1bd882c8aSJames Wright // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2bd882c8aSJames Wright // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3bd882c8aSJames Wright // 4bd882c8aSJames Wright // SPDX-License-Identifier: BSD-2-Clause 5bd882c8aSJames Wright // 6bd882c8aSJames Wright // This file is part of CEED: http://github.com/ceed 7bd882c8aSJames Wright 8bd882c8aSJames Wright #include <ceed/backend.h> 9bd882c8aSJames Wright #include <ceed/ceed.h> 10bd882c8aSJames Wright 11bd882c8aSJames Wright #include <cmath> 12bd882c8aSJames Wright #include <string> 13bd882c8aSJames Wright #include <sycl/sycl.hpp> 14bd882c8aSJames Wright 15bd882c8aSJames Wright #include "ceed-sycl-ref.hpp" 16bd882c8aSJames Wright 17bd882c8aSJames Wright //------------------------------------------------------------------------------ 18bd882c8aSJames Wright // Check if host/device sync is needed 19bd882c8aSJames Wright //------------------------------------------------------------------------------ 20bd882c8aSJames Wright static inline int CeedVectorNeedSync_Sycl(const CeedVector vec, CeedMemType mem_type, bool *need_sync) { 21bd882c8aSJames Wright bool has_valid_array = false; 22dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 23dd64fc84SJeremy L Thompson 24dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 25bd882c8aSJames Wright CeedCallBackend(CeedVectorHasValidArray(vec, &has_valid_array)); 26bd882c8aSJames Wright switch (mem_type) { 27bd882c8aSJames Wright case CEED_MEM_HOST: 28bd882c8aSJames Wright *need_sync = has_valid_array && !impl->h_array; 29bd882c8aSJames Wright break; 30bd882c8aSJames Wright case CEED_MEM_DEVICE: 31bd882c8aSJames Wright *need_sync = has_valid_array && !impl->d_array; 32bd882c8aSJames Wright break; 33bd882c8aSJames Wright } 34bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 35bd882c8aSJames Wright } 36bd882c8aSJames Wright 37bd882c8aSJames Wright //------------------------------------------------------------------------------ 38bd882c8aSJames Wright // Sync host to device 39bd882c8aSJames Wright //------------------------------------------------------------------------------ 40bd882c8aSJames Wright static inline int CeedVectorSyncH2D_Sycl(const CeedVector vec) { 41bd882c8aSJames Wright Ceed ceed; 42bd882c8aSJames Wright Ceed_Sycl *data; 43dd64fc84SJeremy L Thompson CeedSize length; 44dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 45dd64fc84SJeremy L Thompson 46dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 47dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 48bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 49*4e3038a5SJeremy L Thompson CeedCheck(impl->h_array, ceed, CEED_ERROR_BACKEND, "No valid host data to sync to device"); 50bd882c8aSJames Wright 51bd882c8aSJames Wright CeedCallBackend(CeedVectorGetLength(vec, &length)); 52bd882c8aSJames Wright if (impl->d_array_borrowed) { 53bd882c8aSJames Wright impl->d_array = impl->d_array_borrowed; 54bd882c8aSJames Wright } else if (impl->d_array_owned) { 55bd882c8aSJames Wright impl->d_array = impl->d_array_owned; 56bd882c8aSJames Wright } else { 57bd882c8aSJames Wright CeedCallSycl(ceed, impl->d_array_owned = sycl::malloc_device<CeedScalar>(length, data->sycl_device, data->sycl_context)); 58bd882c8aSJames Wright impl->d_array = impl->d_array_owned; 59bd882c8aSJames Wright } 60bd882c8aSJames Wright 61bd882c8aSJames Wright sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 62bd882c8aSJames Wright // Copy from host to device 63bd882c8aSJames Wright sycl::event copy_event = data->sycl_queue.copy<CeedScalar>(impl->h_array, impl->d_array, length, {e}); 64bd882c8aSJames Wright // Wait for copy to finish and handle exceptions. 65bd882c8aSJames Wright CeedCallSycl(ceed, copy_event.wait_and_throw()); 66bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 67bd882c8aSJames Wright } 68bd882c8aSJames Wright 69bd882c8aSJames Wright //------------------------------------------------------------------------------ 70bd882c8aSJames Wright // Sync device to host 71bd882c8aSJames Wright //------------------------------------------------------------------------------ 72bd882c8aSJames Wright static inline int CeedVectorSyncD2H_Sycl(const CeedVector vec) { 73bd882c8aSJames Wright Ceed ceed; 74bd882c8aSJames Wright Ceed_Sycl *data; 75dd64fc84SJeremy L Thompson CeedSize length; 76dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 77dd64fc84SJeremy L Thompson 78dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 79dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 80bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 81bd882c8aSJames Wright 82bd882c8aSJames Wright CeedCheck(impl->d_array, ceed, CEED_ERROR_BACKEND, "No valid device data to sync to host"); 83bd882c8aSJames Wright 84bd882c8aSJames Wright CeedCallBackend(CeedVectorGetLength(vec, &length)); 85bd882c8aSJames Wright if (impl->h_array_borrowed) { 86bd882c8aSJames Wright impl->h_array = impl->h_array_borrowed; 87bd882c8aSJames Wright } else if (impl->h_array_owned) { 88bd882c8aSJames Wright impl->h_array = impl->h_array_owned; 89bd882c8aSJames Wright } else { 90bd882c8aSJames Wright CeedCallBackend(CeedCalloc(length, &impl->h_array_owned)); 91bd882c8aSJames Wright impl->h_array = impl->h_array_owned; 92bd882c8aSJames Wright } 93bd882c8aSJames Wright 94bd882c8aSJames Wright // Order queue 95bd882c8aSJames Wright sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 96bd882c8aSJames Wright // Copy from device to host 97bd882c8aSJames Wright sycl::event copy_event = data->sycl_queue.copy<CeedScalar>(impl->d_array, impl->h_array, length, {e}); 98bd882c8aSJames Wright // Wait for copy to finish and handle exceptions. 