Created by: dyning

@qingqing01 @panyx0718

reduce_sum在CPU和单卡计算时正常，多卡并行计算时与预期的结果不一致，具体代码如下，test_reduce_sum_functions是入口函数：

import numpy as np
def test_data_gen(place):
    import paddle
    import paddle.fluid as fluid
    import random
    input_name = "select_random_idx"
    feed_dict = {}
    feed_dict[input_name] = fluid.LoDTensor()
    lod = [0]
    tmp_data_list = []
    cur_len = 0
    for bno in range(0, 18):
        curdata = np.zeros((1000000, 1), dtype=np.float32)
        select_idx = random.sample(range(1000000), 100)
        curdata[select_idx] = 1.0
        tmp_data_list.append(curdata)
        cur_len += curdata.shape[0]
        lod.append(cur_len)
    tmp_data_set = np.vstack(tmp_data_list)
    total_num = np.sum(tmp_data_set)
    feed_dict[input_name].set(tmp_data_set, place)
    feed_dict[input_name].set_lod([lod])
    return feed_dict, total_num

def test_reduce_sum_functions():
    import paddle
    import paddle.fluid as fluid
    select_random_idx = fluid.layers.data(name='select_random_idx', shape=[1], dtype='float32', lod_level=1)
    sample_num = fluid.layers.reduce_sum(select_random_idx, dim=0) 
    
    place = fluid.CUDAPlace(0)
    #place = fluid.CPUPlace()
    exe = fluid.Executor(place)
    exe.run(fluid.default_startup_program())
    #fluid.memory_optimize(fluid.default_main_program()) 
    train_exe = fluid.ParallelExecutor(use_cuda=True)
    while True:
        feed_dict, total_num = test_data_gen(place)
        results = train_exe.run([sample_num.name], feed=feed_dict)
        tmp = np.array(results[0])
        tmp_sum = np.sum(tmp)
        if tmp_sum != total_num:
            print tmp, total_num
            sys.exit(-1)
    print "ok"

test_reduce_sum_functions()