Yin-Yang: Programming Abstractions for Cross-Domain Multi-Acceleration.

Area of Science:

• Computer Engineering
• Hardware Acceleration
• Software Engineering

Background:

• Field-Programmable Gate Arrays (FPGAs) provide performance and efficiency through domain-specific acceleration.
• Real-world applications frequently span multiple algorithmic domains, necessitating cross-domain acceleration.
• Current FPGA accelerator architectures lack interoperability between different specialized domains.

Purpose of the Study:

• To introduce a novel framework enabling cross-domain multi-acceleration on FPGAs.
• To address the limitations of vertically-specialized FPGA accelerator stacks.
• To facilitate the development of complex applications utilizing multiple heterogeneous FPGA accelerators.

Main Methods:

• Proposal of dual abstractions: Yin for algorithmic specification and Yang for accelerator capability capture.
• Development of the XLVM (eXpressive Low-latency Virtual Machine) dataflow virtual machine.
• Transparent mapping of domain functions to suitable accelerator capabilities via XLVM.

Main Results:

• Demonstrated effectiveness using six real-world cross-domain applications.
• Achieved a 29.4x speedup with the proposed Yin-Yang framework.
• Outperformed the best single-domain acceleration, which yielded a 12.0x speedup.

Conclusions:

• The Yin-Yang abstractions and XLVM effectively enable cross-domain multi-acceleration on FPGAs.
• This approach significantly enhances performance for complex, multi-domain applications.
• It overcomes the limitations of existing, domain-specific FPGA accelerator designs.