openmp/libomptarget/deviceRTLs/interface.h

   1 //===------- interface.h - OpenMP interface definitions ---------- CUDA -*-===//
   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 //  This file contains all the definitions that are relevant to
  10 //  the interface. The first section contains the interface as
  11 //  declared by OpenMP.  The second section includes the compiler
  12 //  specific interfaces.
  13 //
  14 //===----------------------------------------------------------------------===//
  15
  16 #ifndef _INTERFACES_H_
  17 #define _INTERFACES_H_
  18
  19 #include <stddef.h>
  20 #include <stdint.h>
  21
  22 #ifdef __AMDGCN__
  23 #include "amdgcn/src/amdgcn_interface.h"
  24 #endif
  25 #ifdef __CUDACC__
  26 #include "nvptx/src/nvptx_interface.h"
  27 #endif
  28
  29 ////////////////////////////////////////////////////////////////////////////////
  30 // OpenMP interface
  31 ////////////////////////////////////////////////////////////////////////////////
  32
  33 typedef uint64_t omp_nest_lock_t; /* arbitrary type of the right length */
  34
  35 typedef enum omp_sched_t {
  36   omp_sched_static = 1,  /* chunkSize >0 */
  37   omp_sched_dynamic = 2, /* chunkSize >0 */
  38   omp_sched_guided = 3,  /* chunkSize >0 */
  39   omp_sched_auto = 4,    /* no chunkSize */
  40 } omp_sched_t;
  41
  42 typedef enum omp_proc_bind_t {
  43   omp_proc_bind_false = 0,
  44   omp_proc_bind_true = 1,
  45   omp_proc_bind_master = 2,
  46   omp_proc_bind_close = 3,
  47   omp_proc_bind_spread = 4
  48 } omp_proc_bind_t;
  49
  50 EXTERN double omp_get_wtick(void);
  51 EXTERN double omp_get_wtime(void);
  52
  53 EXTERN void omp_set_num_threads(int num);
  54 EXTERN int omp_get_num_threads(void);
  55 EXTERN int omp_get_max_threads(void);
  56 EXTERN int omp_get_thread_limit(void);
  57 EXTERN int omp_get_thread_num(void);
  58 EXTERN int omp_get_num_procs(void);
  59 EXTERN int omp_in_parallel(void);
  60 EXTERN int omp_in_final(void);
  61 EXTERN void omp_set_dynamic(int flag);
  62 EXTERN int omp_get_dynamic(void);
  63 EXTERN void omp_set_nested(int flag);
  64 EXTERN int omp_get_nested(void);
  65 EXTERN void omp_set_max_active_levels(int level);
  66 EXTERN int omp_get_max_active_levels(void);
  67 EXTERN int omp_get_level(void);
  68 EXTERN int omp_get_active_level(void);
  69 EXTERN int omp_get_ancestor_thread_num(int level);
  70 EXTERN int omp_get_team_size(int level);
  71
  72 EXTERN void omp_init_lock(omp_lock_t *lock);
  73 EXTERN void omp_init_nest_lock(omp_nest_lock_t *lock);
  74 EXTERN void omp_destroy_lock(omp_lock_t *lock);
  75 EXTERN void omp_destroy_nest_lock(omp_nest_lock_t *lock);
  76 EXTERN void omp_set_lock(omp_lock_t *lock);
  77 EXTERN void omp_set_nest_lock(omp_nest_lock_t *lock);
  78 EXTERN void omp_unset_lock(omp_lock_t *lock);
  79 EXTERN void omp_unset_nest_lock(omp_nest_lock_t *lock);
  80 EXTERN int omp_test_lock(omp_lock_t *lock);
  81 EXTERN int omp_test_nest_lock(omp_nest_lock_t *lock);
  82
  83 EXTERN void omp_get_schedule(omp_sched_t *kind, int *modifier);
  84 EXTERN void omp_set_schedule(omp_sched_t kind, int modifier);
  85 EXTERN omp_proc_bind_t omp_get_proc_bind(void);
  86 EXTERN int omp_get_cancellation(void);
  87 EXTERN void omp_set_default_device(int deviceId);
  88 EXTERN int omp_get_default_device(void);
  89 EXTERN int omp_get_num_devices(void);
  90 EXTERN int omp_get_num_teams(void);
  91 EXTERN int omp_get_team_num(void);
  92 EXTERN int omp_get_initial_device(void);
  93 EXTERN int omp_get_max_task_priority(void);
  94
