made c_code_cache_version work with GpuCAReduce

theano/sandbox/cuda/basic_ops.py

@@ -1501,9 +1501,16 @@ class GpuCAReduce(GpuOp):
         """ % locals()
 
     def c_code_cache_version(self):
-        # TODO: make this act like tensor.elemwise.CAReduce
-        # needs to include scalar op's cache version in the returned tuple
-        return ()
+        op_version = self.scalar_op.cuda_assign_reduce_code_cache_version()
+        if op_version:
+            # our version is 0
+            rval = [0]
+            rval.extend(op_version)
+            rval = tuple(rval)
+            return rval
+        else:
+            # we can't support caching if the op doesn't
+            return ()
 
     def _op_guard(self):
         """ Raises NotImplementedError if op is not Add """

theano/scalar/basic.py

@@ -823,6 +823,12 @@ class ScalarOp(Op):
         raise NotImplementedError(
                 str(self)+" does not implement cuda_assign_reduce")
 
+    def cuda_assign_reduce_code_cache_version(self):
+        """ Returns a tuple of integers representing the version of the
+        scalar op's cuda_assign_reduce method. Must be unique across ops.
+        An empty tuple means not to use code caching for this op"""
+        return ()
+
     def c_code_cache_version(self):
         return (4,)
 
@@ -1169,6 +1175,10 @@ class Maximum(BinaryScalarOp):
 
     def cuda_assign_reduce(self, left, right):
         return left + ' = max(' + left + ', '+ right + ');'
+    def cuda_assign_reduce_code_cache_version(self):
+        # Maximum's unique identifier is 0
+        # Version 0 of its code
+        return (0,0)
 maximum = Maximum(upcast_out, name='maximum')
 
 
@@ -1235,6 +1245,11 @@ class Add(ScalarOp):
     def cuda_assign_reduce(self, left, right):
         return left + ' += ' + right + ';'
 
+    def cuda_assign_reduce_code_cache_version(self):
+        # Add's unique identifier is 1
+        # Version 0 of its code
+        return (1,0)
+
 add = Add(upcast_out, name='add')