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