Add new operation: BroadcastTensorsOp, test=develop

[jim19930609]

PR types

New features

PR changes

OPs

Describe

• Added BroadcastTensorsOp and BroadcastTensorsGradOp
• Behavior of this Op is aligned with torch.broadcast_tensors
• Supported both CPU and GPU places

API:

paddle.broadcast_tensors() or paddle.tensor.broadcast_tensors()

Code position:

python/paddle/tensor/manipulation.py

Example:

import paddle
import numpy as np

x1 = paddle.to_tensor(np.random.random([1, 2, 3, 4]).astype('float32'))
x2 = paddle.to_tensor(np.random.random([1, 2, 1, 4]).astype('float32'))
x3 = paddle.to_tensor(np.random.random([1, 1, 3, 1]).astype('float32'))
out1, out2, out3 = paddle.broadcast_tensors(input=[x1, x2, x3])
# out1, out2, out3: tensors broadcasted from x1, x2, x3 with shape [1,2,3,4]

CN Doc PR: PaddlePaddle/docs#3606

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[paddle-bot-old]

Sorry to inform you that ac0d714's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[JiabinYang]

Some comments

[TCChenlong]

LGTM

[lanxianghit]

这个值一定需要从python端获取吗

[jim19930609]

用在：core.ops.broadcast_tensors(input, num_inputs)
理论上不需要额外传一个num_inputs进去，可以在C++层从inputs中动态获得。

但是这个"OutNum"是自动代码生成产生的，用来指导创造输出VarBase的数量。不更改自动代码生成逻辑的前提下，很难改这部分逻辑：

类似逻辑的split_op也是从python中获取的"num"，传入core.ops.split()

[jim19930609]

与秋良确认了一下：

1. 目前自动代码生成确实需要产生一个OutNum，后面如果出现较多类似broadcast_tensors的情况，考虑更新自动代码生成逻辑
2. 目前来看多传入一个int num_inputs对性能影响不大

[lanxianghit]

LGTM

[wanghuancoder]

LGTM