F2Asm Learns Exact NVIDIA SASS Encoders for GPU Rewriting

NVIDIA provides disassemblers for its SASS (Streaming Assembler) instruction set but does not publicly offer an assembler for recent data-center GPUs. This absence restricts the ability to perform controlled machine-code rewriting, which is crucial for advanced GPU programming and security analysis. To address this, a new system called F2Asm has been developed. F2Asm is designed to learn precise 128-bit SASS encoders. It achieves this by analyzing paired disassembly and original CUBIN instruction words, employing techniques from F2 linear algebra, specifically Gaussian elimination over F2. This method allows F2Asm to incrementally build a compact basis for the instruction set, detect inconsistencies, and reject invalid inputs. The system also intelligently separates target-specific control bits, relocation rules, and CUBIN metadata from its core learning algorithm. The researchers successfully trained encoders for various NVIDIA architectures, including Hopper SM90/SM90a, Blackwell SM100, and Rubin SM107, using a large corpus of CUBINs from production libraries and CUDA packages. In rigorous round-trip tests, F2Asm was able to reassemble disassembled SASS code, with all compared executable text sections matching the originals exactly. This marks F2Asm as the first open-source NVIDIA SASS assembler to support the Rubin SM107 architecture.