RE: MKLDNN Build (pre: call for contributions to next MXNet release)
+source data to read clearly. > -Original Message- > From: Zhao, Patric [mailto:patric.z...@intel.com] > Sent: Friday, March 16, 2018 8:29 AM > To: 'dev@mxnet.incubator.apache.org' <dev@mxnet.incubator.apache.org> > Cc: Huang, Jin1 <jin1.hu...@intel.com>; Da Zheng <zhengda1...@gmail.com> > Subject: RE: MKLDNN Build (pre: call for contributions to next MXNet release) > > Hi Pedro, > > > > We and Zheng Da tested the performance on Mac laptop (Intel i7 CPU) as > below table. > > The test script is example/image-classification/benchmark_score.py. > > > > MKLDNN shows better performance in MAC too. > > > topo > > batch size > > gcc7 MKLDNN=1 > > gcc7 MKLDNN=0 > > speedup > > alexnet > > 1 > > 38.44 > > 24.49 > > 157% > > 2 > > 50.18 > > 29.62 > > 169% > > 4 > > 59.54 > > 35.40 > > 168% > > 8 > > 70.75 > > 32.49 > > 218% > > 16 > > 73.40 > > 34.93 > > 210% > > 32 > > 47.52 > > 37.19 > > 128% > > vgg16 > > 1 > > 3.80 > > 2.13 > > 178% > > 2 > > 3.52 > > 2.93 > > 120% > > 4 > > 4.54 > > 3.01 > > 151% > > 8 > > 5.12 > > 3.05 > > 168% > > 16 > > 5.37 > > 3.12 > > 172% > > 32 > > 5.36 > > 3.07 > > 175% > > inception-bn > > 1 > > 15.02 > > 14.04 > > 107% > > 2 > > 25.32 > > 15.14 > > 167% > > 4 > > 34.92 > > 14.97 > > 233% > > 8 > > 27.96 > > 15.02 > > 186% > > 16 > > 27.05 > > 15.10 > > 179% > > 32 > > 32.29 > > 15.46 > > 209% > > inception v3 > > 1 > > 10.03 > > 5.89 > > 170% > > 2 > > 11.21 > > 5.99 > > 187% > > 4 > > 13.04 > > 5.89 > > 221% > > 8 > > 12.97 > > 6.01 > > 216% > > 16 > > 12.55 > > 5.13 > > 245% > > 32 > > 12.52 > > 5.65 > > 222% > > resnet-50 > > 1 > > 15.53 > > 7.07 > > 220% > > 2 > > 14.85 > > 7.06 > > 210% > > 4 > > 17.08 > > 7.88 > > 217% > > 8 > > 16.91 > > 8.97 > > 189% > > 16 > > 15.55 > > 9.42 > > 165% > > 32 > > 16.60 > > 7.99 > > 208% > > resnet-152 > > 1 > > 5.90 > > 3.55 > > 166% > > 2 > > 5.64 > > 3.23 > > 175% > > 4 > > 6.11 > > 3.08 > > 198% > > 8 > > 5.22 > > 3.58 > > 146% > > 16 > > 5.28 > > 3.86 > > 137% > > 32 > > 6.20 > > 3.72 > > 167% > > > > > > > > -Original Message- > > > From: Zhao, Patric > > > Sent: Wednesday, March 14, 2018 9:33 PM > > > To: dev@mxnet.incubator.apache.org > > > Cc: Huang, Jin1 <jin1.hu...@intel.com> > > > Subject: MKLDNN Build (pre: call for contributions to next MXNet release) > > > > > > Hi Pedro, > > > > > > Thanks for the good suggestions. I plan to update the install page but a > > slight > > > delay by other things. > > > > > > Actually, I didn't test the performance on OSX since we're lack of Mac > machine. > > > > > > I will try if I can get a Mac in the office. > > > > > > Thanks, > > > > > > --Patric > > > > > > > > > > -Original Message- > > > > From: Pedro Larroy [mailto:pedro.larroy.li...@gmail.com] > > > > Sent: Wednesday, March 14, 2018 9:09 PM > > > > To: dev@mxnet.incubator.apache.org > > > > Subject: Re: call for contributions to next MXNet release > > > > > > > > Hi Patric > > > > > > > > Does it make sense to add instructions to link against mkl in the install > > > docs? > > > > > > > > For example in OSX you can use MKL by downloading > > > > https://github.com/intel/mkl-dnn/releases and adding MKL_RT_LIBRARY > > > > and MKL_INCLUDE_DIR plus -DUSE_MKL_IF_AVAILABLE=ON - > > > > DUSE_MKLML_MKL=ON -DUSE_MKLDNN=ON > > > > > > > > This is not documented in the website. > &g
RE: MKLDNN Build (pre: call for contributions to next MXNet release)
Hi Pedro, We and Zheng Da tested the performance on Mac laptop (Intel i7 CPU) as below table. The test script is example/image-classification/benchmark_score.py. MKLDNN shows better performance in MAC too. topo batch size gcc7 MKLDNN=1 gcc7 MKLDNN=0 speedup alexnet 1 38.44 24.49 157% 2 50.18 29.62 169% 4 59.54 35.40 168% 8 70.75 32.49 218% 16 73.40 34.93 210% 32 47.52 37.19 128% vgg16 1 3.80 2.13 178% 2 3.52 2.93 120% 4 4.54 3.01 151% 8 5.12 3.05 168% 16 5.37 3.12 172% 32 5.36 3.07 175% inception-bn 1 15.02 14.04 107% 2 25.32 15.14 167% 4 34.92 14.97 233% 8 27.96 15.02 186% 16 27.05 15.10 179% 32 32.29 15.46 209% inception v3 1 10.03 5.89 170% 2 11.21 5.99 187% 4 13.04 5.89 221% 8 12.97 6.01 216% 16 12.55 5.13 245% 32 12.52 5.65 222% resnet-50 1 15.53 7.07 220% 2 14.85 7.06 210% 4 17.08 7.88 217% 8 16.91 8.97 189% 16 15.55 9.42 165% 32 16.60 7.99 208% resnet-152 1 5.90 3.55 166% 2 5.64 3.23 175% 4 6.11 3.08 198% 8 5.22 3.58 146% 16 5.28 3.86 137% 32 6.20 3.72 167% > -Original Message- > From: Zhao, Patric > Sent: Wednesday, March 14, 2018 9:33 PM > To: dev@mxnet.incubator.apache.org > Cc: Huang, Jin1 <jin1.hu...