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 <cuda.h> 13 #include <cuda_runtime.h> 14 #include <stdbool.h> 15 #include <string.h> 16 17 #include "../cuda/ceed-cuda-common.h" 18 #include "../cuda/ceed-cuda-compile.h" 19 #include "ceed-cuda-ref.h" 20 21 //------------------------------------------------------------------------------ 22 // Destroy operator 23 //------------------------------------------------------------------------------ 24 static int CeedOperatorDestroy_Cuda(CeedOperator op) { 25 CeedOperator_Cuda *impl; 26 27 CeedCallBackend(CeedOperatorGetData(op, &impl)); 28 29 // Apply data 30 CeedCallBackend(CeedFree(&impl->num_points)); 31 CeedCallBackend(CeedFree(&impl->skip_rstr_in)); 32 CeedCallBackend(CeedFree(&impl->skip_rstr_out)); 33 CeedCallBackend(CeedFree(&impl->apply_add_basis_out)); 34 CeedCallBackend(CeedFree(&impl->input_field_order)); 35 CeedCallBackend(CeedFree(&impl->output_field_order)); 36 CeedCallBackend(CeedFree(&impl->input_states)); 37 38 for (CeedInt i = 0; i < impl->num_inputs; i++) { 39 CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_in[i])); 40 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_in[i])); 41 } 42 CeedCallBackend(CeedFree(&impl->e_vecs_in)); 43 CeedCallBackend(CeedFree(&impl->q_vecs_in)); 44 45 for (CeedInt i = 0; i < impl->num_outputs; i++) { 46 CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_out[i])); 47 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_out[i])); 48 } 49 CeedCallBackend(CeedFree(&impl->e_vecs_out)); 50 CeedCallBackend(CeedFree(&impl->q_vecs_out)); 51 CeedCallBackend(CeedVectorDestroy(&impl->point_coords_elem)); 52 53 // QFunction assembly data 54 for (CeedInt i = 0; i < impl->num_active_in; i++) { 55 CeedCallBackend(CeedVectorDestroy(&impl->qf_active_in[i])); 56 } 57 CeedCallBackend(CeedFree(&impl->qf_active_in)); 58 59 // Diag data 60 if (impl->diag) { 61 Ceed ceed; 62 63 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 64 if (impl->diag->module) { 65 CeedCallCuda(ceed, cuModuleUnload(impl->diag->module)); 66 } 67 if (impl->diag->module_point_block) { 68 CeedCallCuda(ceed, cuModuleUnload(impl->diag->module_point_block)); 69 } 70 CeedCallCuda(ceed, cudaFree(impl->diag->d_eval_modes_in)); 71 CeedCallCuda(ceed, cudaFree(impl->diag->d_eval_modes_out)); 72 CeedCallCuda(ceed, cudaFree(impl->diag->d_identity)); 73 CeedCallCuda(ceed, cudaFree(impl->diag->d_interp_in)); 74 CeedCallCuda(ceed, cudaFree(impl->diag->d_interp_out)); 75 CeedCallCuda(ceed, cudaFree(impl->diag->d_grad_in)); 76 CeedCallCuda(ceed, cudaFree(impl->diag->d_grad_out)); 77 CeedCallCuda(ceed, cudaFree(impl->diag->d_div_in)); 78 CeedCallCuda(ceed, cudaFree(impl->diag->d_div_out)); 79 CeedCallCuda(ceed, cudaFree(impl->diag->d_curl_in)); 80 CeedCallCuda(ceed, cudaFree(impl->diag->d_curl_out)); 81 CeedCallBackend(CeedElemRestrictionDestroy(&impl->diag->diag_rstr)); 82 CeedCallBackend(CeedElemRestrictionDestroy(&impl->diag->point_block_diag_rstr)); 83 CeedCallBackend(CeedVectorDestroy(&impl->diag->elem_diag)); 84 CeedCallBackend(CeedVectorDestroy(&impl->diag->point_block_elem_diag)); 85 } 86 CeedCallBackend(CeedFree(&impl->diag)); 87 88 if (impl->asmb) { 89 Ceed ceed; 90 91 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 92 CeedCallCuda(ceed, cuModuleUnload(impl->asmb->module)); 93 CeedCallCuda(ceed, cudaFree(impl->asmb->d_B_in)); 94 CeedCallCuda(ceed, cudaFree(impl->asmb->d_B_out)); 95 } 96 CeedCallBackend(CeedFree(&impl->asmb)); 97 98 CeedCallBackend(CeedFree(&impl)); 99 return CEED_ERROR_SUCCESS; 100 } 101 102 //------------------------------------------------------------------------------ 103 // Setup infields or outfields 104 //------------------------------------------------------------------------------ 105 static int CeedOperatorSetupFields_Cuda(CeedQFunction qf, CeedOperator op, bool is_input, bool is_at_points, bool *skip_rstr, bool *apply_add_basis, 106 CeedVector *e_vecs, CeedVector *q_vecs, CeedInt num_fields, CeedInt Q, CeedInt num_elem) { 107 Ceed ceed; 108 CeedQFunctionField *qf_fields; 109 CeedOperatorField *op_fields; 110 111 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 112 if (is_input) { 113 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 114 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 115 } else { 116 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 117 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 118 } 119 120 // Loop over fields 121 for (CeedInt i = 0; i < num_fields; i++) { 122 bool is_active = false, is_strided = false, skip_e_vec = false; 123 CeedSize q_size; 124 CeedInt size; 125 CeedEvalMode eval_mode; 126 CeedVector l_vec; 127 CeedElemRestriction elem_rstr; 128 129 // Check whether this field can skip the element restriction: 130 // Input CEED_VECTOR_ACTIVE 131 // Output CEED_VECTOR_ACTIVE without CEED_EVAL_NONE 132 // Input CEED_VECTOR_NONE with CEED_EVAL_WEIGHT 133 // Input passive vectorr with CEED_EVAL_NONE and strided restriction with CEED_STRIDES_BACKEND 134 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &l_vec)); 135 is_active = l_vec == CEED_VECTOR_ACTIVE; 136 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &elem_rstr)); 137 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 138 skip_e_vec = (is_input && is_active) || (is_active && eval_mode != CEED_EVAL_NONE) || (eval_mode == CEED_EVAL_WEIGHT); 139 if (!skip_e_vec && is_input && !is_active && eval_mode == CEED_EVAL_NONE) { 140 CeedCallBackend(CeedElemRestrictionIsStrided(elem_rstr, &is_strided)); 141 if (is_strided) CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_rstr, &skip_e_vec)); 142 } 143 if (skip_e_vec) { 144 e_vecs[i] = NULL; 145 } else { 146 CeedCallBackend(CeedElemRestrictionCreateVector(elem_rstr, NULL, &e_vecs[i])); 147 } 148 149 switch (eval_mode) { 150 case CEED_EVAL_NONE: 151 case CEED_EVAL_INTERP: 152 case CEED_EVAL_GRAD: 153 case CEED_EVAL_DIV: 154 case CEED_EVAL_CURL: 155 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 156 q_size = (CeedSize)num_elem * (CeedSize)Q * (CeedSize)size; 157 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 158 break; 159 case CEED_EVAL_WEIGHT: { 160 CeedBasis basis; 161 162 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 163 q_size = (CeedSize)num_elem * (CeedSize)Q; 164 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 165 if (is_at_points) { 166 CeedInt num_points[num_elem]; 167 168 for (CeedInt i = 0; i < num_elem; i++) num_points[i] = Q; 169 CeedCallBackend( 170 CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, CEED_VECTOR_NONE, q_vecs[i])); 171 } else { 172 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, q_vecs[i])); 173 } 174 break; 175 } 176 } 177 } 178 // Drop duplicate restrictions 179 if (is_input) { 180 for (CeedInt i = 0; i < num_fields; i++) { 181 CeedVector vec_i; 182 CeedElemRestriction rstr_i; 183 184 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec_i)); 185 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr_i)); 186 for (CeedInt j = i + 1; j < num_fields; j++) { 187 CeedVector vec_j; 188 CeedElemRestriction rstr_j; 189 190 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[j], &vec_j)); 191 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[j], &rstr_j)); 192 if (vec_i == vec_j && rstr_i == rstr_j) { 193 if (e_vecs[i]) CeedCallBackend(CeedVectorReferenceCopy(e_vecs[i], &e_vecs[j])); 194 skip_rstr[j] = true; 195 } 196 } 197 } 198 } else { 199 for (CeedInt i = num_fields - 1; i >= 0; i--) { 200 CeedVector vec_i; 201 CeedElemRestriction rstr_i; 202 203 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec_i)); 204 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr_i)); 205 for (CeedInt j = i - 1; j >= 0; j--) { 206 CeedVector vec_j; 207 CeedElemRestriction rstr_j; 208 209 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[j], &vec_j)); 210 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[j], &rstr_j)); 211 if (vec_i == vec_j && rstr_i == rstr_j) { 212 if (e_vecs[i]) CeedCallBackend(CeedVectorReferenceCopy(e_vecs[i], &e_vecs[j])); 213 skip_rstr[j] = true; 214 apply_add_basis[i] = true; 215 } 216 } 217 } 218 } 219 return CEED_ERROR_SUCCESS; 220 } 221 222 //------------------------------------------------------------------------------ 223 // CeedOperator needs to connect all the named fields (be they active or passive) to the named inputs and outputs of its CeedQFunction. 224 //------------------------------------------------------------------------------ 225 static int CeedOperatorSetup_Cuda(CeedOperator op) { 226 Ceed ceed; 227 bool is_setup_done; 228 CeedInt Q, num_elem, num_input_fields, num_output_fields; 229 CeedQFunctionField *qf_input_fields, *qf_output_fields; 230 CeedQFunction qf; 231 CeedOperatorField *op_input_fields, *op_output_fields; 232 CeedOperator_Cuda *impl; 233 234 CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done)); 235 if (is_setup_done) return CEED_ERROR_SUCCESS; 236 237 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 238 CeedCallBackend(CeedOperatorGetData(op, &impl)); 239 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 240 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 241 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 242 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 243 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 244 245 // Allocate 246 CeedCallBackend(CeedCalloc(num_input_fields, &impl->e_vecs_in)); 247 CeedCallBackend(CeedCalloc(num_output_fields, &impl->e_vecs_out)); 248 CeedCallBackend(CeedCalloc(num_input_fields, &impl->skip_rstr_in)); 249 CeedCallBackend(CeedCalloc(num_output_fields, &impl->skip_rstr_out)); 250 CeedCallBackend(CeedCalloc(num_output_fields, &impl->apply_add_basis_out)); 251 CeedCallBackend(CeedCalloc(num_input_fields, &impl->input_field_order)); 252 CeedCallBackend(CeedCalloc(num_output_fields, &impl->output_field_order)); 253 CeedCallBackend(CeedCalloc(num_input_fields, &impl->input_states)); 254 CeedCallBackend(CeedCalloc(num_input_fields, &impl->q_vecs_in)); 255 CeedCallBackend(CeedCalloc(num_output_fields, &impl->q_vecs_out)); 256 impl->num_inputs = num_input_fields; 257 impl->num_outputs = num_output_fields; 258 259 // Set up infield and outfield e-vecs and q-vecs 260 CeedCallBackend( 261 CeedOperatorSetupFields_Cuda(qf, op, true, false, impl->skip_rstr_in, NULL, impl->e_vecs_in, impl->q_vecs_in, num_input_fields, Q, num_elem)); 262 CeedCallBackend(CeedOperatorSetupFields_Cuda(qf, op, false, false, impl->skip_rstr_out, impl->apply_add_basis_out, impl->e_vecs_out, 263 impl->q_vecs_out, num_output_fields, Q, num_elem)); 264 265 // Reorder fields to allow reuse of buffers 266 impl->max_active_e_vec_len = 0; 267 { 268 bool is_ordered[CEED_FIELD_MAX]; 269 CeedInt curr_index = 0; 270 271 for (CeedInt i = 0; i < num_input_fields; i++) is_ordered[i] = false; 