1 // Copyright (c) 2017-2024, 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 CeedCallBackend(CeedFree(&impl->num_points)); 30 CeedCallBackend(CeedFree(&impl->skip_rstr_in)); 31 CeedCallBackend(CeedFree(&impl->skip_rstr_out)); 32 CeedCallBackend(CeedFree(&impl->apply_add_basis_out)); 33 CeedCallBackend(CeedFree(&impl->input_field_order)); 34 CeedCallBackend(CeedFree(&impl->output_field_order)); 35 CeedCallBackend(CeedFree(&impl->input_states)); 36 37 for (CeedInt i = 0; i < impl->num_inputs; i++) { 38 CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_in[i])); 39 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_in[i])); 40 } 41 CeedCallBackend(CeedFree(&impl->e_vecs_in)); 42 CeedCallBackend(CeedFree(&impl->q_vecs_in)); 43 44 for (CeedInt i = 0; i < impl->num_outputs; i++) { 45 CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_out[i])); 46 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_out[i])); 47 } 48 CeedCallBackend(CeedFree(&impl->e_vecs_out)); 49 CeedCallBackend(CeedFree(&impl->q_vecs_out)); 50 CeedCallBackend(CeedVectorDestroy(&impl->point_coords_elem)); 51 52 // QFunction assembly data 53 for (CeedInt i = 0; i < impl->num_active_in; i++) { 54 CeedCallBackend(CeedVectorDestroy(&impl->qf_active_in[i])); 55 } 56 CeedCallBackend(CeedFree(&impl->qf_active_in)); 57 58 // Diag data 59 if (impl->diag) { 60 Ceed ceed; 61 62 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 63 if (impl->diag->module) { 64 CeedCallHip(ceed, hipModuleUnload(impl->diag->module)); 65 } 66 if (impl->diag->module_point_block) { 67 CeedCallHip(ceed, hipModuleUnload(impl->diag->module_point_block)); 68 } 69 CeedCallHip(ceed, hipFree(impl->diag->d_eval_modes_in)); 70 CeedCallHip(ceed, hipFree(impl->diag->d_eval_modes_out)); 71 CeedCallHip(ceed, hipFree(impl->diag->d_identity)); 72 CeedCallHip(ceed, hipFree(impl->diag->d_interp_in)); 73 CeedCallHip(ceed, hipFree(impl->diag->d_interp_out)); 74 CeedCallHip(ceed, hipFree(impl->diag->d_grad_in)); 75 CeedCallHip(ceed, hipFree(impl->diag->d_grad_out)); 76 CeedCallHip(ceed, hipFree(impl->diag->d_div_in)); 77 CeedCallHip(ceed, hipFree(impl->diag->d_div_out)); 78 CeedCallHip(ceed, hipFree(impl->diag->d_curl_in)); 79 CeedCallHip(ceed, hipFree(impl->diag->d_curl_out)); 80 CeedCallBackend(CeedElemRestrictionDestroy(&impl->diag->diag_rstr)); 81 CeedCallBackend(CeedElemRestrictionDestroy(&impl->diag->point_block_diag_rstr)); 82 CeedCallBackend(CeedVectorDestroy(&impl->diag->elem_diag)); 83 CeedCallBackend(CeedVectorDestroy(&impl->diag->point_block_elem_diag)); 84 } 85 CeedCallBackend(CeedFree(&impl->diag)); 86 87 if (impl->asmb) { 88 Ceed ceed; 89 90 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 91 CeedCallHip(ceed, hipModuleUnload(impl->asmb->module)); 92 CeedCallHip(ceed, hipFree(impl->asmb->d_B_in)); 93 CeedCallHip(ceed, hipFree(impl->asmb->d_B_out)); 94 } 95 CeedCallBackend(CeedFree(&impl->asmb)); 96 97 CeedCallBackend(CeedFree(&impl)); 98 return CEED_ERROR_SUCCESS; 99 } 100 101 //------------------------------------------------------------------------------ 102 // Setup infields or outfields 103 //------------------------------------------------------------------------------ 104 static int CeedOperatorSetupFields_Hip(CeedQFunction qf, CeedOperator op, bool is_input, bool is_at_points, bool *skip_rstr, bool *apply_add_basis, 105 CeedVector *e_vecs, CeedVector *q_vecs, CeedInt num_fields, CeedInt Q, CeedInt num_elem) { 106 Ceed ceed; 107 CeedQFunctionField *qf_fields; 108 CeedOperatorField *op_fields; 109 110 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 111 if (is_input) { 112 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 113 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 114 } else { 115 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 116 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 117 } 118 119 // Loop over fields 120 for (CeedInt i = 0; i < num_fields; i++) { 121 bool is_active = false, is_strided = false, skip_e_vec = false; 122 CeedSize q_size; 123 CeedInt size; 124 CeedEvalMode eval_mode; 125 CeedVector l_vec; 126 CeedElemRestriction elem_rstr; 127 128 // Check whether this field can skip the element restriction: 129 // Input CEED_VECTOR_ACTIVE 130 // Output CEED_VECTOR_ACTIVE without CEED_EVAL_NONE 131 // Input CEED_VECTOR_NONE with CEED_EVAL_WEIGHT 132 // Input passive vectorr with CEED_EVAL_NONE and strided restriction with CEED_STRIDES_BACKEND 133 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &l_vec)); 134 is_active = l_vec == CEED_VECTOR_ACTIVE; 135 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &elem_rstr)); 136 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 137 skip_e_vec = (is_input && is_active) || (is_active && eval_mode != CEED_EVAL_NONE) || (eval_mode == CEED_EVAL_WEIGHT); 138 if (!skip_e_vec && is_input && !is_active && eval_mode == CEED_EVAL_NONE) { 139 CeedCallBackend(CeedElemRestrictionIsStrided(elem_rstr, &is_strided)); 140 if (is_strided) CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_rstr, &skip_e_vec)); 141 } 142 if (skip_e_vec) { 143 e_vecs[i] = NULL; 144 } else { 145 CeedCallBackend(CeedElemRestrictionCreateVector(elem_rstr, NULL, &e_vecs[i])); 146 } 147 148 switch (eval_mode) { 149 case CEED_EVAL_NONE: 150 case CEED_EVAL_INTERP: 151 case CEED_EVAL_GRAD: 152 case CEED_EVAL_DIV: 153 case CEED_EVAL_CURL: 154 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 155 q_size = (CeedSize)num_elem * (CeedSize)Q * (CeedSize)size; 156 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 157 break; 158 case CEED_EVAL_WEIGHT: { 159 CeedBasis basis; 160 161 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 162 q_size = (CeedSize)num_elem * (CeedSize)Q; 163 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 164 if (is_at_points) { 165 CeedInt num_points[num_elem]; 166 167 for (CeedInt i = 0; i < num_elem; i++) num_points[i] = Q; 168 CeedCallBackend( 169 CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, CEED_VECTOR_NONE, q_vecs[i])); 170 } else { 171 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, q_vecs[i])); 172 } 173 break; 174 } 175 } 176 } 177 // Drop duplicate restrictions 178 if (is_input) { 179 for (CeedInt i = 0; i < num_fields; i++) { 180 CeedVector vec_i; 181 CeedElemRestriction rstr_i; 182 183 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec_i)); 184 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr_i)); 185 for (CeedInt j = i + 1; j < num_fields; j++) { 186 CeedVector vec_j; 187 CeedElemRestriction rstr_j; 188 189 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[j], &vec_j)); 190 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[j], &rstr_j)); 191 if (vec_i == vec_j && rstr_i == rstr_j) { 192 if (e_vecs[i]) CeedCallBackend(CeedVectorReferenceCopy(e_vecs[i], &e_vecs[j])); 193 skip_rstr[j] = true; 194 } 195 } 196 } 197 } else { 198 for (CeedInt i = num_fields - 1; i >= 0; i--) { 199 CeedVector vec_i; 200 CeedElemRestriction rstr_i; 201 202 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec_i)); 203 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr_i)); 204 for (CeedInt j = i - 1; j >= 0; j--) { 205 CeedVector vec_j; 206 CeedElemRestriction rstr_j; 207 208 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[j], &vec_j)); 209 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[j], &rstr_j)); 210 if (vec_i == vec_j && rstr_i == rstr_j) { 211 if (e_vecs[i]) CeedCallBackend(CeedVectorReferenceCopy(e_vecs[i], &e_vecs[j])); 212 skip_rstr[j] = true; 213 apply_add_basis[i] = true; 214 } 215 } 216 } 217 } 218 return CEED_ERROR_SUCCESS; 219 } 220 221 //------------------------------------------------------------------------------ 222 // CeedOperator needs to connect all the named fields (be they active or passive) to the named inputs and outputs of its CeedQFunction. 223 //------------------------------------------------------------------------------ 224 static int CeedOperatorSetup_Hip(CeedOperator op) { 225 Ceed ceed; 226 bool is_setup_done; 227 CeedInt Q, num_elem, num_input_fields, num_output_fields; 228 CeedQFunctionField *qf_input_fields, *qf_output_fields; 229 CeedQFunction qf; 230 CeedOperatorField *op_input_fields, *op_output_fields; 231 CeedOperator_Hip *impl; 232 233 CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done)); 234 if (is_setup_done) return CEED_ERROR_SUCCESS; 235 236 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 237 CeedCallBackend(CeedOperatorGetData(op, &impl)); 238 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 239 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 240 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 241 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 242 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 243 244 // Allocate 245 CeedCallBackend(CeedCalloc(num_input_fields, &impl->e_vecs_in)); 246 CeedCallBackend(CeedCalloc(num_output_fields, &impl->e_vecs_out)); 247 CeedCallBackend(CeedCalloc(num_input_fields, &impl->skip_rstr_in)); 248 CeedCallBackend(CeedCalloc(num_output_fields, &impl->skip_rstr_out)); 249 CeedCallBackend(CeedCalloc(num_output_fields, &impl->apply_add_basis_out)); 250 CeedCallBackend(CeedCalloc(num_input_fields, &impl->input_field_order)); 251 CeedCallBackend(CeedCalloc(num_output_fields, &impl->output_field_order)); 252 CeedCallBackend(CeedCalloc(num_input_fields, &impl->input_states)); 253 CeedCallBackend(CeedCalloc(num_input_fields, &impl->q_vecs_in)); 254 CeedCallBackend(CeedCalloc(num_output_fields, &impl->q_vecs_out)); 255 impl->num_inputs = num_input_fields; 256 impl->num_outputs = num_output_fields; 257 258 // Set up infield and outfield e-vecs and q-vecs 259 CeedCallBackend( 260 CeedOperatorSetupFields_Hip(qf, op, true, false, impl->skip_rstr_in, NULL, impl->e_vecs_in, impl->q_vecs_in, num_input_fields, Q, num_elem)); 261 CeedCallBackend(CeedOperatorSetupFields_Hip(qf, op, false, false, impl->skip_rstr_out, impl->apply_add_basis_out, impl->e_vecs_out, 262 impl->q_vecs_out, num_output_fields, Q, num_elem)); 263 264 // Reorder fields to allow reuse of buffers 265 impl->max_active_e_vec_len = 0; 266 { 267 bool is_ordered[CEED_FIELD_MAX]; 268 CeedInt curr_index = 0; 269 270 for (CeedInt i = 0; i < num_input_fields; i++) is_ordered[i] = false; 271 for (CeedInt i = 0; i < num_input_fields; i++) { 272 CeedSize e_vec_len_i; 273 CeedVector vec_i; 274 CeedElemRestriction rstr_i; 275 276 if (is_ordered[i]) continue; 277 is_ordered[i] = true; 278 impl->input_field_order[curr_index] = i; 279 curr_index++; 280 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec_i)); 281 if (vec_i == CEED_VECTOR_NONE) continue; // CEED_EVAL_WEIGHT 282 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &rstr_i)); 