@intel.com> > Subject: MKLDNN Build (pre: call for contributions to next MXNet release) > > Hi Pedro, > > Thanks for the good suggestions. I plan to update the install page but a > slight > delay by other things. > > Actually, I didn't test the performance on OSX since we're lack of Mac > machine. > > I will try if I can get a Mac in the office. > > Thanks, > > --Patric > > > > -Original Message- > > From: Pedro Larroy [mailto:pedro.larroy.li...@gmail.com] > > Sent: Wednesday, March 14, 2018 9:09 PM > > To: dev@mxnet.incubator.apache.org > > Subject: Re: call for contributions to next MXNet release > > > > Hi Patric > > > > Does it make sense to add instructions to link against mkl in the install > > docs? > > > > For example in OSX you can use MKL by downloading > > https://github.com/intel/mkl-dnn/releases and adding MKL_RT_LIBRARY > > and MKL_INCLUDE_DIR plus -DUSE_MKL_IF_AVAILABLE=ON - > > DUSE_MKLML_MKL=ON -DUSE_MKLDNN=ON > > > > This is not documented in the website. > > > > What's the performance increase of using MKL in osx when running in > > CPU mode? > > > > Pedro > > > > On Wed, Mar 14, 2018 at 1:54 PM, Zhao, Patric <patric.z...@intel.com> > > wrote: > > > > > My fault, typo for your name, Larroy :( > > > > > > > -Original Message- > > > > From: Zhao, Patric [mailto:patric.z...@intel.com] > > > > Sent: Wednesday, March 14, 2018 8:40 PM > > > > To: dev@mxnet.incubator.apache.org; Huang, Jin1 > > > > <jin1.hu...@intel.com> > > > > Subject: RE: call for contributions to next MXNet release > > > > > > > > Thanks, Larryo. > > > > > > > > This is a PR to fix this issue, > > > > https://github.com/apache/incubator- > > > > mxnet/pull/10075 > > > > The issue will be gone after this PR is merged. > > > > > > > > For the whole logic, we are investigating now and maybe propose a > > > > new approach for Makefile and CMake. > > > > > > > > > -Original Message- > > > > > From: Pedro Larroy [mailto:pedro.larroy.li...@gmail.com] > > > > > Sent: Wednesday, March 14, 2018 8:31 PM > > > > > To: dev@mxnet.incubator.apache.org > > > > > Subject: Re: call for contributions to next MXNet release > > > > > > > > > > Hi > > > > > > > > > > Can you please fix the build logic? > > > > > https://github.com/apache/incubator- > > > > > mxnet/blob/master/CMakeLists.txt#L158 > > > > > > > > > > https://github.com/apache/incubator-mxnet/issues/10072 > > > > > > > > > > It wrongly assumes that you have MKL installed, MKLDNN needs to > > > > > check if MKL is available before, or be disabled by default, > > > > > also in non > > > intel > > > > platforms. > > > > > > > > > > Pedro > > > > > > > > > > On Sun, Mar 11, 2018 at 2:43 PM, Zhao, Patric > > > > > <patric.z...@intel.com> > > >
MKLDNN Build (pre: call for contributions to next MXNet release)
Hi Pedro, Thanks for the good suggestions. I plan to update the install page but a slight delay by other things. Actually, I didn't test the performance on OSX since we're lack of Mac machine. I will try if I can get a Mac in the office. Thanks, --Patric > -Original Message- > From: Pedro Larroy [mailto:pedro.larroy.li...@gmail.com] > Sent: Wednesday, March 14, 2018 9:09 PM > To: dev@mxnet.incubator.apache.org > Subject: Re: call for contributions to next MXNet release > > Hi Patric > > Does it make sense to add instructions to link against mkl in the install > docs? > > For example in OSX you can use MKL by downloading > https://github.com/intel/mkl-dnn/releases and adding MKL_RT_LIBRARY and > MKL_INCLUDE_DIR plus -DUSE_MKL_IF_AVAILABLE=ON - > DUSE_MKLML_MKL=ON -DUSE_MKLDNN=ON > > This is not documented in the website. > > What's the performance increase of using MKL in osx when running in CPU > mode? > > Pedro > > On Wed, Mar 14, 2018 at 1:54 PM, Zhao, Patric> wrote: > > > My fault, typo for your name, Larroy :( > > > > > -Original Message- > > > From: Zhao, Patric [mailto:patric.z...