Exploiting run-time reconfiguration in stencil computation

Abstract

Stencil computation is computationally intensive and required by many applications. This paper proposes an approach to exploit run-time reconfigurability of field-programmable accelerators for stencil computation. System throughput is optimized by partitioning, analysing and scheduling tasks in applications to remove idle functions. To evaluate the proposed approach, Reverse Time Migration (RTM), a high performance application, is developed. Our optimized runtime reconfigurable solution, which targets a Virtex-6 FPGA in a Maxeler MAX3424A system, can achieves an improved throughput of 102.8 GFlop/s, up to two orders of magnitude faster than the CPU reference designs, 1.59 times faster than the best published GPU and FPGA results, and 1.45 times faster than an optimized static implementation.

Elastic Management of Reconfigurable Accelerators

“…In this approach, we would allow dynamic designs to cooperate and willingly give over resources (if requested by the scheduler) after achieving one "unit of work" which is entirely algorithm dependent. For example in an iterative seismic imaging algorithm (such as Reverse Time Migration [22]) the unit of work may be consider one iteration of the outer loop. In case of sorting this could be sorting one subset of the entire input array and so on.…”

Knowledge Transfer in Automatic Optimisation of Reconfigurable Designs

“…Two benchmarks are used: a quadrature-based financial design with customizable precision [7] and a high performance Reverse Time Migration (RTM) design with seven parameters [8]. Each design is first optimised for the Maxeler MPC-X1000 system with Xilinx Virtex-6 XC6VSX475T FPGA.…”

“…This is especially prominent when rms = 0.1 or rms = 0.01 as transfer knowledge does not offer as big an improvement. 5) Reverse Time Migration: The RTM design [8] is a stencilbased design used for seismic imaging and it involves seven parameters. The parameters determine parallelism, numerical precision and balancing of computation and communication through changing of the design architecture.…”

“…The Auto-Transfer algorithm is based on Bayesian optimisation [3]. (Section IV) • Evaluation of the Auto-Transfer algorithm using two case studies: a quadrature design for financial computation [7] and a seismic imaging design [8]. (Section V) II.…”

“…This is followed by design parameterization, along with constraint and optimisation goal specifications. Lastly, either a set of analytical models are constructed and the design is manually optimised [7], [8] or an automatic tool is used [1], [2], [3]. The manual approach the has advantage of being tailored to a particular problem and making use of designer's experience, while a tool has the advantage of being generic and automated.…”