Setting CUDA_NVCC_FLAGS using CMake - c++

Question

I am trying to compile a CUDA project using CMake on Windows. I am not familiar with make or CMake and I have done some reading over the past few days, but I am still not able to figure this out. I get the following error message:
Chri#Riemann-PC /cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64
$ make
[ 1%] Built target lapacktest
[ 1%] Building NVCC (Device) object CMakeFiles/magma.dir/magmablas/magma_generated_zherk_batched_core.cu.o
nvcc fatal : nvcc cannot find a supported version of Microsoft Visual Studio. Only the versions 2010, 2012, and 2013 are supported
CMake Error at magma_generated_zherk_batched_core.cu.o.cmake:207 (message):
Error generating
/cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64/CMakeFiles/magma.dir/magmablas/./magma_generated_zherk_batched_core.cu.o
CMakeFiles/magma.dir/build.make:2058: recipe for target 'CMakeFiles/magma.dir/magmablas/magma_generated_zherk_batched_core.cu.o' failed
make[2]: *** [CMakeFiles/magma.dir/magmablas/magma_generated_zherk_batched_core.cu.o] Error 1
CMakeFiles/Makefile2:105: recipe for target 'CMakeFiles/magma.dir/all' failed
make[1]: *** [CMakeFiles/magma.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2
Using the make VERBOSE=1 command, I get the following error:
cd /cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64/CMakeFiles/magma.dir/magmablas && /usr/bin/cmake.exe -D verbose:BOOL=1 -D build_configuration:STRING= -D generated_file:STRING=/cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64/CMakeFiles/magma.dir/magmablas/./magma_generated_zherk_batched_core.cu.o -D generated_cubin_file:STRING=/cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64/CMakeFiles/magma.dir/magmablas/./magma_generated_zherk_batched_core.cu.o.cubin.txt -P /cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64/CMakeFiles/magma.dir/magmablas/magma_generated_zherk_batched_core.cu.o.cmake
-- Removing /cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64/CMakeFiles/magma.dir/magmablas/./magma_generated_zherk_batched_core.cu.o
/usr/bin/cmake.exe -E remove /cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64/CMakeFiles/magma.dir/magmablas/./magma_generated_zherk_batched_core.cu.o
-- Generating dependency file: /cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64/CMakeFiles/magma.dir/magmablas/magma_generated_zherk_batched_core.cu.o.NVCC-depend
"/cygdrive/c/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v7.5/bin/nvcc.exe" -M -D__CUDACC__ /cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/magmablas/zherk_batched_core.cu -o /cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/build64/CMakeFiles/magma.dir/magmablas/magma_generated_zherk_batched_core.cu.o.NVCC-depend -ccbin /usr/bin/gcc.exe -m64 -Xcompiler ,\"-fopenmp\",\"-Wall\",\"-Wno-unused-function\",\"-g\" -DHAVE_CUBLAS -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_35,code=compute_35 -DNVCC "-I/cygdrive/c/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v7.5/include" "-I/cygdrive/c/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v7.5/include" -I/cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/include -I/cygdrive/c/Apps/ThirdParty/MAGMA/magma-2.0.1/control
nvcc fatal : nvcc cannot find a supported version of Microsoft Visual Studio. Only the versions 2010, 2012, and 2013 are supported
It is still using the gcc compiler (-ccbin /usr/bin/gcc.exe).
I went into the file: C:\Apps\ThirdParty\cygwin64\usr\share\cmake-3.3.2\Modules\FindCUDA.cmake and added the following lines:
set(VS_DIR "C:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\bin")
message(${VS_DIR})
set(CUDA_NVCC_FLAGS "-ccbin ${VS_DIR}" CACHE STRING "Semi-colon delimit multiple arguments.")
I also added the following line to the nvcc.profile file:
CUDA_NVCC_FLAGS += -ccbin "C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin";
Still no luck. I have spent a few days trying to fix this. I am grateful for any help in this matter.

Answer 1

The option to set the host compiler in FindCUDA is CUDA_HOST_COMPILER, see the FindCUDA documentation.
Try
set(CUDA_HOST_COMPILER "<path to compiler executable>")
or alternatively set it when invoking the cmake command
cmake .. -DCUDA_HOST_COMPILER=<path to compiler executable>
I guess the CUDA_HOST_COMPILER option overrides the ccbin settings of CUDA_NVCC_FLAGS. For other options you can use CUDA_NVCC_FLAGS in the way you tried. (Thus, the title is misleading. It is not a problem of how to set CUDA_NVCC_FLAGS.)
I use this often on Linux, however I don't have a Windows environment to test it there.