1 // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3 // 4 // SPDX-License-Identifier: BSD-2-Clause 5 // 6 // This file is part of CEED: http://github.com/ceed 7 8 #include <ceed.h> 9 #include <ceed/backend.h> 10 #include <ceed/jit-tools.h> 11 #include <assert.h> 12 #include <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 for (CeedInt i = 0; i < impl->num_inputs + impl->num_outputs; i++) { 31 CeedCallBackend(CeedVectorDestroy(&impl->e_vecs[i])); 32 } 33 CeedCallBackend(CeedFree(&impl->e_vecs)); 34 35 for (CeedInt i = 0; i < impl->num_inputs; i++) { 36 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_in[i])); 37 } 38 CeedCallBackend(CeedFree(&impl->q_vecs_in)); 39 40 for (CeedInt i = 0; i < impl->num_outputs; i++) { 41 CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_out[i])); 42 } 43 CeedCallBackend(CeedFree(&impl->q_vecs_out)); 44 45 // QFunction assembly data 46 for (CeedInt i = 0; i < impl->num_active_in; i++) { 47 CeedCallBackend(CeedVectorDestroy(&impl->qf_active_in[i])); 48 } 49 CeedCallBackend(CeedFree(&impl->qf_active_in)); 50 51 // Diag data 52 if (impl->diag) { 53 Ceed ceed; 54 55 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 56 CeedCallCuda(ceed, cuModuleUnload(impl->diag->module)); 57 CeedCallBackend(CeedFree(&impl->diag->h_e_mode_in)); 58 CeedCallBackend(CeedFree(&impl->diag->h_e_mode_out)); 59 CeedCallCuda(ceed, cudaFree(impl->diag->d_e_mode_in)); 60 CeedCallCuda(ceed, cudaFree(impl->diag->d_e_mode_out)); 61 CeedCallCuda(ceed, cudaFree(impl->diag->d_identity)); 62 CeedCallCuda(ceed, cudaFree(impl->diag->d_interp_in)); 63 CeedCallCuda(ceed, cudaFree(impl->diag->d_interp_out)); 64 CeedCallCuda(ceed, cudaFree(impl->diag->d_grad_in)); 65 CeedCallCuda(ceed, cudaFree(impl->diag->d_grad_out)); 66 CeedCallBackend(CeedElemRestrictionDestroy(&impl->diag->point_block_rstr)); 67 CeedCallBackend(CeedVectorDestroy(&impl->diag->elem_diag)); 68 CeedCallBackend(CeedVectorDestroy(&impl->diag->point_block_elem_diag)); 69 } 70 CeedCallBackend(CeedFree(&impl->diag)); 71 72 if (impl->asmb) { 73 Ceed ceed; 74 75 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 76 CeedCallCuda(ceed, cuModuleUnload(impl->asmb->module)); 77 CeedCallCuda(ceed, cudaFree(impl->asmb->d_B_in)); 78 CeedCallCuda(ceed, cudaFree(impl->asmb->d_B_out)); 79 } 80 CeedCallBackend(CeedFree(&impl->asmb)); 81 82 CeedCallBackend(CeedFree(&impl)); 83 return CEED_ERROR_SUCCESS; 84 } 85 86 //------------------------------------------------------------------------------ 87 // Setup infields or outfields 88 //------------------------------------------------------------------------------ 89 static int CeedOperatorSetupFields_Cuda(CeedQFunction qf, CeedOperator op, bool is_input, CeedVector *e_vecs, CeedVector *q_vecs, CeedInt e_start, 90 CeedInt num_fields, CeedInt Q, CeedInt num_elem) { 91 Ceed ceed; 92 bool is_strided, skip_restriction; 93 CeedSize q_size; 94 CeedInt dim, size; 95 CeedQFunctionField *qf_fields; 96 CeedOperatorField *op_fields; 97 98 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 99 100 if (is_input) { 101 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 102 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 103 } else { 104 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 105 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 106 } 107 108 // Loop over fields 109 for (CeedInt i = 0; i < num_fields; i++) { 110 CeedEvalMode e_mode; 111 CeedBasis basis; 112 113 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &e_mode)); 114 115 is_strided = false; 116 skip_restriction = false; 117 if (e_mode != CEED_EVAL_WEIGHT) { 118 CeedElemRestriction elem_restr; 119 120 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &elem_restr)); 121 122 // Check whether this field can skip the element restriction: 123 // must be passive input, with e_mode NONE, and have a strided restriction with CEED_STRIDES_BACKEND. 124 125 // First, check whether the field is input or output: 126 if (is_input) { 127 CeedVector vec; 128 129 // Check for passive input: 130 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 131 if (vec != CEED_VECTOR_ACTIVE) { 132 // Check e_mode 133 if (e_mode == CEED_EVAL_NONE) { 134 // Check for strided restriction 135 CeedCallBackend(CeedElemRestrictionIsStrided(elem_restr, &is_strided)); 136 if (is_strided) { 137 // Check if vector is already in preferred backend ordering 138 CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_restr, &skip_restriction)); 139 } 140 } 141 } 142 } 143 if (skip_restriction) { 144 // We do not need an E-Vector, but will use the input field vector's data directly in the operator application. 145 e_vecs[i + e_start] = NULL; 146 } else { 147 CeedCallBackend(CeedElemRestrictionCreateVector(elem_restr, NULL, &e_vecs[i + e_start])); 148 } 149 } 150 151 switch (e_mode) { 152 case CEED_EVAL_NONE: 153 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 154 q_size = (CeedSize)num_elem * Q * size; 155 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 156 break; 157 case CEED_EVAL_INTERP: 158 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 159 q_size = (CeedSize)num_elem * Q * size; 160 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 161 break; 162 case CEED_EVAL_GRAD: 163 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 164 CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size)); 165 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 166 q_size = (CeedSize)num_elem * Q * size; 167 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 168 break; 169 case CEED_EVAL_WEIGHT: // Only on input fields 170 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis)); 171 q_size = (CeedSize)num_elem * Q; 172 CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i])); 173 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, q_vecs[i])); 174 break; 175 case CEED_EVAL_DIV: 176 break; // TODO: Not implemented 177 case CEED_EVAL_CURL: 178 break; // TODO: Not implemented 179 } 180 } 181 return CEED_ERROR_SUCCESS; 182 } 183 184 //------------------------------------------------------------------------------ 185 // CeedOperator needs to connect all the named fields (be they active or passive) to the named inputs and outputs of its CeedQFunction. 