v7.0.2

What's Changed

• Fix for non-constexpr warpSize in ROCm 7.0+

Full Changelog: v7.0.1...v7.0.2

v7.0.1

What's Changed

• Fix partial panel bugs and roctx markers by @noelchalmers in #17

Full Changelog: v7.0.0...v7.0.1

v7.0.0

What's Changed

• GPU-side panel factorization
  • This is a significant change to the usual overlapped CPU+GPU code path.
  • New optimal run parameters may be required on some systems.
• As GPU panel factorization uses the Cooperative Groups API, oversubscription of GPUs is no longer supported
  • Users must ensure to run with at most one process per GPU.
• CPU-GPU binding in run scripts changed to use NUMA affinity
• Added Debian and Linuxmint into the distribution filter by @joelandman in #11

New Contributors

• @joelandman made their first contribution in #11

Full Changelog: v6.1.0...v7.0.0

v6.1.0

What's Changed

• New iteration count input parameter to allow for repeated back-to-back runs
• Support for ROCm 6.0+
• Updated BLIS, UCX, and OpenMPI third party libs built by default.
• Support for OpenMPI v5.0+
• Auto-detect supported gfx archs via ROCm version
• Support for up to 512 CPU cores on single node
• Flux job manager support
• Added toggle for enabling roctx tracing

Full Changelog: v6.0.0...v6.1.0

v6.0.0

Initial public release