283 CeedCallBackend(CeedElemRestrictionGetEVectorSize(rstr_i, &e_vec_len_i)); 284 impl->max_active_e_vec_len = e_vec_len_i > impl->max_active_e_vec_len ? e_vec_len_i : impl->max_active_e_vec_len; 285 for (CeedInt j = i + 1; j < num_input_fields; j++) { 286 CeedVector vec_j; 287 CeedElemRestriction rstr_j; 288 289 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[j], &vec_j)); 290 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[j], &rstr_j)); 291 if (rstr_i == rstr_j && vec_i == vec_j) { 292 is_ordered[j] = true; 293 impl->input_field_order[curr_index] = j; 294 curr_index++; 295 } 296 } 297 } 298 } 299 { 300 bool is_ordered[CEED_FIELD_MAX]; 301 CeedInt curr_index = 0; 302 303 for (CeedInt i = 0; i < num_output_fields; i++) is_ordered[i] = false; 304 for (CeedInt i = 0; i < num_output_fields; i++) { 305 CeedSize e_vec_len_i; 306 CeedVector vec_i; 307 CeedElemRestriction rstr_i; 308 309 if (is_ordered[i]) continue; 310 is_ordered[i] = true; 311 impl->output_field_order[curr_index] = i; 312 curr_index++; 313 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec_i)); 314 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &rstr_i)); 315 CeedCallBackend(CeedElemRestrictionGetEVectorSize(rstr_i, &e_vec_len_i)); 316 impl->max_active_e_vec_len = e_vec_len_i > impl->max_active_e_vec_len ? e_vec_len_i : impl->max_active_e_vec_len; 317 for (CeedInt j = i + 1; j < num_output_fields; j++) { 318 CeedVector vec_j; 319 CeedElemRestriction rstr_j; 320 321 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[j], &vec_j)); 322 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[j], &rstr_j)); 323 if (rstr_i == rstr_j && vec_i == vec_j) { 324 is_ordered[j] = true; 325 impl->output_field_order[curr_index] = j; 326 curr_index++; 327 } 328 } 329 } 330 } 331 CeedCallBackend(CeedOperatorSetSetupDone(op)); 332 return CEED_ERROR_SUCCESS; 333 } 334 335 //------------------------------------------------------------------------------ 336 // Restrict Operator Inputs 337 //------------------------------------------------------------------------------ 338 static inline int CeedOperatorInputRestrict_Hip(CeedOperatorField op_input_field, CeedQFunctionField qf_input_field, CeedInt input_field, 339 CeedVector in_vec, CeedVector active_e_vec, const bool skip_active, CeedOperator_Hip *impl, 340 CeedRequest *request) { 341 bool is_active = false; 342 CeedVector l_vec, e_vec = impl->e_vecs_in[input_field]; 343 344 // Get input vector 345 CeedCallBackend(CeedOperatorFieldGetVector(op_input_field, &l_vec)); 346 is_active = l_vec == CEED_VECTOR_ACTIVE; 347 if (is_active && skip_active) return CEED_ERROR_SUCCESS; 348 if (is_active) { 349 l_vec = in_vec; 350 if (!e_vec) e_vec = active_e_vec; 351 } 352 353 // Restriction action 354 if (e_vec) { 355 // Restrict, if necessary 356 if (!impl->skip_rstr_in[input_field]) { 357 uint64_t state; 358 359 CeedCallBackend(CeedVectorGetState(l_vec, &state)); 360 if (is_active || state != impl->input_states[input_field]) { 361 CeedElemRestriction elem_rstr; 362 363 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_field, &elem_rstr)); 364 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_NOTRANSPOSE, l_vec, e_vec, request)); 365 } 366 impl->input_states[input_field] = state; 367 } 368 } 369 return CEED_ERROR_SUCCESS; 370 } 371 372 //------------------------------------------------------------------------------ 373 // Input Basis Action 374 //------------------------------------------------------------------------------ 375 static inline int CeedOperatorInputBasis_Hip(CeedOperatorField op_input_field, CeedQFunctionField qf_input_field, CeedInt input_field, 376 CeedVector in_vec, CeedVector active_e_vec, CeedInt num_elem, const bool skip_active, 377 CeedOperator_Hip *impl) { 378 bool is_active = false; 379 CeedEvalMode eval_mode; 380 CeedVector l_vec, e_vec = impl->e_vecs_in[input_field], q_vec = impl->q_vecs_in[input_field]; 381 382 // Skip active input 383 CeedCallBackend(CeedOperatorFieldGetVector(op_input_field, &l_vec)); 384 is_active = l_vec == CEED_VECTOR_ACTIVE; 385 if (is_active && skip_active) return CEED_ERROR_SUCCESS; 386 if (is_active) { 387 l_vec = in_vec; 388 if (!e_vec) e_vec = active_e_vec; 389 } 390 391 // Basis action 392 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_field, &eval_mode)); 393 switch (eval_mode) { 394 case CEED_EVAL_NONE: { 395 const CeedScalar *e_vec_array; 396 397 if (e_vec) { 398 CeedCallBackend(CeedVectorGetArrayRead(e_vec, CEED_MEM_DEVICE, &e_vec_array)); 399 } else { 400 CeedCallBackend(CeedVectorGetArrayRead(l_vec, CEED_MEM_DEVICE, &e_vec_array)); 401 } 402 CeedCallBackend(CeedVectorSetArray(q_vec, CEED_MEM_DEVICE, CEED_USE_POINTER, (CeedScalar *)e_vec_array)); 403 break; 404 } 405 case CEED_EVAL_INTERP: 406 case CEED_EVAL_GRAD: 407 case CEED_EVAL_DIV: 408 case CEED_EVAL_CURL: { 409 CeedBasis basis; 410 411 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_field, &basis)); 412 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, eval_mode, e_vec, q_vec)); 413 break; 414 } 415 case CEED_EVAL_WEIGHT: 416 break; // No action 417 } 418 return CEED_ERROR_SUCCESS; 419 } 420 421 //------------------------------------------------------------------------------ 422 // Restore Input Vectors 423 //------------------------------------------------------------------------------ 424 static inline int CeedOperatorInputRestore_Hip(CeedOperatorField op_input_field, CeedQFunctionField qf_input_field, CeedInt input_field, 425 CeedVector in_vec, CeedVector active_e_vec, const bool skip_active, CeedOperator_Hip *impl) { 426 bool is_active = false; 427 CeedEvalMode eval_mode; 428 CeedVector l_vec, e_vec = impl->e_vecs_in[input_field]; 429 430 // Skip active input 431 CeedCallBackend(CeedOperatorFieldGetVector(op_input_field, &l_vec)); 432 is_active = l_vec == CEED_VECTOR_ACTIVE; 433 if (is_active && skip_active) return CEED_ERROR_SUCCESS; 434 if (is_active) { 435 l_vec = in_vec; 436 if (!e_vec) e_vec = active_e_vec; 437 } 438 439 // Restore e-vec 440 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_field, &eval_mode)); 441 if (eval_mode == CEED_EVAL_NONE) { 442 const CeedScalar *e_vec_array; 443 444 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_in[input_field], CEED_MEM_DEVICE, (CeedScalar **)&e_vec_array)); 445 if (e_vec) { 446 CeedCallBackend(CeedVectorRestoreArrayRead(e_vec, &e_vec_array)); 447 } else { 448 CeedCallBackend(CeedVectorRestoreArrayRead(l_vec, &e_vec_array)); 449 } 450 } 451 return CEED_ERROR_SUCCESS; 452 } 453 454 //------------------------------------------------------------------------------ 455 // Apply and add to output 456 //------------------------------------------------------------------------------ 457 static int CeedOperatorApplyAdd_Hip(CeedOperator op, CeedVector in_vec, CeedVector out_vec, CeedRequest *request) { 458 CeedInt Q, num_elem, num_input_fields, num_output_fields; 459 Ceed ceed; 460 CeedVector active_e_vec; 461 CeedQFunctionField *qf_input_fields, *qf_output_fields; 462 CeedQFunction qf; 463 CeedOperatorField *op_input_fields, *op_output_fields; 464 CeedOperator_Hip *impl; 465 466 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 467 CeedCallBackend(CeedOperatorGetData(op, &impl)); 468 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 469 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 470 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 471 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 472 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 473 474 // Setup 475 CeedCallBackend(CeedOperatorSetup_Hip(op)); 476 477 // Work vector 478 CeedCallBackend(CeedGetWorkVector(ceed, impl->max_active_e_vec_len, &active_e_vec)); 479 480 // Process inputs 481 for (CeedInt i = 0; i < num_input_fields; i++) { 482 CeedInt field = impl->input_field_order[i]; 483 484 CeedCallBackend(CeedOperatorInputRestrict_Hip(op_input_fields[field], qf_input_fields[field], field, in_vec, active_e_vec, false, impl, request)); 485 CeedCallBackend(CeedOperatorInputBasis_Hip(op_input_fields[field], qf_input_fields[field], field, in_vec, active_e_vec, num_elem, false, impl)); 486 } 487 488 // Output pointers, as necessary 489 for (CeedInt i = 0; i < num_output_fields; i++) { 490 CeedEvalMode eval_mode; 491 492 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 493 if (eval_mode == CEED_EVAL_NONE) { 494 CeedScalar *e_vec_array; 495 496 CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 497 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_vec_array)); 498 } 499 } 500 501 // Q function 502 CeedCallBackend(CeedQFunctionApply(qf, num_elem * Q, impl->q_vecs_in, impl->q_vecs_out)); 503 504 // Restore input arrays 505 for (CeedInt i = 0; i < num_input_fields; i++) { 506 CeedCallBackend(CeedOperatorInputRestore_Hip(op_input_fields[i], qf_input_fields[i], i, in_vec, active_e_vec, false, impl)); 507 } 508 509 // Output basis and restriction 510 for (CeedInt i = 0; i < num_output_fields; i++) { 511 bool is_active = false; 512 CeedInt field = impl->output_field_order[i]; 513 CeedEvalMode eval_mode; 514 CeedVector l_vec, e_vec = impl->e_vecs_out[field], q_vec = impl->q_vecs_out[field]; 515 CeedElemRestriction elem_rstr; 516 CeedBasis basis; 517 518 // Output vector 519 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[field], &l_vec)); 520 is_active = l_vec == CEED_VECTOR_ACTIVE; 521 if (is_active) { 522 l_vec = out_vec; 523 if (!e_vec) e_vec = active_e_vec; 524 } 525 526 // Basis action 527 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[field], &eval_mode)); 528 switch (eval_mode) { 529 case CEED_EVAL_NONE: 530 break; // No action 531 case CEED_EVAL_INTERP: 532 case CEED_EVAL_GRAD: 533 case CEED_EVAL_DIV: 534 case CEED_EVAL_CURL: 535 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[field], &basis)); 536 if (impl->apply_add_basis_out[field]) { 537 CeedCallBackend(CeedBasisApplyAdd(basis, num_elem, CEED_TRANSPOSE, eval_mode, q_vec, e_vec)); 538 } else { 539 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_TRANSPOSE, eval_mode, q_vec, e_vec)); 540 } 541 break; 542 // LCOV_EXCL_START 543 case CEED_EVAL_WEIGHT: { 544 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 545 // LCOV_EXCL_STOP 546 } 547 } 548 549 // Restore evec 550 if (eval_mode == CEED_EVAL_NONE) { 551 CeedScalar *e_vec_array; 552 553 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 554 CeedCallBackend(CeedVectorRestoreArray(e_vec, &e_vec_array)); 555 } 556 557 // Restrict 558 if (impl->skip_rstr_out[field]) continue; 559 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[field], &elem_rstr)); 560 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_TRANSPOSE, e_vec, l_vec, request)); 561 } 562 563 // Return work vector 564 CeedCallBackend(CeedRestoreWorkVector(ceed, &active_e_vec)); 565 return CEED_ERROR_SUCCESS; 566 } 567 568 //------------------------------------------------------------------------------ 569 // CeedOperator needs to connect all the named fields (be they active or passive) to the named inputs and outputs of its CeedQFunction. 