  95 EXTERN void *llvm_omp_get_dynamic_shared();
  96
  97 ////////////////////////////////////////////////////////////////////////////////
  98 // file below is swiped from kmpc host interface
  99 ////////////////////////////////////////////////////////////////////////////////
 100
 101 ////////////////////////////////////////////////////////////////////////////////
 102 // kmp specific types
 103 ////////////////////////////////////////////////////////////////////////////////
 104
 105 typedef enum kmp_sched_t {
 106   kmp_sched_static_chunk = 33,
 107   kmp_sched_static_nochunk = 34,
 108   kmp_sched_dynamic = 35,
 109   kmp_sched_guided = 36,
 110   kmp_sched_runtime = 37,
 111   kmp_sched_auto = 38,
 112
 113   kmp_sched_static_balanced_chunk = 45,
 114
 115   kmp_sched_static_ordered = 65,
 116   kmp_sched_static_nochunk_ordered = 66,
 117   kmp_sched_dynamic_ordered = 67,
 118   kmp_sched_guided_ordered = 68,
 119   kmp_sched_runtime_ordered = 69,
 120   kmp_sched_auto_ordered = 70,
 121
 122   kmp_sched_distr_static_chunk = 91,
 123   kmp_sched_distr_static_nochunk = 92,
 124   kmp_sched_distr_static_chunk_sched_static_chunkone = 93,
 125
 126   kmp_sched_default = kmp_sched_static_nochunk,
 127   kmp_sched_unordered_first = kmp_sched_static_chunk,
 128   kmp_sched_unordered_last = kmp_sched_auto,
 129   kmp_sched_ordered_first = kmp_sched_static_ordered,
 130   kmp_sched_ordered_last = kmp_sched_auto_ordered,
 131   kmp_sched_distribute_first = kmp_sched_distr_static_chunk,
 132   kmp_sched_distribute_last =
 133       kmp_sched_distr_static_chunk_sched_static_chunkone,
 134
 135   /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers.
 136    * Since we need to distinguish the three possible cases (no modifier,
 137    * monotonic modifier, nonmonotonic modifier), we need separate bits for
 138    * each modifier. The absence of monotonic does not imply nonmonotonic,
 139    * especially since 4.5 says that the behaviour of the "no modifier" case
 140    * is implementation defined in 4.5, but will become "nonmonotonic" in 5.0.
 141    *
 142    * Since we're passing a full 32 bit value, we can use a couple of high
 143    * bits for these flags; out of paranoia we avoid the sign bit.
 144    *
 145    * These modifiers can be or-ed into non-static schedules by the compiler
 146    * to pass the additional information. They will be stripped early in the
 147    * processing in __kmp_dispatch_init when setting up schedules, so
 148    * most of the code won't ever see schedules with these bits set.
 149    */
 150   kmp_sched_modifier_monotonic = (1 << 29),
 151   /**< Set if the monotonic schedule modifier was present */
 152   kmp_sched_modifier_nonmonotonic = (1 << 30),
 153 /**< Set if the nonmonotonic schedule modifier was present */
 154
 155 #define SCHEDULE_WITHOUT_MODIFIERS(s)                                          \
 156   (enum kmp_sched_t)(                                                          \
 157       (s) & ~(kmp_sched_modifier_nonmonotonic | kmp_sched_modifier_monotonic))
 158 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sched_modifier_monotonic) != 0)
 159 #define SCHEDULE_HAS_NONMONOTONIC(s)                                           \
 160   (((s)&kmp_sched_modifier_nonmonotonic) != 0)
 161 #define SCHEDULE_HAS_NO_MODIFIERS(s)                                           \
 162   (((s) & (kmp_sched_modifier_nonmonotonic | kmp_sched_modifier_monotonic)) == \
 163    0)
 164
 165 } kmp_sched_t;
 166
 167 /*!
 168  * Enum for accesseing the reserved_2 field of the ident_t struct below.
