b8911
Breaking Changes📦 llama-cppView on GitHub →
⚠ 2 breaking✨ 6 features🐛 5 fixes🔧 6 symbols
Summary
This release significantly enhances Hexagon backend capabilities by adding comprehensive Op profiling support, including PMU integration and asynchronous operation handling. Several internal naming conventions were updated, such as renaming 'opmask' to 'opstage' and removing 'opsync'.
⚠️ Breaking Changes
- The name of the opmask configuration was changed to opstage. If you were using 'opmask', you must now use 'opstage'.
- The 'opsync' mechanism is removed because the profiler is now fully asynchronous. Code relying on 'opsync' will need refactoring.
Migration Steps
- Rename any usage of 'opmask' to 'opstage'.
- Remove any reliance on the 'opsync' mechanism due to its removal.
✨ New Features
- Added basic and extended Op profiling support for Hexagon.
- Enabled PMU for hex-prof.
- Added support for reading profiling logs from stdin in hex-prof.
- Added a simple profile post-processing tool.
- Made profiler mode configurable, allowing disabling expensive PMU counter reads on older devices.
- Added support for setting profiler PMU events via environment variables.
🐛 Bug Fixes
- Restored HTP_OPMASK_QUEUE functionality.
- Ensured hmx-matmul honors OPMASK_SKIP_COMPUTE.
- Simplified profiler output by removing unnecessary buffer prints.
- Simplified PMU counter formatting.
- Fixed linter warnings and errors in hex-prof.