Fix a test where the gpuarray back-end didn't downcasted python float to float32.

theano/sandbox/gpuarray/tests/test_type.py

@@ -10,7 +10,7 @@ from theano.misc.pkl_utils import CompatUnpickler
 
 from .config import test_ctx_name
 from .test_basic_ops import rand_gpuarray
-from ..type import GpuArrayType
+from ..type import GpuArrayType, gpuarray_shared_constructor
 
 import pygpu
 
@@ -47,8 +47,13 @@ def test_specify_shape():
 
 
 def test_filter_float():
-    s = theano.shared(numpy.array(0.0, dtype='float32'), target=test_ctx_name)
-    theano.function([], updates=[(s, 0.0)])
+    theano.compile.shared_constructor(gpuarray_shared_constructor)
+    try:
+        s = theano.shared(numpy.array(0.0, dtype='float32'),
+                          target=test_ctx_name)
+        theano.function([], updates=[(s, 0.0)])
+    finally:
+        del theano.compile.sharedvalue.shared.constructors[-1]
 
 
 def test_unpickle_gpuarray_as_numpy_ndarray_flag0():

theano/sandbox/gpuarray/type.py

@@ -205,7 +205,13 @@ class GpuArrayType(Type):
                 # (like lists).  We anticipate that the type below
                 # will match and we pass copy=False so it won't make a
                 # second object on the GPU.
-                data = gpuarray.array(data, copy=False, context=self.context)
+                converted_data = gpuarray.array(data, dtype=self.dtype, context=self.context)
+                g_data = gpuarray.array(data, copy=False, context=self.context)
+                if self.values_eq(g_data,
+                                  converted_data,
+                                  force_same_dtype=False):
+                    data = converted_data
+
 
             up_dtype = scalar.upcast(self.dtype, data.dtype)
             if up_dtype == self.dtype:
@@ -263,10 +269,10 @@ class GpuArrayType(Type):
         return GpuFromHost(self.context_name)(other)
 
     @staticmethod
-    def values_eq(a, b):
+    def values_eq(a, b, force_same_dtype=True):
         if a.shape != b.shape:
             return False
-        if a.typecode != b.typecode:
+        if force_same_dtype and a.typecode != b.typecode:
             return False
         a_eq_b = numpy.asarray(compare(a, '==', b))
         if a_eq_b.all():