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
        any([[c for c in var.clients if isinstance(c[0].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 AssertionError("The op %s is on CPU." % node)
            elif config.assert_no_cpu_op == "pdb":
                pdb.set_trace()

    return None

# Register the local_assert_no_cpu_op: 

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
        any([[c for c in var.clients if isinstance(c[0].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 AssertionError("The op %s is on CPU." % node)
            elif config.assert_no_cpu_op == "pdb":
                pdb.set_trace()

    return None

# Register the local_assert_no_cpu_op: 

def local_gpu_subtensor(node):
    if isinstance(node.op, GpuFromHost):
        host_input = node.inputs[0]
        if host_input.owner and \
           isinstance(host_input.owner.op, tensor.Subtensor):
            subt = host_input.owner.op
            x = host_input.owner.inputs[0]
            if len(x.clients) == 1:
                # It mean, the input of the subtensor is used only by
                # the subtensor. We do not want to move the subtensor
                # to the GPU in that case.
                return
            coords = host_input.owner.inputs[1:]
            return [GpuSubtensor(subt.idx_list)(as_cuda_ndarray_variable(x),
                                                *coords)]
    if isinstance(node.op, tensor.Subtensor):
        x = node.inputs[0]
        if (x.owner and
            isinstance(x.owner.op, HostFromGpu) and
            x.dtype == "float32"):

            gpu_x = x.owner.inputs[0]
            if (gpu_x.owner and
                isinstance(gpu_x.owner.op, GpuFromHost) and
                # And it is a shared var or an input of the graph.
                not gpu_x.owner.inputs[0].owner):

                if len(x.clients) == 1:
                    if any([n == 'output' or isinstance(n.op, GpuOp)
                            for n, _ in node.outputs[0].clients]):
                        return
                    else:
                        return [host_from_gpu(as_cuda_ndarray_variable(
                            node.outputs[0]))]
                    return

            gpu_x, = x.owner.inputs
            coords = node.inputs[1:]
            return [host_from_gpu(GpuSubtensor(
                node.op.idx_list)(gpu_x, *coords))]
    return False 

def local_gpu_subtensor(node):
    if isinstance(node.op, GpuFromHost):
        host_input = node.inputs[0]
        if host_input.owner and \
           isinstance(host_input.owner.op, tensor.Subtensor):
            subt = host_input.owner.op
            x = host_input.owner.inputs[0]
            if len(x.clients) == 1:
                # It mean, the input of the subtensor is used only by
                # the subtensor. We do not want to move the subtensor
                # to the GPU in that case.
                return
            coords = host_input.owner.inputs[1:]
            return [GpuSubtensor(subt.idx_list)(as_cuda_ndarray_variable(x),
                                                *coords)]
    if isinstance(node.op, tensor.Subtensor):
        x = node.inputs[0]
        if (x.owner and
            isinstance(x.owner.op, HostFromGpu) and
            x.dtype == "float32"):

            gpu_x = x.owner.inputs[0]
            if (gpu_x.owner and
                isinstance(gpu_x.owner.op, GpuFromHost) and
                # And it is a shared var or an input of the graph.
                not gpu_x.owner.inputs[0].owner):

                if len(x.clients) == 1:
                    if any([n == 'output' or isinstance(n.op, GpuOp)
                            for n, _ in node.outputs[0].clients]):
                        return
                    else:
                        return [host_from_gpu(as_cuda_ndarray_variable(
                            node.outputs[0]))]
                    return

            gpu_x, = x.owner.inputs
            coords = node.inputs[1:]
            return [host_from_gpu(GpuSubtensor(
                node.op.idx_list)(gpu_x, *coords))]
    return False