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.init_requires) = 169 dlsym(dynlib_handle, "__tgt_rtl_init_requires"); 170 *((void **)&R.data_submit_async) = 171 dlsym(dynlib_handle, "__tgt_rtl_data_submit_async"); 172 *((void **)&R.data_retrieve_async) = 173 dlsym(dynlib_handle, "__tgt_rtl_data_retrieve_async"); 174 *((void **)&R.run_region_async) = 175 dlsym(dynlib_handle, "__tgt_rtl_run_target_region_async"); 176 *((void **)&R.run_team_region_async) = 177 dlsym(dynlib_handle, "__tgt_rtl_run_target_team_region_async"); 178 *((void **)&R.synchronize) = dlsym(dynlib_handle, "__tgt_rtl_synchronize"); 179 *((void **)&R.data_exchange) = 180 dlsym(dynlib_handle, "__tgt_rtl_data_exchange"); 181 *((void **)&R.data_exchange_async) = 182 dlsym(dynlib_handle, "__tgt_rtl_data_exchange_async"); 183 *((void **)&R.is_data_exchangable) = 184 dlsym(dynlib_handle, "__tgt_rtl_is_data_exchangable"); 185 *((void **)&R.register_lib) = 186 dlsym(dynlib_handle, "__tgt_rtl_register_lib"); 187 *((void **)&R.unregister_lib) = 188 dlsym(dynlib_handle, "__tgt_rtl_unregister_lib"); 189 *((void **)&R.supports_empty_images) = 190 dlsym(dynlib_handle, "__tgt_rtl_supports_empty_images"); 191 *((void **)&R.set_info_flag) = 192 dlsym(dynlib_handle, "__tgt_rtl_set_info_flag"); 193 *((void **)&R.print_device_info) = 194 dlsym(dynlib_handle, "__tgt_rtl_print_device_info"); 195 *((void **)&R.create_event) = 196 dlsym(dynlib_handle, "__tgt_rtl_create_event"); 197 *((void **)&R.record_event) = 198 dlsym(dynlib_handle, "__tgt_rtl_record_event"); 199 *((void **)&R.wait_event) = dlsym(dynlib_handle, "__tgt_rtl_wait_event"); 200 *((void **)&R.sync_event) = dlsym(dynlib_handle, "__tgt_rtl_sync_event"); 201 *((void **)&R.destroy_event) = 202 dlsym(dynlib_handle, "__tgt_rtl_destroy_event"); 203 *((void **)&R.release_async_info) = 204 dlsym(dynlib_handle, "__tgt_rtl_release_async_info"); 205 *((void **)&R.init_async_info) = 206 dlsym(dynlib_handle, "__tgt_rtl_init_async_info"); 207 *((void **)&R.init_device_info) = 208 dlsym(dynlib_handle, "__tgt_rtl_init_device_info"); 209 } 210 211 DP("RTLs loaded!\n"); 212 213 return; 214 } 215 216 //////////////////////////////////////////////////////////////////////////////// 217 // Functionality for registering libs 218 219 static void RegisterImageIntoTranslationTable(TranslationTable &TT, 220 RTLInfoTy &RTL, 221 __tgt_device_image *image) { 222 223 // same size, as when we increase one, we also increase the other. 224 assert(TT.TargetsTable.size() == TT.TargetsImages.size() && 225 "We should have as many images as we have tables!"); 226 227 // Resize the Targets Table and Images to accommodate the new targets if 228 // required 229 unsigned TargetsTableMinimumSize = RTL.Idx + RTL.NumberOfDevices; 230 231 if (TT.TargetsTable.size() < TargetsTableMinimumSize) { 232 TT.TargetsImages.resize(TargetsTableMinimumSize, 0); 233 TT.TargetsTable.resize(TargetsTableMinimumSize, 0); 234 } 235 236 // Register the image in all devices for this target type. 237 for (int32_t i = 0; i < RTL.NumberOfDevices; ++i) { 238 // If we are changing the image we are also invalidating the target table. 