enable VGG with MKLDNN layers #4310

tensor-tang · 2017-09-21T15:39:19Z

Related #4309

luotao1

这个PR可以拆成两个：一个是VGG相关的，一个是单测中比较网络使用mkldnn前后，有无区别。

luotao1 · 2017-09-25T06:29:35Z

benchmark/paddle/image/run.mkldnn.sh

+#========== mklml ===========#
+train vgg 64 16
+train vgg 128 16
+train vgg 256 16


41行和47行的注释不明白：一个是用mkldnn，一个是没用？
另外，最后一个layer_num的参数，既然传了16进去，前面的也传19进去吧。

这里比较的是，在WITH_MKLDNN和WITH_MKLML都打开的前提下，使用MKLDNN与不使用MKLDNN的区别。

最后一个参数不是layer的层数，是trainer_count的个数。mkldnn默认用一个，所以没写。

luotao1 · 2017-09-25T06:30:18Z

benchmark/paddle/image/run.mkldnn.sh

@@ -0,0 +1,50 @@
+set -e


这个文件名从run.mkldnn.sh改成run_mkldnn.sh

luotao1 · 2017-09-25T06:33:12Z

benchmark/paddle/image/run.mkldnn.sh

+  else
+    use_mkldnn=0
+    log="logs/${topology}-${thread}mklml-${bs}.log"
+  fi


14-20行：thread=1的时候，用mkldnn？否则不用？
那看41-50行，都是用了mkldnn。
trainer_count和use_mkldnn都做为一个参数，在41-50行的时候传进去会比较清楚。

因为当使用MKLDNN的时候只需要一个trainer，MKLDNN会自己使用多线程进行计算，所以才是现在看到的这样。不过我会改一个把接口都留好的，也方便以后大家测试。

luotao1 · 2017-09-25T06:35:22Z

paddle/trainer/tests/CMakeLists.txt

+          ${CMAKE_CURRENT_BINARY_DIR}/test_CompareMKLDNNandCPU
+              --config_file_a=trainer/tests/sample_trainer_config_simple_net.conf --use_mkldnn_a=True
+              --config_file_b=trainer/tests/sample_trainer_config_simple_net.conf --use_mkldnn_b=False
+              --use_gpu=False


49行的use_gpu可以去掉吧。

这个是因为目前使用对比的时候是不希望使用gpu的。

luotao1 · 2017-09-25T06:36:06Z

paddle/trainer/tests/test_CompareTwoNets.cpp


 DECLARE_string(config);
 DECLARE_string(nics);

 DEFINE_string(config_file_a, "", "config of one network to compare");
 DEFINE_string(config_file_b, "", "config of another network to compare");
+DEFINE_bool(use_mkldnn_a, false, "whether to use mkldnn to run network");


whether to use mkldnn to run config_file_a
whether to use mkldnn to run config_file_b

tensor-tang

好的，谢谢。稍后会把更改提交上。

tensor-tang · 2017-09-25T06:53:33Z

benchmark/paddle/image/run.mkldnn.sh

@@ -0,0 +1,50 @@
+set -e


tensor-tang · 2017-09-25T07:22:20Z

benchmark/paddle/image/run.mkldnn.sh

+  else
+    use_mkldnn=0
+    log="logs/${topology}-${thread}mklml-${bs}.log"
+  fi


因为当使用MKLDNN的时候只需要一个trainer，MKLDNN会自己使用多线程进行计算，所以才是现在看到的这样。不过我会改一个把接口都留好的，也方便以后大家测试。

tensor-tang · 2017-09-25T07:23:47Z

benchmark/paddle/image/run.mkldnn.sh

+#========== mklml ===========#
+train vgg 64 16
+train vgg 128 16
+train vgg 256 16


这里比较的是，在WITH_MKLDNN和WITH_MKLML都打开的前提下，使用MKLDNN与不使用MKLDNN的区别。

最后一个参数不是layer的层数，是trainer_count的个数。mkldnn默认用一个，所以没写。

tensor-tang · 2017-09-25T07:28:02Z

paddle/trainer/tests/CMakeLists.txt

+          ${CMAKE_CURRENT_BINARY_DIR}/test_CompareMKLDNNandCPU
+              --config_file_a=trainer/tests/sample_trainer_config_simple_net.conf --use_mkldnn_a=True
+              --config_file_b=trainer/tests/sample_trainer_config_simple_net.conf --use_mkldnn_b=False
+              --use_gpu=False


这个是因为目前使用对比的时候是不希望使用gpu的。

tensor-tang · 2017-09-25T07:28:26Z

paddle/trainer/tests/test_CompareTwoNets.cpp


 DECLARE_string(config);
 DECLARE_string(nics);

 DEFINE_string(config_file_a, "", "config of one network to compare");
 DEFINE_string(config_file_b, "", "config of another network to compare");
+DEFINE_bool(use_mkldnn_a, false, "whether to use mkldnn to run network");


luotao1

LGTM

luotao1 · 2017-09-25T08:07:01Z

benchmark/paddle/image/run_mkldnn.sh

+    thread=`nproc`
+    log="logs/${topology}-${thread}mklml-${bs}.log"
+  else
+    echo "Wrong input $3, use True or False."


12行和16行是多余的。

luotao1 · 2017-09-25T08:07:44Z

benchmark/paddle/image/run_mkldnn.sh

+fi
+
+#========= mkldnn =========#
+# vgg


43行也是多余的。

tensor-tang added 3 commits September 21, 2017 22:06

add some log info

ed27a3b

add vgg and script for mkldnn benchmark

0f7c407

only link iomp when with MKLDNN and MKLML

f96d31d

tensor-tang added 2 commits September 22, 2017 13:19

add KMP setting and default test vgg19

d72f636

add compare simple net of mkldnn and cpu

bea39f6

tensor-tang requested a review from luotao1 September 22, 2017 13:57

luotao1 reviewed Sep 25, 2017

View reviewed changes

tensor-tang commented Sep 25, 2017

View reviewed changes

rename script and modify comments

86a9434

luotao1 approved these changes Sep 25, 2017

View reviewed changes

luotao1 merged commit 330e992 into PaddlePaddle:develop Sep 25, 2017

tensor-tang deleted the vgg branch September 25, 2017 08:20

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

enable VGG with MKLDNN layers #4310

enable VGG with MKLDNN layers #4310

tensor-tang commented Sep 21, 2017 •

edited

Loading

luotao1 left a comment

luotao1 Sep 25, 2017

tensor-tang Sep 25, 2017

luotao1 Sep 25, 2017

tensor-tang Sep 25, 2017

luotao1 Sep 25, 2017

tensor-tang Sep 25, 2017

luotao1 Sep 25, 2017

tensor-tang Sep 25, 2017

luotao1 Sep 25, 2017

tensor-tang Sep 25, 2017

tensor-tang left a comment

tensor-tang Sep 25, 2017

tensor-tang Sep 25, 2017

tensor-tang Sep 25, 2017

tensor-tang Sep 25, 2017

tensor-tang Sep 25, 2017

luotao1 left a comment

luotao1 Sep 25, 2017

luotao1 Sep 25, 2017

enable VGG with MKLDNN layers #4310

enable VGG with MKLDNN layers #4310

Conversation

tensor-tang commented Sep 21, 2017 • edited Loading

luotao1 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tensor-tang left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

luotao1 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tensor-tang commented Sep 21, 2017 •

edited

Loading