tensorflow: compile failed for tf-gpu 1.4, cuda 9, cudnn 7, vc 2017, windows 10.
I would greatly appreciate if anyone could help me out with compiling. I followed #5600 and #13962 in compiling the wheel, however when building i got into 6 types of in total 90 errors. Respectively c2070, c2059, c2064, c2001, c1057, c2146.
My build command was:
cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release ^ -DSWIG_EXECUTABLE=C:/swigwin-3.0.12/swig.exe ^ -DPYTHON_EXECUTABLE=python.exe ^ -DPYTHON_LIBRARIES=C:/ProgramData/Anaconda3/libs/python36_lib ^ -DPYTHON_INCLUDE_DIR=C:/ProgramData/Anaconda3/Include ^ -DNUMPY_INCLUDE_DIR=C:/ProgramData/Anaconda3/lib/site-packages/numpy/core/include ^ -Dtensorflow_ENABLE_GPU=ON ^ -DCUDNN_HOME="C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0" ^ -Dtensorflow_WIN_CPU_SIMD_OPTIONS=/arch:AVX
Everything looks good until the compiler_arch_native_support failed. It manage to configure the build so I continue with: MSBuild /p:Configuration=Release tf_python_build_pip_package.vcxproj
The last 9999 line of the failed log is here. tf-compile-build-error-log.txt
Warmest Regards, Colman
About this issue
- Original URL
- State: closed
- Created 7 years ago
- Comments: 21 (9 by maintainers)
Commits related to this issue
- [CMake] Don't build tests for RE2 Issue #14691 shows a build error on Windows in the RE2 tests. Since we do not run these tests, and they seem to be causing problems on some platforms, do not build t... — committed to tensorflow/tensorflow by mrry 7 years ago
- [CMake] Don't build tests for RE2 (#14696) * [CMake] Don't build tests for RE2 Issue #14691 shows a build error on Windows in the RE2 tests. Since we do not run these tests, and they seem to be ca... — committed to tensorflow/tensorflow by mrry 7 years ago
If you are seeing a failure similar to this one: http://ci.tensorflow.org/job/tf-master-win-gpu-cmake/
This looks like an nvcc bug, and it is reported to NVIDIA. We are waiting for their investigation
@arunmandal53 Just change CUDA host compiler to v140 by the following CMake option (you can also find it in cmak-gui by checking “Advanced”):
NonCUDA code can still be compiled by v141 in VS 2017. It seems that v141 and nvcc is incompatible.
Reference: https://medium.com/@shiweili/building-tensorflow-c-shared-library-on-windows-e79c90e23e6e
http://ci.tensorflow.org/job/tf-master-win-gpu-cmake/ Looks like our continuous build is healthy. But our build uses VS 2015, and we do not set SIMD options. So these may be the cause of the failure you are seeing.
The error message above is helpful, but still does not capture what compiler emits for that build target. Is it possible for you to file a new issue, and maybe share the full build output through pastebin?