Error in compiling cuda code

[devanshshah7560]

I compiled the makefile succesfully

When I run python train_nn.py I get this error:
tensorflow.python.framework.errors_impl.NotFoundError: ./tf_nndistance_so.so: undefined symbol: _ZN10tensorflow12OpDefBuilder5InputESs

I am using a conda environment with python2.7, tensorflow 1.4 and gcc 4.8

This is my makefile:
nvcc = /usr/local/cuda-10.2/bin/nvcc
cudalib = /usr/local/cuda-10.2/lib64/
tensorflow = /home/dpshah2/.conda/envs/PSRNet/lib/python2.7/site-packages/tensorflow/include
tf_lib = /home/dpshah2/.conda/envs/PSRNet/lib/python2.7/site-packages/tensorflow

all: depthestimate/tf_nndistance_so.so depthestimate/render_balls_so.so
.PHONY : all

depthestimate/tf_nndistance_so.so: depthestimate/tf_nndistance_g.cu.o depthestimate/tf_nndistance.cpp
g++ -std=c++11 depthestimate/tf_nndistance.cpp depthestimate/tf_nndistance_g.cu.o -o depthestimate/tf_nndistance_so.so -shared -fPIC -I $(tensorflow) -ltensorflow_framework -L $(tf_lib) -O2 -D_GLIBCXX_USE_CXX11_ABI=0

depthestimate/tf_nndistance_g.cu.o: depthestimate/tf_nndistance_g.cu
$(nvcc) -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++11 -c -o depthestimate/tf_nndistance_g.cu.o depthestimate/tf_nndistance_g.cu -I $(tensorflow) -DGOOGLE_CUDA=1 -x cu -Xcompiler -fPIC -O2

depthestimate/render_balls_so.so: depthestimate/render_balls_so.cpp
g++ -std=c++11 depthestimate/render_balls_so.cpp -o depthestimate/render_balls_so.so -shared -fPIC -O2 -D_GLIBCXX_USE_CXX11_ABI=0

[sbharadwajj]

I have the same issue. It would be great if someone who were able to successfully compile the cuda code post their environment details.

I am using: tensorflow 1.4.0, cuda: 8.0 and cudnn: 6

[justanhduc]

@shahdev @sbharadwajj I have successfully compiled with tf1.1 (downloaded with pip), cuda 8.0 and python 2.7.

[devanshshah7560]

@justanhduc can you share the makefile and how you managed to get it to work?

[justanhduc]

@shahdev I didn't do anything special besides fixing the CUDA path in the makefile. But I notice that I had to use a fresh clone to make it work, as I tried to compile with Python 3 but it failed, then I removed all the .so files and compiled again with Python 2 but it didn't work.
The make file is below

nvcc = ${CUDA_HOME}/bin/nvcc
cudalib = ${CUDA_HOME}/lib64/
tensorflow = ${HOME}/miniconda3/envs/py27/lib/python2.7/site-packages/tensorflow/include

all: depthestimate/tf_nndistance_so.so depthestimate/render_balls_so.so
.PHONY : all

depthestimate/tf_nndistance_so.so: depthestimate/tf_nndistance_g.cu.o depthestimate/tf_nndistance.cpp
	g++ -std=c++11 depthestimate/tf_nndistance.cpp depthestimate/tf_nndistance_g.cu.o -o depthestimate/tf_nndistance_so.so -shared -fPIC -I $(tensorflow) -lcudart -L $(cudalib) -O2 -D_GLIBCXX_USE_CXX11_ABI=0

depthestimate/tf_nndistance_g.cu.o: depthestimate/tf_nndistance_g.cu
	$(nvcc) -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++11 -c -o depthestimate/tf_nndistance_g.cu.o depthestimate/tf_nndistance_g.cu -I $(tensorflow) -DGOOGLE_CUDA=1 -x cu -Xcompiler -fPIC -O2

depthestimate/render_balls_so.so: depthestimate/render_balls_so.cpp
	g++ -std=c++11 depthestimate/render_balls_so.cpp -o depthestimate/render_balls_so.so -shared -fPIC -O2 -D_GLIBCXX_USE_CXX11_ABI=0

[dwindy]

@shahdev I didn't do anything special besides fixing the CUDA path in the makefile. But I notice that I had to use a fresh clone to make it work, as I tried to compile with Python 3 but it failed, then I removed all the .so files and compiled again with Python 2 but it didn't work.
The make file is below

nvcc = ${CUDA_HOME}/bin/nvcc
cudalib = ${CUDA_HOME}/lib64/
tensorflow = ${HOME}/miniconda3/envs/py27/lib/python2.7/site-packages/tensorflow/include

all: depthestimate/tf_nndistance_so.so depthestimate/render_balls_so.so
.PHONY : all

depthestimate/tf_nndistance_so.so: depthestimate/tf_nndistance_g.cu.o depthestimate/tf_nndistance.cpp
	g++ -std=c++11 depthestimate/tf_nndistance.cpp depthestimate/tf_nndistance_g.cu.o -o depthestimate/tf_nndistance_so.so -shared -fPIC -I $(tensorflow) -lcudart -L $(cudalib) -O2 -D_GLIBCXX_USE_CXX11_ABI=0

depthestimate/tf_nndistance_g.cu.o: depthestimate/tf_nndistance_g.cu
	$(nvcc) -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++11 -c -o depthestimate/tf_nndistance_g.cu.o depthestimate/tf_nndistance_g.cu -I $(tensorflow) -DGOOGLE_CUDA=1 -x cu -Xcompiler -fPIC -O2

depthestimate/render_balls_so.so: depthestimate/render_balls_so.cpp
	g++ -std=c++11 depthestimate/render_balls_so.cpp -o depthestimate/render_balls_so.so -shared -fPIC -O2 -D_GLIBCXX_USE_CXX11_ABI=0

Sorry for bothering you.
Could you explain what address the ${CUDA_HOME} stands for?
I am using anaconda, it seems the CUDA was not installed in /usr/local/cuda-x.x/bin/nvcc
But I am not sure where it is.
Thanks in advance

[justanhduc]

Hi @dwindy. You can check where your nvcc is using which nvcc in terminal. Then set ${CUDA_HOME} to be until the bin folder. Good luck!

[dwindy]

Hi @dwindy. You can check where your nvcc is using which nvcc in terminal. Then set ${CUDA_HOME} to be until the bin folder. Good luck!

Hi @justanhduc . Yes, I found it. Thank you!