build: fix CUDA_HOME env variable

[IvanGrigorik]

This PR fixes CUDA_HOME environment variable
Now, users need to to specify it explicitly to build it
This fixes that

[kennykos]

This PR fixes CUDA_HOME environment variable Now, users need to to specify it explicitly to build it This fixes that

How was the CUDA_HOME environment variable broken? What is the specific bug that this is addressing?

[IvanGrigorik]

I'm not sure, but @gliga told me that he was unable to install PyKokkos and got some problems with CUDA_HOME env variable
This PR fixes it, if the CUDA_HOME is not already exported

[JBludau]

I don't think that is going in the right direction

[JBludau]

@gliga could you describe what exactly was your problem? This is kind of nebulous right now

[gliga]

@gliga could you describe what exactly was your problem? This is kind of nebulous right now

I had to

export CUDA_HOME=
export CUDAToolkit_ROOT=
export PATH=
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH

for the install to work.
This should be either in instructions and ideally the script should fail saying which variables should be set.

If we can set those automatically (or in the limit suggest the values to use), that will simplify steps further.

[IvanGrigorik]

@gliga @JBludau I suggest creating a script/environment updater that, if the CUDA option is specified, will install CUDA using the Conda environment. Currently, we may encounter many problems if users have incorrectly installed the CUDA toolkit.

Ping if you think this is a good/bad idea. I will create it if someone will say that this is ok

[JBludau]

I am still not understanding why we need this. This looks like CMake pre version 3. why are we not simply linking to cuda and that populates the flags corresponding to the found cuda toolkit?

[JBludau]

@gliga @JBludau I suggest creating a script/environment updater that, if the CUDA option is specified, will install CUDA using the Conda environment. Currently, we may encounter many problems if users have incorrectly installed the CUDA toolkit.

Ping if you think this is a good/bad idea. I will create it if someone will say that this is ok

I am not sure. In general I don't like autoinstalls but if you would use conda to install it, we would stay within the conda world which I would say is ok.

[IvanGrigorik]

@JBludau I have some issues with pykokkos build now, and it looks like its related to Kokkos itself:

[  1%] Building CXX object _deps/kokkos-build/simd/src/CMakeFiles/kokkossimd.dir/Kokkos_SIMD_dummy.cpp.o
[  2%] Linking CXX shared library libkokkossimd.so
[  2%] Built target kokkossimd
[  3%] Building CXX object _deps/kokkos-build/core/src/CMakeFiles/kokkoscore.dir/impl/Kokkos_Abort.cpp.o
[  4%] Building CXX object _deps/kokkos-build/core/src/CMakeFiles/kokkoscore.dir/impl/Kokkos_CPUDiscovery.cpp.o
[  5%] Building CXX object _deps/kokkos-build/core/src/CMakeFiles/kokkoscore.dir/impl/Kokkos_Command_Line_Parsing.cpp.o
[  6%] Building CXX object _deps/kokkos-build/core/src/CMakeFiles/kokkoscore.dir/impl/Kokkos_Core.cpp.o
~/pykokkos/base/_skbuild/linux-x86_64-3.12/cmake-build/_deps/kokkos-src/core/src/Kokkos_MathematicalFunctions.hpp(301): warning #20013-D: calling a constexpr __host__ function("remainder") from a __host__ __device__ function("remainder") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.

Remark: The warnings can be suppressed with "-diag-suppress <warning-number>"

~/pykokkos/base/_skbuild/linux-x86_64-3.12/cmake-build/_deps/kokkos-src/core/src/Kokkos_MathematicalFunctions.hpp(301): warning #20013-D: calling a constexpr __host__ function("remainder") from a __host__ __device__ function("remainderf") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.

And these checks repeated multiple times.
AFAIK, this is the issue with CUDA declarations, but I don't understand why this issue arouse only after my changes.

[JBludau]

@JBludau I have some issues with pykokkos build now, and it looks like its related to Kokkos itself:

[  1%] Building CXX object _deps/kokkos-build/simd/src/CMakeFiles/kokkossimd.dir/Kokkos_SIMD_dummy.cpp.o
[  2%] Linking CXX shared library libkokkossimd.so
[  2%] Built target kokkossimd
[  3%] Building CXX object _deps/kokkos-build/core/src/CMakeFiles/kokkoscore.dir/impl/Kokkos_Abort.cpp.o
[  4%] Building CXX object _deps/kokkos-build/core/src/CMakeFiles/kokkoscore.dir/impl/Kokkos_CPUDiscovery.cpp.o
[  5%] Building CXX object _deps/kokkos-build/core/src/CMakeFiles/kokkoscore.dir/impl/Kokkos_Command_Line_Parsing.cpp.o
[  6%] Building CXX object _deps/kokkos-build/core/src/CMakeFiles/kokkoscore.dir/impl/Kokkos_Core.cpp.o
~/pykokkos/base/_skbuild/linux-x86_64-3.12/cmake-build/_deps/kokkos-src/core/src/Kokkos_MathematicalFunctions.hpp(301): warning #20013-D: calling a constexpr __host__ function("remainder") from a __host__ __device__ function("remainder") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.

Remark: The warnings can be suppressed with "-diag-suppress <warning-number>"

~/pykokkos/base/_skbuild/linux-x86_64-3.12/cmake-build/_deps/kokkos-src/core/src/Kokkos_MathematicalFunctions.hpp(301): warning #20013-D: calling a constexpr __host__ function("remainder") from a __host__ __device__ function("remainderf") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.

And these checks repeated multiple times. AFAIK, this is the issue with CUDA declarations, but I don't understand why this issue arouse only after my changes.

That is weird ... let me see if I can track that down