Not able to execute make from root user

[janismdhanbad]

Hello, I am trying to install this from root user. I have added cub and json as submodules but still, I am not able to execute the make, it is showing the following exact error

Ubuntu 16.04 with python3

root@instance-2:/home/janpreetds/arboretum# make
"" -gencode=arch=compute_61,code="sm_61,compute_61" -c -DSM610 -Xptxas -v -Xcudafe -# -Xcompiler -ffloat-store -O3 -I_cub/ -I_json/src/ -std=c++11 -Xcompiler -fPIC -Xcompiler -O3 -Xcompiler -fopenmp -Xcompiler -Wall -Xcompiler -funroll-loops -Xcompiler -march=native -ccbin=g++ -o io.o src/io/io.cu
/bin/sh: 1: : Permission denied
makefile:31: recipe for target 'io.o' failed
make: *** [io.o] Error 127

Please help, thanks :)

[sh1ng]

For some reason compiler is empty "". Check that PATH contains location of cuda compiler(nvcc) and check cuda installation in general.

[EricPerbos]

Hello sh1ng

On Ubuntu 16.04 with python 3.6, makefile export sm = 500

$ make
"/usr/bin/nvcc" -c -Xptxas -v -Xcudafe -# -Xcompiler -ffloat-store -O3 -I_cub/ -I_json/src/ -std=c++11 -Xcompiler -fPIC -Xcompiler -O3 -Xcompiler -fopenmp -Xcompiler -Wall -Xcompiler -funroll-loops -Xcompiler -march=native -ccbin=g++ -o io.o src/io/io.cu
/usr/lib/gcc/x86_64-linux-gnu/5/include/mwaitxintrin.h(36): error: identifier "__builtin_ia32_monitorx" is undefined

/usr/lib/gcc/x86_64-linux-gnu/5/include/mwaitxintrin.h(42): error: identifier "__builtin_ia32_mwaitx" is undefined

Front end time 1.40 (CPU) 1.00 (elapsed)
2 errors detected in the compilation of "/tmp/tmpxft_00002054_00000000-7_io.cpp1.ii".
Total compilation time 1.41 (CPU) 1.00 (elapsed)
makefile:31: recipe for target 'io.o' failed
make: *** [io.o] Error 2

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2015 NVIDIA Corporation
Built on Tue_Aug_11_14:27:32_CDT_2015
Cuda compilation tools, release 7.5, V7.5.17

[sh1ng]

related probably
tensorflow/tensorflow#1066

[dsakovych]

@janismdhanbad I had the same issue, cuda and was installed and worked fine with tensorflow, but in some way, there was no path to nvcc. So I just added

export LD_LIBRARY_PATH=/usr/local/cuda/lib
export PATH=$PATH:/usr/local/cuda/bin

to .bashrc and it worked.

[janismdhanbad]

Hello @sh1ng , Thanks a lot, after reinstalling the cuda and following what @kotestyle said, it worked!!
:)

[yifanxie]

@EricPerbos in your case, add '-D_MWAITXINTRIN_H_INCLUDED -D_FORCE_INLINES -D__STRICT_ANSI__' at the end of your CFLAGG statement so it looks like

export CFLAGS = $(NVCCFLAGS) -O3 -I_cub/ -I_json/src/ -std=c++11 -Xcompiler -fPIC -Xcompiler -O3 -Xcompiler -fopenmp -Xcompiler -Wall -Xcompiler -funroll-loops -Xcompiler -march=native -ccbin=$(CXX) -D_MWAITXINTRIN_H_INCLUDED -D_FORCE_INLINES -D__STRICT_ANSI__

In my installation (Ubuntu 16.04, Cuda 7.5), I also come across this issue, just follow the link for the solution there to change the file json.hpp

[lishubin]

maybe a noob qustion. i have gtx1070, what sm value should i assign?
i solve this issue by sudo su, and then
export LD_LIBRARY_PATH=/usr/local/cuda/lib
export PATH=$PATH:/usr/local/cuda/bin

then it works

[janismdhanbad]

@lishubin , I have the same card, you can use the default value of 610, it works for me.

[lishubin]

@janismdhanbad
thanks man!
now i cannot git glone the repo.something is changed :(

[janismdhanbad]

Oh, you're welcome @lishubin
Oh, strange, I don't see any recent commits. Are you following the git --recursive thing? cause that worked for me

[lishubin]

someone from kaggle posted the same issue. the cub repo seems to have changed.. i cannot make this package to run.

[yifanxie]

@lishubin are you still struggling to build it? perhaps you can share the error message. I managed to get to up and running, perhasp with too low SM value, my GPU is only using a fraction of its VRAM

[sh1ng]

cub has been updated, probably irrelevant any more