Alpaka: CUDA 9.2 fails to compile

[tdd11235813]

I want to test alpaka with CUDA 9, especially CUDA 9.2. Regarding CUDA 9.0 and 9.1 see issue #426 .
At the moment I use the modules on ZIH/Taurus cluster. Trying to compile examples/bufferCopy from current develop. At the moment on Taurus CUDA 8.0.61 is only working (with GCC 5.3.0 and Boost 1.64.0). CUDA 9.x requires at least Boost 1.65.1. I still need to build & try Boost with gcc 5.4 as it is missing on Taurus.

|CUDA 9.2.88|Boost >=1.65.1|
|---|---|
|gcc 4.9.3|:x:|
|gcc 5.3|:x:|
|gcc 5.4|:question:|
|gcc 5.5|:x:|
|gcc 6.x|:x:|
|gcc 7.x|:x:|

GCC 7.1 + Boost 1.65.1 + CUDA 9.2.88

Currently Loaded Modules:
  1) modenv/classic (S)   2) cmake/3.10.1   3) git/2.15.1   4) gcc/7.1.0   5) bullxmpi/1.2.8.4   6) boost/1.65.1-gnu7.1   7) cuda/9.2.88

```bash

alpaka/example/bufferCopy/build_cuda

cmake .. && make

Fails with:

```bash
[ 50%] Building NVCC (Device) object CMakeFiles/bufferCopy.dir/src/bufferCopy_generated_bufferCopy.cpp.o
/sw/taurus/libraries/cuda/9.2.88/bin/nvcc /alpaka/example/bufferCopy/src//bufferCopy.cpp -x=cu -c -o /alpaka/example/bufferCopy/build_cuda/CMakeFiles/bufferCopy.dir/src/./bufferCopy_generated_bufferCopy.cpp.o -ccbin /sw/global/compilers/gcc/7.1.0/bin/g++ -m64 -DALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_THREADS_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_FIBERS_ENABLED -DALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED -DALPAKA_ACC_GPU_CUDA_ENABLED -DALPAKA_DEBUG=0 -Xcompiler ,\"-fopenmp\",\"-g\" --expt-extended-lambda --expt-relaxed-constexpr --generate-code arch=compute_30,code=sm_30 --generate-code arch=compute_30,code=compute_30 -std=c++11 --use_fast_math --ftz=false -DNVCC -I/sw/taurus/libraries/cuda/9.2.88/include -I/sw/taurus/libraries/boost/1.65.1-gnu7.1/include -I/alpaka/include
/alpaka/include/alpaka/core/Assert.hpp(102): error: this pragma must immediately precede a declaration

Same error occurs for GCC 5.5 + Boost 1.65.1 + CUDA 9.2.88 and GCC 7.1 + Boost 1.66.0 + CUDA 9.2.88.
This is setup working, but it is only for CUDA 8.

1) modenv/classic (S)   2) gcc/5.3.0   3) bullxmpi/1.2.8.4   4) boost/1.64.0-gnu5.3   5) git/2.15.1   6) cmake/3.10.1   7) cuda/8.0.61

Edit: gcc 5.3 and 4.9.3 added to table.

[ax3l]

Also, as a first try, can you just switch the order of ALPAKA_NO_HOST_ACC_WARNING and ALPAKA_FN_HOST_ACC in the erroring lines?

[ax3l]

Is this a CUDA 9.2 or a NVCC 9.2 issue?

(Can you try if clang -x cuda as a device compiler with CUDA 9.2 is affected as well?)

[ax3l]

@BenjaminW3 let us write in the matrix in README.md not CUDA 8.0+ but instead the range we currently test. This us useful when looking at already released versions if they list CUDA 8.0 - 9.1 then one knows at least what was tested.

That does not mean we should assume things break on new releases, but we can document what we know to work.