 169  */
 170 enum {
 171   /*! Bit set to 1 when in SPMD mode. */
 172   KMP_IDENT_SPMD_MODE = 0x01,
 173   /*! Bit set to 1 when a simplified runtime is used. */
 174   KMP_IDENT_SIMPLE_RT_MODE = 0x02,
 175 };
 176
 177 /*!
 178  * The ident structure that describes a source location.
 179  * The struct is identical to the one in the kmp.h file.
 180  * We maintain the same data structure for compatibility.
 181  */
 182 typedef short kmp_int16;
 183 typedef int kmp_int32;
 184 typedef struct ident {
 185   kmp_int32 reserved_1; /**<  might be used in Fortran; see above  */
 186   kmp_int32 flags;      /**<  also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
 187                            identifies this union member  */
 188   kmp_int32 reserved_2; /**<  not really used in Fortran any more; see above */
 189   kmp_int32 reserved_3; /**<  source[4] in Fortran, do not use for C++  */
 190   char const *psource;  /**<  String describing the source location.
 191                         The string is composed of semi-colon separated fields
 192                         which describe the source file, the function and a pair
 193                         of line numbers that delimit the construct. */
 194 } ident_t;
 195
 196 // parallel defs
 197 typedef ident_t kmp_Ident;
 198 typedef void (*kmp_InterWarpCopyFctPtr)(void *src, int32_t warp_num);
 199 typedef void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t lane_id,
 200                                         int16_t lane_offset,
 201                                         int16_t shortCircuit);
 202 typedef void (*kmp_ListGlobalFctPtr)(void *buffer, int idx, void *reduce_data);
 203
 204 // task defs
 205 typedef struct kmp_TaskDescr kmp_TaskDescr;
 206 typedef int32_t (*kmp_TaskFctPtr)(int32_t global_tid, kmp_TaskDescr *taskDescr);
 207 typedef struct kmp_TaskDescr {
 208   void *sharedPointerTable;   // ptr to a table of shared var ptrs
 209   kmp_TaskFctPtr sub;         // task subroutine
 210   int32_t partId;             // unused
 211   kmp_TaskFctPtr destructors; // destructor of c++ first private
 212 } kmp_TaskDescr;
 213
 214 // sync defs
 215 typedef int32_t kmp_CriticalName[8];
 216
 217 ////////////////////////////////////////////////////////////////////////////////
 218 // external interface
 219 ////////////////////////////////////////////////////////////////////////////////
 220
 221 // parallel
 222 EXTERN int32_t __kmpc_global_thread_num(kmp_Ident *loc);
 223 NOINLINE EXTERN uint8_t __kmpc_parallel_level();
 224
 225 // proc bind
 226 EXTERN void __kmpc_push_proc_bind(kmp_Ident *loc, uint32_t global_tid,
 227                                   int proc_bind);
 228 EXTERN int omp_get_num_places(void);
 229 EXTERN int omp_get_place_num_procs(int place_num);
 230 EXTERN void omp_get_place_proc_ids(int place_num, int *ids);
 231 EXTERN int omp_get_place_num(void);
 232 EXTERN int omp_get_partition_num_places(void);
 233 EXTERN void omp_get_partition_place_nums(int *place_nums);
 234
 235 // for static (no chunk or chunk)
 236 EXTERN void __kmpc_for_static_init_4(kmp_Ident *loc, int32_t global_tid,
 237                                      int32_t sched, int32_t *plastiter,
 238                                      int32_t *plower, int32_t *pupper,
 239                                      int32_t *pstride, int32_t incr,
 240                                      int32_t chunk);
 241 EXTERN void __kmpc_for_static_init_4u(kmp_Ident *loc, int32_t global_tid,
 242                                       int32_t sched, int32_t *plastiter,
 243                                       uint32_t *plower, uint32_t *pupper,
 244                                       int32_t *pstride, int32_t incr,
 245                                       int32_t chunk);
 246 EXTERN void __kmpc_for_static_init_8(kmp_Ident *loc, int32_t global_tid,
 247                                      int32_t sched, int32_t *plastiter,
