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 26 CeedCallBackend(CeedOperatorGetData(op, &impl)); 27 28 // Apply data 29 for (CeedInt i = 0; i < impl->num_inputs + impl->num_outputs; i++) { 30 CeedCallBackend(CeedVectorDestroy(&impl->e_vecs[i])); 31 } 32 CeedCallBackend(CeedFree(&impl->e_vecs)); 33 34 for (CeedInt i = 0; i < impl->num_inputs; i++) { 35 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_in[i])); 36 } 37 CeedCallBackend(CeedFree(&impl->q_vecs_in)); 38 39 for (CeedInt i = 0; i < impl->num_outputs; i++) { 40 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_out[i])); 41 } 42 CeedCallBackend(CeedFree(&impl->q_vecs_out)); 43 44 // QFunction assembly data 45 for (CeedInt i = 0; i < impl->num_active_in; i++) { 46 CeedCallBackend(CeedVectorDestroy(&impl->qf_active_in[i])); 47 } 48 CeedCallBackend(CeedFree(&impl->qf_active_in)); 49 50 // Diag data 51 if (impl->diag) { 52 Ceed ceed; 53 54 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 55 CeedCallHip(ceed, hipModuleUnload(impl->diag->module)); 56 CeedCallBackend(CeedFree(&impl->diag->h_e_mode_in)); 57 CeedCallBackend(CeedFree(&impl->diag->h_e_mode_out)); 58 CeedCallHip(ceed, hipFree(impl->diag->d_e_mode_in)); 59 CeedCallHip(ceed, hipFree(impl->diag->d_e_mode_out)); 60 CeedCallHip(ceed, hipFree(impl->diag->d_identity)); 61 CeedCallHip(ceed, hipFree(impl->diag->d_interp_in)); 62 CeedCallHip(ceed, hipFree(impl->diag->d_interp_out)); 63 CeedCallHip(ceed, hipFree(impl->diag->d_grad_in)); 64 CeedCallHip(ceed, hipFree(impl->diag->d_grad_out)); 65 CeedCallBackend(CeedElemRestrictionDestroy(&impl->diag->point_block_diag_rstr)); 66 CeedCallBackend(CeedVectorDestroy(&impl->diag->elem_diag)); 67 CeedCallBackend(CeedVectorDestroy(&impl->diag->point_block_elem_diag)); 68 } 69 CeedCallBackend(CeedFree(&impl->diag)); 70 71 if (impl->asmb) { 72 Ceed ceed; 73 74 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 75 CeedCallHip(ceed, hipModuleUnload(impl->asmb->module)); 76 CeedCallHip(ceed, hipFree(impl->asmb->d_B_in)); 77 CeedCallHip(ceed, hipFree(impl->asmb->d_B_out)); 78 } 79 CeedCallBackend(CeedFree(&impl->asmb)); 80 81 CeedCallBackend(CeedFree(&impl)); 82 return CEED_ERROR_SUCCESS; 83 } 84 85 //------------------------------------------------------------------------------ 86 // Setup infields or outfields 87 //------------------------------------------------------------------------------ 88 static int CeedOperatorSetupFields_Hip(CeedQFunction qf, CeedOperator op, bool is_input, CeedVector *e_vecs, CeedVector *q_vecs, CeedInt start_e, 89 CeedInt num_fields, CeedInt Q, CeedInt num_elem) { 90 Ceed ceed; 91 CeedQFunctionField *qf_fields; 92 CeedOperatorField *op_fields; 93 94 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 95 if (is_input) { 96 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 97 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 98 } else { 99 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 100 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 101 } 102 103 // Loop over fields 104 for (CeedInt i = 0; i < num_fields; i++) { 105 bool is_strided, skip_restriction; 106 CeedSize q_size; 107 CeedInt dim, size; 108 CeedEvalMode e_mode; 109 CeedVector vec; 110 CeedElemRestriction elem_rstr; 111 CeedBasis basis; 112 113 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &e_mode)); 114 is_strided = false; 115 skip_restriction = false; 116 if (e_mode != CEED_EVAL_WEIGHT) { 117 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &elem_rstr)); 118 119 // Check whether this field can skip the element restriction: 120 // must be passive input, with e_mode NONE, and have a strided restriction with CEED_STRIDES_BACKEND. 121 122 // First, check whether the field is input or output: 123 if (is_input) { 124 // Check for passive input: 125 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 126 if (vec != CEED_VECTOR_ACTIVE) { 127 // Check e_mode 128 if (e_mode == CEED_EVAL_NONE) { 129 // Check for strided restriction 130 CeedCallBackend(CeedElemRestrictionIsStrided(elem_rstr, &is_strided)); 131 if (is_strided) { 132 // Check if vector is already in preferred backend ordering 133 CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_rstr, &skip_restriction)); 134 } 135 } 136 } 137 } 138 if (skip_restriction) { 139 // We do not need an E-Vector, but will use the input field vector's data directly in the operator application. 140 e_vecs[i + start_e] = NULL; 141 } else { 142 CeedCallBackend(CeedElemRestrictionCreateVector(elem_rstr, NULL, &e_vecs[i + start_e])); 143 } 144 } 145 146 switch (e_mode) { 147 case CEED_EVAL_NONE: 148 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 149 q_size = (CeedSize)num_elem * Q * size; 150 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 151 break; 152 case CEED_EVAL_INTERP: 153 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 154 q_size = (CeedSize)num_elem * Q * size; 155 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 156 break; 157 case CEED_EVAL_GRAD: 158 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 159 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 160 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 161 q_size = (CeedSize)num_elem * Q * size; 162 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 163 break; 164 case CEED_EVAL_WEIGHT: // Only on input fields 165 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 166 q_size = (CeedSize)num_elem * Q; 167 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 168 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, q_vecs[i])); 169 break; 170 case CEED_EVAL_DIV: 171 break; // TODO: Not implemented 172 case CEED_EVAL_CURL: 173 break; // TODO: Not implemented 174 } 175 } 176 return CEED_ERROR_SUCCESS; 177 } 178 179 //------------------------------------------------------------------------------ 180 // CeedOperator needs to connect all the named fields (be they active or passive) to the named inputs and outputs of its CeedQFunction. 