compile error: "#" not expected here

[王建堯]

excuse me.

I try to compile matconvnet-1.0-beta17 version and there are some unexpected "#" in cuda code which use cudnn.

Because currently I can not find cudnn v4.

Could the matconvnet works if I remove the part of code for checking cudnn version?

謝謝。

[Andrea Vedaldi]

Hi, I hope this is fixed in the just-released beta17 (the version you used was probably from the devel branch)

[super...@gmail.com]

Hi! I have encountered the same problem, too. Have you figured out how to fixed these errors yet ?

[super...@gmail.com]

Thanks a lot, I have fixed them by removing a few macros in "nnconv_cudnn.cu"

[王建堯]

cuDNN v4 Release Candidate Now Available

https://developer.nvidia.com/cudnn

[王建堯]

i m not sure if the problem is only on windows or not.

but it can works if i choose the branch manually. :)

[張家仁]

My OS is Windows7 64-bit. My matlab version is R2014b, my CUDA version is 6.5 and uses visual studio 2013. 
And I still got this problem with the code in master branch and used cudnn v4.

How can I solve this problem?

王建堯於 2015年12月12日星期六 UTC+8上午10時00分27秒寫道：

[Assaf Mushinsky]

I had the same problem.

There is a problem with the CHECK(x) define in:

nnconv_cudnn.cu

nnbias_cuddn.cu

I have Win8, Matlab2015a, CUDA6.5, CUDNN 4rc, and visual studio 2013.

There is a #if condition to handle interface changes between CUDNN4 and older versions.

In order to do so, they use #if inside CHECK() which don't compile for some reason.

The way to fix it is to put the #if before CHECK (with some redundant code)

for example, instead of: (nnconv_cudnn.cu line 217)

        CHECK(cudnnAddTensor(handle,

#if (CUDNN_VERSION < 4000)

                             CUDNN_ADD_SAME_C,

#endif

                             &alpha,

                             biasesDesc, biases.getMemory() + biasesGrpOffset,

                             &beta,

                             outputDesc, output.getMemory() + outputGrpOffset)) ;

use:

#if (CUDNN_VERSION < 4000)

        CHECK(cudnnAddTensor(handle,

                             CUDNN_ADD_SAME_C,

                             &alpha,

                             biasesDesc, biases.getMemory() + biasesGrpOffset,

                             &beta,

                             outputDesc, output.getMemory() + outputGrpOffset)) ;

#else

        CHECK(cudnnAddTensor(handle,

                             &alpha,

                             biasesDesc, biases.getMemory() + biasesGrpOffset,

                             &beta,

                             outputDesc, output.getMemory() + outputGrpOffset)) ;

#endif

I also had to add #include <algorithm> for std::max().

I hope this helps.

[Karel Lenc]

Hi, thanks for reporting this problem and for the proposed solution!

It was really a bug just related to MSVC compiler when the CuDNN support was enabled...

I've created a new branch on Github called `win-cudnn-fix`, do you think you can test it? I don't have access to a Windows machine with a GPU, so I cannot test whether everything works.

I've also fixed few more bugs related to the MSVC 2010 so I hope I haven't broke anything on the way :)

[Assaf Mushinsky]

Hi Karel,

I've successfully compiled the 'win-cudnn-fix' branch using:

 - visual studio 2013

 - matlab 2015a

 - nvcc compiler

 - cuda 7.5

 - cudnn 4 rc

I had to make the following changes:

In \matconvnet-win-cudnn-fix\matlab\src\bits\impl\nnconv_cudnn.cu I had to add additional include:

#include <algorithm>

Why is the special branch required? 

I could easily compile the main branch using similar changes which can be protected using special flag for windows compilation.

[Karel Lenc]

Hi, the special branch was just temporary so I can check whether it does not break anything for other architectures. It seems to be fine, so I've merged it to master...

Also I've added the include <algorithm> which seems to be needed only for the MSVC2013...

I hope it will work now :)