a Memory Access Network Architecture for Application Specific Hardware
Tuesday April 19, 2005
Hamerschlag Hall A-306
Carnegie Mellon University
High-level synthesis (HLS) tools have shown great potential for
generating high-quality circuits in a timely fashion by bridging
the semantic gap between high-level abstractions and gate-level
implementations. However, a major obstacle to expanding the domain
space of synthesizable applications is the presence of arbitrary
memory reference patterns in the source specification. For example,
pointer aliasing prevents many tools from expanding the subset of
synthesizable ANSI-C. In this talk, I will present SOMA, a framework
for Synthesizing and Optimizing Memory Access in high-level abstractions.
At the core of this framework is a Memory Access Network (MAN) architecture
that inherently enforces dynamic ordering-dependencies between memory
accesses. The design space of MAN architectures for a given application
is large and there are many axes along which the architecture can
be optimized. I will discuss our experience with a couple of MAN
optimizations that we have performed - one, reducing the synchronization
overhead in dynamic memory dependency checking; and two, improving
the average throughput and latency of all accesses in the application
through dynamic memory-parallelism analysis. SOMA has been fully
integrated into the CASH toolflow to automatically generate gate-level
structural Verilog from C programs featuring arbitrary pointers.
Girish received his MS in Computer Science from University of
California Riverside, and he is currently a 3rd year Phd student
in the ECE department at CMU. He is part of the Phoenix reconfigurable
computing group led by Seth Goldstein (who is also his advisor).
The Phoenix group is researching on a C-to-hardware toolflow, and
his focus is on synthesizing (and optimizing) asynchronous circuits
from the compiler IR. He also dabbles in related fields like reconfigurable
computing, computer architecture, compiler design and circuit synthesis.