[Fix] Fix paddle.floor_divide AssertionError when using CUDA 11.2

[Bobholamovic]

PR types

Bug fixes

PR changes

OPs

Describe

How to reproduce

In an environment with CUDA 11.2, the bug can be reproduced by executing the following code:

import paddle

import numpy as np

x = np.array([2, 3, 4])

y = np.array([1, 5, 2])

xp, yp = paddle.to_tensor(x), paddle.to_tensor(y)

rp = paddle.floor_divide(xp, yp)

print(rp)

The running result contains assertion errors even if b is a non-zero number:

Causes

This bug is caused by the calls tostd::trunc(), as the functor gives expected results when these calls are removed.

How I fixed it

My solution is to remove the calls to std::trunc(), since they are NOT necessary here. The reason is two-fold:

1. Currently, paddle.floor_divide() supports only int32 and int64 data, and rounding-to-zero is the defined behavior of integer division in C++.
2. According to link, calling std::trunc() can result in an extra type coercions when input arguments are of integral types, which brings overheads.

It is noteworthy that the floor_divide implementation for integral types in PyTorch is also dependent on C++ integer division, and no std::trunc() is used.

Also note that the name paddle.floor_divide is not so accurate. For divisions involving a negative number and a positive number, the result is rounded to zero instead of performing actual floor division. This PR does not change this behavior.

[paddle-bot]

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.

[chenwhql]

LGTM

[paddle-bot]

你的PR已合入Paddle库，请关注后续测试结果。
Your PR has been merged into the repository. An official integration test will be conducted later. Stay tuned.