Unstable performance with RK3288(armv7a) when use the new implement of winograd4

[BUG1989]

conv1-1_1 input: 56x56x64 kernel:3 stride:1 output: 54x54x64 time:10.404 GOPS:20.664
conv1-1_2 input: 56x56x64 kernel:3 stride:1 output: 54x54x64 time:10.584 GOPS:20.313
conv1-2_1 input: 56x56x64 kernel:3 stride:1 output: 54x54x64 time:34.943 GOPS:6.153
conv1-2_2 input: 56x56x64 kernel:3 stride:1 output: 54x54x64 time:31.684 GOPS:6.785
conv1-3_1 input: 56x56x64 kernel:3 stride:1 output: 54x54x64 time:39.462 GOPS:5.448
conv1-3_2 input: 56x56x64 kernel:3 stride:1 output: 54x54x64 time:39.940 GOPS:5.383

Testing environment:
CPU:RK3288([email protected])
OS:Android 5.1.1
Thread num:2