186 //------------------------------------------------------------------------------ 187 static int CeedOperatorSetup_Cuda(CeedOperator op) { 188 Ceed ceed; 189 bool is_setup_done; 190 CeedInt Q, num_elem, num_input_fields, num_output_fields; 191 CeedQFunctionField *qf_input_fields, *qf_output_fields; 192 CeedQFunction qf; 193 CeedOperatorField *op_input_fields, *op_output_fields; 194 CeedOperator_Cuda *impl; 195 196 CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done)); 197 if (is_setup_done) return CEED_ERROR_SUCCESS; 198 199 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 200 CeedCallBackend(CeedOperatorGetData(op, &impl)); 201 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 202 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 203 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 204 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 205 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 206 207 // Allocate 208 CeedCallBackend(CeedCalloc(num_input_fields + num_output_fields, &impl->e_vecs)); 209 210 CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_in)); 211 CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_out)); 212 213 impl->num_inputs = num_input_fields; 214 impl->num_outputs = num_output_fields; 215 216 // Set up infield and outfield e_vecs and q_vecs 217 // Infields 218 CeedCallBackend(CeedOperatorSetupFields_Cuda(qf, op, true, impl->e_vecs, impl->q_vecs_in, 0, num_input_fields, Q, num_elem)); 219 // Outfields 220 CeedCallBackend(CeedOperatorSetupFields_Cuda(qf, op, false, impl->e_vecs, impl->q_vecs_out, num_input_fields, num_output_fields, Q, num_elem)); 221 222 CeedCallBackend(CeedOperatorSetSetupDone(op)); 223 return CEED_ERROR_SUCCESS; 224 } 225 226 //------------------------------------------------------------------------------ 227 // Setup Operator Inputs 228 //------------------------------------------------------------------------------ 229 static inline int CeedOperatorSetupInputs_Cuda(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields, 230 CeedVector in_vec, const bool skip_active_in, CeedScalar *e_data[2 * CEED_FIELD_MAX], 231 CeedOperator_Cuda *impl, CeedRequest *request) { 232 for (CeedInt i = 0; i < num_input_fields; i++) { 233 CeedEvalMode e_mode; 234 CeedVector vec; 235 CeedElemRestriction elem_restr; 236 237 // Get input vector 238 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 239 if (vec == CEED_VECTOR_ACTIVE) { 240 if (skip_active_in) continue; 241 else vec = in_vec; 242 } 243 244 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &e_mode)); 245 if (e_mode == CEED_EVAL_WEIGHT) { // Skip 246 } else { 247 // Get input element restriction 248 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_restr)); 249 if (vec == CEED_VECTOR_ACTIVE) vec = in_vec; 250 // Restrict, if necessary 251 if (!impl->e_vecs[i]) { 252 // No restriction for this field; read data directly from vec. 253 CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, (const CeedScalar **)&e_data[i])); 254 } else { 255 CeedCallBackend(CeedElemRestrictionApply(elem_restr, CEED_NOTRANSPOSE, vec, impl->e_vecs[i], request)); 256 // Get evec 257 CeedCallBackend(CeedVectorGetArrayRead(impl->e_vecs[i], CEED_MEM_DEVICE, (const CeedScalar **)&e_data[i])); 258 } 259 } 260 } 261 return CEED_ERROR_SUCCESS; 262 } 263 264 //------------------------------------------------------------------------------ 265 // Input Basis Action 266 //------------------------------------------------------------------------------ 267 static inline int CeedOperatorInputBasis_Cuda(CeedInt num_elem, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields, 268 CeedInt num_input_fields, const bool skip_active_in, CeedScalar *e_data[2 * CEED_FIELD_MAX], 269 CeedOperator_Cuda *impl) { 270 for (CeedInt i = 0; i < num_input_fields; i++) { 271 CeedInt elem_size, size; 272 CeedEvalMode e_mode; 273 CeedElemRestriction elem_restr; 274 CeedBasis basis; 275 276 // Skip active input 277 if (skip_active_in) { 278 CeedVector vec; 279 280 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 281 if (vec == CEED_VECTOR_ACTIVE) continue; 282 } 283 // Get elem_size, e_mode, size 284 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_restr)); 285 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_restr, &elem_size)); 286 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &e_mode)); 287 CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size)); 288 // Basis action 289 switch (e_mode) { 290 case CEED_EVAL_NONE: 291 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_data[i])); 292 break; 293 case CEED_EVAL_INTERP: 294 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 295 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_INTERP, impl->e_vecs[i], impl->q_vecs_in[i])); 296 break; 297 case CEED_EVAL_GRAD: 298 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 299 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_NOTRANSPOSE, CEED_EVAL_GRAD, impl->e_vecs[i], impl->q_vecs_in[i])); 300 break; 301 case CEED_EVAL_WEIGHT: 302 break; // No action 303 case CEED_EVAL_DIV: 304 break; // TODO: Not implemented 305 case CEED_EVAL_CURL: 306 break; // TODO: Not implemented 307 } 308 } 309 return CEED_ERROR_SUCCESS; 310 } 311 312 //------------------------------------------------------------------------------ 313 // Restore Input Vectors 314 //------------------------------------------------------------------------------ 315 static inline int CeedOperatorRestoreInputs_Cuda(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields, 316 const bool skip_active_in, CeedScalar *e_data[2 * CEED_FIELD_MAX], CeedOperator_Cuda *impl) { 317 for (CeedInt i = 0; i < num_input_fields; i++) { 318 CeedEvalMode e_mode; 319 CeedVector vec; 320 321 // Skip active input 322 if (skip_active_in) { 323 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 324 if (vec == CEED_VECTOR_ACTIVE) continue; 325 } 326 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &e_mode)); 327 if (e_mode == CEED_EVAL_WEIGHT) { // Skip 328 } else { 329 if (!impl->e_vecs[i]) { // This was a skip_restriction case 330 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 331 CeedCallBackend(CeedVectorRestoreArrayRead(vec, (const CeedScalar **)&e_data[i])); 332 } else { 333 CeedCallBackend(CeedVectorRestoreArrayRead(impl->e_vecs[i], (const CeedScalar **)&e_data[i])); 334 } 335 } 336 } 337 return CEED_ERROR_SUCCESS; 338 } 339 340 //------------------------------------------------------------------------------ 