codon/stdlib/openmp.codon

# Copyright (C) 2022-2024 Exaloop Inc. <https://exaloop.io>
# OpenMP interface
# Ref: https://github.com/llvm/llvm-project/tree/main/openmp

Routine = Function[[i32, cobj], i32]

_KMP_IDENT_IMB = 0x01
_KMP_IDENT_KMPC = 0x02
_KMP_IDENT_AUTOPAR = 0x08
_KMP_IDENT_ATOMIC_REDUCE = 0x10
_KMP_IDENT_BARRIER_EXPL = 0x20
_KMP_IDENT_BARRIER_IMPL = 0x0040
_KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0
_KMP_IDENT_BARRIER_IMPL_FOR = 0x0040
_KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0
_KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140
_KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0
_KMP_IDENT_WORK_LOOP = 0x200
_KMP_IDENT_WORK_SECTIONS = 0x400
_KMP_IDENT_WORK_DISTRIBUTE = 0x800
_KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000
_KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000
_KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000
_KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000
_KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000
_KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000

@tuple
class Lock:
    a1: i32
    a2: i32
    a3: i32
    a4: i32
    a5: i32
    a6: i32
    a7: i32
    a8: i32

    def __new__() -> Lock:
        z = i32(0)
        return Lock(z, z, z, z, z, z, z, z)

@tuple
class Ident:
    reserved_1: i32
    flags: i32
    reserved_2: i32
    reserved_3: i32
    psource: cobj

    def __new__(flags: int = 0, source: str = ";unknown;unknown;0;0;;") -> Ident:
        return Ident(i32(0), i32(flags | _KMP_IDENT_KMPC), i32(0), i32(0), source.ptr)

@tuple
class LRData:
    routine: Routine

@tuple
class Task:
    shareds: cobj
    routine: Routine
    flags: i32
    x: LRData
    y: LRData

@tuple
class TaskWithPrivates:
    task: Task
    data: T
    T: type

@tuple
class TaskReductionInput:
    reduce_shar: cobj
    reduce_orig: cobj
    reduce_size: int
    reduce_init: cobj
    reduce_fini: cobj
    reduce_comb: cobj
    flags: u32

    def __new__(reduce_shar, reduce_orig, reduce_size: int,
                reduce_init: cobj, reduce_comb: cobj):
        return TaskReductionInput(reduce_shar.as_byte(), reduce_orig.as_byte(), reduce_size,
                                  reduce_init, cobj(), reduce_comb, u32(0))

@tuple
class TaskReductionInputArray:
    len: int
    ptr: Ptr[TaskReductionInput]

    def __setitem__(self, idx: int, x: TaskReductionInput):
        self.ptr[idx] = x

_DEFAULT_IDENT = Ident()
_STATIC_LOOP_IDENT = Ident(_KMP_IDENT_WORK_LOOP)
_REDUCTION_IDENT = Ident(_KMP_IDENT_ATOMIC_REDUCE)

def _default_loc():
    return __ptr__(_DEFAULT_IDENT)

_default_loc()

def _static_loop_loc():
    return __ptr__(_STATIC_LOOP_IDENT)

_static_loop_loc()

def _reduction_loc():
    return __ptr__(_REDUCTION_IDENT)

_reduction_loc()

def _critical_begin(loc_ref: Ptr[Ident], gtid: int, lck: Ptr[Lock]):
    from C import __kmpc_critical(Ptr[Ident], i32, Ptr[Lock])
    __kmpc_critical(loc_ref, i32(gtid), lck)

def _critical_end(loc_ref: Ptr[Ident], gtid: int, lck: Ptr[Lock]):
    from C import __kmpc_end_critical(Ptr[Ident], i32, Ptr[Lock])
    __kmpc_end_critical(loc_ref, i32(gtid), lck)

def _single_begin(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_single(Ptr[Ident], i32) -> i32
    return int(__kmpc_single(loc_ref, i32(gtid)))

def _single_end(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_end_single(Ptr[Ident], i32)
    __kmpc_end_single(loc_ref, i32(gtid))

def _master_begin(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_master(Ptr[Ident], i32) -> i32
    return int(__kmpc_master(loc_ref, i32(gtid)))

def _master_end(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_end_master(Ptr[Ident], i32)
    __kmpc_end_master(loc_ref, i32(gtid))

