|
| 1 | +//===-RTLs/generic-64bit/src/rtl.cpp - Target RTLs Implementation - C++ -*-===// |
| 2 | +// |
| 3 | +// The LLVM Compiler Infrastructure |
| 4 | +// |
| 5 | +// This file is dual licensed under the MIT and the University of Illinois Open |
| 6 | +// Source Licenses. See LICENSE.txt for details. |
| 7 | +// |
| 8 | +//===----------------------------------------------------------------------===// |
| 9 | +// |
| 10 | +// RTL for generic 64-bit machine |
| 11 | +// |
| 12 | +//===----------------------------------------------------------------------===// |
| 13 | + |
| 14 | +#include <cassert> |
| 15 | +#include <cstdio> |
| 16 | +#include <cstring> |
| 17 | +#include <cstdlib> |
| 18 | +#include <dlfcn.h> |
| 19 | +#include <ffi.h> |
| 20 | +#include <gelf.h> |
| 21 | +#include <link.h> |
| 22 | +#include <list> |
| 23 | +#include <vector> |
| 24 | + |
| 25 | +#include "omptarget.h" |
| 26 | + |
| 27 | +#ifndef TARGET_NAME |
| 28 | +#define TARGET_NAME Generic ELF - 64bit |
| 29 | +#endif |
| 30 | + |
| 31 | +#ifndef TARGET_ELF_ID |
| 32 | +#define TARGET_ELF_ID 0 |
| 33 | +#endif |
| 34 | + |
| 35 | +#define GETNAME2(name) #name |
| 36 | +#define GETNAME(name) GETNAME2(name) |
| 37 | +#define DP(...) DEBUGP("Target " GETNAME(TARGET_NAME) " RTL", __VA_ARGS__) |
| 38 | + |
| 39 | +#include "../../common/elf_common.c" |
| 40 | + |
| 41 | +#define NUMBER_OF_DEVICES 4 |
| 42 | +#define OFFLOADSECTIONNAME ".omp_offloading.entries" |
| 43 | + |
| 44 | +/// Array of Dynamic libraries loaded for this target. |
| 45 | +struct DynLibTy { |
| 46 | + char *FileName; |
| 47 | + void *Handle; |
| 48 | +}; |
| 49 | + |
| 50 | +/// Keep entries table per device. |
| 51 | +struct FuncOrGblEntryTy { |
| 52 | + __tgt_target_table Table; |
| 53 | +}; |
| 54 | + |
| 55 | +/// Class containing all the device information. |
| 56 | +class RTLDeviceInfoTy { |
| 57 | + std::vector<FuncOrGblEntryTy> FuncGblEntries; |
| 58 | + |
| 59 | +public: |
| 60 | + std::list<DynLibTy> DynLibs; |
| 61 | + |
| 62 | + // Record entry point associated with device. |
| 63 | + void createOffloadTable(int32_t device_id, __tgt_offload_entry *begin, |
| 64 | + __tgt_offload_entry *end) { |
| 65 | + assert(device_id < (int32_t)FuncGblEntries.size() && |
| 66 | + "Unexpected device id!"); |
| 67 | + FuncOrGblEntryTy &E = FuncGblEntries[device_id]; |
| 68 | + |
| 69 | + E.Table.EntriesBegin = begin; |
| 70 | + E.Table.EntriesEnd = end; |
| 71 | + } |
| 72 | + |
| 73 | + // Return true if the entry is associated with device. |
| 74 | + bool findOffloadEntry(int32_t device_id, void *addr) { |
| 75 | + assert(device_id < (int32_t)FuncGblEntries.size() && |
| 76 | + "Unexpected device id!"); |
| 77 | + FuncOrGblEntryTy &E = FuncGblEntries[device_id]; |
| 78 | + |
| 79 | + for (__tgt_offload_entry *i = E.Table.EntriesBegin, *e = E.Table.EntriesEnd; |
