Speaker
Description
Maintaining software portability across diverse accelerated hardware architectures remains a critical challenge for the lattice QCD community. The QDP-JIT library addresses this by providing a transparent, drop-in replacement for QDP++, enabling Chroma applications to run with high efficiency on modern GPUs without requiring any modifications to high-level code.
In this work, we present the evolution of QDP-JIT to a modern compiler infrastructure based on the LLVM IR builder API. Leveraging this infrastructure, we have expanded QDP-JIT beyond its initial NVIDIA support. We detail the integration of the AMDGPU backend and report on our recent efforts targeting Intel GPUs via SPIR-V and the Intel Level Zero runtime API.
We demonstrate the production readiness of this cross-platform stack using full Hybrid Monte Carlo (HMC) simulations featuring $2+1$ dynamical flavors of Wilson Clover fermions at near-physical pion masses on the Aurora supercomputer. While the computationally heavy two-flavor solves are offloaded to the SYCL-based QUDA multigrid solver, Amdahl's law dictates that the remaining HMC operations rapidly dominate the overall trajectory time if left on the CPU. We show how QDP-JIT successfully shifts these remaining execution paths to the Intel GPUs to run smoothly alongside QUDA, bringing down the total trajectory time significantly and establishing a highly efficient, portable pipeline across all three major GPU manufacturers.