackend(CeedFree(&assembly_kernel_path)); 996 CeedCallBackend(CeedFree(&assembly_kernel_source)); 997 998 // Build 'full' B matrices (not 1D arrays used for tensor-product matrices) 999 const CeedScalar *interp_in, *grad_in; 1000 CeedCallBackend(CeedBasisGetInterp(basis_in, &interp_in)); 1001 CeedCallBackend(CeedBasisGetGrad(basis_in, &grad_in)); 1002 1003 // Load into B_in, in order that they will be used in eval_mode 1004 const CeedInt inBytes = size_B_in * sizeof(CeedScalar); 1005 CeedInt mat_start = 0; 1006 CeedCallHip(ceed, hipMalloc((void **)&asmb->d_B_in, inBytes)); 1007 for (int i = 0; i < num_B_in_mats_to_load; i++) { 1008 CeedEvalMode eval_mode = eval_mode_in[i]; 1009 if (eval_mode == CEED_EVAL_INTERP) { 1010 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_in[mat_start], interp_in, esize * nqpts * sizeof(CeedScalar), hipMemcpyHostToDevice)); 1011 mat_start += esize * nqpts; 1012 } else if (eval_mode == CEED_EVAL_GRAD) { 1013 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_in[mat_start], grad_in, dim * esize * nqpts * sizeof(CeedScalar), hipMemcpyHostToDevice)); 1014 mat_start += dim * esize * nqpts; 1015 } 1016 } 1017 1018 const CeedScalar *interp_out, *grad_out; 1019 // Note that this function currently assumes 1 basis, so this should always be true 1020 // for now 1021 if (basis_out == basis_in) { 1022 interp_out = interp_in; 1023 grad_out = grad_in; 1024 } else { 1025 CeedCallBackend(CeedBasisGetInterp(basis_out, &interp_out)); 1026 CeedCallBackend(CeedBasisGetGrad(basis_out, &grad_out)); 1027 } 1028 1029 // Load into B_out, in order that they will be used in eval_mode 1030 const CeedInt outBytes = size_B_out * sizeof(CeedScalar); 1031 mat_start = 0; 1032 CeedCallHip(ceed, hipMalloc((void **)&asmb->d_B_out, outBytes)); 1033 for (int i = 0; i < num_B_out_mats_to_load; i++) { 1034 CeedEvalMode eval_mode = eval_mode_out[i]; 1035 if (eval_mode == CEED_EVAL_INTERP) { 1036 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_out[mat_start], interp_out, esize * nqpts * sizeof(CeedScalar), hipMemcpyHostToDevice)); 1037 mat_start += esize * nqpts; 1038 } else if (eval_mode == CEED_EVAL_GRAD) { 1039 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_out[mat_start], grad_out, dim * esize * nqpts * sizeof(CeedScalar), hipMemcpyHostToDevice)); 1040 mat_start += dim * esize * nqpts; 1041 } 1042 } 1043 return CEED_ERROR_SUCCESS; 1044 } 1045 1046 //------------------------------------------------------------------------------ 1047 // Assemble matrix data for COO matrix of assembled operator. 1048 // The sparsity pattern is set by CeedOperatorLinearAssembleSymbolic. 1049 // 1050 // Note that this (and other assembly routines) currently assume only one active input restriction/basis per operator (could have multiple basis eval 1051 // modes). 1052 // TODO: allow multiple active input restrictions/basis objects 1053 //------------------------------------------------------------------------------ 1054 static int CeedSingleOperatorAssemble_Hip(CeedOperator op, CeedInt offset, CeedVector values) { 1055 Ceed ceed; 1056 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1057 CeedOperator_Hip *impl; 1058 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1059 1060 // Assemble QFunction 1061 CeedVector assembled_qf = NULL; 1062 CeedElemRestriction rstr_q = NULL; 1063 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &rstr_q, CEED_REQUEST_IMMEDIATE)); 1064 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_q)); 1065 CeedScalar *values_array; 1066 CeedCallBackend(CeedVectorGetArray(values, CEED_MEM_DEVICE, &values_array)); 1067 values_array += offset; 1068 const CeedScalar *qf_array; 1069 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &qf_array)); 1070 1071 CeedSize values_length = 0, assembled_qf_length = 0; 1072 CeedCallBackend(CeedVectorGetLength(values, &values_length)); 1073 CeedCallBackend(CeedVectorGetLength(assembled_qf, &assembled_qf_length)); 1074 CeedInt use_ceedsize_idx = 0; 1075 if ((values_length > INT_MAX) || (assembled_qf_length > INT_MAX)) use_ceedsize_idx = 1; 1076 // Setup 1077 if (!impl->asmb) { 1078 CeedCallBackend(CeedSingleOperatorAssembleSetup_Hip(op, use_ceedsize_idx)); 1079 assert(impl->asmb != NULL); 1080 } 1081 1082 // Compute B^T D B 1083 const CeedInt nelem = impl->asmb->nelem; // to satisfy clang-tidy 1084 const CeedInt elemsPerBlock = impl->asmb->elemsPerBlock; 1085 const CeedInt grid = nelem / elemsPerBlock + ((nelem / elemsPerBlock * elemsPerBlock < nelem) ? 1 : 0); 1086 void *args[] = {&impl->asmb->d_B_in, &impl->asmb->d_B_out, &qf_array, &values_array}; 1087 CeedCallBackend( 1088 CeedRunKernelDim_Hip(ceed, impl->asmb->linearAssemble, grid, impl->asmb->block_size_x, impl->asmb->block_size_y, elemsPerBlock, args)); 1089 1090 // Restore arrays 1091 CeedCallBackend(CeedVectorRestoreArray(values, &values_array)); 1092 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &qf_array)); 1093 1094 // Cleanup 1095 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 1096 1097 return CEED_ERROR_SUCCESS; 1098 } 1099 1100 //------------------------------------------------------------------------------ 1101 // Create operator 1102 //------------------------------------------------------------------------------ 1103 int CeedOperatorCreate_Hip(CeedOperator op) { 1104 Ceed ceed; 1105 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1106 CeedOperator_Hip *impl; 1107 1108 CeedCallBackend(CeedCalloc(1, &impl)); 1109 CeedCallBackend(CeedOperatorSetData(op, impl)); 1110 1111 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunction_Hip)); 1112 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunctionUpdate", CeedOperatorLinearAssembleQFunctionUpdate_Hip)); 1113 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddDiagonal", CeedOperatorLinearAssembleAddDiagonal_Hip)); 1114 CeedCallBackend( 1115 CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddPointBlockDiagonal", CeedOperatorLinearAssembleAddPointBlockDiagonal_Hip)); 1116 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleSingle", CeedSingleOperatorAssemble_Hip)); 1117 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Hip)); 1118 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Hip)); 1119 return CEED_ERROR_SUCCESS; 1120 } 1121 1122 //------------------------------------------------------------------------------ 1123