Added test implementation of parallel-vectorize (incomplete).

Various fix.
This commit is contained in:
Siu Kwan Lam 2012-08-05 21:35:02 -07:00
commit 891daa1164
5 changed files with 555 additions and 52 deletions

View file

@ -40,6 +40,10 @@ def _change_block_temporarily(builder, bb):
yield
builder.position_at_end(origbb)
@contextlib.contextmanager
def _change_block_temporarily_dummy(*args):
yield
class _IfElse(object):
def __init__(self, parent, cond):
self.parent = parent
@ -58,14 +62,17 @@ class _IfElse(object):
yield
self._to_close.extend([self._bbif, self._bbelse])
self._to_close.extend([self._bbif, self._bbelse, builder.basic_block])
@contextlib.contextmanager
def otherwise(self):
self.parent.builder.position_at_end(self._bbelse)
builder = self.parent.builder
builder.position_at_end(self._bbelse)
yield
self._to_close.append(builder.basic_block)
def close(self):
self._to_close.append(self.parent.builder.basic_block)
bbend = self.parent.function.append_basic_block('if.end')
builder = self.parent.builder
for bb in self._to_close:
@ -111,7 +118,11 @@ class _Loop(object):
self.parent.builder.branch(self._bbcond)
def close(self):
self.parent.builder.position_at_end(self._bbend)
builder = self.parent.builder
# if not _is_block_terminated(builder.basic_block):
# with _change_block_temporarily(builder, builder.basic_block):
# builder.branch(self._bbend)
builder.position_at_end(self._bbend)
class CBuilder(object):
'''
@ -148,7 +159,10 @@ class CBuilder(object):
def debug(self, *args):
type_mapper = {
'i8' : '%c',
'i16': '%hd',
'i32': '%d',
'i64': '%ld',
'double': '%e',
}
itemsfmt = []
@ -185,9 +199,23 @@ class CBuilder(object):
else:
return CVar(self, ptr)
def var_copy(self, val, name=''):
return self.var(val.type, val, name=name)
def array(self, ty, count, name=''):
with _change_block_temporarily(self.builder, self.declare_block):
if not isinstance(count, lc.Value):
if isinstance(count, int) or isinstance(count, lc.Constant):
contexthelper = _change_block_temporarily
else:
contexthelper = _change_block_temporarily_dummy
with contexthelper(self.builder, self.declare_block):
is_cstruct = _is_cstruct(ty)
if is_cstruct:
cstruct = ty
ty = ty.llvm_type()
if isinstance(count, CValue):
count = count.value
elif not isinstance(count, lc.Value):
count = self.constant(lc.Type.int(), count).value
ptr = self.builder.alloca_array(ty, count, name=name)
return CArray(self, ptr)
@ -341,6 +369,40 @@ class CBuilder(object):
def alignment(self, ty):
return self.target_data.abi_alignment(ty)
def unreachable(self):
self.builder.unreachable()
class CDefinition(CBuilder):
'''
Inherit this class to for defining functions
'''
_name_ = 'unamed' # name of the function; should overide in subclass
_retty_ = lc.Type.void() # return type; can overide in subclass
_argtys_ = [] # a list of tuple(name, type); can overide in subclass
@classmethod
def define(cls, module, **kws):
functype = lc.Type.function(cls._retty_, [v for k, v in cls._argtys_])
func = module.get_or_insert_function(functype, name=cls._name_)
if not func.is_declaration: # already defined?
raise NameError('Function %s has already been defined' % cls.name)
# Name all arguments
for i, (name, _) in enumerate(cls._argtys_):
func.args[i].name = name
# Create builder and populate body
cbuilder = cls(func)
cbuilder.body(*cbuilder.args, **kws)
cbuilder.close()
return func
def body(self):
'''
Overide this function to define the body.
'''
raise NotImplementedError
class CValue(object):
'''
= Signess =
@ -475,13 +537,13 @@ class CValue(object):
return make(builder.bitcast(self.value, ty))
elif self.is_int:
if _is_int(ty):
if self.type.width > ty.width:
if self.type.width < ty.width:
if not unsigned:
return make(self.parent.builder.sext(self.value, ty))
else:
return make(self.parent.builder.zext(self.value, ty))
else:
return make(self.parent.trunc(self.value, ty))
return make(self.parent.builder.trunc(self.value, ty))
elif _is_real(ty):
if not unsigned:
return make(self.parent.builder.sitofp(self.value, ty))
@ -543,7 +605,46 @@ class CValue(object):
def _ensure_is_pointer(self):
if not self.is_pointer:
raise TypeError("Must be a pointer")
raise TypeError("Must be a pointer; got %s" % self.type)
def __getitem__(self, idx):
self._ensure_is_pointer()
if not isinstance(idx, CValue):
idx = self.parent.constant(lc.Type.int(), idx)
bldr = self.parent.builder
ptr = bldr.gep(self.value, [idx.value])
return CVar(self.parent, ptr)
def load(self, volatile=False):
self._ensure_is_pointer()
loaded = self.parent.builder.load(self.value, volatile=volatile)
return CTemp(self.parent, loaded)
def store(self, val, volatile=False):
self._ensure_is_pointer()
self.parent.builder.store(val.value, self.value, volatile=volatile)
def atomic_load(self, ordering, align=None, crossthread=True):
self._ensure_is_pointer()
if align is None:
align = self.parent.alignment(self.type.pointee)
inst = self.parent.builder.atomic_load(self.value, ordering, align,
