diff --git a/parallel_vectorize.py b/parallel_vectorize.py
index bef04eb..b5a03f5 100644
--- a/parallel_vectorize.py
+++ b/parallel_vectorize.py
@@ -95,7 +95,7 @@ class ParallelUFunc(CDefinition):
 
     which should be implemented in subclass or mixin.
     '''
-    _name_   = 'parallel_ufunc_%(ThreadCount)d'
+    #_name_   = 'parallel_ufunc_%(ThreadCount)d'
     _argtys_ = [
         ('func',       C.void_p),
         ('worker',     C.void_p),
@@ -105,8 +105,18 @@ class ParallelUFunc(CDefinition):
         ('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
+        ThreadCount = self.ThreadCount
         common = self.var(ContextCommon, name='common')
         workqueues = self.array(WorkQueue, ThreadCount, name='workqueues')
         contexts = self.array(Context, ThreadCount, name='contexts')
@@ -318,9 +328,7 @@ class UFuncCore(CDefinition):
         '''
         raise NotImplementedError
 
-
 class SpecializedParallelUFunc(CDefinition):
-    _name_ = 'specialized_parallel_ufunc_%(ThreadCount)d_%(FuncName)s'
     _argtys_ = [
         ('args',       C.pointer(C.char_p)),
         ('dimensions', C.pointer(C.intp)),
@@ -328,15 +336,27 @@ class SpecializedParallelUFunc(CDefinition):
         ('data',       C.void_p),
     ]
 
-    def body(self, args, dimensions, steps, data,
-             PUFuncDef, CoreDef, Func, FuncName, ThreadCount=1):
-        pufunc = self.depends(PUFuncDef, ThreadCount=ThreadCount)
-        core = self.depends(CoreDef)
-        func = self.depends(Func)
+    def body(self, args, dimensions, steps, data,):
+        pufunc = self.depends(self.PUFuncDef)
+        core = self.depends(self.CoreDef)
+        func = self.depends(self.FuncDef)
         to_void_p = lambda x: x.cast(C.void_p)
         pufunc(to_void_p(func), to_void_p(core), args, dimensions, steps, data)
         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):
     pthread_t = C.void_p
 
diff --git a/test_parallel_vectorize.py b/test_parallel_vectorize.py
index ed2010d..ab28363 100644
--- a/test_parallel_vectorize.py
+++ b/test_parallel_vectorize.py
@@ -14,6 +14,7 @@ class UFuncCore_D_D(UFuncCore):
     '''
     Specialize UFuncCore for double input, double output.
     '''
+    _name_ = UFuncCore._name_ + '_d_d'
     def _do_work(self, common, item, tid):
         ufunc_type = Type.function(C.double, [C.double])
         ufunc_ptr = CFunc(self, common.func.cast(C.pointer(ufunc_type)).value)
@@ -47,12 +48,10 @@ class Tester(CDefinition):
         ArgCount = 2
         WorkCount = 10000
 
-        sppufunc = self.depends(SpecializedParallelUFunc,
-                                PUFuncDef = ParallelUFuncPosix,
-                                CoreDef = UFuncCore_D_D,
-                                Func = Work_D_D,
-                                FuncName = Work_D_D._name_,
-                                ThreadCount = 2)
+        PUfuncDef = ParallelUFuncPosix.specialize(num_thread=2)
+        SPUF = SpecializedParallelUFunc.specialize(
+                                            PUfuncDef, UFuncCore_D_D, Work_D_D)
+        sppufunc = self.depends(SPUF)
 
         # real work
         NULL = self.constant_null(C.void_p)
diff --git a/test_parallel_vectorize_numpy.py b/test_parallel_vectorize_numpy.py
index dc3557f..7b1ca51 100644
--- a/test_parallel_vectorize_numpy.py
+++ b/test_parallel_vectorize_numpy.py
@@ -4,12 +4,11 @@ import numpy as np
 
 def main():
     module = Module.new(__name__)
-    sppufunc = SpecializedParallelUFunc.define(module,
-                            PUFuncDef = ParallelUFuncPosix,
-                            CoreDef = UFuncCore_D_D,
-                            Func = Work_D_D,
-                            FuncName = Work_D_D._name_,
-                            ThreadCount = 2)
+    PUfuncDef = ParallelUFuncPosix.specialize(num_thread=2)
+    SPUF = SpecializedParallelUFunc.specialize(
+                                        PUfuncDef, UFuncCore_D_D, Work_D_D)
+    sppufunc = SPUF.define(module)
+
     module.verify()
 
     mpm = PassManager.new()