jyapayne/llvmpy
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Add specialize() for cleaner specialization of CDefinition.

Browse source
This commit is contained in:
Siu Kwan Lam 2012-08-06 20:25:36 -07:00
commit 01054c726f
3 changed files with 39 additions and 21 deletions
Download patch file Download diff file Expand all files Collapse all files

38
parallel_vectorize.py
View file

@ -95,7 +95,7 @@ class ParallelUFunc(CDefinition):
which should be implemented in subclass or mixin. which should be implemented in subclass or mixin.
''' '''
_name_ = 'parallel_ufunc_%(ThreadCount)d' #_name_ = 'parallel_ufunc_%(ThreadCount)d'
_argtys_ = [ _argtys_ = [
('func', C.void_p), ('func', C.void_p),
('worker', C.void_p), ('worker', C.void_p),
@ -105,8 +105,18 @@ class ParallelUFunc(CDefinition):
('data', C.void_p), ('data', C.void_p),
] ]
def body(self, func, worker, args, dimensions, steps, data, ThreadCount=1): @classmethod
def specialize(cls, num_thread):
name = 'parallel_ufunc_%d' % num_thread
newcls = type(name, (cls,), {
'_name_' : name,
'ThreadCount': num_thread,
})
return newcls
def body(self, func, worker, args, dimensions, steps, data):
# Setup variables # Setup variables
ThreadCount = self.ThreadCount
common = self.var(ContextCommon, name='common') common = self.var(ContextCommon, name='common')
workqueues = self.array(WorkQueue, ThreadCount, name='workqueues') workqueues = self.array(WorkQueue, ThreadCount, name='workqueues')
contexts = self.array(Context, ThreadCount, name='contexts') contexts = self.array(Context, ThreadCount, name='contexts')
@ -318,9 +328,7 @@ class UFuncCore(CDefinition):
''' '''
raise NotImplementedError raise NotImplementedError
class SpecializedParallelUFunc(CDefinition): class SpecializedParallelUFunc(CDefinition):
_name_ = 'specialized_parallel_ufunc_%(ThreadCount)d_%(FuncName)s'
_argtys_ = [ _argtys_ = [
('args', C.pointer(C.char_p)), ('args', C.pointer(C.char_p)),
('dimensions', C.pointer(C.intp)), ('dimensions', C.pointer(C.intp)),
@ -328,15 +336,27 @@ class SpecializedParallelUFunc(CDefinition):
('data', C.void_p), ('data', C.void_p),
] ]
def body(self, args, dimensions, steps, data, def body(self, args, dimensions, steps, data,):
PUFuncDef, CoreDef, Func, FuncName, ThreadCount=1): pufunc = self.depends(self.PUFuncDef)
pufunc = self.depends(PUFuncDef, ThreadCount=ThreadCount) core = self.depends(self.CoreDef)
core = self.depends(CoreDef) func = self.depends(self.FuncDef)
func = self.depends(Func)
to_void_p = lambda x: x.cast(C.void_p) to_void_p = lambda x: x.cast(C.void_p)
pufunc(to_void_p(func), to_void_p(core), args, dimensions, steps, data) pufunc(to_void_p(func), to_void_p(core), args, dimensions, steps, data)
self.ret() self.ret()
@classmethod
def specialize(cls, pufunc_def, core_def, func_def):
name = 'specialized_%s_%s_%s'% (pufunc_def._name_,
core_def._name_,
func_def._name_)
newcls = type(name, (cls,), {
'_name_' : name,
'PUFuncDef': pufunc_def,
'CoreDef' : core_def,
'FuncDef' : func_def,
})
return newcls
class PThreadAPI(CExternal): class PThreadAPI(CExternal):
pthread_t = C.void_p pthread_t = C.void_p

11
test_parallel_vectorize.py
View file

@ -14,6 +14,7 @@ class UFuncCore_D_D(UFuncCore):
''' '''
Specialize UFuncCore for double input, double output. Specialize UFuncCore for double input, double output.
''' '''
_name_ = UFuncCore._name_ + '_d_d'
def _do_work(self, common, item, tid): def _do_work(self, common, item, tid):
ufunc_type = Type.function(C.double, [C.double]) ufunc_type = Type.function(C.double, [C.double])
ufunc_ptr = CFunc(self, common.func.cast(C.pointer(ufunc_type)).value) ufunc_ptr = CFunc(self, common.func.cast(C.pointer(ufunc_type)).value)
@ -47,12 +48,10 @@ class Tester(CDefinition):
ArgCount = 2 ArgCount = 2
WorkCount = 10000 WorkCount = 10000
sppufunc = self.depends(SpecializedParallelUFunc, PUfuncDef = ParallelUFuncPosix.specialize(num_thread=2)
PUFuncDef = ParallelUFuncPosix, SPUF = SpecializedParallelUFunc.specialize(
CoreDef = UFuncCore_D_D, PUfuncDef, UFuncCore_D_D, Work_D_D)
Func = Work_D_D, sppufunc = self.depends(SPUF)
FuncName = Work_D_D._name_,
ThreadCount = 2)
# real work # real work
NULL = self.constant_null(C.void_p) NULL = self.constant_null(C.void_p)

11
test_parallel_vectorize_numpy.py
View file

@ -4,12 +4,11 @@ import numpy as np
def main(): def main():
module = Module.new(__name__) module = Module.new(__name__)
sppufunc = SpecializedParallelUFunc.define(module, PUfuncDef = ParallelUFuncPosix.specialize(num_thread=2)
PUFuncDef = ParallelUFuncPosix, SPUF = SpecializedParallelUFunc.specialize(
CoreDef = UFuncCore_D_D, PUfuncDef, UFuncCore_D_D, Work_D_D)
Func = Work_D_D, sppufunc = SPUF.define(module)
FuncName = Work_D_D._name_,
ThreadCount = 2)
module.verify() module.verify()
mpm = PassManager.new() mpm = PassManager.new()