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