1 //===----------- rtl.cpp - Target independent OpenMP target RTL -----------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // Functionality for handling RTL plugins. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "rtl.h" 14 #include "device.h" 15 #include "private.h" 16 17 #include <cassert> 18 #include <cstdlib> 19 #include <cstring> 20 #include <dlfcn.h> 21 #include <mutex> 22 #include <string> 23 24 // List of all plugins that can support offloading. 25 static const char *RTLNames[] = { 26 /* PowerPC target */ "libomptarget.rtl.ppc64.so", 27 /* x86_64 target */ "libomptarget.rtl.x86_64.so", 28 /* CUDA target */ "libomptarget.rtl.cuda.so", 29 /* AArch64 target */ "libomptarget.rtl.aarch64.so", 30 /* SX-Aurora VE target */ "libomptarget.rtl.ve.so", 31 /* AMDGPU target */ "libomptarget.rtl.amdgpu.so", 32 /* Remote target */ "libomptarget.rtl.rpc.so", 33 }; 34 35 PluginManager *PM; 36 37 #if OMPTARGET_PROFILE_ENABLED 38 static char *ProfileTraceFile = nullptr; 39 #endif 40 41 __attribute__((constructor(101))) void init() { 42 DP("Init target library!\n"); 43 44 bool UseEventsForAtomicTransfers = true; 45 if (const char *ForceAtomicMap = getenv("LIBOMPTARGET_MAP_FORCE_ATOMIC")) { 46 std::string ForceAtomicMapStr(ForceAtomicMap); 47 if (ForceAtomicMapStr == "false" || ForceAtomicMapStr == "FALSE") 48 UseEventsForAtomicTransfers = false; 49 else if (ForceAtomicMapStr != "true" && ForceAtomicMapStr != "TRUE") 50 fprintf(stderr, 51 "Warning: 'LIBOMPTARGET_MAP_FORCE_ATOMIC' accepts only " 52 "'true'/'TRUE' or 'false'/'FALSE' as options, '%s' ignored\n", 53 ForceAtomicMap); 54 } 55 56 PM = new PluginManager(UseEventsForAtomicTransfers); 57 58 #ifdef OMPTARGET_PROFILE_ENABLED 59 ProfileTraceFile = getenv("LIBOMPTARGET_PROFILE"); 60 // TODO: add a configuration option for time granularity 61 if (ProfileTraceFile) 62 llvm::timeTraceProfilerInitialize(500 /* us */, "libomptarget"); 63 #endif 64 } 65 66 __attribute__((destructor(101))) void deinit() { 67 DP("Deinit target library!\n"); 68 delete PM; 69 70 #ifdef OMPTARGET_PROFILE_ENABLED 71 if (ProfileTraceFile) { 72 // TODO: add env var for file output 73 if (auto E = llvm::timeTraceProfilerWrite(ProfileTraceFile, "-")) 74 fprintf(stderr, "Error writing out the time trace\n"); 75 76 llvm::timeTraceProfilerCleanup(); 77 } 78 #endif 79 } 80 81 void RTLsTy::LoadRTLs() { 82 // Parse environment variable OMP_TARGET_OFFLOAD (if set) 83 PM->TargetOffloadPolicy = 84 (kmp_target_offload_kind_t)__kmpc_get_target_offload(); 85 if (PM->TargetOffloadPolicy == tgt_disabled) { 86 return; 87 } 88 89 DP("Loading RTLs...\n"); 90 91 // Attempt to open all the plugins and, if they exist, check if the interface 92 // is correct and if they are supporting any devices. 93 for (auto *Name : RTLNames) { 94 DP("Loading library '%s'...\n", Name); 95 void *dynlib_handle = dlopen(Name, RTLD_NOW); 96 97 if (!dynlib_handle) { 98 // Library does not exist or cannot be found. 99 DP("Unable to load library '%s': %s!\n", Name, dlerror()); 100 continue; 101 } 102 103 DP("Successfully loaded library '%s'!\n", Name); 104 105 AllRTLs.emplace_back(); 106 107 // Retrieve the RTL information from the runtime library. 108 RTLInfoTy &R = AllRTLs.back(); 109 110 bool ValidPlugin = true; 111 112 if (!(*((void **)&R.is_valid_binary) = 113 dlsym(dynlib_handle, "__tgt_rtl_is_valid_binary"))) 114 ValidPlugin = false; 115 if (!(*((void **)&R.number_of_devices) = 116 dlsym(dynlib_handle, "__tgt_rtl_number_of_devices"))) 117 ValidPlugin = false; 118 if (!