def _ordered_begin(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_ordered(Ptr[Ident], i32)
    __kmpc_ordered(loc_ref, i32(gtid))

def _ordered_end(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_end_ordered(Ptr[Ident], i32)
    __kmpc_end_ordered(loc_ref, i32(gtid))

def _taskwait(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_omp_taskwait(Ptr[Ident], i32)
    __kmpc_omp_taskwait(loc_ref, i32(gtid))

def _taskgroup_begin(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_taskgroup(Ptr[Ident], i32)
    __kmpc_taskgroup(loc_ref, i32(gtid))

def _taskgroup_end(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_end_taskgroup(Ptr[Ident], i32)
    __kmpc_end_taskgroup(loc_ref, i32(gtid))

def _task_alloc_size(
    size_of_task: int,
    size_of_shareds: int,
):
    from C import __kmpc_omp_task_alloc_size(int, int) -> int
    return __kmpc_omp_task_alloc_size(size_of_task, size_of_shareds)

def _task_alloc(
    loc_ref: Ptr[Ident],
    gtid: int,
    flags: int,
    size_of_task: int,
    size_of_shareds: int,
    task_entry: Routine,
):
    from internal.gc import alloc
    from C import __kmpc_omp_task_alloc(Ptr[Ident], i32, i32, int, int, Routine, cobj) -> cobj

    taskdata = alloc(_task_alloc_size(size_of_task, size_of_shareds))
    return __kmpc_omp_task_alloc(
        loc_ref, i32(gtid), i32(flags), size_of_task, size_of_shareds, task_entry, taskdata
    )

def _task_run(loc_ref: Ptr[Ident], gtid: int, new_task: cobj):
    from C import __kmpc_omp_task(Ptr[Ident], i32, cobj) -> i32
    return int(__kmpc_omp_task(loc_ref, i32(gtid), new_task))

def _barrier(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_barrier(Ptr[Ident], i32)
    __kmpc_barrier(loc_ref, i32(gtid))

def _flush(loc_ref: Ptr[Ident]):
    from C import __kmpc_flush(Ptr[Ident])
    __kmpc_flush(loc_ref)

def flush():
    _flush(_default_loc())

def _static_init(
    loc_ref: Ptr[Ident], gtid: int, schedtype: int, loop: range, incr: int, chunk: int
):
    from C import __kmpc_for_static_init_8(Ptr[Ident], i32, i32, Ptr[i32], Ptr[int], Ptr[int], Ptr[int], int, int)
    last = i32(0)
    lower = 0
    upper = len(loop) - 1
    stride = 1
    __kmpc_for_static_init_8(
        loc_ref,
        i32(gtid),
        i32(schedtype),
        __ptr__(last),
        __ptr__(lower),
        __ptr__(upper),
        __ptr__(stride),
        incr,
        chunk,
    )
    return bool(last), range(loop._get(lower), loop._get(upper + 1), loop.step), stride

def _static_fini(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_for_static_fini(Ptr[Ident], i32)
    __kmpc_for_static_fini(loc_ref, i32(gtid))

def _dynamic_init(
    loc_ref: Ptr[Ident], gtid: int, schedtype: int, loop: range, chunk: int
):
    from C import __kmpc_dispatch_init_8(Ptr[Ident], i32, i32, int, int, int, int)
    lower = 0
    upper = len(loop) - 1
    stride = 1
    __kmpc_dispatch_init_8(
        loc_ref, i32(gtid), i32(schedtype), lower, upper, stride, chunk
    )

def _dynamic_next(loc_ref: Ptr[Ident], gtid: int, loop: range):
    from C import __kmpc_dispatch_next_8(Ptr[Ident], i32, Ptr[i32], Ptr[int], Ptr[int], Ptr[int]) -> i32
    last = i32(0)
    lower = 0
    upper = 0
    stride = 0
    more = __kmpc_dispatch_next_8(
        loc_ref,
        i32(gtid),
        __ptr__(last),
        __ptr__(lower),
        __ptr__(upper),
        __ptr__(stride),
    )
    return (
        bool(more),
        bool(last),
        range(loop._get(lower), loop._get(upper + 1), loop.step),
    )

def _dynamic_fini(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_dispatch_fini_8(Ptr[Ident], i32)
    __kmpc_dispatch_fini_8(loc_ref, i32(gtid))

