enable benchmark resnet with MKLDNN #5422
Conversation
benchmark/paddle/image/run_mkldnn.sh
Outdated
| train vgg 128 True | ||
| train vgg 256 True | ||
| total_cores=`ls -l /sys/devices/system/cpu/ | grep "cpu[0-9]*$" | wc -l` | ||
| online_cores=`cat /sys/devices/system/cpu/cpu*/online | grep -o '1' | wc -l` |
There was a problem hiding this comment.
我在服务器Intel(R) Xeon(R) CPU E5-2620 v2 @ 2.10GHz 上运行42和43行：
- 42行是12
- 43行是11
这样说明是关了超线程？
上次的测试中，超线程的影响不是很大。
There was a problem hiding this comment.
上次的测试是因为还有操作系统的因素在里面。并且，在后面高端处理器中这个配置还是会有一定影响的。
There was a problem hiding this comment.
docker我试了是可以的，但是我这边没有mac，没有测试过。
There was a problem hiding this comment.
mac上面测试会报错/sys/devices/system/cpu/: No such file or directory
There was a problem hiding this comment.
谢谢，那请问下，这几个命令有效吗？
grep 'physical id' /proc/cpuinfo |sort -u|wc -l
grep "cpu cores" /proc/cpuinfo|uniq
grep -c processor /proc/cpuinfo
There was a problem hiding this comment.
上面这个命令也是不行的，都会报错：/proc/cpuinfo: No such file or directory
There was a problem hiding this comment.
嗯，是的，我后来也查到了，所以我还是删除了这里HT的判断，免得把事情弄麻烦了。谢谢
benchmark/paddle/image/run_mkldnn.sh
Outdated
| # resnet-50, 101 and 152 | ||
| train resnet 50 $batchsize $use_mkldnn | ||
| train resnet 101 $batchsize $use_mkldnn | ||
| train resnet 152 $batchsize $use_mkldnn |
There was a problem hiding this comment.
- 这里需要测那么多？那这一个脚本跑下来太久了。vgg和resnet各选一个就够了。
- 个人觉得原来全部写开比较好，方便选择和测试，因为还有一个openblas版本要测。
There was a problem hiding this comment.
可以各选一个。
另外关于openblas的测试，只需要第一个for那里去掉True就好了，这样改动也挺少的。不然你还是要删除好几行。
并且加一个topology，每个batchsize需要写一行，这样省code。
|
@luotao1 对于HT的判断，由于在Mac上比较麻烦做到兼容，也有点得不偿失，所以还是去掉了自动判断，保留了以前的做法。 |
No description provided.