(*((void **)&R.init_device) = 119 dlsym(dynlib_handle, "__tgt_rtl_init_device"))) 120 ValidPlugin = false; 121 if (!(*((void **)&R.load_binary) = 122 dlsym(dynlib_handle, "__tgt_rtl_load_binary"))) 123 ValidPlugin = false; 124 if (!(*((void **)&R.data_alloc) = 125 dlsym(dynlib_handle, "__tgt_rtl_data_alloc"))) 126 ValidPlugin = false; 127 if (!(*((void **)&R.data_submit) = 128 dlsym(dynlib_handle, "__tgt_rtl_data_submit"))) 129 ValidPlugin = false; 130 if (!(*((void **)&R.data_retrieve) = 131 dlsym(dynlib_handle, "__tgt_rtl_data_retrieve"))) 132 ValidPlugin = false; 133 if (!(*((void **)&R.data_delete) = 134 dlsym(dynlib_handle, "__tgt_rtl_data_delete"))) 135 ValidPlugin = false; 136 if (!(*((void **)&R.run_region) = 137 dlsym(dynlib_handle, "__tgt_rtl_run_target_region"))) 138 ValidPlugin = false; 139 if (!(*((void **)&R.run_team_region) = 140 dlsym(dynlib_handle, "__tgt_rtl_run_target_team_region"))) 141 ValidPlugin = false; 142 143 // Invalid plugin 144 if (!ValidPlugin) { 145 DP("Invalid plugin as necessary interface is not found.\n"); 146 AllRTLs.pop_back(); 147 continue; 148 } 149 150 // No devices are supported by this RTL? 151 if (!(R.NumberOfDevices = R.number_of_devices())) { 152 // The RTL is invalid! Will pop the object from the RTLs list. 153 DP("No devices supported in this RTL\n"); 154 AllRTLs.pop_back(); 155 continue; 156 } 157 158 R.LibraryHandler = dynlib_handle; 159 160 #ifdef OMPTARGET_DEBUG 161 R.RTLName = Name; 162 #endif 163 164 DP("Registering RTL %s supporting %d devices!\n", R.RTLName.c_str(), 165 R.NumberOfDevices); 166 167 // Optional functions 168 *((void **)&R.deinit_device) = 169 dlsym(dynlib_handle, "__tgt_rtl_deinit_device"); 170 *((void **)&R.init_requires) = 171 dlsym(dynlib_handle, "__tgt_rtl_init_requires"); 172 *((void **)&R.data_submit_async) = 173 dlsym(dynlib_handle, "__tgt_rtl_data_submit_async"); 174 *((void **)&R.data_retrieve_async) = 175 dlsym(dynlib_handle, "__tgt_rtl_data_retrieve_async"); 176 *((void **)&R.run_region_async) = 177 dlsym(dynlib_handle, "__tgt_rtl_run_target_region_async"); 178 *((void **)&R.run_team_region_async) = 179 dlsym(dynlib_handle, "__tgt_rtl_run_target_team_region_async"); 180 *((void **)&R.synchronize) = dlsym(dynlib_handle, "__tgt_rtl_synchronize"); 181 *((void **)&R.data_exchange) = 182 dlsym(dynlib_handle, "__tgt_rtl_data_exchange"); 183 *((void **)&R.data_exchange_async) = 184 dlsym(dynlib_handle, "__tgt_rtl_data_exchange_async"); 185 *((void **)&R.is_data_exchangable) = 186 dlsym(dynlib_handle, "__tgt_rtl_is_data_exchangable"); 187 *((void **)&R.register_lib) = 188 dlsym(dynlib_handle, "__tgt_rtl_register_lib"); 189 *((void **)&R.unregister_lib) = 190 dlsym(dynlib_handle, "__tgt_rtl_unregister_lib"); 191 *((void **)&R.supports_empty_images) = 192 dlsym(dynlib_handle, "__tgt_rtl_supports_empty_images"); 193 *((void **)&R.set_info_flag) = 194 dlsym(dynlib_handle, "__tgt_rtl_set_info_flag"); 195 *((void **)&R.print_device_info) = 196 dlsym(dynlib_handle, "__tgt_rtl_print_device_info"); 197 *((void **)&R.create_event) = 198 dlsym(dynlib_handle, "__tgt_rtl_create_event"); 199 *((void **)&R.record_event) = 200 dlsym(dynlib_handle, "__tgt_rtl_record_event"); 201 *((void **)&R.wait_event) = dlsym(dynlib_handle, "__tgt_rtl_wait_event"); 202 *((void **)&R.sync_event) = dlsym(dynlib_handle, "__tgt_rtl_sync_event"); 203 *((void **)&R.destroy_event) = 204 dlsym(dynlib_handle, "__tgt_rtl_destroy_event"); 205 *((void **)&R.release_async_info) = 206 dlsym(dynlib_handle, "__tgt_rtl_release_async_info"); 207 *((void **)&R.init_async_info) = 208 dlsym(dynlib_handle, "__tgt_rtl_init_async_info"); 209 *((void **)&R.init_device_info) = 210 dlsym(dynlib_handle, "__tgt_rtl_init_device_info"); 211 } 212 213 DP("RTLs loaded!