Issue with SYCL compiler not supporting OpenMP

User discussions
1

GROMACS version: Latest
GROMACS modification: Yes/No

(Installation guide for exotic configurations - GROMACS 2024.1 documentation)
Regarding the SYCL oneApi build having OpenMP issue, there is following suggestion:

For oneAPI 2024.x: -DCMAKE_C_FLAGS="-isystem /opt/intel/oneapi/compiler/latest/opt/compiler/include" -DCMAKE_CXX_FLAGS="-isystem /opt/intel/oneapi/compiler/latest/opt/compiler/include"

However, I’m using open source dpc++ which does not have the suggested paths. Any ideas what I should use instead?

Thanks

2

Hi! If you’re building DPC++ itself, just build it with OpenMP:

python3 ../buildbot/configure.py --llvm-external-projects=openmp [whatever other options]

After building and installing it, GROMACS should detect OpenMP support without any extra flags (you might want to purge GROMACS’s CMake cache in the build directory).

Note: Open-source DPC++ and Unified Runtime are moving fast, so beware of bugs or API changes when using the tip of sycl branch (or any of the nightly snapshots for that matter).

3

Hey,

After trying that, I’m still having the same issue. (Tried to make a fresh build)

CMake Error at cmake/gmxManageOpenMP.cmake:50 (message):
The compiler you are using does not support OpenMP parallelism.  This might
hurt your performance a lot, in particular with GPUs.  Try using a more
recent version, or a different compiler.  If you don't want to use OpenMP,
disable it explicitly with -DGMX_OPENMP=OFF
Call Stack (most recent call first):
CMakeLists.txt:432 (include)

There seems to be something off with the dpc++ build as I cannot find libomp.so anywhere.

4

That’s weird.

I just made a clean build of the open-source Intel LLVM compiler (07bf3c11ebeafbbb408708dd79f578ec7c974423) using

python3 ../buildbot/configure.py --cuda --hip -t release --llvm-external-projects=openmp

After building, I have ./projects/openmp/runtime/src/omp.h and ./lib/libomp.so in the build directory. Installing puts them to {install_prefix}/lib/clang/19/include/omp.h and {install_prefix}/lib/x86_64-unknown-linux-gnu/libomp.so.

5

Hmm, I’m really bad at setting up LLVM, but here is what I got
(with ad59feb6210ae8dab5e73f61e05631261039498b):

$ python3 configure.py -o ../build_default --llvm-external-projects=openmp -t Release
$ python3 compile.py -o ../build_default -j 70

I have omp.h at /build_default/projects/openmp/runtime/src

No libomp.so at /build_default/lib

No omp.h at /build_default/install/lib/clang/19/include

No libomp.so (and not any directories) at /build_default/install/lib

By looking at ccmake, I have CLANG_DEFAULT_OPENMP_RUNTIME libomp, LLVM_ENABLE_PROJECTS clang;sycl;llvm-spirv;opencl;xpti;xptifw;libdevice;sycl-jit;openmp

Could openmp library be somehow statically linked? Anyway, gromacs still complains about the lacking support.

6

Could openmp library be somehow statically linked?

The OpenMP library could perhaps be statically linked to the application (GROMACS), but building Clang should still produce a library file (libomp.so or libomp.a) for that linking to happen. We’re not linking OpenMP to Clang; we build OpenMP library that should later be linked to GROMACS.

Hmm, I’m really bad at setting up LLVM, but here is what I got

IMO, that is a question better answered at intel/llvm · Discussions · GitHub. The problem is that LLVM does not build OpenMP, that’s not related to GROMACS.

That said, compared to your example, I usually directly call

cmake --build . && cmake --build . --target sycl-toolchain sycl-sanitize sycl-trace install deploy-sycl-toolchain

in the build directory instead of using compile.py.

7

Yeah, I agree. Thanks!