[BenjaminW3]

At least nvcc 9.2 using clang 5.0.1 fails with the following errors:

/home/travis/llvm/lib/clang/5.0.1/include/smmintrin.h(674): error: argument of type "const __v2di *" is incompatible with parameter of type "__attribute((vector_size(16))) long long *"
/home/travis/llvm/lib/clang/5.0.1/include/avx2intrin.h(836): error: argument of type "const __v4di_aligned *" is incompatible with parameter of type "__attribute((vector_size(32))) long long *"
/home/travis/llvm/lib/clang/5.0.1/include/avx512fintrin.h(9041): error: argument of type "const __v8di_aligned *" is incompatible with parameter of type "__attribute((vector_size(64))) long long *"

I am currently doing a CI run testing nvcc 9.2 using clang 4.0.0 but I assume similar results.

[tdd11235813]

• tested with gcc5.3 and Boost 1.65.1, same error,
• I do not know if it is a CUDA 9.2 or a NVCC 9.2 issue,
• switched order of ALPAKA_NO_HOST_ACC_WARNING and ALPAKA_FN_HOST_ACC, same error,
• clang 4.0 as host compiler with same error
  

mistakenly tested vectorAdd, but should not be different to bufferCopy

/sw/taurus/libraries/cuda/9.2.88/bin/nvcc /alpaka/example/vectorAdd/src//main.cpp -x=cu -c -o /alpaka/example/vectorAdd/build_cuda/CMakeFiles/vectorAdd.dir/src/./vectorAdd_generated_main.cpp.o -ccbin /sw/global/compilers/llvm/4.0.0/bin/clang++ -m64 -DALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_THREADS_ENABLED -DALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED -DALPAKA_ACC_GPU_CUDA_ENABLED -DALPAKA_DEBUG=0 -Xcompiler ,\"-fopenmp=libiomp5\",\"-g\" --expt-extended-lambda --expt-relaxed-constexpr --generate-code arch=compute_30,code=sm_30 --generate-code arch=compute_30,code=compute_30 -std=c++11 --use_fast_math --ftz=false -DNVCC -I/sw/taurus/libraries/cuda/9.2.88/include -I/alpaka/include -I/software/boost_1.65.1/taurus/include
/alpaka/include/alpaka/core/Assert.hpp(103): error: this pragma must immediately precede a declaration

.

• Clang 5.0 as device compiler (set in ccmake for ALPAKA_CUDA_COMPILER), I have this problem:
  
  
  
  • --cuda-path is correctly set, maybe clang needs something?
  
  

[ 50%] Building CXX object CMakeFiles/bufferCopy.dir/src/bufferCopy.cpp.o
/sw/taurus/eb/Clang/5.0.0-GCC-6.4.0-2.28/bin/clang++  -DALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_THREADS_ENABLED -DALPAKA_ACC_GPU_CUDA_ENABLED -DALPAKA_DEBUG=0 -isystem /sw/taurus/libraries/cuda/9.2.88/include -I/alpaka/include -isystem /software/boost_1.65.1/taurus/include  -fopenmp=libomp   -Werror -Weverything -Wno-c++98-compat -Wno-c++98-compat-pedantic -Wno-disabled-macro-expansion -Wno-global-constructors -Wno-padded --system-header-prefix=boost/ -fopenmp=libomp --cuda-path=/sw/taurus/libraries/cuda/9.2.88 --cuda-gpu-arch=sm_30 -Qunused-arguments -Wno-unused-local-typedef -ffast-math -ffp-contract=fast -std=c++11 -ftemplate-depth=512 -x cuda -o CMakeFiles/bufferCopy.dir/src/bufferCopy.cpp.o -c /alpaka/example/bufferCopy/src/bufferCopy.cpp
clang-5.0: error: cannot find libdevice for sm_30. Provide path to different CUDA installation via --cuda-path, or pass -nocudalib to build without linking with libdevice.

[BenjaminW3]

CUDA 9.2 is now supported with both gcc and clang. I will investigate clang 5.0 in a separate ticket.