Added the new cpu assert opt.

theano/configdefaults.py

@@ -126,6 +126,15 @@ AddConfigVar(
     in_c_key=False)
 
 
+# This flag determines whether or not to raise error/warning message if there is a
+# CPU in the computational graph.
+AddConfigVar(
+    'assert_no_cpu_op',
+    "Raise an error/warning if there is a CPU op in the computational graph.",
+    EnumStr('ignore', 'warn', 'raise', 'pdb', allow_override=False),
+    in_c_key=False)
+
+
 # Do not add FAST_RUN_NOGC to this list (nor any other ALL CAPS shortcut).
 # The way to get FAST_RUN_NOGC is with the flag 'linker=c|py_nogc'.
 # The old all capital letter way of working is deprecated as it is not

theano/sandbox/cuda/opt.py

@@ -446,6 +446,24 @@ def local_gpu_dot_to_dot22(node):
     return False
 
 
+@local_optimizer([gpu_from_host, host_from_gpu])
+def local_assert_no_cpu_op(node):
+    if not isinstance(node.op, GpuOp) and all([var.owner and isinstance(var.owner.op,
+        HostFromGpu) for var in node.inputs]) and all([var.owner and
+            isinstance(var.owner.op, GpuFromHost) for var in node.outputs]):
+            if config.assert_no_cpu_op == "warn":
+                _logger.warning(("CPU op %s is detected in the computational"
+                                 " graph") % node)
+            elif config.assert_no_cpu_op == "raise":
+                raise RuntimeError("The op %s is on CPU." % node)
+            elif config.assert_no_cpu_op == "pdb":
+                import ipdb; ipdb.set_trace()
+    return None
+
+if config.assert_no_cpu_op != "ignore" and config.assert_no_cpu_op:
+    register_opt()(local_assert_no_cpu_op)
+
+
 @register_opt()
 @local_optimizer([theano.ifelse.IfElse, gpu_from_host])
 def local_gpu_lazy_ifelse(node):
@@ -1911,6 +1929,7 @@ gpu_inplace_elemwise_optimizer = tensor.opt.inplace_elemwise_optimizer_op(
 optdb.register('gpu_inplace_elemwise_opt', gpu_inplace_elemwise_optimizer, 75,
                'fast_run', 'inplace', 'gpu_inplace')
 
+
 register_opt()(tensor.opt.local_remove_useless_assert)
 
 register_opt()(tensor.opt.local_shape_to_shape_i)

theano/sandbox/cuda/tests/test_opt.py

@@ -91,6 +91,16 @@ def test_local_gpu_contiguous_gpu_contiguous():
                      if isinstance(node.op, basic_ops.GpuContiguous)])
 
 
+def test_local_assert_no_cpu_op():
+    x = theano.tensor.fscalar()
+    y = theano.tensor.fscalar()
+    z = x * y
+    f = theano.function([x, y], z, mode=mode_with_gpu)
+    topo = f.maker.fgraph.toposort()
+    a_op = [n for n in topo if not isinstance(n.op, cuda.GpuElemwise)]
+    assert len(a_op) == 3
+
+
 def test_int_pow():
     a = CudaNdarrayType([False])()