crossthread=crossthread)
return CTemp(self.parent, inst)
def atomic_store(self, value, ordering, align=None, crossthread=True):
self._ensure_is_pointer()
if align is None:
align = self.parent.alignment(self.type.pointee)
self.parent.builder.atomic_store(value.ptr, self.value, ordering,
align=align, crossthread=crossthread)
def atomic_cmpxchg(self, old, new, ordering, crossthread=True):
self._ensure_is_pointer()
inst = self.parent.builder.atomic_cmpxchg(self.value, old.value,
new.value, ordering,
crossthread=crossthread)
return CTemp(self.parent, inst)
class CFunc(CValue):
@ -552,8 +653,14 @@ class CFunc(CValue):
self.function = func
def __call__(self, *args):
arg_value = _list_values(args)
res = self.parent.builder.call(self.function, arg_value)
arg_values = _list_values(args)
ftype = self.function.type.pointee
for i, (exp, got) in enumerate(zip(ftype.args, arg_values)):
if exp != got.type:
raise TypeError("At call to %s, "
"argument %d mismatch: %s != %s"
% (self.function.name, i, exp, got.type))
res = self.parent.builder.call(self.function, arg_values)
return CTemp(self.parent, res)
@property
@ -601,7 +708,6 @@ class CVar(CValue):
def __imod__(self, rhs):
return self._inplace_binop('mod')(rhs)
def _inplace_bitwise(self, op):
def wrapped(rhs):
res = self._use_bitwise(op)(rhs)
@ -629,46 +735,23 @@ class CVar(CValue):
return self.parent.builder.load(self.ptr)
def assign(self, val):
self._ensure_same_type(val)
self.parent.builder.store(val.value, self.ptr)
@property
def type(self):
return self.ptr.type.pointee
def load(self, volatile=False):
self._ensure_is_pointer()
loaded = self.parent.builder.load(self.value, volatile=volatile)
return CTemp(self.parent, loaded)
def store(self, val, volatile=False):
self._ensure_is_pointer()
self.parent.builder.store(val.value, self.value, volatile=volatile)
def atomic_load(self, ordering, align=None, crossthread=True):
self._ensure_is_pointer()
if align is None:
align = self.parent.alignment(self.type.pointee)
inst = self.parent.builder.atomic_load(self.value, ordering, align,
crossthread=crossthread)
return CTemp(self.parent, inst)
def atomic_store(self, value, ordering, align=None, crossthread=True):
self._ensure_is_pointer()
if align is None:
align = self.parent.alignment(self.type.pointee)
self.parent.builder.atomic_store(value.value, self.value, ordering,
align=align, crossthread=crossthread)
def atomic_cmpxchg(self, old, new, ordering, crossthread=True):
self._ensure_is_pointer()
inst = self.parent.builder.atomic_cmpxchg(self.value, old.value,
new.value, ordering,
crossthread=crossthread)
return CTemp(self.parent, inst)
def reference(self):
return CTemp(self.parent, self.ptr)
def as_struct(self, cstruct_class, volatile=False):
if _is_pointer(self.type):
ptr = self.parent.builder.load(self.ptr, volatile=volatile)
return cstruct_class(self.parent, ptr)
else:
return cstruct_class(self.parent, self.ptr)
class CArray(CValue):
def __init__(self, parent, base):
super(CArray, self).__init__(parent)
@ -678,19 +761,23 @@ class CArray(CValue):
def value(self):
return self.base_ptr
def reference(self):
return CTemp(self.parent, self.base_ptr)
@property
def type(self):
return self.base_ptr.type
def __getitem__(self, idx):
self._ensure_is_pointer()
builder = self.parent.builder
if isinstance(idx, CValue):
idx = idx.value
elif not isinstance(idx, lc.Value):
idx = self.parent.constant(lc.Type.int(), idx).value
ptr = builder.gep(self.value, [idx])
return CVar(self.parent, ptr)
## Moved to CValue
# def __getitem__(self, idx):
# self._ensure_is_pointer()
# builder = self.parent.builder
# if isinstance(idx, CValue):
# idx = idx.value
# elif not isinstance(idx, lc.Value):
# idx = self.parent.constant(lc.Type.int(), idx).value
# ptr = builder.gep(self.value, [idx])
# return CVar(self.parent, ptr)
class CStruct(CValue):
@classmethod
@ -700,8 +787,32 @@ class CStruct(CValue):
def __init__(self, parent, ptr):
super(CStruct, self).__init__(parent)
makeind = lambda x: self.parent.constant(lc.Type.int(), x).value
self.ptr = ptr
for i, (fd, _) in enumerate(self._fields_):
gep = self.parent.builder.gep(ptr, [makeind(0), makeind(i)])
gep.name = "%s.%s" % (type(self).__name__, fd)
setattr(self, fd, CVar(self.parent, gep))
def reference(self):
return CTemp(self.parent, self.ptr)
class CExternal(object):
def __init__(self, cbuilder):
is_func = lambda x: isinstance(x, lc.FunctionType)
non_magic = lambda s: not ( s.startswith('__') and s.endswith('__') )
to_declare = []
for fname in filter(non_magic, vars(type(self))):
ftype = getattr(self, fname)
if is_func(ftype):
to_declare.append((fname, ftype))
mod = cbuilder.function.module
for fname, ftype in to_declare:
func = mod.get_or_insert_function(ftype, name=fname)
if func.type.pointee != ftype:
raise NameError("Function has already been declared "
"with a different type: %s != %s"
% (func.type, ftype) )
setattr(self, fname, CFunc(cbuilder, func))

