It's good to confirm the detected cuDNN. @lukeyeager could you check this?

I pulled the branch and checked it against various cuDNN versions. Seems to work correctly for >= v2, and fail gracefully for older versions:

-- Found cuDNN: .. (include: /usr/local/cudnn/v1/include, library: /usr/local/cudnn/v1/lib64/libcudnn.so)
-- Found cuDNN: .. (include: /usr/local/cudnn/v2-rc3/include, library: /usr/local/cudnn/v2-rc3/lib64/libcudnn.so)
-- Found cuDNN: 2 (include: /usr/local/cudnn/v2/include, library: /usr/local/cudnn/v2/lib64/libcudnn.so)
-- Found cuDNN: 3.0.01 (include: /usr/local/cudnn/v3-beta2/include, library: /usr/local/cudnn/v3-beta2/lib64/libcudnn.so)
-- Found cuDNN: 3.0.01 (include: /usr/local/cudnn/v3-rc/include, library: /usr/local/cudnn/v3-rc/lib64/libcudnn.so)
-- Found cuDNN: 3.0.02 (include: /usr/local/cudnn/v3-rc3/include, library: /usr/local/cudnn/v3-rc3/lib64/libcudnn.so)
-- Found cuDNN: 3.0.07 (include: /usr/local/cudnn/v3/include, library: /usr/local/cudnn/v3/lib64/libcudnn.so)
-- Found cuDNN: 4.0.1 (include: /usr/local/cudnn/v4-20151013/include, library: /usr/local/cudnn/v4-20151013/lib64/libcudnn.so)

I've verified with the cuDNN team that they plan to stick to this style of declaring the version in the header, so this should keep working with future versions.

One suggestion - I would like to see the version number printed in the summary as well:

-- Dependencies:
--   BLAS              :   Yes (Atlas)
--   Boost             :   Yes (ver. 1.54)
--   glog              :   Yes
--   gflags            :   Yes
--   protobuf          :   Yes (ver. 2.5.0)
--   lmdb              :   Yes (ver. 0.9.10)
--   LevelDB           :   Yes (ver. 1.15)
--   Snappy            :   Yes (ver. 1.1.0)
--   OpenCV            :   Yes (ver. 2.4.8)
--   CUDA              :   Yes (ver. 7.0)
-- 
-- NVIDIA CUDA:
--   Target GPU(s)     :   Auto
--   GPU arch(s)       :   sm_35
--   cuDNN             :   Yes

Something like Yes (ver. 3.0.07) instead of just Yes.

@lukeyeager , will work on update in this case.

• I was unaware that cudnn v1 exists.
• Will short the output to align with summaries.

@lukeyeager ,

• Could please provide me the cudnn.h of "v2-rc3" release ?

• v1 clearly have no marks in its header file !

  I think we should skip display any (ver. X) in case where cudnn.h lack of any versioning marks, and i am afraid booth v1 and v2-rc3 fall into this class.

Waiting for "v2-rc3" header file confirmal, if can help me with the header file.

@cbalint13 Caffe doesn't support cuDNN < v3 since #3160. So I wouldn't worry about parsing older versions. I just provided extra info because I had it handy. Failing to parse v1 headers is perfectly fine. Displaying (ver. X) or (ver. ???) or something seems good to me.

@lukeyeager , @shelhamer

First thanks a lot for your feedback ! Lets then settle down:

• (a) v1, v2-rc will be: Yes (ver. ???)
• (b) v2-final will be: Yes (ver. 2)
• (c) v3 and beyond: Yes (ver. 3.0.07)

For (a) or (b) cmake will halt with: cuDNN version >= 3 required. For (c) all will pass.

Do you agree ? If yes, will proceed coding it up.

@lukeyeager ,

• Updated (foce push). Now we have:

-- NVIDIA CUDA:
--   Target GPU(s)     :   Auto
--   GPU arch(s)       :   sm_52
--   cuDNN             :   Yes (ver. 3.0.07)

And for versions below v3 i considered to simply halt and display ???:

-- Found cuDNN: ver. ??? found (include: /usr/local/cuda/include, library: /usr/lib64/libcudnn.so)
CMake Error at cmake/Cuda.cmake:212 (message):
  cuDNN version >3 is required.

for versions below v3 i considered to simply halt and display ???

That sounds reasonable to me, and a nice improvement.

I tried it and it works as you described - errors when version < 3.

Thanks for the build improvement @cbalint13 and guidance @lukeyeager.