\n"); 214 215 return; 216 } 217 218 //////////////////////////////////////////////////////////////////////////////// 219 // Functionality for registering libs 220 221 static void RegisterImageIntoTranslationTable(TranslationTable &TT, 222 RTLInfoTy &RTL, 223 __tgt_device_image *image) { 224 225 // same size, as when we increase one, we also increase the other. 226 assert(TT.TargetsTable.size() == TT.TargetsImages.size() && 227 "We should have as many images as we have tables!"); 228 229 // Resize the Targets Table and Images to accommodate the new targets if 230 // required 231 unsigned TargetsTableMinimumSize = RTL.Idx + RTL.NumberOfDevices; 232 233 if (TT.TargetsTable.size() < TargetsTableMinimumSize) { 234 TT.TargetsImages.resize(TargetsTableMinimumSize, 0); 235 TT.TargetsTable.resize(TargetsTableMinimumSize, 0); 236 } 237 238 // Register the image in all devices for this target type. 239 for (int32_t i = 0; i < RTL.NumberOfDevices; ++i) { 240 // If we are changing the image we are also invalidating the target table. 241 if (TT.TargetsImages[RTL.Idx + i] != image) { 242 TT.TargetsImages[RTL.Idx + i] = image; 243 TT.TargetsTable[RTL.Idx + i] = 0; // lazy initialization of target table. 244 } 245 } 246 } 247 248 //////////////////////////////////////////////////////////////////////////////// 249 // Functionality for registering Ctors/Dtors 250 251 static void RegisterGlobalCtorsDtorsForImage(__tgt_bin_desc *desc, 252 __tgt_device_image *img, 253 RTLInfoTy *RTL) { 254 255 for (int32_t i = 0; i < RTL->NumberOfDevices; ++i) { 256 DeviceTy &Device = *PM->Devices[RTL->Idx + i]; 257 Device.PendingGlobalsMtx.lock(); 258 Device.HasPendingGlobals = true; 259 for (__tgt_offload_entry *entry = img->EntriesBegin; 260 entry != img->EntriesEnd; ++entry) { 261 if (entry->flags & OMP_DECLARE_TARGET_CTOR) { 262 DP("Adding ctor " DPxMOD " to the pending list.\n", 263 DPxPTR(entry->addr)); 264 Device.PendingCtorsDtors[desc].PendingCtors.push_back(entry->addr); 265 } else if (entry->flags & OMP_DECLARE_TARGET_DTOR) { 266 // Dtors are pushed in reverse order so they are executed from end 267 // to beginning when unregistering the library! 268 DP("Adding dtor " DPxMOD " to the pending list.\n", 269 DPxPTR(entry->addr)); 270 Device.PendingCtorsDtors[desc].PendingDtors.push_front(entry->addr); 271 } 272 273 if (entry->flags & OMP_DECLARE_TARGET_LINK) { 274 DP("The \"link\" attribute is not yet supported!\n"); 275 } 276 } 277 Device.PendingGlobalsMtx.unlock(); 278 } 279 } 280 281 void RTLsTy::RegisterRequires(int64_t flags) { 282 // TODO: add more elaborate check. 283 // Minimal check: only set requires flags if previous value 284 // is undefined. This ensures that only the first call to this 285 // function will set the requires flags. All subsequent calls 286 // will be checked for compatibility. 287 assert(flags != OMP_REQ_UNDEFINED && 288 "illegal undefined flag for requires directive!"); 289 if (RequiresFlags == OMP_REQ_UNDEFINED) { 290 RequiresFlags = flags; 291 return; 292 } 293 294 // If multiple compilation units are present enforce 295 // consistency across all of them for require clauses: 296 // - reverse_offload 297 // - unified_address 298 // - unified_shared_memory 299 if ((RequiresFlags & OMP_REQ_REVERSE_OFFLOAD) != 300 (flags & OMP_REQ_REVERSE_OFFLOAD)) { 301 FATAL_MESSAGE0( 302 1, "'#pragma omp requires reverse_offload' not used consistently!"); 