Make caffe occur a err：can not find ：lcudnn.so

Autonomous Machines Jetson & Embedded Systems Jetson TX1
1

Yesterday I compiled caffe,but an error ‘can not find ：lcudnn.so’. but I’m sure I have installed cudnn v2,that is require.Anyone has see a similar err ?Who can help me ?

2

Hi LXJ,

Please install the latest JetPack, which will currently install/flash CUDA8 and cuDNN 5.1 on Jetson TX1, and you should be able to compile caffe successfully.

Thanks

3

I understand your mean, but when I downloaded my caffe, which is a special version,caffe&segnet, it demands the cudnn V2 to match it.Now I don’t know the version of cudnn is decided by CUDA or caffe or the version of ununtu.

Thanks

4

Hi,

We have verified cuda-8.0 and cuDNNv5.1 with latest BVLC caffe.
It is recommended to use this setting.

If you need previous caffe branch, could you tell us which branch or snapshot you used?

5

Oh ! The caffe I have been used was not downloaded form Official website , I think it has been modified by someone and it was named ‘caffe-segnet-segnet-cleaned’. If I don’t use the cuDNNv2 ,I can’t make it at all. If I use cuDNNv2 there is the err:

CXX .build_release/src/caffe/proto/caffe.pb.cc
AR -o .build_release/lib/libcaffe.a
LD -o .build_release/lib/libcaffe.so
/usr/bin/ld: skipping incompatible /usr/local/cuda/lib64/libcudnn.so when searching for -lcudnn
/usr/bin/ld: cannot find -lcudnn
collect2: error: ld returned 1 exit status
make: *** [.build_release/lib/libcaffe.so] Error 1

But the other day I make it success,use the setting:
ubuntu 12.04 32-bit
CUDA 7.0
cuDNN v2

Now my setting is :
ubuntu 12.04 64-bit
CUDA 7.0
cuDNN v2
They all Jteson tx1,and this time it has errors.

6

Hi,

Could you turn off cuDNN compile option to check if this issue is still on?

Edit ‘Makefile.config’

--- USE_CUDNN := 1
+++ # USE_CUDNN := 1
7

OK ！ I will try it again. Now I have made it successfully based on Ubuntu 14.04 32bit(using Jetpack 2.1).
I still don’t understand what has decided the cuDNN ? CUDA or caffe ? Even based on the Ubuntu 14.04 32bit,where I have made it successfully,if I just replace cuDNN V2 to V3 or V4 it cannot go all right.

Thanks!

8

It’s decided by caffe.
If you want to upgrade caffe to cuDNNv3 or further, it’s needed to make sure following files are compatible with upgraded cuDNN.

include/caffe/util/:
cudnn.hpp

src/caffe/layers/:
cudnn_conv_layer.cu
cudnn_relu_layer.cpp
cudnn_relu_layer.cu
cudnn_sigmoid_layer.cpp
cudnn_sigmoid_layer.cu
cudnn_tanh_layer.cpp
cudnn_tanh_layer.cu

include/caffe/layers/:
cudnn_relu_layer.hpp
cudnn_sigmoid_layer.hpp
cudnn_tanh_layer.hpp

9

OK ！Thanks very much ！I understand your mean.
Then I have another problem,compared to cuDNN v2, whether the cuDNN v3 or further will have a better performance ? The other day ,I try to make the same caffe with cuDNN v3,but it’s failed. I think that suggests these above files are not compatible with upgraded cuDNN.If I want to upgrade caffe to cuDNNv3 or further,must I replace the caffe? Or there are methods to modify these files?

10

[quote]

OK ! Thanks very much !
I found the version of Caffe (Dec 2016) with cudnn v5.1:https://github.com/TimoSaemann/caffe-segnet-cudnn5
It has the function SegNet