Problems with AutoDock-GPU compiled on a cluster

[xavgit]

Hi,
I have compiled from the sources the latest version of AutoDock-GPU on the Leonardo cluster.
I have compiled two versions.
The first with DEVICE= CUDA and NUMWI=256 and the second with DEVICE=OCLGPU and NUMWI=256.

I have used the following commands:
module load cuda
module load python
and
module list returns:
Currently Loaded Modulefiles:

1. profile/base 2) python/3.10.8--gcc--11.3.0 3) cuda/11.8

Key:
default-version
Then
export GPU_INCLUDE_PATH=/leonardo/prod/opt/compilers/cuda/11.8/none/include
export GPU_LIBRARY_PATH=/leonardo/prod/opt/compilers/cuda/11.8/none/lib64
make DEVICE= ............................
If I run both the executables without any arguments there is no problem.
I have made a test with only one ligand with the compiled versions and I receive errors.
First I use prepare_gpf4.py and ADFR's autogrid4 and then the AutoDock executables.

For the CUDA version I run:
$ python3 test_ad4gpu.py
sh: line 1: 3637611 Aborted (core dumped) /leonardo/home/userexternal/slemme00/sources/AutoDock-GPU/bin_256wi/autodock_gpu_256wi -x 0 --ffile receptor.maps.fld --lfile DB16260.pdbqt --nrun 100 -N ./docking_res/DB16260.pdbqt_docking_res --gbest 1 > ./docking_res/DB16260.pdbqt_docking_res.log 2>&1
$ less docking_res/DB16260.pdbqt_docking_res.log
autodock_gpu_256wi: ./host/src/performdocking.cpp:128: void setup_gpu_for_docking(GpuData&, GpuTempData&): Assertion `0' failed.

I also have used previous python script with slurm directives and 40 ligands but I receive the same problems.

For the OCLGPU version I run:
$ python3 test_ad4gpu_ocl.py
$ less docking_res_ocl/DB16260.pdbqt_docking_res.log
AutoDock-GPU version: v1.5.3-54-g41083c5e1224d54ad043b62ca53f6618d5e8325d-dirty

Running 1 docking calculation

Kernel source used for development: ./device/calcenergy.cl
Kernel string used for building: ./host/inc/stringify.h
Kernel compilation flags: -I ./device -I ./common -DN256WI -cl-mad-enable
Error: clGetPlatformIDs(): -1001

For the system I use at login I have:
Atos Bull Sequana XH21355 "Da Vinci" Blade -
Red Hat Enterprise Linux 8.6 (Ootpa)

3456 compute nodes with:
- 32 cores Ice Lake at 2.60 GHz
- 4 x NVIDIA Ampere A100 GPUs, 64GB
- 512 GB RAM

Internal Network: Nvidia Mellanox HDR DragonFly++
SLURM 22.05.7

test_ad4gpu.py
import os
os.system( '/leonardo/home/userexternal/slemme00/sources/AutoDock-GPU/bin_256wi/autodock_gpu_256wi -x 0 --ffile receptor.maps.fld --lfile ' + 'DB16260.pdbqt' + ' --nrun 100 -N ' + './docking_res/' + 'DB16260.pdbqt' + '_docking_res --gbest 1 > ./docking_res/' + 'DB16260.pdbqt' + '_docking_res.log 2>&1' )

test_ad4gpu_ocl.py
import os
os.system( '/leonardo/home/userexternal/slemme00/sources/AutoDock-GPU/bin_oclgpu_256/autodock_gpu_256wi -x 0 --ffile receptor.maps.fld --lfile ' + 'DB16260.pdbqt' + ' --nrun 100 -N ' + './docking_res_ocl/' + 'DB16260.pdbqt' + '_docking_res --gbest 1 > ./docking_res_ocl/' + 'DB16260.pdbqt' + '_docking_res.log 2>&1' )

The previous python scripts were modified from a working code on a PC with only one RTX 2080Ti.

What I can do?

Thanks.

Saverio

[atillack]

@xavgit The Cuda runtime error should get resolved compiling with TARGETS="80" (plus other desired compute capabilities if there are other architectures). The OpenCL error you are seeing usually means the OpenCL platform isn't registered (installed) on the system.

[xavgit]

Hi,
all is fine now with your help.

Thanks.

Saverio