def _reduce(
    loc_ref: Ptr[Ident],
    gtid: int,
    reduce_data: T,
    reduce_func: cobj,
    lck: cobj,
    T: type,
):
    from internal.gc import sizeof

    from C import __kmpc_reduce(Ptr[Ident], i32, i32, int, cobj, cobj, cobj) -> i32
    num_vars = staticlen(reduce_data)
    reduce_size = sizeof(T)
    return int(
        __kmpc_reduce(
            loc_ref,
            i32(gtid),
            i32(num_vars),
            reduce_size,
            __ptr__(reduce_data).as_byte(),
            reduce_func,
            lck,
        )
    )

def _end_reduce(loc_ref: Ptr[Ident], gtid: int, lck: cobj):
    from C import __kmpc_end_reduce(Ptr[Ident], i32, cobj)
    __kmpc_end_reduce(loc_ref, i32(gtid), lck)

def _reduce_nowait(
    loc_ref: Ptr[Ident],
    gtid: int,
    reduce_data: T,
    reduce_func: cobj,
    lck: Ptr[Lock],
    T: type,
):
    from internal.gc import sizeof

    from C import __kmpc_reduce_nowait(Ptr[Ident], i32, i32, int, cobj, cobj, Ptr[Lock]) -> i32
    num_vars = staticlen(reduce_data)
    reduce_size = sizeof(T)
    return int(
        __kmpc_reduce_nowait(
            loc_ref,
            i32(gtid),
            i32(num_vars),
            reduce_size,
            __ptr__(reduce_data).as_byte(),
            reduce_func,
            lck,
        )
    )

def _end_reduce_nowait(loc_ref: Ptr[Ident], gtid: int, lck: Ptr[Lock]):
    from C import __kmpc_end_reduce_nowait(Ptr[Ident], i32, Ptr[Lock])
    __kmpc_end_reduce_nowait(loc_ref, i32(gtid), lck)

def _taskred_init(loc_ref: Ptr[Ident], gtid: int, num: int, data):
    from C import __kmpc_taskred_modifier_init(Ptr[Ident], i32, i32, i32, cobj) -> cobj
    return __kmpc_taskred_modifier_init(loc_ref, i32(gtid), i32(0), i32(num), data.as_byte())

def _taskred_fini(loc_ref: Ptr[Ident], gtid: int):
    from C import __kmpc_task_reduction_modifier_fini(Ptr[Ident], i32, i32)
    __kmpc_task_reduction_modifier_fini(loc_ref, i32(gtid), i32(0))

# add tskgrp arg?
def _taskred_data(gtid: int, data):
    from C import __kmpc_task_reduction_get_th_data(i32, cobj, cobj) -> cobj
    T = type(data)
    return T(__kmpc_task_reduction_get_th_data(i32(gtid), cobj(), data.as_byte()))

def _fork_call(microtask: cobj, args):
    from C import __kmpc_fork_call(Ptr[Ident], i32, cobj, ...)
    loc_ref = _default_loc()  # TODO: pass real loc?
    __kmpc_fork_call(loc_ref, i32(1), microtask, __ptr__(args))

def _static_loop_outline_template(gtid_ptr: Ptr[i32], btid_ptr: Ptr[i32], args):
    @nonpure
    def _loop_step():
        return 1

    @nonpure
    def _loop_loc_and_gtid(
        loc_ref: Ptr[Ident], reduction_loc_ref: Ptr[Ident], gtid: int
    ):
        pass

    @nonpure
    def _loop_body_stub(i, args):
        pass

    @nonpure
    def _loop_schedule():
        return (1 << 30) | 35  # nonmonotonic, dynamic chunked

    @nonpure
    def _loop_shared_updates(args):
        pass

    @nonpure
    def _loop_reductions(args):
        pass

    chunk, start, stop, extra = args[0]
    step = _loop_step()
    gtid = int(gtid_ptr[0])
    loc_ref = _default_loc()
    static_loop_loc_ref = _static_loop_loc()
    reduction_loc_ref = _reduction_loc()
    _loop_loc_and_gtid(loc_ref, reduction_loc_ref, gtid)
    loop = range(start, stop, step)
    schedule = _loop_schedule()

    last, subloop, stride = _static_init(
        static_loop_loc_ref, gtid, schedtype=schedule, loop=loop, incr=1, chunk=1
    )
    i = subloop.start
    stop = min(subloop.stop, loop.stop) if step >= 0 else max(subloop.stop, loop.stop)

    while (step >= 0 and i < stop) or (step < 0 and i > stop):
        _loop_body_stub(i, extra)
        i += step
    _static_fini(static_loop_loc_ref, gtid)

    if last:
        _loop_shared_updates(extra)