272 for (CeedInt i = 0; i < num_input_fields; i++) { 273 CeedSize e_vec_len_i; 274 CeedVector vec_i; 275 CeedElemRestriction rstr_i; 276 277 if (is_ordered[i]) continue; 278 is_ordered[i] = true; 279 impl->input_field_order[curr_index] = i; 280 curr_index++; 281 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec_i)); 282 if (vec_i == CEED_VECTOR_NONE) continue; // CEED_EVAL_WEIGHT 283 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &rstr_i)); 284 CeedCallBackend(CeedElemRestrictionGetEVectorSize(rstr_i, &e_vec_len_i)); 285 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; 286 for (CeedInt j = i + 1; j < num_input_fields; j++) { 287 CeedVector vec_j; 288 CeedElemRestriction rstr_j; 289 290 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[j], &vec_j)); 291 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[j], &rstr_j)); 292 if (rstr_i == rstr_j && vec_i == vec_j) { 293 is_ordered[j] = true; 294 impl->input_field_order[curr_index] = j; 295 curr_index++; 296 } 297 } 298 } 299 } 300 { 301 bool is_ordered[CEED_FIELD_MAX]; 302 CeedInt curr_index = 0; 303 304 for (CeedInt i = 0; i < num_output_fields; i++) is_ordered[i] = false; 305 for (CeedInt i = 0; i < num_output_fields; i++) { 306 CeedSize e_vec_len_i; 307 CeedVector vec_i; 308 CeedElemRestriction rstr_i; 309 310 if (is_ordered[i]) continue; 311 is_ordered[i] = true; 312 impl->output_field_order[curr_index] = i; 313 curr_index++; 314 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec_i)); 315 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &rstr_i)); 316 CeedCallBackend(CeedElemRestrictionGetEVectorSize(rstr_i, &e_vec_len_i)); 317 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; 318 for (CeedInt j = i + 1; j < num_output_fields; j++) { 319 CeedVector vec_j; 320 CeedElemRestriction rstr_j; 321 322 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[j], &vec_j)); 323 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[j], &rstr_j)); 324 if (rstr_i == rstr_j && vec_i == vec_j) { 325 is_ordered[j] = true; 326 impl->output_field_order[curr_index] = j; 327 curr_index++; 328 } 329 } 330 } 331 } 332 CeedCallBackend(CeedOperatorSetSetupDone(op)); 333 return CEED_ERROR_SUCCESS; 334 } 335 336 //------------------------------------------------------------------------------ 337 // Restrict Operator Inputs 338 //------------------------------------------------------------------------------ 339 static inline int CeedOperatorInputRestrict_Cuda(CeedOperatorField op_input_field, CeedQFunctionField qf_input_field, CeedInt input_field, 340 CeedVector in_vec, CeedVector active_e_vec, const bool skip_active, CeedOperator_Cuda *impl, 341 CeedRequest *request) { 342 bool is_active = false; 343 CeedVector l_vec, e_vec = impl->e_vecs_in[input_field]; 344 345 // Get input vector 346 CeedCallBackend(CeedOperatorFieldGetVector(op_input_field, &l_vec)); 347 is_active = l_vec == CEED_VECTOR_ACTIVE; 348 if (is_active && skip_active) return CEED_ERROR_SUCCESS; 349 if (is_active) { 350 l_vec = in_vec; 351 if (!e_vec) e_vec = active_e_vec; 352 } 353 354 // Restriction action 355 if (e_vec) { 356 // Restrict, if necessary 357 if (!impl->skip_rstr_in[input_field]) { 358 uint64_t state; 359 360 CeedCallBackend(CeedVectorGetState(l_vec, &state)); 361 if (is_active || state != impl->input_states[input_field]) { 362 CeedElemRestriction elem_rstr; 363 364 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_field, &elem_rstr)); 365 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_NOTRANSPOSE, l_vec, e_vec, request)); 366 } 367 impl->input_states[input_field] = state; 368 } 369 } 370 return CEED_ERROR_SUCCESS; 371 } 372 373 //------------------------------------------------------------------------------ 374 // Input Basis Action 375 //------------------------------------------------------------------------------ 376 static inline int CeedOperatorInputBasis_Cuda(CeedOperatorField op_input_field, CeedQFunctionField qf_input_field, CeedInt input_field, 377 CeedVector in_vec, CeedVector active_e_vec, CeedInt num_elem, const bool skip_active, 378 CeedOperator_Cuda *impl) { 379 bool is_active = false; 380 CeedEvalMode eval_mode; 381 CeedVector l_vec, e_vec = impl->e_vecs_in[input_field], q_vec = impl->q_vecs_in[input_field]; 382 383 // Skip active input 384 CeedCallBackend(CeedOperatorFieldGetVector(op_input_field, &l_vec)); 385 is_active = l_vec == CEED_VECTOR_ACTIVE; 386 if (is_active && skip_active) return CEED_ERROR_SUCCESS; 387 if (is_active) { 388 l_vec = in_vec; 389 if (!e_vec) e_vec = active_e_vec; 390 } 391 392 // Basis action 393 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_field, &eval_mode)); 394 switch (eval_mode) { 395 case CEED_EVAL_NONE: { 396 const CeedScalar *e_vec_array; 397 398 if (e_vec) { 399 CeedCallBackend(CeedVectorGetArrayRead(e_vec, CEED_MEM_DEVICE, &e_vec_array)); 400 } else { 401 CeedCallBackend(CeedVectorGetArrayRead(l_vec, CEED_MEM_DEVICE, &e_vec_array)); 402 } 403 CeedCallBackend(CeedVectorSetArray(q_vec, CEED_MEM_DEVICE, CEED_USE_POINTER, (CeedScalar *)e_vec_array)); 404 break; 405 } 406 case CEED_EVAL_INTERP: 407 case CEED_EVAL_GRAD: 408 case CEED_EVAL_DIV: 409 case CEED_EVAL_CURL: { 410 CeedBasis basis; 411 412 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_field, &basis)); 413 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, eval_mode, e_vec, q_vec)); 414 break; 415 } 416 case CEED_EVAL_WEIGHT: 417 break; // No action 418 } 419 return CEED_ERROR_SUCCESS; 420 } 421 422 //------------------------------------------------------------------------------ 423 // Restore Input Vectors 424 //------------------------------------------------------------------------------ 425 static inline int CeedOperatorInputRestore_Cuda(CeedOperatorField op_input_field, CeedQFunctionField qf_input_field, CeedInt input_field, 426 CeedVector in_vec, CeedVector active_e_vec, const bool skip_active, CeedOperator_Cuda *impl) { 427 bool is_active = false; 428 CeedEvalMode eval_mode; 429 CeedVector l_vec, e_vec = impl->e_vecs_in[input_field]; 430 431 // Skip active input 432 CeedCallBackend(CeedOperatorFieldGetVector(op_input_field, &l_vec)); 433 is_active = l_vec == CEED_VECTOR_ACTIVE; 434 if (is_active && skip_active) return CEED_ERROR_SUCCESS; 435 if (is_active) { 436 l_vec = in_vec; 437 if (!e_vec) e_vec = active_e_vec; 438 } 439 440 // Restore e-vec 441 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_field, &eval_mode)); 442 if (eval_mode == CEED_EVAL_NONE) { 443 const CeedScalar *e_vec_array; 444 445 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_in[input_field], CEED_MEM_DEVICE, (CeedScalar **)&e_vec_array)); 446 if (e_vec) { 447 CeedCallBackend(CeedVectorRestoreArrayRead(e_vec, &e_vec_array)); 448 } else { 449 CeedCallBackend(CeedVectorRestoreArrayRead(l_vec, &e_vec_array)); 450 } 451 } 452 return CEED_ERROR_SUCCESS; 453 } 454 455 //------------------------------------------------------------------------------ 456 // Apply and add to output 457 //------------------------------------------------------------------------------ 458 static int CeedOperatorApplyAdd_Cuda(CeedOperator op, CeedVector in_vec, CeedVector out_vec, CeedRequest *request) { 459 CeedInt Q, num_elem, num_input_fields, num_output_fields; 460 Ceed ceed; 461 CeedVector active_e_vec; 462 CeedQFunctionField *qf_input_fields, *qf_output_fields; 463 CeedQFunction qf; 464 CeedOperatorField *op_input_fields, *op_output_fields; 465 CeedOperator_Cuda *impl; 466 467 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 468 CeedCallBackend(CeedOperatorGetData(op, &impl)); 469 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 470 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 471 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 472 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 473 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 474 475 // Setup 476 CeedCallBackend(CeedOperatorSetup_Cuda(op)); 477 478 // Work vector 479 CeedCallBackend(CeedGetWorkVector(ceed, impl->max_active_e_vec_len, &active_e_vec)); 480 481 // Process inputs 482 for (CeedInt i = 0; i < num_input_fields; i++) { 483 CeedInt field = impl->input_field_order[i]; 484 485 CeedCallBackend( 486 CeedOperatorInputRestrict_Cuda(op_input_fields[field], qf_input_fields[field], field, in_vec, active_e_vec, false, impl, request)); 487 CeedCallBackend(CeedOperatorInputBasis_Cuda(op_input_fields[field], qf_input_fields[field], field, in_vec, active_e_vec, num_elem, false, impl)); 488 } 489 490 // Output pointers, as necessary 491 for (CeedInt i = 0; i < num_output_fields; i++) { 492 CeedEvalMode eval_mode; 493 494 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 495 if (eval_mode == CEED_EVAL_NONE) { 496 CeedScalar *e_vec_array; 497 498 CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 499 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_vec_array)); 500 } 501 } 502 503 // Q function 504 CeedCallBackend(CeedQFunctionApply(qf, num_elem * Q, impl->q_vecs_in, impl->q_vecs_out)); 505 506 // Restore input arrays 507 for (CeedInt i = 0; i < num_input_fields; i++) { 508 CeedCallBackend(CeedOperatorInputRestore_Cuda(op_input_fields[i], qf_input_fields[i], i, in_vec, active_e_vec, false, impl)); 509 } 510 511 // Output basis and restriction 512 for (CeedInt i = 0; i < num_output_fields; i++) { 513 bool is_active = false; 514 CeedInt field = impl->output_field_order[i]; 515 CeedEvalMode eval_mode; 516 CeedVector l_vec, e_vec = impl->e_vecs_out[field], q_vec = impl->q_vecs_out[field]; 517 CeedElemRestriction elem_rstr; 518 CeedBasis basis; 519 520 // Output vector 521 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[field], &l_vec)); 522 is_active = l_vec == CEED_VECTOR_ACTIVE; 523 if (is_active) { 524 l_vec = out_vec; 525 if (!e_vec) e_vec = active_e_vec; 526 } 527 528 // Basis action 529 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[field], &eval_mode)); 530 switch (eval_mode) { 531 case CEED_EVAL_NONE: 532 break; // No action 533 case CEED_EVAL_INTERP: 534 case CEED_EVAL_GRAD: 535 case CEED_EVAL_DIV: 536 case CEED_EVAL_CURL: 537 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[field], &basis)); 538 if (impl->apply_add_basis_out[field]) { 539 CeedCallBackend(CeedBasisApplyAdd(basis, num_elem, CEED_TRANSPOSE, eval_mode, q_vec, e_vec)); 540 } else { 541 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_TRANSPOSE, eval_mode, q_vec, e_vec)); 542 } 543 break; 544 // LCOV_EXCL_START 545 case CEED_EVAL_WEIGHT: { 546 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 547 // LCOV_EXCL_STOP 548 } 549 } 550 551 // Restore evec 552 if (eval_mode == CEED_EVAL_NONE) { 553 CeedScalar *e_vec_array; 554 555 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 556 CeedCallBackend(CeedVectorRestoreArray(e_vec, &e_vec_array)); 557 } 558 559 // Restrict 560 if (impl->skip_rstr_out[field]) continue; 561 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[field], &elem_rstr)); 562 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_TRANSPOSE, e_vec, l_vec, request)); 563 } 564 565 // Return work vector 566 CeedCallBackend(CeedRestoreWorkVector(ceed, &active_e_vec)); 567 return CEED_ERROR_SUCCESS; 568 } 569 570 //------------------------------------------------------------------------------ 571 // CeedOperator needs to connect all the named fields (be they active or passive) to the named inputs and outputs of its CeedQFunction. 