CHARD

About Technology

Code restructuring solution for efficient FPGA implementations.

CHARD is a software that automatically restructures code, transforming C code into a dataflow graph (DFG) and applies graph optimizations to generate C code with high-level synthesis (HLS) directives for FPGA-based hardware accelerators.

Inventors

Afonso Soares Ferreira

João Paiva Cardoso

Licensing Contacts

Daniel Marques Vasconcelos

Problem

Programming FPGAs for hardware acceleration is a complex and time-consuming task that requires specific skills and tools that are different from typical software development.

In addition, current HLS tools allow software developers to use C-like languages to program FPGA, still require hardware expertise and manual code transformations and directives to generate efficient hardware implementations. This affects productivity, performance, and energy efficiency of FPGA-based systems, especially for applications that involve data-intensive and parallel computations, such as image processing, machine learning, and DSP algorithms.

Solution

Code restructuring solution for efficient FPGA implementations.

Our solution is CHARD, a software that automatically restructures code, transforming C code into a dataflow graph (DFG) and applies graph optimizations to generate C code with high-level synthesis (HLS) directives for FPGA-based hardware accelerators. The technology aims to enable software developers to program FPGAs more easily and efficiently, without requiring deep knowledge of hardware design or HLS tools.

The technology is based on a flexible frontend that can handle different input languages and a backend that implements various graph transformations and optimizations, such as folding, unfolding, pipelining, arithmetic optimizations, array partitioning, and data reuse. CHARD is implemented in a framework that can automatically generate HLS-friendly C code from execution traces of the input critical functions, achieving speedups of 30-100 times (vs original unmodified C code) and 2-15 faster (vs C optimized with Vivado HLS directives).