99bd882c8aSJames Wright CeedCallSycl(ceed, copy_event.wait_and_throw()); 100bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 101bd882c8aSJames Wright } 102bd882c8aSJames Wright 103bd882c8aSJames Wright //------------------------------------------------------------------------------ 104bd882c8aSJames Wright // Sync arrays 105bd882c8aSJames Wright //------------------------------------------------------------------------------ 106bd882c8aSJames Wright static int CeedVectorSyncArray_Sycl(const CeedVector vec, CeedMemType mem_type) { 107bd882c8aSJames Wright bool need_sync = false; 108dd64fc84SJeremy L Thompson 109dd64fc84SJeremy L Thompson // Check whether device/host sync is needed 110bd882c8aSJames Wright CeedCallBackend(CeedVectorNeedSync_Sycl(vec, mem_type, &need_sync)); 111bd882c8aSJames Wright if (!need_sync) return CEED_ERROR_SUCCESS; 112bd882c8aSJames Wright 113bd882c8aSJames Wright switch (mem_type) { 114bd882c8aSJames Wright case CEED_MEM_HOST: 115bd882c8aSJames Wright return CeedVectorSyncD2H_Sycl(vec); 116bd882c8aSJames Wright case CEED_MEM_DEVICE: 117bd882c8aSJames Wright return CeedVectorSyncH2D_Sycl(vec); 118bd882c8aSJames Wright } 119bd882c8aSJames Wright return CEED_ERROR_UNSUPPORTED; 120bd882c8aSJames Wright } 121bd882c8aSJames Wright 122bd882c8aSJames Wright //------------------------------------------------------------------------------ 123bd882c8aSJames Wright // Set all pointers as invalid 124bd882c8aSJames Wright //------------------------------------------------------------------------------ 125bd882c8aSJames Wright static inline int CeedVectorSetAllInvalid_Sycl(const CeedVector vec) { 126bd882c8aSJames Wright CeedVector_Sycl *impl; 127bd882c8aSJames Wright 128dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 129bd882c8aSJames Wright impl->h_array = NULL; 130bd882c8aSJames Wright impl->d_array = NULL; 131bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 132bd882c8aSJames Wright } 133bd882c8aSJames Wright 134bd882c8aSJames Wright //------------------------------------------------------------------------------ 135bd882c8aSJames Wright // Check if CeedVector has any valid pointer 136bd882c8aSJames Wright //------------------------------------------------------------------------------ 137bd882c8aSJames Wright static inline int CeedVectorHasValidArray_Sycl(const CeedVector vec, bool *has_valid_array) { 138bd882c8aSJames Wright CeedVector_Sycl *impl; 139dd64fc84SJeremy L Thompson 140bd882c8aSJames Wright CeedCallBackend(CeedVectorGetData(vec, &impl)); 1411c66c397SJeremy L Thompson *has_valid_array = impl->h_array || impl->d_array; 142bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 143bd882c8aSJames Wright } 144bd882c8aSJames Wright 145bd882c8aSJames Wright //------------------------------------------------------------------------------ 146bd882c8aSJames Wright // Check if has array of given type 147bd882c8aSJames Wright //------------------------------------------------------------------------------ 148bd882c8aSJames Wright static inline int CeedVectorHasArrayOfType_Sycl(const CeedVector vec, CeedMemType mem_type, bool *has_array_of_type) { 149bd882c8aSJames Wright CeedVector_Sycl *impl; 150bd882c8aSJames Wright 151dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 152bd882c8aSJames Wright switch (mem_type) { 153bd882c8aSJames Wright case CEED_MEM_HOST: 1541c66c397SJeremy L Thompson *has_array_of_type = impl->h_array_borrowed || impl->h_array_owned; 155bd882c8aSJames Wright break; 156bd882c8aSJames Wright case CEED_MEM_DEVICE: 1571c66c397SJeremy L Thompson *has_array_of_type = impl->d_array_borrowed || impl->d_array_owned; 158bd882c8aSJames Wright break; 159bd882c8aSJames Wright } 160bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 161bd882c8aSJames Wright } 162bd882c8aSJames Wright 163bd882c8aSJames Wright //------------------------------------------------------------------------------ 164bd882c8aSJames Wright // Check if has borrowed array of given type 165bd882c8aSJames Wright //------------------------------------------------------------------------------ 166bd882c8aSJames Wright static inline int CeedVectorHasBorrowedArrayOfType_Sycl(const CeedVector vec, CeedMemType mem_type, bool *has_borrowed_array_of_type) { 167bd882c8aSJames Wright CeedVector_Sycl *impl; 168bd882c8aSJames Wright 169dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 170bd882c8aSJames Wright switch (mem_type) { 171bd882c8aSJames Wright case CEED_MEM_HOST: 1721c66c397SJeremy L Thompson *has_borrowed_array_of_type = impl->h_array_borrowed; 173bd882c8aSJames Wright break; 174bd882c8aSJames Wright case CEED_MEM_DEVICE: 1751c66c397SJeremy L Thompson *has_borrowed_array_of_type = impl->d_array_borrowed; 176bd882c8aSJames Wright break; 177bd882c8aSJames Wright } 178bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 179bd882c8aSJames Wright } 180bd882c8aSJames Wright 181bd882c8aSJames Wright //------------------------------------------------------------------------------ 182bd882c8aSJames Wright // Set array from host 183bd882c8aSJames Wright //------------------------------------------------------------------------------ 184bd882c8aSJames Wright static int CeedVectorSetArrayHost_Sycl(const CeedVector vec, const CeedCopyMode copy_mode, CeedScalar *array) { 185bd882c8aSJames Wright CeedVector_Sycl *impl; 186bd882c8aSJames Wright 187dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 188bd882c8aSJames Wright switch (copy_mode) { 189bd882c8aSJames Wright case CEED_COPY_VALUES: { 190bd882c8aSJames Wright if (!impl->h_array_owned) { 191bd882c8aSJames Wright CeedSize length; 192dd64fc84SJeremy L Thompson 193bd882c8aSJames Wright CeedCallBackend(CeedVectorGetLength(vec, &length)); 194bd882c8aSJames Wright CeedCallBackend(CeedMalloc(length, &impl->h_array_owned)); 195bd882c8aSJames Wright } 196bd882c8aSJames Wright impl->h_array_borrowed = NULL; 197bd882c8aSJames Wright impl->h_array = impl->h_array_owned; 198bd882c8aSJames Wright if (array) { 199bd882c8aSJames