303 } 304 if ((RequiresFlags & OMP_REQ_UNIFIED_ADDRESS) != 305 (flags & OMP_REQ_UNIFIED_ADDRESS)) { 306 FATAL_MESSAGE0( 307 1, "'#pragma omp requires unified_address' not used consistently!"); 308 } 309 if ((RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY) != 310 (flags & OMP_REQ_UNIFIED_SHARED_MEMORY)) { 311 FATAL_MESSAGE0( 312 1, 313 "'#pragma omp requires unified_shared_memory' not used consistently!"); 314 } 315 316 // TODO: insert any other missing checks 317 318 DP("New requires flags %" PRId64 " compatible with existing %" PRId64 "!\n", 319 flags, RequiresFlags); 320 } 321 322 void RTLsTy::initRTLonce(RTLInfoTy &R) { 323 // If this RTL is not already in use, initialize it. 324 if (!R.isUsed && R.NumberOfDevices != 0) { 325 // Initialize the device information for the RTL we are about to use. 326 const size_t Start = PM->Devices.size(); 327 PM->Devices.reserve(Start + R.NumberOfDevices); 328 for (int32_t device_id = 0; device_id < R.NumberOfDevices; device_id++) { 329 PM->Devices.push_back(std::make_unique<DeviceTy>(&R)); 330 // global device ID 331 PM->Devices[Start + device_id]->DeviceID = Start + device_id; 332 // RTL local device ID 333 PM->Devices[Start + device_id]->RTLDeviceID = device_id; 334 } 335 336 // Initialize the index of this RTL and save it in the used RTLs. 337 R.Idx = (UsedRTLs.empty()) 338 ? 0 339 : UsedRTLs.back()->Idx + UsedRTLs.back()->NumberOfDevices; 340 assert((size_t)R.Idx == Start && 341 "RTL index should equal the number of devices used so far."); 342 R.isUsed = true; 343 UsedRTLs.push_back(&R); 344 345 DP("RTL " DPxMOD " has index %d!\n", DPxPTR(R.LibraryHandler), R.Idx); 346 } 347 } 348 349 void RTLsTy::initAllRTLs() { 350 for (auto &R : AllRTLs) 351 initRTLonce(R); 352 } 353 354 void RTLsTy::RegisterLib(__tgt_bin_desc *desc) { 355 PM->RTLsMtx.lock(); 356 // Register the images with the RTLs that understand them, if any. 357 for (int32_t i = 0; i < desc->NumDeviceImages; ++i) { 358 // Obtain the image. 359 __tgt_device_image *img = &desc->DeviceImages[i]; 360 361 RTLInfoTy *FoundRTL = nullptr; 362 363 // Scan the RTLs that have associated images until we find one that supports 364 // the current image. 365 for (auto &R : AllRTLs) { 366 if (!R.is_valid_binary(img)) { 367 DP("Image " DPxMOD " is NOT compatible with RTL %s!\n", 368 DPxPTR(img->ImageStart), R.RTLName.c_str()); 369 continue; 370 } 371 372 DP("Image " DPxMOD " is compatible with RTL %s!\n", 373 DPxPTR(img->ImageStart), R.RTLName.c_str()); 374 375 initRTLonce(R); 376 377 // Initialize (if necessary) translation table for this library. 378 PM->TrlTblMtx.lock(); 379 if (!PM->HostEntriesBeginToTransTable.count(desc->HostEntriesBegin)) { 380 PM->HostEntriesBeginRegistrationOrder.push_back(desc->HostEntriesBegin); 381 TranslationTable &TransTable = 382 (PM->HostEntriesBeginToTransTable)[desc->HostEntriesBegin]; 383 TransTable.HostTable.EntriesBegin = desc->HostEntriesBegin; 384 TransTable.HostTable.EntriesEnd = desc->HostEntriesEnd; 385 } 386 387 // Retrieve translation table for this library. 388 TranslationTable &TransTable = 389 (PM->HostEntriesBeginToTransTable)[desc->HostEntriesBegin]; 390 391 DP("Registering image " DPxMOD " with RTL %s!\n", DPxPTR(img->ImageStart), 392 R.RTLName.c_str()); 393 RegisterImageIntoTranslationTable(TransTable, R, img); 394 PM->TrlTblMtx.unlock(); 395 FoundRTL = &R; 396 397 // Load ctors/dtors for static objects 398 RegisterGlobalCtorsDtorsForImage(desc, img, FoundRTL); 399 400 // if an RTL was found we are done - proceed to register the next image 401 break; 402 } 403 404 if (!FoundRTL) { 405 DP("No RTL found for image " DPxMOD "!