Optimizing the LiGen Drug Discovery Pipeline for Intel Max GPUs

High-throughput virtual screening is a fundamental technique in modern drug discovery, enabling the identification of promising drug candidates by evaluating millions of ligand-protein interactions in silico. Achieving high performance and scalability in such workflows requires efficient exploitation of parallel architectures. LiGen is a high-performance virtual screening application designed to accelerate molecular docking and scoring computations. Initially implemented in CUDA to fully exploit NVIDIA GPUs, LiGen has been ported to SYCL to extend its support for heterogeneous architectures. In this work, we enhance the LiGen SYCL codebase through a performance-portable implementation to maximize the ligand throughput based on the new SYCL features introduced in oneAPI. Our optimized implementation leverages portable features that abstract the underlying hardware resources, enabling dynamic adaptation of the number of ligands processed per kernel to the characteristics of the target device. We further extended our implementation with architecture-specific optimizations for Intel GPUs, focusing on sub-group size tuning and efficient General Register File (GRF) utilization to maximize ligand throughput. Our experimental evaluation compares our optimized SYCL implementations against the manually-tuned CUDA and SYCL baselines. Results show that our version achieves up to 1.69x throughput compared to the SYCL baseline without requiring manual tuning, while the version with Intel-specific optimizations achieves a throughput of up to 2.42x.