[PTen] Compatible runtime performance optimization

[chenwhql]

PR types

Performance optimization

PR changes

Others

Describe

[PTen] Compatible runtime performance optimization

目前PTen执行兼容态由于引入了pten kernelContext及SmallVector构造析构，pten::DenseTensor的构造析构，导致调度性能下降。

测试代码：

import paddle
import numpy as np
import yep

paddle.set_device("cpu")
x_data = np.random.uniform(0.1, 1, [10]).astype(np.float32)
y_data = np.random.uniform(1, 3, [10]).astype(np.float32)

x = paddle.to_tensor(x_data)
y = paddle.to_tensor(y_data)

yep.start("dot.prof")
for i in range(1000000):
    z = paddle.dot(x, y)
yep.stop()

现develop核心执行函数火焰图如下：

• Run: 24.73%，3.73s

因此，本PR尝试对这一问题进行优化，主要通过缓存KernelContext、DenseTensor解决问题，能够避免大部分不必要的开销。

但由于兼容态存在两种Tensor（fluid Tensor和pten DenseTensor），所以至少会引入Tensor成员和shared_ptr的拷贝构造及析构开销，现阶段难以避免，本PR修改后的核心执行函数火焰图如下：

• Run: 18.03%, 2.33s

tqdm测试，测试代码中循环部分改为：

for i in tqdm(range(1000000)):
    z = paddle.dot(x, y)

现develop的数据：

本PR数据：

在demo上的执行性能约提升27%

[paddle-bot-old]

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

[JiabinYang]

some comment

[JiabinYang]

It seems we don't need it.

[chenwhql]

done, thx

[JiabinYang]

Maybe cause value error, if we resize tensor in kernel. Anyway, we can solve it in the future

[chenwhql]

I know this point, will change kernel output share rule in next pr

[JiabinYang]

LGTM

[zhiqiu]

LGTM for operator.h