570 //------------------------------------------------------------------------------ 571 static int CeedOperatorSetupAtPoints_Hip(CeedOperator op) { 572 Ceed ceed; 573 bool is_setup_done; 574 CeedInt max_num_points = -1, num_elem, num_input_fields, num_output_fields; 575 CeedQFunctionField *qf_input_fields, *qf_output_fields; 576 CeedQFunction qf; 577 CeedOperatorField *op_input_fields, *op_output_fields; 578 CeedOperator_Hip *impl; 579 580 CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done)); 581 if (is_setup_done) return CEED_ERROR_SUCCESS; 582 583 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 584 CeedCallBackend(CeedOperatorGetData(op, &impl)); 585 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 586 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 587 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 588 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 589 { 590 CeedElemRestriction rstr_points = NULL; 591 592 CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, NULL)); 593 CeedCallBackend(CeedElemRestrictionGetMaxPointsInElement(rstr_points, &max_num_points)); 594 CeedCallBackend(CeedCalloc(num_elem, &impl->num_points)); 595 for (CeedInt e = 0; e < num_elem; e++) { 596 CeedInt num_points_elem; 597 598 CeedCallBackend(CeedElemRestrictionGetNumPointsInElement(rstr_points, e, &num_points_elem)); 599 impl->num_points[e] = num_points_elem; 600 } 601 } 602 impl->max_num_points = max_num_points; 603 604 // Allocate 605 CeedCallBackend(CeedCalloc(num_input_fields, &impl->e_vecs_in)); 606 CeedCallBackend(CeedCalloc(num_output_fields, &impl->e_vecs_out)); 607 CeedCallBackend(CeedCalloc(num_input_fields, &impl->skip_rstr_in)); 608 CeedCallBackend(CeedCalloc(num_output_fields, &impl->skip_rstr_out)); 609 CeedCallBackend(CeedCalloc(num_output_fields, &impl->apply_add_basis_out)); 610 CeedCallBackend(CeedCalloc(num_input_fields, &impl->input_field_order)); 611 CeedCallBackend(CeedCalloc(num_output_fields, &impl->output_field_order)); 612 CeedCallBackend(CeedCalloc(num_input_fields, &impl->input_states)); 613 CeedCallBackend(CeedCalloc(num_input_fields, &impl->q_vecs_in)); 614 CeedCallBackend(CeedCalloc(num_output_fields, &impl->q_vecs_out)); 615 impl->num_inputs = num_input_fields; 616 impl->num_outputs = num_output_fields; 617 618 // Set up infield and outfield e-vecs and q-vecs 619 CeedCallBackend(CeedOperatorSetupFields_Hip(qf, op, true, true, impl->skip_rstr_in, NULL, impl->e_vecs_in, impl->q_vecs_in, num_input_fields, 620 max_num_points, num_elem)); 621 CeedCallBackend(CeedOperatorSetupFields_Hip(qf, op, false, true, impl->skip_rstr_out, impl->apply_add_basis_out, impl->e_vecs_out, impl->q_vecs_out, 622 num_output_fields, max_num_points, num_elem)); 623 624 // Reorder fields to allow reuse of buffers 625 impl->max_active_e_vec_len = 0; 626 { 627 bool is_ordered[CEED_FIELD_MAX]; 628 CeedInt curr_index = 0; 629 630 for (CeedInt i = 0; i < num_input_fields; i++) is_ordered[i] = false; 631 for (CeedInt i = 0; i < num_input_fields; i++) { 632 CeedSize e_vec_len_i; 633 CeedVector vec_i; 634 CeedElemRestriction rstr_i; 635 636 if (is_ordered[i]) continue; 637 is_ordered[i] = true; 638 impl->input_field_order[curr_index] = i; 639 curr_index++; 640 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec_i)); 641 if (vec_i == CEED_VECTOR_NONE) continue; // CEED_EVAL_WEIGHT 642 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &rstr_i)); 643 CeedCallBackend(CeedElemRestrictionGetEVectorSize(rstr_i, &e_vec_len_i)); 644 impl->max_active_e_vec_len = e_vec_len_i > impl->max_active_e_vec_len ? e_vec_len_i : impl->max_active_e_vec_len; 645 for (CeedInt j = i + 1; j < num_input_fields; j++) { 646 CeedVector vec_j; 647 CeedElemRestriction rstr_j; 648 649 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[j], &vec_j)); 650 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[j], &rstr_j)); 651 if (rstr_i == rstr_j && vec_i == vec_j) { 652 is_ordered[j] = true; 653 impl->input_field_order[curr_index] = j; 654 curr_index++; 655 } 656 } 657 } 658 } 659 { 660 bool is_ordered[CEED_FIELD_MAX]; 661 CeedInt curr_index = 0; 662 663 for (CeedInt i = 0; i < num_output_fields; i++) is_ordered[i] = false; 664 for (CeedInt i = 0; i < num_output_fields; i++) { 665 CeedSize e_vec_len_i; 666 CeedVector vec_i; 667 CeedElemRestriction rstr_i; 668 669 if (is_ordered[i]) continue; 670 is_ordered[i] = true; 671 impl->output_field_order[curr_index] = i; 672 curr_index++; 673 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec_i)); 674 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &rstr_i)); 675 CeedCallBackend(CeedElemRestrictionGetEVectorSize(rstr_i, &e_vec_len_i)); 676 impl->max_active_e_vec_len = e_vec_len_i > impl->max_active_e_vec_len ? e_vec_len_i : impl->max_active_e_vec_len; 677 for (CeedInt j = i + 1; j < num_output_fields; j++) { 678 CeedVector vec_j; 679 CeedElemRestriction rstr_j; 680 681 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[j], &vec_j)); 682 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[j], &rstr_j)); 683 if (rstr_i == rstr_j && vec_i == vec_j) { 684 is_ordered[j] = true; 685 impl->output_field_order[curr_index] = j; 686 curr_index++; 687 } 688 } 689 } 690 } 691 692 CeedCallBackend(CeedOperatorSetSetupDone(op)); 693 return CEED_ERROR_SUCCESS; 694 } 695 696 //------------------------------------------------------------------------------ 697 // Input Basis Action AtPoints 698 //------------------------------------------------------------------------------ 699 static inline int CeedOperatorInputBasisAtPoints_Hip(CeedOperatorField op_input_field, CeedQFunctionField qf_input_field, CeedInt input_field, 700 CeedVector in_vec, CeedVector active_e_vec, CeedInt num_elem, const CeedInt *num_points, 701 const bool skip_active, CeedOperator_Hip *impl) { 702 bool is_active = false; 703 CeedEvalMode eval_mode; 704 CeedVector l_vec, e_vec = impl->e_vecs_in[input_field], q_vec = impl->q_vecs_in[input_field]; 705 706 // Skip active input 707 CeedCallBackend(CeedOperatorFieldGetVector(op_input_field, &l_vec)); 708 is_active = l_vec == CEED_VECTOR_ACTIVE; 709 if (is_active && skip_active) return CEED_ERROR_SUCCESS; 710 if (is_active) { 711 l_vec = in_vec; 712 if (!e_vec) e_vec = active_e_vec; 713 } 714 715 // Basis action 716 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_field, &eval_mode)); 717 switch (eval_mode) { 718 case CEED_EVAL_NONE: { 719 const CeedScalar *e_vec_array; 720 721 if (e_vec) { 722 CeedCallBackend(CeedVectorGetArrayRead(e_vec, CEED_MEM_DEVICE, &e_vec_array)); 723 } else { 724 CeedCallBackend(CeedVectorGetArrayRead(l_vec, CEED_MEM_DEVICE, &e_vec_array)); 725 } 726 CeedCallBackend(CeedVectorSetArray(q_vec, CEED_MEM_DEVICE, CEED_USE_POINTER, (CeedScalar *)e_vec_array)); 727 break; 728 } 729 case CEED_EVAL_INTERP: 730 case CEED_EVAL_GRAD: 731 case CEED_EVAL_DIV: 732 case CEED_EVAL_CURL: { 733 CeedBasis basis; 734 735 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_field, &basis)); 736 CeedCallBackend(CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_NOTRANSPOSE, eval_mode, impl->point_coords_elem, e_vec, q_vec)); 737 break; 738 } 739 case CEED_EVAL_WEIGHT: 740 break; // No action 741 } 742 return CEED_ERROR_SUCCESS; 743 } 744 745 //------------------------------------------------------------------------------ 746 // Apply and add to output AtPoints 747 //------------------------------------------------------------------------------ 748 static int CeedOperatorApplyAddAtPoints_Hip(CeedOperator op, CeedVector in_vec, CeedVector out_vec, CeedRequest *request) { 749 CeedInt max_num_points, *num_points, num_elem, num_input_fields, num_output_fields; 750 Ceed ceed; 751 CeedVector active_e_vec; 752 CeedQFunctionField *qf_input_fields, *qf_output_fields; 753 CeedQFunction qf; 754 CeedOperatorField *op_input_fields, *op_output_fields; 755 CeedOperator_Hip *impl; 756 757 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 758 CeedCallBackend(CeedOperatorGetData(op, &impl)); 759 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 760 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 761 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 762 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 763 764 // Setup 765 CeedCallBackend(CeedOperatorSetupAtPoints_Hip(op)); 766 num_points = impl->num_points; 767 max_num_points = impl->max_num_points; 768 769 // Work vector 770 CeedCallBackend(CeedGetWorkVector(ceed, impl->max_active_e_vec_len, &active_e_vec)); 771 772 // Get point coordinates 773 if (!impl->point_coords_elem) { 774 CeedVector point_coords = NULL; 775 CeedElemRestriction rstr_points = NULL; 776 777 CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, &point_coords)); 778 CeedCallBackend(CeedElemRestrictionCreateVector(rstr_points, NULL, &impl->point_coords_elem)); 779 CeedCallBackend(CeedElemRestrictionApply(rstr_points, CEED_NOTRANSPOSE, point_coords, impl->point_coords_elem, request)); 780 } 781 782 // Process inputs 783 for (CeedInt i = 0; i < num_input_fields; i++) { 784 CeedInt field = impl->input_field_order[i]; 785 786 CeedCallBackend(CeedOperatorInputRestrict_Hip(op_input_fields[field], qf_input_fields[field], field, in_vec, active_e_vec, false, impl, request)); 787 CeedCallBackend(CeedOperatorInputBasisAtPoints_Hip(op_input_fields[field], qf_input_fields[field], field, in_vec, active_e_vec, num_elem, 788 num_points, false, impl)); 789 } 790 791 // Output pointers, as necessary 792 for (CeedInt i = 0; i < num_output_fields; i++) { 793 CeedEvalMode eval_mode; 794 795 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 796 if (eval_mode == CEED_EVAL_NONE) { 797 CeedScalar *e_vec_array; 798 799 CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 800 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_vec_array)); 801 } 802 } 803 804 // Q function 805 CeedCallBackend(CeedQFunctionApply(qf, num_elem * max_num_points, impl->q_vecs_in, impl->q_vecs_out)); 806 807 // Restore input arrays 808 for (CeedInt i = 0; i < num_input_fields; i++) { 809 CeedCallBackend(CeedOperatorInputRestore_Hip(op_input_fields[i], qf_input_fields[i], i, in_vec, active_e_vec, false, impl)); 810 } 811 812 // Output basis and restriction 813 for (CeedInt i = 0; i < num_output_fields; i++) { 814 bool is_active = false; 815 CeedInt field = impl->output_field_order[i]; 816 CeedEvalMode eval_mode; 817 CeedVector l_vec, e_vec = impl->e_vecs_out[field], q_vec = impl->q_vecs_out[field]; 818 CeedElemRestriction elem_rstr; 819 CeedBasis basis; 820 821 // Output vector 822 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[field], &l_vec)); 823 is_active = l_vec == CEED_VECTOR_ACTIVE; 824 if (is_active) { 825 l_vec = out_vec; 826 if (!e_vec) e_vec = active_e_vec; 827 } 828 829 // Basis action 830 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[field], &eval_mode)); 831 switch (eval_mode) { 832 case CEED_EVAL_NONE: 833 break; // No action 834 case CEED_EVAL_INTERP: 835 case CEED_EVAL_GRAD: 836 case CEED_EVAL_DIV: 837 case CEED_EVAL_CURL: 838 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[field], &basis)); 839 if (impl->apply_add_basis_out[field]) { 840 CeedCallBackend(CeedBasisApplyAddAtPoints(basis, num_elem, num_points, CEED_TRANSPOSE, eval_mode, impl->point_coords_elem, q_vec, e_vec)); 841 } else { 842 CeedCallBackend(CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_TRANSPOSE, eval_mode, impl->point_coords_elem, q_vec, e_vec)); 843 } 844 break; 845 // LCOV_EXCL_START 846 case CEED_EVAL_WEIGHT: { 847 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 848 // LCOV_EXCL_STOP 849 } 850 } 851 852 // Restore evec 853 if (eval_mode == CEED_EVAL_NONE) { 854 CeedScalar *e_vec_array; 855 856 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 857 CeedCallBackend(CeedVectorRestoreArray(e_vec, &e_vec_array)); 858 } 859 860 // Restrict 861 if (impl->skip_rstr_out[field]) continue; 862 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[field], &elem_rstr)); 863 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_TRANSPOSE, e_vec, l_vec, request)); 864 } 865 866 // Restore work vector 867 CeedCallBackend(CeedRestoreWorkVector(ceed, &active_e_vec)); 868 return CEED_ERROR_SUCCESS; 869 } 870 871 //------------------------------------------------------------------------------ 872 // Linear QFunction Assembly Core 873 //------------------------------------------------------------------------------ 874 static inline int CeedOperatorLinearAssembleQFunctionCore_Hip(CeedOperator op, bool build_objects, CeedVector *assembled, CeedElemRestriction *rstr, 875 CeedRequest *request) { 876 Ceed ceed, ceed_parent; 877 CeedInt num_active_in, num_active_out, Q, num_elem, num_input_fields, num_output_fields, size; 878 CeedScalar *assembled_array; 879 CeedVector *active_inputs; 880 CeedQFunctionField *qf_input_fields, *qf_output_fields; 881 CeedQFunction qf; 882 CeedOperatorField *op_input_fields, *op_output_fields; 883 CeedOperator_Hip *impl; 884 885 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 886 CeedCallBackend(CeedOperatorGetFallbackParentCeed(op, &ceed_parent)); 887 CeedCallBackend(CeedOperatorGetData(op, &impl)); 888 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 889 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 890 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 891 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 892 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 893 active_inputs = impl->qf_active_in; 894 num_active_in = impl->num_active_in, num_active_out = impl->num_active_out; 895 896 // Setup 897 CeedCallBackend(CeedOperatorSetup_Hip(op)); 898 899 // Process inputs 900 for (CeedInt i = 0; i < num_input_fields; i++) { 901 CeedCallBackend(CeedOperatorInputRestrict_Hip(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, true, impl, request)); 902 CeedCallBackend(CeedOperatorInputBasis_Hip(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, num_elem, true, impl)); 903 } 904 905 // Count number of active input fields 906 if (!num_active_in) { 907 for (CeedInt i = 0; i < num_input_fields; i++) { 908 CeedScalar *q_vec_array; 909 CeedVector l_vec; 910 911 // Check if active input 912 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &l_vec)); 913 if (l_vec == CEED_VECTOR_ACTIVE) { 914 CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size)); 915 CeedCallBackend(CeedVectorSetValue(impl->q_vecs_in[i], 0.0)); 916 CeedCallBackend(CeedVectorGetArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, &q_vec_array)); 917 CeedCallBackend(CeedRealloc(num_active_in + size, &active_inputs)); 918 for (CeedInt field = 0; field < size; field++) { 919 CeedSize q_size = (CeedSize)Q * num_elem; 920 921 CeedCallBackend(CeedVectorCreate(ceed, q_size, &active_inputs[num_active_in + field])); 922 CeedCallBackend( 923 CeedVectorSetArray(active_inputs[num_active_in + field], CEED_MEM_DEVICE, CEED_USE_POINTER, &q_vec_array[field * Q * num_elem])); 924 } 925 num_active_in += size; 926 CeedCallBackend(CeedVectorRestoreArray(impl->q_vecs_in[i], &q_vec_array)); 927 } 928 } 929 impl->num_active_in = num_active_in; 930 impl->qf_active_in = active_inputs; 931 } 932 933 // Count number of active output fields 934 if (!num_active_out) { 935 for (CeedInt i = 0; i < num_output_fields; i++) { 936 CeedVector l_vec; 937 938 // Check if active output 939 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &l_vec)); 940 if (l_vec == CEED_VECTOR_ACTIVE) { 941 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size)); 942 num_active_out += size; 943 } 944 } 945 impl->num_active_out = num_active_out; 946 } 947 948 // Check sizes 949 CeedCheck(num_active_in > 0 && num_active_out > 0, ceed, CEED_ERROR_BACKEND, "Cannot assemble QFunction without active inputs and outputs"); 950 951 // Build objects if needed 952 if (build_objects) { 953 CeedSize l_size = (CeedSize)num_elem * Q * num_active_in * num_active_out; 954 CeedInt strides[3] = {1, num_elem * Q, Q}; /* *NOPAD* */ 955 956 // Create output restriction 957 CeedCallBackend(CeedElemRestrictionCreateStrided(ceed_parent, num_elem, Q, num_active_in * num_active_out, 958 (CeedSize)num_active_in * (CeedSize)num_active_out * (CeedSize)num_elem * (CeedSize)Q, strides, 959 rstr)); 960 // Create assembled vector 961 CeedCallBackend(CeedVectorCreate(ceed_parent, l_size, assembled)); 962 } 963 CeedCallBackend(CeedVectorSetValue(*assembled, 0.0)); 964 CeedCallBackend(CeedVectorGetArray(*assembled, CEED_MEM_DEVICE, &assembled_array)); 965 966 // Assemble QFunction 967 for (CeedInt in = 0; in < num_active_in; in++) { 968 // Set Inputs 969 CeedCallBackend(CeedVectorSetValue(active_inputs[in], 1.0)); 970 if (num_active_in > 1) { 971 CeedCallBackend(CeedVectorSetValue(active_inputs[(in + num_active_in - 1) % num_active_in], 0.0)); 972 } 973 // Set Outputs 974 for (CeedInt out = 0; out < num_output_fields; out++) { 975 CeedVector l_vec; 976 977 // Get output vector 978 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &l_vec)); 979 // Check if active output 980 if (l_vec == CEED_VECTOR_ACTIVE) { 981 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, CEED_USE_POINTER, assembled_array)); 982 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[out], &size)); 983 assembled_array += size * Q * num_elem; // Advance the pointer by the size of the output 984 } 985 } 986 // Apply QFunction 987 CeedCallBackend(CeedQFunctionApply(qf, Q * num_elem, impl->q_vecs_in, impl->q_vecs_out)); 988 } 989 990 // Un-set output q-vecs to prevent accidental overwrite of Assembled 991 for (CeedInt out = 0; out < num_output_fields; out++) { 992 CeedVector l_vec; 993 994 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &l_vec)); 995 if (l_vec == CEED_VECTOR_ACTIVE) { 996 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, NULL)); 997 } 998 } 999 1000 // Restore input arrays 1001 for (CeedInt i = 0; i < num_input_fields; i++) { 1002 CeedCallBackend(CeedOperatorInputRestore_Hip(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, true, impl)); 1003 } 1004 1005 // Restore output 1006 CeedCallBackend(CeedVectorRestoreArray(*assembled, &assembled_array)); 1007 return CEED_ERROR_SUCCESS; 1008 } 1009 1010 //------------------------------------------------------------------------------ 1011 // Assemble Linear QFunction 1012 //------------------------------------------------------------------------------ 1013 static int CeedOperatorLinearAssembleQFunction_Hip(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 1014 return CeedOperatorLinearAssembleQFunctionCore_Hip(op, true, assembled, rstr, request); 1015 } 1016 1017 //------------------------------------------------------------------------------ 1018 // Update Assembled Linear QFunction 1019 //------------------------------------------------------------------------------ 1020 static int CeedOperatorLinearAssembleQFunctionUpdate_Hip(CeedOperator op, CeedVector assembled, CeedElemRestriction rstr, CeedRequest *request) { 1021 return CeedOperatorLinearAssembleQFunctionCore_Hip(op, false, &assembled, &rstr, request); 1022 } 1023 1024 //------------------------------------------------------------------------------ 1025 // Assemble Diagonal Setup 1026 //------------------------------------------------------------------------------ 1027 static inline int CeedOperatorAssembleDiagonalSetup_Hip(CeedOperator op) { 1028 Ceed ceed; 1029 CeedInt num_input_fields, num_output_fields, num_eval_modes_in = 0, num_eval_modes_out = 0; 1030 CeedInt q_comp, num_nodes, num_qpts; 1031 CeedEvalMode *eval_modes_in = NULL, *eval_modes_out = NULL; 1032 CeedBasis basis_in = NULL, basis_out = NULL; 1033 CeedQFunctionField *qf_fields; 1034 CeedQFunction qf; 1035 CeedOperatorField *op_fields; 1036 CeedOperator_Hip *impl; 1037 1038 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1039 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1040 CeedCallBackend(CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields)); 1041 1042 // Determine active input basis 1043 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 1044 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 1045 for (CeedInt i = 0; i < num_input_fields; i++) { 1046 CeedVector vec; 1047 1048 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1049 if (vec == CEED_VECTOR_ACTIVE) { 1050 CeedBasis basis; 1051 CeedEvalMode eval_mode; 1052 1053 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 1054 CeedCheck(!basis_in || basis_in == basis, ceed, CEED_ERROR_BACKEND, 1055 "Backend does not implement operator diagonal assembly with multiple active bases"); 1056 basis_in = basis; 1057 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1058 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_in, eval_mode, &q_comp)); 1059 if (eval_mode != CEED_EVAL_WEIGHT) { 1060 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF assembly 1061 CeedCallBackend(CeedRealloc(num_eval_modes_in + q_comp, &eval_modes_in)); 1062 for (CeedInt d = 0; d < q_comp; d++) eval_modes_in[num_eval_modes_in + d] = eval_mode; 1063 num_eval_modes_in += q_comp; 1064 } 1065 } 1066 } 1067 1068 // Determine active output basis 1069 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 1070 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 1071 for (CeedInt i = 0; i < num_output_fields; i++) { 1072 CeedVector vec; 1073 1074 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1075 if (vec == CEED_VECTOR_ACTIVE) { 1076 CeedBasis basis; 1077 CeedEvalMode eval_mode; 1078 1079 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 1080 CeedCheck(!basis_out || basis_out == basis, ceed, CEED_ERROR_BACKEND, 1081 "Backend does not implement operator diagonal assembly with multiple active bases"); 1082 basis_out = basis; 1083 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1084 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_out, eval_mode, &q_comp)); 1085 if (eval_mode != CEED_EVAL_WEIGHT) { 1086 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF assembly 1087 CeedCallBackend(CeedRealloc(num_eval_modes_out + q_comp, &eval_modes_out)); 1088 for (CeedInt d = 0; d < q_comp; d++) eval_modes_out[num_eval_modes_out + d] = eval_mode; 1089 num_eval_modes_out += q_comp; 1090 } 1091 } 1092 } 1093 1094 // Operator data struct 1095 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1096 CeedCallBackend(CeedCalloc(1, &impl->diag)); 1097 CeedOperatorDiag_Hip *diag = impl->diag; 1098 1099 // Basis matrices 1100 CeedCallBackend(CeedBasisGetNumNodes(basis_in, &num_nodes)); 1101 if (basis_in == CEED_BASIS_NONE) num_qpts = num_nodes; 1102 else CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 1103 const CeedInt interp_bytes = num_nodes * num_qpts * sizeof(CeedScalar); 1104 const CeedInt eval_modes_bytes = sizeof(CeedEvalMode); 1105 bool has_eval_none = false; 1106 1107 // CEED_EVAL_NONE 1108 for (CeedInt i = 0; i < num_eval_modes_in; i++) has_eval_none = has_eval_none || (eval_modes_in[i] == CEED_EVAL_NONE); 1109 for (CeedInt i = 0; i < num_eval_modes_out; i++) has_eval_none = has_eval_none || (eval_modes_out[i] == CEED_EVAL_NONE); 1110 if (has_eval_none) { 1111 CeedScalar *identity = NULL; 1112 1113 CeedCallBackend(CeedCalloc(num_nodes * num_qpts, &identity)); 1114 for (CeedInt i = 0; i < (num_nodes < num_qpts ? num_nodes : num_qpts); i++) identity[i * num_nodes + i] = 1.0; 1115 CeedCallHip(ceed, hipMalloc((void **)&diag->d_identity, interp_bytes)); 1116 CeedCallHip(ceed, hipMemcpy(diag->d_identity, identity, interp_bytes, hipMemcpyHostToDevice)); 1117 CeedCallBackend(CeedFree(&identity)); 1118 } 1119 1120 // CEED_EVAL_INTERP, CEED_EVAL_GRAD, CEED_EVAL_DIV, and CEED_EVAL_CURL 1121 for (CeedInt in = 0; in < 2; in++) { 1122 CeedFESpace fespace; 1123 CeedBasis basis = in ? basis_in : basis_out; 1124 1125 CeedCallBackend(CeedBasisGetFESpace(basis, &fespace)); 1126 switch (fespace) { 1127 case CEED_FE_SPACE_H1: { 1128 CeedInt q_comp_interp, q_comp_grad; 1129 const CeedScalar *interp, *grad; 1130 CeedScalar *d_interp, *d_grad; 1131 1132 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 1133 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_grad)); 1134 1135 CeedCallBackend(CeedBasisGetInterp(basis, &interp)); 1136 CeedCallHip(ceed, hipMalloc((void **)&d_interp, interp_bytes * q_comp_interp)); 1137 CeedCallHip(ceed, hipMemcpy(d_interp, interp, interp_bytes * q_comp_interp, hipMemcpyHostToDevice)); 1138 CeedCallBackend(CeedBasisGetGrad(basis, &grad)); 1139 CeedCallHip(ceed, hipMalloc((void **)&d_grad, interp_bytes * q_comp_grad)); 1140 CeedCallHip(ceed, hipMemcpy(d_grad, grad, interp_bytes * q_comp_grad, hipMemcpyHostToDevice)); 1141 if (in) { 1142 diag->d_interp_in = d_interp; 1143 diag->d_grad_in = d_grad; 1144 } else { 1145 diag->d_interp_out = d_interp; 1146 diag->d_grad_out = d_grad; 1147 } 1148 } break; 1149 case CEED_FE_SPACE_HDIV: { 1150 CeedInt q_comp_interp, q_comp_div; 1151 const CeedScalar *interp, *div; 1152 CeedScalar *d_interp, *d_div; 1153 1154 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 1155 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_DIV, &q_comp_div)); 1156 1157 CeedCallBackend(CeedBasisGetInterp(basis, &interp)); 1158 CeedCallHip(ceed, hipMalloc((void **)&d_interp, interp_bytes * q_comp_interp)); 1159 CeedCallHip(ceed, hipMemcpy(d_interp, interp, interp_bytes * q_comp_interp, hipMemcpyHostToDevice)); 1160 CeedCallBackend(CeedBasisGetDiv(basis, &div)); 1161 CeedCallHip(ceed, hipMalloc((void **)&d_div, interp_bytes * q_comp_div)); 1162 CeedCallHip(ceed, hipMemcpy(d_div, div, interp_bytes * q_comp_div, hipMemcpyHostToDevice)); 1163 if (in) { 1164 diag->d_interp_in = d_interp; 1165 diag->d_div_in = d_div; 1166 } else { 1167 diag->d_interp_out = d_interp; 1168 diag->d_div_out = d_div; 1169 } 1170 } break; 1171 case CEED_FE_SPACE_HCURL: { 1172 CeedInt q_comp_interp, q_comp_curl; 1173 const CeedScalar *interp, *curl; 1174 CeedScalar *d_interp, *d_curl; 1175 1176 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 1177 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_CURL, &q_comp_curl)); 1178 1179 CeedCallBackend(CeedBasisGetInterp(basis, &interp)); 1180 CeedCallHip(ceed, hipMalloc((void **)&d_interp, interp_bytes * q_comp_interp)); 1181 CeedCallHip(ceed, hipMemcpy(d_interp, interp, interp_bytes * q_comp_interp, hipMemcpyHostToDevice)); 1182 CeedCallBackend(CeedBasisGetCurl(basis, &curl)); 1183 CeedCallHip(ceed, hipMalloc((void **)&d_curl, interp_bytes * q_comp_curl)); 1184 CeedCallHip(ceed, hipMemcpy(d_curl, curl, interp_bytes * q_comp_curl, hipMemcpyHostToDevice)); 1185 if (in) { 1186 diag->d_interp_in = d_interp; 1187 diag->d_curl_in = d_curl; 1188 } else { 1189 diag->d_interp_out = d_interp; 1190 diag->d_curl_out = d_curl; 1191 } 1192 } break; 1193 } 1194 } 1195 1196 // Arrays of eval_modes 1197 CeedCallHip(ceed, hipMalloc((void **)&diag->d_eval_modes_in, num_eval_modes_in * eval_modes_bytes)); 1198 CeedCallHip(ceed, hipMemcpy(diag->d_eval_modes_in, eval_modes_in, num_eval_modes_in * eval_modes_bytes, hipMemcpyHostToDevice)); 1199 CeedCallHip(ceed, hipMalloc((void **)&diag->d_eval_modes_out, num_eval_modes_out * eval_modes_bytes)); 1200 CeedCallHip(ceed, hipMemcpy(diag->d_eval_modes_out, eval_modes_out, num_eval_modes_out * eval_modes_bytes, hipMemcpyHostToDevice)); 1201 CeedCallBackend(CeedFree(&eval_modes_in)); 1202 CeedCallBackend(CeedFree(&eval_modes_out)); 1203 return CEED_ERROR_SUCCESS; 1204 } 1205 1206 //------------------------------------------------------------------------------ 1207 // Assemble Diagonal Setup (Compilation) 1208 //------------------------------------------------------------------------------ 1209 static inline int CeedOperatorAssembleDiagonalSetupCompile_Hip(CeedOperator op, CeedInt use_ceedsize_idx, const bool is_point_block) { 1210 Ceed ceed; 1211 char *diagonal_kernel_source; 1212 const char *diagonal_kernel_path; 1213 CeedInt num_input_fields, num_output_fields, num_eval_modes_in = 0, num_eval_modes_out = 0; 1214 CeedInt num_comp, q_comp, num_nodes, num_qpts; 1215 CeedBasis basis_in = NULL, basis_out = NULL; 1216 CeedQFunctionField *qf_fields; 1217 CeedQFunction qf; 1218 CeedOperatorField *op_fields; 1219 CeedOperator_Hip *impl; 1220 1221 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1222 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1223 CeedCallBackend(CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields)); 1224 1225 // Determine active input basis 1226 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 1227 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 1228 for (CeedInt i = 0; i < num_input_fields; i++) { 1229 CeedVector vec; 1230 1231 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1232 if (vec == CEED_VECTOR_ACTIVE) { 1233 CeedEvalMode eval_mode; 1234 1235 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_in)); 1236 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1237 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_in, eval_mode, &q_comp)); 1238 if (eval_mode != CEED_EVAL_WEIGHT) { 1239 num_eval_modes_in += q_comp; 1240 } 1241 } 1242 } 1243 1244 // Determine active output basis 1245 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 1246 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 1247 for (CeedInt i = 0; i < num_output_fields; i++) { 1248 CeedVector vec; 1249 1250 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1251 if (vec == CEED_VECTOR_ACTIVE) { 1252 CeedEvalMode eval_mode; 1253 1254 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_out)); 1255 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1256 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_out, eval_mode, &q_comp)); 1257 if (eval_mode != CEED_EVAL_WEIGHT) { 1258 num_eval_modes_out += q_comp; 1259 } 1260 } 1261 } 1262 1263 // Operator data struct 1264 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1265 CeedOperatorDiag_Hip *diag = impl->diag; 1266 1267 // Assemble kernel 1268 hipModule_t *module = is_point_block ? &diag->module_point_block : &diag->module; 1269 CeedInt elems_per_block = 1; 1270 CeedCallBackend(CeedBasisGetNumNodes(basis_in, &num_nodes)); 1271 CeedCallBackend(CeedBasisGetNumComponents(basis_in, &num_comp)); 1272 if (basis_in == CEED_BASIS_NONE) num_qpts = num_nodes; 1273 else CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 1274 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-ref-operator-assemble-diagonal.h", &diagonal_kernel_path)); 1275 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Kernel Source -----\n"); 1276 CeedCallBackend(CeedLoadSourceToBuffer(ceed, diagonal_kernel_path, &diagonal_kernel_source)); 1277 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Source Complete! -----\n"); 1278 CeedCallHip(ceed, CeedCompile_Hip(ceed, diagonal_kernel_source, module, 8, "NUM_EVAL_MODES_IN", num_eval_modes_in, "NUM_EVAL_MODES_OUT", 1279 num_eval_modes_out, "NUM_COMP", num_comp, "NUM_NODES", num_nodes, "NUM_QPTS", num_qpts, "USE_CEEDSIZE", 1280 use_ceedsize_idx, "USE_POINT_BLOCK", is_point_block ? 