CUDNN v4 upgrade. Optionally, CNMEM used for GPU pool allocation. #42

borisfom · 2015-10-13T22:09:49Z

Note: CNMEM(https://github.com/NVIDIA/cnmem.git) submodule added at caffe/cnmem.

Attention:

./cnmem is not getting populated by default with 'git clone' or pull.
To populate added CNMEM submodule, do 'git submodule update'.
Alernatively, CNMEM can be pulled externally from https://github.com/NVIDIA/cnmem.git.
Caffe does not make any assumptions about actual CNMEM code location at the time.
CNMEM has to be built separately. Refer to CNMEM's README.md
CMNEM pool is turned on by default. It can be turned off by adding 'false' parameter to the constructor of caffe::MemoryManager object in caffe_main.cpp and/or test_caffe_main.cpp.

Add argmax_param "axis" to maximise output along the specified axis

488f126 Conflicts: Makefile Makefile.config.example src/caffe/layers/cudnn_conv_layer.cpp src/caffe/layers/cudnn_conv_layer.cu src/caffe/layers/cudnn_lcn_layer.cpp src/caffe/layers/cudnn_lcn_layer.cu

438add5 Conflicts: CMakeLists.txt cmake/Templates/CaffeConfig.cmake.in

38d6baf Conflicts: src/caffe/layers/cudnn_conv_layer.cpp src/caffe/layers/cudnn_conv_layer.cu

dummy alloc / free in MemoryHandlerActivator to ensure that the memory pool has been set up before any potential operations

add badge for travis build and license

Useful for exporting models from MATLAB (e.g. MatConvNet) to Caffe

Add a caffe.io.write_mean function to the MATLAB interface

SilenceLayer Backward bugfix (fixes BVLC#3151)

…fixes Documentation References ImageNet Sample Location

git-subtree-dir: .3rdparty/cnmem git-subtree-split: e817a7a2889389eaf934fe66a9744878b900acba

…/cnmem'

lukeyeager · 2015-10-20T22:19:13Z

.travis.yml

@@ -8,6 +8,9 @@ env:
    - WITH_CUDA=true WITH_CMAKE=true WITH_IO=true
    - WITH_CUDA=false WITH_CMAKE=false WITH_IO=false
    - WITH_CUDA=false WITH_CMAKE=true WITH_IO=false PYTHON_VERSION=3
+#  Currently there is no way to install cudnn via apt-get. Uncomment wjen it's available.


check spelling

lukeyeager · 2015-10-20T22:30:26Z

Woohoo, all the tests are passing! Now lets clean up the commit history and get this merged!

There are several commits that are unrelated to this PR - how did those get in there?
There are a lot of useless "fix" and "merge" commits - let's squash those together.

We don't need to get it down to a single commit - we can preserve authorship - but let's pare it down to a reasonable number of commits with clear commit messages explaining the contribution of each.

Also, the .gitmodules file is still there even after you removed it. Not sure what's going on there.

borisfom · 2015-10-20T23:17:58Z

Done going through the notes - I have removed .gitmodules, too!
Please squash commits as you see fit!

lukeyeager · 2015-10-20T23:18:41Z

Awesome thanks. Ok, I'll give it a shot.

lukeyeager · 2015-10-21T00:48:44Z

Ok I give up. If I want to remove or squash commits I have to use git rebase -i, but I can't do that without removing all merge commits and attempting to turn the history into a linear stream of commits, which leads to lots of tricky merge commits. Not worth it.

On the bright side, we still have a working cuDNN v4 integration!

CUDNN v4 upgrade. Optionally, CNMEM used for GPU pool allocation.

borisfom · 2015-10-21T00:54:42Z

Great news, thanks Luke!

From: Luke Yeager <[email protected]mailto:[email protected]>
Reply-To: NVIDIA/caffe <[email protected]mailto:[email protected]>
Date: Tuesday, October 20, 2015 at 5:48 PM
To: NVIDIA/caffe <[email protected]mailto:[email protected]>
Cc: Boris Fomitchev <[email protected]mailto:[email protected]>
Subject: Re: [caffe] CUDNN v4 upgrade. Optionally, CNMEM used for GPU pool allocation. (#42)

Ok I give up. If I want to remove or squash commits I have to use git rebase -i, but I can't do that without removing all merge commits and attempting to turn the history into a linear stream of commits, which leads to lots of tricky merge commits. Not worth it.

