1 // Copyright (c) 2017-2024, Lawrence Livermore National Security, LLC and other 2 // CEED contributors. All Rights Reserved. See the top-level LICENSE and NOTICE 3 // files for details. 4 // 5 // SPDX-License-Identifier: BSD-2-Clause 6 // 7 // This file is part of CEED: http://github.com/ceed 8 9 #include <ceed/backend.h> 10 #include <ceed/ceed.h> 11 12 #include <cassert> 13 #include <string> 14 #include <sycl/sycl.hpp> 15 16 #include "../sycl/ceed-sycl-compile.hpp" 17 #include "ceed-sycl-ref.hpp" 18 19 class CeedOperatorSyclLinearDiagonal; 20 class CeedOperatorSyclLinearAssemble; 21 class CeedOperatorSyclLinearAssembleFallback; 22 23 //------------------------------------------------------------------------------ 24 // Get Basis Emode Pointer 25 //------------------------------------------------------------------------------ 26 void CeedOperatorGetBasisPointer_Sycl(const CeedScalar **basis_ptr, CeedEvalMode e_mode, const CeedScalar *identity, const CeedScalar *interp, 27 const CeedScalar *grad) { 28 switch (e_mode) { 29 case CEED_EVAL_NONE: 30 *basis_ptr = identity; 31 break; 32 case CEED_EVAL_INTERP: 33 *basis_ptr = interp; 34 break; 35 case CEED_EVAL_GRAD: 36 *basis_ptr = grad; 37 break; 38 case CEED_EVAL_WEIGHT: 39 case CEED_EVAL_DIV: 40 case CEED_EVAL_CURL: 41 break; // Caught by QF Assembly 42 } 43 } 44 45 //------------------------------------------------------------------------------ 46 // Destroy operator 47 //------------------------------------------------------------------------------ 48 static int CeedOperatorDestroy_Sycl(CeedOperator op) { 49 Ceed ceed; 50 Ceed_Sycl *sycl_data; 51 CeedOperator_Sycl *impl; 52 53 CeedCallBackend(CeedOperatorGetData(op, &impl)); 54 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 55 CeedCallBackend(CeedGetData(ceed, &sycl_data)); 56 57 // Apply data 58 for (CeedInt i = 0; i < impl->num_e_in + impl->num_e_out; i++) { 59 CeedCallBackend(CeedVectorDestroy(&impl->e_vecs[i])); 60 } 61 CeedCallBackend(CeedFree(&impl->e_vecs)); 62 63 for (CeedInt i = 0; i < impl->num_e_in; i++) { 64 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_in[i])); 65 } 66 CeedCallBackend(CeedFree(&impl->q_vecs_in)); 67 68 for (CeedInt i = 0; i < impl->num_e_out; i++) { 69 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_out[i])); 70 } 71 CeedCallBackend(CeedFree(&impl->q_vecs_out)); 72 73 // QFunction assembly data 74 for (CeedInt i = 0; i < impl->num_active_in; i++) { 75 CeedCallBackend(CeedVectorDestroy(&impl->qf_active_in[i])); 76 } 77 CeedCallBackend(CeedFree(&impl->qf_active_in)); 78 79 // Diag data 80 if (impl->diag) { 81 CeedCallBackend(CeedFree(&impl->diag->h_e_mode_in)); 82 CeedCallBackend(CeedFree(&impl->diag->h_e_mode_out)); 83 84 CeedCallSycl(ceed, sycl_data->sycl_queue.wait_and_throw()); 85 CeedCallSycl(ceed, sycl::free(impl->diag->d_e_mode_in, sycl_data->sycl_context)); 86 CeedCallSycl(ceed, sycl::free(impl->diag->d_e_mode_out, sycl_data->sycl_context)); 87 CeedCallSycl(ceed, sycl::free(impl->diag->d_identity, sycl_data->sycl_context)); 88 CeedCallSycl(ceed, sycl::free(impl->diag->d_interp_in, sycl_data->sycl_context)); 89 CeedCallSycl(ceed, sycl::free(impl->diag->d_interp_out, sycl_data->sycl_context)); 90 CeedCallSycl(ceed, sycl::free(impl->diag->d_grad_in, sycl_data->sycl_context)); 91 CeedCallSycl(ceed, sycl::free(impl->diag->d_grad_out, sycl_data->sycl_context)); 92 CeedCallBackend(CeedElemRestrictionDestroy(&impl->diag->point_block_diag_rstr)); 93 94 CeedCallBackend(CeedVectorDestroy(&impl->diag->elem_diag)); 95 CeedCallBackend(CeedVectorDestroy(&impl->diag->point_block_elem_diag)); 96 } 97 CeedCallBackend(CeedFree(&impl->diag)); 98 99 if (impl->asmb) { 100 CeedCallSycl(ceed, sycl_data->sycl_queue.wait_and_throw()); 101 CeedCallSycl(ceed, sycl::free(impl->asmb->d_B_in, sycl_data->sycl_context)); 102 CeedCallSycl(ceed, sycl::free(impl->asmb->d_B_out, sycl_data->sycl_context)); 103 } 104 CeedCallBackend(CeedFree(&impl->asmb)); 105 106 CeedCallBackend(CeedFree(&impl)); 107 return CEED_ERROR_SUCCESS; 108 } 109 110 //------------------------------------------------------------------------------ 111 // Setup infields or outfields 112 //------------------------------------------------------------------------------ 113 static int CeedOperatorSetupFields_Sycl(CeedQFunction qf, CeedOperator op, bool is_input, CeedVector *e_vecs, CeedVector *q_vecs, CeedInt start_e, 114 CeedInt num_fields, CeedInt Q, CeedInt num_elem) { 115 Ceed ceed; 116 CeedSize q_size; 117 bool is_strided, skip_restriction; 118 CeedInt dim, size; 119 CeedOperatorField *op_fields; 120 CeedQFunctionField *qf_fields; 121 122 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 123 if (is_input) { 124 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 125 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 126 } else { 127 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 128 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 129 } 130 131 // Loop over fields 132 for (CeedInt i = 0; i < num_fields; i++) { 133 CeedEvalMode e_mode; 134 CeedVector vec; 135 CeedElemRestriction rstr; 136 CeedBasis basis; 137 138 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &e_mode)); 139 140 is_strided = false; 141 skip_restriction = false; 142 if (e_mode != CEED_EVAL_WEIGHT) { 143 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr)); 144 145 // Check whether this field can skip the element restriction: 146 // must be passive input, with e_mode NONE, and have a strided restriction with CEED_STRIDES_BACKEND. 147 148 // First, check whether the field is input or output: 149 if (is_input) { 150 // Check for passive input: 151 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 152 if (vec != CEED_VECTOR_ACTIVE) { 153 // Check e_mode 154 if (e_mode == CEED_EVAL_NONE) { 155 // Check for is_strided restriction 156 CeedCallBackend(CeedElemRestrictionIsStrided(rstr, &is_strided)); 157 if (is_strided) { 158 // Check if vector is already in preferred backend ordering 159 CeedCallBackend(CeedElemRestrictionHasBackendStrides(rstr, &skip_restriction)); 160 } 161 } 162 } 163 } 164 if (skip_restriction) { 165 // We do not need an E-Vector, but will use the input field vector's data directly in the operator application 166 e_vecs[i + start_e] = NULL; 167 } else { 168 CeedCallBackend(CeedElemRestrictionCreateVector(rstr, NULL, &e_vecs[i + start_e])); 169 } 170 } 171 172 switch (e_mode) { 173 case CEED_EVAL_NONE: 174 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 175 q_size = (CeedSize)num_elem * Q * size; 176 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 177 break; 178 case CEED_EVAL_INTERP: 179 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 180 q_size = (CeedSize)num_elem * Q * size; 181 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 182 break; 183 case CEED_EVAL_GRAD: 184 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 185 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 186 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 187 q_size = (CeedSize)num_elem * Q * size; 188 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 189 break; 190 case CEED_EVAL_WEIGHT: // Only on input fields 191 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 192 q_size = (CeedSize)num_elem * Q; 193 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 194 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, q_vecs[i])); 195 break; 196 case CEED_EVAL_DIV: 197 break; // TODO: Not implemented 198 case CEED_EVAL_CURL: 199 break; // TODO: Not implemented 200 } 201 } 202 return CEED_ERROR_SUCCESS; 203 } 204 205 //------------------------------------------------------------------------------ 206 // CeedOperator needs to connect all the named fields (be they active or 207 // passive) to the named inputs and outputs of its CeedQFunction. 