Wright CeedSize length; 200672b0f2aSSebastian Grimberg size_t bytes; 201dd64fc84SJeremy L Thompson 202bd882c8aSJames Wright CeedCallBackend(CeedVectorGetLength(vec, &length)); 203672b0f2aSSebastian Grimberg bytes = length * sizeof(CeedScalar); 204bd882c8aSJames Wright memcpy(impl->h_array, array, bytes); 205bd882c8aSJames Wright } 206bd882c8aSJames Wright } break; 207bd882c8aSJames Wright case CEED_OWN_POINTER: 208bd882c8aSJames Wright CeedCallBackend(CeedFree(&impl->h_array_owned)); 209bd882c8aSJames Wright impl->h_array_owned = array; 210bd882c8aSJames Wright impl->h_array_borrowed = NULL; 211bd882c8aSJames Wright impl->h_array = array; 212bd882c8aSJames Wright break; 213bd882c8aSJames Wright case CEED_USE_POINTER: 214bd882c8aSJames Wright CeedCallBackend(CeedFree(&impl->h_array_owned)); 215bd882c8aSJames Wright impl->h_array_borrowed = array; 216bd882c8aSJames Wright impl->h_array = array; 217bd882c8aSJames Wright break; 218bd882c8aSJames Wright } 219bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 220bd882c8aSJames Wright } 221bd882c8aSJames Wright 222bd882c8aSJames Wright //------------------------------------------------------------------------------ 223bd882c8aSJames Wright // Set array from device 224bd882c8aSJames Wright //------------------------------------------------------------------------------ 225bd882c8aSJames Wright static int CeedVectorSetArrayDevice_Sycl(const CeedVector vec, const CeedCopyMode copy_mode, CeedScalar *array) { 226bd882c8aSJames Wright Ceed ceed; 227bd882c8aSJames Wright Ceed_Sycl *data; 228dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 229dd64fc84SJeremy L Thompson 230dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 231dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 232bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 233bd882c8aSJames Wright 234bd882c8aSJames Wright // Order queue 235bd882c8aSJames Wright sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 236bd882c8aSJames Wright 237bd882c8aSJames Wright switch (copy_mode) { 238bd882c8aSJames Wright case CEED_COPY_VALUES: { 239bd882c8aSJames Wright CeedSize length; 240dd64fc84SJeremy L Thompson 241bd882c8aSJames Wright CeedCallBackend(CeedVectorGetLength(vec, &length)); 242bd882c8aSJames Wright if (!impl->d_array_owned) { 243bd882c8aSJames Wright CeedCallSycl(ceed, impl->d_array_owned = sycl::malloc_device<CeedScalar>(length, data->sycl_device, data->sycl_context)); 244bd882c8aSJames Wright } 2453ce2313bSJeremy L Thompson impl->d_array = impl->d_array_owned; 246bd882c8aSJames Wright if (array) { 247bd882c8aSJames Wright sycl::event copy_event = data->sycl_queue.copy<CeedScalar>(array, impl->d_array, length, {e}); 248bd882c8aSJames Wright // Wait for copy to finish and handle exceptions. 249bd882c8aSJames Wright CeedCallSycl(ceed, copy_event.wait_and_throw()); 250bd882c8aSJames Wright } 251bd882c8aSJames Wright } break; 252bd882c8aSJames Wright case CEED_OWN_POINTER: 253bd882c8aSJames Wright if (impl->d_array_owned) { 254bd882c8aSJames Wright // Wait for all work to finish before freeing memory 255bd882c8aSJames Wright CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 256bd882c8aSJames Wright CeedCallSycl(ceed, sycl::free(impl->d_array_owned, data->sycl_context)); 257bd882c8aSJames Wright } 258bd882c8aSJames Wright impl->d_array_owned = array; 259bd882c8aSJames Wright impl->d_array_borrowed = NULL; 260bd882c8aSJames Wright impl->d_array = array; 261bd882c8aSJames Wright break; 262bd882c8aSJames Wright case CEED_USE_POINTER: 263bd882c8aSJames Wright if (impl->d_array_owned) { 264bd882c8aSJames Wright // Wait for all work to finish before freeing memory 265bd882c8aSJames Wright CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 266bd882c8aSJames Wright CeedCallSycl(ceed, sycl::free(impl->d_array_owned, data->sycl_context)); 267bd882c8aSJames Wright } 268bd882c8aSJames Wright impl->d_array_owned = NULL; 269bd882c8aSJames Wright impl->d_array_borrowed = array; 270bd882c8aSJames Wright impl->d_array = array; 271bd882c8aSJames Wright break; 272bd882c8aSJames Wright } 273bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 274bd882c8aSJames Wright } 275bd882c8aSJames Wright 276bd882c8aSJames Wright //------------------------------------------------------------------------------ 277bd882c8aSJames Wright // Set the array used by a vector, 278bd882c8aSJames Wright // freeing any previously allocated array if applicable 279bd882c8aSJames Wright //------------------------------------------------------------------------------ 280bd882c8aSJames Wright static int CeedVectorSetArray_Sycl(const CeedVector vec, const CeedMemType mem_type, const CeedCopyMode copy_mode, CeedScalar *array) { 281bd882c8aSJames Wright Ceed ceed; 282bd882c8aSJames Wright CeedVector_Sycl *impl; 283dd64fc84SJeremy L Thompson 284dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 285bd882c8aSJames Wright CeedCallBackend(CeedVectorGetData(vec, &impl)); 286bd882c8aSJames Wright 287bd882c8aSJames Wright CeedCallBackend(CeedVectorSetAllInvalid_Sycl(vec)); 288bd882c8aSJames Wright switch (mem_type) { 289bd882c8aSJames Wright case CEED_MEM_HOST: 290bd882c8aSJames Wright return CeedVectorSetArrayHost_Sycl(vec, copy_mode, array); 291bd882c8aSJames Wright case CEED_MEM_DEVICE: 292bd882c8aSJames Wright return CeedVectorSetArrayDevice_Sycl(vec, copy_mode, array); 293bd882c8aSJames Wright } 294bd882c8aSJames Wright return CEED_ERROR_UNSUPPORTED; 295bd882c8aSJames Wright } 296bd882c8aSJames Wright 297bd882c8aSJames Wright //------------------------------------------------------------------------------ 298bd882c8aSJames Wright // Set host array to value 299bd882c8aSJames Wright //------------------------------------------------------------------------------ 3006ca0f394SUmesh Unnikrishnan static int CeedHostSetValue_Sycl(CeedScalar *h_array, CeedSize length, CeedScalar val) { 3016ca0f394SUmesh Unnikrishnan for (CeedSize i = 0; i < length; i++) h_array[i] = val; 302bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 303bd882c8aSJames Wright } 304bd882c8aSJames Wright 305bd882c8aSJames Wright //------------------------------------------------------------------------------ 306bd882c8aSJames Wright // Set device array to value 307bd882c8aSJames Wright //------------------------------------------------------------------------------ 3086ca0f394SUmesh Unnikrishnan static int CeedDeviceSetValue_Sycl(sycl::queue &sycl_queue, CeedScalar *d_array, CeedSize length, CeedScalar val) { 309bd882c8aSJames Wright // Order queue 310bd882c8aSJames Wright sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 311bd882c8aSJames Wright sycl_queue.fill(d_array, val, length, {e}); 312bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 313bd882c8aSJames Wright } 314bd882c8aSJames Wright 315bd882c8aSJames Wright //------------------------------------------------------------------------------ 316bd882c8aSJames Wright // Set a vector to a value, 317bd882c8aSJames Wright //------------------------------------------------------------------------------ 318bd882c8aSJames Wright static int CeedVectorSetValue_Sycl(CeedVector vec, CeedScalar val) { 319bd882c8aSJames Wright Ceed ceed; 320bd882c8aSJames Wright Ceed_Sycl *data; 321dd64fc84SJeremy L Thompson CeedSize length; 322dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 323dd64fc84SJeremy L Thompson 324dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 325dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 326dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetLength(vec, &length)); 327bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 328bd882c8aSJames Wright 329bd882c8aSJames Wright // Set value for synced device/host array 330bd882c8aSJames Wright if (!impl->d_array && !impl->h_array) { 331bd882c8aSJames Wright if (impl->d_array_borrowed) { 332bd882c8aSJames Wright impl->d_array = impl->d_array_borrowed; 333bd882c8aSJames Wright } else if (impl->h_array_borrowed) { 334bd882c8aSJames Wright impl->h_array = impl->h_array_borrowed; 335bd882c8aSJames Wright } else if (impl->d_array_owned) { 336bd882c8aSJames Wright impl->d_array = impl->d_array_owned; 337bd882c8aSJames Wright } else if (impl->h_array_owned) { 338bd882c8aSJames Wright impl->h_array = impl->h_array_owned; 339bd882c8aSJames Wright } else { 340bd882c8aSJames Wright CeedCallBackend(CeedVectorSetArray(vec, CEED_MEM_DEVICE, CEED_COPY_VALUES, NULL)); 341bd882c8aSJames Wright } 342bd882c8aSJames Wright } 343bd882c8aSJames Wright if (impl->d_array) { 344bd882c8aSJames Wright CeedCallBackend(CeedDeviceSetValue_Sycl(data->sycl_queue, impl->d_array, length, val)); 345bd882c8aSJames Wright impl->h_array = NULL; 346bd882c8aSJames Wright } 347bd882c8aSJames Wright if (impl->h_array) { 348bd882c8aSJames Wright CeedCallBackend(CeedHostSetValue_Sycl(impl->h_array, length, val)); 349bd882c8aSJames Wright impl->d_array = NULL; 350bd882c8aSJames Wright } 351bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 352bd882c8aSJames Wright } 353bd882c8aSJames Wright 354bd882c8aSJames Wright //------------------------------------------------------------------------------ 355bd882c8aSJames Wright // Vector Take Array 356bd882c8aSJames Wright //------------------------------------------------------------------------------ 357bd882c8aSJames Wright static int CeedVectorTakeArray_Sycl(CeedVector vec, CeedMemType mem_type, CeedScalar **array) { 358bd882c8aSJames Wright Ceed ceed; 359bd882c8aSJames Wright Ceed_Sycl *data; 360dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 361dd64fc84SJeremy L Thompson 362dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 363dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 364bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 365bd882c8aSJames Wright 366bd882c8aSJames Wright // Order queue 367bd882c8aSJames Wright data->sycl_queue.ext_oneapi_submit_barrier(); 368bd882c8aSJames Wright 369bd882c8aSJames Wright // Sync array to requested mem_type 370bd882c8aSJames Wright CeedCallBackend(CeedVectorSyncArray(vec, mem_type)); 371bd882c8aSJames Wright 372bd882c8aSJames Wright // Update pointer 373bd882c8aSJames Wright switch (mem_type) { 374bd882c8aSJames Wright case CEED_MEM_HOST: 375bd882c8aSJames Wright (*array) = impl->h_array_borrowed; 376bd882c8aSJames Wright impl->h_array_borrowed = NULL; 377bd882c8aSJames Wright impl->h_array = NULL; 378bd882c8aSJames Wright break; 379bd882c8aSJames Wright case CEED_MEM_DEVICE: 380bd882c8aSJames Wright (*array) = impl->d_array_borrowed; 381bd882c8aSJames Wright impl->d_array_borrowed = NULL; 382bd882c8aSJames Wright impl->d_array = NULL; 383bd882c8aSJames Wright break; 384bd882c8aSJames Wright } 385bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 386bd882c8aSJames Wright } 387bd882c8aSJames Wright 388bd882c8aSJames Wright //------------------------------------------------------------------------------ 389bd882c8aSJames Wright // Core logic for array syncronization for GetArray. 