Quantitative MRI (qMRI) has great potential to transform clinical radiology by offering higher-quality medical images while reducing acquisition times. This enables faster diagnoses by radiologists and shorter scanning times for patients. However, the computational demands of qMRI algorithms are significant, often causing image reconstruction to take hours and thus hindering clinical adoption.

We present COMPAS, a composable toolkit of high-performance qMRI primitives for developing state-of-the-art qMRI methods. COMPAS hides the technical complexity required to achieve near-real-time performance while providing an easy-to-use interface for both C++ and Julia.

COMPAS integrates several cutting-edge technologies, including work developed at the Netherlands eScience Center. We use Kernel Tuner to auto-tune the performance of individual GPU kernels. We develop KMM, a parallel dataflow and memory-manager layer for multi-GPU systems that minimizes data transfers, reuses GPU allocations, and overlaps computation with communication. We also perform selected operations in low precision to increase performance at the cost of a minimal loss in numerical accuracy. Finally, by targeting both CUDA and HIP, we support AMD and NVIDIA GPUs with a single codebase.

We present results using Snellius (NVIDIA H100) and LUMI (AMD MI250X) supercomputers, reducing reconstruction times from hours to nearly one minute, making qMRI ready for potential use in clinical trials.