1 : 0, "BLOCK_SIZE", num_nodes * elems_per_block)); 1281 CeedCallHip(ceed, CeedGetKernel_Hip(ceed, *module, "LinearDiagonal", is_point_block ? &diag->LinearPointBlock : &diag->LinearDiagonal)); 1282 CeedCallBackend(CeedFree(&diagonal_kernel_path)); 1283 CeedCallBackend(CeedFree(&diagonal_kernel_source)); 1284 return CEED_ERROR_SUCCESS; 1285 } 1286 1287 //------------------------------------------------------------------------------ 1288 // Assemble Diagonal Core 1289 //------------------------------------------------------------------------------ 1290 static inline int CeedOperatorAssembleDiagonalCore_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request, const bool is_point_block) { 1291 Ceed ceed; 1292 CeedInt num_elem, num_nodes; 1293 CeedScalar *elem_diag_array; 1294 const CeedScalar *assembled_qf_array; 1295 CeedVector assembled_qf = NULL, elem_diag; 1296 CeedElemRestriction assembled_rstr = NULL, rstr_in, rstr_out, diag_rstr; 1297 CeedOperator_Hip *impl; 1298 1299 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1300 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1301 1302 // Assemble QFunction 1303 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &assembled_rstr, request)); 1304 CeedCallBackend(CeedElemRestrictionDestroy(&assembled_rstr)); 1305 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &assembled_qf_array)); 1306 1307 // Setup 1308 if (!impl->diag) CeedCallBackend(CeedOperatorAssembleDiagonalSetup_Hip(op)); 1309 CeedOperatorDiag_Hip *diag = impl->diag; 1310 1311 assert(diag != NULL); 1312 1313 // Assemble kernel if needed 1314 if ((!is_point_block && !diag->LinearDiagonal) || (is_point_block && !diag->LinearPointBlock)) { 1315 CeedSize assembled_length, assembled_qf_length; 1316 CeedInt use_ceedsize_idx = 0; 1317 CeedCallBackend(CeedVectorGetLength(assembled, &assembled_length)); 1318 CeedCallBackend(CeedVectorGetLength(assembled_qf, &assembled_qf_length)); 1319 if ((assembled_length > INT_MAX) || (assembled_qf_length > INT_MAX)) use_ceedsize_idx = 1; 1320 1321 CeedCallBackend(CeedOperatorAssembleDiagonalSetupCompile_Hip(op, use_ceedsize_idx, is_point_block)); 1322 } 1323 1324 // Restriction and diagonal vector 1325 CeedCallBackend(CeedOperatorGetActiveElemRestrictions(op, &rstr_in, &rstr_out)); 1326 CeedCheck(rstr_in == rstr_out, ceed, CEED_ERROR_BACKEND, 1327 "Cannot assemble operator diagonal with different input and output active element restrictions"); 1328 if (!is_point_block && !diag->diag_rstr) { 1329 CeedCallBackend(CeedElemRestrictionCreateUnsignedCopy(rstr_out, &diag->diag_rstr)); 1330 CeedCallBackend(CeedElemRestrictionCreateVector(diag->diag_rstr, NULL, &diag->elem_diag)); 1331 } else if (is_point_block && !diag->point_block_diag_rstr) { 1332 CeedCallBackend(CeedOperatorCreateActivePointBlockRestriction(rstr_out, &diag->point_block_diag_rstr)); 1333 CeedCallBackend(CeedElemRestrictionCreateVector(diag->point_block_diag_rstr, NULL, &diag->point_block_elem_diag)); 1334 } 1335 diag_rstr = is_point_block ? diag->point_block_diag_rstr : diag->diag_rstr; 1336 elem_diag = is_point_block ? diag->point_block_elem_diag : diag->elem_diag; 1337 CeedCallBackend(CeedVectorSetValue(elem_diag, 0.0)); 1338 1339 // Only assemble diagonal if the basis has nodes, otherwise inputs are null pointers 1340 CeedCallBackend(CeedElemRestrictionGetElementSize(diag_rstr, &num_nodes)); 1341 if (num_nodes > 0) { 1342 // Assemble element operator diagonals 1343 CeedCallBackend(CeedElemRestrictionGetNumElements(diag_rstr, &num_elem)); 1344 CeedCallBackend(CeedVectorGetArray(elem_diag, CEED_MEM_DEVICE, &elem_diag_array)); 1345 1346 // Compute the diagonal of B^T D B 1347 CeedInt elems_per_block = 1; 1348 CeedInt grid = CeedDivUpInt(num_elem, elems_per_block); 1349 void *args[] = {(void *)&num_elem, &diag->d_identity, &diag->d_interp_in, &diag->d_grad_in, &diag->d_div_in, 1350 &diag->d_curl_in, &diag->d_interp_out, &diag->d_grad_out, &diag->d_div_out, &diag->d_curl_out, 1351 &diag->d_eval_modes_in, &diag->d_eval_modes_out, &assembled_qf_array, &elem_diag_array}; 1352 1353 if (is_point_block) { 1354 CeedCallBackend(CeedRunKernelDim_Hip(ceed, diag->LinearPointBlock, grid, num_nodes, 1, elems_per_block, args)); 1355 } else { 1356 CeedCallBackend(CeedRunKernelDim_Hip(ceed, diag->LinearDiagonal, grid, num_nodes, 1, elems_per_block, args)); 1357 } 1358 1359 // Restore arrays 1360 CeedCallBackend(CeedVectorRestoreArray(elem_diag, &elem_diag_array)); 1361 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 1362 } 1363 1364 // Assemble local operator diagonal 1365 CeedCallBackend(CeedElemRestrictionApply(diag_rstr, CEED_TRANSPOSE, elem_diag, assembled, request)); 1366 1367 // Cleanup 1368 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 1369 return CEED_ERROR_SUCCESS; 1370 } 1371 1372 //------------------------------------------------------------------------------ 1373 // Assemble Linear Diagonal 1374 //------------------------------------------------------------------------------ 1375 static int CeedOperatorLinearAssembleAddDiagonal_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request) { 1376 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Hip(op, assembled, request, false)); 1377 return CEED_ERROR_SUCCESS; 1378 } 1379 1380 //------------------------------------------------------------------------------ 1381 // Assemble Linear Point Block Diagonal 1382 //------------------------------------------------------------------------------ 1383 static int CeedOperatorLinearAssembleAddPointBlockDiagonal_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request) { 1384 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Hip(op, assembled, request, true)); 1385 return CEED_ERROR_SUCCESS; 1386 } 1387 1388 //------------------------------------------------------------------------------ 1389 // Single Operator Assembly Setup 1390 //------------------------------------------------------------------------------ 1391 static int CeedSingleOperatorAssembleSetup_Hip(CeedOperator op, CeedInt use_ceedsize_idx) { 1392 Ceed ceed; 1393 Ceed_Hip *Hip_data; 1394 char *assembly_kernel_source; 1395 const char *assembly_kernel_path; 1396 CeedInt num_input_fields, num_output_fields, num_eval_modes_in = 0, num_eval_modes_out = 0; 1397 CeedInt elem_size_in, num_qpts_in = 0, num_comp_in, elem_size_out, num_qpts_out, num_comp_out, q_comp; 1398 CeedEvalMode *eval_modes_in = NULL, *eval_modes_out = NULL; 1399 CeedElemRestriction rstr_in = NULL, rstr_out = NULL; 1400 CeedBasis basis_in = NULL, basis_out = NULL; 1401 CeedQFunctionField *qf_fields; 1402 CeedQFunction qf; 1403 CeedOperatorField *input_fields, *output_fields; 1404 CeedOperator_Hip *impl; 1405 1406 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1407 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1408 1409 // Get intput and output fields 1410 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 1411 1412 // Determine active input basis eval mode 1413 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1414 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 1415 for (CeedInt i = 0; i < num_input_fields; i++) { 1416 CeedVector vec; 1417 1418 CeedCallBackend(CeedOperatorFieldGetVector(input_fields[i], &vec)); 1419 if (vec == CEED_VECTOR_ACTIVE) { 1420 CeedBasis basis; 1421 CeedEvalMode eval_mode; 1422 1423 CeedCallBackend(CeedOperatorFieldGetBasis(input_fields[i], &basis)); 1424 CeedCheck(!basis_in || basis_in == basis, ceed, CEED_ERROR_BACKEND, "Backend does not implement operator assembly with multiple active bases"); 1425 basis_in = basis; 1426 CeedCallBackend(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr_in)); 1427 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_in, &elem_size_in)); 1428 if (basis_in == CEED_BASIS_NONE) num_qpts_in = elem_size_in; 1429 else CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts_in)); 1430 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1431 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_in, eval_mode, &q_comp)); 1432 if (eval_mode != CEED_EVAL_WEIGHT) { 1433 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF Assembly 1434 CeedCallBackend(CeedRealloc(num_eval_modes_in + q_comp, &eval_modes_in)); 1435 for (CeedInt d = 0; d < q_comp; d++) { 1436 eval_modes_in[num_eval_modes_in + d] = eval_mode; 1437 } 1438 num_eval_modes_in += q_comp; 1439 } 1440 } 1441 } 1442 1443 // Determine active output basis; basis_out and rstr_out only used if same as input, TODO 1444 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 1445 for (CeedInt i = 0; i < num_output_fields; i++) { 1446 CeedVector vec; 1447 1448 CeedCallBackend(CeedOperatorFieldGetVector(output_fields[i], &vec)); 1449 if (vec == CEED_VECTOR_ACTIVE) { 1450 CeedBasis basis; 1451 CeedEvalMode eval_mode; 1452 1453 CeedCallBackend(CeedOperatorFieldGetBasis(output_fields[i], &basis)); 1454 CeedCheck(!basis_out || basis_out == basis, ceed, CEED_ERROR_BACKEND, 1455 "Backend does not implement operator assembly with multiple active bases"); 1456 basis_out = basis; 1457 CeedCallBackend(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr_out)); 1458 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_out, &elem_size_out)); 1459 if (basis_out == CEED_BASIS_NONE) num_qpts_out = elem_size_out; 1460 else CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_out, &num_qpts_out)); 1461 CeedCheck(num_qpts_in == num_qpts_out, ceed, CEED_ERROR_UNSUPPORTED, 1462 "Active input and output bases must have the same number of quadrature points"); 1463 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1464 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_out, eval_mode, &q_comp)); 1465 if (eval_mode != CEED_EVAL_WEIGHT) { 1466 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF Assembly 1467 CeedCallBackend(CeedRealloc(num_eval_modes_out + q_comp, &eval_modes_out)); 1468 for (CeedInt d = 0; d < q_comp; d++) { 1469 eval_modes_out[num_eval_modes_out + d] = eval_mode; 1470 } 1471 num_eval_modes_out += q_comp; 1472 } 1473 } 1474 } 1475 CeedCheck(num_eval_modes_in > 0 && num_eval_modes_out > 0, ceed, CEED_ERROR_UNSUPPORTED, "Cannot assemble operator without inputs/outputs"); 1476 1477 CeedCallBackend(CeedCalloc(1, &impl->asmb)); 1478 CeedOperatorAssemble_Hip *asmb = impl->asmb; 1479 asmb->elems_per_block = 1; 1480 asmb->block_size_x = elem_size_in; 1481 asmb->block_size_y = elem_size_out; 1482 1483 CeedCallBackend(CeedGetData(ceed, &Hip_data)); 1484 bool fallback = asmb->block_size_x * asmb->block_size_y * asmb->elems_per_block > Hip_data->device_prop.maxThreadsPerBlock; 1485 1486 if (fallback) { 1487 // Use fallback kernel with 1D threadblock 1488 asmb->block_size_y = 1; 1489 } 1490 1491 // Compile kernels 1492 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr_in, &num_comp_in)); 1493 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr_out, &num_comp_out)); 1494 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-ref-operator-assemble.h", &assembly_kernel_path)); 1495 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Kernel Source -----\n"); 1496 CeedCallBackend(CeedLoadSourceToBuffer(ceed, assembly_kernel_path, &assembly_kernel_source)); 1497 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Source Complete! -----\n"); 1498 CeedCallBackend(CeedCompile_Hip(ceed, assembly_kernel_source, &asmb->module, 10, "NUM_EVAL_MODES_IN", num_eval_modes_in, "NUM_EVAL_MODES_OUT", 