View file

@ -31,7 +31,7 @@ MAP_CTYPES = {
class CExecutor(object):
def __init__(self, mod_or_engine):
if isinstance(mod_or_engine, Module):
self.engine = le.EngineBuilder.new(mod_or_engine).create()
self.engine = le.EngineBuilder.new(mod_or_engine).opt(3).create()
else:
self.engine = mod_or_engine

View file

@ -16,6 +16,7 @@ double = Type.double()
pointer = Type.pointer
void_p = pointer(char)
char_p = pointer(char)
# platform dependent

318
parallel_vectorize.py Normal file
View file

@ -0,0 +1,318 @@
from llvm.core import *
from llvm.passes import *
from llvm_cbuilder import *
import llvm_cbuilder.shortnames as C
class WorkQueue(CStruct):
_fields_ = [
('next', C.intp),
('last', C.intp),
('lock', C.int),
]
def Lock(self):
with self.parent.loop() as loop:
with loop.condition() as setcond:
unlocked = self.parent.constant(self.lock.type, 0)
locked = self.parent.constant(self.lock.type, 1)
res = self.lock.reference().atomic_cmpxchg(unlocked, locked,
ordering='acquire')
setcond( res != unlocked )
with loop.body():
pass
def Unlock(self):
unlocked = self.parent.constant(self.lock.type, 0)
locked = self.parent.constant(self.lock.type, 1)
res = self.lock.reference().atomic_cmpxchg(locked, unlocked,
ordering='release')
with self.parent.ifelse( res != locked ) as ifelse:
with ifelse.then():
# This shall kill the program
self.parent.unreachable()
class ContextCommon(CStruct):
_fields_ = [
# loop ufunc args
('args', C.pointer(C.char_p)),
('dimensions', C.pointer(C.intp)),
('steps', C.pointer(C.intp)),
('data', C.void_p),
# specifics for work queues
('func', C.void_p),
('num_thread', C.int),
('workqueues', C.pointer(WorkQueue.llvm_type())),
]
class Context(CStruct):
_fields_ = [
('common', C.pointer(ContextCommon.llvm_type())),
('id', C.int),
('completed', C.intp),
]
class ParallelUFunc(CDefinition):
_name_ = 'parallel_ufunc'
_retty_ = C.void
_argtys_ = [
('func', C.void_p),
('worker', C.void_p),
('args', C.pointer(C.char_p)),
('dimensions', C.pointer(C.intp)),
('steps', C.pointer(C.intp)),
('data', C.void_p),
]
def body(self, func, worker, args, dimensions, steps, data, ThreadCount=1):
common = self.var(ContextCommon, name='common')
workqueues = self.array(WorkQueue, ThreadCount, name='workqueues')
contexts = self.array(Context, ThreadCount, name='contexts')
num_thread = self.var(C.int, ThreadCount, name='num_thread')
common.args.assign(args)
common.dimensions.assign(dimensions)
common.steps.assign(steps)
common.data.assign(data)
common.func.assign(func)
common.num_thread.assign(num_thread.cast(C.int))
common.workqueues.assign(workqueues.reference())
N = dimensions[0]
ChunkSize = self.var_copy(N / num_thread.cast(N.type))
ChunkSize_NULL = self.constant_null(ChunkSize.type)
with self.ifelse(ChunkSize == ChunkSize_NULL) as ifelse:
with ifelse.then():
ChunkSize.assign(self.constant(ChunkSize.type, 1))
num_thread.assign(N.cast(num_thread.type))
self._populate_workqueues(workqueues, N, ChunkSize, num_thread)
self._populate_context(contexts, common, num_thread)
self._dispatch_worker(worker, contexts, num_thread)
# Check for race condition
total_completed = self.var(C.intp, 0, name='total_completed')
for t in range(ThreadCount):
cur_ctxt = contexts[t].as_struct(Context)
total_completed += cur_ctxt.completed
self.debug(cur_ctxt.id, 'completed', cur_ctxt.completed)
with self.ifelse( total_completed == N ) as ifelse:
with ifelse.then():
self.debug("All is well!")
with ifelse.otherwise():
self.debug("ERROR: race occurred!")
self.ret()
def _populate_workqueues(self, workqueues, N, ChunkSize, num_thread):
i = self.var(num_thread.type, 0, name='i')
ONE = self.constant(i.type, 1)
with self.loop() as loop:
with loop.condition() as setcond:
setcond( i < num_thread )
with loop.body():
cur_wq = workqueues[i].as_struct(WorkQueue)
cur_wq.next.assign(i.cast(ChunkSize.type) * ChunkSize)
cur_wq.last.assign((i + ONE).cast(ChunkSize.type) * ChunkSize)
cur_wq.lock.assign(self.constant(C.int, 0))
# increment
i += ONE
# end loop
last_wq = workqueues[num_thread - ONE].as_struct(WorkQueue)
last_wq.last.assign(N)
def _populate_context(self, contexts, common, num_thread):
i = self.var(num_thread.type, 0, name='i')
ONE = self.constant(i.type, 1)
with self.loop() as loop:
with loop.condition() as setcond:
setcond( i < num_thread )
with loop.body():
