BarraCUDA Open-source CUDA compiler targeting AMD GPUs (github.com)

Article summary

BarraCUDA is an open-source CUDA, HIP, and Triton compiler that targets multiple GPU and CPU architectures, allowing for the compilation of CUDA C, HIP, or Triton source code into AMD RDNA 2/3/4 binaries, NVIDIA PTX, or native x86-64 code. The compiler is written in C99 and has no dependencies, including LLVM. It is designed to be a lightweight and flexible alternative to existing compilers. The project is named after Kathleen Booth, a pioneer in computer science.

Main themes

• GPU Computing
• Open-Source Compilers
• CUDA Support
• AMD vs NVIDIA
• Compiler Development
• GPU Architecture

What commenters say

• The lack of CUDA support on AMD is a strategic decision rather than a technical limitation, as supporting CUDA would only benefit NVIDIA.
• The project's decision to use C99 and avoid dependencies like LLVM is seen as a strength, allowing for a more lightweight and flexible compiler.
• Some commentators argue that ZLUDA is a more practical choice for replacing CUDA, despite BarraCUDA's impressive technical achievements.
• Others believe that AMD has given up on competing with NVIDIA in the GPU market, citing the company's lack of support for certain features and technologies.
• There is disagreement over the relevance of CUDA, with some arguing it is still essential and others claiming it is becoming less important.
• The project's custom instruction encoding without LLVM is seen as a notable achievement, but the C subset limitation is a significant drawback for many potential users.
• Some commentators think that AMD's focus on APUs and integrated graphics is a viable strategy, while others believe the company needs to improve its discrete GPU offerings to compete with NVIDIA.
• The discussion also touches on the challenges of developing and maintaining open-source projects, including the risk of being acquired or shut down by larger companies.