572 //------------------------------------------------------------------------------ 573 static int CeedOperatorSetupAtPoints_Cuda(CeedOperator op) { 574 Ceed ceed; 575 bool is_setup_done; 576 CeedInt max_num_points = -1, num_elem, num_input_fields, num_output_fields; 577 CeedQFunctionField *qf_input_fields, *qf_output_fields; 578 CeedQFunction qf; 579 CeedOperatorField *op_input_fields, *op_output_fields; 580 CeedOperator_Cuda *impl; 581 582 CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done)); 583 if (is_setup_done) return CEED_ERROR_SUCCESS; 584 585 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 586 CeedCallBackend(CeedOperatorGetData(op, &impl)); 587 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 588 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 589 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 590 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 591 { 592 CeedElemRestriction rstr_points = NULL; 593 594 CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, NULL)); 595 CeedCallBackend(CeedElemRestrictionGetMaxPointsInElement(rstr_points, &max_num_points)); 596 CeedCallBackend(CeedCalloc(num_elem, &impl->num_points)); 597 for (CeedInt e = 0; e < num_elem; e++) { 598 CeedInt num_points_elem; 599 600 CeedCallBackend(CeedElemRestrictionGetNumPointsInElement(rstr_points, e, &num_points_elem)); 601 impl->num_points[e] = num_points_elem; 602 } 603 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_points)); 604 } 605 impl->max_num_points = max_num_points; 606 607 // Allocate 608 CeedCallBackend(CeedCalloc(num_input_fields, &impl->e_vecs_in)); 609 CeedCallBackend(CeedCalloc(num_output_fields, &impl->e_vecs_out)); 610 CeedCallBackend(CeedCalloc(num_input_fields, &impl->skip_rstr_in)); 611 CeedCallBackend(CeedCalloc(num_output_fields, &impl->skip_rstr_out)); 612 CeedCallBackend(CeedCalloc(num_output_fields, &impl->apply_add_basis_out)); 613 CeedCallBackend(CeedCalloc(num_input_fields, &impl->input_field_order)); 614 CeedCallBackend(CeedCalloc(num_output_fields, &impl->output_field_order)); 615 CeedCallBackend(CeedCalloc(num_input_fields, &impl->input_states)); 616 CeedCallBackend(CeedCalloc(num_input_fields, &impl->q_vecs_in)); 617 CeedCallBackend(CeedCalloc(num_output_fields, &impl->q_vecs_out)); 618 impl->num_inputs = num_input_fields; 619 impl->num_outputs = num_output_fields; 620 621 // Set up infield and outfield e-vecs and q-vecs 622 CeedCallBackend(CeedOperatorSetupFields_Cuda(qf, op, true, true, impl->skip_rstr_in, NULL, impl->e_vecs_in, impl->q_vecs_in, num_input_fields, 623 max_num_points, num_elem)); 624 CeedCallBackend(CeedOperatorSetupFields_Cuda(qf, op, false, true, impl->skip_rstr_out, impl->apply_add_basis_out, impl->e_vecs_out, 625 impl->q_vecs_out, num_output_fields, max_num_points, num_elem)); 626 627 // Reorder fields to allow reuse of buffers 628 impl->max_active_e_vec_len = 0; 629 { 630 bool is_ordered[CEED_FIELD_MAX]; 631 CeedInt curr_index = 0; 632 633 for (CeedInt i = 0; i < num_input_fields; i++) is_ordered[i] = false; 634 for (CeedInt i = 0; i < num_input_fields; i++) { 635 CeedSize e_vec_len_i; 636 CeedVector vec_i; 637 CeedElemRestriction rstr_i; 638 639 if (is_ordered[i]) continue; 640 is_ordered[i] = true; 641 impl->input_field_order[curr_index] = i; 642 curr_index++; 643 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec_i)); 644 if (vec_i == CEED_VECTOR_NONE) continue; // CEED_EVAL_WEIGHT 645 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &rstr_i)); 646 CeedCallBackend(CeedElemRestrictionGetEVectorSize(rstr_i, &e_vec_len_i)); 647 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; 648 for (CeedInt j = i + 1; j < num_input_fields; j++) { 649 CeedVector vec_j; 650 CeedElemRestriction rstr_j; 651 652 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[j], &vec_j)); 653 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[j], &rstr_j)); 654 if (rstr_i == rstr_j && vec_i == vec_j) { 655 is_ordered[j] = true; 656 impl->input_field_order[curr_index] = j; 657 curr_index++; 658 } 659 } 660 } 661 } 662 { 663 bool is_ordered[CEED_FIELD_MAX]; 664 CeedInt curr_index = 0; 665 666 for (CeedInt i = 0; i < num_output_fields; i++) is_ordered[i] = false; 667 for (CeedInt i = 0; i < num_output_fields; i++) { 668 CeedSize e_vec_len_i; 669 CeedVector vec_i; 670 CeedElemRestriction rstr_i; 671 672 if (is_ordered[i]) continue; 673 is_ordered[i] = true; 674 impl->output_field_order[curr_index] = i; 675 curr_index++; 676 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec_i)); 677 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &rstr_i)); 678 CeedCallBackend(CeedElemRestrictionGetEVectorSize(rstr_i, &e_vec_len_i)); 679 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; 680 for (CeedInt j = i + 1; j < num_output_fields; j++) { 681 CeedVector vec_j; 682 CeedElemRestriction rstr_j; 683 684 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[j], &vec_j)); 685 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[j], &rstr_j)); 686 if (rstr_i == rstr_j && vec_i == vec_j) { 687 is_ordered[j] = true; 688 impl->output_field_order[curr_index] = j; 689 curr_index++; 690 } 691 } 692 } 693 } 694 695 CeedCallBackend(CeedOperatorSetSetupDone(op)); 696 return CEED_ERROR_SUCCESS; 697 } 698 699 //------------------------------------------------------------------------------ 700 // Input Basis Action AtPoints 701 //------------------------------------------------------------------------------ 702 static inline int CeedOperatorInputBasisAtPoints_Cuda(CeedOperatorField op_input_field, CeedQFunctionField qf_input_field, CeedInt input_field, 703 CeedVector in_vec, CeedVector active_e_vec, CeedInt num_elem, const CeedInt *num_points, 704 const bool skip_active, CeedOperator_Cuda *impl) { 705 bool is_active = false; 706 CeedEvalMode eval_mode; 707 CeedVector l_vec, e_vec = impl->e_vecs_in[input_field], q_vec = impl->q_vecs_in[input_field]; 708 709 // Skip active input 710 CeedCallBackend(CeedOperatorFieldGetVector(op_input_field, &l_vec)); 711 is_active = l_vec == CEED_VECTOR_ACTIVE; 712 if (is_active && skip_active) return CEED_ERROR_SUCCESS; 713 if (is_active) { 714 l_vec = in_vec; 715 if (!e_vec) e_vec = active_e_vec; 716 } 717 718 // Basis action 719 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_field, &eval_mode)); 720 switch (eval_mode) { 721 case CEED_EVAL_NONE: { 722 const CeedScalar *e_vec_array; 723 724 if (e_vec) { 725 CeedCallBackend(CeedVectorGetArrayRead(e_vec, CEED_MEM_DEVICE, &e_vec_array)); 726 } else { 727 CeedCallBackend(CeedVectorGetArrayRead(l_vec, CEED_MEM_DEVICE, &e_vec_array)); 728 } 729 CeedCallBackend(CeedVectorSetArray(q_vec, CEED_MEM_DEVICE, CEED_USE_POINTER, (CeedScalar *)e_vec_array)); 730 break; 731 } 732 case CEED_EVAL_INTERP: 733 case CEED_EVAL_GRAD: 734 case CEED_EVAL_DIV: 735 case CEED_EVAL_CURL: { 736 CeedBasis basis; 737 738 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_field, &basis)); 739 CeedCallBackend(CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_NOTRANSPOSE, eval_mode, impl->point_coords_elem, e_vec, q_vec)); 740 break; 741 } 742 case CEED_EVAL_WEIGHT: 743 break; // No action 744 } 745 return CEED_ERROR_SUCCESS; 746 } 747 748 //------------------------------------------------------------------------------ 749 // Apply and add to output AtPoints 750 //------------------------------------------------------------------------------ 751 static int CeedOperatorApplyAddAtPoints_Cuda(CeedOperator op, CeedVector in_vec, CeedVector out_vec, CeedRequest *request) { 752 CeedInt max_num_points, *num_points, num_elem, num_input_fields, num_output_fields; 753 Ceed ceed; 754 CeedVector active_e_vec; 755 CeedQFunctionField *qf_input_fields, *qf_output_fields; 756 CeedQFunction qf; 757 CeedOperatorField *op_input_fields, *op_output_fields; 758 CeedOperator_Cuda *impl; 759 760 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 761 CeedCallBackend(CeedOperatorGetData(op, &impl)); 762 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 763 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 764 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 765 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 766 767 // Setup 768 CeedCallBackend(CeedOperatorSetupAtPoints_Cuda(op)); 769 num_points = impl->num_points; 770 max_num_points = impl->max_num_points; 771 772 // Work vector 773 CeedCallBackend(CeedGetWorkVector(ceed, impl->max_active_e_vec_len, &active_e_vec)); 774 775 // Get point coordinates 776 if (!impl->point_coords_elem) { 777 CeedVector point_coords = NULL; 778 CeedElemRestriction rstr_points = NULL; 779 780 CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, &point_coords)); 781 CeedCallBackend(CeedElemRestrictionCreateVector(rstr_points, NULL, &impl->point_coords_elem)); 782 CeedCallBackend(CeedElemRestrictionApply(rstr_points, CEED_NOTRANSPOSE, point_coords, impl->point_coords_elem, request)); 783 CeedCallBackend(CeedVectorDestroy(&point_coords)); 784 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_points)); 785 } 786 787 // Process inputs 788 for (CeedInt i = 0; i < num_input_fields; i++) { 789 CeedInt field = impl->input_field_order[i]; 790 791 CeedCallBackend( 792 CeedOperatorInputRestrict_Cuda(op_input_fields[field], qf_input_fields[field], field, in_vec, active_e_vec, false, impl, request)); 793 CeedCallBackend(CeedOperatorInputBasisAtPoints_Cuda(op_input_fields[field], qf_input_fields[field], field, in_vec, active_e_vec, num_elem, 794 num_points, false, impl)); 795 } 796 797 // Output pointers, as necessary 798 for (CeedInt i = 0; i < num_output_fields; i++) { 799 CeedEvalMode eval_mode; 800 801 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 802 if (eval_mode == CEED_EVAL_NONE) { 803 CeedScalar *e_vec_array; 804 805 CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 806 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_vec_array)); 807 } 808 } 809 810 // Q function 811 CeedCallBackend(CeedQFunctionApply(qf, num_elem * max_num_points, impl->q_vecs_in, impl->q_vecs_out)); 812 813 // Restore input arrays 814 for (CeedInt i = 0; i < num_input_fields; i++) { 815 