 248                                      int64_t *plower, int64_t *pupper,
 249                                      int64_t *pstride, int64_t incr,
 250                                      int64_t chunk);
 251 EXTERN void __kmpc_for_static_init_8u(kmp_Ident *loc, int32_t global_tid,
 252                                       int32_t sched, int32_t *plastiter1,
 253                                       uint64_t *plower, uint64_t *pupper,
 254                                       int64_t *pstride, int64_t incr,
 255                                       int64_t chunk);
 256 // distribute static (no chunk or chunk)
 257 EXTERN void __kmpc_distribute_static_init_4(kmp_Ident *loc, int32_t global_tid,
 258                                             int32_t sched, int32_t *plastiter,
 259                                             int32_t *plower, int32_t *pupper,
 260                                             int32_t *pstride, int32_t incr,
 261                                             int32_t chunk);
 262 EXTERN void __kmpc_distribute_static_init_4u(kmp_Ident *loc, int32_t global_tid,
 263                                              int32_t sched, int32_t *plastiter,
 264                                              uint32_t *plower, uint32_t *pupper,
 265                                              int32_t *pstride, int32_t incr,
 266                                              int32_t chunk);
 267 EXTERN void __kmpc_distribute_static_init_8(kmp_Ident *loc, int32_t global_tid,
 268                                             int32_t sched, int32_t *plastiter,
 269                                             int64_t *plower, int64_t *pupper,
 270                                             int64_t *pstride, int64_t incr,
 271                                             int64_t chunk);
 272 EXTERN void __kmpc_distribute_static_init_8u(kmp_Ident *loc, int32_t global_tid,
 273                                              int32_t sched, int32_t *plastiter1,
 274                                              uint64_t *plower, uint64_t *pupper,
 275                                              int64_t *pstride, int64_t incr,
 276                                              int64_t chunk);
 277 EXTERN
 278 void __kmpc_for_static_init_4_simple_spmd(kmp_Ident *loc, int32_t global_tid,
 279                                           int32_t sched, int32_t *plastiter,
 280                                           int32_t *plower, int32_t *pupper,
 281                                           int32_t *pstride, int32_t incr,
 282                                           int32_t chunk);
 283 EXTERN
 284 void __kmpc_for_static_init_4u_simple_spmd(kmp_Ident *loc, int32_t global_tid,
 285                                            int32_t sched, int32_t *plastiter,
 286                                            uint32_t *plower, uint32_t *pupper,
 287                                            int32_t *pstride, int32_t incr,
 288                                            int32_t chunk);
 289 EXTERN
 290 void __kmpc_for_static_init_8_simple_spmd(kmp_Ident *loc, int32_t global_tid,
 291                                           int32_t sched, int32_t *plastiter,
 292                                           int64_t *plower, int64_t *pupper,
 293                                           int64_t *pstride, int64_t incr,
 294                                           int64_t chunk);
 295 EXTERN
 296 void __kmpc_for_static_init_8u_simple_spmd(kmp_Ident *loc, int32_t global_tid,
 297                                            int32_t sched, int32_t *plastiter1,
 298                                            uint64_t *plower, uint64_t *pupper,
 299                                            int64_t *pstride, int64_t incr,
 300                                            int64_t chunk);
 301 EXTERN
 302 void __kmpc_for_static_init_4_simple_generic(kmp_Ident *loc, int32_t global_tid,