239 if (TT.TargetsImages[RTL.Idx + i] != image) { 240 TT.TargetsImages[RTL.Idx + i] = image; 241 TT.TargetsTable[RTL.Idx + i] = 0; // lazy initialization of target table. 242 } 243 } 244 } 245 246 //////////////////////////////////////////////////////////////////////////////// 247 // Functionality for registering Ctors/Dtors 248 249 static void RegisterGlobalCtorsDtorsForImage(__tgt_bin_desc *desc, 250 __tgt_device_image *img, 251 RTLInfoTy *RTL) { 252 253 for (int32_t i = 0; i < RTL->NumberOfDevices; ++i) { 254 DeviceTy &Device = *PM->Devices[RTL->Idx + i]; 255 Device.PendingGlobalsMtx.lock(); 256 Device.HasPendingGlobals = true; 257 for (__tgt_offload_entry *entry = img->EntriesBegin; 258 entry != img->EntriesEnd; ++entry) { 259 if (entry->flags & OMP_DECLARE_TARGET_CTOR) { 260 DP("Adding ctor " DPxMOD " to the pending list.\n", 261 DPxPTR(entry->addr)); 262 Device.PendingCtorsDtors[desc].PendingCtors.push_back(entry->addr); 263 } else if (entry->flags & OMP_DECLARE_TARGET_DTOR) { 264 // Dtors are pushed in reverse order so they are executed from end 265 // to beginning when unregistering the library! 266 DP("Adding dtor " DPxMOD " to the pending list.\n", 267 DPxPTR(entry->addr)); 268 Device.PendingCtorsDtors[desc].PendingDtors.push_front(entry->addr); 269 } 270 271 if (entry->flags & OMP_DECLARE_TARGET_LINK) { 272 DP("The \"link\" attribute is not yet supported!\n"); 273 } 274 } 275 Device.PendingGlobalsMtx.unlock(); 276 } 277 } 278 279 void RTLsTy::RegisterRequires(int64_t flags) { 280 // TODO: add more elaborate check. 281 // Minimal check: only set requires flags if previous value 282 // is undefined. This ensures that only the first call to this 283 // function will set the requires flags. All subsequent calls 284 // will be checked for compatibility. 285 assert(flags != OMP_REQ_UNDEFINED && 286 "illegal undefined flag for requires directive!"); 287 if (RequiresFlags == OMP_REQ_UNDEFINED) { 288 RequiresFlags = flags; 289 return; 290 } 291 292 // If multiple compilation units are present enforce 293 // consistency across all of them for require clauses: 294 // - reverse_offload 295 // - unified_address 296 // - unified_shared_memory 297 if ((RequiresFlags & OMP_REQ_REVERSE_OFFLOAD) != 298 (flags & OMP_REQ_REVERSE_OFFLOAD)) { 299 FATAL_MESSAGE0( 300 1, "'#pragma omp requires reverse_offload' not used consistently!"); 301 } 302 if ((RequiresFlags & OMP_REQ_UNIFIED_ADDRESS) != 303 (flags & OMP_REQ_UNIFIED_ADDRESS)) { 304 FATAL_MESSAGE0( 305 1, "'#pragma omp requires unified_address' not used consistently!"); 306 } 307 if ((RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY) != 308 (flags & OMP_REQ_UNIFIED_SHARED_MEMORY)) { 309 FATAL_MESSAGE0( 310 1, 311 "'#pragma omp requires unified_shared_memory' not used consistently!"); 312 } 313 314 // TODO: insert any other missing checks 315 316 DP("New requires flags %" PRId64 " compatible with existing %" PRId64 "!