CeedCallBackend(CeedOperatorInputRestore_Cuda(op_input_fields[i], qf_input_fields[i], i, in_vec, active_e_vec, false, impl)); 816 } 817 818 // Output basis and restriction 819 for (CeedInt i = 0; i < num_output_fields; i++) { 820 bool is_active = false; 821 CeedInt field = impl->output_field_order[i]; 822 CeedEvalMode eval_mode; 823 CeedVector l_vec, e_vec = impl->e_vecs_out[field], q_vec = impl->q_vecs_out[field]; 824 CeedElemRestriction elem_rstr; 825 CeedBasis basis; 826 827 // Output vector 828 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[field], &l_vec)); 829 is_active = l_vec == CEED_VECTOR_ACTIVE; 830 if (is_active) { 831 l_vec = out_vec; 832 if (!e_vec) e_vec = active_e_vec; 833 } 834 835 // Basis action 836 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[field], &eval_mode)); 837 switch (eval_mode) { 838 case CEED_EVAL_NONE: 839 break; // No action 840 case CEED_EVAL_INTERP: 841 case CEED_EVAL_GRAD: 842 case CEED_EVAL_DIV: 843 case CEED_EVAL_CURL: 844 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[field], &basis)); 845 if (impl->apply_add_basis_out[field]) { 846 CeedCallBackend(CeedBasisApplyAddAtPoints(basis, num_elem, num_points, CEED_TRANSPOSE, eval_mode, impl->point_coords_elem, q_vec, e_vec)); 847 } else { 848 CeedCallBackend(CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_TRANSPOSE, eval_mode, impl->point_coords_elem, q_vec, e_vec)); 849 } 850 break; 851 // LCOV_EXCL_START 852 case CEED_EVAL_WEIGHT: { 853 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 854 // LCOV_EXCL_STOP 855 } 856 } 857 858 // Restore evec 859 if (eval_mode == CEED_EVAL_NONE) { 860 CeedScalar *e_vec_array; 861 862 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 863 CeedCallBackend(CeedVectorRestoreArray(e_vec, &e_vec_array)); 864 } 865 866 // Restrict 867 if (impl->skip_rstr_out[field]) continue; 868 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[field], &elem_rstr)); 869 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_TRANSPOSE, e_vec, l_vec, request)); 870 } 871 872 // Restore work vector 873 CeedCallBackend(CeedRestoreWorkVector(ceed, &active_e_vec)); 874 return CEED_ERROR_SUCCESS; 875 } 876 877 //------------------------------------------------------------------------------ 878 // Linear QFunction Assembly Core 879 //------------------------------------------------------------------------------ 880 static inline int CeedOperatorLinearAssembleQFunctionCore_Cuda(CeedOperator op, bool build_objects, CeedVector *assembled, CeedElemRestriction *rstr, 881 CeedRequest *request) { 882 Ceed ceed, ceed_parent; 883 CeedInt num_active_in, num_active_out, Q, num_elem, num_input_fields, num_output_fields, size; 884 CeedScalar *assembled_array; 885 CeedVector *active_inputs; 886 CeedQFunctionField *qf_input_fields, *qf_output_fields; 887 CeedQFunction qf; 888 CeedOperatorField *op_input_fields, *op_output_fields; 889 CeedOperator_Cuda *impl; 890 891 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 892 CeedCallBackend(CeedOperatorGetFallbackParentCeed(op, &ceed_parent)); 893 CeedCallBackend(CeedOperatorGetData(op, &impl)); 894 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 895 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 896 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 897 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 898 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 899 active_inputs = impl->qf_active_in; 900 num_active_in = impl->num_active_in, num_active_out = impl->num_active_out; 901 902 // Setup 903 CeedCallBackend(CeedOperatorSetup_Cuda(op)); 904 905 // Process inputs 906 for (CeedInt i = 0; i < num_input_fields; i++) { 907 CeedCallBackend(CeedOperatorInputRestrict_Cuda(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, true, impl, request)); 908 CeedCallBackend(CeedOperatorInputBasis_Cuda(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, num_elem, true, impl)); 909 } 910 911 // Count number of active input fields 912 if (!num_active_in) { 913 for (CeedInt i = 0; i < num_input_fields; i++) { 914 CeedScalar *q_vec_array; 915 CeedVector l_vec; 916 917 // Check if active input 918 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &l_vec)); 919 if (l_vec == CEED_VECTOR_ACTIVE) { 920 CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size)); 921 CeedCallBackend(CeedVectorSetValue(impl->q_vecs_in[i], 0.0)); 922 CeedCallBackend(CeedVectorGetArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, &q_vec_array)); 923 CeedCallBackend(CeedRealloc(num_active_in + size, &active_inputs)); 924 for (CeedInt field = 0; field < size; field++) { 925 CeedSize q_size = (CeedSize)Q * num_elem; 926 927 CeedCallBackend(CeedVectorCreate(ceed, q_size, &active_inputs[num_active_in + field])); 928 CeedCallBackend( 929 CeedVectorSetArray(active_inputs[num_active_in + field], CEED_MEM_DEVICE, CEED_USE_POINTER, &q_vec_array[field * Q * num_elem])); 930 } 931 num_active_in += size; 932 CeedCallBackend(CeedVectorRestoreArray(impl->q_vecs_in[i], &q_vec_array)); 933 } 934 } 935 impl->num_active_in = num_active_in; 936 impl->qf_active_in = active_inputs; 937 } 938 939 // Count number of active output fields 940 if (!num_active_out) { 941 for (CeedInt i = 0; i < num_output_fields; i++) { 942 CeedVector l_vec; 943 944 // Check if active output 945 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &l_vec)); 946 if (l_vec == CEED_VECTOR_ACTIVE) { 947 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size)); 948 num_active_out += size; 949 } 950 } 951 impl->num_active_out = num_active_out; 952 } 953 954 // Check sizes 955 CeedCheck(num_active_in > 0 && num_active_out > 0, ceed, CEED_ERROR_BACKEND, "Cannot assemble QFunction without active inputs and outputs"); 956 957 // Build objects if needed 958 if (build_objects) { 959 CeedSize l_size = (CeedSize)num_elem * Q * num_active_in * num_active_out; 960 CeedInt strides[3] = {1, num_elem * Q, Q}; /* *NOPAD* */ 961 962 // Create output restriction 963 CeedCallBackend(CeedElemRestrictionCreateStrided(ceed_parent, num_elem, Q, num_active_in * num_active_out, 964 (CeedSize)num_active_in * (CeedSize)num_active_out * (CeedSize)num_elem * (CeedSize)Q, strides, 965 rstr)); 966 // Create assembled vector 967 CeedCallBackend(CeedVectorCreate(ceed_parent, l_size, assembled)); 968 } 969 CeedCallBackend(CeedVectorSetValue(*assembled, 0.0)); 970 CeedCallBackend(CeedVectorGetArray(*assembled, CEED_MEM_DEVICE, &assembled_array)); 971 972 // Assemble QFunction 973 for (CeedInt in = 0; in < num_active_in; in++) { 974 // Set Inputs 975 CeedCallBackend(CeedVectorSetValue(active_inputs[in], 1.0)); 976 if (num_active_in > 1) { 977 CeedCallBackend(CeedVectorSetValue(active_inputs[(in + num_active_in - 1) % num_active_in], 0.0)); 978 } 979 // Set Outputs 980 for (CeedInt out = 0; out < num_output_fields; out++) { 981 CeedVector l_vec; 982 983 // Get output vector 984 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &l_vec)); 985 // Check if active output 986 if (l_vec == CEED_VECTOR_ACTIVE) { 987 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, CEED_USE_POINTER, assembled_array)); 988 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[out], &size)); 989 assembled_array += size * Q * num_elem; // Advance the pointer by the size of the output 990 } 991 } 992 // Apply QFunction 993 CeedCallBackend(CeedQFunctionApply(qf, Q * num_elem, impl->q_vecs_in, impl->q_vecs_out)); 994 } 995 996 // Un-set output q-vecs to prevent accidental overwrite of Assembled 997 for (CeedInt out = 0; out < num_output_fields; out++) { 998 CeedVector l_vec; 999 1000 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &l_vec)); 1001 if (l_vec == CEED_VECTOR_ACTIVE) { 1002 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, NULL)); 1003 } 1004 } 1005 1006 // Restore input arrays 1007 for (CeedInt i = 0; i < num_input_fields; i++) { 1008 CeedCallBackend(CeedOperatorInputRestore_Cuda(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, true, impl)); 1009 } 1010 1011 // Restore output 1012 CeedCallBackend(CeedVectorRestoreArray(*assembled, &assembled_array)); 1013 return CEED_ERROR_SUCCESS; 1014 } 1015 1016 //------------------------------------------------------------------------------ 1017 // Assemble Linear QFunction 1018 //------------------------------------------------------------------------------ 1019 static int CeedOperatorLinearAssembleQFunction_Cuda(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 1020 return CeedOperatorLinearAssembleQFunctionCore_Cuda(op, true, assembled, rstr, request); 1021 } 1022 1023 //------------------------------------------------------------------------------ 1024 // Update Assembled Linear QFunction 1025 //------------------------------------------------------------------------------ 1026 static int CeedOperatorLinearAssembleQFunctionUpdate_Cuda(CeedOperator op, CeedVector assembled, CeedElemRestriction rstr, CeedRequest *request) { 1027 return CeedOperatorLinearAssembleQFunctionCore_Cuda(op, false, &assembled, &rstr, request); 1028 } 1029 1030 //------------------------------------------------------------------------------ 1031 // Assemble Diagonal Setup 1032 //------------------------------------------------------------------------------ 1033 static inline int CeedOperatorAssembleDiagonalSetup_Cuda(CeedOperator op) { 1034 Ceed ceed; 1035 CeedInt num_input_fields, num_output_fields, num_eval_modes_in = 0, num_eval_modes_out = 0; 1036 CeedInt q_comp, num_nodes, num_qpts; 1037 CeedEvalMode *eval_modes_in = NULL, *eval_modes_out = NULL; 1038 CeedBasis basis_in = NULL, basis_out = NULL; 1039 CeedQFunctionField *qf_fields; 1040 CeedQFunction qf; 1041 CeedOperatorField *op_fields; 1042 CeedOperator_Cuda *impl; 1043 1044 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1045 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1046 CeedCallBackend(CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields)); 1047 1048 // Determine active input basis 1049 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 1050 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 1051 for (CeedInt i = 0; i < num_input_fields; i++) { 1052 CeedVector vec; 1053 1054 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1055 if (vec == CEED_VECTOR_ACTIVE) { 1056 CeedBasis basis; 1057 CeedEvalMode eval_mode; 1058 1059 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 1060 CeedCheck(!basis_in || basis_in == basis, ceed, CEED_ERROR_BACKEND, 1061 "Backend does not implement operator diagonal assembly with multiple active bases"); 1062 basis_in = basis; 1063 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1064 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_in, eval_mode, &q_comp)); 1065 if (eval_mode != CEED_EVAL_WEIGHT) { 1066 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF assembly 1067 CeedCallBackend(CeedRealloc(num_eval_modes_in + q_comp, &eval_modes_in)); 1068 for (CeedInt d = 0; d < q_comp; d++) eval_modes_in[num_eval_modes_in + d] = eval_mode; 1069 num_eval_modes_in += q_comp; 1070 } 1071 } 1072 } 1073 1074 // Determine active output basis 1075 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 1076 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 1077 for (CeedInt i = 0; i < num_output_fields; i++) { 1078 CeedVector vec; 1079 1080 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1081 if (vec == CEED_VECTOR_ACTIVE) { 1082 CeedBasis basis; 1083 CeedEvalMode eval_mode; 1084 1085 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 1086 CeedCheck(!basis_out || basis_out == basis, ceed, CEED_ERROR_BACKEND, 1087 "Backend does not implement operator diagonal assembly with multiple active bases"); 1088 basis_out = basis; 1089 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1090 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_out, eval_mode, &q_comp)); 1091 if (eval_mode != CEED_EVAL_WEIGHT) { 1092 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF assembly 1093 CeedCallBackend(CeedRealloc(num_eval_modes_out + q_comp, &eval_modes_out)); 1094 for (CeedInt d = 0; d < q_comp; d++) eval_modes_out[num_eval_modes_out + d] = eval_mode; 1095 num_eval_modes_out += q_comp; 1096 } 1097 } 1098 } 1099 1100 // Operator data struct 1101 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1102 CeedCallBackend(CeedCalloc(1, &impl->diag)); 1103 CeedOperatorDiag_Cuda *diag = impl->diag; 1104 1105 // Basis matrices 1106 CeedCallBackend(CeedBasisGetNumNodes(basis_in, &num_nodes)); 1107 if (basis_in == CEED_BASIS_NONE) num_qpts = num_nodes; 1108 else CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 1109 const CeedInt interp_bytes = num_nodes * num_qpts * sizeof(CeedScalar); 1110 const CeedInt eval_modes_bytes = sizeof(CeedEvalMode); 1111 bool has_eval_none = false; 1112 1113 // CEED_EVAL_NONE 1114 for (CeedInt i = 0; i < num_eval_modes_in; i++) has_eval_none = has_eval_none || (eval_modes_in[i] == CEED_EVAL_NONE); 1115 for (CeedInt i = 0; i < num_eval_modes_out; i++) has_eval_none = has_eval_none || (eval_modes_out[i] == CEED_EVAL_NONE); 1116 if (has_eval_none) { 1117 CeedScalar *identity = NULL; 1118 1119 CeedCallBackend(CeedCalloc(num_nodes * num_qpts, &identity)); 1120 for (CeedInt i = 0; i < (num_nodes < num_qpts ? num_nodes : num_qpts); i++) identity[i * num_nodes + i] = 1.0; 1121 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_identity, interp_bytes)); 1122 CeedCallCuda(ceed, cudaMemcpy(diag->d_identity, identity, interp_bytes, cudaMemcpyHostToDevice)); 1123 CeedCallBackend(CeedFree(&identity)); 1124 } 1125 1126 // CEED_EVAL_INTERP, CEED_EVAL_GRAD, CEED_EVAL_DIV, and CEED_EVAL_CURL 1127 for (CeedInt in = 0; in < 2; in++) { 1128 CeedFESpace fespace; 1129 CeedBasis basis = in ? basis_in : basis_out; 1130 1131 CeedCallBackend(CeedBasisGetFESpace(basis, &fespace)); 1132 switch (fespace) { 1133 case CEED_FE_SPACE_H1: { 1134 CeedInt q_comp_interp, q_comp_grad; 1135 const CeedScalar *interp, *grad; 1136 CeedScalar *d_interp, *d_grad; 1137 1138 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 1139 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_grad)); 1140 1141 CeedCallBackend(CeedBasisGetInterp(basis, &interp)); 1142 CeedCallCuda(ceed, cudaMalloc((void **)&d_interp, interp_bytes * q_comp_interp)); 1143 CeedCallCuda(ceed, cudaMemcpy(d_interp, interp, interp_bytes * q_comp_interp, cudaMemcpyHostToDevice)); 1144 CeedCallBackend(CeedBasisGetGrad(basis, &grad)); 1145 CeedCallCuda(ceed, cudaMalloc((void **)&d_grad, interp_bytes * q_comp_grad)); 1146 CeedCallCuda(ceed, cudaMemcpy(d_grad, grad, interp_bytes * q_comp_grad, cudaMemcpyHostToDevice)); 1147 if (in) { 1148 diag->d_interp_in = d_interp; 1149 diag->d_grad_in = d_grad; 1150 } else { 1151 diag->d_interp_out = d_interp; 1152 diag->d_grad_out = d_grad; 1153 } 1154 } break; 1155 case CEED_FE_SPACE_HDIV: { 1156 CeedInt q_comp_interp, q_comp_div; 1157 const CeedScalar *interp, *div; 1158 CeedScalar *d_interp, *d_div; 1159 1160 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 1161 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_DIV, &q_comp_div)); 1162 1163 CeedCallBackend(CeedBasisGetInterp(basis, &interp)); 1164 CeedCallCuda(ceed, cudaMalloc((void **)&d_interp, interp_bytes * q_comp_interp)); 1165 CeedCallCuda(ceed, cudaMemcpy(d_interp, interp, interp_bytes * q_comp_interp, cudaMemcpyHostToDevice)); 1166 CeedCallBackend(CeedBasisGetDiv(basis, &div)); 1167 CeedCallCuda(ceed, cudaMalloc((void **)&d_div, interp_bytes * q_comp_div)); 1168 CeedCallCuda(ceed, cudaMemcpy(d_div, div, interp_bytes * q_comp_div, cudaMemcpyHostToDevice)); 1169 if (in) { 1170 diag->d_interp_in = d_interp; 1171 diag->d_div_in = d_div; 1172 } else { 1173 diag->d_interp_out = d_interp; 1174 diag->d_div_out = d_div; 1175 } 1176 } break; 1177 case CEED_FE_SPACE_HCURL: { 1178 CeedInt q_comp_interp, q_comp_curl; 1179 const CeedScalar *interp, *curl; 1180 CeedScalar *d_interp, *d_curl; 1181 1182 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 1183 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_CURL, &q_comp_curl)); 1184 1185 CeedCallBackend(CeedBasisGetInterp(basis, &interp)); 1186 CeedCallCuda(ceed, cudaMalloc((void **)&d_interp, interp_bytes * q_comp_interp)); 1187 CeedCallCuda(ceed, cudaMemcpy(d_interp, interp, interp_bytes * q_comp_interp, cudaMemcpyHostToDevice)); 1188 CeedCallBackend(CeedBasisGetCurl(basis, &curl)); 1189 CeedCallCuda(ceed, cudaMalloc((void **)&d_curl, interp_bytes * q_comp_curl)); 1190 CeedCallCuda(ceed, cudaMemcpy(d_curl, curl, interp_bytes * q_comp_curl, cudaMemcpyHostToDevice)); 1191 if (in) { 1192 diag->d_interp_in = d_interp; 1193 diag->d_curl_in = d_curl; 1194 } else { 1195 diag->d_interp_out = d_interp; 1196 diag->d_curl_out = d_curl; 1197 } 1198 } break; 1199 } 1200 } 1201 1202 // Arrays of eval_modes 1203 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_eval_modes_in, num_eval_modes_in * eval_modes_bytes)); 1204 CeedCallCuda(ceed, cudaMemcpy(diag->d_eval_modes_in, eval_modes_in, num_eval_modes_in * eval_modes_bytes, cudaMemcpyHostToDevice)); 1205 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_eval_modes_out, num_eval_modes_out * eval_modes_bytes)); 1206 CeedCallCuda(ceed, cudaMemcpy(diag->d_eval_modes_out, eval_modes_out, num_eval_modes_out * eval_modes_bytes, cudaMemcpyHostToDevice)); 1207 CeedCallBackend(CeedFree(&eval_modes_in)); 1208 CeedCallBackend(CeedFree(&eval_modes_out)); 1209 return CEED_ERROR_SUCCESS; 1210 } 1211 1212 //------------------------------------------------------------------------------ 1213 // Assemble Diagonal Setup (Compilation) 1214 //------------------------------------------------------------------------------ 1215 static inline int CeedOperatorAssembleDiagonalSetupCompile_Cuda(CeedOperator op, CeedInt use_ceedsize_idx, const bool is_point_block) { 1216 Ceed ceed; 1217 char *diagonal_kernel_source; 1218 const char *diagonal_kernel_path; 1219 CeedInt num_input_fields, num_output_fields, num_eval_modes_in = 0, num_eval_modes_out = 0; 1220 CeedInt num_comp, q_comp, num_nodes, num_qpts; 1221 CeedBasis basis_in = NULL, basis_out = NULL; 1222 CeedQFunctionField *qf_fields; 1223 CeedQFunction qf; 1224 CeedOperatorField *op_fields; 1225 CeedOperator_Cuda *impl; 1226 1227 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1228 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1229 CeedCallBackend(CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields)); 1230 1231 // Determine active input basis 1232 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 1233 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 1234 for (CeedInt i = 0; i < num_input_fields; i++) { 1235 CeedVector vec; 1236 1237 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1238 if (vec == CEED_VECTOR_ACTIVE) { 1239 CeedEvalMode eval_mode; 1240 1241 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_in)); 1242 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1243 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_in, eval_mode, &q_comp)); 1244 if (eval_mode != CEED_EVAL_WEIGHT) { 1245 num_eval_modes_in += q_comp; 1246 } 1247 } 1248 } 1249 1250 // Determine active output basis 1251 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 1252 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 1253 for (CeedInt i = 0; i < num_output_fields; i++) { 1254 CeedVector vec; 1255 1256 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 1257 if (vec == CEED_VECTOR_ACTIVE) { 1258 CeedEvalMode eval_mode; 1259 1260 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_out)); 1261 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1262 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_out, eval_mode, &q_comp)); 1263 if (eval_mode != CEED_EVAL_WEIGHT) { 1264 num_eval_modes_out += q_comp; 1265 } 1266 } 1267 } 1268 1269 // Operator data struct 1270 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1271 CeedOperatorDiag_Cuda *diag = impl->diag; 1272 1273 // Assemble kernel 1274 CUmodule *module = is_point_block ? &diag->module_point_block : &diag->module; 1275 CeedInt elems_per_block = 1; 1276 CeedCallBackend(CeedBasisGetNumNodes(basis_in, &num_nodes)); 1277 CeedCallBackend(CeedBasisGetNumComponents(basis_in, &num_comp)); 1278 if (basis_in == CEED_BASIS_NONE) num_qpts = num_nodes; 1279 else CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 1280 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-operator-assemble-diagonal.h", &diagonal_kernel_path)); 1281 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Kernel Source -----\n"); 1282 CeedCallBackend(CeedLoadSourceToBuffer(ceed, diagonal_kernel_path, &diagonal_kernel_source)); 1283 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Source Complete! -----\n"); 1284 CeedCallCuda(ceed, CeedCompile_Cuda(ceed, diagonal_kernel_source, module, 8, "NUM_EVAL_MODES_IN", num_eval_modes_in, "NUM_EVAL_MODES_OUT", 1285 num_eval_modes_out, "NUM_COMP", num_comp, "NUM_NODES", num_nodes, "NUM_QPTS", num_qpts, "USE_CEEDSIZE", 1286 use_ceedsize_idx, "USE_POINT_BLOCK", is_point_block ? 