341 // Apply and add to output 342 //------------------------------------------------------------------------------ 343 static int CeedOperatorApplyAdd_Cuda(CeedOperator op, CeedVector in_vec, CeedVector out_vec, CeedRequest *request) { 344 CeedOperator_Cuda *impl; 345 CeedInt Q, num_elem, elem_size, num_input_fields, num_output_fields, size; 346 CeedEvalMode e_mode; 347 CeedScalar *e_data[2 * CEED_FIELD_MAX] = {NULL}; 348 CeedOperatorField *op_input_fields, *op_output_fields; 349 CeedQFunctionField *qf_input_fields, *qf_output_fields; 350 CeedQFunction qf; 351 352 CeedCallBackend(CeedOperatorGetData(op, &impl)); 353 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 354 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 355 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 356 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 357 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 358 359 // Setup 360 CeedCallBackend(CeedOperatorSetup_Cuda(op)); 361 362 // Input e_vecs and Restriction 363 CeedCallBackend(CeedOperatorSetupInputs_Cuda(num_input_fields, qf_input_fields, op_input_fields, in_vec, false, e_data, impl, request)); 364 365 // Input basis apply if needed 366 CeedCallBackend(CeedOperatorInputBasis_Cuda(num_elem, qf_input_fields, op_input_fields, num_input_fields, false, e_data, impl)); 367 368 // Output pointers, as necessary 369 for (CeedInt i = 0; i < num_output_fields; i++) { 370 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &e_mode)); 371 if (e_mode == CEED_EVAL_NONE) { 372 // Set the output Q-Vector to use the E-Vector data directly. 373 CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs[i + impl->num_inputs], CEED_MEM_DEVICE, &e_data[i + num_input_fields])); 374 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_DEVICE, CEED_USE_POINTER, e_data[i + num_input_fields])); 375 } 376 } 377 378 // Q function 379 CeedCallBackend(CeedQFunctionApply(qf, num_elem * Q, impl->q_vecs_in, impl->q_vecs_out)); 380 381 // Output basis apply if needed 382 for (CeedInt i = 0; i < num_output_fields; i++) { 383 CeedElemRestriction elem_restr; 384 CeedBasis basis; 385 386 // Get elem_size, e_mode, size 387 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_restr)); 388 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_restr, &elem_size)); 389 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &e_mode)); 390 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size)); 391 // Basis action 392 switch (e_mode) { 393 case CEED_EVAL_NONE: 394 break; 395 case CEED_EVAL_INTERP: 396 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 397 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_TRANSPOSE, CEED_EVAL_INTERP, impl->q_vecs_out[i], impl->e_vecs[i + impl->num_inputs])); 398 break; 399 case CEED_EVAL_GRAD: 400 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 401 CeedCallBackend(CeedBasisApply(basis, num_elem, CEED_TRANSPOSE, CEED_EVAL_GRAD, impl->q_vecs_out[i], impl->e_vecs[i + impl->num_inputs])); 402 break; 403 // LCOV_EXCL_START 404 case CEED_EVAL_WEIGHT: { 405 Ceed ceed; 406 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 407 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode"); 408 break; // Should not occur 409 } 410 case CEED_EVAL_DIV: 411 break; // TODO: Not implemented 412 case CEED_EVAL_CURL: 413 break; // TODO: Not implemented 414 // LCOV_EXCL_STOP 415 } 416 } 417 418 // Output restriction 419 for (CeedInt i = 0; i < num_output_fields; i++) { 420 CeedVector vec; 421 CeedElemRestriction elem_restr; 422 423 // Restore evec 424 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &e_mode)); 425 if (e_mode == CEED_EVAL_NONE) { 426 CeedCallBackend(CeedVectorRestoreArray(impl->e_vecs[i + impl->num_inputs], &e_data[i + num_input_fields])); 427 } 428 // Get output vector 429 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 430 // Restrict 431 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_restr)); 432 // Active 433 if (vec == CEED_VECTOR_ACTIVE) vec = out_vec; 434 435 CeedCallBackend(CeedElemRestrictionApply(elem_restr, CEED_TRANSPOSE, impl->e_vecs[i + impl->num_inputs], vec, request)); 436 } 437 438 // Restore input arrays 439 CeedCallBackend(CeedOperatorRestoreInputs_Cuda(num_input_fields, qf_input_fields, op_input_fields, false, e_data, impl)); 440 return CEED_ERROR_SUCCESS; 441 } 442 443 //------------------------------------------------------------------------------ 444 // Core code for assembling linear QFunction 445 //------------------------------------------------------------------------------ 446 static inline int CeedOperatorLinearAssembleQFunctionCore_Cuda(CeedOperator op, bool build_objects, CeedVector *assembled, CeedElemRestriction *rstr, 447 CeedRequest *request) { 448 Ceed ceed, ceed_parent; 449 bool is_identity_qf; 450 CeedInt num_active_in, num_active_out, Q, num_elem, num_input_fields, num_output_fields, size; 451 CeedSize q_size; 452 CeedScalar *assembled_array, *e_data[2 * CEED_FIELD_MAX] = {NULL}; 453 CeedVector *active_inputs; 454 CeedQFunctionField *qf_input_fields, *qf_output_fields; 455 CeedQFunction qf; 456 CeedOperatorField *op_input_fields, *op_output_fields; 457 CeedOperator_Cuda *impl; 458 459 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 460 CeedCallBackend(CeedOperatorGetFallbackParentCeed(op, &ceed_parent)); 461 CeedCallBackend(CeedOperatorGetData(op, &impl)); 462 CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q)); 463 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 464 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 465 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 466 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 467 active_inputs = impl->qf_active_in; 468 num_active_in = impl->num_active_in, num_active_out = impl->num_active_out; 469 470 // Setup 471 CeedCallBackend(CeedOperatorSetup_Cuda(op)); 472 473 // Check for identity 474 CeedCallBackend(CeedQFunctionIsIdentity(qf, &is_identity_qf)); 475 CeedCheck(!is_identity_qf, ceed, CEED_ERROR_BACKEND, "Assembling identity QFunctions not supported"); 476 477 // Input e_vecs and Restriction 478 CeedCallBackend(CeedOperatorSetupInputs_Cuda(num_input_fields, qf_input_fields, op_input_fields, NULL, true, e_data, impl, request)); 479 480 // Count number of active input fields 481 if (!num_active_in) { 482 