import tensorflow as tf
print(tf.sysconfig.get_compile_flags(),'\n')
print(tf.sysconfig.get_link_flags())
Running this gave the error: NotFoundError: undefined symbol: _ZTIN10tensorflow8OpKernelE
It seems to be essentially caused by my versions of CUDA and Tensorflow being newer than the one the author tested on (mine are 9.0 and 1.11.0), These helped me solve the problem: these answers, Tensorflow's guide, and this answer.
I could not use the solutions given in the above articles directly because for some reason my system (Linux Ubuntu 16.04 LTS) would give the error "No such file or directory" for ${TF_CFLAGS[@]} and ${TF_LFLAGS[@]}. It needs the space afer-L and -I. Here's how I fixed it:
['-I/usr/local/lib/python2.7/dist-packages/tensorflow/include', '-I/usr/local/lib/python2.7/dist-packages/tensorflow/include/external/nsync/public', '-D_GLIBCXX_USE_CXX11_ABI=1']
['-L/usr/local/lib/python2.7/dist-packages/tensorflow', '-ltensorflow_framework']
I manually replaced what was -I $(tensorflow) with the stuff from the first line, added the stuff from the second line to the g++ commands, and changed -D_GLIBCXX_USE_CXX11_ABI=0 to -D_GLIBCXX_USE_CXX11_ABI=1.
makefile:
nvcc=/usr/local/cuda-9.0/bin/nvcc
cudalib=/usr/local/cuda-9.0/lib64
nsync=/usr/local/lib/python2.7/dist-packages/tensorflow/include/external/nsync/public
TF_INC=/usr/local/lib/python2.7/dist-packages/tensorflow/include
TF_LIB=/usr/local/lib/python2.7/dist-packages/tensorflow/
all: tf_approxmatch_so.so tf_approxmatch_g.cu.o tf_nndistance_so.so tf_nndistance_g.cu.o
tf_approxmatch_so.so: tf_approxmatch_g.cu.o tf_approxmatch.cpp
g++ -std=c++11 tf_approxmatch.cpp tf_approxmatch_g.cu.o -o tf_approxmatch_so.so -shared -fPIC -I $(TF_INC) -I $(nsync) -lcudart -L $(cudalib) -L $(TF_LIB) -ltensorflow_framework -O2 -D_GLIBCXX_USE_CXX11_ABI=1
tf_approxmatch_g.cu.o: tf_approxmatch_g.cu
$(nvcc) -std=c++11 -c -o tf_approxmatch_g.cu.o tf_approxmatch_g.cu -I $(TF_INC) -I $(nsync) -DGOOGLE_CUDA=1 -x cu -Xcompiler -fPIC -O2 -D_GLIBCXX_USE_CXX11_ABI=1
tf_nndistance_so.so: tf_nndistance_g.cu.o tf_nndistance.cpp
g++ -std=c++11 tf_nndistance.cpp tf_nndistance_g.cu.o -o tf_nndistance_so.so -shared -fPIC -I $(TF_INC) -I $(nsync) -lcudart -L $(cudalib) -L $(TF_LIB) -ltensorflow_framework -O2 -D_GLIBCXX_USE_CXX11_ABI=1
tf_nndistance_g.cu.o: tf_nndistance_g.cu
$(nvcc) -std=c++11 -c -o tf_nndistance_g.cu.o tf_nndistance_g.cu -I $(TF_INC) -I $(nsync) -DGOOGLE_CUDA=1 -x cu -Xcompiler -fPIC -O2 -D_GLIBCXX_USE_CXX11_ABI=1
clean:
rm tf_approxmatch_so.so
rm tf_nndistance_so.so
rm *.cu.o
charlesq34/pointnet-autoencoder#12 (comment)
import tensorflow as tf
print(tf.sysconfig.get_compile_flags(),'\n')
print(tf.sysconfig.get_link_flags())
In my case, output:
['-I/home/sohee/anaconda3/envs/py3/lib/python3.6/site-packages/tensorflow/include', '-D_GLIBCXX_USE_CXX11_ABI=1']
['-L/home/sohee/anaconda3/envs/py3/lib/python3.6/site-packages/tensorflow', '-l:libtensorflow_framework.so.1']
make ~/pointnet-autoencoder/tf_ops/tf_compile.sh
CUDA_ROOT=/usr/local/cuda-10.0
TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')
TF_LIB=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib())')
CXX11=-D_GLIBCXX_USE_CXX11_ABI=1
ltensor=-l:libtensorflow_framework.so.1
$CUDA_ROOT/bin/nvcc approxmatch/tf_approxmatch_g.cu -o approxmatch/tf_approxmatch_g.cu.o -c -O2 -DGOOGLE_CUDA=1 -x cu -Xcompiler -fPIC
# TF>=1.4.0
g++ -std=c++11 approxmatch/tf_approxmatch.cpp approxmatch/tf_approxmatch_g.cu.o -o approxmatch/tf_approxmatch_so.so -shared -fPIC -I$TF_INC/ -I$TF_INC/external/nsync/public -L$TF_LIB $ltensor -I$CUDA_ROOT/include -lcudart -L$CUDA_ROOT/lib64/ -O2 $CXX11
$CUDA_ROOT/bin/nvcc nn_distance/tf_nndistance_g.cu -o nn_distance/tf_nndistance_g.cu.o -c -O2 -DGOOGLE_CUDA=1 -x cu -Xcompiler -fPIC
g++ -std=c++11 nn_distance/tf_nndistance.cpp nn_distance/tf_nndistance_g.cu.o -o nn_distance/tf_nndistance_so.so -shared -fPIC -I$TF_INC/ -I$TF_INC/external/nsync/public -L$TF_LIB $ltensor -I$CUDA_ROOT/include -lcudart -L$CUDA_ROOT/lib64/ -O2 $CXX11
$ bash tf_compile.sh