GROMACS-SYCL on Intel GPUs

GROMACS version: 2021 (master)
GROMACS modification: No

Hi,

My goal is to demonstrate the feasibility of SYCL implementation of GROMACS on various hardwares.
For instance, on Intel DevCloud, with lasted llvm-intel compiler, I got the following message related to dummy kernel:

WARNING: While sanity checking device #5, Unable to run dummy kernel on device Intel(R) Iris(R) Xe MAX Graphics [0x4905]: Unhandled type of command group -59 (CL_INVALID_OPERATION)


Program: gmx mdrun, version 2021-sycl
Source file: src/gromacs/taskassignment/findallgputasks.cpp (line 86)
MPI rank: 0 (out of 2)

Fatal error:
Cannot run short-ranged nonbonded interactions on a GPU because no GPU is
detected.

Following a similar discussion regarding CUDA support, I am checking other merge requests related to SYCL. Due to my unfamiliarity with GROMACS development, I am not sure which pull requrest should be merged.
Being able to run a simple test with GROMACS-SYCL would be suffice at this stage.

This looks like a software stack issue, can you test whether you can run other programs on the same node?

Snapshot releases are made from the stable 2021 branch, you can use those. Otherwise, you can try the latest master which has more features ported to SYCL, but expect it to change quite a lot as works is ongoing on these features.

Thanks for the suggestions.

It was indeed a issue with software stacks.
I was able to build gromacs using the following cmake options:

$ cmake \
-DGMX_BUILD_OWN_FFTW=ON
-DGMX_GPU=SYCL
-DCMAKE_C_COMPILER=icx
-DCMAKE_CXX_COMPILER=icpx \
-DCMAKE_INSTALL_PREFIX=…

Here, cmake/3.16.3 works while higher version incorrectly identified dpcpp as ‘IntelLLVM 2021.2.0’
For now I am satisfied that at least the binary can runs a simple simulation.

Regards.

1 Like