for (CeedInt i = 0; i < num_input_fields; i++) { 483 CeedScalar *q_vec_array; 484 CeedVector vec; 485 486 // Get input vector 487 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 488 // Check if active input 489 if (vec == CEED_VECTOR_ACTIVE) { 490 CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size)); 491 CeedCallBackend(CeedVectorSetValue(impl->q_vecs_in[i], 0.0)); 492 CeedCallBackend(CeedVectorGetArray(impl->q_vecs_in[i], CEED_MEM_DEVICE, &q_vec_array)); 493 CeedCallBackend(CeedRealloc(num_active_in + size, &active_inputs)); 494 for (CeedInt field = 0; field < size; field++) { 495 q_size = (CeedSize)Q * num_elem; 496 CeedCallBackend(CeedVectorCreate(ceed, q_size, &active_inputs[num_active_in + field])); 497 CeedCallBackend( 498 CeedVectorSetArray(active_inputs[num_active_in + field], CEED_MEM_DEVICE, CEED_USE_POINTER, &q_vec_array[field * Q * num_elem])); 499 } 500 num_active_in += size; 501 CeedCallBackend(CeedVectorRestoreArray(impl->q_vecs_in[i], &q_vec_array)); 502 } 503 } 504 impl->num_active_in = num_active_in; 505 impl->qf_active_in = active_inputs; 506 } 507 508 // Count number of active output fields 509 if (!num_active_out) { 510 for (CeedInt i = 0; i < num_output_fields; i++) { 511 CeedVector vec; 512 513 // Get output vector 514 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 515 // Check if active output 516 if (vec == CEED_VECTOR_ACTIVE) { 517 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size)); 518 num_active_out += size; 519 } 520 } 521 impl->num_active_out = num_active_out; 522 } 523 524 // Check sizes 525 CeedCheck(num_active_in > 0 && num_active_out > 0, ceed, CEED_ERROR_BACKEND, "Cannot assemble QFunction without active inputs and outputs"); 526 527 // Build objects if needed 528 if (build_objects) { 529 // Create output restriction 530 CeedInt strides[3] = {1, num_elem * Q, Q}; /* *NOPAD* */ 531 CeedCallBackend(CeedElemRestrictionCreateStrided(ceed_parent, num_elem, Q, num_active_in * num_active_out, 532 num_active_in * num_active_out * num_elem * Q, strides, rstr)); 533 // Create assembled vector 534 CeedSize l_size = (CeedSize)num_elem * Q * num_active_in * num_active_out; 535 CeedCallBackend(CeedVectorCreate(ceed_parent, l_size, assembled)); 536 } 537 CeedCallBackend(CeedVectorSetValue(*assembled, 0.0)); 538 CeedCallBackend(CeedVectorGetArray(*assembled, CEED_MEM_DEVICE, &assembled_array)); 539 540 // Input basis apply 541 CeedCallBackend(CeedOperatorInputBasis_Cuda(num_elem, qf_input_fields, op_input_fields, num_input_fields, true, e_data, impl)); 542 543 // Assemble QFunction 544 for (CeedInt in = 0; in < num_active_in; in++) { 545 // Set Inputs 546 CeedCallBackend(CeedVectorSetValue(active_inputs[in], 1.0)); 547 if (num_active_in > 1) { 548 CeedCallBackend(CeedVectorSetValue(active_inputs[(in + num_active_in - 1) % num_active_in], 0.0)); 549 } 550 // Set Outputs 551 for (CeedInt out = 0; out < num_output_fields; out++) { 552 CeedVector vec; 553 554 // Get output vector 555 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec)); 556 // Check if active output 557 if (vec == CEED_VECTOR_ACTIVE) { 558 CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, CEED_USE_POINTER, assembled_array)); 559 CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[out], &size)); 560 assembled_array += size * Q * num_elem; // Advance the pointer by the size of the output 561 } 562 } 563 // Apply QFunction 564 CeedCallBackend(CeedQFunctionApply(qf, Q * num_elem, impl->q_vecs_in, impl->q_vecs_out)); 565 } 566 567 // Un-set output q_vecs to prevent accidental overwrite of Assembled 568 for (CeedInt out = 0; out < num_output_fields; out++) { 569 CeedVector vec; 570 571 // Get output vector 572 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec)); 573 // Check if active output 574 if (vec == CEED_VECTOR_ACTIVE) { 575 CeedCallBackend(CeedVectorTakeArray(impl->q_vecs_out[out], CEED_MEM_DEVICE, NULL)); 576 } 577 } 578 579 // Restore input arrays 580 CeedCallBackend(CeedOperatorRestoreInputs_Cuda(num_input_fields, qf_input_fields, op_input_fields, true, e_data, impl)); 581 582 // Restore output 583 CeedCallBackend(CeedVectorRestoreArray(*assembled, &assembled_array)); 584 return CEED_ERROR_SUCCESS; 585 } 586 587 //------------------------------------------------------------------------------ 588 // Assemble Linear QFunction 589 //------------------------------------------------------------------------------ 590 static int CeedOperatorLinearAssembleQFunction_Cuda(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) { 591 return CeedOperatorLinearAssembleQFunctionCore_Cuda(op, true, assembled, rstr, request); 592 } 593 594 //------------------------------------------------------------------------------ 595 // Update Assembled Linear QFunction 596 //------------------------------------------------------------------------------ 597 static int CeedOperatorLinearAssembleQFunctionUpdate_Cuda(CeedOperator op, CeedVector assembled, CeedElemRestriction rstr, CeedRequest *request) { 598 return CeedOperatorLinearAssembleQFunctionCore_Cuda(op, false, &assembled, &rstr, request); 599 } 600 601 //------------------------------------------------------------------------------ 602 // Create point block restriction 603 //------------------------------------------------------------------------------ 604 static int CreatePointBlockRestriction(CeedElemRestriction rstr, CeedElemRestriction *point_block_rstr) { 605 Ceed ceed; 606 CeedSize l_size; 607 CeedInt num_elem, num_comp, elem_size, comp_stride, *point_block_offsets; 608 const CeedInt *offsets; 609 610 CeedCallBackend(CeedElemRestrictionGetCeed(rstr, &ceed)); 611 CeedCallBackend(CeedElemRestrictionGetOffsets(rstr, CEED_MEM_HOST, &offsets)); 612 613 // Expand offsets 614 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr, &num_elem)); 615 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr, &num_comp)); 616 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr, &elem_size)); 617 CeedCallBackend(CeedElemRestrictionGetCompStride(rstr, &comp_stride)); 618 CeedCallBackend(CeedElemRestrictionGetLVectorSize(rstr, &l_size)); 619 CeedInt shift = num_comp; 620 621 if (comp_stride != 1) shift *= num_comp; 622 CeedCallBackend(CeedCalloc(num_elem * elem_size, &point_block_offsets)); 623 for (CeedInt i = 0; i < num_elem * elem_size; i++) { 624 