\n", 317 flags, RequiresFlags); 318 } 319 320 void RTLsTy::initRTLonce(RTLInfoTy &R) { 321 // If this RTL is not already in use, initialize it. 322 if (!R.isUsed && R.NumberOfDevices != 0) { 323 // Initialize the device information for the RTL we are about to use. 324 const size_t Start = PM->Devices.size(); 325 PM->Devices.reserve(Start + R.NumberOfDevices); 326 for (int32_t device_id = 0; device_id < R.NumberOfDevices; device_id++) { 327 PM->Devices.push_back(std::make_unique<DeviceTy>(&R)); 328 // global device ID 329 PM->Devices[Start + device_id]->DeviceID = Start + device_id; 330 // RTL local device ID 331 PM->Devices[Start + device_id]->RTLDeviceID = device_id; 332 } 333 334 // Initialize the index of this RTL and save it in the used RTLs. 335 R.Idx = (UsedRTLs.empty()) 336 ? 0 337 : UsedRTLs.back()->Idx + UsedRTLs.back()->NumberOfDevices; 338 assert((size_t)R.Idx == Start && 339 "RTL index should equal the number of devices used so far."); 340 R.isUsed = true; 341 UsedRTLs.push_back(&R); 342 343 DP("RTL " DPxMOD " has index %d!\n", DPxPTR(R.LibraryHandler), R.Idx); 344 } 345 } 346 347 void RTLsTy::initAllRTLs() { 348 for (auto &R : AllRTLs) 349 initRTLonce(R); 350 } 351 352 void RTLsTy::RegisterLib(__tgt_bin_desc *desc) { 353 PM->RTLsMtx.lock(); 354 // Register the images with the RTLs that understand them, if any. 355 for (int32_t i = 0; i < desc->NumDeviceImages; ++i) { 356 // Obtain the image. 357 __tgt_device_image *img = &desc->DeviceImages[i]; 358 359 RTLInfoTy *FoundRTL = nullptr; 360 361 // Scan the RTLs that have associated images until we find one that supports 362 // the current image. 363 for (auto &R : AllRTLs) { 364 if (!R.is_valid_binary(img)) { 365 DP("Image " DPxMOD " is NOT compatible with RTL %s!\n", 366 DPxPTR(img->ImageStart), R.RTLName.c_str()); 367 continue; 368 } 369 370 DP("Image " DPxMOD " is compatible with RTL %s!\n", 371 DPxPTR(img->ImageStart), R.RTLName.c_str()); 372 373 initRTLonce(R); 374 375 // Initialize (if necessary) translation table for this library. 376 PM->TrlTblMtx.lock(); 377 if (!PM->HostEntriesBeginToTransTable.count(desc->HostEntriesBegin)) { 378 PM->HostEntriesBeginRegistrationOrder.push_back(desc->HostEntriesBegin); 379 TranslationTable &TransTable = 380 (PM->HostEntriesBeginToTransTable)[desc->HostEntriesBegin]; 381 TransTable.HostTable.EntriesBegin = desc->HostEntriesBegin; 382 TransTable.HostTable.EntriesEnd = desc->HostEntriesEnd; 383 } 384 385 // Retrieve translation table for this library. 386 TranslationTable &TransTable = 387 (PM->HostEntriesBeginToTransTable)[desc->HostEntriesBegin]; 388 389 DP("Registering image " DPxMOD " with RTL %s!\n", DPxPTR(img->ImageStart), 390 R.RTLName.c_str()); 391 RegisterImageIntoTranslationTable(TransTable, R, img); 392 PM->TrlTblMtx.unlock(); 393 FoundRTL = &R; 394 395 // Load ctors/dtors for static objects 396 RegisterGlobalCtorsDtorsForImage(desc, img, FoundRTL); 397 398 // if an RTL was found we are done - proceed to register the next image 399 break; 400 } 401 402 if (!FoundRTL) { 403 DP("No RTL found for image " DPxMOD "!\n", DPxPTR(img->ImageStart)); 404 } 405 } 406 PM->RTLsMtx.unlock(); 407 408 DP("Done registering entries!\n"); 409 } 410 411 void RTLsTy::UnregisterLib(__tgt_bin_desc *desc) { 412 DP("Unloading target library!\n"); 413 414 PM->RTLsMtx.lock(); 415 // Find which RTL understands each image, if any. 416 for (int32_t i = 0; i < desc->NumDeviceImages; ++i) { 417 // Obtain the image. 