    _loop_reductions(extra)

def _static_chunked_loop_outline_template(gtid_ptr: Ptr[i32], btid_ptr: Ptr[i32], args):
    @nonpure
    def _loop_step():
        return 1

    @nonpure
    def _loop_loc_and_gtid(
        loc_ref: Ptr[Ident], reduction_loc_ref: Ptr[Ident], gtid: int
    ):
        pass

    @nonpure
    def _loop_body_stub(i, args):
        pass

    @nonpure
    def _loop_schedule():
        return (1 << 30) | 35  # nonmonotonic, dynamic chunked

    @nonpure
    def _loop_shared_updates(args):
        pass

    @nonpure
    def _loop_reductions(args):
        pass

    chunk, start, stop, extra = args[0]
    step = _loop_step()
    gtid = int(gtid_ptr[0])
    loc_ref = _default_loc()
    static_loop_loc_ref = _static_loop_loc()
    reduction_loc_ref = _reduction_loc()
    _loop_loc_and_gtid(loc_ref, reduction_loc_ref, gtid)
    loop = range(start, stop, step)
    schedule = _loop_schedule()

    last, subloop, stride = _static_init(
        static_loop_loc_ref, gtid, schedtype=schedule, loop=loop, incr=1, chunk=chunk
    )
    start = subloop.start
    stop = min(subloop.stop, loop.stop) if step >= 0 else max(subloop.stop, loop.stop)

    while (step >= 0 and start < loop.stop) or (step < 0 and start > loop.stop):
        i = start
        while (step >= 0 and i < stop) or (step < 0 and i > stop):
            _loop_body_stub(i, extra)
            i += step

        start += stride * step
        stop += stride * step
        stop = min(stop, loop.stop) if step >= 0 else max(stop, loop.stop)
    _static_fini(static_loop_loc_ref, gtid)

    if last:
        _loop_shared_updates(extra)

    _loop_reductions(extra)

def _dynamic_loop_outline_template(gtid_ptr: Ptr[i32], btid_ptr: Ptr[i32], args):
    @nonpure
    def _loop_step():
        return 1

    @nonpure
    def _loop_loc_and_gtid(
        loc_ref: Ptr[Ident], reduction_loc_ref: Ptr[Ident], gtid: int
    ):
        pass

    @nonpure
    def _loop_body_stub(i, args):
        pass

    @nonpure
    def _loop_schedule():
        return (1 << 30) | 35  # nonmonotonic, dynamic chunked

    @nonpure
    def _loop_shared_updates(args):
        pass

    @nonpure
    def _loop_reductions(args):
        pass

    @nonpure
    def _loop_ordered():
        return False

    chunk, start, stop, extra = args[0]
    step = _loop_step()
    gtid = int(gtid_ptr[0])
    loc_ref = _default_loc()
    reduction_loc_ref = _reduction_loc()
    _loop_loc_and_gtid(loc_ref, reduction_loc_ref, gtid)
    loop = range(start, stop, step)
    schedule = _loop_schedule()
    ordered = _loop_ordered()

    _dynamic_init(loc_ref, gtid, schedtype=schedule, loop=loop, chunk=chunk)
    while True:
        more, last, subloop = _dynamic_next(loc_ref, gtid, loop)
        if not more:
            break
        i = subloop.start
        while (step >= 0 and i < subloop.stop) or (step < 0 and i > subloop.stop):
            _loop_body_stub(i, extra)
            i += step
            if ordered:
                _dynamic_fini(loc_ref, gtid)
        if last:
            _loop_shared_updates(extra)

    _loop_reductions(extra)

# P = privates; tuple of types
# S = shareds; tuple of pointers
def _spawn_and_run_task(
    loc_ref: Ptr[Ident], gtid: int, routine: cobj, priv: P, shared: S, P: type, S: type
):
    from internal.gc import sizeof

