Incline rocm 5.6.0 dev #53
Conversation
cameronrutherford
force-pushed
the
incline-rocm-5.6.0-dev
branch
from
4026a4d
to
1a13ab4
Compare
buildsystem/spack/incline/spack.yaml
Outdated
| cc: amdclang | ||
| cxx: amdclang++ | ||
| f77: /share/apps/gcc/8.4.0/bin/gfortran |
There was a problem hiding this comment.
Here we pray amdclang is on the path @ryandanehy
cameronrutherford
force-pushed
the
incline-rocm-5.6.0-dev
branch
from
63e28ee
to
693e498
Compare
| unset ROCR_VISIBLE_DEVICES | ||
| unset CUDA_VISIBLE_DEVICES | ||
| unset GPU_DEVICE_ORDINAL |
There was a problem hiding this comment.
Crazy that this is the fix we need for CI...
| - compiler: | ||
| spec: clang@16.0.0-rocm5.6.0 | ||
| paths: | ||
| cc: amdclang | ||
| cxx: amdclang++ | ||
| f77: /share/apps/gcc/8.4.0/bin/gfortran | ||
| fc: /share/apps/gcc/8.4.0/bin/gfortran | ||
| flags: | ||
| cxxflags: --gcc-toolchain=/share/apps/gcc/8.4.0 | ||
| operating_system: centos7 | ||
| target: x86_64 | ||
| modules: [] |
There was a problem hiding this comment.
@ryandanehy this would be way easier if we just managed compilers as modules within spack like ORNL does...
| # TODO - This is a bit heavy-handed, but otherwise you get gcc which is not ideal | ||
| # - if(NOT CMAKE_C_COMPILER) wasn't working at top of file... | ||
| set(CMAKE_C_COMPILER @CMAKE_C_COMPILER@) | ||
| set(CMAKE_CXX_COMPILER @CMAKE_CXX_COMPILER@) |
There was a problem hiding this comment.
Some extra eyes here would be appreciated...
| # This is just an agly hack to make HIP build work | ||
| get_target_property(hip_includes hip::device INTERFACE_INCLUDE_DIRECTORIES) | ||
| message(STATUS "HIP include directories: ${hip_includes}") | ||
| include_directories(${hip_includes}) | ||
| target_include_directories(ReSolve::resolve_hip INTERFACE $<BUILD_INTERFACE:${hip_includes}>) |
There was a problem hiding this comment.
This way we don't propagate hard-coded install dir
There was a problem hiding this comment.
Although this is still a hack...
| LinSolverDirectCuSolverGLU.hpp | ||
| LinSolverDirectCuSolverRf.hpp | ||
| LinSolverDirectRocSolverRf.hpp | ||
| LinSolverDirectKLU.hpp | ||
| LinSolverIterativeFGMRES.hpp |
There was a problem hiding this comment.
Maybe these should be installed based on the backend configured?
| @@ -8,7 +8,6 @@ | |||
| #include <rocblas/rocblas.h> | |||
| #include <rocsolver/rocsolver.h> | |||
| #include <hip/hip_runtime.h> | |||
| #include <roctracer/roctx.h> | |||
There was a problem hiding this comment.
Spack installation of roctx sucks. There is no exported CMake configuration to consume. I was able to install the libs with packages roctracer-dev and roctracer-dev-api, but it wasn't immediately obvious of a good way to support this without an obvious hack
| f77: /share/apps/gcc/8.4.0/bin/gfortran | ||
| fc: /share/apps/gcc/8.4.0/bin/gfortran | ||
| flags: | ||
| cxxflags: --gcc-toolchain=/share/apps/gcc/8.4.0 |
There was a problem hiding this comment.
FYI spack refuses to function if there is a training slash here ( --gcc-toolchain=/share/apps/gcc/8.4.0/)
Co-authored-by: pelesh <peless@ornl.gov>
No description provided.