optimize __call__ to make dygraph faster

[JiabinYang]

PR types

Performance optimization

PR changes

Others

Describe

This PR optimize __call__ in paddle.Layers to optimize dygraph performance without to_static.
We test multi-layer linear performance with code below:

import os
import paddle.fluid.core as core
from paddle import _C_ops
import paddle
import numpy as np
from time import time

num_runs = 100000

os.environ["CUDA_VISIBLE_DEVICES"] = ""
paddle.set_device("cpu")

class FluidMatmulx2(paddle.nn.Layer):
    def __init__(self):
        super(FluidMatmulx2, self).__init__()

        arrW1 = np.ones([4, 128]).astype('float32')
        self.W1 = paddle.to_tensor(arrW1, 'float32', core.CPUPlace())
        self.W1.stop_gradient = False
        
        arrW2 = np.ones([128, 2]).astype('float32')
        self.W2 = paddle.to_tensor(arrW2, 'float32', core.CPUPlace())
        self.W2.stop_gradient = False

    def forward(self, obs):
        Out1 = _C_ops.matmul_v2(obs, self.W1, 'trans_x', False, 'trans_y', False)
        Out = _C_ops.matmul_v2(Out1, self.W2, 'trans_x', False, 'trans_y', False)
        return Out

if __name__ == "__main__":
    input_data = np.ones([32, 4]).astype('float32')
    
    ###########
    # Warm Up #
    ###########
    data_paddle = paddle.to_tensor(input_data.astype(np.float32))
    fluid_matmul = FluidMatmulx2()
    for _ in range(num_runs):
        fluid_matmul.forward(data_paddle)
    
    ###############
    # Performance #
    ###############
    # Fluid Matmul Forward
    data_paddle = paddle.to_tensor(input_data.astype(np.float32))
    ts = time()
    for _ in range(num_runs):
        out = fluid_matmul.forward(data_paddle)
    te = time()
    print("Fluid Matmul Forward: ", 1e6*(te-ts))
    
    # Fluid Matmul Call
    data_paddle = paddle.to_tensor(input_data.astype(np.float32))
    ts = time()
    for _ in range(num_runs):
        out = fluid_matmul(data_paddle)
    te = time()
    print("Fluid Matmul Call: ", 1e6*(te-ts))

And we have result looks like below:

This indicates that we have over 20% time cost on python __call__, and with some profile and test, we found that python decorator context used here is pretty slow.

So, we just use if statement to optimize it. And we can get performance improve as below:

[paddle-bot-old]

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

[zhiqiu]

LGTM

[woshilinsixu]

why I cannot import name '_C_ops' from 'paddle'