Proposal for research projects using the BEE2 board:
(MESCAL research group, Prof. Kurt Keutzer, UC Berkeley)


MESCAL research goals: 

The aim of the MESCAL project is to advance a disciplined approach for
deploying concurrent applications on programmable multiprocessor
platforms.  Our solution is to provide a domain specific programming
model and a software development framework that allows efficient and
productive implementations of applications on to a family of fully
programmable multiprocessor microarchitectures.

We previously deployed representative network and media applications
on FPGA-based soft multiprocessor systems using the Microblaze/PowerPC
cores and the FSL/OPB communication links on the Xilinx ML-310 and XUP
boards.  However, we were limited to implementations containing 5-10
processors due to the compute area and memory constraints on these
boards.  The BEE2 board is a viable target for building larger, more
diverse soft multiprocessor systems to showcase realistic compute and
memory intensive applications.  In particular, we intend to
investigate the following two applications and deploy them on soft
multiprocessors on the BEE2 board.


Project 1: Multiprocessor implementation of statistical machine
learning algorithms for image recognition

Image recognition has been identified as a key compute intensive and
highly parallel emerging application under Intel's RMS taxonomy.
Recognition draws on a class of statistical machine learning
algorithms to classify objects or events of interest to the user.  The
objective of this project is to deploy component classification of
natural sceneries on a soft multiprocessor system on the BEE2.  We
intend to evaluate task and data level parallelism and memory data
layout for this class of learning algorithms.


Project 2: A parallel implementation of H.264 decoding on soft
multiprocessors

H.264 is an advanced video codec with increasingly control centric
routines and tight feedback dependencies that is representative of the
trend in video applications.  It poses significantly more
parallelization challenges than the highly data centric multimedia
applications that we previously implemented.  Parallelizing H.264
decoding for soft multiprocessors will provide insights into task and
data level parallelism and memory data layout for future multimedia
codecs.  Based on this experience, we intend to develop a design space
exploration framework for multimedia applications from a
MATLAB/Simulink domain specific language onto parallel platforms.