181 //------------------------------------------------------------------------------ 182 static int CeedOperatorSetup_Hip(CeedOperator op) { 183 Ceed ceed; 184 bool is_setup_done; 185 CeedInt Q, num_elem, num_input_fields, num_output_fields; 186 CeedQFunctionField *qf_input_fields, *qf_output_fields; 187 CeedQFunction qf; 188 CeedOperatorField *op_input_fields, *op_output_fields; 189 CeedOperator_Hip *impl; 190 191 CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done)); 192 if (is_setup_done) return CEED_ERROR_SUCCESS; 193 194 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 195 CeedCallBackend(CeedOperatorGetData(op, &impl)); 196 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 197 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 198 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 199 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 200 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 201 202 // Allocate 203 CeedCallBackend(CeedCalloc(num_input_fields + num_output_fields, &impl->e_vecs)); 204 205 CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_in)); 206 CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_out)); 207 208 impl->num_inputs = num_input_fields; 209 impl->num_outputs = num_output_fields; 210 211 // Set up infield and outfield e_vecs and q_vecs 212 // Infields 213 CeedCallBackend(CeedOperatorSetupFields_Hip(qf, op, true, impl->e_vecs, impl->q_vecs_in, 0, num_input_fields, Q, num_elem)); 214 215 // Outfields 216 CeedCallBackend(CeedOperatorSetupFields_Hip(qf, op, false, impl->e_vecs, impl->q_vecs_out, num_input_fields, num_output_fields, Q, num_elem)); 217 218 CeedCallBackend(CeedOperatorSetSetupDone(op)); 219 return CEED_ERROR_SUCCESS; 220 } 221 222 //------------------------------------------------------------------------------ 223 // Setup Operator Inputs 224 //------------------------------------------------------------------------------ 225 static inline int CeedOperatorSetupInputs_Hip(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields, 226 CeedVector in_vec, const bool skip_active, CeedScalar *e_data[2 * CEED_FIELD_MAX], 227 CeedOperator_Hip *impl, CeedRequest *request) { 228 for (CeedInt i = 0; i < num_input_fields; i++) { 229 CeedEvalMode e_mode; 230 CeedVector vec; 231 CeedElemRestriction elem_rstr; 232 233 // Get input vector 234 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 235 if (vec == CEED_VECTOR_ACTIVE) { 236 if (skip_active) continue; 237 else vec = in_vec; 238 } 239 240 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &e_mode)); 241 if (e_mode == CEED_EVAL_WEIGHT) { // Skip 242 } else { 243 // Get input vector 244 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 245 // Get input element restriction 246 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_rstr)); 247 if (vec == CEED_VECTOR_ACTIVE) vec = in_vec; 248 // Restrict, if necessary 249 if (!impl->e_vecs[i]) { 250 // No restriction for this field; read data directly from vec. 251 CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, (const CeedScalar **)&e_data[i])); 252 } else { 253 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_NOTRANSPOSE, vec, impl->e_vecs[i], request)); 254 // Get evec 255 CeedCallBackend(CeedVectorGetArrayRead(impl->e_vecs[i], CEED_MEM_DEVICE, (const CeedScalar **)&e_data[i])); 256 } 257 } 258 } 259 return CEED_ERROR_SUCCESS; 260 } 261 262 //------------------------------------------------------------------------------ 263 // Input Basis Action 264 //------------------------------------------------------------------------------ 265 static inline int CeedOperatorInputBasis_Hip(CeedInt num_elem, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields, 266 CeedInt num_input_fields, const bool skip_active, CeedScalar *e_data[2 * CEED_FIELD_MAX], 267 CeedOperator_Hip *impl) { 268 for (CeedInt i = 0; i < num_input_fields; i++) { 269 CeedInt elem_size, size; 270 CeedEvalMode e_mode; 271 CeedElemRestriction elem_rstr; 272 CeedBasis basis; 273 274 // Skip active input 275 if (skip_active) { 276 CeedVector vec; 277 278 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 279 if (vec == CEED_VECTOR_ACTIVE) continue; 280 } 281 // Get elem_size, e_mode, size 282 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_rstr)); 283 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 284 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &e_mode)); 285 CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size)); 286 // Basis action 287 switch (e_mode) { 288 case CEED_EVAL_NONE: 289 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_data[i])); 290 break; 291 case CEED_EVAL_INTERP: 292 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 293 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_INTERP, impl->e_vecs[i], impl->q_vecs_in[i])); 294 break; 295 case CEED_EVAL_GRAD: 296 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 297 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_GRAD, impl->e_vecs[i], impl->q_vecs_in[i])); 298 break; 299 case CEED_EVAL_WEIGHT: 300 break; // No action 301 case CEED_EVAL_DIV: 302 break; // TODO: Not implemented 303 case CEED_EVAL_CURL: 304 break; // TODO: Not implemented 305 } 306 } 307 return CEED_ERROR_SUCCESS; 308 } 309 310 //------------------------------------------------------------------------------ 311 // Restore Input Vectors 312 //------------------------------------------------------------------------------ 313 static inline int CeedOperatorRestoreInputs_Hip(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields, 314 const bool skip_active, CeedScalar *e_data[2 * CEED_FIELD_MAX], CeedOperator_Hip *impl) { 315 for (CeedInt i = 0; i < num_input_fields; i++) { 316 CeedEvalMode e_mode; 317 CeedVector vec; 318 // Skip active input 319 if (skip_active) { 320 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 321 if (vec == CEED_VECTOR_ACTIVE) continue; 322 } 323 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &e_mode)); 324 if (e_mode == CEED_EVAL_WEIGHT) { // Skip 325 } else { 326 if (!impl->e_vecs[i]) { // This was a skip_restriction case 327 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 328 CeedCallBackend(CeedVectorRestoreArrayRead(vec, (const CeedScalar **)&e_data[i])); 329 } else { 330 CeedCallBackend(CeedVectorRestoreArrayRead(impl->e_vecs[i], (const CeedScalar **)&e_data[i])); 331 } 332 } 333 } 334 return CEED_ERROR_SUCCESS; 335 } 336 337 //------------------------------------------------------------------------------ 