    TaskThunk = TaskWithPrivates[P]
    flags = 1
    size_of_kmp_task_t = sizeof(TaskThunk)
    size_of_privs = sizeof(P)
    size_of_shareds = sizeof(S)
    loc_ref = _default_loc()

    task = Ptr[TaskThunk](
        _task_alloc(
            loc_ref, gtid, flags, size_of_kmp_task_t, size_of_shareds, Routine(routine)
        )
    )
    if staticlen(shared) != 0:
        shared_ptr = task[0].task.shareds
        str.memcpy(shared_ptr, __ptr__(shared).as_byte(), size_of_shareds)
    if staticlen(priv) != 0:
        priv_ptr = task.as_byte() + sizeof(Task)
        str.memcpy(priv_ptr, __ptr__(priv).as_byte(), size_of_privs)

    _task_run(loc_ref, gtid, task.as_byte())

# Note: this is different than OpenMP's "taskloop" -- this template simply
# spawns a new task for each loop iteration.
def _task_loop_outline_template(gtid_ptr: Ptr[i32], btid_ptr: Ptr[i32], args):
    def _routine_stub(gtid: i32, data: cobj, P: type, S: type):
        @nonpure
        def _task_loop_body_stub(gtid: int, priv, shared):
            pass

        task = Ptr[TaskWithPrivates[P]](data)[0]
        priv = task.data
        gtid64 = int(gtid)
        if staticlen(S) != 0:
            shared = Ptr[S](task.task.shareds)[0]
            _task_loop_body_stub(gtid64, priv, shared)
        else:
            shared = ()
            _task_loop_body_stub(gtid64, priv, shared)
        return i32(0)

    @nonpure
    def _loop_loc_and_gtid(
        loc_ref: Ptr[Ident], reduction_loc_ref: Ptr[Ident], gtid: int
    ):
        pass

    @nonpure
    def _fix_privates_and_shareds(i, priv, shared):
        return priv, shared

    @nonpure
    def _taskred_setup(args):
        pass

    @nonpure
    def _taskred_finish():
        pass

    @nonpure
    def _loop_reductions(args):
        pass

    iterable, priv, shared = args[0]
    P = type(priv)
    S = type(shared)

    gtid = int(gtid_ptr[0])
    loc_ref = _default_loc()
    reduction_loc_ref = _reduction_loc()
    _loop_loc_and_gtid(loc_ref, reduction_loc_ref, gtid)

    _taskred_setup(shared)

    if _single_begin(loc_ref, gtid) != 0:
        _taskgroup_begin(loc_ref, gtid)
        try:
            for i in iterable:
                priv_fixed, shared_fixed = _fix_privates_and_shareds(i, priv, shared)
                _spawn_and_run_task(
                    loc_ref, gtid, _routine_stub(P=P, S=S, ...).__raw__(), priv_fixed, shared_fixed
                )
        finally:
            _taskgroup_end(loc_ref, gtid)
            _single_end(loc_ref, gtid)

    _taskred_finish()
    _loop_reductions(shared)
    _barrier(loc_ref, gtid)

@pure
def get_num_threads():
    from C import omp_get_num_threads() -> i32
    return int(omp_get_num_threads())

@pure
def get_thread_num():
    from C import omp_get_thread_num() -> i32
    return int(omp_get_thread_num())

@pure
def get_max_threads():
    from C import omp_get_max_threads() -> i32
    return int(omp_get_max_threads())

@pure
def get_num_procs():
    from C import omp_get_num_procs() -> i32
    return int(omp_get_num_procs())

def set_num_threads(num_threads: int):
    from C import omp_set_num_threads(i32)
    omp_set_num_threads(i32(num_threads))

@pure
def in_parallel():
    from C import omp_in_parallel() -> i32
    return bool(omp_in_parallel())

def set_dynamic(dynamic_threads: bool = True):
    from C import omp_set_dynamic(i32)
    omp_set_dynamic(i32(1 if dynamic_threads else 0))

@pure
def get_dynamic():
    from C import omp_get_dynamic() -> i32
    return bool(omp_get_dynamic())