418 __tgt_device_image *img = &desc->DeviceImages[i]; 419 420 RTLInfoTy *FoundRTL = NULL; 421 422 // Scan the RTLs that have associated images until we find one that supports 423 // the current image. We only need to scan RTLs that are already being used. 424 for (auto *R : UsedRTLs) { 425 426 assert(R->isUsed && "Expecting used RTLs."); 427 428 if (!R->is_valid_binary(img)) { 429 DP("Image " DPxMOD " is NOT compatible with RTL " DPxMOD "!\n", 430 DPxPTR(img->ImageStart), DPxPTR(R->LibraryHandler)); 431 continue; 432 } 433 434 DP("Image " DPxMOD " is compatible with RTL " DPxMOD "!\n", 435 DPxPTR(img->ImageStart), DPxPTR(R->LibraryHandler)); 436 437 FoundRTL = R; 438 439 // Execute dtors for static objects if the device has been used, i.e. 440 // if its PendingCtors list has been emptied. 441 for (int32_t i = 0; i < FoundRTL->NumberOfDevices; ++i) { 442 DeviceTy &Device = *PM->Devices[FoundRTL->Idx + i]; 443 Device.PendingGlobalsMtx.lock(); 444 if (Device.PendingCtorsDtors[desc].PendingCtors.empty()) { 445 AsyncInfoTy AsyncInfo(Device); 446 for (auto &dtor : Device.PendingCtorsDtors[desc].PendingDtors) { 447 int rc = target(nullptr, Device, dtor, 0, nullptr, nullptr, nullptr, 448 nullptr, nullptr, nullptr, 1, 1, true /*team*/, 449 AsyncInfo); 450 if (rc != OFFLOAD_SUCCESS) { 451 DP("Running destructor " DPxMOD " failed.\n", DPxPTR(dtor)); 452 } 453 } 454 // Remove this library's entry from PendingCtorsDtors 455 Device.PendingCtorsDtors.erase(desc); 456 // All constructors have been issued, wait for them now. 457 if (AsyncInfo.synchronize() != OFFLOAD_SUCCESS) 458 DP("Failed synchronizing destructors kernels.\n"); 459 } 460 Device.PendingGlobalsMtx.unlock(); 461 } 462 463 DP("Unregistered image " DPxMOD " from RTL " DPxMOD "!\n", 464 DPxPTR(img->ImageStart), DPxPTR(R->LibraryHandler)); 465 466 break; 467 } 468 469 // if no RTL was found proceed to unregister the next image 470 if (!FoundRTL) { 471 DP("No RTLs in use support the image " DPxMOD "!\n", 472 DPxPTR(img->ImageStart)); 473 } 474 } 475 PM->RTLsMtx.unlock(); 476 DP("Done unregistering images!\n"); 477 478 // Remove entries from PM->HostPtrToTableMap 479 PM->TblMapMtx.lock(); 480 for (__tgt_offload_entry *cur = desc->HostEntriesBegin; 481 cur < desc->HostEntriesEnd; ++cur) { 482 PM->HostPtrToTableMap.erase(cur->addr); 483 } 484 485 // Remove translation table for this descriptor. 486 auto TransTable = 487 PM->HostEntriesBeginToTransTable.find(desc->HostEntriesBegin); 488 if (TransTable != PM->HostEntriesBeginToTransTable.end()) { 489 DP("Removing translation table for descriptor " DPxMOD "\n", 490 DPxPTR(desc->HostEntriesBegin)); 491 PM->HostEntriesBeginToTransTable.erase(TransTable); 492 } else { 493 DP("Translation table for descriptor " DPxMOD " cannot be found, probably " 494 "it has been already removed.\n", 495 DPxPTR(desc->HostEntriesBegin)); 496 } 497 498 PM->TblMapMtx.unlock(); 499 500 // TODO: Remove RTL and the devices it manages if it's not used anymore? 501 // TODO: Write some RTL->unload_image(...) function? 502 503 DP("Done unregistering library!\n"); 504 } 505