On the bright side, we still have a working cuDNN v4 integration!

Reply to this email directly or view it on GitHubhttps://github.com//pull/42#issuecomment-149745429.

This email message is for the sole use of the intended recipient(s) and may contain
confidential information. Any unauthorized review, use, disclosure or distribution
is prohibited. If you are not the intended recipient, please contact the sender by

reply email and destroy all copies of the original message.

lukeyeager · 2015-10-22T00:37:42Z

Makefile

+  endif
+endif
+
+PROJECT_DIR=$(PWD)


@borisfom why did you add this PROJECT_DIR variable? It broke my debian packaging. Is it really necessary?

I needed so set paths to absolute, let’s discuss this.

From: Luke Yeager [mailto:[email protected]]
Sent: Wednesday, October 21, 2015 5:38 PM
To: NVIDIA/caffe
Cc: Boris Fomitchev
Subject: Re: [caffe] CUDNN v4 upgrade. Optionally, CNMEM used for GPU pool allocation. (#42)

In Makefilehttps://github.com//pull/42#discussion_r42700544:

@@ -7,12 +7,27 @@ $(error $(CONFIG_FILE) not found. See $(CONFIG_FILE).example.)

endif

include $(CONFIG_FILE)

+# Rectify input parameters

+ifeq ($(CPU_ONLY),1)

USE_CUDNN=0

USE_CNMEM=0

+endif

+ifeq ($(USE_CUDNN),1)

+# CNMEM is ON by default in CUDNN is ON

ifeq ($(USE_CNMEM),)

USE_CNMEM=1

endif

+endif

+PROJECT_DIR=$(PWD)

@borisfomhttps://github.com/borisfom why did you add this PROJECT_DIR variable? It broke my debian packaging. Is it really necessary?

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/42/files#r42700544.

This email message is for the sole use of the intended recipient(s) and may contain
confidential information. Any unauthorized review, use, disclosure or distribution
is prohibited. If you are not the intended recipient, please contact the sender by

reply email and destroy all copies of the original message.

timmeinhardt and others added 30 commits September 25, 2015 12:05

Add argmax_param axis

6c02c8b

Implement ArgMaxLayer forward_cpu and reshape for axis param

c77d5e5

Update ArgMaxLayer documentation for axis param

9b2d267

Generalise ArgMaxLayerTest bottom blob shape

a2a5e22

Implement ArgMaxLayerTest for axis param

def3d3c

test

9b903b5

Initial cuDNN v3 support

c8073dc

cuDNN v4 compatibility fix

c089a2e

Merge pull request BVLC#3069 from timmeinhardt/argmax

01e15d0

Add argmax_param "axis" to maximise output along the specified axis

Add CNMeM support

441bddc

488f126 Conflicts: Makefile Makefile.config.example src/caffe/layers/cudnn_conv_layer.cpp src/caffe/layers/cudnn_conv_layer.cu src/caffe/layers/cudnn_lcn_layer.cpp src/caffe/layers/cudnn_lcn_layer.cu

Cmake support for CNMeM

fd0d640

438add5 Conflicts: CMakeLists.txt cmake/Templates/CaffeConfig.cmake.in

Move to centralized cuDNN handle

6aaf7b1

38d6baf Conflicts: src/caffe/layers/cudnn_conv_layer.cpp src/caffe/layers/cudnn_conv_layer.cu