390bd882c8aSJames Wright // If a different memory type is most up to date, this will perform a copy 391bd882c8aSJames Wright //------------------------------------------------------------------------------ 392bd882c8aSJames Wright static int CeedVectorGetArrayCore_Sycl(const CeedVector vec, const CeedMemType mem_type, CeedScalar **array) { 393bd882c8aSJames Wright Ceed ceed; 394bd882c8aSJames Wright CeedVector_Sycl *impl; 395dd64fc84SJeremy L Thompson 396dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 397bd882c8aSJames Wright CeedCallBackend(CeedVectorGetData(vec, &impl)); 398bd882c8aSJames Wright 399bd882c8aSJames Wright // Sync array to requested mem_type 400bd882c8aSJames Wright CeedCallBackend(CeedVectorSyncArray(vec, mem_type)); 401bd882c8aSJames Wright 402bd882c8aSJames Wright // Update pointer 403bd882c8aSJames Wright switch (mem_type) { 404bd882c8aSJames Wright case CEED_MEM_HOST: 405bd882c8aSJames Wright *array = impl->h_array; 406bd882c8aSJames Wright break; 407bd882c8aSJames Wright case CEED_MEM_DEVICE: 408bd882c8aSJames Wright *array = impl->d_array; 409bd882c8aSJames Wright break; 410bd882c8aSJames Wright } 411bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 412bd882c8aSJames Wright } 413ff1e7120SSebastian Grimberg 414bd882c8aSJames Wright //------------------------------------------------------------------------------ 415bd882c8aSJames Wright // Get read-only access to a vector via the specified mem_type 416bd882c8aSJames Wright //------------------------------------------------------------------------------ 417bd882c8aSJames Wright static int CeedVectorGetArrayRead_Sycl(const CeedVector vec, const CeedMemType mem_type, const CeedScalar **array) { 418bd882c8aSJames Wright return CeedVectorGetArrayCore_Sycl(vec, mem_type, (CeedScalar **)array); 419bd882c8aSJames Wright } 420bd882c8aSJames Wright 421bd882c8aSJames Wright //------------------------------------------------------------------------------ 422bd882c8aSJames Wright // Get read/write access to a vector via the specified mem_type 423bd882c8aSJames Wright //------------------------------------------------------------------------------ 424bd882c8aSJames Wright static int CeedVectorGetArray_Sycl(const CeedVector vec, const CeedMemType mem_type, CeedScalar **array) { 425bd882c8aSJames Wright CeedVector_Sycl *impl; 426dd64fc84SJeremy L Thompson 427bd882c8aSJames Wright CeedCallBackend(CeedVectorGetData(vec, &impl)); 428bd882c8aSJames Wright CeedCallBackend(CeedVectorGetArrayCore_Sycl(vec, mem_type, array)); 429bd882c8aSJames Wright CeedCallBackend(CeedVectorSetAllInvalid_Sycl(vec)); 430bd882c8aSJames Wright switch (mem_type) { 431bd882c8aSJames Wright case CEED_MEM_HOST: 432bd882c8aSJames Wright impl->h_array = *array; 433bd882c8aSJames Wright break; 434bd882c8aSJames Wright case CEED_MEM_DEVICE: 435bd882c8aSJames Wright impl->d_array = *array; 436bd882c8aSJames Wright break; 437bd882c8aSJames Wright } 438bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 439bd882c8aSJames Wright } 440bd882c8aSJames Wright 441bd882c8aSJames Wright //------------------------------------------------------------------------------ 442bd882c8aSJames Wright // Get write access to a vector via the specified mem_type 443bd882c8aSJames Wright //------------------------------------------------------------------------------ 444bd882c8aSJames Wright static int CeedVectorGetArrayWrite_Sycl(const CeedVector vec, const CeedMemType mem_type, CeedScalar **array) { 445bd882c8aSJames Wright bool has_array_of_type = true; 446dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 447dd64fc84SJeremy L Thompson 448dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 449bd882c8aSJames Wright CeedCallBackend(CeedVectorHasArrayOfType_Sycl(vec, mem_type, &has_array_of_type)); 450bd882c8aSJames Wright if (!has_array_of_type) { 451bd882c8aSJames Wright // Allocate if array is not yet allocated 452bd882c8aSJames Wright CeedCallBackend(CeedVectorSetArray(vec, mem_type, CEED_COPY_VALUES, NULL)); 453bd882c8aSJames Wright } else { 454bd882c8aSJames Wright // Select dirty array 455bd882c8aSJames Wright switch (mem_type) { 456bd882c8aSJames Wright case CEED_MEM_HOST: 457bd882c8aSJames Wright if (impl->h_array_borrowed) impl->h_array = impl->h_array_borrowed; 458bd882c8aSJames Wright else impl->h_array = impl->h_array_owned; 459bd882c8aSJames Wright break; 460bd882c8aSJames Wright case CEED_MEM_DEVICE: 461bd882c8aSJames Wright if (impl->d_array_borrowed) impl->d_array = impl->d_array_borrowed; 462bd882c8aSJames Wright else impl->d_array = impl->d_array_owned; 463bd882c8aSJames Wright } 464bd882c8aSJames Wright } 465bd882c8aSJames Wright return CeedVectorGetArray_Sycl(vec, mem_type, array); 466bd882c8aSJames Wright } 467bd882c8aSJames Wright 468bd882c8aSJames Wright //------------------------------------------------------------------------------ 469bd882c8aSJames Wright // Get the norm of a CeedVector 470bd882c8aSJames Wright //------------------------------------------------------------------------------ 471bd882c8aSJames Wright static int CeedVectorNorm_Sycl(CeedVector vec, CeedNormType type, CeedScalar *norm) { 472bd882c8aSJames Wright Ceed ceed; 473bd882c8aSJames Wright Ceed_Sycl *data; 474dd64fc84SJeremy L Thompson CeedSize length; 475dd64fc84SJeremy L Thompson const CeedScalar *d_array; 476dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 477dd64fc84SJeremy L Thompson 478dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 479dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 480dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetLength(vec, &length)); 481bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 482bd882c8aSJames Wright 483bd882c8aSJames Wright // Compute norm 484bd882c8aSJames Wright CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &d_array)); 485bd882c8aSJames Wright switch (type) { 486bd882c8aSJames Wright case CEED_NORM_1: { 487bd882c8aSJames Wright // Order queue 488bd882c8aSJames Wright sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 489bd882c8aSJames Wright auto sumReduction = sycl::reduction(impl->reduction_norm, sycl::plus<>(), {sycl::property::reduction::initialize_to_identity{}}); 490bd882c8aSJames Wright data->sycl_queue.parallel_for(length, {e}, sumReduction, [=](sycl::id<1> i, auto &sum) { sum += abs(d_array[i]); }).wait_and_throw(); 491bd882c8aSJames Wright } break; 492bd882c8aSJames Wright case CEED_NORM_2: { 493bd882c8aSJames Wright // Order queue 494bd882c8aSJames Wright sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 