cur_ctxt = contexts[i].as_struct(Context)
cur_ctxt.common.assign(common.reference())
cur_ctxt.id.assign(i)
cur_ctxt.completed.assign(self.constant_null(
cur_ctxt.completed.type))
# increment
i += ONE
class ParallelUFuncPosix(ParallelUFunc):
def _dispatch_worker(self, worker, contexts, num_thread):
api = PThreadAPI(self)
NULL = self.constant_null(C.void_p)
threads = self.array(api.pthread_t, num_thread, name='threads')
# self.debug("launch threads")
# TODO error handling
i = self.var(num_thread.type, 0, name='i')
ONE = self.constant(i.type, 1)
with self.loop() as loop:
with loop.condition() as setcond:
setcond( i < num_thread )
with loop.body():
api.pthread_create(threads[i].reference(), NULL, worker,
contexts[i].reference().cast(C.void_p))
i += ONE
# self.debug("join threads")
i.assign(self.constant_null(i.type))
with self.loop() as loop:
with loop.condition() as setcond:
setcond( i < num_thread )
with loop.body():
api.pthread_join(threads[i], NULL)
i += ONE
# self.debug("closing")
class UFuncCore(CDefinition):
_name_ = 'ufunc_worker'
_argtys_ = [
('context', C.pointer(Context.llvm_type())),
]
def body(self, context):
context = context.as_struct(Context)
common = context.common.as_struct(ContextCommon)
tid = context.id
# self.debug("start thread", tid, "/", common.num_thread)
workqueue = common.workqueues[tid].as_struct(WorkQueue)
self._do_workqueue(common, workqueue, tid, context.completed)
self._do_work_stealing(common, tid, context.completed) # optional
self.ret()
def _do_workqueue(self, common, workqueue, tid, completed):
'''
Process local workqueue.
'''
ZERO = self.constant_null(C.int)
with self.loop() as loop:
with loop.condition() as setcond:
setcond(ZERO == ZERO) # loop forever
with loop.body():
workqueue.Lock()
# Critical section
item = self.var_copy(workqueue.next, name='item')
workqueue.next += self.constant(item.type, 1)
last = self.var_copy(workqueue.last, name='last')
# Release
workqueue.Unlock()
with self.ifelse( item >= last ) as ifelse:
with ifelse.then():
loop.break_loop()
self._do_work(common, item, tid, completed)
def _do_work_stealing(self, common, tid, completed):
'''
Steal work from other workqueues
'''
# self.debug("start work stealing", tid)
incomplete_thread_ct = self.var(C.int, 1)
ITC_NULL = self.constant_null(incomplete_thread_ct.type)
with self.loop() as loop:
with loop.condition() as setcond:
setcond( incomplete_thread_ct != ITC_NULL )
with loop.body():
incomplete_thread_ct.assign(ITC_NULL)
self._do_work_stealing_innerloop(common, incomplete_thread_ct,
tid, completed)
def _do_work_stealing_innerloop(self, common, incomplete_thread_ct, tid,
completed):
i = self.var(C.int, 0)
with self.loop() as loop:
with loop.condition() as setcond:
setcond( i < common.num_thread )
with loop.body():
with self.ifelse( i != tid ) as ifelse:
with ifelse.then():
otherqueue = common.workqueues[i].as_struct(WorkQueue)
self._do_work_stealing_check(common, otherqueue,
incomplete_thread_ct, tid,
completed)
# increment
i += self.constant(i.type, 1)
def _do_work_stealing_check(self, common, otherqueue,
incomplete_thread_ct, tid, completed):
otherqueue.Lock()
ONE = self.constant(otherqueue.last.type, 1)
ITC_ONE = self.constant(incomplete_thread_ct.type, 1)
with self.ifelse(otherqueue.next < otherqueue.last) as ifelse:
with ifelse.then():
otherqueue.last -= ONE
item = self.var_copy(otherqueue.last)
otherqueue.Unlock()
# Released
self._do_work(common, item, tid, completed)
# Mark incomplete thread
incomplete_thread_ct.assign(ITC_ONE)
with ifelse.otherwise():
otherqueue.Unlock()
def _do_work(self, common, item, tid, completed):
# self.debug(" tid", tid, "item", item)
completed += self.constant(completed.type, 1)
class PThreadAPI(CExternal):
pthread_t = C.void_p
pthread_create = Type.function(C.int,
[C.pointer(pthread_t), # thread_t
C.void_p, # thread attr
C.void_p, # function
C.void_p]) # arg
pthread_join = Type.function(C.int, [C.void_p, C.void_p])
def _main():
NUM_THREAD = 2
module = Module.new(__name__)
fpm = FunctionPassManager.new(module)
PassManagerBuilder.new().populate(fpm)
f1 = ParallelUFuncPosix.define(module, ThreadCount=NUM_THREAD)
f2 = UFuncCore.define(module)
#print(module)
# print(f1)
print(f2)
module.verify()
fpm.run(f1)
fpm.run(f2)
print('optimized'.center(80,'-'))
print(f2)
if __name__ == '__main__':
_main()