Turn off cuDNN in TravisCI script

4a57fff

Fix potential CNMEM_NOT_INITIALIZED errors

62582cb

dummy alloc / free in MemoryHandlerActivator to ensure that the memory pool has been set up before any potential operations

add badge for travis build and license

98cc023

Merge pull request BVLC#3133 from kashefy/badge

6eae122

add badge for travis build and license

unused variables warning

f0d6549

Fix invalid argument error in tests

848bfda

Add a caffe.io.write_mean function to the MATLAB interface

552a84a

Useful for exporting models from MATLAB (e.g. MatConvNet) to Caffe

Merge pull request BVLC#3058 from zoharby/master

92dc4e6

Add a caffe.io.write_mean function to the MATLAB interface

SilenceLayer Backward bugfix (fixes BVLC#3151)

64f948a

Merge pull request BVLC#3152 from jeffdonahue/silence-backward-fix

8ef3d63

SilenceLayer Backward bugfix (fixes BVLC#3151)

fixes BVLC#3163

e061546

Merge pull request BVLC#3164 from eerwitt/layer-documentation-sample-…

04c7c36

…fixes Documentation References ImageNet Sample Location

CUDNN config improved

02630e0

Initial cuDNN v3 support

ffee008

Fixing memory handler

17b90a2

Fixing memory handler

ee9ab75

CNMEM fix

dfd65ed

Renamed handler class

cda2e18

borisfom added 12 commits October 15, 2015 14:28

CPU_ONLY fix

00d6a1b

macro redefinition fix

b2daa56

Names normalized

906002c

deleted submodule

d12b5cc

Squashed '.3rdparty/cnmem/' content from commit e817a7a

6480847

git-subtree-dir: .3rdparty/cnmem git-subtree-split: e817a7a2889389eaf934fe66a9744878b900acba

Merge commit '6480847b61b0f004bfd69616cd82761a64711b61' as '.3rdparty…

d25e6a5

…/cnmem'

CNMEM integration

8a16cbd

Travis build has Cmake 2.8.7

8635051

Removing Cmake

ce1c463

CNMEM conditional include fixed

3f462d1

Rectified cnmem cmake, renamed 3rd party targets

80af4b1

Fixing CUDNN options

7d5d1cf

lukeyeager reviewed Oct 20, 2015
View reviewed changes

borisfom added 2 commits October 20, 2015 16:15

Cleanup for PR submission

e191830

Removed submodules remainders

0792af7

lukeyeager added a commit that referenced this pull request Oct 21, 2015

Merge pull request #42 from drnikolaev/caffe-0.14

c3cd5c8

CUDNN v4 upgrade. Optionally, CNMEM used for GPU pool allocation.

lukeyeager merged commit c3cd5c8 into NVIDIA:caffe-0.14 Oct 21, 2015

lukeyeager reviewed Oct 22, 2015
View reviewed changes

lukeyeager mentioned this pull request Nov 6, 2015

inference not working on production server NVIDIA/DIGITS#297

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

CUDNN v4 upgrade. Optionally, CNMEM used for GPU pool allocation. #42

CUDNN v4 upgrade. Optionally, CNMEM used for GPU pool allocation. #42

borisfom commented Oct 13, 2015

lukeyeager Oct 20, 2015

borisfom Oct 20, 2015

lukeyeager commented Oct 20, 2015

borisfom commented Oct 20, 2015

lukeyeager commented Oct 20, 2015

lukeyeager commented Oct 21, 2015

borisfom commented Oct 21, 2015

lukeyeager Oct 22, 2015

borisfom Oct 22, 2015

CUDNN v4 upgrade. Optionally, CNMEM used for GPU pool allocation. #42

CUDNN v4 upgrade. Optionally, CNMEM used for GPU pool allocation. #42

Conversation

borisfom commented Oct 13, 2015

lukeyeager Oct 20, 2015

Choose a reason for hiding this comment

borisfom Oct 20, 2015

Choose a reason for hiding this comment

lukeyeager commented Oct 20, 2015

borisfom commented Oct 20, 2015

lukeyeager commented Oct 20, 2015

lukeyeager commented Oct 21, 2015

borisfom commented Oct 21, 2015

reply email and destroy all copies of the original message.

lukeyeager Oct 22, 2015

Choose a reason for hiding this comment

borisfom Oct 22, 2015

Choose a reason for hiding this comment

reply email and destroy all copies of the original message.