495bd882c8aSJames Wright auto sumReduction = sycl::reduction(impl->reduction_norm, sycl::plus<>(), {sycl::property::reduction::initialize_to_identity{}}); 496bd882c8aSJames Wright data->sycl_queue.parallel_for(length, {e}, sumReduction, [=](sycl::id<1> i, auto &sum) { sum += (d_array[i] * d_array[i]); }).wait_and_throw(); 497bd882c8aSJames Wright } break; 498bd882c8aSJames Wright case CEED_NORM_MAX: { 499bd882c8aSJames Wright // Order queue 500bd882c8aSJames Wright sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 501bd882c8aSJames Wright auto maxReduction = sycl::reduction(impl->reduction_norm, sycl::maximum<>(), {sycl::property::reduction::initialize_to_identity{}}); 502bd882c8aSJames Wright data->sycl_queue.parallel_for(length, {e}, maxReduction, [=](sycl::id<1> i, auto &max) { max.combine(abs(d_array[i])); }).wait_and_throw(); 503bd882c8aSJames Wright } break; 504bd882c8aSJames Wright } 505bd882c8aSJames Wright // L2 norm - square root over reduced value 506bd882c8aSJames Wright if (type == CEED_NORM_2) *norm = sqrt(*impl->reduction_norm); 507bd882c8aSJames Wright else *norm = *impl->reduction_norm; 508bd882c8aSJames Wright CeedCallBackend(CeedVectorRestoreArrayRead(vec, &d_array)); 509bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 510bd882c8aSJames Wright } 511bd882c8aSJames Wright 512bd882c8aSJames Wright //------------------------------------------------------------------------------ 513bd882c8aSJames Wright // Take reciprocal of a vector on host 514bd882c8aSJames Wright //------------------------------------------------------------------------------ 5156ca0f394SUmesh Unnikrishnan static int CeedHostReciprocal_Sycl(CeedScalar *h_array, CeedSize length) { 5166ca0f394SUmesh Unnikrishnan for (CeedSize i = 0; i < length; i++) { 517bd882c8aSJames Wright if (std::fabs(h_array[i]) > CEED_EPSILON) h_array[i] = 1. / h_array[i]; 518bd882c8aSJames Wright } 519bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 520bd882c8aSJames Wright } 521bd882c8aSJames Wright 522bd882c8aSJames Wright //------------------------------------------------------------------------------ 523bd882c8aSJames Wright // Take reciprocal of a vector on device 524bd882c8aSJames Wright //------------------------------------------------------------------------------ 5256ca0f394SUmesh Unnikrishnan static int CeedDeviceReciprocal_Sycl(sycl::queue &sycl_queue, CeedScalar *d_array, CeedSize length) { 526bd882c8aSJames Wright // Order queue 527bd882c8aSJames Wright sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 528bd882c8aSJames Wright sycl_queue.parallel_for(length, {e}, [=](sycl::id<1> i) { 529bd882c8aSJames Wright if (std::fabs(d_array[i]) > CEED_EPSILON) d_array[i] = 1. / d_array[i]; 530bd882c8aSJames Wright }); 531bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 532bd882c8aSJames Wright } 533bd882c8aSJames Wright 534bd882c8aSJames Wright //------------------------------------------------------------------------------ 535bd882c8aSJames Wright // Take reciprocal of a vector 536bd882c8aSJames Wright //------------------------------------------------------------------------------ 537bd882c8aSJames Wright static int CeedVectorReciprocal_Sycl(CeedVector vec) { 538bd882c8aSJames Wright Ceed ceed; 539bd882c8aSJames Wright Ceed_Sycl *data; 540dd64fc84SJeremy L Thompson CeedSize length; 541dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 542dd64fc84SJeremy L Thompson 543dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 544dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 545dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetLength(vec, &length)); 546bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 547bd882c8aSJames Wright 548bd882c8aSJames Wright // Set value for synced device/host array 549bd882c8aSJames Wright if (impl->d_array) CeedCallBackend(CeedDeviceReciprocal_Sycl(data->sycl_queue, impl->d_array, length)); 550bd882c8aSJames Wright if (impl->h_array) CeedCallBackend(CeedHostReciprocal_Sycl(impl->h_array, length)); 551bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 552bd882c8aSJames Wright } 553bd882c8aSJames Wright 554bd882c8aSJames Wright //------------------------------------------------------------------------------ 555bd882c8aSJames Wright // Compute x = alpha x on the host 556bd882c8aSJames Wright //------------------------------------------------------------------------------ 5576ca0f394SUmesh Unnikrishnan static int CeedHostScale_Sycl(CeedScalar *x_array, CeedScalar alpha, CeedSize length) { 5586ca0f394SUmesh Unnikrishnan for (CeedSize i = 0; i < length; i++) x_array[i] *= alpha; 559bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 560bd882c8aSJames Wright } 561bd882c8aSJames Wright 562bd882c8aSJames Wright //------------------------------------------------------------------------------ 563bd882c8aSJames Wright // Compute x = alpha x on device 564bd882c8aSJames Wright //------------------------------------------------------------------------------ 5656ca0f394SUmesh Unnikrishnan static int CeedDeviceScale_Sycl(sycl::queue &sycl_queue, CeedScalar *x_array, CeedScalar alpha, CeedSize length) { 566bd882c8aSJames Wright // Order queue 567bd882c8aSJames Wright sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 568bd882c8aSJames Wright sycl_queue.parallel_for(length, {e}, [=](sycl::id<1> i) { x_array[i] *= alpha; }); 569bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 570bd882c8aSJames Wright } 571bd882c8aSJames Wright 572bd882c8aSJames Wright //------------------------------------------------------------------------------ 573bd882c8aSJames Wright // Compute x = alpha x 574bd882c8aSJames Wright //------------------------------------------------------------------------------ 575bd882c8aSJames Wright static int CeedVectorScale_Sycl(CeedVector x, CeedScalar alpha) { 576bd882c8aSJames Wright