point_block_offsets[i] = offsets[i] * shift; 625 } 626 627 // Create new restriction 628 CeedCallBackend(CeedElemRestrictionCreate(ceed, num_elem, elem_size, num_comp * num_comp, 1, l_size * num_comp, CEED_MEM_HOST, CEED_OWN_POINTER, 629 point_block_offsets, point_block_rstr)); 630 631 // Cleanup 632 CeedCallBackend(CeedElemRestrictionRestoreOffsets(rstr, &offsets)); 633 return CEED_ERROR_SUCCESS; 634 } 635 636 //------------------------------------------------------------------------------ 637 // Assemble diagonal setup 638 //------------------------------------------------------------------------------ 639 static inline int CeedOperatorAssembleDiagonalSetup_Cuda(CeedOperator op, const bool is_point_block, CeedInt use_ceedsize_idx) { 640 Ceed ceed; 641 char *diagonal_kernel_path, *diagonal_kernel_source; 642 CeedInt num_input_fields, num_output_fields, num_e_mode_in = 0, num_comp = 0, dim = 1, num_e_mode_out = 0, num_nodes, num_qpts; 643 CeedEvalMode *e_mode_in = NULL, *e_mode_out = NULL; 644 CeedElemRestriction rstr_in = NULL, rstr_out = NULL; 645 CeedBasis basis_in = NULL, basis_out = NULL; 646 CeedQFunctionField *qf_fields; 647 CeedQFunction qf; 648 CeedOperatorField *op_fields; 649 CeedOperator_Cuda *impl; 650 651 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 652 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 653 CeedCallBackend(CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields)); 654 655 // Determine active input basis 656 CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL)); 657 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 658 for (CeedInt i = 0; i < num_input_fields; i++) { 659 CeedVector vec; 660 661 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 662 if (vec == CEED_VECTOR_ACTIVE) { 663 CeedEvalMode e_mode; 664 CeedElemRestriction rstr; 665 666 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_in)); 667 CeedCallBackend(CeedBasisGetNumComponents(basis_in, &num_comp)); 668 CeedCallBackend(CeedBasisGetDimension(basis_in, &dim)); 669 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr)); 670 CeedCheck(!rstr_in || rstr_in == rstr, ceed, CEED_ERROR_BACKEND, 671 "Backend does not implement multi-field non-composite operator diagonal assembly"); 672 rstr_in = rstr; 673 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &e_mode)); 674 switch (e_mode) { 675 case CEED_EVAL_NONE: 676 case CEED_EVAL_INTERP: 677 CeedCallBackend(CeedRealloc(num_e_mode_in + 1, &e_mode_in)); 678 e_mode_in[num_e_mode_in] = e_mode; 679 num_e_mode_in += 1; 680 break; 681 case CEED_EVAL_GRAD: 682 CeedCallBackend(CeedRealloc(num_e_mode_in + dim, &e_mode_in)); 683 for (CeedInt d = 0; d < dim; d++) e_mode_in[num_e_mode_in + d] = e_mode; 684 num_e_mode_in += dim; 685 break; 686 case CEED_EVAL_WEIGHT: 687 case CEED_EVAL_DIV: 688 case CEED_EVAL_CURL: 689 break; // Caught by QF Assembly 690 } 691 } 692 } 693 694 // Determine active output basis 695 CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields)); 696 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 697 for (CeedInt i = 0; i < num_output_fields; i++) { 698 CeedVector vec; 699 700 CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec)); 701 if (vec == CEED_VECTOR_ACTIVE) { 702 CeedEvalMode e_mode; 703 CeedElemRestriction rstr; 704 705 CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis_out)); 706 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr)); 707 CeedCheck(!rstr_out || rstr_out == rstr, ceed, CEED_ERROR_BACKEND, 708 "Backend does not implement multi-field non-composite operator diagonal assembly"); 709 rstr_out = rstr; 710 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &e_mode)); 711 switch (e_mode) { 712 case CEED_EVAL_NONE: 713 case CEED_EVAL_INTERP: 714 CeedCallBackend(CeedRealloc(num_e_mode_out + 1, &e_mode_out)); 715 e_mode_out[num_e_mode_out] = e_mode; 716 num_e_mode_out += 1; 717 break; 718 case CEED_EVAL_GRAD: 719 CeedCallBackend(CeedRealloc(num_e_mode_out + dim, &e_mode_out)); 720 for (CeedInt d = 0; d < dim; d++) e_mode_out[num_e_mode_out + d] = e_mode; 721 num_e_mode_out += dim; 722 break; 723 case CEED_EVAL_WEIGHT: 724 case CEED_EVAL_DIV: 725 case CEED_EVAL_CURL: 726 break; // Caught by QF Assembly 727 } 728 } 729 } 730 731 // Operator data struct 732 CeedCallBackend(CeedOperatorGetData(op, &impl)); 733 CeedCallBackend(CeedCalloc(1, &impl->diag)); 734 CeedOperatorDiag_Cuda *diag = impl->diag; 735 736 diag->basis_in = basis_in; 737 diag->basis_out = basis_out; 738 diag->h_e_mode_in = e_mode_in; 739 diag->h_e_mode_out = e_mode_out; 740 diag->num_e_mode_in = num_e_mode_in; 741 diag->num_e_mode_out = num_e_mode_out; 742 743 // Assemble kernel 744 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-operator-assemble-diagonal.h", &diagonal_kernel_path)); 745 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Kernel Source -----\n"); 746 CeedCallBackend(CeedLoadSourceToBuffer(ceed, diagonal_kernel_path, &diagonal_kernel_source)); 747 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Diagonal Assembly Source Complete! -----\n"); 748 CeedCallBackend(CeedBasisGetNumNodes(basis_in, &num_nodes)); 749 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 750 diag->num_nodes = num_nodes; 751 CeedCallCuda(ceed, 752 CeedCompile_Cuda(ceed, diagonal_kernel_source, &diag->module, 6, "NUM_E_MODE_IN", num_e_mode_in, "NUM_E_MODE_OUT", num_e_mode_out, 753 "NUM_NODES", num_nodes, "NUM_QPTS", num_qpts, "NUM_COMP", num_comp, "USE_CEEDSIZE", use_ceedsize_idx)); 754 CeedCallCuda(ceed, CeedGetKernel_Cuda(ceed, diag->module, "linearDiagonal", &diag->linearDiagonal)); 755 CeedCallCuda(ceed, CeedGetKernel_Cuda(ceed, diag->module, "linearPointBlockDiagonal", &diag->linearPointBlock)); 756 CeedCallBackend(CeedFree(&diagonal_kernel_path)); 757 CeedCallBackend(CeedFree(&diagonal_kernel_source)); 758 759 // Basis matrices 760 const CeedInt q_bytes = num_qpts * sizeof(CeedScalar); 761 const CeedInt interp_bytes = q_bytes * num_nodes; 762 const CeedInt grad_bytes = q_bytes * num_nodes * dim; 763 const CeedInt e_mode_bytes = sizeof(CeedEvalMode); 764 const CeedScalar *interp_in, *interp_out, *grad_in, *grad_out; 765 766 // CEED_EVAL_NONE 767 CeedScalar *identity = NULL; 768 bool is_eval_none = false; 769 770 for (CeedInt i = 0; i < num_e_mode_in; i++) is_eval_none = is_eval_none || (e_mode_in[i] == CEED_EVAL_NONE); 771 for (CeedInt i = 0; i < num_e_mode_out; i++) is_eval_none = is_eval_none || (e_mode_out[i] == CEED_EVAL_NONE); 772 if (is_eval_none) { 773 CeedCallBackend(CeedCalloc(num_qpts * num_nodes, &identity)); 774 for (CeedInt i = 0; i < (num_nodes < num_qpts ? num_nodes : num_qpts); i++) identity[i * num_nodes + i] = 1.0; 775 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_identity, interp_bytes)); 776 CeedCallCuda(ceed, cudaMemcpy(diag->d_identity, identity, interp_bytes, cudaMemcpyHostToDevice)); 777 } 778 779 // CEED_EVAL_INTERP 780 CeedCallBackend(CeedBasisGetInterp(basis_in, &interp_in)); 781 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_interp_in, interp_bytes)); 782 CeedCallCuda(ceed, cudaMemcpy(diag->d_interp_in, interp_in, interp_bytes, cudaMemcpyHostToDevice)); 783 CeedCallBackend(CeedBasisGetInterp(basis_out, &interp_out)); 784 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_interp_out, interp_bytes)); 785 CeedCallCuda(ceed, cudaMemcpy(diag->d_interp_out, interp_out, interp_bytes, cudaMemcpyHostToDevice)); 786 787 // CEED_EVAL_GRAD 788 CeedCallBackend(CeedBasisGetGrad(basis_in, &grad_in)); 789 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_grad_in, grad_bytes)); 790 CeedCallCuda(ceed, cudaMemcpy(diag->d_grad_in, grad_in, grad_bytes, cudaMemcpyHostToDevice)); 791 CeedCallBackend(CeedBasisGetGrad(basis_out, &grad_out)); 792 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_grad_out, grad_bytes)); 793 CeedCallCuda(ceed, cudaMemcpy(diag->d_grad_out, grad_out, grad_bytes, cudaMemcpyHostToDevice)); 794 795 // Arrays of e_modes 796 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_e_mode_in, num_e_mode_in * e_mode_bytes)); 797 CeedCallCuda(ceed, cudaMemcpy(diag->d_e_mode_in, e_mode_in, num_e_mode_in * e_mode_bytes, cudaMemcpyHostToDevice)); 798 CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_e_mode_out, num_e_mode_out * e_mode_bytes)); 799 CeedCallCuda(ceed, cudaMemcpy(diag->d_e_mode_out, e_mode_out, num_e_mode_out * e_mode_bytes, cudaMemcpyHostToDevice)); 800 801 // Restriction 802 diag->diag_rstr = rstr_out; 803 return CEED_ERROR_SUCCESS; 804 } 805 806 //------------------------------------------------------------------------------ 807 // Assemble diagonal common code 808 //------------------------------------------------------------------------------ 809 static inline int CeedOperatorAssembleDiagonalCore_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request, const bool is_point_block) { 810 Ceed ceed; 811 CeedSize assembled_length = 0, assembled_qf_length = 0; 812 CeedInt use_ceedsize_idx = 0, num_elem; 813 CeedScalar *elem_diag_array; 814 const CeedScalar *assembled_qf_array; 815 CeedVector assembled_qf = NULL; 816 CeedElemRestriction rstr = NULL; 817 CeedOperator_Cuda *impl; 818 819 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 820 CeedCallBackend(CeedOperatorGetData(op, &impl)); 821 822 // Assemble QFunction 823 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &rstr, request)); 824 CeedCallBackend(CeedElemRestrictionDestroy(&rstr)); 825 826 CeedCallBackend(CeedVectorGetLength(assembled, &assembled_length)); 827 CeedCallBackend(CeedVectorGetLength(assembled_qf, &assembled_qf_length)); 828 if ((assembled_length > INT_MAX) || (assembled_qf_length > INT_MAX)) use_ceedsize_idx = 1; 829 830 // Setup 831 if (!impl->diag) CeedCallBackend(CeedOperatorAssembleDiagonalSetup_Cuda(op, is_point_block, use_ceedsize_idx)); 832 CeedOperatorDiag_Cuda *diag = impl->diag; 833 834 assert(diag != NULL); 835 836 // Restriction 837 if (is_point_block && !diag->point_block_rstr) { 838 CeedElemRestriction point_block_rstr; 839 840 CeedCallBackend(CreatePointBlockRestriction(diag->diag_rstr, &point_block_rstr)); 841 diag->point_block_rstr = point_block_rstr; 842 } 843 CeedElemRestriction diag_rstr = is_point_block ? diag->point_block_rstr : diag->diag_rstr; 844 845 // Create diagonal vector 846 CeedVector elem_diag = is_point_block ? diag->point_block_elem_diag : diag->elem_diag; 847 848 if (!elem_diag) { 849 CeedCallBackend(CeedElemRestrictionCreateVector(diag_rstr, NULL, &elem_diag)); 850 if (is_point_block) diag->point_block_elem_diag = elem_diag; 851 else diag->elem_diag = elem_diag; 852 } 853 CeedCallBackend(CeedVectorSetValue(elem_diag, 0.0)); 854 855 // Assemble element operator diagonals 856 CeedCallBackend(CeedVectorGetArray(elem_diag, CEED_MEM_DEVICE, &elem_diag_array)); 857 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &assembled_qf_array)); 858 CeedCallBackend(CeedElemRestrictionGetNumElements(diag_rstr, &num_elem)); 859 860 // Compute the diagonal of B^T D B 861 int elem_per_block = 1; 862 int grid = num_elem / elem_per_block + ((num_elem / elem_per_block * elem_per_block < num_elem) ? 1 : 0); 863 void *args[] = {(void *)&num_elem, &diag->d_identity, &diag->d_interp_in, &diag->d_grad_in, &diag->d_interp_out, 864 &diag->d_grad_out, &diag->d_e_mode_in, &diag->d_e_mode_out, &assembled_qf_array, &elem_diag_array}; 865 if (is_point_block) { 866 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, diag->linearPointBlock, grid, diag->num_nodes, 1, elem_per_block, args)); 867 } else { 868 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, diag->linearDiagonal, grid, diag->num_nodes, 1, elem_per_block, args)); 869 } 870 871 // Restore arrays 872 CeedCallBackend(CeedVectorRestoreArray(elem_diag, &elem_diag_array)); 873 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &assembled_qf_array)); 874 875 // Assemble local operator diagonal 876 CeedCallBackend(CeedElemRestrictionApply(diag_rstr, CEED_TRANSPOSE, elem_diag, assembled, request)); 877 878 // Cleanup 879 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 880 return CEED_ERROR_SUCCESS; 881 } 882 883 //------------------------------------------------------------------------------ 884 // Assemble Linear Diagonal 885 //------------------------------------------------------------------------------ 886 static int CeedOperatorLinearAssembleAddDiagonal_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request) { 887 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Cuda(op, assembled, request, false)); 888 return CEED_ERROR_SUCCESS; 889 } 890 891 //------------------------------------------------------------------------------ 892 // Assemble Linear Point Block Diagonal 893 //------------------------------------------------------------------------------ 