Memory-centric customization toward FPGA accelerators

For realizing a human-centered super smart society called Society 5.0, it is very important to attain highperformance and high-efficiency computing through custom computing systems specialized for particular application domains. On the other hand, customization of systems incurs challenges in productivity and costs for the hardware and software development process. To overcome these difficulty, we investigate the ways to detect locality of processing inherent in application programs and develop high-level optimization frameworks that automatically map it to FPGA accelerators.

Scientific modeling of system performance and quality of the performance

We investigate the ways to profile, analyze and predict behaviors of program execution to understand the software execution performance. Here, we develop profilers, performance models, and simulators of hierarchical memories.

Automation of customization and co-design driven by mathematical optimization and

machine learning techniques We investigate the ways to automate co-design processes for domain-specific customization. These are implemented upon the current software stack such as compilers, code translators, and optimizers. These are also strongly tackled with mathematical optimization, machine learning, and deep learning techniques.