 303                                              int32_t sched, int32_t *plastiter,
 304                                              int32_t *plower, int32_t *pupper,
 305                                              int32_t *pstride, int32_t incr,
 306                                              int32_t chunk);
 307 EXTERN
 308 void __kmpc_for_static_init_4u_simple_generic(
 309     kmp_Ident *loc, int32_t global_tid, int32_t sched, int32_t *plastiter,
 310     uint32_t *plower, uint32_t *pupper, int32_t *pstride, int32_t incr,
 311     int32_t chunk);
 312 EXTERN
 313 void __kmpc_for_static_init_8_simple_generic(kmp_Ident *loc, int32_t global_tid,
 314                                              int32_t sched, int32_t *plastiter,
 315                                              int64_t *plower, int64_t *pupper,
 316                                              int64_t *pstride, int64_t incr,
 317                                              int64_t chunk);
 318 EXTERN
 319 void __kmpc_for_static_init_8u_simple_generic(
 320     kmp_Ident *loc, int32_t global_tid, int32_t sched, int32_t *plastiter1,
 321     uint64_t *plower, uint64_t *pupper, int64_t *pstride, int64_t incr,
 322     int64_t chunk);
 323
 324 EXTERN void __kmpc_for_static_fini(kmp_Ident *loc, int32_t global_tid);
 325
 326 EXTERN void __kmpc_distribute_static_fini(kmp_Ident *loc, int32_t global_tid);
 327
 328 // for dynamic
 329 EXTERN void __kmpc_dispatch_init_4(kmp_Ident *loc, int32_t global_tid,
 330                                    int32_t sched, int32_t lower, int32_t upper,
 331                                    int32_t incr, int32_t chunk);
 332 EXTERN void __kmpc_dispatch_init_4u(kmp_Ident *loc, int32_t global_tid,
 333                                     int32_t sched, uint32_t lower,
 334                                     uint32_t upper, int32_t incr,
 335                                     int32_t chunk);
 336 EXTERN void __kmpc_dispatch_init_8(kmp_Ident *loc, int32_t global_tid,
 337                                    int32_t sched, int64_t lower, int64_t upper,
 338                                    int64_t incr, int64_t chunk);
 339 EXTERN void __kmpc_dispatch_init_8u(kmp_Ident *loc, int32_t global_tid,
 340                                     int32_t sched, uint64_t lower,
 341                                     uint64_t upper, int64_t incr,
 342                                     int64_t chunk);
 343
 344 EXTERN int __kmpc_dispatch_next_4(kmp_Ident *loc, int32_t global_tid,
 345                                   int32_t *plastiter, int32_t *plower,
 346                                   int32_t *pupper, int32_t *pstride);
 347 EXTERN int __kmpc_dispatch_next_4u(kmp_Ident *loc, int32_t global_tid,
 348                                    int32_t *plastiter, uint32_t *plower,
 349                                    uint32_t *pupper, int32_t *pstride);
 350 EXTERN int __kmpc_dispatch_next_8(kmp_Ident *loc, int32_t global_tid,
 351                                   int32_t *plastiter, int64_t *plower,
 352                                   int64_t *pupper, int64_t *pstride);
 353 EXTERN int __kmpc_dispatch_next_8u(kmp_Ident *loc, int32_t global_tid,
 354                                    int32_t *plastiter, uint64_t *plower,
 355                                    uint64_t *pupper, int64_t *pstride);
 356
 357 EXTERN void __kmpc_dispatch_fini_4(kmp_Ident *loc, int32_t global_tid);
 358 EXTERN void __kmpc_dispatch_fini_4u(kmp_Ident *loc, int32_t global_tid);
 359 EXTERN void __kmpc_dispatch_fini_8(kmp_Ident *loc, int32_t global_tid);
 360 EXTERN void __kmpc_dispatch_fini_8u(kmp_Ident *loc, int32_t global_tid);
 361
 362 // reduction
 363 EXTERN void __kmpc_nvptx_end_reduce(int32_t global_tid);
 364 EXTERN void __kmpc_nvptx_end_reduce_nowait(int32_t global_tid);
 365 EXTERN int32_t __kmpc_nvptx_parallel_reduce_nowait_v2(
 366     kmp_Ident *loc, int32_t global_tid, int32_t num_vars, size_t reduce_size,
