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 <math.h> 11 #include <stdbool.h> 12 #include <string.h> 13 #include <hip/hip_runtime.h> 14 15 #include "../hip/ceed-hip-common.h" 16 #include "ceed-hip-ref.h" 17 18 //------------------------------------------------------------------------------ 19 // Check if host/device sync is needed 20 //------------------------------------------------------------------------------ 21 static inline int CeedVectorNeedSync_Hip(const CeedVector vec, CeedMemType mem_type, bool *need_sync) { 22 CeedVector_Hip *impl; 23 bool has_valid_array = false; 24 25 CeedCallBackend(CeedVectorGetData(vec, &impl)); 26 CeedCallBackend(CeedVectorHasValidArray(vec, &has_valid_array)); 27 switch (mem_type) { 28 case CEED_MEM_HOST: 29 *need_sync = has_valid_array && !impl->h_array; 30 break; 31 case CEED_MEM_DEVICE: 32 *need_sync = has_valid_array && !impl->d_array; 33 break; 34 } 35 return CEED_ERROR_SUCCESS; 36 } 37 38 //------------------------------------------------------------------------------ 39 // Sync host to device 40 //------------------------------------------------------------------------------ 41 static inline int CeedVectorSyncH2D_Hip(const CeedVector vec) { 42 Ceed ceed; 43 CeedSize length; 44 size_t bytes; 45 CeedVector_Hip *impl; 46 47 CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 48 CeedCallBackend(CeedVectorGetData(vec, &impl)); 49 50 CeedCheck(impl->h_array, ceed, CEED_ERROR_BACKEND, "No valid host data to sync to device"); 51 52 CeedCallBackend(CeedVectorGetLength(vec, &length)); 53 bytes = length * sizeof(CeedScalar); 54 if (impl->d_array_borrowed) { 55 impl->d_array = impl->d_array_borrowed; 56 } else if (impl->d_array_owned) { 57 impl->d_array = impl->d_array_owned; 58 } else { 59 CeedCallHip(ceed, hipMalloc((void **)&impl->d_array_owned, bytes)); 60 impl->d_array = impl->d_array_owned; 61 } 62 CeedCallHip(ceed, hipMemcpy(impl->d_array, impl->h_array, bytes, hipMemcpyHostToDevice)); 63 return CEED_ERROR_SUCCESS; 64 } 65 66 //------------------------------------------------------------------------------ 67 // Sync device to host 68 //------------------------------------------------------------------------------ 69 static inline int CeedVectorSyncD2H_Hip(const CeedVector vec) { 70 Ceed ceed; 71 CeedSize length; 72 size_t bytes; 73 CeedVector_Hip *impl; 74 75 CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 76 CeedCallBackend(CeedVectorGetData(vec, &impl)); 77 78 CeedCheck(impl->d_array, ceed, CEED_ERROR_BACKEND, "No valid device data to sync to host"); 79 80 if (impl->h_array_borrowed) { 81 impl->h_array = impl->h_array_borrowed; 82 } else if (impl->h_array_owned) { 83 impl->h_array = impl->h_array_owned; 84 } else { 85 CeedSize length; 86 87 CeedCallBackend(CeedVectorGetLength(vec, &length)); 88 CeedCallBackend(CeedCalloc(length, &impl->h_array_owned)); 89 impl->h_array = impl->h_array_owned; 90 } 91 92 CeedCallBackend(CeedVectorGetLength(vec, &length)); 93 bytes = length * sizeof(CeedScalar); 94 CeedCallHip(ceed, hipMemcpy(impl->h_array, impl->d_array, bytes, hipMemcpyDeviceToHost)); 95 return CEED_ERROR_SUCCESS; 96 } 97 98 //------------------------------------------------------------------------------ 99 // Sync arrays 100 //------------------------------------------------------------------------------ 101 static int CeedVectorSyncArray_Hip(const CeedVector vec, CeedMemType mem_type) { 102 bool need_sync = false; 103 104 // Check whether device/host sync is needed 105 CeedCallBackend(CeedVectorNeedSync_Hip(vec, mem_type, &need_sync)); 106 if (!need_sync) return CEED_ERROR_SUCCESS; 107 108 switch (mem_type) { 109 case CEED_MEM_HOST: 110 return CeedVectorSyncD2H_Hip(vec); 111 case CEED_MEM_DEVICE: 112 return CeedVectorSyncH2D_Hip(vec); 113 } 114 return CEED_ERROR_UNSUPPORTED; 115 } 116 117 //------------------------------------------------------------------------------ 118 // Set all pointers as invalid 119 //------------------------------------------------------------------------------ 120 static inline int CeedVectorSetAllInvalid_Hip(const CeedVector vec) { 121 CeedVector_Hip *impl; 122 123 CeedCallBackend(CeedVectorGetData(vec, &impl)); 124 impl->h_array = NULL; 125 impl->d_array = NULL; 126 return CEED_ERROR_SUCCESS; 127 } 128 129 //------------------------------------------------------------------------------ 