338 // Apply and add to output 339 //------------------------------------------------------------------------------ 340 static int CeedOperatorApplyAdd_Hip(CeedOperator op, CeedVector in_vec, CeedVector out_vec, CeedRequest *request) { 341 CeedInt Q, num_elem, elem_size, num_input_fields, num_output_fields, size; 342 CeedScalar *e_data[2 * CEED_FIELD_MAX] = {NULL}; 343 CeedQFunctionField *qf_input_fields, *qf_output_fields; 344 CeedQFunction qf; 345 CeedOperatorField *op_input_fields, *op_output_fields; 346 CeedOperator_Hip *impl; 347 348 CeedCallBackend(CeedOperatorGetData(op, &impl)); 349 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 350 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 351 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 352 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 353 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 354 355 // Setup 356 CeedCallBackend(CeedOperatorSetup_Hip(op)); 357 358 // Input Evecs and Restriction 359 CeedCallBackend(CeedOperatorSetupInputs_Hip(num_input_fields, qf_input_fields, op_input_fields, in_vec, false, e_data, impl, request)); 360 361 // Input basis apply if needed 362 CeedCallBackend(CeedOperatorInputBasis_Hip(num_elem, qf_input_fields, op_input_fields, num_input_fields, false, e_data, impl)); 363 364 // Output pointers, as necessary 365 for (CeedInt i = 0; i < num_output_fields; i++) { 366 CeedEvalMode e_mode; 367 368 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &e_mode)); 369 if (e_mode == CEED_EVAL_NONE) { 370 // Set the output Q-Vector to use the E-Vector data directly. 371 CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs[i + impl->num_inputs], CEED_MEM_DEVICE, &e_data[i + num_input_fields])); 372 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_data[i + num_input_fields])); 373 } 374 } 375 376 // Q function 377 CeedCallBackend(CeedQFunctionApply(qf, num_elem * Q, impl->q_vecs_in, impl->q_vecs_out)); 378 379 // Output basis apply if needed 380 for (CeedInt i = 0; i < num_output_fields; i++) { 381 CeedEvalMode e_mode; 382 CeedElemRestriction elem_rstr; 383 CeedBasis basis; 384 385 // Get elem_size, e_mode, size 386 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_rstr)); 387 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 388 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &e_mode)); 389 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size)); 390 // Basis action 391 switch (e_mode) { 392 case CEED_EVAL_NONE: 393 break; 394 case CEED_EVAL_INTERP: 395 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 396 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_TRANSPOSE, CEED_EVAL_INTERP, impl->q_vecs_out[i], impl->e_vecs[i + impl->num_inputs])); 397 break; 398 case CEED_EVAL_GRAD: 399 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 400 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_TRANSPOSE, CEED_EVAL_GRAD, impl->q_vecs_out[i], impl->e_vecs[i + impl->num_inputs])); 401 break; 402 // LCOV_EXCL_START 403 case CEED_EVAL_WEIGHT: { 404 Ceed ceed; 405 406 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 407 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 408 break; // Should not occur 409 } 410 case CEED_EVAL_DIV: 411 break; // TODO: Not implemented 412 case CEED_EVAL_CURL: 413 break; // TODO: Not implemented 414 // LCOV_EXCL_STOP 415 } 416 } 417 418 // Output restriction 419 for (CeedInt i = 0; i < num_output_fields; i++) { 420 CeedEvalMode e_mode; 421 CeedVector vec; 422 CeedElemRestriction elem_rstr; 423 424 // Restore evec 425 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &e_mode)); 426 if (e_mode == CEED_EVAL_NONE) { 427 CeedCallBackend(CeedVectorRestoreArray(impl->e_vecs[i + impl->num_inputs], &e_data[i + num_input_fields])); 428 } 429 // Get output vector 430 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 431 // Restrict 432 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_rstr)); 433 // Active 434 if (vec == CEED_VECTOR_ACTIVE) vec = out_vec; 435 436 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_TRANSPOSE, impl->e_vecs[i + impl->num_inputs], vec, request)); 437 } 438 439 // Restore input arrays 440 CeedCallBackend(CeedOperatorRestoreInputs_Hip(num_input_fields, qf_input_fields, op_input_fields, false, e_data, impl)); 441 return CEED_ERROR_SUCCESS; 442 } 443 444 //------------------------------------------------------------------------------ 445 // Core code for assembling linear QFunction 446 //------------------------------------------------------------------------------ 447 static inline int CeedOperatorLinearAssembleQFunctionCore_Hip(CeedOperator op, bool build_objects, CeedVector *assembled, CeedElemRestriction *rstr, 448 CeedRequest *request) { 449 Ceed ceed, ceed_parent; 450 bool is_identity_qf; 451 CeedSize q_size; 452 CeedInt num_active_in, num_active_out, Q, num_elem, num_input_fields, num_output_fields, size; 453 CeedScalar *assembled_array, *e_data[2 * CEED_FIELD_MAX] = {NULL}; 454 CeedVector *active_in; 455 CeedQFunctionField *qf_input_fields, *qf_output_fields; 456 CeedQFunction qf; 457 CeedOperatorField *op_input_fields, *op_output_fields; 458 CeedOperator_Hip *impl; 459 460 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 461 CeedCallBackend(CeedOperatorGetFallbackParentCeed(op, &ceed_parent)); 462 CeedCallBackend(CeedOperatorGetData(op, &impl)); 463 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 464 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 465 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 466 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 467 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 468 active_in = impl->qf_active_in; 469 num_active_in = impl->num_active_in; 470 num_active_out = impl->num_active_out; 471 472 // Setup 473 CeedCallBackend(CeedOperatorSetup_Hip(op)); 474 475 // Check for identity 476 CeedCallBackend(CeedQFunctionIsIdentity(qf, &is_identity_qf)); 477 CeedCheck(!is_identity_qf, ceed, CEED_ERROR_BACKEND, "Assembling identity QFunctions not supported"); 478 479 // Input Evecs and Restriction 480 CeedCallBackend(CeedOperatorSetupInputs_Hip(num_input_fields, qf_input_fields, op_input_fields, NULL, true, e_data, impl, request)); 