Ceed ceed; 577bd882c8aSJames Wright Ceed_Sycl *data; 578dd64fc84SJeremy L Thompson CeedSize length; 579dd64fc84SJeremy L Thompson CeedVector_Sycl *x_impl; 580dd64fc84SJeremy L Thompson 581dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(x, &ceed)); 582dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(x, &x_impl)); 583dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetLength(x, &length)); 584bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 585bd882c8aSJames Wright 586bd882c8aSJames Wright // Set value for synced device/host array 587bd882c8aSJames Wright if (x_impl->d_array) CeedCallBackend(CeedDeviceScale_Sycl(data->sycl_queue, x_impl->d_array, alpha, length)); 588bd882c8aSJames Wright if (x_impl->h_array) CeedCallBackend(CeedHostScale_Sycl(x_impl->h_array, alpha, length)); 589bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 590bd882c8aSJames Wright } 591bd882c8aSJames Wright 592bd882c8aSJames Wright //------------------------------------------------------------------------------ 593bd882c8aSJames Wright // Compute y = alpha x + y on the host 594bd882c8aSJames Wright //------------------------------------------------------------------------------ 5956ca0f394SUmesh Unnikrishnan static int CeedHostAXPY_Sycl(CeedScalar *y_array, CeedScalar alpha, CeedScalar *x_array, CeedSize length) { 5966ca0f394SUmesh Unnikrishnan for (CeedSize i = 0; i < length; i++) y_array[i] += alpha * x_array[i]; 597bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 598bd882c8aSJames Wright } 599bd882c8aSJames Wright 600bd882c8aSJames Wright //------------------------------------------------------------------------------ 601bd882c8aSJames Wright // Compute y = alpha x + y on device 602bd882c8aSJames Wright //------------------------------------------------------------------------------ 6036ca0f394SUmesh Unnikrishnan static int CeedDeviceAXPY_Sycl(sycl::queue &sycl_queue, CeedScalar *y_array, CeedScalar alpha, CeedScalar *x_array, CeedSize length) { 604bd882c8aSJames Wright // Order queue 605bd882c8aSJames Wright sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 606bd882c8aSJames Wright sycl_queue.parallel_for(length, {e}, [=](sycl::id<1> i) { y_array[i] += alpha * x_array[i]; }); 607bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 608bd882c8aSJames Wright } 609bd882c8aSJames Wright 610bd882c8aSJames Wright //------------------------------------------------------------------------------ 611bd882c8aSJames Wright // Compute y = alpha x + y 612bd882c8aSJames Wright //------------------------------------------------------------------------------ 613bd882c8aSJames Wright static int CeedVectorAXPY_Sycl(CeedVector y, CeedScalar alpha, CeedVector x) { 614bd882c8aSJames Wright Ceed ceed; 615dd64fc84SJeremy L Thompson Ceed_Sycl *data; 616dd64fc84SJeremy L Thompson CeedSize length; 617bd882c8aSJames Wright CeedVector_Sycl *y_impl, *x_impl; 618dd64fc84SJeremy L Thompson 619dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(y, &ceed)); 620bd882c8aSJames Wright CeedCallBackend(CeedVectorGetData(y, &y_impl)); 621bd882c8aSJames Wright CeedCallBackend(CeedVectorGetData(x, &x_impl)); 622bd882c8aSJames Wright CeedCallBackend(CeedVectorGetLength(y, &length)); 623bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 624bd882c8aSJames Wright 625bd882c8aSJames Wright // Set value for synced device/host array 626bd882c8aSJames Wright if (y_impl->d_array) { 627bd882c8aSJames Wright CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_DEVICE)); 628bd882c8aSJames Wright CeedCallBackend(CeedDeviceAXPY_Sycl(data->sycl_queue, y_impl->d_array, alpha, x_impl->d_array, length)); 629bd882c8aSJames Wright } 630bd882c8aSJames Wright if (y_impl->h_array) { 631bd882c8aSJames Wright CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_HOST)); 632bd882c8aSJames Wright CeedCallBackend(CeedHostAXPY_Sycl(y_impl->h_array, alpha, x_impl->h_array, length)); 633bd882c8aSJames Wright } 634bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 635bd882c8aSJames Wright } 636bd882c8aSJames Wright 637bd882c8aSJames Wright //------------------------------------------------------------------------------ 638bd882c8aSJames Wright // Compute the pointwise multiplication w = x .* y on the host 639bd882c8aSJames Wright //------------------------------------------------------------------------------ 6406ca0f394SUmesh Unnikrishnan static int CeedHostPointwiseMult_Sycl(CeedScalar *w_array, CeedScalar *x_array, CeedScalar *y_array, CeedSize length) { 6416ca0f394SUmesh Unnikrishnan for (CeedSize i = 0; i < length; i++) w_array[i] = x_array[i] * y_array[i]; 642bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 643bd882c8aSJames Wright } 644bd882c8aSJames Wright 645bd882c8aSJames Wright //------------------------------------------------------------------------------ 646bd882c8aSJames Wright // Compute the pointwise multiplication w = x .* y on device (impl in .cu file) 647bd882c8aSJames Wright //------------------------------------------------------------------------------ 6486ca0f394SUmesh Unnikrishnan static int CeedDevicePointwiseMult_Sycl(sycl::queue &sycl_queue, CeedScalar *w_array, CeedScalar *x_array, CeedScalar *y_array, CeedSize length) { 649bd882c8aSJames Wright // Order queue 650bd882c8aSJames Wright sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 651bd882c8aSJames Wright sycl_queue.parallel_for(length, {e}, [=](sycl::id<1> i) { w_array[i] = x_array[i] * y_array[i]; }); 652bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 653bd882c8aSJames Wright } 654bd882c8aSJames Wright 655bd882c8aSJames Wright //------------------------------------------------------------------------------ 656bd882c8aSJames Wright // Compute the pointwise multiplication w = x .