@pure
def get_cancellation():
    from C import omp_get_cancellation() -> i32
    return bool(omp_get_cancellation())

def set_schedule(kind: str, chunk_size: int = 0):
    from C import omp_set_schedule(i32, i32)
    if kind == "static":
        omp_set_schedule(i32(1), i32(chunk_size))
    elif kind == "dynamic":
        omp_set_schedule(i32(2), i32(chunk_size))
    elif kind == "guided":
        omp_set_schedule(i32(3), i32(chunk_size))
    elif kind == "auto":
        if chunk_size != 0:
            raise ValueError("cannot specify chunk size for auto schedule")
        omp_set_schedule(i32(4), i32(chunk_size))
    else:
        raise ValueError(
            "invalid schedule kind; valid ones are: 'static', 'dynamic', 'guided', 'auto'"
        )

@pure
def get_schedule():
    from C import omp_get_schedule(Ptr[i32], Ptr[i32])
    kind_code = i32(0)
    chunk_size = i32(0)
    omp_get_schedule(__ptr__(kind_code), __ptr__(chunk_size))
    idx = int(kind_code)
    kind = (
        ("static", "dynamic", "guided", "auto")[idx - 1] if 1 < idx <= 4 else "unknown"
    )
    return kind, int(chunk_size)

@pure
def get_thread_limit():
    from C import omp_get_thread_limit() -> i32
    return int(omp_get_thread_limit())

def set_max_active_levels(max_levels: int):
    from C import omp_set_max_active_levels(i32)
    omp_set_max_active_levels(i32(max_levels))

@pure
def get_max_active_levels():
    from C import omp_get_max_active_levels() -> i32
    return int(omp_get_max_active_levels())

@pure
def get_level():
    from C import omp_get_level() -> i32
    return int(omp_get_level())

@pure
def get_ancestor_thread_num(level: int):
    from C import omp_get_ancestor_thread_num(i32) -> i32
    return int(omp_get_ancestor_thread_num(i32(level)))

@pure
def get_team_size(level: int):
    from C import omp_get_team_size(i32) -> i32
    return int(omp_get_team_size(i32(level)))

@pure
def get_active_level():
    from C import omp_get_active_level() -> i32
    return int(omp_get_active_level())

@pure
def in_final():
    from C import omp_in_final() -> i32
    return bool(omp_in_final())

@pure
def get_proc_bind():
    from C import omp_get_proc_bind() -> i32
    result = int(omp_get_proc_bind())
    if result < 0 or result > 4:
        return "unknown"
    return ("false", "true", "master", "close", "spread")[result]

def set_default_device(device_num: int):
    from C import omp_set_default_device(i32)
    omp_set_default_device(i32(device_num))

@pure
def get_default_device():
    from C import omp_get_default_device() -> i32
    return int(omp_get_default_device())

@pure
def get_num_devices():
    from C import omp_get_num_devices() -> i32
    return int(omp_get_num_devices())

@pure
def get_num_teams():
    from C import omp_get_num_teams() -> i32
    return int(omp_get_num_teams())

@pure
def get_team_num():
    from C import omp_get_team_num() -> i32
    return int(omp_get_team_num())

@pure
def is_initial_device():
    from C import omp_is_initial_device() -> i32
    return bool(omp_is_initial_device())

@pure
def get_wtime():
    from C import omp_get_wtime() -> float
    return omp_get_wtime()

@pure
def get_wtick():
    from C import omp_get_wtick() -> float
    return omp_get_wtick()

def single(func):
    def _wrapper(*args, **kwargs):
        gtid = get_thread_num()
        loc = _default_loc()
        if _single_begin(loc, gtid) != 0:
            try:
                func(*args, **kwargs)
            finally:
                _single_end(loc, gtid)

    return _wrapper

def master(func):
    def _wrapper(*args, **kwargs):
        gtid = get_thread_num()
        loc = _default_loc()
        if _master_begin(loc, gtid) != 0:
            try:
                func(*args, **kwargs)
            finally:
                _master_end(loc, gtid)

    return _wrapper

def ordered(func):
    def _wrapper(*args, **kwargs):
        gtid = get_thread_num()
        loc = _default_loc()
        _ordered_begin(loc, gtid)
        try:
            func(*args, **kwargs)
        finally:
            _ordered_end(loc, gtid)

    return _wrapper

_default_lock = Lock()

def critical(func):
    def _wrapper(*args, **kwargs):
        gtid = get_thread_num()
        loc = _default_loc()
        _critical_begin(loc, gtid, __ptr__(_default_lock))
        try:
            func(*args, **kwargs)
        finally:
            _critical_end(loc, gtid, __ptr__(_default_lock))

    return _wrapper

def _push_num_threads(num_threads: int):
    from C import __kmpc_push_num_threads(Ptr[Ident], i32, i32)
    gtid = get_thread_num()
    loc = _default_loc()
    __kmpc_push_num_threads(loc, i32(gtid), i32(num_threads))

