Doesn't work if cuda is not installed in usr/local

If cuda is not installed in ‘usr/local’, it does not run correctly. This is due to lines 42 and 46 in ops.py

Note also that lines 37-40 are not used in ops.py, and therefore neither is line 17 in config.ini.

My simple fix to this was to replace lines 37 to 46 in ops.py with the following:

    out, err = subprocess.Popen(['which', 'nvcc'], stdout=subprocess.PIPE).communicate()
    cuda_dir = out.decode().split('/cuda')[0]

    nvcc_cmd = "nvcc -std=c++11 -c -o {} {} {} -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC -I " + cuda_dir + " --expt-relaxed-constexpr"
    nvcc_cmd = nvcc_cmd.format(" ".join([fn_cu_o, fn_cu_cc]),
                               tf_inc, tf_lib)
    subprocess.check_output(nvcc_cmd, shell=True)
    gcc_cmd = "{} -std=c++11 -shared -o {} {} -fPIC -L " + cuda_dir + "/cuda/lib64 -lcudart {} -O2 -D GOOGLE_CUDA=1"

I also removed line 17 from config.ini

Finally, I know this is a custom config file, but I would also suggest perhaps changing line 15 in config.ini to

g++ = g++

I have tested on 3 different machines, all with different versions of linux, tensorflow, and cuda. With these small changes, your code runs immediately following a clone from this repo on all of them (after copying over the config.ini file of course.)

Just a few small suggestions to make things as out-the-box as possible! 😃

[bragilee]

Hi @DeckerDai ,

sorry about late. It has been a long time since I did not work on this problem. Previously I tried to run this repo then I moved to other projects. For my case, I only make the cuda version+cudnn version compatible with this repo. I am not quite sure about what other factors will have effects on compilation, for example, gcc. I suggest you can start to follow exactly the same versions used in this repo.

Thank you.

[simonmeister]

Many thanks! I changed the code as suggested. Sorry for taking so long to respond - I am very busy with new projects.