View file

@ -0,0 +1,73 @@
from parallel_vectorize import *
'''
void parallel_ufunc(void * func, void * worker,
char **args, npy_intp *dimensions,
npy_intp *steps, void *data)
'''
class Tester(CDefinition):
_name_ = 'tester'
def body(self):
# depends
module = self.function.module
ThreadCount = 2
ArgCount = 2
WorkCount = 1000000
parallel_ufunc = CFunc(self, ParallelUFuncPosix.define(module,
ThreadCount=ThreadCount))
worker = CFunc(self, UFuncCore.define(module))
# real work
NULL = self.constant_null(C.void_p)
args = self.array(C.char_p, 2, name='args')
for t in range(ThreadCount):
args_for_thread = self.array(C.double, 4)
args[t].assign(args_for_thread.cast(C.char_p))
dims = self.array(C.intp, 1, name='dims')
dims[0].assign(self.constant(C.intp, WorkCount))
steps = self.array(C.intp, ArgCount, name='steps')
for c in range(ArgCount):
steps[c].assign(self.constant(C.intp, 8))
parallel_ufunc(NULL, worker.cast(C.void_p), args, dims, steps, NULL)
self.ret()
def main():
NUM_THREAD = 2
module = Module.new(__name__)
mpm = PassManager.new()
pmbuilder = PassManagerBuilder.new()
pmbuilder.opt_level = 3
pmbuilder.populate(mpm)
fntester = Tester.define(module)
# print(module)
module.verify()
mpm.run(module)
print('optimized'.center(80,'-'))
print(module)
# run
exe = CExecutor(module)
exe.engine.get_pointer_to_function(fntester)
func = exe.get_ctype_function(fntester, 'void')
func()
if __name__ == '__main__':
main()