208 //------------------------------------------------------------------------------ 209 static int CeedOperatorSetup_Sycl(CeedOperator op) { 210 Ceed ceed; 211 bool is_setup_done; 212 CeedInt Q, num_elem, num_input_fields, num_output_fields; 213 CeedQFunctionField *qf_input_fields, *qf_output_fields; 214 CeedQFunction qf; 215 CeedOperatorField *op_input_fields, *op_output_fields; 216 CeedOperator_Sycl *impl; 217 218 CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done)); 219 if (is_setup_done) return CEED_ERROR_SUCCESS; 220 221 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 222 CeedCallBackend(CeedOperatorGetData(op, &impl)); 223 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 224 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 225 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 226 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 227 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 228 229 // Allocate 230 CeedCallBackend(CeedCalloc(num_input_fields + num_output_fields, &impl->e_vecs)); 231 232 CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_in)); 233 CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_out)); 234 235 impl->num_e_in = num_input_fields; 236 impl->num_e_out = num_output_fields; 237 238 // Set up infield and outfield e_vecs and q_vecs 239 // Infields 240 CeedCallBackend(CeedOperatorSetupFields_Sycl(qf, op, true, impl->e_vecs, impl->q_vecs_in, 0, num_input_fields, Q, num_elem)); 241 // Outfields 242 CeedCallBackend(CeedOperatorSetupFields_Sycl(qf, op, false, impl->e_vecs, impl->q_vecs_out, num_input_fields, num_output_fields, Q, num_elem)); 243 244 CeedCallBackend(CeedOperatorSetSetupDone(op)); 245 return CEED_ERROR_SUCCESS; 246 } 247 248 //------------------------------------------------------------------------------ 249 // Setup Operator Inputs 250 //------------------------------------------------------------------------------ 251 static inline int CeedOperatorSetupInputs_Sycl(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields, 252 CeedVector in_vec, const bool skip_active, CeedScalar *e_data[2 * CEED_FIELD_MAX], 253 CeedOperator_Sycl *impl, CeedRequest *request) { 254 for (CeedInt i = 0; i < num_input_fields; i++) { 255 CeedEvalMode e_mode; 256 CeedVector vec; 257 CeedElemRestriction rstr; 258 259 // Get input vector 260 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 261 if (vec == CEED_VECTOR_ACTIVE) { 262 if (skip_active) continue; 263 else vec = in_vec; 264 } 265 266 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &e_mode)); 267 if (e_mode == CEED_EVAL_WEIGHT) { // Skip 268 } else { 269 // Get input vector 270 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 271 // Get input element restriction 272 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &rstr)); 273 if (vec == CEED_VECTOR_ACTIVE) vec = in_vec; 274 // Restrict, if necessary 275 if (!impl->e_vecs[i]) { 276 // No restriction for this field; read data directly from vec. 277 CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, (const CeedScalar **)&e_data[i])); 278 } else { 279 CeedCallBackend(CeedElemRestrictionApply(rstr, CEED_NOTRANSPOSE, vec, impl->e_vecs[i], request)); 280 // Get evec 281 CeedCallBackend(CeedVectorGetArrayRead(impl->e_vecs[i], CEED_MEM_DEVICE, (const CeedScalar **)&e_data[i])); 282 } 283 } 284 } 285 return CEED_ERROR_SUCCESS; 286 } 287 288 //------------------------------------------------------------------------------ 289 // Input Basis Action 290 //------------------------------------------------------------------------------ 291 static inline int CeedOperatorInputBasis_Sycl(CeedInt num_elem, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields, 292 CeedInt num_input_fields, const bool skip_active, CeedScalar *e_data[2 * CEED_FIELD_MAX], 293 CeedOperator_Sycl *impl) { 294 for (CeedInt i = 0; i < num_input_fields; i++) { 295 CeedInt elem_size, size; 296 CeedElemRestriction rstr; 297 CeedEvalMode e_mode; 298 CeedBasis basis; 299 300 // Skip active input 301 if (skip_active) { 302 CeedVector vec; 303 304 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 305 if (vec == CEED_VECTOR_ACTIVE) continue; 306 } 307 // Get elem_size, e_mode, size 308 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &rstr)); 309 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr, &elem_size)); 310 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &e_mode)); 311 CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size)); 312 // Basis action 313 switch (e_mode) { 314 case CEED_EVAL_NONE: 315 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_data[i])); 316 break; 317 case CEED_EVAL_INTERP: 318 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 319 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_INTERP, impl->e_vecs[i], impl->q_vecs_in[i])); 320 break; 321 case CEED_EVAL_GRAD: 322 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 323 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_GRAD, impl->e_vecs[i], impl->q_vecs_in[i])); 324 break; 325 case CEED_EVAL_WEIGHT: 326 break; // No action 327 case CEED_EVAL_DIV: 328 break; // TODO: Not implemented 329 case CEED_EVAL_CURL: 330 break; // TODO: Not implemented 331 } 332 } 333 return CEED_ERROR_SUCCESS; 334 } 335 336 //------------------------------------------------------------------------------ 337 // Restore Input Vectors 338 //------------------------------------------------------------------------------ 339 static inline int CeedOperatorRestoreInputs_Sycl(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields, 340 const bool skip_active, CeedScalar *e_data[2 * CEED_FIELD_MAX], CeedOperator_Sycl *impl) { 341 for (CeedInt i = 0; i < num_input_fields; i++) { 342 CeedEvalMode e_mode; 343 CeedVector vec; 344 345 // Skip active input 346 if (skip_active) { 347 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 348 if (vec == CEED_VECTOR_ACTIVE) continue; 349 } 350 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &e_mode)); 351 if (e_mode == CEED_EVAL_WEIGHT) { // Skip 352 } else { 353 if (!impl->e_vecs[i]) { // This was a skip_restriction case 354 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 355 CeedCallBackend(CeedVectorRestoreArrayRead(vec, (const CeedScalar **)&e_data[i])); 356 } else { 357 CeedCallBackend(CeedVectorRestoreArrayRead(impl->e_vecs[i], (const CeedScalar **)&e_data[i])); 358 } 359 } 360 } 361 return CEED_ERROR_SUCCESS; 362 } 363 364 //------------------------------------------------------------------------------ 365 // Apply and add to output 366 //------------------------------------------------------------------------------ 367 static int CeedOperatorApplyAdd_Sycl(CeedOperator op, CeedVector in_vec, CeedVector out_vec, CeedRequest *request) { 368 CeedInt Q, num_elem, elem_size, num_input_fields, num_output_fields, size; 369 CeedEvalMode eval_mode; 370 CeedScalar *e_data[2 * CEED_FIELD_MAX] = {0}; 371 CeedQFunctionField *qf_input_fields, *qf_output_fields; 372 CeedQFunction qf; 373 CeedOperatorField *op_input_fields, *op_output_fields; 374 CeedOperator_Sycl *impl; 375 376 CeedCallBackend(CeedOperatorGetData(op, &impl)); 