Hi Martin,
I was just able to compile Amber24 with a similar setup:
Currently Loaded Modules:
1) gcc/11.3.0 3) openmpi/5.0.5-slurm 5) cmake/3.28.3
2) cuda/12.2.0 4) pnetcdf/1.12.3-slurm
Here is my CMAKE line:
cmake /u/droe/amber/amber24_src \
-DCMAKE_INSTALL_PREFIX=$INSTALLDIR \
-DCMAKE_BUILD_TYPE=Release \
-DCOMPILER=GNU \
-DCMAKE_INSTALL_MESSAGE=NEVER \
-DMPI=TRUE -DCUDA=TRUE -DOPENMP=TRUE \
-DPnetCDF_LIBRARY=$PNETCDF_HOME/lib/libpnetcdf.a
-DPnetCDF_INCLUDE_DIR=$PNETCDF_HOME/include \
-DINSTALL_TESTS=TRUE \
-DFORCE_INTERNAL_LIBS=readline \
-DBUILD_PYTHON=TRUE -DDOWNLOAD_MINICONDA=TRUE -DMINICONDA_USE_PY3=TRUE
(I always force internal readline because I've had lots of issues when
I try to use external readline libraries).
Are you running cmake in an empty directory? If not, maybe some old
files from your CUDA11 build are messing up the CUDA 12 build. If
that's not the issue, maybe try different flavors of gcc11/cuda12 if
you have them available? Good luck!
-Dan
On Wed, Oct 16, 2024 at 1:54 PM Martin Cuma via AMBER <amber.ambermd.org> wrote:
>
> I am continuing to work on building Amber on our Rocky Linux 8 cluster, I disabled Perl which got me over the perlmol error I got earlier.
>
> Now I am getting the following error, when using CUDA 12.4.
> [ 39%] Building NVCC (Device) object src/pmemd/src/cuda/CMakeFiles/pmemd_cuda_SPFP_mpi.dir/pmemd_cuda_SPFP_mpi_generated_kNLRadixSort.cu.o
> /uufs/chpc.utah.edu/sys/srcdir/r8/amber/24/amber24_src/src/pmemd/src/./cub/device/dispatch/../../agent/../iterator/tex_ref_input_iterator.cuh(88): error: texture is not a template
> typedef texture<TextureWord> TexRef;
> ^
>
> /uufs/chpc.utah.edu/sys/srcdir/r8/amber/24/amber24_src/src/pmemd/src/./cub/device/dispatch/../../agent/../iterator/tex_ref_input_iterator.cuh(100): error: identifier "cudaBindTexture" is undefined
> return (cub::Debug((cudaError_t) (cudaBindTexture(&offset, ref, d_in)), "/uufs/chpc.utah.edu/sys/srcdir/r8/amber/24/amber24_src/src/pmemd/src/./cub/device/dispatch/../../agent/../iterator/tex_ref_input_iterator.cuh", 100));
> ^
>
> /uufs/chpc.utah.edu/sys/srcdir/r8/amber/24/amber24_src/src/pmemd/src/./cub/device/dispatch/../../agent/../iterator/tex_ref_input_iterator.cuh(109): error: identifier "cudaUnbindTexture" is undefined
> return cub::Debug((cudaError_t) (cudaUnbindTexture(ref)), "/uufs/chpc.utah.edu/sys/srcdir/r8/amber/24/amber24_src/src/pmemd/src/./cub/device/dispatch/../../agent/../iterator/tex_ref_input_iterator.cuh", 109);
> ^
>
> 3 errors detected in the compilation of "/uufs/chpc.utah.edu/sys/srcdir/r8/amber/24/amber24_src/src/pmemd/src/cuda/kNLRadixSort.cu".
> CMake Error at pmemd_cuda_SPFP_mpi_generated_kNLRadixSort.cu.o.RELEASE.cmake:278 (message):
> Error generating file
> /uufs/chpc.utah.edu/sys/srcdir/r8/amber/24/amber24_src_updated/build1/src/pmemd/src/cuda/CMakeFiles/pmemd_cuda_SPFP_mpi.dir//./pmemd_cuda_SPFP_mpi_generated_kNLRadixSort.cu.o
>
> Based on https://forums.developer.nvidia.com/t/cuda-12-0-still-support-for-texture-reference-support-for-pascal-architecture-warp-synchronous-programming/237284/3, it appears that cudaBindTexture has been removed in CUDA 12.0, but, it does appear that AMBER is supporting CUDA >12.0 so I am wondering what may be going on with my build, perhaps it's picking up a wrong preprocessor path? Though the CMakeCache.txt has the right CUDA env. vars in it, e.g. CUDA_VERSION:STRING=12.4
>
> AMBER does build with CUDA 11.8, so, I can stick with that but it'd be nice to have it working with the newer GPUs. These are the modules I have loaded:
> 1) chpc/1.0 (S) 3) openmpi/5.0.3 5) intel-oneapi-mkl/2024.0.0
> 2) gcc/11.2.0 4) cmake/3.26.0 6) cuda/12.4.0
>
> Thanks,
> Martin
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Oct 17 2024 - 08:30:01 PDT
