[Prim] add reduce_as op for paddle

[zeroRains]

PR Category

Others

PR Types

New features

Description

由于broadcast类算子的反向拆解无法适配动态shape，

在静态shape中的例子如下，

x.shape = [2,3]
y.shape = [1,3]
z = x+y 
# z.shape = [2,3]
# dz -> dy
reduce_dim = get_reduce_dim(y.shape, z.shape)
# reduce_dim = [0]
dy = sum(dz, reduce_dim, True)

但在动态shape的编译期拿不到reduce_dim，于是要添加reduce_as算子，其伪代码如下：

def reduce_as(x: Tensor, y: Tensor):
      reduce_dim = get_reduce_dim(x.shape, y.shape)
      res = sum(x, reduce_dim)
      return res

有了这个算子之后，动态shape的场景写法如下：

x.shape = [-1,3]
y.shape = [1,3]
z = x+y 
# z.shape = [-1,3]
# dz -> dy
# dz.shape = [-1,3]
dy = reduce_as(dz, y)

这个pr就是实现对reduce_as op的添加

[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.

[paddle-bot]

❌ The PR is not created using PR's template. You can refer to this Demo.
Please use PR's template, it helps save our maintainers' time so that more developers get helped.

[cyber-pioneer]

这个算子目前没有反向拆解，这里是不能设置的

[zeroRains]

Done

[cyber-pioneer]

补充动态shape的单侧

[zeroRains]

由于在本地cuda 11.2环境, 本地cuda12.0环境 以及CI环境都无法复现test_assign_pos_op 单测出错的场景，现将该PR的实现部分进行拆解，分成多个子PR，确认是哪个部分的改动引起了这个单测出错。这些PR并不需要Review和Merge具体PR如下：

reduce_as 前向计算： #63652

[Xreki]

LGTM for op-benchmark ci

[sunzhongkai588]

LGTM，中文也同步增加文档吧

[XieYunshen]

LGTM

[jeff41404]

can we support other data type of complex64/128、uint8 andint8?

[zeroRains]

后续的PR再进行补充吧