1 : 0, "BLOCK_SIZE", num_nodes * elems_per_block)); 1287 CeedCallCuda(ceed, CeedGetKernel_Cuda(ceed, *module, "LinearDiagonal", is_point_block ? &diag->LinearPointBlock : &diag->LinearDiagonal)); 1288 CeedCallBackend(CeedFree(&diagonal_kernel_path)); 1289 CeedCallBackend(CeedFree(&diagonal_kernel_source)); 1290 return CEED_ERROR_SUCCESS; 1291 } 1292 1293 //------------------------------------------------------------------------------ 1294 // Assemble Diagonal Core 1295 //------------------------------------------------------------------------------ 1296 static inline int CeedOperatorAssembleDiagonalCore_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request, const bool is_point_block) { 1297 Ceed ceed; 1298 CeedInt num_elem, num_nodes; 1299 CeedScalar *elem_diag_array; 1300 const CeedScalar *assembled_qf_array; 1301 CeedVector assembled_qf = NULL, elem_diag; 1302 CeedElemRestriction assembled_rstr = NULL, rstr_in, rstr_out, diag_rstr; 1303 CeedOperator_Cuda *impl; 1304 1305 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1306 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1307 1308 // Assemble QFunction 1309 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &assembled_rstr, request)); 1310 CeedCallBackend(CeedElemRestrictionDestroy(&assembled_rstr)); 1311 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &assembled_qf_array)); 1312 1313 // Setup 1314 if (!impl->diag) CeedCallBackend(CeedOperatorAssembleDiagonalSetup_Cuda(op)); 1315 CeedOperatorDiag_Cuda *diag = impl->diag; 1316 1317 assert(diag != NULL); 1318 1319 // Assemble kernel if needed 1320 if ((!is_point_block && !diag->LinearDiagonal) || (is_point_block && !diag->LinearPointBlock)) { 1321 CeedSize assembled_length, assembled_qf_length; 1322 CeedInt use_ceedsize_idx = 0; 1323 CeedCallBackend(CeedVectorGetLength(assembled, &assembled_length)); 1324 CeedCallBackend(CeedVectorGetLength(assembled_qf, &assembled_qf_length)); 1325 if ((assembled_length > INT_MAX) || (assembled_qf_length > INT_MAX)) use_ceedsize_idx = 1; 1326 1327 CeedCallBackend(CeedOperatorAssembleDiagonalSetupCompile_Cuda(op, use_ceedsize_idx, is_point_block)); 1328 } 1329 1330 // Restriction and diagonal vector 1331 CeedCallBackend(CeedOperatorGetActiveElemRestrictions(op, &rstr_in, &rstr_out)); 1332 CeedCheck(rstr_in == rstr_out, ceed, CEED_ERROR_BACKEND, 1333 "Cannot assemble operator diagonal with different input and output active element restrictions"); 1334 if (!is_point_block && !diag->diag_rstr) { 1335 CeedCallBackend(CeedElemRestrictionCreateUnsignedCopy(rstr_out, &diag->diag_rstr)); 1336 CeedCallBackend(CeedElemRestrictionCreateVector(diag->diag_rstr, NULL, &diag->elem_diag)); 1337 } else if (is_point_block && !diag->point_block_diag_rstr) { 1338 CeedCallBackend(CeedOperatorCreateActivePointBlockRestriction(rstr_out, &diag->point_block_diag_rstr)); 1339 CeedCallBackend(CeedElemRestrictionCreateVector(diag->point_block_diag_rstr, NULL, &diag->point_block_elem_diag)); 1340 } 1341 diag_rstr = is_point_block ? diag->point_block_diag_rstr : diag->diag_rstr; 1342 elem_diag = is_point_block ? diag->point_block_elem_diag : diag->elem_diag; 1343 CeedCallBackend(CeedVectorSetValue(elem_diag, 0.0)); 1344 1345 // Only assemble diagonal if the basis has nodes, otherwise inputs are null pointers 1346 CeedCallBackend(CeedElemRestrictionGetElementSize(diag_rstr, &num_nodes)); 1347 if (num_nodes > 0) { 1348 // Assemble element operator diagonals 1349 CeedCallBackend(CeedElemRestrictionGetNumElements(diag_rstr, &num_elem)); 1350 CeedCallBackend(CeedVectorGetArray(elem_diag, CEED_MEM_DEVICE, &elem_diag_array)); 1351 1352 // Compute the diagonal of B^T D B 1353 CeedInt elems_per_block = 1; 1354 CeedInt grid = CeedDivUpInt(num_elem, elems_per_block); 1355 void *args[] = {(void *)&num_elem, &diag->d_identity, &diag->d_interp_in, &diag->d_grad_in, &diag->d_div_in, 1356 &diag->d_curl_in, &diag->d_interp_out, &diag->d_grad_out, &diag->d_div_out, &diag->d_curl_out, 1357 &diag->d_eval_modes_in, &diag->d_eval_modes_out, &assembled_qf_array, &elem_diag_array}; 1358 1359 if (is_point_block) { 1360 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, diag->LinearPointBlock, grid, num_nodes, 1, elems_per_block, args)); 1361 } else { 1362 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, diag->LinearDiagonal, grid, num_nodes, 1, elems_per_block, args)); 1363 } 1364 1365 // Restore arrays 1366 CeedCallBackend(CeedVectorRestoreArray(elem_diag, &elem_diag_array)); 1367 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 1368 } 1369 1370 // Assemble local operator diagonal 1371 CeedCallBackend(CeedElemRestrictionApply(diag_rstr, CEED_TRANSPOSE, elem_diag, assembled, request)); 1372 1373 // Cleanup 1374 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 1375 return CEED_ERROR_SUCCESS; 1376 } 1377 1378 //------------------------------------------------------------------------------ 1379 // Assemble Linear Diagonal 1380 //------------------------------------------------------------------------------ 1381 static int CeedOperatorLinearAssembleAddDiagonal_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request) { 1382 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Cuda(op, assembled, request, false)); 1383 return CEED_ERROR_SUCCESS; 1384 } 1385 1386 //------------------------------------------------------------------------------ 1387 // Assemble Linear Point Block Diagonal 1388 //------------------------------------------------------------------------------ 1389 static int CeedOperatorLinearAssembleAddPointBlockDiagonal_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request) { 1390 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Cuda(op, assembled, request, true)); 1391 return CEED_ERROR_SUCCESS; 1392 } 1393 1394 //------------------------------------------------------------------------------ 1395 // Single Operator Assembly Setup 1396 //------------------------------------------------------------------------------ 1397 static int CeedSingleOperatorAssembleSetup_Cuda(CeedOperator op, CeedInt use_ceedsize_idx) { 1398 Ceed ceed; 1399 Ceed_Cuda *cuda_data; 1400 char *assembly_kernel_source; 1401 const char *assembly_kernel_path; 1402 CeedInt num_input_fields, num_output_fields, num_eval_modes_in = 0, num_eval_modes_out = 0; 1403 CeedInt elem_size_in, num_qpts_in = 0, num_comp_in, elem_size_out, num_qpts_out, num_comp_out, q_comp; 1404 CeedEvalMode *eval_modes_in = NULL, *eval_modes_out = NULL; 1405 CeedElemRestriction rstr_in = NULL, rstr_out = NULL; 1406 CeedBasis basis_in = NULL, basis_out = NULL; 1407 CeedQFunctionField *qf_fields; 1408 CeedQFunction qf; 1409 CeedOperatorField *input_fields, *output_fields; 1410 CeedOperator_Cuda *impl; 1411 1412 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1413 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1414 1415 // Get intput and output fields 1416 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 1417 1418 // Determine active input basis eval mode 1419 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1420 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 1421 for (CeedInt i = 0; i < num_input_fields; i++) { 1422 CeedVector vec; 1423 1424 CeedCallBackend(CeedOperatorFieldGetVector(input_fields[i], &vec)); 1425 if (vec == CEED_VECTOR_ACTIVE) { 1426 CeedBasis basis; 1427 CeedEvalMode eval_mode; 1428 1429 CeedCallBackend(CeedOperatorFieldGetBasis(input_fields[i], &basis)); 1430 CeedCheck(!basis_in || basis_in == basis, ceed, CEED_ERROR_BACKEND, "Backend does not implement operator assembly with multiple active bases"); 1431 basis_in = basis; 1432 CeedCallBackend(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr_in)); 1433 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_in, &elem_size_in)); 1434 if (basis_in == CEED_BASIS_NONE) num_qpts_in = elem_size_in; 1435 else CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts_in)); 1436 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1437 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_in, eval_mode, &q_comp)); 1438 if (eval_mode != CEED_EVAL_WEIGHT) { 1439 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF Assembly 1440 CeedCallBackend(CeedRealloc(num_eval_modes_in + q_comp, &eval_modes_in)); 1441 for (CeedInt d = 0; d < q_comp; d++) { 1442 eval_modes_in[num_eval_modes_in + d] = eval_mode; 1443 } 1444 num_eval_modes_in += q_comp; 1445 } 1446 } 1447 } 1448 1449 // Determine active output basis; basis_out and rstr_out only used if same as input, TODO 1450 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 1451 for (CeedInt i = 0; i < num_output_fields; i++) { 1452 CeedVector vec; 1453 1454 CeedCallBackend(CeedOperatorFieldGetVector(output_fields[i], &vec)); 1455 if (vec == CEED_VECTOR_ACTIVE) { 1456 CeedBasis basis; 1457 CeedEvalMode eval_mode; 1458 1459 CeedCallBackend(CeedOperatorFieldGetBasis(output_fields[i], &basis)); 1460 CeedCheck(!basis_out || basis_out == basis, ceed, CEED_ERROR_BACKEND, 1461 "Backend does not implement operator assembly with multiple active bases"); 1462 basis_out = basis; 1463 CeedCallBackend(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr_out)); 1464 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_out, &elem_size_out)); 1465 if (basis_out == CEED_BASIS_NONE) num_qpts_out = elem_size_out; 1466 else CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_out, &num_qpts_out)); 1467 CeedCheck(num_qpts_in == num_qpts_out, ceed, CEED_ERROR_UNSUPPORTED, 1468 "Active input and output bases must have the same number of quadrature points"); 1469 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 1470 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_out, eval_mode, &q_comp)); 1471 if (eval_mode != CEED_EVAL_WEIGHT) { 1472 // q_comp = 1 if CEED_EVAL_NONE, CEED_EVAL_WEIGHT caught by QF Assembly 1473 CeedCallBackend(CeedRealloc(num_eval_modes_out + q_comp, &eval_modes_out)); 1474 for (CeedInt d = 0; d < q_comp; d++) { 1475 eval_modes_out[num_eval_modes_out + d] = eval_mode; 1476 } 1477 num_eval_modes_out += q_comp; 1478 } 1479 } 1480 } 1481 CeedCheck(num_eval_modes_in > 0 && num_eval_modes_out > 0, ceed, CEED_ERROR_UNSUPPORTED, "Cannot assemble operator without inputs/outputs"); 1482 1483 CeedCallBackend(CeedCalloc(1, &impl->asmb)); 1484 CeedOperatorAssemble_Cuda *asmb = impl->asmb; 1485 asmb->elems_per_block = 1; 1486 asmb->block_size_x = elem_size_in; 1487 asmb->block_size_y = elem_size_out; 1488 1489 CeedCallBackend(CeedGetData(ceed, &cuda_data)); 1490 bool fallback = asmb->block_size_x * asmb->block_size_y * asmb->elems_per_block > cuda_data->device_prop.maxThreadsPerBlock; 1491 1492 if (fallback) { 1493 // Use fallback kernel with 1D threadblock 1494 asmb->block_size_y = 1; 1495 } 1496 1497 // Compile kernels 1498 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr_in, &num_comp_in)); 1499 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr_out, &num_comp_out)); 1500 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-operator-assemble.h", &assembly_kernel_path)); 1501 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Kernel Source -----\n"); 1502 CeedCallBackend(CeedLoadSourceToBuffer(ceed, assembly_kernel_path, &assembly_kernel_source)); 1503 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Source Complete! -----\n"); 1504 CeedCallBackend(CeedCompile_Cuda(ceed, assembly_kernel_source, &asmb->module, 10, "NUM_EVAL_MODES_IN", num_eval_modes_in, "NUM_EVAL_MODES_OUT", 1505 