481 482 // Count number of active input fields 483 if (!num_active_in) { 484 for (CeedInt i = 0; i < num_input_fields; i++) { 485 CeedScalar *q_vec_array; 486 CeedVector vec; 487 488 // Get input vector 489 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 490 // Check if active input 491 if (vec == CEED_VECTOR_ACTIVE) { 492 CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size)); 493 CeedCallBackend(CeedVectorSetValue(impl->q_vecs_in[i], 0.0)); 494 CeedCallBackend(CeedVectorGetArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, &q_vec_array)); 495 CeedCallBackend(CeedRealloc(num_active_in + size, &active_in)); 496 for (CeedInt field = 0; field < size; field++) { 497 q_size = (CeedSize)Q * num_elem; 498 CeedCallBackend(CeedVectorCreate(ceed, q_size, &active_in[num_active_in + field])); 499 CeedCallBackend( 500 CeedVectorSetArray(active_in[num_active_in + field], CEED_MEM_DEVICE, CEED_USE_POINTER, &q_vec_array[field * Q * num_elem])); 501 } 502 num_active_in += size; 503 CeedCallBackend(CeedVectorRestoreArray(impl->q_vecs_in[i], &q_vec_array)); 504 } 505 } 506 impl->num_active_in = num_active_in; 507 impl->qf_active_in = active_in; 508 } 509 510 // Count number of active output fields 511 if (!num_active_out) { 512 for (CeedInt i = 0; i < num_output_fields; i++) { 513 CeedVector vec; 514 515 // Get output vector 516 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 517 // Check if active output 518 if (vec == CEED_VECTOR_ACTIVE) { 519 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size)); 520 num_active_out += size; 521 } 522 } 523 impl->num_active_out = num_active_out; 524 } 525 526 // Check sizes 527 CeedCheck(num_active_in > 0 && num_active_out > 0, ceed, CEED_ERROR_BACKEND, "Cannot assemble QFunction without active inputs and outputs"); 528 529 // Build objects if needed 530 if (build_objects) { 531 // Create output restriction 532 CeedSize l_size = (CeedSize)num_elem * Q * num_active_in * num_active_out; 533 CeedInt strides[3] = {1, num_elem * Q, Q}; /* *NOPAD* */ 534 535 CeedCallBackend(CeedElemRestrictionCreateStrided(ceed_parent, num_elem, Q, num_active_in * num_active_out, 536 num_active_in * num_active_out * num_elem * Q, strides, rstr)); 537 // Create assembled vector 538 CeedCallBackend(CeedVectorCreate(ceed_parent, l_size, assembled)); 539 } 540 CeedCallBackend(CeedVectorSetValue(*assembled, 0.0)); 541 CeedCallBackend(CeedVectorGetArray(*assembled, CEED_MEM_DEVICE, &assembled_array)); 542 543 // Input basis apply 544 CeedCallBackend(CeedOperatorInputBasis_Hip(num_elem, qf_input_fields, op_input_fields, num_input_fields, true, e_data, impl)); 545 546 // Assemble QFunction 547 for (CeedInt in = 0; in < num_active_in; in++) { 548 // Set Inputs 549 CeedCallBackend(CeedVectorSetValue(active_in[in], 1.0)); 550 if (num_active_in > 1) { 551 CeedCallBackend(CeedVectorSetValue(active_in[(in + num_active_in - 1) % num_active_in], 0.0)); 552 } 553 // Set Outputs 554 for (CeedInt out = 0; out < num_output_fields; out++) { 555 CeedVector vec; 556 557 // Get output vector 558 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec)); 559 // Check if active output 560 if (vec == CEED_VECTOR_ACTIVE) { 561 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, CEED_USE_POINTER, assembled_array)); 562 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[out], &size)); 563 assembled_array += size * Q * num_elem; // Advance the pointer by the size of the output 564 } 565 } 566 // Apply QFunction 567 CeedCallBackend(CeedQFunctionApply(qf, Q * num_elem, impl->q_vecs_in, impl->q_vecs_out)); 568 } 569 570 // Un-set output Qvecs to prevent accidental overwrite of Assembled 571 for (CeedInt out = 0; out < num_output_fields; out++) { 572 CeedVector vec; 573 574 // Get output vector 575 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec)); 576 // Check if active output 577 if (vec == CEED_VECTOR_ACTIVE) { 578 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, NULL)); 579 } 580 } 581 582 // Restore input arrays 583 CeedCallBackend(CeedOperatorRestoreInputs_Hip(num_input_fields, qf_input_fields, op_input_fields, true, e_data, impl)); 584 585 // Restore output 586 CeedCallBackend(CeedVectorRestoreArray(*assembled, &assembled_array)); 587 return CEED_ERROR_SUCCESS; 588 } 589 590 //------------------------------------------------------------------------------ 591 // Assemble Linear QFunction 592 //------------------------------------------------------------------------------ 593 static int CeedOperatorLinearAssembleQFunction_Hip(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 594 return CeedOperatorLinearAssembleQFunctionCore_Hip(op, true, assembled, rstr, request); 595 } 596 597 //------------------------------------------------------------------------------ 598 // Update Assembled Linear QFunction 599 //------------------------------------------------------------------------------ 600 static int CeedOperatorLinearAssembleQFunctionUpdate_Hip(CeedOperator op, CeedVector assembled, CeedElemRestriction rstr, CeedRequest *request) { 601 return CeedOperatorLinearAssembleQFunctionCore_Hip(op, false, &assembled, &rstr, request); 602 } 603 604 //------------------------------------------------------------------------------ 605 // Create point block restriction 606 //------------------------------------------------------------------------------ 607 static int CreatePBRestriction(CeedElemRestriction rstr, CeedElemRestriction *point_block_rstr) { 608 Ceed ceed; 609 CeedSize l_size; 610 CeedInt num_elem, num_comp, elem_size, comp_stride, *point_block_offsets; 611 const CeedInt *offsets; 612 613 CeedCallBackend(CeedElemRestrictionGetCeed(rstr, &ceed)); 614 CeedCallBackend(CeedElemRestrictionGetOffsets(rstr, CEED_MEM_HOST, &offsets)); 615 616 // Expand offsets 617 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr, &num_elem)); 618 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr, &num_comp)); 619 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr, &elem_size)); 620 CeedCallBackend(CeedElemRestrictionGetCompStride(rstr, &comp_stride)); 621 CeedCallBackend(CeedElemRestrictionGetLVectorSize(rstr, &l_size)); 622 CeedInt shift = num_comp; 623 624 if (comp_stride != 1) shift *= num_comp; 