Disable c code for some float16 GPU ops.

theano/gpuarray/opt.py

@@ -1060,7 +1060,7 @@ def local_gpua_advanced_incsubtensor(op, context_name, inputs, outputs):
     set_instead_of_inc = op.set_instead_of_inc
 
     compute_capability = int(context.bin_id[-2])
-    if compute_capability >= 2 and x.ndim == 1 and y.ndim == 0:
+    if compute_capability >= 2 and x.ndim == 1 and y.ndim == 0 and inputs[0].dtype != 'float16':
         x = x.dimshuffle(0, 'x')
         y = y.dimshuffle('x', 'x')
         ret = GpuAdvancedIncSubtensor1_dev20(

theano/gpuarray/subtensor.py

@@ -589,7 +589,7 @@ class GpuAdvancedIncSubtensor1(Op):
     Implement AdvancedIncSubtensor1 on the gpu.
 
     """
-    _f16_ok = True
+    _f16_ok = False
     __props__ = ('inplace', 'set_instead_of_inc')
     params_type = gpu_context_type
 
@@ -799,7 +799,7 @@ class GpuAdvancedIncSubtensor1_dev20(GpuKernelBase, HideC,
     only avail on compute capability 2.0 and more recent.
 
     """
-    _f16_ok = True
+    _f16_ok = False
 
     def make_node(self, x, y, ilist):
         """

theano/gpuarray/tests/test_subtensor.py

@@ -136,7 +136,8 @@ def test_advinc_subtensor1_vector_scalar():
                           name='y')
         expr = tensor.advanced_inc_subtensor1(x, y, [0, 2])
         f = theano.function([y], expr, mode=mode_with_gpu)
-        assert sum([isinstance(node.op, GpuAdvancedIncSubtensor1_dev20)
+        assert sum([isinstance(node.op, (GpuAdvancedIncSubtensor1_dev20,
+                                         GpuAdvancedIncSubtensor1))
                     for node in f.maker.fgraph.toposort()]) == 1
         rval = f(yval)
         rep = xval.copy()