\n", DPxPTR(img->ImageStart)); 406 } 407 } 408 PM->RTLsMtx.unlock(); 409 410 DP("Done registering entries!\n"); 411 } 412 413 void RTLsTy::UnregisterLib(__tgt_bin_desc *desc) { 414 DP("Unloading target library!\n"); 415 416 PM->RTLsMtx.lock(); 417 // Find which RTL understands each image, if any. 418 for (int32_t i = 0; i < desc->NumDeviceImages; ++i) { 419 // Obtain the image. 420 __tgt_device_image *img = &desc->DeviceImages[i]; 421 422 RTLInfoTy *FoundRTL = NULL; 423 424 // Scan the RTLs that have associated images until we find one that supports 425 // the current image. We only need to scan RTLs that are already being used. 426 for (auto *R : UsedRTLs) { 427 428 assert(R->isUsed && "Expecting used RTLs."); 429 430 if (!R->is_valid_binary(img)) { 431 DP("Image " DPxMOD " is NOT compatible with RTL " DPxMOD "!\n", 432 DPxPTR(img->ImageStart), DPxPTR(R->LibraryHandler)); 433 continue; 434 } 435 436 DP("Image " DPxMOD " is compatible with RTL " DPxMOD "!\n", 437 DPxPTR(img->ImageStart), DPxPTR(R->LibraryHandler)); 438 439 FoundRTL = R; 440 441 // Execute dtors for static objects if the device has been used, i.e. 442 // if its PendingCtors list has been emptied. 443 for (int32_t i = 0; i < FoundRTL->NumberOfDevices; ++i) { 444 DeviceTy &Device = *PM->Devices[FoundRTL->Idx + i]; 445 Device.PendingGlobalsMtx.lock(); 446 if (Device.PendingCtorsDtors[desc].PendingCtors.empty()) { 447 AsyncInfoTy AsyncInfo(Device); 448 for (auto &dtor : Device.PendingCtorsDtors[desc].PendingDtors) { 449 int rc = target(nullptr, Device, dtor, 0, nullptr, nullptr, nullptr, 450 nullptr, nullptr, nullptr, 1, 1, true /*team*/, 451 AsyncInfo); 452 if (rc != OFFLOAD_SUCCESS) { 453 DP("Running destructor " DPxMOD " failed.\n", DPxPTR(dtor)); 454 } 455 } 456 // Remove this library's entry from PendingCtorsDtors 457 Device.PendingCtorsDtors.erase(desc); 458 // All constructors have been issued, wait for them now. 459 if (AsyncInfo.synchronize() != OFFLOAD_SUCCESS) 460 DP("Failed synchronizing destructors kernels.\n"); 461 } 462 Device.PendingGlobalsMtx.unlock(); 463 } 464 465 DP("Unregistered image " DPxMOD " from RTL " DPxMOD "!\n", 466 DPxPTR(img->ImageStart), DPxPTR(R->LibraryHandler)); 467 468 break; 469 } 470 471 // if no RTL was found proceed to unregister the next image 472 if (!FoundRTL) { 473 DP("No RTLs in use support the image " DPxMOD "!\n", 474 DPxPTR(img->ImageStart)); 475 } 476 } 477 PM->RTLsMtx.unlock(); 478 DP("Done unregistering images!\n"); 479 480 // Remove entries from PM->HostPtrToTableMap 481 PM->TblMapMtx.lock(); 482 for (__tgt_offload_entry *cur = desc->HostEntriesBegin; 483 cur < desc->HostEntriesEnd; ++cur) { 484 PM->HostPtrToTableMap.erase(cur->addr); 485 } 486 487 // Remove translation table for this descriptor. 488 auto TransTable = 489 PM->HostEntriesBeginToTransTable.find(desc->HostEntriesBegin); 490 if (TransTable != PM->HostEntriesBeginToTransTable.end()) { 491 DP("Removing translation table for descriptor " DPxMOD "\n", 492 DPxPTR(desc->HostEntriesBegin)); 493 PM->HostEntriesBeginToTransTable.erase(TransTable); 494 } else { 495 DP("Translation table for descriptor " DPxMOD " cannot be found, probably " 496 "it has been already removed.\n", 497 DPxPTR(desc->HostEntriesBegin)); 498 } 499 500 PM->TblMapMtx.unlock(); 501 502 // TODO: Remove RTL and the devices it manages if it's not used anymore? 503 // TODO: Write some RTL->unload_image(...) function? 504 505 DP("Done unregistering library!\n"); 506 } 507