1499 num_eval_modes_out, "NUM_COMP_IN", num_comp_in, "NUM_COMP_OUT", num_comp_out, "NUM_NODES_IN", elem_size_in, 1500 "NUM_NODES_OUT", elem_size_out, "NUM_QPTS", num_qpts_in, "BLOCK_SIZE", 1501 asmb->block_size_x * asmb->block_size_y * asmb->elems_per_block, "BLOCK_SIZE_Y", asmb->block_size_y, "USE_CEEDSIZE", 1502 use_ceedsize_idx)); 1503 CeedCallBackend(CeedGetKernel_Hip(ceed, asmb->module, "LinearAssemble", &asmb->LinearAssemble)); 1504 CeedCallBackend(CeedFree(&assembly_kernel_path)); 1505 CeedCallBackend(CeedFree(&assembly_kernel_source)); 1506 1507 // Load into B_in, in order that they will be used in eval_modes_in 1508 { 1509 const CeedInt in_bytes = elem_size_in * num_qpts_in * num_eval_modes_in * sizeof(CeedScalar); 1510 CeedInt d_in = 0; 1511 CeedEvalMode eval_modes_in_prev = CEED_EVAL_NONE; 1512 bool has_eval_none = false; 1513 CeedScalar *identity = NULL; 1514 1515 for (CeedInt i = 0; i < num_eval_modes_in; i++) { 1516 has_eval_none = has_eval_none || (eval_modes_in[i] == CEED_EVAL_NONE); 1517 } 1518 if (has_eval_none) { 1519 CeedCallBackend(CeedCalloc(elem_size_in * num_qpts_in, &identity)); 1520 for (CeedInt i = 0; i < (elem_size_in < num_qpts_in ? elem_size_in : num_qpts_in); i++) identity[i * elem_size_in + i] = 1.0; 1521 } 1522 1523 CeedCallHip(ceed, hipMalloc((void **)&asmb->d_B_in, in_bytes)); 1524 for (CeedInt i = 0; i < num_eval_modes_in; i++) { 1525 const CeedScalar *h_B_in; 1526 1527 CeedCallBackend(CeedOperatorGetBasisPointer(basis_in, eval_modes_in[i], identity, &h_B_in)); 1528 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_in, eval_modes_in[i], &q_comp)); 1529 if (q_comp > 1) { 1530 if (i == 0 || eval_modes_in[i] != eval_modes_in_prev) d_in = 0; 1531 else h_B_in = &h_B_in[(++d_in) * elem_size_in * num_qpts_in]; 1532 } 1533 eval_modes_in_prev = eval_modes_in[i]; 1534 1535 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_in[i * elem_size_in * num_qpts_in], h_B_in, elem_size_in * num_qpts_in * sizeof(CeedScalar), 1536 hipMemcpyHostToDevice)); 1537 } 1538 1539 if (identity) { 1540 CeedCallBackend(CeedFree(&identity)); 1541 } 1542 } 1543 1544 // Load into B_out, in order that they will be used in eval_modes_out 1545 { 1546 const CeedInt out_bytes = elem_size_out * num_qpts_out * num_eval_modes_out * sizeof(CeedScalar); 1547 CeedInt d_out = 0; 1548 CeedEvalMode eval_modes_out_prev = CEED_EVAL_NONE; 1549 bool has_eval_none = false; 1550 CeedScalar *identity = NULL; 1551 1552 for (CeedInt i = 0; i < num_eval_modes_out; i++) { 1553 has_eval_none = has_eval_none || (eval_modes_out[i] == CEED_EVAL_NONE); 1554 } 1555 if (has_eval_none) { 1556 CeedCallBackend(CeedCalloc(elem_size_out * num_qpts_out, &identity)); 1557 for (CeedInt i = 0; i < (elem_size_out < num_qpts_out ? elem_size_out : num_qpts_out); i++) identity[i * elem_size_out + i] = 1.0; 1558 } 1559 1560 CeedCallHip(ceed, hipMalloc((void **)&asmb->d_B_out, out_bytes)); 1561 for (CeedInt i = 0; i < num_eval_modes_out; i++) { 1562 const CeedScalar *h_B_out; 1563 1564 CeedCallBackend(CeedOperatorGetBasisPointer(basis_out, eval_modes_out[i], identity, &h_B_out)); 1565 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_out, eval_modes_out[i], &q_comp)); 1566 if (q_comp > 1) { 1567 if (i == 0 || eval_modes_out[i] != eval_modes_out_prev) d_out = 0; 1568 else h_B_out = &h_B_out[(++d_out) * elem_size_out * num_qpts_out]; 1569 } 1570 eval_modes_out_prev = eval_modes_out[i]; 1571 1572 CeedCallHip(ceed, hipMemcpy(&asmb->d_B_out[i * elem_size_out * num_qpts_out], h_B_out, elem_size_out * num_qpts_out * sizeof(CeedScalar), 1573 hipMemcpyHostToDevice)); 1574 } 1575 1576 if (identity) { 1577 CeedCallBackend(CeedFree(&identity)); 1578 } 1579 } 1580 return CEED_ERROR_SUCCESS; 1581 } 1582 1583 //------------------------------------------------------------------------------ 1584 // Assemble matrix data for COO matrix of assembled operator. 1585 // The sparsity pattern is set by CeedOperatorLinearAssembleSymbolic. 1586 // 1587 // Note that this (and other assembly routines) currently assume only one active input restriction/basis per operator 1588 // (could have multiple basis eval modes). 1589 // TODO: allow multiple active input restrictions/basis objects 1590 //------------------------------------------------------------------------------ 1591 static int CeedSingleOperatorAssemble_Hip(CeedOperator op, CeedInt offset, CeedVector values) { 1592 Ceed ceed; 1593 CeedSize values_length = 0, assembled_qf_length = 0; 1594 CeedInt use_ceedsize_idx = 0, num_elem_in, num_elem_out, elem_size_in, elem_size_out; 1595 CeedScalar *values_array; 1596 const CeedScalar *assembled_qf_array; 1597 CeedVector assembled_qf = NULL; 1598 CeedElemRestriction assembled_rstr = NULL, rstr_in, rstr_out; 1599 CeedRestrictionType rstr_type_in, rstr_type_out; 1600 const bool *orients_in = NULL, *orients_out = NULL; 1601 const CeedInt8 *curl_orients_in = NULL, *curl_orients_out = NULL; 1602 CeedOperator_Hip *impl; 1603 1604 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1605 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1606 1607 // Assemble QFunction 1608 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &assembled_rstr, CEED_REQUEST_IMMEDIATE)); 1609 CeedCallBackend(CeedElemRestrictionDestroy(&assembled_rstr)); 1610 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &assembled_qf_array)); 1611 1612 CeedCallBackend(CeedVectorGetLength(values, &values_length)); 1613 CeedCallBackend(CeedVectorGetLength(assembled_qf, &assembled_qf_length)); 1614 if ((values_length > INT_MAX) || (assembled_qf_length > INT_MAX)) use_ceedsize_idx = 1; 1615 1616 // Setup 1617 if (!impl->asmb) CeedCallBackend(CeedSingleOperatorAssembleSetup_Hip(op, use_ceedsize_idx)); 1618 CeedOperatorAssemble_Hip *asmb = impl->asmb; 1619 1620 assert(asmb != NULL); 1621 1622 // Assemble element operator 1623 CeedCallBackend(CeedVectorGetArray(values, CEED_MEM_DEVICE, &values_array)); 1624 values_array += offset; 1625 1626 CeedCallBackend(CeedOperatorGetActiveElemRestrictions(op, &rstr_in, &rstr_out)); 1627 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr_in, &num_elem_in)); 1628 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_in, &elem_size_in)); 1629 1630 CeedCallBackend(CeedElemRestrictionGetType(rstr_in, &rstr_type_in)); 1631 if (rstr_type_in == CEED_RESTRICTION_ORIENTED) { 1632 CeedCallBackend(CeedElemRestrictionGetOrientations(rstr_in, CEED_MEM_DEVICE, &orients_in)); 1633 } else if (rstr_type_in == CEED_RESTRICTION_CURL_ORIENTED) { 1634 CeedCallBackend(CeedElemRestrictionGetCurlOrientations(rstr_in, CEED_MEM_DEVICE, &curl_orients_in)); 1635 } 1636 1637 if (rstr_in != rstr_out) { 1638 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr_out, &num_elem_out)); 1639 CeedCheck(num_elem_in == num_elem_out, ceed, CEED_ERROR_UNSUPPORTED, 1640 "Active input and output operator restrictions must have the same number of elements"); 1641 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_out, &elem_size_out)); 1642 1643 CeedCallBackend(CeedElemRestrictionGetType(rstr_out, &rstr_type_out)); 1644 if (rstr_type_out == CEED_RESTRICTION_ORIENTED) { 1645 CeedCallBackend(CeedElemRestrictionGetOrientations(rstr_out, CEED_MEM_DEVICE, &orients_out)); 1646 } else if (rstr_type_out == CEED_RESTRICTION_CURL_ORIENTED) { 1647 CeedCallBackend(CeedElemRestrictionGetCurlOrientations(rstr_out, CEED_MEM_DEVICE, &curl_orients_out)); 1648 } 1649 } else { 1650 elem_size_out = elem_size_in; 1651 orients_out = orients_in; 1652 curl_orients_out = curl_orients_in; 1653 } 1654 1655 // Compute B^T D B 1656 CeedInt shared_mem = 1657 ((curl_orients_in || curl_orients_out ? elem_size_in * elem_size_out : 0) + (curl_orients_in ? elem_size_in * asmb->block_size_y : 0)) * 1658 sizeof(CeedScalar); 1659 CeedInt grid = CeedDivUpInt(num_elem_in, asmb->elems_per_block); 1660 void *args[] = {(void *)&num_elem_in, &asmb->d_B_in, &asmb->d_B_out, &orients_in, &curl_orients_in, 1661 &orients_out, &curl_orients_out, &assembled_qf_array, &values_array}; 1662 1663 CeedCallBackend( 1664 CeedRunKernelDimShared_Hip(ceed, asmb->LinearAssemble, grid, asmb->block_size_x, asmb->block_size_y, asmb->elems_per_block, shared_mem, args)); 1665 1666 // Restore arrays 1667 CeedCallBackend(CeedVectorRestoreArray(values, &values_array)); 1668 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 1669 1670 // Cleanup 1671 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 1672 if (rstr_type_in == CEED_RESTRICTION_ORIENTED) { 1673 CeedCallBackend(CeedElemRestrictionRestoreOrientations(rstr_in, &orients_in)); 1674 } else if (rstr_type_in == CEED_RESTRICTION_CURL_ORIENTED) { 1675 CeedCallBackend(CeedElemRestrictionRestoreCurlOrientations(rstr_in, &curl_orients_in)); 1676 } 1677 if (rstr_in != rstr_out) { 1678 if (rstr_type_out == CEED_RESTRICTION_ORIENTED) { 1679 CeedCallBackend(CeedElemRestrictionRestoreOrientations(rstr_out, &orients_out)); 1680 } else if (rstr_type_out == CEED_RESTRICTION_CURL_ORIENTED) { 1681 CeedCallBackend(CeedElemRestrictionRestoreCurlOrientations(rstr_out, &curl_orients_out)); 1682 } 1683 } 1684 return CEED_ERROR_SUCCESS; 1685 } 1686 1687 //------------------------------------------------------------------------------ 1688 // Assemble Linear QFunction AtPoints 1689 //------------------------------------------------------------------------------ 1690 static int CeedOperatorLinearAssembleQFunctionAtPoints_Hip(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 1691 return CeedError(CeedOperatorReturnCeed(op), CEED_ERROR_BACKEND, "Backend does not implement CeedOperatorLinearAssembleQFunction"); 1692 } 1693 1694 //------------------------------------------------------------------------------ 1695 // Assemble Linear Diagonal AtPoints 1696 //------------------------------------------------------------------------------ 1697 static int CeedOperatorLinearAssembleAddDiagonalAtPoints_Hip(CeedOperator op, CeedVector assembled, CeedRequest *request) { 1698 CeedInt max_num_points, *num_points, num_elem, num_input_fields, num_output_fields; 1699 Ceed ceed; 1700 CeedVector active_e_vec_in, active_e_vec_out; 1701 CeedQFunctionField *qf_input_fields, *qf_output_fields; 1702 CeedQFunction qf; 1703 CeedOperatorField *op_input_fields, *op_output_fields; 1704 CeedOperator_Hip *impl; 1705 1706 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1707 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1708 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1709 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 1710 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 1711 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 1712 1713 // Setup 1714 