377 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 378 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 379 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 380 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 381 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 382 383 // Setup 384 CeedCallBackend(CeedOperatorSetup_Sycl(op)); 385 386 // Input Evecs and Restriction 387 CeedCallBackend(CeedOperatorSetupInputs_Sycl(num_input_fields, qf_input_fields, op_input_fields, in_vec, false, e_data, impl, request)); 388 389 // Input basis apply if needed 390 CeedCallBackend(CeedOperatorInputBasis_Sycl(num_elem, qf_input_fields, op_input_fields, num_input_fields, false, e_data, impl)); 391 392 // Output pointers, as necessary 393 for (CeedInt i = 0; i < num_output_fields; i++) { 394 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 395 if (eval_mode == CEED_EVAL_NONE) { 396 // Set the output Q-Vector to use the E-Vector data directly 397 CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs[i + impl->num_e_in], CEED_MEM_DEVICE, &e_data[i + num_input_fields])); 398 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_data[i + num_input_fields])); 399 } 400 } 401 402 // Q function 403 CeedCallBackend(CeedQFunctionApply(qf, num_elem * Q, impl->q_vecs_in, impl->q_vecs_out)); 404 405 // Output basis apply if needed 406 for (CeedInt i = 0; i < num_output_fields; i++) { 407 CeedElemRestriction rstr; 408 CeedBasis basis; 409 410 // Get elem_size, eval_mode, size 411 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &rstr)); 412 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr, &elem_size)); 413 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 414 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size)); 415 // Basis action 416 switch (eval_mode) { 417 case CEED_EVAL_NONE: 418 break; 419 case CEED_EVAL_INTERP: 420 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 421 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_TRANSPOSE, CEED_EVAL_INTERP, impl->q_vecs_out[i], impl->e_vecs[i + impl->num_e_in])); 422 break; 423 case CEED_EVAL_GRAD: 424 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 425 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_TRANSPOSE, CEED_EVAL_GRAD, impl->q_vecs_out[i], impl->e_vecs[i + impl->num_e_in])); 426 break; 427 // LCOV_EXCL_START 428 case CEED_EVAL_WEIGHT: 429 Ceed ceed; 430 431 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 432 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 433 break; // Should not occur 434 case CEED_EVAL_DIV: 435 case CEED_EVAL_CURL: { 436 Ceed ceed; 437 438 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 439 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 440 break; // Should not occur 441 } 442 // LCOV_EXCL_STOP 443 } 444 } 445 446 // Output restriction 447 for (CeedInt i = 0; i < num_output_fields; i++) { 448 CeedVector vec; 449 CeedElemRestriction rstr; 450 451 // Restore evec 452 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 453 if (eval_mode == CEED_EVAL_NONE) { 454 CeedCallBackend(CeedVectorRestoreArray(impl->e_vecs[i + impl->num_e_in], &e_data[i + num_input_fields])); 455 } 456 // Get output vector 457 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 458 // Restrict 459 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &rstr)); 460 // Active 461 if (vec == CEED_VECTOR_ACTIVE) vec = out_vec; 462 463 CeedCallBackend(CeedElemRestrictionApply(rstr, CEED_TRANSPOSE, impl->e_vecs[i + impl->num_e_in], vec, request)); 464 } 465 466 // Restore input arrays 467 CeedCallBackend(CeedOperatorRestoreInputs_Sycl(num_input_fields, qf_input_fields, op_input_fields, false, e_data, impl)); 468 return CEED_ERROR_SUCCESS; 469 } 470 471 //------------------------------------------------------------------------------ 472 // Core code for assembling linear QFunction 473 //------------------------------------------------------------------------------ 474 static inline int CeedOperatorLinearAssembleQFunctionCore_Sycl(CeedOperator op, bool build_objects, CeedVector *assembled, CeedElemRestriction *rstr, 475 CeedRequest *request) { 476 Ceed ceed, ceed_parent; 477 CeedSize q_size; 478 CeedInt num_active_in, num_active_out, Q, num_elem, num_input_fields, num_output_fields, size; 479 CeedScalar *assembled_array, *e_data[2 * CEED_FIELD_MAX] = {NULL}; 480 CeedVector *active_in; 481 CeedQFunctionField *qf_input_fields, *qf_output_fields; 482 CeedQFunction qf; 483 CeedOperatorField *op_input_fields, *op_output_fields; 484 CeedOperator_Sycl *impl; 485 486 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 487 CeedCallBackend(CeedOperatorGetFallbackParentCeed(op, &ceed_parent)); 488 CeedCallBackend(CeedOperatorGetData(op, &impl)); 489 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 490 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 491 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 492 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 493 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 494 active_in = impl->qf_active_in; 495 num_active_in = impl->num_active_in, num_active_out = impl->num_active_out; 496 497 // Setup 498 CeedCallBackend(CeedOperatorSetup_Sycl(op)); 499 500 // Input Evecs and Restriction 501 CeedCallBackend(CeedOperatorSetupInputs_Sycl(num_input_fields, qf_input_fields, op_input_fields, NULL, true, e_data, impl, request)); 502 503 // Count number of active input fields 504 if (!num_active_in) { 505 for (CeedInt i = 0; i < num_input_fields; i++) { 506 CeedScalar *q_vec_array; 507 CeedVector vec; 508 509 // Get input vector 510 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 511 // Check if active input 512 if (vec == CEED_VECTOR_ACTIVE) { 513 CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size)); 514 CeedCallBackend(CeedVectorSetValue(impl->q_vecs_in[i], 0.0)); 515 CeedCallBackend(CeedVectorGetArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, &q_vec_array)); 516 CeedCallBackend(CeedRealloc(num_active_in + size, &active_in)); 517 for (CeedInt field = 0; field < size; field++) { 518 q_size = (CeedSize)Q * num_elem; 519 CeedCallBackend(CeedVectorCreate(ceed, q_size, &active_in[num_active_in + field])); 520 CeedCallBackend( 521 CeedVectorSetArray(active_in[num_active_in + field], CEED_MEM_DEVICE, CEED_USE_POINTER, &q_vec_array[field * Q * num_elem])); 522 } 523 num_active_in += size; 524 CeedCallBackend(CeedVectorRestoreArray(impl->q_vecs_in[i], &q_vec_array)); 525 } 526 } 527 impl->num_active_in = num_active_in; 528 impl->qf_active_in = active_in; 529 } 530 531 // Count number of active output fields 532 if (!num_active_out) { 533 for (CeedInt i = 0; i < num_output_fields; i++) { 534 CeedVector vec; 535 536 // Get output vector 537 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 538 // Check if active output 539 if (vec == CEED_VECTOR_ACTIVE) { 540 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size)); 541 num_active_out += size; 542 } 543 } 544 impl->num_active_out = num_active_out; 545 } 546 547 // Check sizes 548 CeedCheck(num_active_in > 0 && num_active_out > 0, ceed, CEED_ERROR_BACKEND, "Cannot assemble QFunction without active inputs and outputs"); 549 550 // Build objects if needed 551 if (build_objects) { 552 CeedSize l_size = (CeedSize)num_elem * Q * num_active_in * num_active_out; 553 CeedInt strides[3] = {1, num_elem * Q, Q}; /* *NOPAD* */ 554 555 // Create output restriction 556 CeedCallBackend(CeedElemRestrictionCreateStrided(ceed_parent, num_elem, Q, num_active_in * num_active_out, 557 num_active_in * num_active_out * num_elem * Q, strides, rstr)); 558 // Create assembled vector 559 CeedCallBackend(CeedVectorCreate(ceed_parent, l_size, assembled)); 560 } 561 CeedCallBackend(CeedVectorSetValue(*assembled, 0.0)); 562 CeedCallBackend(CeedVectorGetArray(*assembled, CEED_MEM_DEVICE, &assembled_array)); 563 564 // Input basis apply 565 CeedCallBackend(CeedOperatorInputBasis_Sycl(num_elem, qf_input_fields, op_input_fields, num_input_fields, true, e_data, impl)); 566 567 // Assemble QFunction 568 for (CeedInt in = 0; in < num_active_in; in++) { 569 // Set Inputs 570 CeedCallBackend(CeedVectorSetValue(active_in[in], 1.0)); 571 if (num_active_in > 1) { 572 CeedCallBackend(CeedVectorSetValue(active_in[(in + num_active_in - 1) % num_active_in], 0.0)); 573 } 574 // Set Outputs 575 for (CeedInt out = 0; out < num_output_fields; out++) { 576 CeedVector vec; 577 578 // Get output vector 579 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec)); 580 // Check if active output 581 if (vec == CEED_VECTOR_ACTIVE) { 582 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, CEED_USE_POINTER, assembled_array)); 583 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[out], &size)); 584 assembled_array += size * Q * num_elem; // Advance the pointer by the size of the output 585 } 586 } 587 // Apply QFunction 588 CeedCallBackend(CeedQFunctionApply(qf, Q * num_elem, impl->q_vecs_in, impl->q_vecs_out)); 589 } 590 591 // Un-set output Qvecs to prevent accidental overwrite of Assembled 592 for (CeedInt out = 0; out < num_output_fields; out++) { 593 CeedVector vec; 594 595 // Get output vector 596 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec)); 597 // Check if active output 598 if (vec == CEED_VECTOR_ACTIVE) { 599 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, NULL)); 600 } 601 } 602 603 // Restore input arrays 604 CeedCallBackend(CeedOperatorRestoreInputs_Sycl(num_input_fields, qf_input_fields, op_input_fields, true, e_data, impl)); 605 606 // Restore output 607 CeedCallBackend(CeedVectorRestoreArray(*assembled, &assembled_array)); 608 return CEED_ERROR_SUCCESS; 609 } 610 611 //------------------------------------------------------------------------------ 612 // Assemble Linear QFunction 613 //------------------------------------------------------------------------------ 614 static int CeedOperatorLinearAssembleQFunction_Sycl(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 615 return CeedOperatorLinearAssembleQFunctionCore_Sycl(op, true, assembled, rstr, request); 616 } 617 618 //------------------------------------------------------------------------------ 619 // Update Assembled Linear QFunction 620 //------------------------------------------------------------------------------ 621 static int CeedOperatorLinearAssembleQFunctionUpdate_Sycl(CeedOperator op, CeedVector assembled, CeedElemRestriction rstr, CeedRequest *request) { 622 return CeedOperatorLinearAssembleQFunctionCore_Sycl(op, false, &assembled, &rstr, request); 623 } 624 625 //------------------------------------------------------------------------------ 626 // Assemble diagonal setup 627 //------------------------------------------------------------------------------ 628 static inline int CeedOperatorAssembleDiagonalSetup_Sycl(CeedOperator op) { 629 Ceed ceed; 630 Ceed_Sycl *sycl_data; 631 CeedInt num_input_fields, num_output_fields, num_e_mode_in = 0, num_comp = 0, dim = 1, num_e_mode_out = 0; 632 CeedEvalMode *e_mode_in = NULL, *e_mode_out = NULL; 633 CeedBasis basis_in = NULL, basis_out = NULL; 634 CeedElemRestriction rstr_in = NULL, rstr_out = NULL; 635 CeedQFunctionField *qf_fields; 636 CeedQFunction qf; 637 CeedOperatorField *op_fields; 638 CeedOperator_Sycl *impl; 639 640 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 641 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 642 CeedCallBackend(CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields)); 643 644 // Determine active input basis 645 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 646 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 647 for (CeedInt i = 0; i < num_input_fields; i++) { 648 CeedVector vec; 649 650 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 651 if (vec == CEED_VECTOR_ACTIVE) { 652 CeedEvalMode e_mode; 653 CeedElemRestriction rstr; 654 655 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_in)); 656 CeedCallBackend(CeedBasisGetNumComponents(basis_in, &num_comp)); 657 CeedCallBackend(CeedBasisGetDimension(basis_in, &dim)); 658 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr)); 659 CeedCheck(!rstr_in || rstr_in == rstr, ceed, CEED_ERROR_BACKEND, 660 "Backend does not implement multi-field non-composite operator diagonal assembly"); 661 rstr_in = rstr; 662 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &e_mode)); 663 switch (e_mode) { 664 case CEED_EVAL_NONE: 665 case CEED_EVAL_INTERP: 666 CeedCallBackend(CeedRealloc(num_e_mode_in + 1, &e_mode_in)); 667 e_mode_in[num_e_mode_in] = e_mode; 668 num_e_mode_in += 1; 669 break; 670 case CEED_EVAL_GRAD: 671 CeedCallBackend(CeedRealloc(num_e_mode_in + dim, &e_mode_in)); 672 for (CeedInt d = 0; d < dim; d++) e_mode_in[num_e_mode_in + d] = e_mode; 673 num_e_mode_in += dim; 674 break; 675 case CEED_EVAL_WEIGHT: 676 case CEED_EVAL_DIV: 677 case CEED_EVAL_CURL: 678 break; // Caught by QF Assembly 679 } 680 } 681 } 682 683 // Determine active output basis 684 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 685 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 686 for (CeedInt i = 0; i < num_output_fields; i++) { 687 CeedVector vec; 688 689 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 690 if (vec == CEED_VECTOR_ACTIVE) { 691 CeedEvalMode e_mode; 692 CeedElemRestriction rstr; 693 694 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_out)); 695 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr)); 696 CeedCheck(!rstr_out || rstr_out == rstr, ceed, CEED_ERROR_BACKEND, 697 "Backend does not implement multi-field non-composite operator diagonal assembly"); 698 rstr_out = rstr; 699 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &e_mode)); 700 switch (e_mode) { 701 case CEED_EVAL_NONE: 702 case CEED_EVAL_INTERP: 703 CeedCallBackend(CeedRealloc(num_e_mode_out + 1, &e_mode_out)); 704 e_mode_out[num_e_mode_out] = e_mode; 705 num_e_mode_out += 1; 706 break; 707 case CEED_EVAL_GRAD: 708 CeedCallBackend(CeedRealloc(num_e_mode_out + dim, &e_mode_out)); 709 for (CeedInt d = 0; d < dim; d++) e_mode_out[num_e_mode_out + d] = e_mode; 710 num_e_mode_out += dim; 711 break; 712 case CEED_EVAL_WEIGHT: 713 case CEED_EVAL_DIV: 714 case CEED_EVAL_CURL: 715 break; // Caught by QF Assembly 716 } 717 } 718 } 719 720 // Operator data struct 721 CeedCallBackend(CeedOperatorGetData(op, &impl)); 722 CeedCallBackend(CeedGetData(ceed, &sycl_data)); 723 CeedCallBackend(CeedCalloc(1, &impl->diag)); 724 CeedOperatorDiag_Sycl *diag = impl->diag; 725 726 diag->basis_in = basis_in; 727 diag->basis_out = basis_out; 728 diag->h_e_mode_in = e_mode_in; 729 diag->h_e_mode_out = e_mode_out; 730 diag->num_e_mode_in = num_e_mode_in; 731 diag->num_e_mode_out = num_e_mode_out; 732 733 // Kernel parameters 734 CeedInt num_nodes, num_qpts; 735 CeedCallBackend(CeedBasisGetNumNodes(basis_in, &num_nodes)); 736 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 737 diag->num_nodes = num_nodes; 738 diag->num_qpts = num_qpts; 739 diag->num_comp = num_comp; 740 741 // Basis matrices 742 const CeedInt i_len = num_qpts * num_nodes; 743 const CeedInt g_len = num_qpts * num_nodes * dim; 744 const CeedScalar *interp_in, *interp_out, *grad_in, *grad_out; 745 746 // CEED_EVAL_NONE 747 CeedScalar *identity = NULL; 748 bool has_eval_none = false; 749 for (CeedInt i = 0; i < num_e_mode_in; i++) has_eval_none = has_eval_none || (e_mode_in[i] == CEED_EVAL_NONE); 750 for (CeedInt i = 0; i < num_e_mode_out; i++) has_eval_none = has_eval_none || (e_mode_out[i] == CEED_EVAL_NONE); 751 752 std::vector<sycl::event> e; 753 754 if (!sycl_data->sycl_queue.is_in_order()) e = {sycl_data->sycl_queue.ext_oneapi_submit_barrier()}; 755 756 std::vector<sycl::event> copy_events; 757 758 if (has_eval_none) { 759 CeedCallBackend(CeedCalloc(num_qpts * num_nodes, &identity)); 760 for (CeedSize i = 0; i < (num_nodes < num_qpts ? num_nodes : num_qpts); i++) identity[i * num_nodes + i] = 1.0; 761 CeedCallSycl(ceed, diag->d_identity = sycl::malloc_device<CeedScalar>(i_len, sycl_data->sycl_device, sycl_data->sycl_context)); 762 sycl::event identity_copy = sycl_data->sycl_queue.copy<CeedScalar>(identity, diag->d_identity, i_len, e); 763 copy_events.push_back(identity_copy); 764 } 765 766 // CEED_EVAL_INTERP 767 CeedCallBackend(CeedBasisGetInterp(basis_in, &interp_in)); 768 CeedCallSycl(ceed, diag->d_interp_in = sycl::malloc_device<CeedScalar>(i_len, sycl_data->sycl_device, sycl_data->sycl_context)); 769 sycl::event interp_in_copy = sycl_data->sycl_queue.copy<CeedScalar>(interp_in, diag->d_interp_in, i_len, e); 770 copy_events.push_back(interp_in_copy); 771 772 CeedCallBackend(CeedBasisGetInterp(basis_out, &interp_out)); 773 CeedCallSycl(ceed, diag->d_interp_out = sycl::malloc_device<CeedScalar>(i_len, sycl_data->sycl_device, sycl_data->sycl_context)); 774 sycl::event interp_out_copy = sycl_data->sycl_queue.copy<CeedScalar>(interp_out, diag->d_interp_out, i_len, e); 775 copy_events.push_back(interp_out_copy); 776 777 // CEED_EVAL_GRAD 778 CeedCallBackend(CeedBasisGetGrad(basis_in, &grad_in)); 779 CeedCallSycl(ceed, diag->d_grad_in = sycl::malloc_device<CeedScalar>(g_len, sycl_data->sycl_device, sycl_data->sycl_context)); 780 sycl::event grad_in_copy = sycl_data->sycl_queue.copy<CeedScalar>(grad_in, diag->d_grad_in, g_len, e); 781 copy_events.push_back(grad_in_copy); 782 783 CeedCallBackend(CeedBasisGetGrad(basis_out, &grad_out)); 784 CeedCallSycl(ceed, diag->d_grad_out = sycl::malloc_device<CeedScalar>(g_len, sycl_data->sycl_device, sycl_data->sycl_context)); 785 sycl::event grad_out_copy = sycl_data->sycl_queue.copy<CeedScalar>(grad_out, diag->d_grad_out, g_len, e); 786 copy_events.push_back(grad_out_copy); 787 788 // Arrays of e_modes 789 CeedCallSycl(ceed, diag->d_e_mode_in = sycl::malloc_device<CeedEvalMode>(num_e_mode_in, sycl_data->sycl_device, sycl_data->sycl_context)); 790 sycl::event e_mode_in_copy = sycl_data->sycl_queue.copy<CeedEvalMode>(e_mode_in, diag->d_e_mode_in, num_e_mode_in, e); 791 copy_events.push_back(e_mode_in_copy); 792 793 CeedCallSycl(ceed, diag->d_e_mode_out = sycl::malloc_device<CeedEvalMode>(num_e_mode_out, sycl_data->sycl_device, sycl_data->sycl_context)); 794 sycl::event e_mode_out_copy = sycl_data->sycl_queue.copy<CeedEvalMode>(e_mode_out, diag->d_e_mode_out, num_e_mode_out, e); 795 copy_events.push_back(e_mode_out_copy); 796 797 // Restriction 798 diag->diag_rstr = rstr_out; 799 800 // Wait for all copies to complete and handle exceptions 801 CeedCallSycl(ceed, sycl::event::wait_and_throw(copy_events)); 802 return CEED_ERROR_SUCCESS; 803 } 804 805 //------------------------------------------------------------------------------ 806 // Kernel for diagonal assembly 807 //------------------------------------------------------------------------------ 808 static int CeedOperatorLinearDiagonal_Sycl(sycl::queue &sycl_queue, const bool is_point_block, const CeedInt num_elem, 809 const CeedOperatorDiag_Sycl *diag, const CeedScalar *assembled_qf_array, CeedScalar *elem_diag_array) { 810 const CeedSize num_nodes = diag->num_nodes; 811 const CeedSize num_qpts = diag->num_qpts; 812 const CeedSize num_comp = diag->num_comp; 813 const CeedSize num_e_mode_in = diag->num_e_mode_in; 814 const CeedSize num_e_mode_out = diag->num_e_mode_out; 815 const CeedScalar *identity = diag->d_identity; 816 const CeedScalar *interp_in = diag->d_interp_in; 817 const CeedScalar *grad_in = diag->d_grad_in; 818 const CeedScalar *interp_out = diag->d_interp_out; 819 const CeedScalar *grad_out = diag->d_grad_out; 820 const CeedEvalMode *e_mode_in = diag->d_e_mode_in; 821 const CeedEvalMode *e_mode_out = diag->d_e_mode_out; 822 823 sycl::range<1> kernel_range(num_elem * num_nodes); 824 825 std::vector<sycl::event> e; 826 827 if (!sycl_queue.is_in_order()) e = {sycl_queue.ext_oneapi_submit_barrier()}; 828 829 sycl_queue.parallel_for<CeedOperatorSyclLinearDiagonal>(kernel_range, e, [=](sycl::id<1> idx) { 830 const CeedInt tid = idx % num_nodes; 831 const CeedInt e = idx / num_nodes; 832 833 // Compute the diagonal of B^T D B 834 // Each element 835 CeedInt d_out = -1; 836 // Each basis eval mode pair 837 for (CeedSize e_out = 0; e_out < num_e_mode_out; e_out++) { 838 const CeedScalar *bt = NULL; 839 840 if (e_mode_out[e_out] == CEED_EVAL_GRAD) ++d_out; 841 CeedOperatorGetBasisPointer_Sycl(&bt, e_mode_out[e_out], identity, interp_out, &grad_out[d_out * num_qpts * num_nodes]); 842 CeedInt d_in = -1; 843 844 for (CeedSize e_in = 0; e_in < num_e_mode_in; e_in++) { 845 const CeedScalar *b = NULL; 846 847 if (e_mode_in[e_in] == CEED_EVAL_GRAD) ++d_in; 848 CeedOperatorGetBasisPointer_Sycl(&b, e_mode_in[e_in], identity, interp_in, &grad_in[d_in * num_qpts * num_nodes]); 849 // Each component 850 for (CeedSize comp_out = 0; comp_out < num_comp; comp_out++) { 851 // Each qpoint/node pair 852 if (is_point_block) { 853 // Point Block Diagonal 854 for (CeedInt comp_in = 0; comp_in < num_comp; comp_in++) { 855 CeedScalar e_value = 0.0; 856 857 for (CeedSize q = 0; q < num_qpts; q++) { 858 const CeedScalar qf_value = 859 assembled_qf_array[((((e_in * num_comp + comp_in) * num_e_mode_out + e_out) * num_comp + comp_out) * num_elem + e) * num_qpts + 860 q]; 861 862 e_value += bt[q * num_nodes + tid] * qf_value * b[q * num_nodes + tid]; 863 } 864 elem_diag_array[((comp_out * num_comp + comp_in) * num_elem + e) * num_nodes + tid] += e_value; 865 } 866 } else { 867 // Diagonal Only 868 CeedScalar e_value = 0.0; 869 870 for (CeedSize q = 0; q < num_qpts; q++) { 871 const CeedScalar qf_value = 872 assembled_qf_array[((((e_in * num_comp + comp_out) * num_e_mode_out + e_out) * num_comp + comp_out) * num_elem + e) * num_qpts + q]; 873 e_value += bt[q * num_nodes + tid] * qf_value * b[q * num_nodes + tid]; 874 } 875 elem_diag_array[(comp_out * num_elem + e) * num_nodes + tid] += e_value; 876 } 877 } 878 } 879 } 880 }); 881 return CEED_ERROR_SUCCESS; 882 } 883 884 //------------------------------------------------------------------------------ 885 // Assemble diagonal common code 886 //------------------------------------------------------------------------------ 887 static inline int CeedOperatorAssembleDiagonalCore_Sycl(CeedOperator op, CeedVector assembled, CeedRequest *request, const bool is_point_block) { 888 Ceed ceed; 889 Ceed_Sycl *sycl_data; 890 CeedInt num_elem; 891 CeedScalar *elem_diag_array; 892 const CeedScalar *assembled_qf_array; 893 CeedVector assembled_qf = NULL; 894 CeedElemRestriction rstr = NULL; 895 