Add support to gpu reduce for acc and output dtype.

theano/sandbox/gpuarray/opt.py

@@ -311,9 +311,10 @@ def local_gpua_careduce(node):
     if isinstance(node.op.scalar_op, (scalar.Add, scalar.Mul,
                                       scalar.Maximum, scalar.Minimum)):
         x, = node.inputs
-        greduce = GpuCAReduceCuda(node.op.scalar_op, axis=node.op.axis)
-        if x.dtype != "float32":
-            return
+        greduce = GpuCAReduceCuda(
+            node.op.scalar_op, axis=node.op.axis,
+            dtype=getattr(node.op, 'dtype', None),
+            acc_dtype=getattr(node.op, 'acc_dtype', None))
         gvar = greduce(x)
         #We need to have the make node called, otherwise the mask can
         #be None

theano/sandbox/gpuarray/tests/test_elemwise.py

@@ -68,9 +68,10 @@ class test_GpuCAReduceCPY(test_CAReduce):
 
 
 class test_GpuCAReduceCuda(test_GpuCAReduceCPY):
-    dtypes = ["float32"]
+    dtypes = ["float32", "int64"]
+    dtypes = []
     bin_dtypes = ["uint8", "int8"]
-    bin_dtypes = []
+
     cases = [((5, 6), None),
              ((5, 6), (0, 1)),
              ((5, 6), (0, )),