Summary

SASS King addresses the critical gap in understanding modern NVIDIA GPU SASS (native instruction set) following significant architectural changes. It provides kernel engineers and researchers with a structured knowledge base and reverse-engineering methodology to analyze SASS dumps, identify compiler patterns, and link binary structures to source-level optimizations, enhancing transparency for performance analysis on newer NVIDIA architectures.

How It Works

The project employs a systematic methodology combining controlled micro-kernels with production-kernel audits. By isolating compiler decisions through single-parameter variations, it generates detailed SASS evidence. This data populates a knowledge base and a formal pattern library of reusable audit signatures. The approach uses strict claim tagging ([OBS], [INF], [HYP], [RES], [GAP]) for evidence traceability and prioritizes pattern-based audits over purely bit-level disassembly.

Quick Start & Requirements

• Primary Interaction: A research repository; direct installation commands (pip, Docker) are not specified. Users engage with the corpus, knowledge base, and pattern library.
• Prerequisites: Requires NVIDIA CUDA toolkit (nvcc, cuobjdump), Nsight Compute, and external tools like gpuasm.com and redplait/denvdis. Initial focus is SM120 / SM120a (consumer Blackwell).
• Documentation: Key entry points include docs/START_HERE.md and patterns/README.md.

Highlighted Details

• Phase 3 Pattern Library: Features 29 formalized, reusable SASS signatures with explanations, variants, anti-patterns, and confidence levels for audit building blocks.
• Tensor-Core Focus: Comprehensive studies cover SM120 tensor-core operations (HMMA, QMMA, OMMA) and matrix memory pipelines (LDSM, STSM), including epilogues and sparse metadata.
• Cross-Architecture Expansion: Plans to expand methodology and patterns across multiple NVIDIA architectures (SM80, SM86, SM89, SM90a, SM100a).
• Evidence-Driven: Utilizes strict claim tagging ([OBS], [INF], [HYP], [RES], [GAP]) to ensure technical claims are tied to observed or inferred evidence.

Maintenance & Community

Authored by Florian Mattana. Contributions are welcomed via CONTRIBUTING.md. No specific community channels or sponsorship details are mentioned.

Licensing & Compatibility

The license type is not explicitly stated in the provided README text. Compatibility is focused on NVIDIA GPU architectures.

Limitations & Caveats

SASS ISA coverage is not complete; runtime layout decoding, full control-code bit placement, and cross-architecture replay are identified as future work. The project is actively developing (Phase 4: Production Audits is the next major step), indicating it is not yet a fully mature, stable toolset for all use cases.