| 80 | + i < e; ++i) { |
| 81 | + if (i->addr == addr) |
| 82 | + return true; |
| 83 | + } |
| 84 | + |
| 85 | + return false; |
| 86 | + } |
| 87 | + |
| 88 | + // Return the pointer to the target entries table. |
| 89 | + __tgt_target_table *getOffloadEntriesTable(int32_t device_id) { |
| 90 | + assert(device_id < (int32_t)FuncGblEntries.size() && |
| 91 | + "Unexpected device id!"); |
| 92 | + FuncOrGblEntryTy &E = FuncGblEntries[device_id]; |
| 93 | + |
| 94 | + return &E.Table; |
| 95 | + } |
| 96 | + |
| 97 | + RTLDeviceInfoTy(int32_t num_devices) { FuncGblEntries.resize(num_devices); } |
| 98 | + |
| 99 | + ~RTLDeviceInfoTy() { |
| 100 | + // Close dynamic libraries |
| 101 | + for (auto &lib : DynLibs) { |
| 102 | + if (lib.Handle) { |
| 103 | + dlclose(lib.Handle); |
| 104 | + remove(lib.FileName); |
| 105 | + } |
| 106 | + } |
| 107 | + } |
| 108 | +}; |
| 109 | + |
| 110 | +static RTLDeviceInfoTy DeviceInfo(NUMBER_OF_DEVICES); |
| 111 | + |
| 112 | +#ifdef __cplusplus |
| 113 | +extern "C" { |
| 114 | +#endif |
| 115 | + |
| 116 | +int32_t __tgt_rtl_is_valid_binary(__tgt_device_image *image) { |
| 117 | +// If we don't have a valid ELF ID we can just fail. |
| 118 | +#if TARGET_ELF_ID < 1 |
| 119 | + return 0; |
| 120 | +#else |
| 121 | + return elf_check_machine(image, TARGET_ELF_ID); |
| 122 | +#endif |
| 123 | +} |
| 124 | + |
| 125 | +int32_t __tgt_rtl_number_of_devices() { return NUMBER_OF_DEVICES; } |
| 126 | + |
| 127 | +int32_t __tgt_rtl_init_device(int32_t device_id) { return OFFLOAD_SUCCESS; } |
| 128 | + |
| 129 | +__tgt_target_table *__tgt_rtl_load_binary(int32_t device_id, |
| 130 | + __tgt_device_image *image) { |
| 131 | + |
| 132 | + DP("Dev %d: load binary from " DPxMOD " image\n", device_id, |
| 133 | + DPxPTR(image->ImageStart)); |
| 134 | + |
| 135 | + assert(device_id >= 0 && device_id < NUMBER_OF_DEVICES && "bad dev id"); |
| 136 | + |
| 137 | + size_t ImageSize = (size_t)image->ImageEnd - (size_t)image->ImageStart; |
| 138 | + size_t NumEntries = (size_t)(image->EntriesEnd - image->EntriesBegin); |
| 139 | + DP("Expecting to have %zd entries defined.\n", NumEntries); |
| 140 | + |
| 141 | + // Is the library version incompatible with the header file? |
| 142 | + if (elf_version(EV_CURRENT) == EV_NONE) { |
| 143 | + DP("Incompatible ELF library!\n"); |
| 144 | + return NULL; |
| 145 | + } |
| 146 | + |
| 147 | + // Obtain elf handler |
| 148 | + Elf *e = elf_memory((char *)image->ImageStart, ImageSize); |
| 149 | + if (!e) { |
| 150 | + DP("Unable to get ELF handle: %s!\n", elf_errmsg(-1)); |
| 151 | + return NULL; |
| 152 | + } |
| 153 | + |
| 154 | + if (elf_kind(e) != ELF_K_ELF) { |
| 155 | + DP("Invalid Elf kind!\n"); |
| 156 | + elf_end(e); |
| 157 | + return NULL; |
| 158 | + } |
| 159 | + |
| 160 | + // Find the entries section offset |