CeedCallBackend(CeedOperatorSetupAtPoints_Hip(op)); 1715 num_points = impl->num_points; 1716 max_num_points = impl->max_num_points; 1717 1718 // Work vector 1719 CeedCallBackend(CeedGetWorkVector(ceed, impl->max_active_e_vec_len, &active_e_vec_in)); 1720 CeedCallBackend(CeedGetWorkVector(ceed, impl->max_active_e_vec_len, &active_e_vec_out)); 1721 { 1722 CeedSize length_in, length_out; 1723 1724 CeedCallBackend(CeedVectorGetLength(active_e_vec_in, &length_in)); 1725 CeedCallBackend(CeedVectorGetLength(active_e_vec_out, &length_out)); 1726 // Need input e_vec to be longer 1727 if (length_in < length_out) { 1728 CeedVector temp = active_e_vec_in; 1729 1730 active_e_vec_in = active_e_vec_out; 1731 active_e_vec_out = temp; 1732 } 1733 } 1734 1735 // Get point coordinates 1736 if (!impl->point_coords_elem) { 1737 CeedVector point_coords = NULL; 1738 CeedElemRestriction rstr_points = NULL; 1739 1740 CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, &point_coords)); 1741 CeedCallBackend(CeedElemRestrictionCreateVector(rstr_points, NULL, &impl->point_coords_elem)); 1742 CeedCallBackend(CeedElemRestrictionApply(rstr_points, CEED_NOTRANSPOSE, point_coords, impl->point_coords_elem, request)); 1743 } 1744 1745 // Process inputs 1746 for (CeedInt i = 0; i < num_input_fields; i++) { 1747 CeedCallBackend(CeedOperatorInputRestrict_Hip(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, true, impl, request)); 1748 CeedCallBackend(CeedOperatorInputBasisAtPoints_Hip(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, num_elem, num_points, true, impl)); 1749 } 1750 1751 // Clear active input Qvecs 1752 for (CeedInt i = 0; i < num_input_fields; i++) { 1753 CeedVector vec; 1754 1755 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 1756 if (vec != CEED_VECTOR_ACTIVE) continue; 1757 CeedCallBackend(CeedVectorSetValue(impl->q_vecs_in[i], 0.0)); 1758 } 1759 1760 // Output pointers, as necessary 1761 for (CeedInt i = 0; i < num_output_fields; i++) { 1762 CeedEvalMode eval_mode; 1763 1764 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1765 if (eval_mode == CEED_EVAL_NONE) { 1766 CeedScalar *e_vec_array; 1767 1768 CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 1769 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_vec_array)); 1770 } 1771 } 1772 1773 // Loop over active fields 1774 for (CeedInt i = 0; i < num_input_fields; i++) { 1775 bool is_active_at_points = true; 1776 CeedInt elem_size = 1, num_comp_active = 1, e_vec_size = 0; 1777 CeedRestrictionType rstr_type; 1778 CeedVector l_vec; 1779 CeedElemRestriction elem_rstr; 1780 1781 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &l_vec)); 1782 // -- Skip non-active input 1783 if (l_vec != CEED_VECTOR_ACTIVE) continue; 1784 1785 // -- Get active restriction type 1786 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_rstr)); 1787 CeedCallBackend(CeedElemRestrictionGetType(elem_rstr, &rstr_type)); 1788 is_active_at_points = rstr_type == CEED_RESTRICTION_POINTS; 1789 if (!is_active_at_points) CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 1790 else elem_size = max_num_points; 1791 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp_active)); 1792 1793 e_vec_size = elem_size * num_comp_active; 1794 for (CeedInt s = 0; s < e_vec_size; s++) { 1795 bool is_active_input = false; 1796 CeedEvalMode eval_mode; 1797 CeedVector l_vec, q_vec = impl->q_vecs_in[i]; 1798 CeedBasis basis; 1799 1800 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &l_vec)); 1801 // Skip non-active input 1802 is_active_input = l_vec == CEED_VECTOR_ACTIVE; 1803 if (!is_active_input) continue; 1804 1805 // Update unit vector 1806 if (s == 0) CeedCallBackend(CeedVectorSetValue(active_e_vec_in, 0.0)); 1807 else CeedCallBackend(CeedVectorSetValueStrided(active_e_vec_in, s - 1, e_vec_size, 0.0)); 1808 CeedCallBackend(CeedVectorSetValueStrided(active_e_vec_in, s, e_vec_size, 1.0)); 1809 1810 // Basis action 1811 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1812 switch (eval_mode) { 1813 case CEED_EVAL_NONE: { 1814 const CeedScalar *e_vec_array; 1815 1816 CeedCallBackend(CeedVectorGetArrayRead(active_e_vec_in, CEED_MEM_DEVICE, &e_vec_array)); 1817 CeedCallBackend(CeedVectorSetArray(q_vec, CEED_MEM_DEVICE, CEED_USE_POINTER, (CeedScalar *)e_vec_array)); 1818 break; 1819 } 1820 case CEED_EVAL_INTERP: 1821 case CEED_EVAL_GRAD: 1822 case CEED_EVAL_DIV: 1823 case CEED_EVAL_CURL: 1824 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 1825 CeedCallBackend( 1826 CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_NOTRANSPOSE, eval_mode, impl->point_coords_elem, active_e_vec_in, q_vec)); 1827 break; 1828 case CEED_EVAL_WEIGHT: 1829 break; // No action 1830 } 1831 1832 // Q function 1833 CeedCallBackend(CeedQFunctionApply(qf, num_elem * max_num_points, impl->q_vecs_in, impl->q_vecs_out)); 1834 1835 // Output basis apply if needed 1836 for (CeedInt j = 0; j < num_output_fields; j++) { 1837 bool is_active_output = false; 1838 CeedInt elem_size = 0; 1839 CeedRestrictionType rstr_type; 1840 CeedEvalMode eval_mode; 1841 CeedVector l_vec, e_vec = impl->e_vecs_out[j], q_vec = impl->q_vecs_out[j]; 1842 CeedElemRestriction elem_rstr; 1843 CeedBasis basis; 1844 1845 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[j], &l_vec)); 1846 // ---- Skip non-active output 1847 is_active_output = l_vec == CEED_VECTOR_ACTIVE; 1848 if (!is_active_output) continue; 1849 if (!e_vec) e_vec = active_e_vec_out; 1850 1851 // ---- Check if elem size matches 1852 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[j], &elem_rstr)); 1853 CeedCallBackend(CeedElemRestrictionGetType(elem_rstr, &rstr_type)); 1854 if (is_active_at_points && rstr_type != CEED_RESTRICTION_POINTS) continue; 1855 if (rstr_type == CEED_RESTRICTION_POINTS) { 1856 CeedCallBackend(CeedElemRestrictionGetMaxPointsInElement(elem_rstr, &elem_size)); 1857 } else { 1858 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 1859 } 1860 { 1861 CeedInt num_comp = 0; 1862 1863 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 1864 if (e_vec_size != num_comp * elem_size) continue; 1865 } 1866 1867 // Basis action 1868 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[j], &eval_mode)); 1869 switch (eval_mode) { 1870 case CEED_EVAL_NONE: { 1871 CeedScalar *e_vec_array; 1872 1873 CeedCallBackend(CeedVectorTakeArray(q_vec, CEED_MEM_DEVICE, &e_vec_array)); 1874 CeedCallBackend(CeedVectorRestoreArray(e_vec, &e_vec_array)); 1875 break; 1876 } 1877 case CEED_EVAL_INTERP: 1878 case CEED_EVAL_GRAD: 1879 case CEED_EVAL_DIV: 1880 case CEED_EVAL_CURL: 1881 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[j], &basis)); 1882 CeedCallBackend(CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_TRANSPOSE, eval_mode, impl->point_coords_elem, q_vec, e_vec)); 1883 break; 1884 // LCOV_EXCL_START 1885 case CEED_EVAL_WEIGHT: { 1886 return CeedError(CeedOperatorReturnCeed(op), CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 1887 // LCOV_EXCL_STOP 1888 } 1889 } 1890 1891 // Mask output e-vec 1892 CeedCallBackend(CeedVectorPointwiseMult(e_vec, active_e_vec_in, e_vec)); 1893 1894 // Restrict 1895 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[j], &elem_rstr)); 1896 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_TRANSPOSE, e_vec, assembled, request)); 1897 1898 // Reset q_vec for 1899 if (eval_mode == CEED_EVAL_NONE) { 1900 CeedScalar *e_vec_array; 1901 1902 CeedCallBackend(CeedVectorGetArrayWrite(e_vec, CEED_MEM_DEVICE, &e_vec_array)); 1903 CeedCallBackend(CeedVectorSetArray(q_vec, CEED_MEM_DEVICE, CEED_USE_POINTER, e_vec_array)); 1904 } 1905 } 1906 1907 // Reset vec 1908 if (s == e_vec_size - 1 && i != num_input_fields - 1) CeedCallBackend(CeedVectorSetValue(q_vec, 0.0)); 1909 } 1910 } 1911 1912 // Restore CEED_EVAL_NONE 1913 for (CeedInt i = 0; i < num_output_fields; i++) { 1914 CeedEvalMode eval_mode; 1915 1916 // Get eval_mode 1917 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1918 1919 // Restore evec 1920 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1921 if (eval_mode == CEED_EVAL_NONE) { 1922 CeedScalar *e_vec_array; 1923 1924 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, &e_vec_array)); 1925 CeedCallBackend(CeedVectorRestoreArray(impl->e_vecs_in[i], &e_vec_array)); 1926 } 1927 } 1928 1929 // Restore input arrays 1930 for (CeedInt i = 0; i < num_input_fields; i++) { 1931 CeedCallBackend(CeedOperatorInputRestore_Hip(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, true, impl)); 1932 } 1933 return CEED_ERROR_SUCCESS; 1934 } 1935 1936 //------------------------------------------------------------------------------ 1937 // Create operator 1938 //------------------------------------------------------------------------------ 1939 int CeedOperatorCreate_Hip(CeedOperator op) { 1940 Ceed ceed; 1941 CeedOperator_Hip *impl; 1942 1943 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1944 CeedCallBackend(CeedCalloc(1, &impl)); 1945 CeedCallBackend(CeedOperatorSetData(op, impl)); 1946 1947 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunction_Hip)); 1948 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunctionUpdate", CeedOperatorLinearAssembleQFunctionUpdate_Hip)); 1949 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddDiagonal", CeedOperatorLinearAssembleAddDiagonal_Hip)); 1950 CeedCallBackend( 1951 CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddPointBlockDiagonal", CeedOperatorLinearAssembleAddPointBlockDiagonal_Hip)); 1952 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleSingle", CeedSingleOperatorAssemble_Hip)); 1953 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Hip)); 1954 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Hip)); 1955 return CEED_ERROR_SUCCESS; 1956 } 1957 1958 //------------------------------------------------------------------------------ 1959 // Create operator AtPoints 1960 //------------------------------------------------------------------------------ 1961 int CeedOperatorCreateAtPoints_Hip(CeedOperator op) { 1962 Ceed ceed; 1963 CeedOperator_Hip *impl; 1964 1965 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1966 CeedCallBackend(CeedCalloc(1, &impl)); 1967 CeedCallBackend(CeedOperatorSetData(op, impl)); 1968 1969 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunctionAtPoints_Hip)); 1970 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddDiagonal", CeedOperatorLinearAssembleAddDiagonalAtPoints_Hip)); 1971 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAddAtPoints_Hip)); 1972 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Hip)); 1973 return CEED_ERROR_SUCCESS; 1974 } 1975 1976 //------------------------------------------------------------------------------ 1977