Do not move all scalar float on the GPU.

theano/gpuarray/opt.py

@@ -703,12 +703,6 @@ def local_gpua_elemwise(op, context_name, inputs, outputs):
     if len(outputs) > 1:
         return
 
-    # We move float* scalar only if the outputs is used on the GPU.
-    # This will trigger a backward pass when needed, but will prevent
-    # many useless transfer to only compute GpuElemwise on scalar.
-    if outputs[0].ndim == 0 and len([c for c, _ in outputs[0].clients
-                                     if isinstance(c.op, GpuFromHost)]) == 0:
-        return
     have_cuda = False
     have_opencl = False
     if inputs and isinstance(inputs[0].type, GpuArrayType):

theano/gpuarray/tests/test_opt.py

@@ -502,10 +502,10 @@ def test_not_useless_scalar_gpuelemwise():
         x = np.random.randn(32, 32).astype(np.float32)
         m1 = theano.shared(np.random.randn(32, 32).astype(np.float32))
         loss = (X - tensor.dot(X, m1)).norm(L=2)
-        lr0 = .001
+        lr = theano.shared(np.asarray(.001, dtype=np.float32))
         grad = tensor.grad(loss, m1)
 
-        train = theano.function(inputs=[X], updates=[(m1, m1 - lr1 * grad)],
+        train = theano.function(inputs=[X], updates=[(m1, m1 - lr * grad)],
                                 mode=mode_with_gpu)
         train(x)
         topo = train.maker.fgraph.toposort()

theano/gpuarray/type.py

@@ -38,8 +38,8 @@ def move_to_gpu(data):
     # We don't support complex on the GPU
     if str(data.dtype) in tensor.basic.complex_dtypes:
         return False
-    # We don't want scalar int on the GPU.
-    if data.ndim == 0 and str(data.dtype) in tensor.basic.discrete_dtypes:
+    # We don't want scalars on the GPU.
+    if data.ndim == 0:
         return False
     return True