fix the cumsum bug for large size

[wawltor]

PR types

Function optimization

PR changes

OPs

Describe

fix the cumsum bug for large size.
problem: In the cumsum kernel used the two dims block to cumsum the input tensor, but the max of thread in the dim of block.y is 65535, in some case the dim of block.y exceed 65535, the computation of the kernel will be shutdown.

solution: Use the 1 dim of block to compute.

import numpy as np
import paddle

#  Small shape Tensor,  the result is correct 
x = np.random.randn(2, 128, 128)

x_paddle = paddle.to_tensor(x.copy(), stop_gradient=False)
print('paddle cumsum; ', x_paddle.cumsum(0).abs().sum().numpy())

# Big shape Tensor, the result is error 
x = np.random.randn(2, 1280, 1280)
x_paddle = paddle.to_tensor(x.copy(), stop_gradient=False)
print('paddle cumsum; ', x_paddle.cumsum(0).abs().sum().numpy())

[paddle-bot-old]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[joey12300]

LGTM