Modern High-Level Synthesis for Complex Data Science Applications

Field Programmable Gate Arrays (FPGAs) have become one of the key accelerators for data analytics and machine learning in Data Center, and they are seeing increased adoption in heterogeneous high-performance computing sys-tems for complex workflows that couple scientific simulation with data science. FPGAs have traditionally been pro-grammed with hardware description languages, requiring significant engineering efforts and long development times. Today, the availability of new high-level synthesis (HLS) tools to generate accelerators starting from high-level specifi-cations provides easier access to FPGAs and preserves programmer productivity. However, the conventional HLS flow typically starts from languages such as C, C++, or OpenCL, heavily annotated to provide information for the hardware generation, still leaving a significant gap with respect to the (Python based) data science frameworks. This tutorial will discuss HLS to accelerate data science on FPGAs, highlighting key methodologies, trends, ad-vantages, benefits, but also gaps that still need to be closed. The tutorial will provide a hands-on experience of the SOftware Defined Accelerators (SODA) Synthesizer, a toolchain composed of SODA-OPT, an opensource front-end and optimizer that interface with productive programming data science frameworks in Python, and Bambu, the most advanced open-source HLS tool available, able to generate optimized accelerators for data-intensive kernels.

---

Contributors:

• Nicolas Bohm Agostini (Pacific Northwest National Laboratory)
• Serena Curzel (Pacific Northwest National Laboratory)
• Michele Fiorito (Politecnico di Milano)
• Vito Giovanni Castellana (Pacific Northwest National Laboratory)
• Marco Minutoli (Pacific Northwest National Lab)
• Fabrizio Ferrandi (Politecnico di Milano)
• Antonino Tumeo (Pacific Northwest National Laboratory)