130 // Check if CeedVector has any valid pointer 131 //------------------------------------------------------------------------------ 132 static inline int CeedVectorHasValidArray_Hip(const CeedVector vec, bool *has_valid_array) { 133 CeedVector_Hip *impl; 134 135 CeedCallBackend(CeedVectorGetData(vec, &impl)); 136 *has_valid_array = impl->h_array || impl->d_array; 137 return CEED_ERROR_SUCCESS; 138 } 139 140 //------------------------------------------------------------------------------ 141 // Check if has array of given type 142 //------------------------------------------------------------------------------ 143 static inline int CeedVectorHasArrayOfType_Hip(const CeedVector vec, CeedMemType mem_type, bool *has_array_of_type) { 144 CeedVector_Hip *impl; 145 146 CeedCallBackend(CeedVectorGetData(vec, &impl)); 147 switch (mem_type) { 148 case CEED_MEM_HOST: 149 *has_array_of_type = impl->h_array_borrowed || impl->h_array_owned; 150 break; 151 case CEED_MEM_DEVICE: 152 *has_array_of_type = impl->d_array_borrowed || impl->d_array_owned; 153 break; 154 } 155 return CEED_ERROR_SUCCESS; 156 } 157 158 //------------------------------------------------------------------------------ 159 // Check if has borrowed array of given type 160 //------------------------------------------------------------------------------ 161 static inline int CeedVectorHasBorrowedArrayOfType_Hip(const CeedVector vec, CeedMemType mem_type, bool *has_borrowed_array_of_type) { 162 CeedVector_Hip *impl; 163 164 CeedCallBackend(CeedVectorGetData(vec, &impl)); 165 switch (mem_type) { 166 case CEED_MEM_HOST: 167 *has_borrowed_array_of_type = impl->h_array_borrowed; 168 break; 169 case CEED_MEM_DEVICE: 170 *has_borrowed_array_of_type = impl->d_array_borrowed; 171 break; 172 } 173 return CEED_ERROR_SUCCESS; 174 } 175 176 //------------------------------------------------------------------------------ 177 // Set array from host 178 //------------------------------------------------------------------------------ 179 static int CeedVectorSetArrayHost_Hip(const CeedVector vec, const CeedCopyMode copy_mode, CeedScalar *array) { 180 CeedSize length; 181 CeedVector_Hip *impl; 182 183 CeedCallBackend(CeedVectorGetData(vec, &impl)); 184 CeedCallBackend(CeedVectorGetLength(vec, &length)); 185 186 CeedCallBackend(CeedSetHostCeedScalarArray(array, copy_mode, length, (const CeedScalar **)&impl->h_array_owned, 187 (const CeedScalar **)&impl->h_array_borrowed, (const CeedScalar **)&impl->h_array)); 188 return CEED_ERROR_SUCCESS; 189 } 190 191 //------------------------------------------------------------------------------ 192 // Set array from device 193 //------------------------------------------------------------------------------ 194 static int CeedVectorSetArrayDevice_Hip(const CeedVector vec, const CeedCopyMode copy_mode, CeedScalar *array) { 195 CeedSize length; 196 Ceed ceed; 197 CeedVector_Hip *impl; 198 199 CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 200 CeedCallBackend(CeedVectorGetData(vec, &impl)); 201 CeedCallBackend(CeedVectorGetLength(vec, &length)); 202 203 CeedCallBackend(CeedSetDeviceCeedScalarArray_Hip(ceed, array, copy_mode, length, (const CeedScalar **)&impl->d_array_owned, 204 (const CeedScalar **)&impl->d_array_borrowed, (const CeedScalar **)&impl->d_array)); 205 return CEED_ERROR_SUCCESS; 206 } 207 208 //------------------------------------------------------------------------------ 209 // Set the array used by a vector, 210 // freeing any previously allocated array if applicable 211 //------------------------------------------------------------------------------ 212 static int CeedVectorSetArray_Hip(const CeedVector vec, const CeedMemType mem_type, const CeedCopyMode copy_mode, CeedScalar *array) { 213 CeedVector_Hip *impl; 214 215 CeedCallBackend(CeedVectorGetData(vec, &impl)); 216 CeedCallBackend(CeedVectorSetAllInvalid_Hip(vec)); 217 switch (mem_type) { 218 case CEED_MEM_HOST: 219 return CeedVectorSetArrayHost_Hip(vec, copy_mode, array); 220 case CEED_MEM_DEVICE: 221 return CeedVectorSetArrayDevice_Hip(vec, copy_mode, array); 222 } 223 return CEED_ERROR_UNSUPPORTED; 224 } 225 226 //------------------------------------------------------------------------------ 227 // Copy host array to value strided 228 //------------------------------------------------------------------------------ 229 static