Created by: sneaxiy

以下代码use_constant为True和False时，性能差异很大。

import paddle.fluid as fluid
from paddle.fluid.backward import append_backward
import numpy as np
import time

C = 3
batch_size = 32
scale = 224

use_constant = True
print('use_constant {}'.format(use_constant))

shape = [batch_size, C, scale, scale]

place = fluid.CUDAPlace(0)

image = fluid.layers.data(name='image', shape=shape, dtype='float32', append_batch_size=False)
image.stop_gradient = False

w = float(3.0)

if not use_constant:
    w = fluid.layers.create_global_var(
        shape=[scale, scale],
        value=w,
        dtype='float32',
        persistable=True)

out = image * w 

loss = fluid.layers.reduce_mean(out)

append_backward(loss)

for op in fluid.default_main_program().global_block().ops:
    print(op.type)

exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())

compiled_prog = fluid.CompiledProgram(fluid.default_main_program())

in_tensor = fluid.LoDTensor()
in_tensor.set(np.random.random(size=shape).astype('float32'), place)

start_t = time.time()
for _ in range(1000): 
    exe.run(compiled_prog, feed={image.name: in_tensor}, fetch_list=[out])
end_t = time.time()

print('time cost: {}'.format(end_t - start_t))

Related to #19023 (closed).

怀疑是FastElemwiseGradBroadcast1CUDAKernel导致的问题。