@llvm
def _atomic_int_add(a: Ptr[int], b: int) -> None:
    %old = atomicrmw add ptr %a, i64 %b monotonic
    ret {} {}

def _atomic_int_mul(a: Ptr[int], b: int):
    from C import __kmpc_atomic_fixed8_mul(Ptr[Ident], i32, Ptr[int], int)
    __kmpc_atomic_fixed8_mul(_default_loc(), i32(0), a, b)

@llvm
def _atomic_int_and(a: Ptr[int], b: int) -> None:
    %old = atomicrmw and ptr %a, i64 %b monotonic
    ret {} {}

@llvm
def _atomic_int_or(a: Ptr[int], b: int) -> None:
    %old = atomicrmw or ptr %a, i64 %b monotonic
    ret {} {}

@llvm
def _atomic_int_xor(a: Ptr[int], b: int) -> None:
    %old = atomicrmw xor ptr %a, i64 %b monotonic
    ret {} {}

@llvm
def _atomic_int_min(a: Ptr[int], b: int) -> None:
    %old = atomicrmw min ptr %a, i64 %b monotonic
    ret {} {}

@llvm
def _atomic_int_max(a: Ptr[int], b: int) -> None:
    %old = atomicrmw max ptr %a, i64 %b monotonic
    ret {} {}

def _atomic_float_add(a: Ptr[float], b: float) -> None:
    from C import __kmpc_atomic_float8_add(Ptr[Ident], i32, Ptr[float], float)
    __kmpc_atomic_float8_add(_default_loc(), i32(0), a, b)

def _atomic_float_mul(a: Ptr[float], b: float):
    from C import __kmpc_atomic_float8_mul(Ptr[Ident], i32, Ptr[float], float)
    __kmpc_atomic_float8_mul(_default_loc(), i32(0), a, b)

def _atomic_float_min(a: Ptr[float], b: float):
    from C import __kmpc_atomic_float8_min(Ptr[Ident], i32, Ptr[float], float)
    __kmpc_atomic_float8_min(_default_loc(), i32(0), a, b)

def _atomic_float_max(a: Ptr[float], b: float):
    from C import __kmpc_atomic_float8_max(Ptr[Ident], i32, Ptr[float], float)
    __kmpc_atomic_float8_max(_default_loc(), i32(0), a, b)

def _atomic_float32_add(a: Ptr[float32], b: float32) -> None:
    from C import __kmpc_atomic_float4_add(Ptr[Ident], i32, Ptr[float32], float32)
    __kmpc_atomic_float4_add(_default_loc(), i32(0), a, b)

def _atomic_float32_mul(a: Ptr[float32], b: float32):
    from C import __kmpc_atomic_float4_mul(Ptr[Ident], i32, Ptr[float32], float32)
    __kmpc_atomic_float4_mul(_default_loc(), i32(0), a, b)

def _atomic_float32_min(a: Ptr[float32], b: float32) -> None:
    from C import __kmpc_atomic_float4_min(Ptr[Ident], i32, Ptr[float32], float32)
    __kmpc_atomic_float4_min(_default_loc(), i32(0), a, b)

def _atomic_float32_max(a: Ptr[float32], b: float32) -> None:
    from C import __kmpc_atomic_float4_max(Ptr[Ident], i32, Ptr[float32], float32)
    __kmpc_atomic_float4_max(_default_loc(), i32(0), a, b)

def _range_len(start: int, stop: int, step: int):
    if step > 0 and start < stop:
        return 1 + (stop - 1 - start) // step
    elif step < 0 and start > stop:
        return 1 + (start - 1 - stop) // (-step)
    else:
        return 0

def for_par(
    num_threads: int = -1,
    chunk_size: int = -1,
    schedule: Static[str] = "static",
    ordered: Static[int] = False,
    collapse: Static[int] = 0,
    gpu: Static[int] = False,
):
    pass