Browse Teach Curate

Automatic Level-3 BLAS GPU Offloading on Unified-Memory Architectures

A runtime tool that uses dynamic binary instrumentation to intercept level-3 BLAS calls in an unmodified, already-compiled program and redirect them to the GPU on cache-coherent unified-memory systems, employing a Device First-Use data-placement policy inspired by OpenMP NUMA first-touch to minimize CPU-GPU transfers and deliver large speedups on scientific codes.

Data Text

2501.00279

This paper presents SCILIB-Accel, a tool that automatically moves dense linear-algebra work onto the GPU on unified-memory systems such as NVIDIA Grace-Hopper, where CPU and GPU share cache-coherent …