| 161 | + Elf_Scn *section = 0; |
| 162 | + Elf64_Off entries_offset = 0; |
| 163 | + |
| 164 | + size_t shstrndx; |
| 165 | + |
| 166 | + if (elf_getshdrstrndx(e, &shstrndx)) { |
| 167 | + DP("Unable to get ELF strings index!\n"); |
| 168 | + elf_end(e); |
| 169 | + return NULL; |
| 170 | + } |
| 171 | + |
| 172 | + while ((section = elf_nextscn(e, section))) { |
| 173 | + GElf_Shdr hdr; |
| 174 | + gelf_getshdr(section, &hdr); |
| 175 | + |
| 176 | + if (!strcmp(elf_strptr(e, shstrndx, hdr.sh_name), OFFLOADSECTIONNAME)) { |
| 177 | + entries_offset = hdr.sh_addr; |
| 178 | + break; |
| 179 | + } |
| 180 | + } |
| 181 | + |
| 182 | + if (!entries_offset) { |
| 183 | + DP("Entries Section Offset Not Found\n"); |
| 184 | + elf_end(e); |
| 185 | + return NULL; |
| 186 | + } |
| 187 | + |
| 188 | + DP("Offset of entries section is (" DPxMOD ").\n", DPxPTR(entries_offset)); |
| 189 | + |
| 190 | + // load dynamic library and get the entry points. We use the dl library |
| 191 | + // to do the loading of the library, but we could do it directly to avoid the |
| 192 | + // dump to the temporary file. |
| 193 | + // |
| 194 | + // 1) Create tmp file with the library contents. |
| 195 | + // 2) Use dlopen to load the file and dlsym to retrieve the symbols. |
| 196 | + char tmp_name[] = "/tmp/tmpfile_XXXXXX"; |
| 197 | + int tmp_fd = mkstemp(tmp_name); |
| 198 | + |
| 199 | + if (tmp_fd == -1) { |
| 200 | + elf_end(e); |
| 201 | + return NULL; |
| 202 | + } |
| 203 | + |
| 204 | + FILE *ftmp = fdopen(tmp_fd, "wb"); |
| 205 | + |
| 206 | + if (!ftmp) { |
| 207 | + elf_end(e); |
| 208 | + return NULL; |
| 209 | + } |
| 210 | + |
| 211 | + fwrite(image->ImageStart, ImageSize, 1, ftmp); |
| 212 | + fclose(ftmp); |
| 213 | + |
| 214 | + DynLibTy Lib = {tmp_name, dlopen(tmp_name, RTLD_LAZY)}; |
| 215 | + |
| 216 | + if (!Lib.Handle) { |
| 217 | + DP("Target library loading error: %s\n", dlerror()); |
| 218 | + elf_end(e); |
| 219 | + return NULL; |
| 220 | + } |
| 221 | + |
| 222 | + DeviceInfo.DynLibs.push_back(Lib); |
| 223 | + |
| 224 | + struct link_map *libInfo = (struct link_map *)Lib.Handle; |
| 225 | + |
| 226 | + // The place where the entries info is loaded is the library base address |
| 227 | + // plus the offset determined from the ELF file. |
| 228 | + Elf64_Addr entries_addr = libInfo->l_addr + entries_offset; |
| 229 | + |
| 230 | + DP("Pointer to first entry to be loaded is (" DPxMOD ").\n", |
| 231 | + DPxPTR(entries_addr)); |
| 232 | + |
| 233 | + // Table of pointers to all the entries in the target. |
| 234 | + __tgt_offload_entry *entries_table = (__tgt_offload_entry *)entries_addr; |
| 235 | + |
| 236 | + __tgt_offload_entry *entries_begin = &entries_table[0]; |
| 237 | + __tgt_offload_entry *entries_end = entries_begin + NumEntries; |
| 238 | + |