625 CeedCallBackend(CeedCalloc(num_elem * elem_size, &point_block_offsets)); 626 for (CeedInt i = 0; i < num_elem * elem_size; i++) point_block_offsets[i] = offsets[i] * shift; 627 628 // Create new restriction 629 CeedCallBackend(CeedElemRestrictionCreate(ceed, num_elem, elem_size, num_comp * num_comp, 1, l_size * num_comp, CEED_MEM_HOST, CEED_OWN_POINTER, 630 point_block_offsets, point_block_rstr)); 631 632 // Cleanup 633 CeedCallBackend(CeedElemRestrictionRestoreOffsets(rstr, &offsets)); 634 return CEED_ERROR_SUCCESS; 635 } 636 637 //------------------------------------------------------------------------------ 638 // Assemble diagonal setup 639 //------------------------------------------------------------------------------ 640 static inline int CeedOperatorAssembleDiagonalSetup_Hip(CeedOperator op, const bool is_point_block, CeedInt use_ceedsize_idx) { 641 Ceed ceed; 642 char *diagonal_kernel_path, *diagonal_kernel_source; 643 CeedInt num_input_fields, num_output_fields, num_e_mode_in = 0, num_comp = 0, dim = 1, num_e_mode_out = 0; 644 CeedEvalMode *e_mode_in = NULL, *e_mode_out = NULL; 645 CeedElemRestriction rstr_in = NULL, rstr_out = NULL; 646 CeedBasis basis_in = NULL, basis_out = NULL; 647 CeedQFunctionField *qf_fields; 648 CeedQFunction qf; 649 CeedOperatorField *op_fields; 650 CeedOperator_Hip *impl; 651 652 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 653 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 654 CeedCallBackend(CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields)); 655 656 // Determine active input basis 657 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 658 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 659 for (CeedInt i = 0; i < num_input_fields; i++) { 660 CeedVector vec; 661 662 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 663 if (vec == CEED_VECTOR_ACTIVE) { 664 CeedEvalMode e_mode; 665 CeedElemRestriction rstr; 666 667 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_in)); 668 CeedCallBackend(CeedBasisGetNumComponents(basis_in, &num_comp)); 669 CeedCallBackend(CeedBasisGetDimension(basis_in, &dim)); 670 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr)); 671 CeedCheck(!rstr_in || rstr_in == rstr, ceed, CEED_ERROR_BACKEND, 672 "Backend does not implement multi-field non-composite operator diagonal assembly"); 673 rstr_in = rstr; 674 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &e_mode)); 675 switch (e_mode) { 676 case CEED_EVAL_NONE: 677 case CEED_EVAL_INTERP: 678 CeedCallBackend(CeedRealloc(num_e_mode_in + 1, &e_mode_in)); 679 e_mode_in[num_e_mode_in] = e_mode; 680 num_e_mode_in += 1; 681 break; 682 case CEED_EVAL_GRAD: 683 CeedCallBackend(CeedRealloc(num_e_mode_in + dim, &e_mode_in)); 684 for (CeedInt d = 0; d < dim; d++) e_mode_in[num_e_mode_in + d] = e_mode; 685 num_e_mode_in += dim; 686 break; 687 case CEED_EVAL_WEIGHT: 688 case CEED_EVAL_DIV: 689 case CEED_EVAL_CURL: 690 break; // Caught by QF Assembly 691 } 692 } 693 } 694 695 // Determine active output basis 696 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 697 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 698 for (CeedInt i = 0; i < num_output_fields; i++) { 699 CeedVector vec; 700 701 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 702 if (vec == CEED_VECTOR_ACTIVE) { 703 CeedEvalMode e_mode; 704 CeedElemRestriction rstr; 705 706 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_out)); 707 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr)); 708 CeedCheck(!rstr_out || rstr_out == rstr, ceed, CEED_ERROR_BACKEND, 709 "Backend does not implement multi-field non-composite operator diagonal assembly"); 710 rstr_out = rstr; 711 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &e_mode)); 712 switch (e_mode) { 713 case CEED_EVAL_NONE: 714 case CEED_EVAL_INTERP: 715 CeedCallBackend(CeedRealloc(num_e_mode_out + 1, &e_mode_out)); 716 e_mode_out[num_e_mode_out] = e_mode; 717 num_e_mode_out += 1; 718 break; 719 case CEED_EVAL_GRAD: 720 CeedCallBackend(CeedRealloc(num_e_mode_out + dim, &e_mode_out)); 721 for (CeedInt d = 0; d < dim; d++) e_mode_out[num_e_mode_out + d] = e_mode; 722 num_e_mode_out += dim; 723 break; 724 case CEED_EVAL_WEIGHT: 725 case CEED_EVAL_DIV: 726 case CEED_EVAL_CURL: 727 break; // Caught by QF Assembly 728 } 729 } 730 } 731 732 // Operator data struct 733 CeedCallBackend(CeedOperatorGetData(op, &impl)); 734 CeedCallBackend(CeedCalloc(1, &impl->diag)); 735 CeedOperatorDiag_Hip *diag = impl->diag; 736 737 diag->basis_in = basis_in; 738 diag->basis_out = basis_out; 739 diag->h_e_mode_in = e_mode_in; 740 diag->h_e_mode_out = e_mode_out; 741 diag->num_e_mode_in = num_e_mode_in; 742 diag->num_e_mode_out = num_e_mode_out; 743 744 // Assemble kernel 745 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-ref-operator-assemble-diagonal.h", &diagonal_kernel_path)); 746 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Kernel Source -----\n"); 747 CeedCallBackend(CeedLoadSourceToBuffer(ceed, diagonal_kernel_path, &diagonal_kernel_source)); 748 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Source Complete! -----\n"); 749 CeedInt num_modes, num_qpts; 750 CeedCallBackend(CeedBasisGetNumNodes(basis_in, &num_modes)); 751 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 752 diag->num_modes = num_modes; 753 CeedCallBackend(CeedCompile_Hip(ceed, diagonal_kernel_source, &diag->module, 6, "NUMEMODEIN", num_e_mode_in, "NUMEMODEOUT", num_e_mode_out, 754 "NNODES", num_modes, "NQPTS", num_qpts, "NCOMP", num_comp, "CEEDSIZE", use_ceedsize_idx)); 755 CeedCallBackend(CeedGetKernel_Hip(ceed, diag->module, "linearDiagonal", &diag->linearDiagonal)); 756 CeedCallBackend(CeedGetKernel_Hip(ceed, diag->module, "linearPointBlockDiagonal", &diag->linearPointBlock)); 757 CeedCallBackend(CeedFree(&diagonal_kernel_path)); 758 CeedCallBackend(CeedFree(&diagonal_kernel_source)); 759 760 // Basis matrices 761 const CeedInt q_bytes = num_qpts * sizeof(CeedScalar); 762 const CeedInt interp_bytes = q_bytes * num_modes; 763 const CeedInt grad_bytes = q_bytes * num_modes * dim; 764 const CeedInt e_mode_bytes = sizeof(CeedEvalMode); 765 const CeedScalar *interp_in, *interp_out, *grad_in, *grad_out; 766 767 // CEED_EVAL_NONE 768 