CeedOperator_Sycl *impl; 896 897 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 898 CeedCallBackend(CeedOperatorGetData(op, &impl)); 899 CeedCallBackend(CeedGetData(ceed, &sycl_data)); 900 901 // Assemble QFunction 902 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &rstr, request)); 903 CeedCallBackend(CeedElemRestrictionDestroy(&rstr)); 904 905 // Setup 906 if (!impl->diag) { 907 CeedCallBackend(CeedOperatorAssembleDiagonalSetup_Sycl(op)); 908 } 909 CeedOperatorDiag_Sycl *diag = impl->diag; 910 911 assert(diag != NULL); 912 913 // Restriction 914 if (is_point_block && !diag->point_block_diag_rstr) { 915 CeedCallBackend(CeedOperatorCreateActivePointBlockRestriction(diag->diag_rstr, &diag->point_block_diag_rstr)); 916 } 917 CeedElemRestriction diag_rstr = is_point_block ? diag->point_block_diag_rstr : diag->diag_rstr; 918 919 // Create diagonal vector 920 CeedVector elem_diag = is_point_block ? diag->point_block_elem_diag : diag->elem_diag; 921 922 if (!elem_diag) { 923 CeedCallBackend(CeedElemRestrictionCreateVector(diag_rstr, NULL, &elem_diag)); 924 if (is_point_block) diag->point_block_elem_diag = elem_diag; 925 else diag->elem_diag = elem_diag; 926 } 927 CeedCallBackend(CeedVectorSetValue(elem_diag, 0.0)); 928 929 // Assemble element operator diagonals 930 CeedCallBackend(CeedVectorGetArray(elem_diag, CEED_MEM_DEVICE, &elem_diag_array)); 931 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &assembled_qf_array)); 932 CeedCallBackend(CeedElemRestrictionGetNumElements(diag_rstr, &num_elem)); 933 934 // Compute the diagonal of B^T D B 935 CeedCallBackend(CeedOperatorLinearDiagonal_Sycl(sycl_data->sycl_queue, is_point_block, num_elem, diag, assembled_qf_array, elem_diag_array)); 936 937 // Wait for queue to complete and handle exceptions 938 sycl_data->sycl_queue.wait_and_throw(); 939 940 // Restore arrays 941 CeedCallBackend(CeedVectorRestoreArray(elem_diag, &elem_diag_array)); 942 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 943 944 // Assemble local operator diagonal 945 CeedCallBackend(CeedElemRestrictionApply(diag_rstr, CEED_TRANSPOSE, elem_diag, assembled, request)); 946 947 // Cleanup 948 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 949 return CEED_ERROR_SUCCESS; 950 } 951 952 //------------------------------------------------------------------------------ 953 // Assemble Linear Diagonal 954 //------------------------------------------------------------------------------ 955 static int CeedOperatorLinearAssembleAddDiagonal_Sycl(CeedOperator op, CeedVector assembled, CeedRequest *request) { 956 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Sycl(op, assembled, request, false)); 957 return CEED_ERROR_SUCCESS; 958 } 959 960 //------------------------------------------------------------------------------ 961 // Assemble Linear Point Block Diagonal 962 //------------------------------------------------------------------------------ 963 static int CeedOperatorLinearAssembleAddPointBlockDiagonal_Sycl(CeedOperator op, CeedVector assembled, CeedRequest *request) { 964 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Sycl(op, assembled, request, true)); 965 return CEED_ERROR_SUCCESS; 966 } 967 968 //------------------------------------------------------------------------------ 969 // Single operator assembly setup 970 //------------------------------------------------------------------------------ 971 static int CeedSingleOperatorAssembleSetup_Sycl(CeedOperator op) { 972 Ceed ceed; 973 CeedInt num_input_fields, num_output_fields, num_e_mode_in = 0, dim = 1, num_B_in_mats_to_load = 0, size_B_in = 0, num_e_mode_out = 0, 974 num_B_out_mats_to_load = 0, size_B_out = 0, num_qpts = 0, elem_size = 0, num_elem, num_comp, 975 mat_start = 0; 976 CeedEvalMode *eval_mode_in = NULL, *eval_mode_out = NULL; 977 const CeedScalar *interp_in, *grad_in; 978 CeedElemRestriction rstr_in = NULL, rstr_out = NULL; 979 CeedBasis basis_in = NULL, basis_out = NULL; 980 CeedQFunctionField *qf_fields; 981 CeedQFunction qf; 982 CeedOperatorField *input_fields, *output_fields; 983 CeedOperator_Sycl *impl; 984 985 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 986 CeedCallBackend(CeedOperatorGetData(op, &impl)); 987 988 // Get input and output fields 989 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 990 991 // Determine active input basis eval mode 992 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 993 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 994 // Note that the kernel will treat each dimension of a gradient action separately; 995 // i.e., when an active input has a CEED_EVAL_GRAD mode, num_ e_mode_in will increment by dim. 996 // However, for the purposes of load_ing the B matrices, it will be treated as one mode, and we will load/copy the entire gradient matrix at once, 997 // so num_B_in_mats_to_load will be incremented by 1. 998 for (CeedInt i = 0; i < num_input_fields; i++) { 999 CeedEvalMode eval_mode; 1000 CeedVector vec; 1001 1002 CeedCallBackend(CeedOperatorFieldGetVector(input_fields[i], &vec)); 1003 if (vec == CEED_VECTOR_ACTIVE) { 1004 CeedCallBackend(CeedOperatorFieldGetBasis(input_fields[i], &basis_in)); 1005 CeedCallBackend(CeedBasisGetDimension(basis_in, &dim)); 1006 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 1007 CeedCallBackend(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr_in)); 1008 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_in, &elem_size)); 1009 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1010 if (eval_mode != CEED_EVAL_NONE) { 1011 CeedCallBackend(CeedRealloc(num_B_in_mats_to_load + 1, &eval_mode_in)); 1012 eval_mode_in[num_B_in_mats_to_load] = eval_mode; 1013 num_B_in_mats_to_load += 1; 1014 if (eval_mode == CEED_EVAL_GRAD) { 1015 num_e_mode_in += dim; 1016 size_B_in += dim * elem_size * num_qpts; 1017 } else { 1018 num_e_mode_in += 1; 1019 size_B_in += elem_size * num_qpts; 1020 } 1021 } 1022 } 1023 } 1024 1025 // Determine active output basis; basis_out and rstr_out only used if same as input, TODO 1026 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 1027 for (CeedInt i = 0; i < num_output_fields; i++) { 1028 CeedEvalMode eval_mode; 1029 CeedVector vec; 1030 1031 CeedCallBackend(CeedOperatorFieldGetVector(output_fields[i], &vec)); 1032 if (vec == CEED_VECTOR_ACTIVE) { 1033 CeedCallBackend(CeedOperatorFieldGetBasis(output_fields[i], &basis_out)); 1034 CeedCallBackend(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr_out)); 1035 CeedCheck(!rstr_out || rstr_out == rstr_in, ceed, CEED_ERROR_BACKEND, "Backend does not implement multi-field non-composite operator assembly"); 1036 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1037 if (eval_mode != CEED_EVAL_NONE) { 1038 CeedCallBackend(CeedRealloc(num_B_out_mats_to_load + 1, &eval_mode_out)); 1039 eval_mode_out[num_B_out_mats_to_load] = eval_mode; 1040 num_B_out_mats_to_load += 1; 1041 if (eval_mode == CEED_EVAL_GRAD) { 1042 num_e_mode_out += dim; 1043 size_B_out += dim * elem_size * num_qpts; 1044 } else { 1045 num_e_mode_out += 1; 1046 size_B_out += elem_size * num_qpts; 1047 } 1048 } 1049 } 1050 } 1051 CeedCheck(num_e_mode_in > 0 && num_e_mode_out > 0, ceed, CEED_ERROR_UNSUPPORTED, "Cannot assemble operator without inputs/outputs"); 1052 1053 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr_in, &num_elem)); 1054 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr_in, &num_comp)); 1055 1056 CeedCallBackend(CeedCalloc(1, &impl->asmb)); 1057 