* y 657bd882c8aSJames Wright //------------------------------------------------------------------------------ 658bd882c8aSJames Wright static int CeedVectorPointwiseMult_Sycl(CeedVector w, CeedVector x, CeedVector y) { 659bd882c8aSJames Wright Ceed ceed; 660dd64fc84SJeremy L Thompson Ceed_Sycl *data; 661dd64fc84SJeremy L Thompson CeedSize length; 662bd882c8aSJames Wright CeedVector_Sycl *w_impl, *x_impl, *y_impl; 663dd64fc84SJeremy L Thompson 664dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(w, &ceed)); 665bd882c8aSJames Wright CeedCallBackend(CeedVectorGetData(w, &w_impl)); 666bd882c8aSJames Wright CeedCallBackend(CeedVectorGetData(x, &x_impl)); 667bd882c8aSJames Wright CeedCallBackend(CeedVectorGetData(y, &y_impl)); 668bd882c8aSJames Wright CeedCallBackend(CeedVectorGetLength(w, &length)); 669bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 670bd882c8aSJames Wright 671bd882c8aSJames Wright // Set value for synced device/host array 672bd882c8aSJames Wright if (!w_impl->d_array && !w_impl->h_array) { 673bd882c8aSJames Wright CeedCallBackend(CeedVectorSetValue(w, 0.0)); 674bd882c8aSJames Wright } 675bd882c8aSJames Wright if (w_impl->d_array) { 676bd882c8aSJames Wright CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_DEVICE)); 677bd882c8aSJames Wright CeedCallBackend(CeedVectorSyncArray(y, CEED_MEM_DEVICE)); 678bd882c8aSJames Wright CeedCallBackend(CeedDevicePointwiseMult_Sycl(data->sycl_queue, w_impl->d_array, x_impl->d_array, y_impl->d_array, length)); 679bd882c8aSJames Wright } 680bd882c8aSJames Wright if (w_impl->h_array) { 681bd882c8aSJames Wright CeedCallBackend(CeedVectorSyncArray(x, CEED_MEM_HOST)); 682bd882c8aSJames Wright CeedCallBackend(CeedVectorSyncArray(y, CEED_MEM_HOST)); 683bd882c8aSJames Wright CeedCallBackend(CeedHostPointwiseMult_Sycl(w_impl->h_array, x_impl->h_array, y_impl->h_array, length)); 684bd882c8aSJames Wright } 685bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 686bd882c8aSJames Wright } 687bd882c8aSJames Wright 688bd882c8aSJames Wright //------------------------------------------------------------------------------ 689bd882c8aSJames Wright // Destroy the vector 690bd882c8aSJames Wright //------------------------------------------------------------------------------ 691bd882c8aSJames Wright static int CeedVectorDestroy_Sycl(const CeedVector vec) { 692bd882c8aSJames Wright Ceed ceed; 693bd882c8aSJames Wright Ceed_Sycl *data; 694dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 695dd64fc84SJeremy L Thompson 696dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 697dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetData(vec, &impl)); 698bd882c8aSJames Wright CeedCallBackend(CeedGetData(ceed, &data)); 699bd882c8aSJames Wright 700bd882c8aSJames Wright // Wait for all work to finish before freeing memory 701bd882c8aSJames Wright CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 702bd882c8aSJames Wright CeedCallSycl(ceed, sycl::free(impl->d_array_owned, data->sycl_context)); 703bd882c8aSJames Wright CeedCallSycl(ceed, sycl::free(impl->reduction_norm, data->sycl_context)); 704bd882c8aSJames Wright 705bd882c8aSJames Wright CeedCallBackend(CeedFree(&impl->h_array_owned)); 706bd882c8aSJames Wright CeedCallBackend(CeedFree(&impl)); 707bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 708bd882c8aSJames Wright } 709bd882c8aSJames Wright 710bd882c8aSJames Wright //------------------------------------------------------------------------------ 711bd882c8aSJames Wright // Create a vector of the specified length (does not allocate memory) 712bd882c8aSJames Wright //------------------------------------------------------------------------------ 713bd882c8aSJames Wright int CeedVectorCreate_Sycl(CeedSize n, CeedVector vec) { 714bd882c8aSJames Wright Ceed ceed; 715bd882c8aSJames Wright Ceed_Sycl *data; 716dd64fc84SJeremy L Thompson CeedVector_Sycl *impl; 717bd882c8aSJames Wright 718dd64fc84SJeremy L Thompson CeedCallBackend(CeedVectorGetCeed(vec, &ceed)); 719dd64fc84SJeremy L Thompson CeedCallBackend(CeedGetData(ceed, &data)); 720bd882c8aSJames Wright CeedCallBackend(CeedCalloc(1, &impl)); 721bd882c8aSJames Wright CeedCallSycl(ceed, impl->reduction_norm = sycl::malloc_host<CeedScalar>(1, data->sycl_context)); 722bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "HasValidArray", CeedVectorHasValidArray_Sycl)); 723bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "HasBorrowedArrayOfType", CeedVectorHasBorrowedArrayOfType_Sycl)); 724bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "SetArray", CeedVectorSetArray_Sycl)); 725bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "TakeArray", CeedVectorTakeArray_Sycl)); 726bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "SetValue", CeedVectorSetValue_Sycl)); 727bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "SyncArray", CeedVectorSyncArray_Sycl)); 728bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "GetArray", CeedVectorGetArray_Sycl)); 729bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "GetArrayRead", CeedVectorGetArrayRead_Sycl)); 730bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "GetArrayWrite", CeedVectorGetArrayWrite_Sycl)); 731bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "Norm", CeedVectorNorm_Sycl)); 732bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "Reciprocal", CeedVectorReciprocal_Sycl)); 733bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "AXPY", CeedVectorAXPY_Sycl)); 734bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "Scale", CeedVectorScale_Sycl)); 735bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "PointwiseMult", CeedVectorPointwiseMult_Sycl)); 736bd882c8aSJames Wright CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Vector", vec, "Destroy", CeedVectorDestroy_Sycl)); 737bd882c8aSJames Wright CeedCallBackend(CeedVectorSetData(vec, impl)); 738bd882c8aSJames Wright return CEED_ERROR_SUCCESS; 739bd882c8aSJames Wright } 740