int CeedHostCopyStrided_Hip(CeedScalar *h_array, CeedSize start, CeedSize step, CeedSize length, CeedScalar *h_copy_array) { 230 for (CeedSize i = start; i < length; i += step) h_copy_array[i] = h_array[i]; 231 return CEED_ERROR_SUCCESS; 232 } 233 234 //------------------------------------------------------------------------------ 235 // Copy device array to value strided (impl in .hip.cpp file) 236 //------------------------------------------------------------------------------ 237 int CeedDeviceCopyStrided_Hip(CeedScalar *d_array, CeedSize start, CeedSize step, CeedSize length, CeedScalar *d_copy_array); 238 239 //------------------------------------------------------------------------------ 240 // Copy a vector to a value strided 241 //------------------------------------------------------------------------------ 242 static int CeedVectorCopyStrided_Hip(CeedVector vec, CeedSize start, CeedSize step, CeedVector vec_copy) { 243 CeedSize length; 244 CeedVector_Hip *impl; 245 246 CeedCallBackend(CeedVectorGetData(vec, &impl)); 247 CeedCallBackend(CeedVectorGetLength(vec, &length)); 248 // Set value for synced device/host array 249 if (impl->d_array) { 250 CeedScalar *copy_array; 251 252 CeedCallBackend(CeedVectorGetArray(vec_copy, CEED_MEM_DEVICE, ©_array)); 253 CeedCallBackend(CeedDeviceCopyStrided_Hip(impl->d_array, start, step, length, copy_array)); 254 CeedCallBackend(CeedVectorRestoreArray(vec_copy, ©_array)); 255 } else if (impl->h_array) { 256 CeedScalar *copy_array; 257 258 CeedCallBackend(CeedVectorGetArray(vec_copy, CEED_MEM_HOST, ©_array)); 259 CeedCallBackend(CeedHostCopyStrided_Hip(impl->h_array, start, step, length, copy_array)); 260 CeedCallBackend(CeedVectorRestoreArray(vec_copy, ©_array)); 261 } else { 262 return CeedError(CeedVectorReturnCeed(vec), CEED_ERROR_BACKEND, "CeedVector must have valid data set"); 263 } 264 return CEED_ERROR_SUCCESS; 265 } 266 267 //------------------------------------------------------------------------------ 268 // Set host array to value 269 //------------------------------------------------------------------------------ 270 static int CeedHostSetValue_Hip(CeedScalar *h_array, CeedSize length, CeedScalar val) { 271 for (CeedSize i = 0; i < length; i++) h_array[i] = val; 272 return CEED_ERROR_SUCCESS; 273 } 274 275 //------------------------------------------------------------------------------ 276 // Set device array to value (impl in .hip file) 277 //------------------------------------------------------------------------------ 278 int CeedDeviceSetValue_Hip(CeedScalar *d_array, CeedSize length, CeedScalar val); 279 280 //------------------------------------------------------------------------------ 281 // Set a vector to a value 282 //------------------------------------------------------------------------------ 283 static int CeedVectorSetValue_Hip(CeedVector vec, CeedScalar val) { 284 CeedSize length; 285 CeedVector_Hip *impl; 286 287 CeedCallBackend(CeedVectorGetData(vec, &impl)); 288 CeedCallBackend(CeedVectorGetLength(vec, &length)); 289 // Set value for synced device/host array 290 if (!impl->d_array && !impl->h_array) { 291 if (impl->d_array_borrowed) { 292 impl->d_array = impl->d_array_borrowed; 293 } else if (impl->h_array_borrowed) { 294 impl->h_array = impl->h_array_borrowed; 295 } else if (impl->d_array_owned) { 296 impl->d_array = impl->d_array_owned; 297 } else if (impl->h_array_owned) { 298 impl->h_array = impl->h_array_owned; 299 } else { 300 CeedCallBackend(CeedVectorSetArray(vec, CEED_MEM_DEVICE, CEED_COPY_VALUES, NULL)); 301 } 302 } 303 if (impl->d_array) { 304 CeedCallBackend(CeedDeviceSetValue_Hip(impl->d_array, length, val)); 305 impl->h_array = NULL; 306 } 307 if (impl->h_array) { 308 CeedCallBackend(CeedHostSetValue_Hip(impl->h_array, length, val)); 309 impl->d_array = NULL; 310 } 311 return CEED_ERROR_SUCCESS; 312 } 313 314 //------------------------------------------------------------------------------ 315 // Set host array to value strided 316 //------------------------------------------------------------------------------ 317 static int CeedHostSetValueStrided_Hip(CeedScalar *h_array, CeedSize start, CeedSize step, CeedSize length, CeedScalar val) { 318 for (CeedSize i = start; i < length; i += step) h_array[i] = val; 319 return CEED_ERROR_SUCCESS; 320 } 321 322 //------------------------------------------------------------------------------ 323 // Set device array to value strided (impl in .hip.cpp file) 324 //------------------------------------------------------------------------------ 325 int CeedDeviceSetValueStrided_Hip(CeedScalar *d_array, CeedSize start, CeedSize step, CeedSize length, CeedScalar val); 326 327 //------------------------------------------------------------------------------ 328 // Set a vector to a value strided 329 //------------------------------------------------------------------------------ 330 static int CeedVectorSetValueStrided_Hip(CeedVector vec, CeedSize start, CeedSize step, CeedScalar val) { 331 CeedSize length; 332 CeedVector_Hip *impl; 333 334 CeedCallBackend(CeedVectorGetData(vec, &impl)); 335 CeedCallBackend(CeedVectorGetLength(vec, &length)); 336 // Set value for synced device/host array 337 if (impl->d_array) { 338 CeedCallBackend(CeedDeviceSetValueStrided_Hip(impl->d_array, start, step, length, val)); 339 impl->h_array = NULL; 340 } else if (impl->h_array) { 341 CeedCallBackend(CeedHostSetValueStrided_Hip(impl->h_array, start, step, length, val)); 342 impl->d_array = NULL; 343 } else { 344 return CeedError(CeedVectorReturnCeed(vec), CEED_ERROR_BACKEND, "CeedVector must have valid data set"); 345 } 346 return CEED_ERROR_SUCCESS; 347 } 348 349 //------------------------------------------------------------------------------ 350 // Vector Take Array 351 //------------------------------------------------------------------------------ 352 static int CeedVectorTakeArray_Hip(CeedVector vec, CeedMemType mem_type, CeedScalar **array) { 353 CeedVector_Hip *impl; 354 355 CeedCallBackend(CeedVectorGetData(vec, &impl)); 356 357 // Sync array to requested mem_type 358 CeedCallBackend(CeedVectorSyncArray(vec, mem_type)); 359 360 // Update pointer 361 switch (mem_type) { 362 case CEED_MEM_HOST: 363 (*array) = impl->h_array_borrowed; 364 impl->h_array_borrowed = NULL; 365 impl->h_array = NULL; 366 break; 367 case CEED_MEM_DEVICE: 368 (*array) = impl->d_array_borrowed; 369 impl->d_array_borrowed = NULL; 370 impl->d_array = NULL; 371 break; 372 } 373 return CEED_ERROR_SUCCESS; 374 } 375 376 //------------------------------------------------------------------------------ 377 // Core logic for array syncronization for GetArray. 378 // If a different memory type is most up to date, this will perform a copy 379 //------------------------------------------------------------------------------ 380 static int CeedVectorGetArrayCore_Hip(const CeedVector vec, const CeedMemType mem_type, CeedScalar **array) { 381 CeedVector_Hip *impl; 382 383 CeedCallBackend(CeedVectorGetData(vec, &impl)); 384 385 // Sync array to requested mem_type 386 CeedCallBackend(CeedVectorSyncArray(vec, mem_type)); 387 388 // Update pointer 389 switch (mem_type) { 390 case CEED_MEM_HOST: 391 *array = impl->h_array; 392 break; 393 case CEED_MEM_DEVICE: 394 *array = impl->d_array; 395 break; 396 } 397 return CEED_ERROR_SUCCESS; 398 } 399 400 //------------------------------------------------------------------------------ 401 // Get read-only access to a vector via the specified mem_type 402 //------------------------------------------------------------------------------ 403 static int CeedVectorGetArrayRead_Hip(const CeedVector vec, const CeedMemType mem_type, const CeedScalar **array) { 404 return CeedVectorGetArrayCore_Hip(vec, mem_type, (CeedScalar **)array); 405 } 406 407 //------------------------------------------------------------------------------ 408 // Get read/write access to a vector via the specified mem_type 409 //------------------------------------------------------------------------------ 410 static int CeedVectorGetArray_Hip(const CeedVector vec, const CeedMemType mem_type, CeedScalar **array) { 411 CeedVector_Hip *impl; 412 413 CeedCallBackend(CeedVectorGetData(vec, &impl)); 414 CeedCallBackend(CeedVectorGetArrayCore_Hip(vec, mem_type, array)); 415 CeedCallBackend(CeedVectorSetAllInvalid_Hip(vec)); 416 switch (mem_type) { 417 case CEED_MEM_HOST: 418 impl->h_array = *array; 419 break; 420 case CEED_MEM_DEVICE: 421 impl->d_array = *array; 422 break; 423 } 424 return CEED_ERROR_SUCCESS; 425 } 426 427 //------------------------------------------------------------------------------ 428 // Get write access to a vector via the specified mem_type 429 //------------------------------------------------------------------------------ 430 static int CeedVectorGetArrayWrite_Hip(const CeedVector vec, const CeedMemType mem_type, CeedScalar **array) { 431 bool has_array_of_type = true; 432 CeedVector_Hip *impl; 433 434 CeedCallBackend(CeedVectorGetData(vec, &impl)); 435 CeedCallBackend(CeedVectorHasArrayOfType_Hip(vec, mem_type, &has_array_of_type)); 436 if (!has_array_of_type) { 437 // Allocate if array is not yet allocated 438 CeedCallBackend(CeedVectorSetArray(vec, mem_type, CEED_COPY_VALUES, NULL)); 439 } else { 440 // Select dirty array 441 switch (mem_type) { 442 case CEED_MEM_HOST: 443 if (impl->h_array_borrowed) impl->h_array = impl->h_array_borrowed; 444 else impl->h_array = impl->h_array_owned; 445 break; 446 case CEED_MEM_DEVICE: 447 if (impl->d_array_borrowed) impl->d_array = impl->d_array_borrowed; 448 else impl->d_array = impl->d_array_owned; 449 } 450 } 451 return CeedVectorGetArray_Hip(vec, mem_type, array); 452 } 453 454 //------------------------------------------------------------------------------ 455 // Get the norm of a CeedVector 456 //------------------------------------------------------------------------------ 457 static int CeedVectorNorm_Hip(CeedVector vec, CeedNormType type, CeedScalar *norm) { 458 Ceed ceed; 459 CeedSize length, num_calls; 460 const CeedScalar *d_array; 461 CeedVector_Hip *impl; 462 hipblasHandle_t handle; 463 464 CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 465 CeedCallBackend(CeedVectorGetData(vec, &impl)); 466 CeedCallBackend(CeedVectorGetLength(vec, &length)); 467 CeedCallBackend(CeedGetHipblasHandle_Hip(ceed, &handle)); 468 469 // Is the vector too long to handle with int32? If so, we will divide 470 // it up into "int32-sized" subsections and make repeated BLAS calls. 471 num_calls = length / INT_MAX; 472 if (length % INT_MAX > 0) num_calls += 1; 473 474 // Compute norm 475 CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &d_array)); 476 switch (type) { 477 case CEED_NORM_1: { 478 *norm = 0.0; 479 if (CEED_SCALAR_TYPE == CEED_SCALAR_FP32) { 480 float sub_norm = 0.0; 481 float *d_array_start; 482 483 for (CeedInt i = 0; i < num_calls; i++) { 484 d_array_start = (float *)d_array + (CeedSize)(i)*INT_MAX; 485 CeedSize remaining_length = length - (CeedSize)(i)*INT_MAX; 486 CeedInt sub_length = (i == num_calls - 1) ? (CeedInt)(remaining_length) : INT_MAX; 487 488 CeedCallHipblas(ceed, hipblasSasum(handle, (CeedInt)sub_length, (float *)d_array_start, 1, &sub_norm)); 489 *norm += sub_norm; 490 } 491 } else { 492 double sub_norm = 0.0; 493 double *d_array_start; 494 495 for (CeedInt i = 0; i < num_calls; i++) { 496 d_array_start = (double *)d_array + (CeedSize)(i)*INT_MAX; 497 CeedSize remaining_length = length - (CeedSize)(i)*INT_MAX; 498 CeedInt sub_length = (i == num_calls - 1) ? (CeedInt)(remaining_length) : INT_MAX; 499 500 CeedCallHipblas(ceed, hipblasDasum(handle, (CeedInt)sub_length, (double *)d_array_start, 1, &sub_norm)); 501 *norm += sub_norm; 502 } 503 } 504 break; 505 } 506 case CEED_NORM_2: { 507 if (CEED_SCALAR_TYPE == CEED_SCALAR_FP32) { 508 float sub_norm = 0.0, norm_sum = 0.0; 509 float *d_array_start; 510 511 for (CeedInt i = 0; i < num_calls; i++) { 512 d_array_start = (float *)d_array + (CeedSize)(i)*INT_MAX; 513 CeedSize remaining_length = length - (CeedSize)(i)*INT_MAX; 514 CeedInt sub_length = (i == num_calls - 1) ? (CeedInt)(remaining_length) : INT_MAX; 515 516 CeedCallHipblas(ceed, hipblasSnrm2(handle, (CeedInt)sub_length, (float *)d_array_start, 1, &sub_norm)); 517 norm_sum += sub_norm * sub_norm; 518 } 519 *norm = sqrt(norm_sum); 520 } else { 521 double sub_norm = 0.0, norm_sum = 0.0; 522 double *d_array_start; 523 524 for (CeedInt i = 0; i < num_calls; i++) { 525 d_array_start = (double *)d_array + (CeedSize)(i)*INT_MAX; 526 CeedSize remaining_length = length - (CeedSize)(i)*INT_MAX; 527 CeedInt sub_length = (i == num_calls - 1) ? (CeedInt)(remaining_length) : INT_MAX; 528 529 CeedCallHipblas(ceed, hipblasDnrm2(handle, (CeedInt)sub_length, (double *)d_array_start, 1, &sub_norm)); 530 norm_sum += sub_norm * sub_norm; 531 } 532 *norm = sqrt(norm_sum); 533 } 534 break; 535 } 536 case CEED_NORM_MAX: { 537 CeedInt index; 538 539 if (CEED_SCALAR_TYPE == CEED_SCALAR_FP32) { 540 float sub_max = 0.0, current_max = 0.0; 541 float *d_array_start; 542 for (CeedInt i = 