Patch opencl opt

[neiltian-tencent]

优化前：
densenet.tnnproto time cost: min = 101.097 ms | max = 103.606 ms | avg = 101.830 ms
inception_v3.tnnproto time cost: min = 134.573 ms | max = 138.854 ms | avg = 136.250 ms
inception_v4.tnnproto time cost: min = 286.451 ms | max = 296.450 ms | avg = 288.725 ms
mobilenet_v1.tnnproto time cost: min = 14.575 ms | max = 15.999 ms | avg = 15.419 ms
mobilenet_v2.tnnproto time cost: min = 12.284 ms | max = 14.734 ms | avg = 13.378 ms
resnet50.tnnproto time cost: min = 95.722 ms | max = 98.836 ms | avg = 97.148 ms
shufflenet_v2_x0.5.tnnproto time cost: min = 10.028 ms | max = 15.307 ms | avg = 12.291 ms
squeezenet_v1.0.tnnproto time cost: min = 24.530 ms | max = 26.495 ms | avg = 25.799 ms
squeezenet_v1.1.tnnproto time cost: min = 11.673 ms | max = 13.708 ms | avg = 12.651 ms

优化后：
densenet.tnnproto time cost: min = 110.838 ms | max = 113.314 ms | avg = 111.700 ms
inception_v3.tnnproto time cost: min = 104.419 ms | max = 106.725 ms | avg = 105.715 ms
inception_v4.tnnproto time cost: min = 225.869 ms | max = 228.219 ms | avg = 227.117 ms
mobilenet_v1.tnnproto time cost: min = 14.149 ms | max = 16.269 ms | avg = 15.220 ms
mobilenet_v2.tnnproto time cost: min = 11.753 ms | max = 14.660 ms | avg = 13.403 ms
resnet50.tnnproto time cost: min = 74.531 ms | max = 76.266 ms | avg = 75.497 ms
shufflenet_v2_x0.5.tnnproto time cost: min = 8.186 ms | max = 13.541 ms | avg = 10.607 ms
squeezenet_v1.0.tnnproto time cost: min = 24.187 ms | max = 32.484 ms | avg = 25.297 ms
squeezenet_v1.1.tnnproto time cost: min = 11.549 ms | max = 13.001 ms | avg = 12.349 ms

主要解决inception_v3, inception_v4以及resnet网络性能差问题，densenet网络性能有一定下降，后续需进一步优化。

[codecov-io]

Codecov Report

Merging #515 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #515   +/-   ##
=======================================
  Coverage   24.48%   24.48%           
=======================================
  Files         287      287           
  Lines        9026     9026           
=======================================
  Hits         2210     2210           
  Misses       6816     6816           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update c8100db...c7120d3. Read the comment docs.