894 static int CeedOperatorLinearAssembleAddPointBlockDiagonal_Cuda(CeedOperator op, CeedVector assembled, CeedRequest *request) { 895 CeedCallBackend(CeedOperatorAssembleDiagonalCore_Cuda(op, assembled, request, true)); 896 return CEED_ERROR_SUCCESS; 897 } 898 899 //------------------------------------------------------------------------------ 900 // Single operator assembly setup 901 //------------------------------------------------------------------------------ 902 static int CeedSingleOperatorAssembleSetup_Cuda(CeedOperator op, CeedInt use_ceedsize_idx) { 903 Ceed ceed; 904 char *assembly_kernel_path, *assembly_kernel_source; 905 CeedInt num_input_fields, num_output_fields, num_e_mode_in = 0, dim = 1, num_B_in_mats_to_load = 0, size_B_in = 0, num_qpts = 0, elem_size = 0, 906 num_e_mode_out = 0, num_B_out_mats_to_load = 0, size_B_out = 0, num_elem, num_comp; 907 CeedEvalMode *eval_mode_in = NULL, *eval_mode_out = NULL; 908 CeedElemRestriction rstr_in = NULL, rstr_out = NULL; 909 CeedBasis basis_in = NULL, basis_out = NULL; 910 CeedQFunctionField *qf_fields; 911 CeedQFunction qf; 912 CeedOperatorField *input_fields, *output_fields; 913 CeedOperator_Cuda *impl; 914 915 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 916 CeedCallBackend(CeedOperatorGetData(op, &impl)); 917 918 // Get intput and output fields 919 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &input_fields, &num_output_fields, &output_fields)); 920 921 // Determine active input basis eval mode 922 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 923 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL)); 924 // Note that the kernel will treat each dimension of a gradient action separately; 925 // i.e., when an active input has a CEED_EVAL_GRAD mode, num_e_mode_in will increment by dim. 926 // However, for the purposes of loading the B matrices, it will be treated as one mode, and we will load/copy the entire gradient matrix at once, so 927 // num_B_in_mats_to_load will be incremented by 1. 928 for (CeedInt i = 0; i < num_input_fields; i++) { 929 CeedVector vec; 930 931 CeedCallBackend(CeedOperatorFieldGetVector(input_fields[i], &vec)); 932 if (vec == CEED_VECTOR_ACTIVE) { 933 CeedEvalMode eval_mode; 934 935 CeedCallBackend(CeedOperatorFieldGetBasis(input_fields[i], &basis_in)); 936 CeedCallBackend(CeedBasisGetDimension(basis_in, &dim)); 937 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis_in, &num_qpts)); 938 CeedCallBackend(CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr_in)); 939 CeedCallBackend(CeedElemRestrictionGetElementSize(rstr_in, &elem_size)); 940 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 941 if (eval_mode != CEED_EVAL_NONE) { 942 CeedCallBackend(CeedRealloc(num_B_in_mats_to_load + 1, &eval_mode_in)); 943 eval_mode_in[num_B_in_mats_to_load] = eval_mode; 944 num_B_in_mats_to_load += 1; 945 if (eval_mode == CEED_EVAL_GRAD) { 946 num_e_mode_in += dim; 947 size_B_in += dim * elem_size * num_qpts; 948 } else { 949 num_e_mode_in += 1; 950 size_B_in += elem_size * num_qpts; 951 } 952 } 953 } 954 } 955 956 // Determine active output basis; basis_out and rstr_out only used if same as input, TODO 957 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields)); 958 for (CeedInt i = 0; i < num_output_fields; i++) { 959 CeedVector vec; 960 961 CeedCallBackend(CeedOperatorFieldGetVector(output_fields[i], &vec)); 962 if (vec == CEED_VECTOR_ACTIVE) { 963 CeedEvalMode eval_mode; 964 965 CeedCallBackend(CeedOperatorFieldGetBasis(output_fields[i], &basis_out)); 966 CeedCallBackend(CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr_out)); 967 CeedCheck(!rstr_out || rstr_out == rstr_in, ceed, CEED_ERROR_BACKEND, "Backend does not implement multi-field non-composite operator assembly"); 968 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode)); 969 if (eval_mode != CEED_EVAL_NONE) { 970 CeedCallBackend(CeedRealloc(num_B_out_mats_to_load + 1, &eval_mode_out)); 971 eval_mode_out[num_B_out_mats_to_load] = eval_mode; 972 num_B_out_mats_to_load += 1; 973 if (eval_mode == CEED_EVAL_GRAD) { 974 num_e_mode_out += dim; 975 size_B_out += dim * elem_size * num_qpts; 976 } else { 977 num_e_mode_out += 1; 978 size_B_out += elem_size * num_qpts; 979 } 980 } 981 } 982 } 983 CeedCheck(num_e_mode_in > 0 && num_e_mode_out > 0, ceed, CEED_ERROR_UNSUPPORTED, "Cannot assemble operator without inputs/outputs"); 984 985 CeedCallBackend(CeedElemRestrictionGetNumElements(rstr_in, &num_elem)); 986 CeedCallBackend(CeedElemRestrictionGetNumComponents(rstr_in, &num_comp)); 987 988 CeedCallBackend(CeedCalloc(1, &impl->asmb)); 989 CeedOperatorAssemble_Cuda *asmb = impl->asmb; 990 asmb->num_elem = num_elem; 991 992 // Compile kernels 993 int elem_per_block = 1; 994 asmb->elem_per_block = elem_per_block; 995 CeedInt block_size = elem_size * elem_size * elem_per_block; 996 Ceed_Cuda *cuda_data; 997 998 CeedCallBackend(CeedGetData(ceed, &cuda_data)); 999 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-operator-assemble.h", &assembly_kernel_path)); 1000 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Kernel Source -----\n"); 1001 CeedCallBackend(CeedLoadSourceToBuffer(ceed, assembly_kernel_path, &assembly_kernel_source)); 1002 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Assembly Source Complete! -----\n"); 1003 bool fallback = block_size > cuda_data->device_prop.maxThreadsPerBlock; 1004 1005 if (fallback) { 1006 // Use fallback kernel with 1D threadblock 1007 block_size = elem_size * elem_per_block; 1008 asmb->block_size_x = elem_size; 1009 asmb->block_size_y = 1; 1010 } else { // Use kernel with 2D threadblock 1011 asmb->block_size_x = elem_size; 1012 asmb->block_size_y = elem_size; 1013 } 1014 CeedCallBackend(CeedCompile_Cuda(ceed, assembly_kernel_source, &asmb->module, 8, "NUM_ELEM", num_elem, "NUM_E_MODE_IN", num_e_mode_in, 1015 "NUM_E_MODE_OUT", num_e_mode_out, "NUM_QPTS", num_qpts, "NUM_NODES", elem_size, "BLOCK_SIZE", block_size, 1016 "NUM_COMP", num_comp, "USE_CEEDSIZE", use_ceedsize_idx)); 1017 CeedCallBackend(CeedGetKernel_Cuda(ceed, asmb->module, fallback ? "linearAssembleFallback" : "linearAssemble", &asmb->linearAssemble)); 1018 CeedCallBackend(CeedFree(&assembly_kernel_path)); 1019 CeedCallBackend(CeedFree(&assembly_kernel_source)); 1020 1021 // Build 'full' B matrices (not 1D arrays used for