num_eval_modes_out, "NUM_COMP_IN", num_comp_in, "NUM_COMP_OUT", num_comp_out, "NUM_NODES_IN", elem_size_in, 1506 "NUM_NODES_OUT", elem_size_out, "NUM_QPTS", num_qpts_in, "BLOCK_SIZE", 1507 asmb->block_size_x * asmb->block_size_y * asmb->elems_per_block, "BLOCK_SIZE_Y", asmb->block_size_y, 1508 "USE_CEEDSIZE", use_ceedsize_idx)); 1509 CeedCallBackend(CeedGetKernel_Cuda(ceed, asmb->module, "LinearAssemble", &asmb->LinearAssemble)); 1510 CeedCallBackend(CeedFree(&assembly_kernel_path)); 1511 CeedCallBackend(CeedFree(&assembly_kernel_source)); 1512 1513 // Load into B_in, in order that they will be used in eval_modes_in 1514 { 1515 const CeedInt in_bytes = elem_size_in * num_qpts_in * num_eval_modes_in * sizeof(CeedScalar); 1516 CeedInt d_in = 0; 1517 CeedEvalMode eval_modes_in_prev = CEED_EVAL_NONE; 1518 bool has_eval_none = false; 1519 CeedScalar *identity = NULL; 1520 1521 for (CeedInt i = 0; i < num_eval_modes_in; i++) { 1522 has_eval_none = has_eval_none || (eval_modes_in[i] == CEED_EVAL_NONE); 1523 } 1524 if (has_eval_none) { 1525 CeedCallBackend(CeedCalloc(elem_size_in * num_qpts_in, &identity)); 1526 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; 1527 } 1528 1529 CeedCallCuda(ceed, cudaMalloc((void **)&asmb->d_B_in, in_bytes)); 1530 for (CeedInt i = 0; i < num_eval_modes_in; i++) { 1531 const CeedScalar *h_B_in; 1532 1533 CeedCallBackend(CeedOperatorGetBasisPointer(basis_in, eval_modes_in[i], identity, &h_B_in)); 1534 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_in, eval_modes_in[i], &q_comp)); 1535 if (q_comp > 1) { 1536 if (i == 0 || eval_modes_in[i] != eval_modes_in_prev) d_in = 0; 1537 else h_B_in = &h_B_in[(++d_in) * elem_size_in * num_qpts_in]; 1538 } 1539 eval_modes_in_prev = eval_modes_in[i]; 1540 1541 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_in[i * elem_size_in * num_qpts_in], h_B_in, elem_size_in * num_qpts_in * sizeof(CeedScalar), 1542 cudaMemcpyHostToDevice)); 1543 } 1544 CeedCallBackend(CeedFree(&identity)); 1545 } 1546 CeedCallBackend(CeedFree(&eval_modes_in)); 1547 1548 // Load into B_out, in order that they will be used in eval_modes_out 1549 { 1550 const CeedInt out_bytes = elem_size_out * num_qpts_out * num_eval_modes_out * sizeof(CeedScalar); 1551 CeedInt d_out = 0; 1552 CeedEvalMode eval_modes_out_prev = CEED_EVAL_NONE; 1553 bool has_eval_none = false; 1554 CeedScalar *identity = NULL; 1555 1556 for (CeedInt i = 0; i < num_eval_modes_out; i++) { 1557 has_eval_none = has_eval_none || (eval_modes_out[i] == CEED_EVAL_NONE); 1558 } 1559 if (has_eval_none) { 1560 CeedCallBackend(CeedCalloc(elem_size_out * num_qpts_out, &identity)); 1561 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; 1562 } 1563 1564 CeedCallCuda(ceed, cudaMalloc((void **)&asmb->d_B_out, out_bytes)); 1565 for (CeedInt i = 0; i < num_eval_modes_out; i++) { 1566 const CeedScalar *h_B_out; 1567 1568 CeedCallBackend(CeedOperatorGetBasisPointer(basis_out, eval_modes_out[i], identity, &h_B_out)); 1569 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis_out, eval_modes_out[i], &q_comp)); 1570 if (q_comp > 1) { 1571 if (i == 0 || eval_modes_out[i] != eval_modes_out_prev) d_out = 0; 1572 else h_B_out = &h_B_out[(++d_out) * elem_size_out * num_qpts_out]; 1573 } 1574 eval_modes_out_prev = eval_modes_out[i]; 1575 1576 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_out[i * elem_size_out * num_qpts_out], h_B_out, elem_size_out * num_qpts_out * sizeof(CeedScalar), 1577 cudaMemcpyHostToDevice)); 1578 } 1579 CeedCallBackend(CeedFree(&identity)); 1580 } 1581 CeedCallBackend(CeedFree(&eval_modes_out)); 1582 return CEED_ERROR_SUCCESS; 1583 } 1584 1585 //------------------------------------------------------------------------------ 1586 // Assemble matrix data for COO matrix of assembled operator. 1587 // The sparsity pattern is set by CeedOperatorLinearAssembleSymbolic. 1588 // 1589 // Note that this (and other assembly routines) currently assume only one active input restriction/basis per operator 1590 // (could have multiple basis eval modes). 1591 // TODO: allow multiple active input restrictions/basis objects 1592 //------------------------------------------------------------------------------ 1593 static int CeedSingleOperatorAssemble_Cuda(CeedOperator op, CeedInt offset, CeedVector values) { 1594 Ceed ceed; 1595 CeedSize values_length = 0, assembled_qf_length = 0; 1596 CeedInt use_ceedsize_idx = 0, num_elem_in, num_elem_out, elem_size_in, elem_size_out; 1597 CeedScalar *values_array; 1598 const CeedScalar *assembled_qf_array; 1599 CeedVector assembled_qf = NULL; 1600 CeedElemRestriction assembled_rstr = NULL, rstr_in, rstr_out; 1601 CeedRestrictionType rstr_type_in, rstr_type_out; 1602 const bool *orients_in = NULL, *orients_out = NULL; 1603 const CeedInt8 *curl_orients_in = NULL, *curl_orients_out = NULL; 1604 CeedOperator_Cuda *impl; 1605 1606 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1607 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1608 1609 // Assemble QFunction 1610 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &assembled_rstr, CEED_REQUEST_IMMEDIATE)); 1611 CeedCallBackend(CeedElemRestrictionDestroy(&assembled_rstr)); 1612 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &assembled_qf_array)); 1613 1614 CeedCallBackend(CeedVectorGetLength(values, &values_length)); 1615 CeedCallBackend(CeedVectorGetLength(assembled_qf, &assembled_qf_length)); 1616 if ((values_length > INT_MAX) || (assembled_qf_length > INT_MAX)) use_ceedsize_idx = 1; 1617 1618 // Setup 1619 if (!impl->asmb) CeedCallBackend(CeedSingleOperatorAssembleSetup_Cuda(op, use_ceedsize_idx)); 1620 CeedOperatorAssemble_Cuda *asmb = impl->asmb; 1621 1622 assert(asmb != NULL); 1623 1624 // Assemble element operator 1625 CeedCallBackend(CeedVectorGetArray(values, CEED_MEM_DEVICE, &values_array)); 1626 values_array += offset; 1627 1628 CeedCallBackend(CeedOperatorGetActiveElemRestrictions(op, &rstr_in, &rstr_out)); 1629 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr_in, &num_elem_in)); 1630 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_in, &elem_size_in)); 1631 1632 CeedCallBackend(CeedElemRestrictionGetType(rstr_in, &rstr_type_in)); 1633 if (rstr_type_in == CEED_RESTRICTION_ORIENTED) { 1634 CeedCallBackend(CeedElemRestrictionGetOrientations(rstr_in, CEED_MEM_DEVICE, &orients_in)); 1635 } else if (rstr_type_in == CEED_RESTRICTION_CURL_ORIENTED) { 1636 CeedCallBackend(CeedElemRestrictionGetCurlOrientations(rstr_in, CEED_MEM_DEVICE, &curl_orients_in)); 1637 } 1638 1639 if (rstr_in != rstr_out) { 1640 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr_out, &num_elem_out)); 1641 CeedCheck(num_elem_in == num_elem_out, ceed, CEED_ERROR_UNSUPPORTED, 1642 "Active input and output operator restrictions must have the same number of elements"); 1643 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_out, &elem_size_out)); 1644 1645 CeedCallBackend(CeedElemRestrictionGetType(rstr_out, &rstr_type_out)); 1646 if (rstr_type_out == CEED_RESTRICTION_ORIENTED) { 1647 CeedCallBackend(CeedElemRestrictionGetOrientations(rstr_out, CEED_MEM_DEVICE, &orients_out)); 1648 } else if (rstr_type_out == CEED_RESTRICTION_CURL_ORIENTED) { 1649 CeedCallBackend(CeedElemRestrictionGetCurlOrientations(rstr_out, CEED_MEM_DEVICE, &curl_orients_out)); 1650 } 1651 } else { 1652 elem_size_out = elem_size_in; 1653 orients_out = orients_in; 1654 curl_orients_out = curl_orients_in; 1655 } 1656 1657 // Compute B^T D B 1658 CeedInt shared_mem = 1659 ((curl_orients_in || curl_orients_out ? elem_size_in * elem_size_out : 0) + (curl_orients_in ? elem_size_in * asmb->block_size_y : 0)) * 1660 sizeof(CeedScalar); 1661 CeedInt grid = CeedDivUpInt(num_elem_in, asmb->elems_per_block); 1662 void *args[] = {(void *)&num_elem_in, &asmb->d_B_in, &asmb->d_B_out, &orients_in, &curl_orients_in, 1663 &orients_out, &curl_orients_out, &assembled_qf_array, &values_array}; 1664 1665 CeedCallBackend( 1666 CeedRunKernelDimShared_Cuda(ceed, asmb->LinearAssemble, grid, asmb->block_size_x, asmb->block_size_y, asmb->elems_per_block, shared_mem, args)); 1667 1668 // Restore arrays 1669 CeedCallBackend(CeedVectorRestoreArray(values, &values_array)); 1670 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 1671 1672 // Cleanup 1673 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 1674 if (rstr_type_in == CEED_RESTRICTION_ORIENTED) { 1675 CeedCallBackend(CeedElemRestrictionRestoreOrientations(rstr_in, &orients_in)); 1676 } else if (rstr_type_in == CEED_RESTRICTION_CURL_ORIENTED) { 1677 CeedCallBackend(CeedElemRestrictionRestoreCurlOrientations(rstr_in, &curl_orients_in)); 1678 } 1679 if (rstr_in != rstr_out) { 1680 if (rstr_type_out == CEED_RESTRICTION_ORIENTED) { 1681 CeedCallBackend(CeedElemRestrictionRestoreOrientations(rstr_out, &orients_out)); 1682 } else if (rstr_type_out == CEED_RESTRICTION_CURL_ORIENTED) { 1683 CeedCallBackend(CeedElemRestrictionRestoreCurlOrientations(rstr_out, &curl_orients_out)); 1684 } 1685 } 1686 return CEED_ERROR_SUCCESS; 1687 } 1688 1689 //------------------------------------------------------------------------------ 1690 // Assemble Linear QFunction AtPoints 1691 //------------------------------------------------------------------------------ 1692 static int CeedOperatorLinearAssembleQFunctionAtPoints_Cuda(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 1693 return CeedError(CeedOperatorReturnCeed(op), CEED_ERROR_BACKEND, "Backend does not implement CeedOperatorLinearAssembleQFunction"); 1694 } 1695 1696 //------------------------------------------------------------------------------ 1697 // Assemble Linear Diagonal AtPoints 1698 //------------------------------------------------------------------------------ 1699 static int CeedOperatorLinearAssembleAddDiagonalAtPoints_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request) { 1700 CeedInt max_num_points, *num_points, num_elem, num_input_fields, num_output_fields; 1701 Ceed ceed; 1702 CeedVector active_e_vec_in, active_e_vec_out; 1703 CeedQFunctionField *qf_input_fields, *qf_output_fields; 1704 CeedQFunction qf; 1705 CeedOperatorField *op_input_fields, *op_output_fields; 1706 CeedOperator_Cuda *impl; 1707 1708 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1709 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1710 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1711 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 1712 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 1713 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 1714 1715 // Setup 1716 CeedCallBackend(CeedOperatorSetupAtPoints_Cuda(op)); 1717 num_points = impl->num_points; 1718 max_num_points = impl->max_num_points; 1719 1720 // Work vector 1721 CeedCallBackend(CeedGetWorkVector(ceed, impl->max_active_e_vec_len, &active_e_vec_in)); 1722 