JACC: Leveraging Performance Portability with the Just-in-Time Julia Language

Tentative Abstract: We present JACC as a performance-portable metaprogramming model for the just-in-time Julia language. JACC. provides a unified and lightweight front end across different back ends available in Julia (Base.Threads, CUDA, AMDGPU, and OneAPI) to enable the same Julia code to run on different CPU and GPU targets. We evaluated the performance of JACC for common HPC kernels (e.g., AXPY, DOT, Conjugate Gradient algorithm, lattice-Boltzmann method) on modern HPC architectures found in some of the most advanced supercomputers today, including Frontier, Aurora, and Perlmutter equipped with a variety of architectures from different vendors-including AMD EPYC 7742 Rome CPUs, AMD Mi100 GPUs, NVIDIA A100 GPUs, and Intel Max 1550 GPUs. We show that the proposed programming model incurs only a negligible overhead versus Julia's vendor-specific solutions. We report speedups for the GPU implementations over the CPU implementations with no extra cost to programmability for the kernel granularity.