Evaluating high-level programming models for FPGA platforms

Research topic and goals

The end of the Moore’s law poses a significant challenge for scientific computing: from the mid 2020s, performance will not improve any more from the CMOS technology progress. Reconfigurable computing presents the unique opportunity of allowing performance progress by customization while still serving a large variety of applications, offering a true co-design vehicle. However, its adoption in scientific computing still faces the lack of high-level parallel programming abstraction and the extreme difficulty of achieving high performance with existing compilation stacks.

We focus on node-level parallelism; and our objective is to demonstrate, for a set of JLESC applications accelerated with emerging high-level synthesis technology, significant performance/watt improvement compared with CPUs and GPUs of the comparable technology.

The advent of high-level synthesis technology such as OpenCL, OpenMP and OpenACC for FPGAs eases the FPGA adoption process in HPC. However, we have little expertise to port current existing computational kernel and have little knowledge on optimizations. In this project, we will implement key application kernels using high level synthesis tools, study optimization techniques and conduct performance evaluations. For optimization, we will evaluate existing source-to-source optimizers such as Merlin compiler and understand its pros and cons. We simultaneously study manual optimizations, that are algorithm-aware, to accelerate kernels significantly.

Results for 2015/2016

We organized an international workshop on FPGA for scientific simulation and data analytics, at Argonne on January 2016, that attracted application developers, researchers in computer science and applied mathematics, the main FPGA vendors (Xilinx and Altera), and HPC application users.

Results for 2016/2017

We gave presentations on our FPGA work at the H2RC workshop and the post-Moore workshop in SC16. We organized the second international workshop on FPGA for scientific simulation and data analytics at NCSA in Urbana, Illinois.

Visits and meetings

Franck Cappello (ANL) visited BSC for two weeks in July 2016 and spent two days for this project.

Impact and publications

Franck Cappello (ANL), Internal report on HPC trends: focus on Reconfigurable Architecture. ANL technical report. As part of the EDF contract for the JLESC. The team is submitting a proposal to DOE on this topic.

    Future plans

    We will study performance portability and code optimization techniques on selected application kernels using the latest generation FPGA platforms such as Intel Arria10 and possibly submit results to relevant venues. We will also evaluate the Merlin compiler, the best commercial compiler for FPGA accelerators, for a comparison purpose.

    References