CeedOperatorAssemble_Sycl *asmb = impl->asmb; 1058 asmb->num_elem = num_elem; 1059 1060 Ceed_Sycl *sycl_data; 1061 CeedCallBackend(CeedGetData(ceed, &sycl_data)); 1062 1063 // Kernel setup 1064 int elems_per_block = 1; 1065 asmb->elems_per_block = elems_per_block; 1066 asmb->block_size_x = elem_size; 1067 asmb->block_size_y = elem_size; 1068 asmb->num_e_mode_in = num_e_mode_in; 1069 asmb->num_e_mode_out = num_e_mode_out; 1070 asmb->num_qpts = num_qpts; 1071 asmb->num_nodes = elem_size; 1072 asmb->block_size = elem_size * elem_size * elems_per_block; 1073 asmb->num_comp = num_comp; 1074 1075 // Build 'full' B matrices (not 1D arrays used for tensor-product matrices 1076 CeedCallBackend(CeedBasisGetInterp(basis_in, &interp_in)); 1077 CeedCallBackend(CeedBasisGetGrad(basis_in, &grad_in)); 1078 1079 // Load into B_in, in order that they will be used in eval_mode 1080 CeedCallSycl(ceed, asmb->d_B_in = sycl::malloc_device<CeedScalar>(size_B_in, sycl_data->sycl_device, sycl_data->sycl_context)); 1081 for (int i = 0; i < num_B_in_mats_to_load; i++) { 1082 CeedEvalMode eval_mode = eval_mode_in[i]; 1083 1084 if (eval_mode == CEED_EVAL_INTERP) { 1085 std::vector<sycl::event> e; 1086 1087 if (!sycl_data->sycl_queue.is_in_order()) e = {sycl_data->sycl_queue.ext_oneapi_submit_barrier()}; 1088 sycl_data->sycl_queue.copy<CeedScalar>(interp_in, &asmb->d_B_in[mat_start], elem_size * num_qpts, e); 1089 mat_start += elem_size * num_qpts; 1090 } else if (eval_mode == CEED_EVAL_GRAD) { 1091 std::vector<sycl::event> e; 1092 1093 if (!sycl_data->sycl_queue.is_in_order()) e = {sycl_data->sycl_queue.ext_oneapi_submit_barrier()}; 1094 sycl_data->sycl_queue.copy<CeedScalar>(grad_in, &asmb->d_B_in[mat_start], dim * elem_size * num_qpts, e); 1095 mat_start += dim * elem_size * num_qpts; 1096 } 1097 } 1098 1099 const CeedScalar *interp_out, *grad_out; 1100 // Note that this function currently assumes 1 basis, so this should always be true 1101 // for now 1102 if (basis_out == basis_in) { 1103 interp_out = interp_in; 1104 grad_out = grad_in; 1105 } else { 1106 CeedCallBackend(CeedBasisGetInterp(basis_out, &interp_out)); 1107 CeedCallBackend(CeedBasisGetGrad(basis_out, &grad_out)); 1108 } 1109 1110 // Load into B_out, in order that they will be used in eval_mode 1111 mat_start = 0; 1112 CeedCallSycl(ceed, asmb->d_B_out = sycl::malloc_device<CeedScalar>(size_B_out, sycl_data->sycl_device, sycl_data->sycl_context)); 1113 for (int i = 0; i < num_B_out_mats_to_load; i++) { 1114 CeedEvalMode eval_mode = eval_mode_out[i]; 1115 1116 if (eval_mode == CEED_EVAL_INTERP) { 1117 std::vector<sycl::event> e; 1118 1119 if (!sycl_data->sycl_queue.is_in_order()) e = {sycl_data->sycl_queue.ext_oneapi_submit_barrier()}; 1120 sycl_data->sycl_queue.copy<CeedScalar>(interp_out, &asmb->d_B_out[mat_start], elem_size * num_qpts, e); 1121 mat_start += elem_size * num_qpts; 1122 } else if (eval_mode == CEED_EVAL_GRAD) { 1123 std::vector<sycl::event> e; 1124 1125 if (!sycl_data->sycl_queue.is_in_order()) e = {sycl_data->sycl_queue.ext_oneapi_submit_barrier()}; 1126 sycl_data->sycl_queue.copy<CeedScalar>(grad_out, &asmb->d_B_out[mat_start], dim * elem_size * num_qpts, e); 1127 mat_start += dim * elem_size * num_qpts; 1128 } 1129 } 1130 return CEED_ERROR_SUCCESS; 1131 } 1132 1133 //------------------------------------------------------------------------------ 1134 // Matrix assembly kernel for low-order elements (3D thread block) 1135 //------------------------------------------------------------------------------ 1136 static int CeedOperatorLinearAssemble_Sycl(sycl::queue &sycl_queue, const CeedOperator_Sycl *impl, const CeedScalar *qf_array, 1137 CeedScalar *values_array) { 1138 // This kernels assumes B_in and B_out have the same number of quadrature points and basis points. 1139 // TODO: expand to more general cases 1140 CeedOperatorAssemble_Sycl *asmb = impl->asmb; 1141 const CeedInt num_elem = asmb->num_elem; 1142 const CeedSize num_nodes = asmb->num_nodes; 1143 const CeedSize num_comp = asmb->num_comp; 1144 const CeedSize num_qpts = asmb->num_qpts; 1145 const CeedSize num_e_mode_in = asmb->num_e_mode_in; 1146 const CeedSize num_e_mode_out = asmb->num_e_mode_out; 1147 1148 // Strides for final output ordering, determined by the reference (inference) implementation of the symbolic assembly, slowest --> fastest: element, 1149 // comp_in, comp_out, node_row, node_col 1150 const CeedSize comp_out_stride = num_nodes * num_nodes; 1151 const CeedSize comp_in_stride = comp_out_stride * num_comp; 1152 const CeedSize e_stride = comp_in_stride * num_comp; 1153 // Strides for QF array, slowest --> fastest: e_mode_in, comp_in, e_mode_out, comp_out, elem, qpt 1154 const CeedSize q_e_stride = num_qpts; 1155 const CeedSize q_comp_out_stride = num_elem * q_e_stride; 1156 const CeedSize q_e_mode_out_stride = q_comp_out_stride * num_comp; 1157 const CeedSize q_comp_in_stride = q_e_mode_out_stride * num_e_mode_out; 1158 const CeedSize q_e_mode_in_stride = q_comp_in_stride * num_comp; 1159 1160 CeedScalar *B_in, *B_out; 1161 B_in = asmb->d_B_in; 1162 B_out = asmb->d_B_out; 1163 const CeedInt block_size_x = asmb->block_size_x; 1164 const CeedInt block_size_y = asmb->block_size_y; 1165 1166 sycl::range<3> kernel_range(num_elem, block_size_y, block_size_x); 1167 1168 std::vector<sycl::event> e; 1169 1170 if (!sycl_queue.is_in_order()) e = {sycl_queue.ext_oneapi_submit_barrier()}; 1171 sycl_queue.parallel_for<CeedOperatorSyclLinearAssemble>(kernel_range, e, [=](sycl::id<3> idx) { 1172 const int e = idx.get(0); // Element index 1173 const int l = idx.get(1); // The output column index of each B^TDB operation 1174 const int i = idx.get(2); // The output row index of each B^TDB operation 1175 // such that we have (Bout^T)_ij D_jk Bin_kl = C_il 1176 for (CeedSize comp_in = 0; comp_in < num_comp; comp_in++) { 1177 for (CeedSize comp_out = 0; comp_out < num_comp; comp_out++) { 1178 CeedScalar result = 0.0; 1179 CeedSize qf_index_comp = q_comp_in_stride * comp_in + q_comp_out_stride * comp_out + q_e_stride * e; 1180 1181 for (CeedSize e_mode_in = 0; e_mode_in < num_e_mode_in; e_mode_in++) { 1182 CeedSize b_in_index = e_mode_in * num_qpts * num_nodes; 1183 1184 for (CeedSize e_mode_out = 0; e_mode_out < num_e_mode_out; e_mode_out++) { 1185 CeedSize b_out_index = e_mode_out * num_qpts * num_nodes; 1186 CeedSize qf_index = qf_index_comp + q_e_mode_out_stride * e_mode_out + q_e_mode_in_stride * e_mode_in; 1187 1188 // Perform the B^T D B operation for this 'chunk' of D (the qf_array) 1189 for (CeedSize j = 0; j < num_qpts; j++) { 1190 result += B_out[b_out_index + j * num_nodes + i] * qf_array[qf_index + j] * B_in[b_in_index + j * num_nodes + l]; 1191 } 1192 } // end of e_mode_out 1193 } // end of e_mode_in 1194 CeedSize val_index = comp_in_stride * comp_in + comp_out_stride * comp_out + e_stride * e + num_nodes * i + l; 1195 1196 values_array[val_index] = result; 1197 } // end of out component 1198 } // end of in component 1199 }); 1200 return CEED_ERROR_SUCCESS; 1201 } 1202 1203 //------------------------------------------------------------------------------ 1204 // Fallback kernel for larger orders (1D thread block) 1205 //------------------------------------------------------------------------------ 1206 /* 1207 static int CeedOperatorLinearAssembleFallback_Sycl(sycl::queue &sycl_queue, const CeedOperator_Sycl *impl, const CeedScalar *qf_array, 1208 CeedScalar *values_array) { 1209 // This kernel assumes B_in and B_out have the same number of quadrature points and basis points. 