0; i < num_calls; i++) { 543 d_array_start = (float *)d_array + (CeedSize)(i)*INT_MAX; 544 CeedSize remaining_length = length - (CeedSize)(i)*INT_MAX; 545 CeedInt sub_length = (i == num_calls - 1) ? (CeedInt)(remaining_length) : INT_MAX; 546 547 CeedCallHipblas(ceed, hipblasIsamax(handle, (CeedInt)sub_length, (float *)d_array_start, 1, &index)); 548 CeedCallHip(ceed, hipMemcpy(&sub_max, d_array_start + index - 1, sizeof(CeedScalar), hipMemcpyDeviceToHost)); 549 if (fabs(sub_max) > current_max) current_max = fabs(sub_max); 550 } 551 *norm = current_max; 552 } else { 553 double sub_max = 0.0, current_max = 0.0; 554 double *d_array_start; 555 556 for (CeedInt i = 0; i < num_calls; i++) { 557 d_array_start = (double *)d_array + (CeedSize)(i)*INT_MAX; 558 CeedSize remaining_length = length - (CeedSize)(i)*INT_MAX; 559 CeedInt sub_length = (i == num_calls - 1) ? (CeedInt)(remaining_length) : INT_MAX; 560 561 CeedCallHipblas(ceed, hipblasIdamax(handle, (CeedInt)sub_length, (double *)d_array_start, 1, &index)); 562 CeedCallHip(ceed, hipMemcpy(&sub_max, d_array_start + index - 1, sizeof(CeedScalar), hipMemcpyDeviceToHost)); 563 if (fabs(sub_max) > current_max) current_max = fabs(sub_max); 564 } 565 *norm = current_max; 566 } 567 break; 568 } 569 } 570 CeedCallBackend(CeedVectorRestoreArrayRead(vec, &d_array)); 571 return CEED_ERROR_SUCCESS; 572 } 573 574 //------------------------------------------------------------------------------ 575 // Take reciprocal of a vector on host 576 //------------------------------------------------------------------------------ 577 static int CeedHostReciprocal_Hip(CeedScalar *h_array, CeedSize length) { 578 for (CeedSize i = 0; i < length; i++) { 579 if (fabs(h_array[i]) > CEED_EPSILON) h_array[i] = 1. / h_array[i]; 580 } 581 return CEED_ERROR_SUCCESS; 582 } 583 584 //------------------------------------------------------------------------------ 585 // Take reciprocal of a vector on device (impl in .hip.cpp file) 586 //------------------------------------------------------------------------------ 587 int CeedDeviceReciprocal_Hip(CeedScalar *d_array, CeedSize length); 588 589 //------------------------------------------------------------------------------ 590 // Take reciprocal of a vector 591 //------------------------------------------------------------------------------ 592 static int CeedVectorReciprocal_Hip(CeedVector vec) { 593 CeedSize length; 594 CeedVector_Hip *impl; 595 596 CeedCallBackend(CeedVectorGetData(vec, &impl)); 597 CeedCallBackend(CeedVectorGetLength(vec, &length)); 598 // Set value for synced device/host array 599 if (impl->d_array) CeedCallBackend(CeedDeviceReciprocal_Hip(impl->d_array, length)); 600 if (impl->h_array) CeedCallBackend(CeedHostReciprocal_Hip(impl->h_array, length)); 601 return CEED_ERROR_SUCCESS; 602 } 603 604 //------------------------------------------------------------------------------ 605 // Compute x = alpha x on the host 606 //------------------------------------------------------------------------------ 607 static int CeedHostScale_Hip(CeedScalar *x_array, CeedScalar alpha, CeedSize length) { 608 for (CeedSize i = 0; i < length; i++) x_array[i] *= alpha; 609 return CEED_ERROR_SUCCESS; 610 } 611 612 //------------------------------------------------------------------------------ 613 // Compute x = alpha x on device (impl in .hip.cpp file) 614 //------------------------------------------------------------------------------ 615 int CeedDeviceScale_Hip(CeedScalar *x_array, CeedScalar alpha, CeedSize length); 616 617 //------------------------------------------------------------------------------ 618 // Compute x = alpha x 619 //------------------------------------------------------------------------------ 620 static int CeedVectorScale_Hip(CeedVector x, CeedScalar alpha) { 621 CeedSize length; 622 CeedVector_Hip *x_impl; 623 624 CeedCallBackend(CeedVectorGetData(x, &x_impl)); 625 CeedCallBackend(CeedVectorGetLength(x, &length)); 626 // Set value for synced device/host array 627 if (x_impl->d_array) CeedCallBackend(CeedDeviceScale_Hip(x_impl->d_array, alpha, length)); 628 if (x_impl->h_array) CeedCallBackend(CeedHostScale_Hip(x_impl->h_array, alpha, length)); 629 return CEED_ERROR_SUCCESS; 630 } 631 632 //------------------------------------------------------------------------------ 633 // Compute y = alpha x + y