conv_transpose gpu版本filter_size>3的条件下固定输入和参数，多次运行得到的结果不同。（在Kernel里调试发现： 在cudnnConvolutionBackwardData之前打印的input和filter相同，经过cudnnConvolutionBackwardData计算之后的output不同）
paddle环境:1.5.2和1.6.1都会出现。
cuda：9.0
cudnn：7.5

测试代码：

import paddle.fluid as fluid
def test(place, data):
    input = fluid.layers.data(dtype='float32', shape=[None, 128, 8, 8], name='data')
    param_attr = fluid.ParamAttr(name='tc_w', initializer=fluid.initializer.Constant(1.0), trainable=True)
    result = fluid.layers.conv2d_transpose(input, num_filters=128, bias_attr=None, param_attr=param_attr,
        filter_size=4, act=None)

    exe = fluid.Executor(place)
    exe.run(fluid.default_startup_program())

    main_program=fluid.default_main_program())

    res, = exe.run(feed={'data':data}, fetch_list=[result])
    return res

import numpy
import numpy as np
numpy.random.seed(13)
#data = numpy.loadtxt("result_batch_norm_81.tmp_3.txt").reshape((1, 128, 8, 8)).astype(np.float32)
data = np.random.rand(1, 128, 8, 8).astype(np.float32)
res_gpu = test(fluid.CUDAPlace(0), data)
np.savetxt("result1_c.txt", res_gpu[0].reshape((-1, 1)))
print res_gpu

结果对比：