Jim Cownie | 33f7b24 | 2014-04-09 15:40:23 +0000 | [diff] [blame] | 1 | //===----------------------------------------------------------------------===// |
| 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 | |
| 11 | #include "compiler_if_host.h" |
| 12 | |
| 13 | #include <malloc.h> |
| 14 | #ifndef TARGET_WINNT |
| 15 | #include <alloca.h> |
| 16 | #endif // TARGET_WINNT |
| 17 | |
| 18 | // Global counter on host. |
| 19 | // This variable is used if P2OPT_offload_do_data_persistence == 2. |
| 20 | // The variable used to identify offload constructs contained in one procedure. |
| 21 | // Increment of OFFLOAD_CALL_COUNT is inserted at entries of HOST routines with |
| 22 | // offload constructs. |
| 23 | static int offload_call_count = 0; |
| 24 | |
| 25 | extern "C" OFFLOAD OFFLOAD_TARGET_ACQUIRE( |
| 26 | TARGET_TYPE target_type, |
| 27 | int target_number, |
| 28 | int is_optional, |
| 29 | _Offload_status* status, |
| 30 | const char* file, |
| 31 | uint64_t line |
| 32 | ) |
| 33 | { |
| 34 | bool retval; |
| 35 | OFFLOAD ofld; |
| 36 | |
| 37 | // initialize status |
| 38 | if (status != 0) { |
| 39 | status->result = OFFLOAD_UNAVAILABLE; |
| 40 | status->device_number = -1; |
| 41 | status->data_sent = 0; |
| 42 | status->data_received = 0; |
| 43 | } |
| 44 | |
| 45 | // make sure libray is initialized |
| 46 | retval = __offload_init_library(); |
| 47 | |
| 48 | // OFFLOAD_TIMER_INIT must follow call to __offload_init_library |
| 49 | OffloadHostTimerData * timer_data = OFFLOAD_TIMER_INIT(file, line); |
| 50 | |
| 51 | OFFLOAD_TIMER_START(timer_data, c_offload_host_total_offload); |
| 52 | |
| 53 | OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize); |
| 54 | |
Alp Toker | c2d5e61 | 2014-06-01 18:28:36 +0000 | [diff] [blame] | 55 | // initialize all devices is init_type is on_offload_all |
Jim Cownie | 33f7b24 | 2014-04-09 15:40:23 +0000 | [diff] [blame] | 56 | if (retval && __offload_init_type == c_init_on_offload_all) { |
| 57 | for (int i = 0; i < mic_engines_total; i++) { |
| 58 | mic_engines[i].init(); |
| 59 | } |
| 60 | } |
| 61 | OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize); |
| 62 | |
| 63 | OFFLOAD_TIMER_START(timer_data, c_offload_host_target_acquire); |
| 64 | |
| 65 | if (target_type == TARGET_HOST) { |
| 66 | // Host always available |
| 67 | retval = true; |
| 68 | } |
| 69 | else if (target_type == TARGET_MIC) { |
| 70 | if (target_number >= -1) { |
| 71 | if (retval) { |
| 72 | if (target_number >= 0) { |
| 73 | // User provided the device number |
| 74 | target_number = target_number % mic_engines_total; |
| 75 | } |
| 76 | else { |
| 77 | // use device 0 |
| 78 | target_number = 0; |
| 79 | } |
| 80 | |
| 81 | // reserve device in ORSL |
| 82 | if (is_optional) { |
| 83 | if (!ORSL::try_reserve(target_number)) { |
| 84 | target_number = -1; |
| 85 | } |
| 86 | } |
| 87 | else { |
| 88 | if (!ORSL::reserve(target_number)) { |
| 89 | target_number = -1; |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | // initialize device |
| 94 | if (target_number >= 0 && |
| 95 | __offload_init_type == c_init_on_offload) { |
| 96 | OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize); |
| 97 | mic_engines[target_number].init(); |
| 98 | OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize); |
| 99 | } |
| 100 | } |
| 101 | else { |
| 102 | // fallback to CPU |
