Add the reverse version of local_gpu_contiguous

theano/sandbox/cuda/opt.py

@@ -2281,13 +2281,18 @@ def local_gpu_contiguous_gpu_contiguous(node):
 
 
 @register_opt('fast_compile')
-@local_optimizer([tensor.extra_ops.CpuContiguous])
+@local_optimizer([GpuFromHost, tensor.extra_ops.CpuContiguous])
 def local_gpu_contiguous(node):
     if isinstance(node.op, tensor.extra_ops.CpuContiguous):
         x, = node.inputs
         if x.owner and isinstance(x.owner.op, HostFromGpu):
             gpu_x, = x.owner.inputs
             return [tensor.as_tensor_variable(gpu_contiguous(gpu_x))]
+    if isinstance(node.op, GpuFromHost):
+        x, = node.inputs
+        if x.owner and isinstance(x.owner.op, tensor.extra_ops.CpuContiguous):
+            gpu_x, = x.owner.inputs
+            return [gpu_contiguous(gpu_x)]
     return False
 
 

theano/sandbox/cuda/tests/test_opt.py

@@ -111,10 +111,11 @@ def test_local_gpu_contiguous_gpu_contiguous():
 def test_local_gpu_contiguous():
     a = tensor.fmatrix()
     o = tensor.extra_ops.cpu_contiguous(a)
-    f = theano.function([a], o, mode=mode_with_gpu)
-    assert 1 == len([node for node in f.maker.fgraph.toposort()
-                     if isinstance(node.op, basic_ops.GpuContiguous)])
-    f([[2.]])
+    for o in [o, cuda.gpu_from_host(o)]:
+        f = theano.function([a], o, mode=mode_with_gpu)
+        assert 1 == len([node for node in f.maker.fgraph.toposort()
+                         if isinstance(node.op, basic_ops.GpuContiguous)])
+        f([[2.]])
 
 
 def test_local_assert_no_cpu_op():