CeedScalar *identity = NULL; 769 bool is_eval_none = false; 770 771 for (CeedInt i = 0; i < num_e_mode_in; i++) is_eval_none = is_eval_none || (e_mode_in[i] == CEED_EVAL_NONE); 772 for (CeedInt i = 0; i < num_e_mode_out; i++) is_eval_none = is_eval_none || (e_mode_out[i] == CEED_EVAL_NONE); 773 if (is_eval_none) { 774 CeedCallBackend(CeedCalloc(num_qpts * num_modes, &identity)); 775 for (CeedInt i = 0; i < (num_modes < num_qpts ? num_modes : num_qpts); i++) identity[i * num_modes + i] = 1.0; 776 CeedCallHip(ceed, hipMalloc((void **)&diag->d_identity, interp_bytes)); 777 CeedCallHip(ceed, hipMemcpy(diag->d_identity, identity, interp_bytes, hipMemcpyHostToDevice)); 778 } 779 780 // CEED_EVAL_INTERP 781 CeedCallBackend(CeedBasisGetInterp(basis_in, &interp_in)); 782 CeedCallHip(ceed, hipMalloc((void **)&diag->d_interp_in, interp_bytes)); 783 CeedCallHip(ceed, hipMemcpy(diag->d_interp_in, interp_in, interp_bytes, hipMemcpyHostToDevice)); 784 CeedCallBackend(CeedBasisGetInterp(basis_out, &interp_out)); 785 CeedCallHip(ceed, hipMalloc((void **)&diag->d_interp_out, interp_bytes)); 786 CeedCallHip(ceed, hipMemcpy(diag->d_interp_out, interp_out, interp_bytes, hipMemcpyHostToDevice)); 787 788 // CEED_EVAL_GRAD 789 CeedCallBackend(CeedBasisGetGrad(basis_in, &grad_in)); 790 CeedCallHip(ceed, hipMalloc((void **)&diag->d_grad_in, grad_bytes)); 791 CeedCallHip(ceed, hipMemcpy(diag->d_grad_in, grad_in, grad_bytes, hipMemcpyHostToDevice)); 792 CeedCallBackend(CeedBasisGetGrad(basis_out, &grad_out)); 793 CeedCallHip(ceed, hipMalloc((void **)&diag->d_grad_out, grad_bytes)); 794 CeedCallHip(ceed, hipMemcpy(diag->d_grad_out, grad_out, grad_bytes, hipMemcpyHostToDevice)); 795 796 // Arrays of e_modes 797 CeedCallHip(ceed, hipMalloc((void **)&diag->d_e_mode_in, num_e_mode_in * e_mode_bytes)); 798 CeedCallHip(ceed, hipMemcpy(diag->d_e_mode_in, e_mode_in, num_e_mode_in * e_mode_bytes, hipMemcpyHostToDevice)); 799 CeedCallHip(ceed, hipMalloc((void **)&diag->d_e_mode_out, num_e_mode_out * e_mode_bytes)); 800 CeedCallHip(ceed, hipMemcpy(diag->d_e_mode_out, e_mode_out, num_e_mode_out * e_mode_bytes, hipMemcpyHostToDevice)); 801 802 // Restriction 803 diag->diag_rstr = rstr_out; 804 return CEED_ERROR_SUCCESS; 805 } 806 807 //------------------------------------------------------------------------------ 808 // Assemble diagonal common code 809 //------------------------------------------------------------------------------ 810 static inline int CeedOperatorAssembleDiagonalCore_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request, const bool is_point_block) { 811 Ceed ceed; 812 CeedSize assembled_length = 0, assembled_qf_length = 0; 813 CeedInt use_ceedsize_idx = 0, num_elem; 814 CeedScalar *elem_diag_array; 815 const CeedScalar *assembled_qf_array; 816 CeedVector assembled_qf = NULL; 817 CeedElemRestriction rstr = NULL; 818 CeedOperator_Hip *impl; 819 820 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 821 CeedCallBackend(CeedOperatorGetData(op, &impl)); 822 823 // Assemble QFunction 824 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &rstr, request)); 825 CeedCallBackend(CeedElemRestrictionDestroy(&rstr)); 826 827 CeedCallBackend(CeedVectorGetLength(assembled, &assembled_length)); 828 CeedCallBackend(CeedVectorGetLength(assembled_qf, &assembled_qf_length)); 829 if ((assembled_length > INT_MAX) || (assembled_qf_length > INT_MAX)) use_ceedsize_idx = 1; 830 831 // Setup 832 if (!impl->diag) CeedCallBackend(CeedOperatorAssembleDiagonalSetup_Hip(op, is_point_block, use_ceedsize_idx)); 833 CeedOperatorDiag_Hip *diag = impl->diag; 834 assert(diag != NULL); 835 836 // Restriction 837 if (is_point_block && !diag->point_block_diag_rstr) { 838 CeedElemRestriction point_block_diag_rstr; 839 840 CeedCallBackend(CreatePBRestriction(diag->diag_rstr, &point_block_diag_rstr)); 841 diag->point_block_diag_rstr = point_block_diag_rstr; 842 } 843 CeedElemRestriction diag_rstr = is_point_block ? diag->point_block_diag_rstr : diag->diag_rstr; 844 845 // Create diagonal vector 846 CeedVector elem_diag = is_point_block ? diag->point_block_elem_diag : diag->elem_diag; 847 848 if (!elem_diag) { 849 // Element diagonal vector 850 CeedCallBackend(CeedElemRestrictionCreateVector(diag_rstr, NULL, &elem_diag)); 851 if (is_point_block) diag->point_block_elem_diag = elem_diag; 852 else diag->elem_diag = elem_diag; 853 } 854 CeedCallBackend(CeedVectorSetValue(elem_diag, 0.0)); 855 856 // Assemble element operator diagonals 857 CeedCallBackend(CeedVectorGetArray(elem_diag, CEED_MEM_DEVICE, &elem_diag_array)); 858 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &assembled_qf_array)); 859 CeedCallBackend(CeedElemRestrictionGetNumElements(diag_rstr, &num_elem)); 860 861 // Compute the diagonal of B^T D B 862 int elem_per_block = 1; 863 int grid = num_elem / elem_per_block + ((num_elem / elem_per_block * elem_per_block < num_elem) ? 1 : 0); 864 void *args[] = {(void *)&num_elem, &diag->d_identity, &diag->d_interp_in, &diag->d_grad_in, &diag->d_interp_out, 865 &diag->d_grad_out, &diag->d_e_mode_in, &diag->d_e_mode_out, &assembled_qf_array, &elem_diag_array}; 866 867 if (is_point_block) { 868 CeedCallBackend(CeedRunKernelDim_Hip(ceed, diag->linearPointBlock, grid, diag->num_modes, 1, elem_per_block, args)); 869 } else { 870 CeedCallBackend(CeedRunKernelDim_Hip(ceed, diag->linearDiagonal, grid, diag->num_modes, 1, elem_per_block, args)); 871 } 872 873 // Restore arrays 874 CeedCallBackend(CeedVectorRestoreArray(elem_diag, &elem_diag_array)); 875 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 876 877 // Assemble local operator diagonal 878 CeedCallBackend(CeedElemRestrictionApply(diag_rstr, CEED_TRANSPOSE, elem_diag, assembled, request)); 879 880 // Cleanup 881 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 882 return CEED_ERROR_SUCCESS; 883 } 884 885 //------------------------------------------------------------------------------ 886 // Assemble Linear Diagonal 887 //------------------------------------------------------------------------------ 888 static int CeedOperatorLinearAssembleAddDiagonal_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request) { 889 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Hip(op, assembled, request, false)); 