on the host 634 //------------------------------------------------------------------------------ 635 static int CeedHostAXPY_Hip(CeedScalar *y_array, CeedScalar alpha, CeedScalar *x_array, CeedSize length) { 636 for (CeedSize i = 0; i < length; i++) y_array[i] += alpha * x_array[i]; 637 return CEED_ERROR_SUCCESS; 638 } 639 640 //------------------------------------------------------------------------------ 641 // Compute y = alpha x + y on device (impl in .hip.cpp file) 642 //------------------------------------------------------------------------------ 643 int CeedDeviceAXPY_Hip(CeedScalar *y_array, CeedScalar alpha, CeedScalar *x_array, CeedSize length); 644 645 //------------------------------------------------------------------------------ 646 // Compute y = alpha x + y 647 //------------------------------------------------------------------------------ 648 static int CeedVectorAXPY_Hip(CeedVector y, CeedScalar alpha, CeedVector x) { 649 CeedSize length; 650 CeedVector_Hip *y_impl, *x_impl; 651 652 CeedCallBackend(CeedVectorGetData(y, &y_impl)); 653 CeedCallBackend(CeedVectorGetData(x, &x_impl)); 654 CeedCallBackend(CeedVectorGetLength(y, &length)); 655 // Set value for synced device/host array 656 if (y_impl->d_array) { 657 CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_DEVICE)); 658 CeedCallBackend(CeedDeviceAXPY_Hip(y_impl->d_array, alpha, x_impl->d_array, length)); 659 } 660 if (y_impl->h_array) { 661 CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_HOST)); 662 CeedCallBackend(CeedHostAXPY_Hip(y_impl->h_array, alpha, x_impl->h_array, length)); 663 } 664 return CEED_ERROR_SUCCESS; 665 } 666 667 //------------------------------------------------------------------------------ 668 // Compute y = alpha x + beta y on the host 669 //------------------------------------------------------------------------------ 670 static int CeedHostAXPBY_Hip(CeedScalar *y_array, CeedScalar alpha, CeedScalar beta, CeedScalar *x_array, CeedSize length) { 671 for (CeedSize i = 0; i < length; i++) y_array[i] = alpha * x_array[i] + beta * y_array[i]; 672 return CEED_ERROR_SUCCESS; 673 } 674 675 //------------------------------------------------------------------------------ 676 // Compute y = alpha x + beta y on device (impl in .hip.cpp file) 677 //------------------------------------------------------------------------------ 678 int CeedDeviceAXPBY_Hip(CeedScalar *y_array, CeedScalar alpha, CeedScalar beta, CeedScalar *x_array, CeedSize length); 679 680 //------------------------------------------------------------------------------ 681 // Compute y = alpha x + beta y 682 //------------------------------------------------------------------------------ 683 static int CeedVectorAXPBY_Hip(CeedVector y, CeedScalar alpha, CeedScalar beta, CeedVector x) { 684 CeedSize length; 685 CeedVector_Hip *y_impl, *x_impl; 686 687 CeedCallBackend(CeedVectorGetData(y, &y_impl)); 688 CeedCallBackend(CeedVectorGetData(x, &x_impl)); 689 CeedCallBackend(CeedVectorGetLength(y, &length)); 690 // Set value for synced device/host array 691 if (y_impl->d_array) { 692 CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_DEVICE)); 693 CeedCallBackend(CeedDeviceAXPBY_Hip(y_impl->d_array, alpha, beta, x_impl->d_array, length)); 694 } 695 if (y_impl->h_array) { 696 CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_HOST)); 697 CeedCallBackend(CeedHostAXPBY_Hip(y_impl->h_array, alpha, beta, x_impl->h_array, length)); 698 } 699 return CEED_ERROR_SUCCESS; 700 } 701 702 //------------------------------------------------------------------------------ 703 // Compute the pointwise multiplication w = x .* y on the host 704 //------------------------------------------------------------------------------ 705 static int CeedHostPointwiseMult_Hip(CeedScalar *w_array, CeedScalar *x_array, CeedScalar *y_array, CeedSize length) { 706 for (CeedSize i = 0; i < length; i++) w_array[i] = x_array[i] * y_array[i]; 707 return CEED_ERROR_SUCCESS; 708 } 709 710 //------------------------------------------------------------------------------ 711 // Compute the pointwise multiplication w = x .* y on device (impl in .hip.cpp file) 712 //------------------------------------------------------------------------------ 713 int CeedDevicePointwiseMult_Hip(CeedScalar *w_array, CeedScalar *x_array, CeedScalar *y_array, CeedSize length); 714 715 //------------------------------------------------------------------------------ 716 // Compute the pointwise multiplication w = x .