| 239 | + if (!entries_begin) { |
| 240 | + DP("Can't obtain entries begin\n"); |
| 241 | + elf_end(e); |
| 242 | + return NULL; |
| 243 | + } |
| 244 | + |
| 245 | + DP("Entries table range is (" DPxMOD ")->(" DPxMOD ")\n", |
| 246 | + DPxPTR(entries_begin), DPxPTR(entries_end)); |
| 247 | + DeviceInfo.createOffloadTable(device_id, entries_begin, entries_end); |
| 248 | + |
| 249 | + elf_end(e); |
| 250 | + |
| 251 | + return DeviceInfo.getOffloadEntriesTable(device_id); |
| 252 | +} |
| 253 | + |
| 254 | +void *__tgt_rtl_data_alloc(int32_t device_id, int64_t size) { |
| 255 | + void *ptr = malloc(size); |
| 256 | + return ptr; |
| 257 | +} |
| 258 | + |
| 259 | +int32_t __tgt_rtl_data_submit(int32_t device_id, void *tgt_ptr, void *hst_ptr, |
| 260 | + int64_t size) { |
| 261 | + memcpy(tgt_ptr, hst_ptr, size); |
| 262 | + return OFFLOAD_SUCCESS; |
| 263 | +} |
| 264 | + |
| 265 | +int32_t __tgt_rtl_data_retrieve(int32_t device_id, void *hst_ptr, void *tgt_ptr, |
| 266 | + int64_t size) { |
| 267 | + memcpy(hst_ptr, tgt_ptr, size); |
| 268 | + return OFFLOAD_SUCCESS; |
| 269 | +} |
| 270 | + |
| 271 | +int32_t __tgt_rtl_data_delete(int32_t device_id, void *tgt_ptr) { |
| 272 | + free(tgt_ptr); |
| 273 | + return OFFLOAD_SUCCESS; |
| 274 | +} |
| 275 | + |
| 276 | +int32_t __tgt_rtl_run_target_team_region(int32_t device_id, void *tgt_entry_ptr, |
| 277 | + void **tgt_args, int32_t arg_num, int32_t team_num, int32_t thread_limit, |
| 278 | + uint64_t loop_tripcount /*not used*/) { |
| 279 | + // ignore team num and thread limit. |
| 280 | + |
| 281 | + // Use libffi to launch execution. |
| 282 | + ffi_cif cif; |
| 283 | + |
| 284 | + // All args are references. |
| 285 | + std::vector<ffi_type *> args_types(arg_num, &ffi_type_pointer); |
| 286 | + std::vector<void *> args(arg_num); |
| 287 | + |
| 288 | + for (int32_t i = 0; i < arg_num; ++i) |
| 289 | + args[i] = &tgt_args[i]; |
| 290 | + |
| 291 | + ffi_status status = ffi_prep_cif(&cif, FFI_DEFAULT_ABI, arg_num, |
| 292 | + &ffi_type_void, &args_types[0]); |
| 293 | + |
| 294 | + assert(status == FFI_OK && "Unable to prepare target launch!"); |
| 295 | + |
| 296 | + if (status != FFI_OK) |
| 297 | + return OFFLOAD_FAIL; |
| 298 | + |
| 299 | + DP("Running entry point at " DPxMOD "...\n", DPxPTR(tgt_entry_ptr)); |
| 300 | + |
| 301 | + ffi_call(&cif, FFI_FN(tgt_entry_ptr), NULL, &args[0]); |
| 302 | + return OFFLOAD_SUCCESS; |
| 303 | +} |
| 304 | + |
| 305 | +int32_t __tgt_rtl_run_target_region(int32_t device_id, void *tgt_entry_ptr, |
| 306 | + void **tgt_args, int32_t arg_num) { |
| 307 | + // use one team and one thread. |
| 308 | + return __tgt_rtl_run_target_team_region(device_id, tgt_entry_ptr, tgt_args, |
| 309 | + arg_num, 1, 1, 0); |
| 310 | +} |
| 311 | + |
| 312 | +#ifdef __cplusplus |
| 313 | +} |
| 314 | +#endif |
0 commit comments