890 return CEED_ERROR_SUCCESS; 891 } 892 893 //------------------------------------------------------------------------------ 894 // Assemble Linear Point Block Diagonal 895 //------------------------------------------------------------------------------ 896 static int CeedOperatorLinearAssembleAddPointBlockDiagonal_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request) { 897 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Hip(op, assembled, request, true)); 898 return CEED_ERROR_SUCCESS; 899 } 900 901 //------------------------------------------------------------------------------ 902 // Single operator assembly setup 903 //------------------------------------------------------------------------------ 904 static int CeedSingleOperatorAssembleSetup_Hip(CeedOperator op, CeedInt use_ceedsize_idx) { 905 Ceed ceed; 906 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_qpts = 0, elem_size = 0, 907 num_e_mode_out = 0, num_B_out_mats_to_load = 0, size_B_out = 0, num_elem, num_comp; 908 CeedEvalMode *eval_mode_in = NULL, *eval_mode_out = NULL; 909 CeedElemRestriction rstr_in = NULL, rstr_out = NULL; 910 CeedBasis basis_in = NULL, basis_out = NULL; 911 CeedQFunctionField *qf_fields; 912 CeedQFunction qf; 913 CeedOperatorField *input_fields, *output_fields; 914 CeedOperator_Hip *impl; 915 916 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 917 CeedCallBackend(CeedOperatorGetData(op, &impl)); 918 919 // Get intput and output fields 920 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 921 922 // Determine active input basis eval mode 923 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 924 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 925 // Note that the kernel will treat each dimension of a gradient action separately; 926 // i.e., when an active input has a CEED_EVAL_GRAD mode, num_e_mode_in will increment by dim. 927 // 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 928 // num_B_in_mats_to_load will be incremented by 1. 929 for (CeedInt i = 0; i < num_input_fields; i++) { 930 CeedVector vec; 931 932 CeedCallBackend(CeedOperatorFieldGetVector(input_fields[i], &vec)); 933 if (vec == CEED_VECTOR_ACTIVE) { 934 CeedEvalMode eval_mode; 935 936 CeedCallBackend(CeedOperatorFieldGetBasis(input_fields[i], &basis_in)); 937 CeedCallBackend(CeedBasisGetDimension(basis_in, &dim)); 938 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 939 CeedCallBackend(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr_in)); 940 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_in, &elem_size)); 941 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 942 if (eval_mode != CEED_EVAL_NONE) { 943 CeedCallBackend(CeedRealloc(num_B_in_mats_to_load + 1, &eval_mode_in)); 944 eval_mode_in[num_B_in_mats_to_load] = eval_mode; 945 num_B_in_mats_to_load += 1; 946 if (eval_mode == CEED_EVAL_GRAD) { 947 num_e_mode_in += dim; 948 size_B_in += dim * elem_size * num_qpts; 949 } else { 950 num_e_mode_in += 1; 951 size_B_in += elem_size * num_qpts; 952 } 953 } 954 } 955 } 956 957 // Determine active output basis; basis_out and rstr_out only used if same as input, TODO 958 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 959 for (CeedInt i = 0; i < num_output_fields; i++) { 960 CeedVector vec; 961 962 CeedCallBackend(CeedOperatorFieldGetVector(output_fields[i], &vec)); 963 if (vec == CEED_VECTOR_ACTIVE) { 964 CeedEvalMode eval_mode; 965 966 CeedCallBackend(CeedOperatorFieldGetBasis(output_fields[i], &basis_out)); 967 CeedCallBackend(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr_out)); 968 CeedCheck(!rstr_out || rstr_out == rstr_in, ceed, CEED_ERROR_BACKEND, "Backend does not implement multi-field non-composite operator assembly"); 969 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 970 if (eval_mode != CEED_EVAL_NONE) { 971 CeedCallBackend(CeedRealloc(num_B_out_mats_to_load + 1, &eval_mode_out)); 972 eval_mode_out[num_B_out_mats_to_load] = eval_mode; 973 num_B_out_mats_to_load += 1; 974 if (eval_mode == CEED_EVAL_GRAD) { 975 num_e_mode_out += dim; 976 size_B_out += dim * elem_size * num_qpts; 977 } else { 978 num_e_mode_out += 1; 979 size_B_out += elem_size * num_qpts; 980 } 981 } 982 } 983 } 984 985 CeedCheck(num_e_mode_in > 0 && num_e_mode_out > 0, ceed, CEED_ERROR_UNSUPPORTED, "Cannot assemble operator without inputs/outputs"); 986 987 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr_in, &num_elem)); 988 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr_in, &num_comp)); 989 990 CeedCallBackend(CeedCalloc(1, &impl->asmb)); 991 CeedOperatorAssemble_Hip *asmb = impl->asmb; 992 asmb->num_elem = num_elem; 993 994 // Compile kernels 995 int elem_per_block = 1; 996 asmb->elem_per_block = elem_per_block; 997 CeedInt block_size = elem_size * elem_size * elem_per_block; 998 char *assembly_kernel_path, *assembly_kernel_source; 999 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-ref-operator-assemble.h", &assembly_kernel_path)); 1000 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Kernel Source -----\n"); 1001 CeedCallBackend(CeedLoadSourceToBuffer(ceed, assembly_kernel_path, &assembly_kernel_source)); 1002 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Source Complete! -----\n"); 1003 bool fallback = block_size > 1024; 1004 if (fallback) { // Use fallback kernel with 1D threadblock 1005 block_size = elem_size * elem_per_block; 1006 asmb->block_size_x = elem_size; 1007 asmb->block_size_y = 1; 1008 } else { // Use kernel with 2D threadblock 1009 asmb->block_size_x = elem_size; 1010 asmb->block_size_y = elem_size; 1011 } 1012 CeedCallBackend(CeedCompile_Hip(ceed, assembly_kernel_source, &asmb->module, 8, "NELEM", num_elem, "NUMEMODEIN", num_e_mode_in, "NUMEMODEOUT", 1013 num_e_mode_out, "NQPTS", num_qpts, "NNODES", elem_size, "BLOCK_SIZE", block_size, "NCOMP", num_comp, "CEEDSIZE", 1014 use_ceedsize_idx)); 1015 CeedCallBackend(CeedGetKernel_Hip(ceed, asmb->module, fallback ? "linearAssembleFallback" : "linearAssemble", &asmb->linearAssemble)); 1016 CeedCallBackend(CeedFree(&assembly_kernel_path)); 1017 CeedCallBackend(CeedFree(&assembly_kernel_source)); 1018 1019 // Build 'full' B matrices (not 1D arrays used for tensor-product matrices) 1020 const