tensor-product matrices) 1022 const CeedScalar *interp_in, *grad_in; 1023 1024 CeedCallBackend(CeedBasisGetInterp(basis_in, &interp_in)); 1025 CeedCallBackend(CeedBasisGetGrad(basis_in, &grad_in)); 1026 1027 // Load into B_in, in order that they will be used in eval_mode 1028 const CeedInt inBytes = size_B_in * sizeof(CeedScalar); 1029 CeedInt mat_start = 0; 1030 1031 CeedCallCuda(ceed, cudaMalloc((void **)&asmb->d_B_in, inBytes)); 1032 for (int i = 0; i < num_B_in_mats_to_load; i++) { 1033 CeedEvalMode eval_mode = eval_mode_in[i]; 1034 1035 if (eval_mode == CEED_EVAL_INTERP) { 1036 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_in[mat_start], interp_in, elem_size * num_qpts * sizeof(CeedScalar), cudaMemcpyHostToDevice)); 1037 mat_start += elem_size * num_qpts; 1038 } else if (eval_mode == CEED_EVAL_GRAD) { 1039 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_in[mat_start], grad_in, dim * elem_size * num_qpts * sizeof(CeedScalar), cudaMemcpyHostToDevice)); 1040 mat_start += dim * elem_size * num_qpts; 1041 } 1042 } 1043 1044 const CeedScalar *interp_out, *grad_out; 1045 1046 // Note that this function currently assumes 1 basis, so this should always be true for now 1047 if (basis_out == basis_in) { 1048 interp_out = interp_in; 1049 grad_out = grad_in; 1050 } else { 1051 CeedCallBackend(CeedBasisGetInterp(basis_out, &interp_out)); 1052 CeedCallBackend(CeedBasisGetGrad(basis_out, &grad_out)); 1053 } 1054 1055 // Load into B_out, in order that they will be used in eval_mode 1056 const CeedInt outBytes = size_B_out * sizeof(CeedScalar); 1057 mat_start = 0; 1058 1059 CeedCallCuda(ceed, cudaMalloc((void **)&asmb->d_B_out, outBytes)); 1060 for (int i = 0; i < num_B_out_mats_to_load; i++) { 1061 CeedEvalMode eval_mode = eval_mode_out[i]; 1062 1063 if (eval_mode == CEED_EVAL_INTERP) { 1064 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_out[mat_start], interp_out, elem_size * num_qpts * sizeof(CeedScalar), cudaMemcpyHostToDevice)); 1065 mat_start += elem_size * num_qpts; 1066 } else if (eval_mode == CEED_EVAL_GRAD) { 1067 CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_out[mat_start], grad_out, dim * elem_size * num_qpts * sizeof(CeedScalar), cudaMemcpyHostToDevice)); 1068 mat_start += dim * elem_size * num_qpts; 1069 } 1070 } 1071 return CEED_ERROR_SUCCESS; 1072 } 1073 1074 //------------------------------------------------------------------------------ 1075 // Assemble matrix data for COO matrix of assembled operator. 1076 // The sparsity pattern is set by CeedOperatorLinearAssembleSymbolic. 1077 // 1078 // Note that this (and other assembly routines) currently assume only one active input restriction/basis per operator (could have multiple basis eval 1079 // modes). 1080 // TODO: allow multiple active input restrictions/basis objects 1081 //------------------------------------------------------------------------------ 1082 static int CeedSingleOperatorAssemble_Cuda(CeedOperator op, CeedInt offset, CeedVector values) { 1083 Ceed ceed; 1084 CeedSize values_length = 0, assembled_qf_length = 0; 1085 CeedInt use_ceedsize_idx = 0; 1086 CeedScalar *values_array; 1087 const CeedScalar *qf_array; 1088 CeedVector assembled_qf = NULL; 1089 CeedElemRestriction rstr_q = NULL; 1090 CeedOperator_Cuda *impl; 1091 1092 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1093 CeedCallBackend(CeedOperatorGetData(op, &impl)); 1094 1095 // Assemble QFunction 1096 CeedCallBackend(CeedOperatorLinearAssembleQFunctionBuildOrUpdate(op, &assembled_qf, &rstr_q, CEED_REQUEST_IMMEDIATE)); 1097 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_q)); 1098 CeedCallBackend(CeedVectorGetArray(values, CEED_MEM_DEVICE, &values_array)); 1099 values_array += offset; 1100 CeedCallBackend(CeedVectorGetArrayRead(assembled_qf, CEED_MEM_DEVICE, &qf_array)); 1101 1102 CeedCallBackend(CeedVectorGetLength(values, &values_length)); 1103 CeedCallBackend(CeedVectorGetLength(assembled_qf, &assembled_qf_length)); 1104 if ((values_length > INT_MAX) || (assembled_qf_length > INT_MAX)) use_ceedsize_idx = 1; 1105 // Setup 1106 if (!impl->asmb) { 1107 CeedCallBackend(CeedSingleOperatorAssembleSetup_Cuda(op, use_ceedsize_idx)); 1108 assert(impl->asmb != NULL); 1109 } 1110 1111 // Compute B^T D B 1112 const CeedInt num_elem = impl->asmb->num_elem; 1113 const CeedInt elem_per_block = impl->asmb->elem_per_block; 1114 const CeedInt grid = num_elem / elem_per_block + ((num_elem / elem_per_block * elem_per_block < num_elem) ? 1 : 0); 1115 void *args[] = {&impl->asmb->d_B_in, &impl->asmb->d_B_out, &qf_array, &values_array}; 1116 1117 CeedCallBackend( 1118 CeedRunKernelDim_Cuda(ceed, impl->asmb->linearAssemble, grid, impl->asmb->block_size_x, impl->asmb->block_size_y, elem_per_block, args)); 1119 1120 // Restore arrays 1121 CeedCallBackend(CeedVectorRestoreArray(values, &values_array)); 1122 CeedCallBackend(CeedVectorRestoreArrayRead(assembled_qf, &qf_array)); 1123 1124 // Cleanup 1125 CeedCallBackend(CeedVectorDestroy(&assembled_qf)); 1126 return CEED_ERROR_SUCCESS; 1127 } 1128 1129 //------------------------------------------------------------------------------ 1130 // Create operator 1131 //------------------------------------------------------------------------------ 1132 int CeedOperatorCreate_Cuda(CeedOperator op) { 1133 Ceed ceed; 1134 CeedOperator_Cuda *impl; 1135 1136 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1137 CeedCallBackend(CeedCalloc(1, &impl)); 1138 CeedCallBackend(CeedOperatorSetData(op, impl)); 1139 1140 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunction_Cuda)); 1141 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunctionUpdate", CeedOperatorLinearAssembleQFunctionUpdate_Cuda)); 1142 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddDiagonal", CeedOperatorLinearAssembleAddDiagonal_Cuda)); 1143 CeedCallBackend( 1144 CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleAddPointBlockDiagonal", CeedOperatorLinearAssembleAddPointBlockDiagonal_Cuda)); 1145 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleSingle", CeedSingleOperatorAssemble_Cuda)); 1146 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Cuda)); 1147 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Cuda)); 1148 return CEED_ERROR_SUCCESS; 1149 } 1150 1151 //------------------------------------------------------------------------------ 1152