CeedCallBackend(CeedGetWorkVector(ceed, impl->max_active_e_vec_len, &active_e_vec_out)); 1723 { 1724 CeedSize length_in, length_out; 1725 1726 CeedCallBackend(CeedVectorGetLength(active_e_vec_in, &length_in)); 1727 CeedCallBackend(CeedVectorGetLength(active_e_vec_out, &length_out)); 1728 // Need input e_vec to be longer 1729 if (length_in < length_out) { 1730 CeedVector temp = active_e_vec_in; 1731 1732 active_e_vec_in = active_e_vec_out; 1733 active_e_vec_out = temp; 1734 } 1735 } 1736 1737 // Get point coordinates 1738 if (!impl->point_coords_elem) { 1739 CeedVector point_coords = NULL; 1740 CeedElemRestriction rstr_points = NULL; 1741 1742 CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, &point_coords)); 1743 CeedCallBackend(CeedElemRestrictionCreateVector(rstr_points, NULL, &impl->point_coords_elem)); 1744 CeedCallBackend(CeedElemRestrictionApply(rstr_points, CEED_NOTRANSPOSE, point_coords, impl->point_coords_elem, request)); 1745 CeedCallBackend(CeedVectorDestroy(&point_coords)); 1746 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_points)); 1747 } 1748 1749 // Process inputs 1750 for (CeedInt i = 0; i < num_input_fields; i++) { 1751 CeedCallBackend(CeedOperatorInputRestrict_Cuda(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, true, impl, request)); 1752 CeedCallBackend(CeedOperatorInputBasisAtPoints_Cuda(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, num_elem, num_points, true, impl)); 1753 } 1754 1755 // Clear active input Qvecs 1756 for (CeedInt i = 0; i < num_input_fields; i++) { 1757 CeedVector vec; 1758 1759 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 1760 if (vec != CEED_VECTOR_ACTIVE) continue; 1761 CeedCallBackend(CeedVectorSetValue(impl->q_vecs_in[i], 0.0)); 1762 } 1763 1764 // Output pointers, as necessary 1765 for (CeedInt i = 0; i < num_output_fields; i++) { 1766 CeedEvalMode eval_mode; 1767 1768 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1769 if (eval_mode == CEED_EVAL_NONE) { 1770 CeedScalar *e_vec_array; 1771 1772 CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs_out[i], CEED_MEM_DEVICE, &e_vec_array)); 1773 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_vec_array)); 1774 } 1775 } 1776 1777 // Loop over active fields 1778 for (CeedInt i = 0; i < num_input_fields; i++) { 1779 bool is_active_at_points = true; 1780 CeedInt elem_size = 1, num_comp_active = 1, e_vec_size = 0; 1781 CeedRestrictionType rstr_type; 1782 CeedVector l_vec; 1783 CeedElemRestriction elem_rstr; 1784 1785 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &l_vec)); 1786 // -- Skip non-active input 1787 if (l_vec != CEED_VECTOR_ACTIVE) continue; 1788 1789 // -- Get active restriction type 1790 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_rstr)); 1791 CeedCallBackend(CeedElemRestrictionGetType(elem_rstr, &rstr_type)); 1792 is_active_at_points = rstr_type == CEED_RESTRICTION_POINTS; 1793 if (!is_active_at_points) CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 1794 else elem_size = max_num_points; 1795 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp_active)); 1796 1797 e_vec_size = elem_size * num_comp_active; 1798 for (CeedInt s = 0; s < e_vec_size; s++) { 1799 bool is_active_input = false; 1800 CeedEvalMode eval_mode; 1801 CeedVector l_vec, q_vec = impl->q_vecs_in[i]; 1802 CeedBasis basis; 1803 1804 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &l_vec)); 1805 // Skip non-active input 1806 is_active_input = l_vec == CEED_VECTOR_ACTIVE; 1807 if (!is_active_input) continue; 1808 1809 // Update unit vector 1810 if (s == 0) CeedCallBackend(CeedVectorSetValue(active_e_vec_in, 0.0)); 1811 else CeedCallBackend(CeedVectorSetValueStrided(active_e_vec_in, s - 1, e_vec_size, 0.0)); 1812 CeedCallBackend(CeedVectorSetValueStrided(active_e_vec_in, s, e_vec_size, 1.0)); 1813 1814 // Basis action 1815 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1816 switch (eval_mode) { 1817 case CEED_EVAL_NONE: { 1818 const CeedScalar *e_vec_array; 1819 1820 CeedCallBackend(CeedVectorGetArrayRead(active_e_vec_in, CEED_MEM_DEVICE, &e_vec_array)); 1821 CeedCallBackend(CeedVectorSetArray(q_vec, CEED_MEM_DEVICE, CEED_USE_POINTER, (CeedScalar *)e_vec_array)); 1822 break; 1823 } 1824 case CEED_EVAL_INTERP: 1825 case CEED_EVAL_GRAD: 1826 case CEED_EVAL_DIV: 1827 case CEED_EVAL_CURL: 1828 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 1829 CeedCallBackend( 1830 CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_NOTRANSPOSE, eval_mode, impl->point_coords_elem, active_e_vec_in, q_vec)); 1831 break; 1832 case CEED_EVAL_WEIGHT: 1833 break; // No action 1834 } 1835 1836 // Q function 1837 CeedCallBackend(CeedQFunctionApply(qf, num_elem * max_num_points, impl->q_vecs_in, impl->q_vecs_out)); 1838 1839 // Output basis apply if needed 1840 for (CeedInt j = 0; j < num_output_fields; j++) { 1841 bool is_active_output = false; 1842 CeedInt elem_size = 0; 1843 CeedRestrictionType rstr_type; 1844 CeedEvalMode eval_mode; 1845 CeedVector l_vec, e_vec = impl->e_vecs_out[j], q_vec = impl->q_vecs_out[j]; 1846 CeedElemRestriction elem_rstr; 1847 CeedBasis basis; 1848 1849 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[j], &l_vec)); 1850 // ---- Skip non-active output 1851 is_active_output = l_vec == CEED_VECTOR_ACTIVE; 1852 if (!is_active_output) continue; 1853 if (!e_vec) e_vec = active_e_vec_out; 1854 1855 // ---- Check if elem size matches 1856 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[j], &elem_rstr)); 1857 CeedCallBackend(CeedElemRestrictionGetType(elem_rstr, &rstr_type)); 1858 if (is_active_at_points && rstr_type != CEED_RESTRICTION_POINTS) continue; 1859 if (rstr_type == CEED_RESTRICTION_POINTS) { 1860 CeedCallBackend(CeedElemRestrictionGetMaxPointsInElement(elem_rstr, &elem_size)); 1861 } else { 1862 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 1863 } 1864 { 1865 CeedInt num_comp = 0; 1866 1867 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 1868 if (e_vec_size != num_comp * elem_size) continue; 1869 } 1870 1871 // Basis action 1872 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[j], &eval_mode)); 1873 switch (eval_mode) { 1874 case CEED_EVAL_NONE: { 1875 CeedScalar *e_vec_array; 1876 1877 CeedCallBackend(CeedVectorTakeArray(q_vec, CEED_MEM_DEVICE, &e_vec_array)); 1878 CeedCallBackend(CeedVectorRestoreArray(e_vec, &e_vec_array)); 1879 break; 1880 } 1881 case CEED_EVAL_INTERP: 1882 case CEED_EVAL_GRAD: 1883 case CEED_EVAL_DIV: 1884 case CEED_EVAL_CURL: 1885 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[j], &basis)); 1886 CeedCallBackend(CeedBasisApplyAtPoints(basis, num_elem, num_points, CEED_TRANSPOSE, eval_mode, impl->point_coords_elem, q_vec, e_vec)); 1887 break; 1888 // LCOV_EXCL_START 1889 case CEED_EVAL_WEIGHT: { 1890 return CeedError(CeedOperatorReturnCeed(op), CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 1891 // LCOV_EXCL_STOP 1892 } 1893 } 1894 1895 // Mask output e-vec 1896 CeedCallBackend(CeedVectorPointwiseMult(e_vec, active_e_vec_in, e_vec)); 1897 1898 // Restrict 1899 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[j], &elem_rstr)); 1900 CeedCallBackend(CeedElemRestrictionApply(elem_rstr, CEED_TRANSPOSE, e_vec, assembled, request)); 1901 1902 // Reset q_vec for 1903 if (eval_mode == CEED_EVAL_NONE) { 1904 CeedScalar *e_vec_array; 1905 1906 CeedCallBackend(CeedVectorGetArrayWrite(e_vec, CEED_MEM_DEVICE, &e_vec_array)); 1907 CeedCallBackend(CeedVectorSetArray(q_vec, CEED_MEM_DEVICE, CEED_USE_POINTER, e_vec_array)); 1908 } 1909 } 1910 1911 // Reset vec 1912 if (s == e_vec_size - 1 && i != num_input_fields - 1) CeedCallBackend(CeedVectorSetValue(q_vec, 0.0)); 1913 } 1914 } 1915 1916 // Restore CEED_EVAL_NONE 1917 for (CeedInt i = 0; i < num_output_fields; i++) { 1918 CeedEvalMode eval_mode; 1919 1920 // Get eval_mode 1921 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1922 1923 // Restore evec 1924 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1925 if (eval_mode == CEED_EVAL_NONE) { 1926 CeedScalar *e_vec_array; 1927 1928 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, &e_vec_array)); 1929 CeedCallBackend(CeedVectorRestoreArray(impl->e_vecs_in[i], &e_vec_array)); 1930 } 1931 } 1932 1933 // Restore input arrays 1934 for (CeedInt i = 0; i < num_input_fields; i++) { 1935 CeedCallBackend(CeedOperatorInputRestore_Cuda(op_input_fields[i], qf_input_fields[i], i, NULL, NULL, true, impl)); 1936 } 1937 1938 // Restore work vector 1939 CeedCallBackend(CeedRestoreWorkVector(ceed, &active_e_vec_in)); 1940 CeedCallBackend(CeedRestoreWorkVector(ceed, &active_e_vec_out)); 1941 return CEED_ERROR_SUCCESS; 1942 } 1943 1944 //------------------------------------------------------------------------------ 1945 // Create operator 1946 //------------------------------------------------------------------------------ 1947 int CeedOperatorCreate_Cuda(CeedOperator op) { 1948 Ceed ceed; 1949 CeedOperator_Cuda *impl; 1950 1951 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1952 CeedCallBackend(CeedCalloc(1, &impl)); 1953 CeedCallBackend(CeedOperatorSetData(op, impl)); 1954 1955 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunction_Cuda)); 1956 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunctionUpdate", CeedOperatorLinearAssembleQFunctionUpdate_Cuda)); 1957 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddDiagonal", CeedOperatorLinearAssembleAddDiagonal_Cuda)); 1958 CeedCallBackend( 1959 CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddPointBlockDiagonal", CeedOperatorLinearAssembleAddPointBlockDiagonal_Cuda)); 1960 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleSingle", CeedSingleOperatorAssemble_Cuda)); 1961 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Cuda)); 1962 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Cuda)); 1963 return CEED_ERROR_SUCCESS; 1964 } 1965 1966 //------------------------------------------------------------------------------ 1967 // Create operator AtPoints 1968 //------------------------------------------------------------------------------ 1969 int CeedOperatorCreateAtPoints_Cuda(CeedOperator op) { 1970 Ceed ceed; 1971 CeedOperator_Cuda *impl; 1972 1973 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1974 CeedCallBackend(CeedCalloc(1, &impl)); 1975 CeedCallBackend(CeedOperatorSetData(op, impl)); 1976 1977 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunctionAtPoints_Cuda)); 1978 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddDiagonal", CeedOperatorLinearAssembleAddDiagonalAtPoints_Cuda)); 1979 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAddAtPoints_Cuda)); 1980 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Cuda)); 1981 return CEED_ERROR_SUCCESS; 1982 } 1983 1984 //------------------------------------------------------------------------------ 1985