CeedScalar *interp_in, *grad_in; 1021 CeedCallBackend(CeedBasisGetInterp(basis_in, &interp_in)); 1022 CeedCallBackend(CeedBasisGetGrad(basis_in, &grad_in)); 1023 1024 // Load into B_in, in order that they will be used in eval_mode 1025 const CeedInt in_bytes = size_B_in * sizeof(CeedScalar); 1026 CeedInt mat_start = 0; 1027 1028 CeedCallHip(ceed, hipMalloc((void **)&asmb->d_B_in, in_bytes)); 1029 for (int i = 0; i < num_B_in_mats_to_load; i++) { 1030 CeedEvalMode eval_mode = eval_mode_in[i]; 1031 if (eval_mode == CEED_EVAL_INTERP) { 1032 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_in[mat_start], interp_in, elem_size * num_qpts * sizeof(CeedScalar), hipMemcpyHostToDevice)); 1033 mat_start += elem_size * num_qpts; 1034 } else if (eval_mode == CEED_EVAL_GRAD) { 1035 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_in[mat_start], grad_in, dim * elem_size * num_qpts * sizeof(CeedScalar), hipMemcpyHostToDevice)); 1036 mat_start += dim * elem_size * num_qpts; 1037 } 1038 } 1039 1040 const CeedScalar *interp_out, *grad_out; 1041 1042 // Note that this function currently assumes 1 basis, so this should always be true for now 1043 if (basis_out == basis_in) { 1044 interp_out = interp_in; 1045 grad_out = grad_in; 1046 } else { 1047 CeedCallBackend(CeedBasisGetInterp(basis_out, &interp_out)); 1048 CeedCallBackend(CeedBasisGetGrad(basis_out, &grad_out)); 1049 } 1050 1051 // Load into B_out, in order that they will be used in eval_mode 1052 const CeedInt out_bytes = size_B_out * sizeof(CeedScalar); 1053 1054 mat_start = 0; 1055 CeedCallHip(ceed, hipMalloc((void **)&asmb->d_B_out, out_bytes)); 1056 for (int i = 0; i < num_B_out_mats_to_load; i++) { 1057 CeedEvalMode eval_mode = eval_mode_out[i]; 1058 if (eval_mode == CEED_EVAL_INTERP) { 1059 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_out[mat_start], interp_out, elem_size * num_qpts * sizeof(CeedScalar), hipMemcpyHostToDevice)); 1060 mat_start += elem_size * num_qpts; 1061 } else if (eval_mode == CEED_EVAL_GRAD) { 1062 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_out[mat_start], grad_out, dim * elem_size * num_qpts * sizeof(CeedScalar), hipMemcpyHostToDevice)); 1063 mat_start += dim * elem_size * num_qpts; 1064 } 1065 } 1066 return CEED_ERROR_SUCCESS; 1067 } 1068 1069 //------------------------------------------------------------------------------ 1070 // Assemble matrix data for COO matrix of assembled operator. 1071 // The sparsity pattern is set by CeedOperatorLinearAssembleSymbolic. 1072 // 1073 // Note that this (and other assembly routines) currently assume only one active input restriction/basis per operator (could have multiple basis eval 1074 // modes). 1075 // TODO: allow multiple active input restrictions/basis objects 1076 //------------------------------------------------------------------------------ 1077 static int CeedSingleOperatorAssemble_Hip(CeedOperator op, CeedInt offset, CeedVector values) { 1078 Ceed ceed; 1079 CeedSize values_length = 0, assembled_qf_length = 0; 1080 CeedInt use_ceedsize_idx = 0; 1081 CeedScalar *values_array; 1082 const CeedScalar *qf_array; 1083 CeedVector assembled_qf = NULL; 1084 CeedElemRestriction rstr_q = NULL; 1085 CeedOperator_Hip *impl; 1086 1087 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1088 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1089 1090 // Assemble QFunction 1091 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &rstr_q, CEED_REQUEST_IMMEDIATE)); 1092 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_q)); 1093 CeedCallBackend(CeedVectorGetArray(values, CEED_MEM_DEVICE, &values_array)); 1094 values_array += offset; 1095 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &qf_array)); 1096 1097 CeedCallBackend(CeedVectorGetLength(values, &values_length)); 1098 CeedCallBackend(CeedVectorGetLength(assembled_qf, &assembled_qf_length)); 1099 if ((values_length > INT_MAX) || (assembled_qf_length > INT_MAX)) use_ceedsize_idx = 1; 1100 // Setup 1101 if (!impl->asmb) { 1102 CeedCallBackend(CeedSingleOperatorAssembleSetup_Hip(op, use_ceedsize_idx)); 1103 assert(impl->asmb != NULL); 1104 } 1105 1106 // Compute B^T D B 1107 const CeedInt num_elem = impl->asmb->num_elem; 1108 const CeedInt elem_per_block = impl->asmb->elem_per_block; 1109 const CeedInt grid = num_elem / elem_per_block + ((num_elem / elem_per_block * elem_per_block < num_elem) ? 1 : 0); 1110 void *args[] = {&impl->asmb->d_B_in, &impl->asmb->d_B_out, &qf_array, &values_array}; 1111 1112 CeedCallBackend( 1113 CeedRunKernelDim_Hip(ceed, impl->asmb->linearAssemble, grid, impl->asmb->block_size_x, impl->asmb->block_size_y, elem_per_block, args)); 1114 1115 // Restore arrays 1116 CeedCallBackend(CeedVectorRestoreArray(values, &values_array)); 1117 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &qf_array)); 1118 1119 // Cleanup 1120 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 1121 return CEED_ERROR_SUCCESS; 1122 } 1123 1124 //------------------------------------------------------------------------------ 1125 // Create operator 1126 //------------------------------------------------------------------------------ 1127 int CeedOperatorCreate_Hip(CeedOperator op) { 1128 Ceed ceed; 1129 CeedOperator_Hip *impl; 1130 1131 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1132 CeedCallBackend(CeedCalloc(1, &impl)); 1133 CeedCallBackend(CeedOperatorSetData(op, impl)); 1134 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunction_Hip)); 1135 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunctionUpdate", CeedOperatorLinearAssembleQFunctionUpdate_Hip)); 1136 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddDiagonal", CeedOperatorLinearAssembleAddDiagonal_Hip)); 1137 CeedCallBackend( 1138 CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddPointBlockDiagonal", CeedOperatorLinearAssembleAddPointBlockDiagonal_Hip)); 1139 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleSingle", CeedSingleOperatorAssemble_Hip)); 1140 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Hip)); 1141 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Hip)); 1142 return CEED_ERROR_SUCCESS; 1143 } 1144 1145 //------------------------------------------------------------------------------ 1146