| 103 | target_number = -1; |
| 104 | } |
| 105 | |
| 106 | if (target_number < 0 || !retval) { |
| 107 | if (!is_optional && status == 0) { |
| 108 | LIBOFFLOAD_ERROR(c_device_is_not_available); |
| 109 | exit(1); |
| 110 | } |
| 111 | |
| 112 | retval = false; |
| 113 | } |
| 114 | } |
| 115 | else { |
| 116 | LIBOFFLOAD_ERROR(c_invalid_device_number); |
| 117 | exit(1); |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | if (retval) { |
| 122 | ofld = new OffloadDescriptor(target_number, status, |
| 123 | !is_optional, false, timer_data); |
| 124 | OFFLOAD_TIMER_HOST_MIC_NUM(timer_data, target_number); |
| 125 | Offload_Report_Prolog(timer_data); |
| 126 | OFFLOAD_DEBUG_TRACE_1(2, timer_data->offload_number, c_offload_start, |
| 127 | "Starting offload: target_type = %d, " |
| 128 | "number = %d, is_optional = %d\n", |
| 129 | target_type, target_number, is_optional); |
| 130 | |
| 131 | OFFLOAD_TIMER_STOP(timer_data, c_offload_host_target_acquire); |
| 132 | } |
| 133 | else { |
| 134 | ofld = NULL; |
| 135 | |
| 136 | OFFLOAD_TIMER_STOP(timer_data, c_offload_host_target_acquire); |
| 137 | OFFLOAD_TIMER_STOP(timer_data, c_offload_host_total_offload); |
| 138 | offload_report_free_data(timer_data); |
| 139 | } |
| 140 | |
| 141 | return ofld; |
| 142 | } |
| 143 | |
| 144 | extern "C" OFFLOAD OFFLOAD_TARGET_ACQUIRE1( |
| 145 | const int* device_num, |
| 146 | const char* file, |
| 147 | uint64_t line |
| 148 | ) |
| 149 | { |
| 150 | int target_number; |
| 151 | |
| 152 | // make sure libray is initialized and at least one device is available |
| 153 | if (!__offload_init_library()) { |
| 154 | LIBOFFLOAD_ERROR(c_device_is_not_available); |
| 155 | exit(1); |
| 156 | } |
| 157 | |
| 158 | // OFFLOAD_TIMER_INIT must follow call to __offload_init_library |
| 159 | |
| 160 | OffloadHostTimerData * timer_data = OFFLOAD_TIMER_INIT(file, line); |
| 161 | |
| 162 | OFFLOAD_TIMER_START(timer_data, c_offload_host_total_offload); |
| 163 | |
| 164 | OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize); |
| 165 | |
| 166 | if (__offload_init_type == c_init_on_offload_all) { |
| 167 | for (int i = 0; i < mic_engines_total; i++) { |
| 168 | mic_engines[i].init(); |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize); |
| 173 | |
| 174 | OFFLOAD_TIMER_START(timer_data, c_offload_host_target_acquire); |
| 175 | |
| 176 | // use default device number if it is not provided |
| 177 | if (device_num != 0) { |
| 178 | target_number = *device_num; |
| 179 | } |
| 180 | else { |
| 181 | target_number = __omp_device_num; |
| 182 | } |
| 183 | |
| 184 | // device number should be a non-negative integer value |
| 185 | if (target_number < 0) { |
| 186 | LIBOFFLOAD_ERROR(c_omp_invalid_device_num); |
| 187 | exit(1); |
| 188 | } |
| 189 | |
| 190 | // should we do this for OpenMP? |
| 191 | target_number %= mic_engines_total; |
| 192 | |
| 193 | // reserve device in ORSL |
| 194 | if (!ORSL::reserve(target_number)) { |
| 195 | LIBOFFLOAD_ERROR(c_device_is_not_available); |
| 196 | exit(1); |
| 197 | } |
| 198 | |
| 199 | // initialize device(s) |
| 200 | OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize); |
| 201 | |
| 202 | if (__offload_init_type == c_init_on_offload) { |
| 203 | mic_engines[target_number].init(); |
| 204 | } |
| 205 | |
| 206 | OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize); |
