使用现在 ncnn 的库比去年 8 月份的 ncnn 库网络推理时间变长很多。。。

[popper0912]

你好 nihui， 使用现在 ncnn 的库比去年 8 月份的 ncnn 库网络推理时间变长很多。我使用的是 caffe 训练的模型，然后转化为 ncnn 库，进行安卓部署。发现使用新库推理时间加长了好几倍，很郁闷，不知道是哪里的问题，还请大神帮助。

[nihui]

应该是哪里部署有问题 你用的是 release 里预编译的静态库吗？

[shineway14]

我的也是

[onexuan]

我也对比过，20180830版本开始变慢了

[sunbinbin1991]

+1，等消息

[BUG1989]

@popper0912 @onexuan @sunbinbin1991 @shineway14 请各位大佬提供以下信息，方便nihui进行分析：

• 测试所用的硬件平台
• benchncnn的log信息

谢谢~~

[SophieChang66]

我的也是cortex A53, 20190731的master下的ncnn比去年的慢了，20190731版本的编译加了openmp,之前的版本没有openmp是关掉的（当时不是我编译的，据说开了openmp和没开一样，不能使用多线程，所以关了）

[nihui]

你们几位，编译下最新的master代码跑benchncnn看看呢？

[onexuan]

没对比benchncnn，只对比目前项目的网络，因为实际运行的网络才有对比性，环境arm7，20190908比20180704慢200ms+

[qaz734913414]

+1，等消息

[SophieChang66]

@nihui 在cortex A53上，20190908和master_20191105的benchmark耗时如下： 20190908版本 loop_count = 20 num_threads = 1 powersave = 0 gpu_device = -1 squeezenet min = 367.37 max = 372.50 avg = 370.06 squeezenet_int8 min = 248.58 max = 252.11 avg = 249.59 mobilenet min = 468.06 max = 472.17 avg = 470.06 mobilenet_int8 min = 404.65 max = 497.18 avg = 430.71 mobilenet_v2 min = 808.40 max = 816.33 avg = 810.73 mobilenet_v3 min = 539.69 max = 543.54 avg = 541.43 shufflenet min = 374.77 max = 377.58 avg = 376.14 mnasnet min = 516.65 max = 519.84 avg = 518.36 proxylessnasnet min = 824.25 max = 827.05 avg = 826.02 googlenet min = 1092.32 max = 1332.60 avg = 1147.07 googlenet_int8 min = 812.46 max = 989.07 avg = 976.71 resnet18 min = 889.95 max = 893.65 avg = 891.96 resnet18_int8 min = 911.78 max = 917.99 avg = 915.17 alexnet min = 1707.62 max = 2038.28 avg = 1882.43 master_20191105: loop_count = 20 num_threads = 1 powersave = 0 gpu_device = -1 squeezenet min = 202.50 max = 206.68 avg = 204.41 squeezenet_int8 min = 249.73 max = 254.57 avg = 252.17 mobilenet min = 308.42 max = 313.81 avg = 310.99 mobilenet_int8 min = 410.07 max = 415.68 avg = 412.29 mobilenet_v2 min = 220.95 max = 224.73 avg = 222.38 mobilenet_v3 min = 194.93 max = 197.22 avg = 196.19 shufflenet min = 142.34 max = 145.66 avg = 144.09 shufflenet_v2 min = 165.68 max = 167.96 avg = 166.53 mnasnet min = 211.42 max = 214.96 avg = 213.30 proxylessnasnet min = 270.10 max = 273.17 avg = 271.71 googlenet min = 840.55 max = 847.38 avg = 843.26 googlenet_int8 min = 807.18 max = 831.52 avg = 811.44 resnet18 min = 660.73 max = 664.10 avg = 662.01 resnet18_int8 min = 759.34 max = 760.61 avg = 760.03 alexnet min = 880.08 max = 1064.36 avg = 1018.88 有很多模型差异很大，都是在板子空闲状态下，前后分别跑的.

[nihui]

@nihui 在cortex A53上，20190908和master_20191105的benchmark耗时如下： 20190908版本 loop_count = 20 num_threads = 1 powersave = 0 gpu_device = -1 squeezenet min = 367.37 max = 372.50 avg = 370.06 squeezenet_int8 min = 248.58 max = 252.11 avg = 249.59 mobilenet min = 468.06 max = 472.17 avg = 470.06 mobilenet_int8 min = 404.65 max = 497.18 avg = 430.71 mobilenet_v2 min = 808.40 max = 816.33 avg = 810.73 mobilenet_v3 min = 539.69 max = 543.54 avg = 541.43 shufflenet min = 374.77 max = 377.58 avg = 376.14 mnasnet min = 516.65 max = 519.84 avg = 518.36 proxylessnasnet min = 824.25 max = 827.05 avg = 826.02 googlenet min = 1092.32 max = 1332.60 avg = 1147.07 googlenet_int8 min = 812.46 max = 989.07 avg = 976.71 resnet18 min = 889.95 max = 893.65 avg = 891.96 resnet18_int8 min = 911.78 max = 917.99 avg = 915.17 alexnet min = 1707.62 max = 2038.28 avg = 1882.43 master_20191105: loop_count = 20 num_threads = 1 powersave = 0 gpu_device = -1 squeezenet min = 202.50 max = 206.68 avg = 204.41 squeezenet_int8 min = 249.73 max = 254.57 avg = 252.17 mobilenet min = 308.42 max = 313.81 avg = 310.99 mobilenet_int8 min = 410.07 max = 415.68 avg = 412.29 mobilenet_v2 min = 220.95 max = 224.73 avg = 222.38 mobilenet_v3 min = 194.93 max = 197.22 avg = 196.19 shufflenet min = 142.34 max = 145.66 avg = 144.09 shufflenet_v2 min = 165.68 max = 167.96 avg = 166.53 mnasnet min = 211.42 max = 214.96 avg = 213.30 proxylessnasnet min = 270.10 max = 273.17 avg = 271.71 googlenet min = 840.55 max = 847.38 avg = 843.26 googlenet_int8 min = 807.18 max = 831.52 avg = 811.44 resnet18 min = 660.73 max = 664.10 avg = 662.01 resnet18_int8 min = 759.34 max = 760.61 avg = 760.03 alexnet min = 880.08 max = 1064.36 avg = 1018.88 有很多模型差异很大，都是在板子空闲状态下，前后分别跑的.

那么现在的 master 和 20180704 比是不是更快了呢...

[SophieChang66]

@nihui 20180704版本的benchmark如下： loop_count = 20 num_threads = 1 powersave = 0 squeezenet min = 243.62 max = 246.16 avg = 244.87 mobilenet min = 399.12 max = 400.93 avg = 400.00 mobilenet_v2 min = 341.33 max = 346.81 avg = 344.55 shufflenet min = 162.92 max = 166.17 avg = 164.35 googlenet min = 996.89 max = 1000.29 avg = 998.58 resnet18 min = 1066.68 max = 1071.64 avg = 1068.82 alexnet min = 1676.75 max = 1678.85 avg = 1677.91 现在的master在benchmark上，相同模型耗时比20180704和20190908的都少。不同的版本耗时差异很大，这些版本有哪些本质上不同的呢？没有对比过源码~

[nihui]

https://github.com/Tencent/ncnn/releases/tag/20191113 最新版本应该比以前的都快了