Repository navigation
🧠 TL;DR
- ✅ Expanded IRON programming capabilities, ObjectFIFO fusion, and dynamic reconfiguration
- ✅ New YOLO26n NPU2 implementation, Llama 3.2 1B improvements, and attention kernels
- ✅ Faster compilation, lower host dispatch overhead, and improved kernel performance
- ✅ Expanded HSA packaging, CERT runtime sequencing, and tracing support
- ✅ Stronger DMA, routing, ObjectFIFO verification, and developer tooling
🔍 Highlights
🚀 IRON & Dynamic Runtime
- Added
ObjectFifoPool, support for multiple device configurations, and improved placement and runtime merging for on-chip operator fusion → @andrej, @hunhoffe - Expanded dynamic reconfiguration with a new programming example and benchmark, full-ELF initialization/teardown, and per-call DMA repeat counts → @andrej, @jskimko, @hunhoffe
- Added explicit route pinning via
aie.flow'sviaattribute and improved packet-flow routing to avoid arbiter deadlocks → @andrej, @hunhoffe - Improved ObjectFIFO lowering, runtime DMA task configuration, and dynamic-shape programming examples → @jgmelber, @atassis
⚡ Kernels, Models & Performance
- Implemented the full YOLO26n detector in a single NPU2 configuration, using a dataflow programming approach → @hunhoffe
- Improved Llama 3.2 1B A8W8 decode and prefill, and expanded attention kernels for AIE2/AIE2P → @hunhoffe
- Added and optimized softmax, argmax, attention, vision, and quantized kernels, alongside expanded AIE2P vector lowering → @hunhoffe, @atassis, @erwei-xilinx
- Reduced host dispatch overhead in
CachedXRTRuntimeand improved compilation time and peak memory usage for large fused designs → @hunhoffe
⚙️ Compiler, Runtime & Correctness
- Improved core stack sizing using compiler measurements and added support for preserving MLIR source locations from IRON/Python through
aieccfor better diagnostics → @hunhoffe - Expanded CERT runtime support with deterministic multi-microcontroller sequencing, global barriers, and DMA chain generation → @fifield
- Improved HSACO packaging with full-ELF kernel selection by name and AIE4 architecture support → @ypapadop-amd
- Strengthened ObjectFIFO, DMA, and routing verification, including detection of stalled loops, unsafe BD reuse, incomplete transfers, and routing deadlocks → @hunhoffe
- Expanded AIEVec lowering and numerical correctness for AIE2P, including shuffle, elementwise, transcendental, and contraction operations → @erwei-xilinx
🛠️ Developer Experience & CI
- Improved IRON Python API typing, compiler diagnostics, source locations, and standalone programming examples → @hunhoffe, @saadafsy
- Added a maintainer dashboard and expanded nightly kernel, routing, and compiler reliability checks → @hunhoffe
- Improved Windows CI, on-device test retries, Peano-only builds, and build-system reliability → @jgmelber, @andrej, @atassis, @hunhoffe
- Continued dependency updates, security hardening, and documentation cleanup → multiple contributors
⚠️ Migration Notes
- Compiler verification is stricter. Invalid ObjectFIFO, DMA, and routing configurations that previously compiled may now produce diagnostics.
- IRON Python APIs continue to evolve. Downstream users relying on internal APIs may need to accommodate ongoing changes to runtime, kernel, and ObjectFIFO abstractions.
🔭 Looking Ahead
Several programming-model areas remain under active development:
- Continued DMA lowering and verification consolidation may introduce changes or deprecation around
npu.dma_memcpy_nd. - ObjectFIFO APIs and lowering will continue evolving as restructuring and operator-fusion support progress.
- IRON Python APIs will continue expanding around dynamic runtime sequences, reconfiguration, and host-driven execution.
🙌 New Contributors
Thank you to everyone who contributed to the v1.4.5 release!
Full Changelog: v1.4.4...v1.4.5