@intel.com] > > > Sent: Wednesday, March 14, 2018 8:40 PM > > > To: dev@mxnet.incubator.apache.org; Huang, Jin1 > > > > > > Subject: RE: call for contributions to next MXNet release > > > > > > Thanks, Larryo. > > > > > > This is a PR to fix this issue, https://github.com/apache/incubator- > > > mxnet/pull/10075 > > > The issue will be gone after this PR is merged. > > > > > > For the whole logic, we are investigating now and maybe propose a > > > new approach for Makefile and CMake. > > > > > > > -Original Message- > > > > From: Pedro Larroy [mailto:pedro.larroy.li...@gmail.com] > > > > Sent: Wednesday, March 14, 2018 8:31 PM > > > > To: dev@mxnet.incubator.apache.org > > > > Subject: Re: call for contributions to next MXNet release > > > > > > > > Hi > > > > > > > > Can you please fix the build logic? > > > > https://github.com/apache/incubator- > > > > mxnet/blob/master/CMakeLists.txt#L158 > > > > > > > > https://github.com/apache/incubator-mxnet/issues/10072 > > > > > > > > It wrongly assumes that you have MKL installed, MKLDNN needs to > > > > check if MKL is available before, or be disabled by default, also > > > > in non > > intel > > > platforms. > > > > > > > > Pedro > > > > > > > > On Sun, Mar 11, 2018 at 2:43 PM, Zhao, Patric > > > > > > > > wrote: > > > > > > > > > Steffen, really thanks for the helps and I totally agree with > > > > > you and Marco's suggestions. > > > > > I will summarize the open issues and status for the review soon. > > > > > > > > > > > What is your github handle? > > > > > Github ID: Patric Zhao, https://github.com/pengzhao-intel/ Da > > Zheng, > > > > > https://github.com/zheng-da/ > > > > > > > > > > > - do you have design docs we can link to? Does the doc cover > > > > > > packaging, > > > > > use > > > > > > of MXNet with full MKL (I understand there are license issues, > > > > > > but we do > > > > > have > > > > > > users who are or plan to use MXNet with the complete MKL > > > > > > package, not > > > > > just > > > > > > DNN subset)? > > > > > Design doc: > > > > > https://cwiki.apache.org/confluence/display/MXNET/The+ > > > > > design+of+MKLDNN+integration > > > > > It's a good suggestion for the full MKL package and we will add > > > > > the related part in the doc. > > > > > > > > > > > - do you have performance measurements (or plan to measure) to > > > > > > include in release notes? > > > > > Yes, we have tested the performance on all C4/C5 instances > > > > > (8x,4x,2x,xlarge,large). And the accuracy are verified as well. > > > > > We plan to publish these data both in the release note and mxnet > > > > > website ( https://mxnet.incubator.apache.org/faq/perf.html). > > > > > > > > > > > - should we talk about the package at the Apr 24th meetup in > > Seattle? > > > > > We're based in Shanghai, China; so maybe @Da Zheng? > > > > > > > > > > > > > > > > -Original Message- > > > > > > From: Steffen Rochel [mailto:steffenroc...@gmail.com] > > > > > > Sent: Sunday, March 11, 2018 6:11 PM > > > > > > To: dev@mxnet.incubator.apache.org > > > > > > Cc: marco.g.ab...@googlemail.com > > > > > > Subject: Re: call for contributions to next MXNet release > > > > > > > > > > > > Patric - added MKL-DNN to the project list. What is your > > > > > > github > > handle? > > > > > > I do agree with Marco that we need to resolve disable tests, > > > > > > broken > > > > > features > > > > > > etc. Jira looks like the right answer to create a list of > > > > > > known > > issues. > > > > > > Questions I do have: > > > > > > - do you have design docs we can link to? Does the doc cover > > > > > > packaging, > > > > > use > > > > > > of MXNet with full MKL (I understand there are license issues, > > > > > > but we do > > > > > have > > > > > > users who are or plan to use