search

LLNL / hiop

HPC solver for nonlinear optimization problems
Other
210 stars 42 forks source link

`hiop@develop%clang-rocm` unable to build #667

Open cameronrutherford opened 1 year ago

cameronrutherford commented 1 year ago

Building HiOp spec hiop@develop%clang@15.0.0-rocm5.3.0 cxxflags="--gcc-toolchain=/share/apps/gcc/8.4.0/" ~cuda~deepchecking~ginkgo~ipo~jsrun+kron+mpi+raja+rocm~shared+sparse amdgpu_target=gfx908 build_system=cmake build_type=Release generator=make arch=linux-centos7-zen on ROCm based platform Incline results in multiple errors based on unused variables - https://gitlab.pnnl.gov/exasgd/frameworks/exago-github-mirror/-/jobs/130334/raw

I understand that compiler flag changes have caused some API changes, and perhaps I am triggering a new CMake change. I will try a few builds of previous HiOp versions perhaps to find when this regression happened.

There could also be Incline specific bugs, and it is yet to be tested on Crusher.

Same build errors are not seen on Deception and Ascent.

cameronrutherford commented 1 year ago

Relevant ExaGO PR - https://github.com/pnnl/ExaGO/pull/20.

Likely going to stick with making 0.7.2 the version ExaGO 1.6.0 targets for now, and then work on 1.0.0+ compatibility in 1.6.1...

cc @abhyshr

nychiang commented 1 year ago

Similar issue to #668 This happens due to the additional c++ flag set(CMAKE_CXX_FLAGS "-Wall -Werror") introduced in PR #653, which has only been tested on CUDA platforms. I will work on the fix of this.

nychiang commented 1 year ago

@cameronrutherford The main problem in this issue is that we used cflag -NDEBUG together with -Wall -Werror. I can reproduce this error in other platforms and have a nasty fix to it (See PR #672). We'd like to introduce error control to HiOp, but it may take some time. I think the fastest fix to this issue, and for you to compile everything on crush, is to remove either cflag NDEBUG or line set(CMAKE_CXX_FLAGS "-Wall -Werror").

Please let me know which way you prefer.

cameronrutherford commented 11 months ago

@jaelynlitz - can you also make sure this is resolved / not an issue during your incline testing? This might be something small to close out?

jaelynlitz commented 11 months ago

I believe this has been fixed for the moment with #673. So this can probably be closed unless there is a desire to actually remove the unused variables flagged by the compiler.