Experiments AD-GPU DPCPP

This repository contains instructions and corresponding scripts for performing experiments.

Instructions

1. Preparation

Clone this experiments repository:

git clone https://github.com/LeoCollab/experiments-adgpu-dpcpp.git

Clone the AutoDock-GPU code repository, which contains DPCPP as well as CUDA and OpenCL versions:

git clone https://github.com/emascarenhas/AutoDock-GPU.git

2. Compilation

Move into the code folder:

cd AutoDock-GPU

2.1. Compile DPCPP version

Set the required oneAPI environment variables by executing initialization script (NOT required on DevCloud):

source /opt/intel/oneapi/setvars.sh

Compile DPCPP code, and move generated binaries into the experiments folder:

make DEVICE=XeGPU NUMWI=32 && mv bin/autodock_xegpu_32wi bin/autodock_xegpu_32wi_dpcpp && cp bin/autodock_xegpu_32wi_dpcpp ../experiments-adgpu-dpcpp/
make DEVICE=XeGPU NUMWI=64 && mv bin/autodock_xegpu_64wi bin/autodock_xegpu_64wi_dpcpp && cp bin/autodock_xegpu_64wi_dpcpp ../experiments-adgpu-dpcpp/
make DEVICE=XeGPU NUMWI=128 && mv bin/autodock_xegpu_128wi bin/autodock_xegpu_128wi_dpcpp && cp bin/autodock_xegpu_128wi_dpcpp ../experiments-adgpu-dpcpp/
make DEVICE=XeGPU NUMWI=256 && mv bin/autodock_xegpu_256wi bin/autodock_xegpu_256wi_dpcpp && cp bin/autodock_xegpu_256wi_dpcpp ../experiments-adgpu-dpcpp/

2.2. Compile OpenCL and CUDA versions

Switch to git branch holding original OpenCL and CUDA versions (DPCPP code was translated from this equivalent CUDA code):

git checkout v1.5.3

Set environment variables (adapt the paths to your systems' installation):

export GPU_INCLUDE_PATH=/usr/local/cuda/include
export GPU_LIBRARY_PATH=/usr/local/cuda/lib64

Compile codes, rename the produced binaries conveniently (for the experiment scripts to work), and move generated binaries into the experiments folder:

OpenCL

make DEVICE=OCLGPU NUMWI=32 && mv bin/autodock_gpu_32wi bin/autodock_gpu_32wi_ocl && cp bin/autodock_gpu_32wi_ocl ../experiments-adgpu-dpcpp/
make DEVICE=OCLGPU NUMWI=64 && mv bin/autodock_gpu_64wi bin/autodock_gpu_64wi_ocl && cp bin/autodock_gpu_64wi_ocl ../experiments-adgpu-dpcpp/
make DEVICE=OCLGPU NUMWI=128 && mv bin/autodock_gpu_128wi bin/autodock_gpu_128wi_ocl && cp bin/autodock_gpu_128wi_ocl ../experiments-adgpu-dpcpp/
make DEVICE=OCLGPU NUMWI=256 && mv bin/autodock_gpu_256wi bin/autodock_gpu_256wi_ocl && cp bin/autodock_gpu_256wi_ocl ../experiments-adgpu-dpcpp/

CUDA

CUDA code for modern NVIDIA GPU architectures might require to also specify their compute capability via the TARGETS="XX" compile option. See reference.

For the A100 GPU or newer devices (TARGETS="80"):

make DEVICE=GPU NUMWI=32 TARGETS="80" && mv bin/autodock_gpu_32wi bin/autodock_gpu_32wi_cuda && cp bin/autodock_gpu_32wi_cuda ../experiments-adgpu-dpcpp/
make DEVICE=GPU NUMWI=64 TARGETS="80" && mv bin/autodock_gpu_64wi bin/autodock_gpu_64wi_cuda && cp bin/autodock_gpu_64wi_cuda ../experiments-adgpu-dpcpp/
make DEVICE=GPU NUMWI=128 TARGETS="80" && mv bin/autodock_gpu_128wi bin/autodock_gpu_128wi_cuda && cp bin/autodock_gpu_128wi_cuda ../experiments-adgpu-dpcpp/
make DEVICE=GPU NUMWI=256 TARGETS="80" && mv bin/autodock_gpu_256wi bin/autodock_gpu_256wi_cuda && cp bin/autodock_gpu_256wi_cuda ../experiments-adgpu-dpcpp/

For older GPUs (no need for extra options):

make DEVICE=GPU NUMWI=32 && mv bin/autodock_gpu_32wi bin/autodock_gpu_32wi_cuda && cp bin/autodock_gpu_32wi_cuda ../experiments-adgpu-dpcpp/
make DEVICE=GPU NUMWI=64 && mv bin/autodock_gpu_64wi bin/autodock_gpu_64wi_cuda && cp bin/autodock_gpu_64wi_cuda ../experiments-adgpu-dpcpp/
make DEVICE=GPU NUMWI=128 && mv bin/autodock_gpu_128wi bin/autodock_gpu_128wi_cuda && cp bin/autodock_gpu_128wi_cuda ../experiments-adgpu-dpcpp/
make DEVICE=GPU NUMWI=256 && mv bin/autodock_gpu_256wi bin/autodock_gpu_256wi_cuda && cp bin/autodock_gpu_256wi_cuda ../experiments-adgpu-dpcpp/

3. Evaluation

Move into the experiments folder:

cd experiments-adgpu-dpcpp

Clone git submodule repository containing input data-set automatically:

./1_prepare_inputs.sh

Evaluate the performance for different {DPCPP work-group} / {OpenCL work-group} / {CUDA block} sizes:

./2_evaluate_numwi.sh

4. Collection

Collect metrics and store them into xlsx files.

Execute Python script within a virtual environment. See reference.

python3 -mvenv env

source env/bin/activate

python3 parse_adgpu_dlg.py <folder>

deactivate