* y 717 //------------------------------------------------------------------------------ 718 static int CeedVectorPointwiseMult_Hip(CeedVector w, CeedVector x, CeedVector y) { 719 CeedSize length; 720 CeedVector_Hip *w_impl, *x_impl, *y_impl; 721 722 CeedCallBackend(CeedVectorGetData(w, &w_impl)); 723 CeedCallBackend(CeedVectorGetData(x, &x_impl)); 724 CeedCallBackend(CeedVectorGetData(y, &y_impl)); 725 CeedCallBackend(CeedVectorGetLength(w, &length)); 726 727 // Set value for synced device/host array 728 if (!w_impl->d_array && !w_impl->h_array) { 729 CeedCallBackend(CeedVectorSetValue(w, 0.0)); 730 } 731 if (w_impl->d_array) { 732 CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_DEVICE)); 733 CeedCallBackend(CeedVectorSyncArray(y, CEED_MEM_DEVICE)); 734 CeedCallBackend(CeedDevicePointwiseMult_Hip(w_impl->d_array, x_impl->d_array, y_impl->d_array, length)); 735 } 736 if (w_impl->h_array) { 737 CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_HOST)); 738 CeedCallBackend(CeedVectorSyncArray(y, CEED_MEM_HOST)); 739 CeedCallBackend(CeedHostPointwiseMult_Hip(w_impl->h_array, x_impl->h_array, y_impl->h_array, length)); 740 } 741 return CEED_ERROR_SUCCESS; 742 } 743 744 //------------------------------------------------------------------------------ 745 // Destroy the vector 746 //------------------------------------------------------------------------------ 747 static int CeedVectorDestroy_Hip(const CeedVector vec) { 748 CeedVector_Hip *impl; 749 750 CeedCallBackend(CeedVectorGetData(vec, &impl)); 751 CeedCallHip(CeedVectorReturnCeed(vec), hipFree(impl->d_array_owned)); 752 CeedCallBackend(CeedFree(&impl->h_array_owned)); 753 CeedCallBackend(CeedFree(&impl)); 754 return CEED_ERROR_SUCCESS; 755 } 756 757 //------------------------------------------------------------------------------ 758 // Create a vector of the specified length (does not allocate memory) 759 //------------------------------------------------------------------------------ 760 int CeedVectorCreate_Hip(CeedSize n, CeedVector vec) { 761 CeedVector_Hip *impl; 762 Ceed ceed; 763 764 CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 765 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "HasValidArray", CeedVectorHasValidArray_Hip)); 766 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "HasBorrowedArrayOfType", CeedVectorHasBorrowedArrayOfType_Hip)); 767 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "SetArray", CeedVectorSetArray_Hip)); 768 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "TakeArray", CeedVectorTakeArray_Hip)); 769 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "CopyStrided", (int (*)())CeedVectorCopyStrided_Hip)); 770 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "SetValue", (int (*)())CeedVectorSetValue_Hip)); 771 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "SetValueStrided", (int (*)())CeedVectorSetValueStrided_Hip)); 772 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "SyncArray", CeedVectorSyncArray_Hip)); 773 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "GetArray", CeedVectorGetArray_Hip)); 774 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "GetArrayRead", CeedVectorGetArrayRead_Hip)); 775 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "GetArrayWrite", CeedVectorGetArrayWrite_Hip)); 776 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "Norm", CeedVectorNorm_Hip)); 777 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "Reciprocal", CeedVectorReciprocal_Hip)); 778 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "Scale", (int (*)())CeedVectorScale_Hip)); 779 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "AXPY", (int (*)())CeedVectorAXPY_Hip)); 780 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "AXPBY", (int (*)())CeedVectorAXPBY_Hip)); 781 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "PointwiseMult", CeedVectorPointwiseMult_Hip)); 782 CeedCallBackend(CeedSetBackendFunction(ceed, "Vector", vec, "Destroy", CeedVectorDestroy_Hip)); 783 CeedCallBackend(CeedCalloc(1, &impl)); 784 CeedCallBackend(CeedVectorSetData(vec, impl)); 785 return CEED_ERROR_SUCCESS; 786 } 787 788 //------------------------------------------------------------------------------ 789