1210 // TODO: expand to more general cases 1211 CeedOperatorAssemble_Sycl *asmb = impl->asmb; 1212 const CeedInt num_elem = asmb->num_elem; 1213 const CeedInt num_nodes = asmb->num_nodes; 1214 const CeedInt num_comp = asmb->num_comp; 1215 const CeedInt num_qpts = asmb->num_qpts; 1216 const CeedInt num_e_mode_in = asmb->num_e_mode_in; 1217 const CeedInt num_e_mode_out = asmb->num_e_mode_out; 1218 1219 // Strides for final output ordering, determined by the reference (interface) implementation of the symbolic assembly, slowest --> fastest: elememt, 1220 // comp_in, comp_out, node_row, node_col 1221 const CeedInt comp_out_stride = num_nodes * num_nodes; 1222 const CeedInt comp_in_stride = comp_out_stride * num_comp; 1223 const CeedInt e_stride = comp_in_stride * num_comp; 1224 // Strides for QF array, slowest --> fastest: e_mode_in, comp_in, e_mode_out, comp_out, elem, qpt 1225 const CeedInt q_e_stride = num_qpts; 1226 const CeedInt q_comp_out_stride = num_elem * q_e_stride; 1227 const CeedInt q_e_mode_out_stride = q_comp_out_stride * num_comp; 1228 const CeedInt q_comp_in_stride = q_e_mode_out_stride * num_e_mode_out; 1229 const CeedInt q_e_mode_in_stride = q_comp_in_stride * num_comp; 1230 1231 CeedScalar *B_in, *B_out; 1232 B_in = asmb->d_B_in; 1233 B_out = asmb->d_B_out; 1234 const CeedInt elems_per_block = asmb->elems_per_block; 1235 const CeedInt block_size_x = asmb->block_size_x; 1236 const CeedInt block_size_y = asmb->block_size_y; // This will be 1 for the fallback kernel 1237 1238 const CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 1239 sycl::range<3> local_range(block_size_x, block_size_y, elems_per_block); 1240 sycl::range<3> global_range(grid * block_size_x, block_size_y, elems_per_block); 1241 sycl::nd_range<3> kernel_range(global_range, local_range); 1242 1243 sycl_queue.parallel_for<CeedOperatorSyclLinearAssembleFallback>(kernel_range, [=](sycl::nd_item<3> work_item) { 1244 const CeedInt blockIdx = work_item.get_group(0); 1245 const CeedInt gridDimx = work_item.get_group_range(0); 1246 const CeedInt threadIdx = work_item.get_local_id(0); 1247 const CeedInt threadIdz = work_item.get_local_id(2); 1248 const CeedInt blockDimz = work_item.get_local_range(2); 1249 1250 const int l = threadIdx; // The output column index of each B^TDB operation 1251 // such that we have (Bout^T)_ij D_jk Bin_kl = C_il 1252 for (CeedInt e = blockIdx * blockDimz + threadIdz; e < num_elem; e += gridDimx * blockDimz) { 1253 for (CeedInt comp_in = 0; comp_in < num_comp; comp_in++) { 1254 for (CeedInt comp_out = 0; comp_out < num_comp; comp_out++) { 1255 for (CeedInt i = 0; i < num_nodes; i++) { 1256 CeedScalar result = 0.0; 1257 CeedInt qf_index_comp = q_comp_in_stride * comp_in + q_comp_out_stride * comp_out + q_e_stride * e; 1258 for (CeedInt e_mode_in = 0; e_mode_in < num_e_mode_in; e_mode_in++) { 1259 CeedInt b_in_index = e_mode_in * num_qpts * num_nodes; 1260 for (CeedInt e_mode_out = 0; e_mode_out < num_e_mode_out; e_mode_out++) { 1261 CeedInt b_out_index = e_mode_out * num_qpts * num_nodes; 1262 CeedInt qf_index = qf_index_comp + q_e_mode_out_stride * e_mode_out + q_e_mode_in_stride * e_mode_in; 1263 // Perform the B^T D B operation for this 'chunk' of D (the qf_array) 1264 for (CeedInt j = 0; j < num_qpts; j++) { 1265 result += B_out[b_out_index + j * num_nodes + i] * qf_array[qf_index + j] * B_in[b_in_index + j * num_nodes + l]; 1266 } 1267 } // end of e_mode_out 1268 } // end of e_mode_in 1269 CeedInt val_index = comp_in_stride * comp_in + comp_out_stride * comp_out + e_stride * e + num_nodes * i + l; 1270 values_array[val_index] = result; 1271 } // end of loop over element node index, i 1272 } // end of out component 1273 } // end of in component 1274 } // end of element loop 1275 }); 1276 return CEED_ERROR_SUCCESS; 1277 }*/ 1278 1279 //------------------------------------------------------------------------------ 1280 // Assemble matrix data for COO matrix of assembled operator. 1281 // The sparsity pattern is set by CeedOperatorLinearAssembleSymbolic. 1282 // 1283 // Note that this (and other assembly routines) currently assume only one active 1284 // input restriction/basis per operator (could have multiple basis eval modes). 1285 // TODO: allow multiple active input restrictions/basis objects 1286 //------------------------------------------------------------------------------ 1287 static int CeedSingleOperatorAssemble_Sycl(CeedOperator op, CeedInt offset, CeedVector values) { 1288 Ceed ceed; 1289 Ceed_Sycl *sycl_data; 1290 CeedScalar *values_array; 1291 const CeedScalar *qf_array; 1292 CeedVector assembled_qf = NULL; 1293 CeedElemRestriction rstr_q = NULL; 1294 CeedOperator_Sycl *impl; 1295 1296 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1297 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1298 CeedCallBackend(CeedGetData(ceed, &sycl_data)); 1299 1300 // Setup 1301 if (!impl->asmb) { 1302 CeedCallBackend(CeedSingleOperatorAssembleSetup_Sycl(op)); 1303 assert(impl->asmb != NULL); 1304 } 1305 1306 // Assemble QFunction 1307 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &rstr_q, CEED_REQUEST_IMMEDIATE)); 1308 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_q)); 1309 CeedCallBackend(CeedVectorGetArrayWrite(values, CEED_MEM_DEVICE, &values_array)); 1310 values_array += offset; 1311 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &qf_array)); 1312 1313 // Compute B^T D B 1314 CeedCallBackend(CeedOperatorLinearAssemble_Sycl(sycl_data->sycl_queue, impl, qf_array, values_array)); 1315 1316 // Wait for kernels to be completed 1317 // Kris: Review if this is necessary -- enqueing an async barrier may be sufficient 1318 sycl_data->sycl_queue.wait_and_throw(); 1319 1320 // Restore arrays 1321 CeedCallBackend(CeedVectorRestoreArray(values, &values_array)); 1322 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &qf_array)); 1323 1324 // Cleanup 1325 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 1326 return CEED_ERROR_SUCCESS; 1327 } 1328 1329 //------------------------------------------------------------------------------ 1330 // Create operator 1331 //------------------------------------------------------------------------------ 1332 int CeedOperatorCreate_Sycl(CeedOperator op) { 1333 Ceed ceed; 1334 CeedOperator_Sycl *impl; 1335 1336 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1337 1338 CeedCallBackend(CeedCalloc(1, &impl)); 1339 CeedCallBackend(CeedOperatorSetData(op, impl)); 1340 1341 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunction_Sycl)); 1342 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Operator", op, "LinearAssembleQFunctionUpdate", CeedOperatorLinearAssembleQFunctionUpdate_Sycl)); 1343 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Operator", op, "LinearAssembleAddDiagonal", CeedOperatorLinearAssembleAddDiagonal_Sycl)); 1344 CeedCallBackend( 1345 CeedSetBackendFunctionCpp(ceed, "Operator", op, "LinearAssembleAddPointBlockDiagonal", CeedOperatorLinearAssembleAddPointBlockDiagonal_Sycl)); 1346 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Operator", op, "LinearAssembleSingle", CeedSingleOperatorAssemble_Sycl)); 1347 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Sycl)); 1348 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Sycl)); 1349 return CEED_ERROR_SUCCESS; 1350 } 1351 1352 //------------------------------------------------------------------------------ 1353