| 207 | |
| 208 | OFFLOAD ofld = |
| 209 | new OffloadDescriptor(target_number, 0, true, true, timer_data); |
| 210 | |
| 211 | OFFLOAD_TIMER_HOST_MIC_NUM(timer_data, target_number); |
| 212 | |
| 213 | Offload_Report_Prolog(timer_data); |
| 214 | |
| 215 | OFFLOAD_DEBUG_TRACE_1(2, timer_data->offload_number, c_offload_start, |
| 216 | "Starting OpenMP offload, device = %d\n", |
| 217 | target_number); |
| 218 | |
| 219 | OFFLOAD_TIMER_STOP(timer_data, c_offload_host_target_acquire); |
| 220 | |
| 221 | return ofld; |
| 222 | } |
| 223 | |
| 224 | int offload_offload_wrap( |
| 225 | OFFLOAD ofld, |
| 226 | const char *name, |
| 227 | int is_empty, |
| 228 | int num_vars, |
| 229 | VarDesc *vars, |
| 230 | VarDesc2 *vars2, |
| 231 | int num_waits, |
| 232 | const void **waits, |
| 233 | const void **signal, |
| 234 | int entry_id, |
| 235 | const void *stack_addr |
| 236 | ) |
| 237 | { |
| 238 | bool ret = ofld->offload(name, is_empty, vars, vars2, num_vars, |
| 239 | waits, num_waits, signal, entry_id, stack_addr); |
| 240 | if (!ret || signal == 0) { |
| 241 | delete ofld; |
| 242 | } |
| 243 | return ret; |
| 244 | } |
| 245 | |
| 246 | extern "C" int OFFLOAD_OFFLOAD1( |
| 247 | OFFLOAD ofld, |
| 248 | const char *name, |
| 249 | int is_empty, |
| 250 | int num_vars, |
| 251 | VarDesc *vars, |
| 252 | VarDesc2 *vars2, |
| 253 | int num_waits, |
| 254 | const void **waits, |
| 255 | const void **signal |
| 256 | ) |
| 257 | { |
| 258 | return offload_offload_wrap(ofld, name, is_empty, |
| 259 | num_vars, vars, vars2, |
| 260 | num_waits, waits, |
| 261 | signal, NULL, NULL); |
| 262 | } |
| 263 | |
| 264 | extern "C" int OFFLOAD_OFFLOAD2( |
| 265 | OFFLOAD ofld, |
| 266 | const char *name, |
| 267 | int is_empty, |
| 268 | int num_vars, |
| 269 | VarDesc *vars, |
| 270 | VarDesc2 *vars2, |
| 271 | int num_waits, |
| 272 | const void** waits, |
| 273 | const void** signal, |
| 274 | int entry_id, |
| 275 | const void *stack_addr |
| 276 | ) |
| 277 | { |
| 278 | return offload_offload_wrap(ofld, name, is_empty, |
| 279 | num_vars, vars, vars2, |
| 280 | num_waits, waits, |
| 281 | signal, entry_id, stack_addr); |
| 282 | } |
| 283 | |
| 284 | extern "C" int OFFLOAD_OFFLOAD( |
| 285 | OFFLOAD ofld, |
| 286 | const char *name, |
| 287 | int is_empty, |
| 288 | int num_vars, |
| 289 | VarDesc *vars, |
| 290 | VarDesc2 *vars2, |
| 291 | int num_waits, |
| 292 | const void **waits, |
| 293 | const void *signal, |
| 294 | int entry_id, |
| 295 | const void *stack_addr |
| 296 | ) |
| 297 | { |
| 298 | // signal is passed by reference now |
| 299 | const void **signal_new = (signal != 0) ? &signal : 0; |
| 300 | const void **waits_new = 0; |
| 301 | int num_waits_new = 0; |
| 302 | |
| 303 | // remove NULL values from the list of signals to wait for |
| 304 | if (num_waits > 0) { |
| 305 | waits_new = (const void**) alloca(sizeof(void*) * num_waits); |
| 306 | for (int i = 0; i < num_waits; i++) { |
| 307 | if (waits[i] != 0) { |
| 308 | waits_new[num_waits_new++] = waits[i]; |
| 309 | } |
| 310 | } |
| 311 | } |
| 312 | |
| 313 | return OFFLOAD_OFFLOAD1(ofld, name, is_empty, |
| 314 | num_vars, vars, vars2, |
| 315 | num_waits_new, waits_new, |
| 316 | signal_new); |
| 317 | } |
| 318 | |
| 319 | extern "C" int OFFLOAD_CALL_COUNT() |
| 320 | { |
| 321 | offload_call_count++; |
| 322 | return offload_call_count; |
| 323 | } |