 367     void *reduce_data, kmp_ShuffleReductFctPtr shflFct,
 368     kmp_InterWarpCopyFctPtr cpyFct);
 369 EXTERN int32_t __kmpc_nvptx_teams_reduce_nowait_v2(
 370     kmp_Ident *loc, int32_t global_tid, void *global_buffer,
 371     int32_t num_of_records, void *reduce_data, kmp_ShuffleReductFctPtr shflFct,
 372     kmp_InterWarpCopyFctPtr cpyFct, kmp_ListGlobalFctPtr lgcpyFct,
 373     kmp_ListGlobalFctPtr lgredFct, kmp_ListGlobalFctPtr glcpyFct,
 374     kmp_ListGlobalFctPtr glredFct);
 375 EXTERN int32_t __kmpc_shuffle_int32(int32_t val, int16_t delta, int16_t size);
 376 EXTERN int64_t __kmpc_shuffle_int64(int64_t val, int16_t delta, int16_t size);
 377
 378 // sync barrier
 379 EXTERN void __kmpc_barrier(kmp_Ident *loc_ref, int32_t tid);
 380 EXTERN void __kmpc_barrier_simple_spmd(kmp_Ident *loc_ref, int32_t tid);
 381 EXTERN void __kmpc_barrier_simple_generic(kmp_Ident *loc_ref, int32_t tid);
 382 EXTERN int32_t __kmpc_cancel_barrier(kmp_Ident *loc, int32_t global_tid);
 383
 384 // single
 385 EXTERN int32_t __kmpc_single(kmp_Ident *loc, int32_t global_tid);
 386 EXTERN void __kmpc_end_single(kmp_Ident *loc, int32_t global_tid);
 387
 388 // sync
 389 EXTERN int32_t __kmpc_master(kmp_Ident *loc, int32_t global_tid);
 390 EXTERN void __kmpc_end_master(kmp_Ident *loc, int32_t global_tid);
 391 EXTERN void __kmpc_ordered(kmp_Ident *loc, int32_t global_tid);
 392 EXTERN void __kmpc_end_ordered(kmp_Ident *loc, int32_t global_tid);
 393 EXTERN void __kmpc_critical(kmp_Ident *loc, int32_t global_tid,
 394                             kmp_CriticalName *crit);
 395 EXTERN void __kmpc_end_critical(kmp_Ident *loc, int32_t global_tid,
 396                                 kmp_CriticalName *crit);
 397 EXTERN void __kmpc_flush(kmp_Ident *loc);
 398
 399 // vote
 400 EXTERN uint64_t __kmpc_warp_active_thread_mask(void);
 401 // syncwarp
 402 EXTERN void __kmpc_syncwarp(uint64_t);
 403
 404 // tasks
 405 EXTERN kmp_TaskDescr *__kmpc_omp_task_alloc(kmp_Ident *loc, uint32_t global_tid,
 406                                             int32_t flag,
 407                                             size_t sizeOfTaskInclPrivate,
 408                                             size_t sizeOfSharedTable,
 409                                             kmp_TaskFctPtr sub);
 410 EXTERN int32_t __kmpc_omp_task(kmp_Ident *loc, uint32_t global_tid,
 411                                kmp_TaskDescr *newLegacyTaskDescr);
 412 EXTERN int32_t __kmpc_omp_task_with_deps(kmp_Ident *loc, uint32_t global_tid,
 413                                          kmp_TaskDescr *newLegacyTaskDescr,
 414                                          int32_t depNum, void *depList,
 415                                          int32_t noAliasDepNum,
 416                                          void *noAliasDepList);
 417 EXTERN void __kmpc_omp_task_begin_if0(kmp_Ident *loc, uint32_t global_tid,
 418                                       kmp_TaskDescr *newLegacyTaskDescr);
 419 EXTERN void __kmpc_omp_task_complete_if0(kmp_Ident *loc, uint32_t global_tid,
 420                                          kmp_TaskDescr *newLegacyTaskDescr);
 421 EXTERN void __kmpc_omp_wait_deps(kmp_Ident *loc, uint32_t global_tid,
 422                                  int32_t depNum, void *depList,
 423                                  int32_t noAliasDepNum, void *noAliasDepList);
 424 EXTERN void __kmpc_taskgroup(kmp_Ident *loc, uint32_t global_tid);
 425 EXTERN void __kmpc_end_taskgroup(kmp_Ident *loc, uint32_t global_tid);
 426 EXTERN int32_t __kmpc_omp_taskyield(kmp_Ident *loc, uint32_t global_tid,
 427                                     int end_part);
 428 EXTERN int32_t __kmpc_omp_taskwait(kmp_Ident *loc, uint32_t global_tid);
 429 EXTERN void __kmpc_taskloop(kmp_Ident *loc, uint32_t global_tid,
 430                             kmp_TaskDescr *newKmpTaskDescr, int if_val,
 431                             uint64_t *lb, uint64_t *ub, int64_t st, int nogroup,
