Department of Electronic and Computer Engineering Seminar - Accelerator Programming for Data Specialization

In today's computing landscape, applications are highly data-intensive, while hardware is becoming ever more heterogeneous. Achieving high performance and efficiency requires tailoring data formats, placements, and numeric types to suit both applications and hardware. However, current programming models and compilers lack adequate support for data specialization, presenting a daunting challenge that remains out of reach for most programmers to address.

This talk discusses our ongoing research on evolvable data specialization, employing a co-design methodology spanning programming models, compilers, and hardware acceleration. I'll introduce UniSparse (OOPSLA'24), an intermediate language offering a unified abstraction for customizing sparse matrix/tensor formats. Unlike existing frameworks, UniSparse separates logical representation from low-level memory layout, facilitating concise format customizations through well-defined primitives. We validate our approach through experiments on various hardware targets, including an Intel CPU, an NVIDIA GPU, an AMD FPGA, and a simulated processing-in-memory (PIM) device. Then, I'll cover Allo (PLDI'24), a new programming model and compiler for building efficient dataflow accelerators. Allo decouples specification of custom data types and data placement from algorithm specification, and encapsulates them as a set of composable primitives. Our evaluation shows that Allo can significantly outperform state-of-the-art high-level synthesis tools and accelerator design languages on commonly used benchmarks. Allo also enables productive generation of complex model-specific LLM accelerators on FPGAs, producing low latency and high energy efficiency compared to commercial GPUs.