Closed scrouzet closed 4 years ago
This is the output of nvcc -V
. Still the setup.py uses the 10.0 CUDA COMPILER.
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Sun_Jul_28_19:07:16_PDT_2019
Cuda compilation tools, release 10.1, V10.1.243
Can you try adding '-DCMAKE_CUDA_COMPILER=/usr/local/cuda-10.1/bin/nvcc' in cmake_args in setup.py and run again.
For example: cmake_args = ['-DCMAKE_LIBRARY_OUTPUT_DIRECTORY=' + extdir + "/spconv", '-DCMAKE_PREFIX_PATH=' + LIBTORCH_ROOT, '-DPYBIND11_PYTHON_VERSION={}'.format(PYTHON_VERSION), '-DCMAKE_CUDA_COMPILER=/usr/local/cuda-10.1/bin/nvcc', '-DSPCONV_BuildTests=OFF', '-DCMAKE_CUDA_FLAGS="--expt-relaxed-constexpr"']
@gauravjindal3513 I did it and it worked.
There was then a new error nvcc fatal : Unknown option 'Wall'
, but this was solved by commenting 2 lines in the file Caffe2Targets.cmake
(pytorch package) as indicated in https://github.com/traveller59/spconv/issues/69.
Thanks!
When trying to build the wheel for spconv I get this error that I do not know how to solve.
The machine is running Ubuntu 18.04, and this is in a miniconda env named pointpillars with Python 3.8 (could this be the issue?). I see in the output that it relies on cuda10-0 to compile, while this env is supposed to use cuda10-1 (I installed pytorch with the 10-1 toolkit). the cuda simlink to cuda10-1 so if this is what I should change, it may rather be inside spconv (or elsewhere?), thanks for letting me know.
Let me know if other specific info would be required for you to help. Thanks in advance of your help.