BarraCUDA Open-source CUDA compiler targeting AMD GPUs
Open-source tool compiles .cu files directly to AMD machine code with zero LLVM dependency.
Deep Dive
Developer Ethan L. Miller built BarraCUDA, an open-source CUDA compiler written in 15,000 lines of C99. It bypasses NVIDIA's ecosystem by compiling .cu files directly to GFX11 machine code for AMD RDNA 3 GPUs, generating ELF .hsaco binaries. The tool features its own lexer, parser, and 1,700 lines of hand-written instruction selection, supporting core CUDA features like __shared__ memory, atomics, and warp intrinsics without translation layers.
Why It Matters
This could break NVIDIA's CUDA lock-in, letting developers run existing GPU code on AMD hardware without costly rewrites.