 432                             int32_t sched, uint64_t grainsize, void *task_dup);
 433
 434 // cancel
 435 EXTERN int32_t __kmpc_cancellationpoint(kmp_Ident *loc, int32_t global_tid,
 436                                         int32_t cancelVal);
 437 EXTERN int32_t __kmpc_cancel(kmp_Ident *loc, int32_t global_tid,
 438                              int32_t cancelVal);
 439
 440 // non standard
 441 EXTERN int32_t __kmpc_target_init(ident_t *Ident, int8_t Mode,
 442                                   bool UseGenericStateMachine,
 443                                   bool RequiresFullRuntime);
 444 EXTERN void __kmpc_target_deinit(ident_t *Ident, int8_t Mode,
 445                                  bool RequiresFullRuntime);
 446 EXTERN void __kmpc_kernel_prepare_parallel(void *WorkFn,
 447                                            int32_t NumThreadsClause);
 448 EXTERN bool __kmpc_kernel_parallel(void **WorkFn);
 449 EXTERN void __kmpc_kernel_end_parallel();
 450
 451 EXTERN void __kmpc_data_sharing_init_stack();
 452 EXTERN void __kmpc_begin_sharing_variables(void ***GlobalArgs, size_t nArgs);
 453 EXTERN void __kmpc_end_sharing_variables();
 454 EXTERN void __kmpc_get_shared_variables(void ***GlobalArgs);
 455
 456 /// Entry point to start a new parallel region.
 457 ///
 458 /// \param ident       The source identifier.
 459 /// \param global_tid  The global thread ID.
 460 /// \param if_expr     The if(expr), or 1 if none given.
 461 /// \param num_threads The num_threads(expr), or -1 if none given.
 462 /// \param proc_bind   The proc_bind, or `proc_bind_default` if none given.
 463 /// \param fn          The outlined parallel region function.
 464 /// \param wrapper_fn  The worker wrapper function of fn.
 465 /// \param args        The pointer array of arguments to fn.
 466 /// \param nargs       The number of arguments to fn.
 467 NOINLINE EXTERN void __kmpc_parallel_51(ident_t *ident, kmp_int32 global_tid,
 468                                         kmp_int32 if_expr,
 469                                         kmp_int32 num_threads, int proc_bind,
 470                                         void *fn, void *wrapper_fn, void **args,
 471                                         size_t nargs);
 472
 473 // SPMD execution mode interrogation function.
 474 EXTERN int8_t __kmpc_is_spmd_exec_mode();
 475
 476 /// Return true if the hardware thread id \p Tid represents the OpenMP main
 477 /// thread in generic mode outside of a parallel region.
 478 EXTERN int8_t __kmpc_is_generic_main_thread(kmp_int32 Tid);
 479
 480 /// Return true if the hardware thread id \p Tid represents the OpenMP main
 481 /// thread in generic mode.
 482 EXTERN int8_t __kmpc_is_generic_main_thread_id(kmp_int32 Tid);
 483
 484 EXTERN void __kmpc_get_team_static_memory(int16_t isSPMDExecutionMode,
 485                                           const void *buf, size_t size,
 486                                           int16_t is_shared, const void **res);
 487
 488 EXTERN void __kmpc_restore_team_static_memory(int16_t isSPMDExecutionMode,
 489                                               int16_t is_shared);
 490
 491 /// Allocate \p Bytes in "shareable" memory and return the address. Needs to be
 492 /// called balanced with __kmpc_free_shared like a stack (push/pop). Can be
 493 /// called by any thread, allocation happens per-thread.
 494 EXTERN void *__kmpc_alloc_shared(uint64_t Bytes);
 495
 496 /// Deallocate \p Ptr. Needs to be called balanced with __kmpc_alloc_shared like
 497 /// a stack (push/pop). Can be called by any thread. \p Ptr must be allocated by
 498 /// __kmpc_alloc_shared by the same thread. \p Bytes contains the size of the
 499 /// paired allocation to make memory management easier.
 500 EXTERN void __kmpc_free_shared(void *Ptr, size_t Bytes);
 501
 502 /// Get a pointer to the dynamic shared memory buffer in the device.
 503 EXTERN void *__kmpc_get_dynamic_shared();
 504
 505 #endif