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