ROUGE-V is an open-source AOT compiler designed to translate PTX code generated by the NVIDIA nvcc 13.4 compiler into LLVM IR. The project follows a clean-room development approach, avoiding the use of proprietary NVIDIA binaries or headers. It supports compilation for x86-64, AMD, and RISC-V architectures.
The tool chain involves four steps: compiling .cu source code to PTX via nvcc, parsing PTX into LLVM IR using the ptx2ir translator, assembling the final binary with clang, and executing it on hardware. The compiler has been tested on a GeForce MX450, achieving bit-exact results compared to host references.
Development challenges included grammar discrepancies in modern PTX output, such as single-line entry signatures and specific instruction handling for arithmetic and predicates. The author implemented fixes for address-space bugs and data corruption issues identified during testing on real GPU hardware.
The project currently supports seven canonical kernels, including vadd and gemm_tile, and provides a mechanism for cross-compilation to AMD and RISC-V. The author emphasizes that the project focuses on bit-identity rather than performance metrics and is licensed under Apache 2.0 with LLVM-exception.
