Lift GpuAllocEmpty

theano/gof/op.py

@@ -779,6 +779,10 @@ class Op(utils.object2, PureOp, CLinkerOp):
     def _props(self):
         return tuple(getattr(self, a) for a in self.__props__)
 
+    def _props_dict(self):
+        return dict([(a, getattr(self, a))
+                     for a in self.__props__])
+
     def __hash__(self):
         if hasattr(self, '__props__'):
             return hash((type(self), self._props()))

theano/sandbox/gpuarray/opt.py

@@ -273,6 +273,13 @@ def local_gpuaalloc(node, context_name):
     return GpuAlloc(context_name)(*node.inputs)
 
 
+@register_opt('fast_compile')
+@op_lifter([tensor.AllocEmpty])
+def local_gpuaallocempty(node, context_name):
+    return GpuAllocEmpty(context_name=context_name,
+                         **node.op._props_dict())(*node.inputs)
+
+
 @register_opt()
 @local_optimizer([GpuAlloc])
 def local_gpualloc_memset_0(node):

theano/sandbox/gpuarray/tests/test_opt.py

@@ -9,7 +9,8 @@ from theano.tensor.tests import test_basic
 import theano.sandbox.gpuarray
 from .. import basic_ops
 from ..type import GpuArrayType, gpuarray_shared_constructor, get_context
-from ..basic_ops import GpuAlloc, GpuReshape, GpuFromHost, host_from_gpu
+from ..basic_ops import (
+    GpuAlloc, GpuAllocEmpty, GpuReshape, GpuFromHost, host_from_gpu)
 from ..elemwise import GpuCAReduceCuda, GpuCAReduceCPY, GpuElemwise
 from ..subtensor import GpuSubtensor
 
@@ -151,6 +152,29 @@ def test_local_gpualloc_memset_0():
     assert (numpy.asarray(f(2)) == 1).all()
 
 
+def test_local_gpualloc_empty():
+    i = theano.tensor.iscalar()
+    ii = theano.tensor.iscalar()
+
+    # Test with vector
+    a = tensor.AllocEmpty('float32')(i)
+    f = theano.function([i], a, mode=mode_with_gpu)
+    topo = f.maker.fgraph.toposort()
+    assert len(topo) == 2
+    assert isinstance(topo[0].op, GpuAllocEmpty)
+    # This return not initilized data, so we can only check the shape
+    assert f(3).shape == (3,)
+
+    # Test with matrix
+    a = tensor.AllocEmpty('float32')(i, ii)
+    f = theano.function([i, ii], a, mode=mode_with_gpu)
+    topo = f.maker.fgraph.toposort()
+    assert len(topo) == 2
+    assert isinstance(topo[0].op, GpuAllocEmpty)
+    # This return not initilized data, so we can only check the shape
+    assert f(3, 4).shape == (3, 4)
+
+
 def test_rebroadcast():
     d = numpy.random.rand(10, 10).astype('float32')
     v = theano.tensor.fmatrix()