amd-oss-knowledge

The AMD open-source NPU chain, documented and verified — a curated knowledge base distilled from the local research workspace ~/amd-oss (AMD research repos cloned and analyzed — the NPU chain + Xilinx/AMD toolchain mirrors, Aug/Sep 2026 snapshot).

Entry point: INDEX.md.

What this is

The complete, bit-exact-verified pipeline from a PyTorch model to hardware:

Brevitas → QONNX → FINN → RTL/HLS → chip/RTL simulation → stitched FPGA

plus the Ryzen AI NPU software-stack mapping (MLIR-AIE / IRON / Triton-XDNA / FastFlowLM) and the measured receipts behind the 1bit-MONSTER HRX2 llama.cpp work (zero-copy decode win, prefill regression debate).

Repository layout

Path	Contents
INDEX.md	Curated knowledge index (start here)
DESCENT.md	The Full Descent — the verified chain, hop by hop
ANALYSIS.md	Consolidated technical analysis (master table + NPU stack)
structure.txt	Per-repo structural pass
gh_meta.txt	Repo metadata snapshot
docs/fastflowlm/	FastFlowLM docs (AMD's NPU-native runtime, built on IRON)
docs/hrx-loom/	Loom Programming Guide (ROCm/hrx-system, mirrors rocm.github.io/hrx-system/loom/)
docs/mlir-air/	MLIR-AIR published docs (Xilinx/mlir-air, mirrors xilinx.github.io/mlir-air/dev/)

Canonical sources

• The local workspace ~/amd-oss/ (this repo is a published snapshot of its docs; DESCENT.md/ANALYSIS.md are authored there).
• The 1bit-MONSTER HRX2 research log lives in the 1bit-MONSTER repo under research/ws12-hrx-loom/ (round-by-round measured numbers) and docs/research/ (hybrid-prefill-decode, IOMMU PerfOpt).
• FastFlowLM docs under docs/fastflowlm/ are copied from ROCm/FastFlowLM (Apache-2.0).
• Loom guide under docs/hrx-loom/ is mirrored from ROCm/hrx-system loom/docs/src (publishes https://rocm.github.io/hrx-system/loom/).
• MLIR-AIR docs under docs/mlir-air/ are mirrored from Xilinx/mlir-air docs/ published pages (publishes https://xilinx.github.io/mlir-air/dev/).

Refresh

DESCENT.md/ANALYSIS.md are regenerated in ~/amd-oss/; re-copy them here and push to refresh this snapshot.