Gpu reshape opt.

theano/sandbox/cuda/basic_ops.py

@@ -2445,6 +2445,8 @@ class GpuReshape(tensor.Reshape, GpuOp):
 
     # __hash__, __eq__, __str__ come from tensor.Subtensor
     def make_node(self, x, shp):
+        x = as_cuda_ndarray_variable(x)
+        shp = tensor.as_tensor_variable(shp)
         host_reshaped = host_from_gpu(x).reshape(shp, ndim=self.ndim)
         return Apply(self, [x, shp],
                      [CudaNdarrayType(host_reshaped.broadcastable)()])

theano/sandbox/cuda/opt.py

@@ -944,6 +944,35 @@ def local_gpu_reshape(node):
     return False
 
 
+@local_optimizer([GpuReshape])
+def local_gpu_reshape_chain(node):
+    """
+    GuReshape(GpuReshape(shape1),shape2) -> GpuReshape(shape2)
+
+    """
+    if not tensor.opt.opt.check_chain(node, GpuReshape, GpuReshape):
+        return False
+
+    # TODO: this can permit a failing program to run by eliminating
+    #       the lower reshape
+    rval = node.op(node.inputs[0].owner.inputs[0], node.inputs[1])
+    # It might happen that the desired output of this node has a broadcastable
+    # pattern that does not match that of 'rval'. This is when originally, we
+    # were able to figure out that one of the dimensions of the reshape is one,
+    # but some other transformation replaced the shape by one for which this
+    # cannot be guessed.
+    # We should try to figure out why we lost the information about this
+    # constant value... but in the meantime, better not apply this
+    # optimization.
+    if rval.broadcastable == node.outputs[0].broadcastable:
+        return [rval]
+    else:
+        return False
+gpu_cut_copies.register('cut_local_gpu_reshape_chain',
+                        local_gpu_reshape_chain,
+                        'fast_run', 'gpu')
+
+
 @register_opt()
 @local_optimizer([gpu_from_host, tensor.Flatten])
 def local_gpu_flatten(node):

theano/sandbox/cuda/tests/test_opt.py

@@ -819,6 +819,16 @@ class test_diag(theano.tensor.tests.test_nlinalg.test_diag):
               self).__init__(name)
 
 
+def test_local_gpu_reshape():
+    mode = mode_with_gpu
+    a = tensor.fmatrix()
+    b = basic_ops.GpuReshape(3)(a, [2, 3, 4])
+    c = basic_ops.GpuReshape(1)(b, [24])
+    f = theano.function([a], c, mode=mode)
+    topo = f.maker.fgraph.toposort()
+    assert sum(isinstance(node.op, basic_ops.GpuReshape) for node in topo) == 1
+
+
 if __name__ == '__main__':
     test_gpualloc()
     test_opt_gpujoin_onlyajoin()