GABIND: a GA approach to allocation and binding for the high-level synthesis of data paths

Abstract

We present here a technique for allocation and binding for data path synthesis (DPS) using a Genetic Algorithm (GA) approach. This GA uses an unconventional crossover mechanism relying on a force directed data path binding completion algorithm. The data path is synthesized using some supplied design parameters. A bus-based interconnection scheme, use of multi-port memories, and provision for multicycling and pipelining are the main features of this system. The method presented here has been applied to standard benchmark examples and the results obtained are promising. Index Terms-Allocation, binding, data path synthesis. I. INTRODUCTION Data path synthesis (DPS) involves scheduling of operations followed by allocation and binding. The latter step consists of several sub-tasks which include determining the mix of functional units (FUs) grouping variables and assigning these variable clusters to storage units, memory port assignment when multi-port memories are used in the design, mapping operations to the FUs, and mapping transfers to buses, when buses are used. The present work is concerned with the allocation and binding aspects of DPS. Earlier techniques attempted to solve these sub-tasks independently. Contemporary techniques attempt to handle the sub-tasks and other tasks such as scheduling in larger groups to ensure better optimality of the final design. However, all these sub-tasks are known to be NP-Complete and amalgamating many of these as a single problem is computationally prohibitive. It is therefore desirable to be able to solve the sub-tasks in overlapping combinations and move from one set of sub-tasks to another with a set of solutions rather than just a single solution. Multiple heuristics and randomization schemes may be used to find multiple solutions that are of the same cost or are nondominating. These concerns have motivated us to develop a Genetic Algorithm (GA), called GABIND, for synthesizing optimized data paths from a given scheduled data flow graph. GABIND builds on previously developed successful heuristics, such as force [1], by incorporating them into the GA. Several researchers have worked on the DPS problems and several systems such as HAL [1], SAST [2], Facet [3], STAR [4], SAM [5], and Vital-NS [6] have been developed to solve the problem. While all techniques attempt to optimize schedule time and cost of storage and FUs, current techniques place an emphasis on interconnect optimization. HAL, SAM, Vital-NS, and STAR are some of the systems that perform interconnect optimization, in addition to the other DPS related optimizations. HAL was the first to make use of the force-directed algorithm to perform scheduling and data path optimizations. SAM combines scheduling, allocation, and mapping in a single algorithm. The algorithm uses the notion of force [1] to measure the effect that a tentative scheduling of an operation would have on the resource requirements. VITAL-NS performs scheduling, allocation and binding sub-tasks of DPS. FU registers and buses are partially allocated Manu...

Speeding-Up Expensive Evaluations in High-Level Synthesis Using Solution Modeling and Fitness Inheritance

“…In addition, to compute the fitness of the evolved solution we replace the usual expensive evaluation process with a cost model coupled with an inheritance fitness scheme. In particular, our approach extends previous works on the application of EAs to HLS [14,21,24,25] basically in three respects: (i) while in previous works focused on evolutionary approaches to optimize a human designed objective function, we exploit NSGA-II [9], a multi-objective genetic algorithm, to perform a fully automated design space exploration; (ii) we exploit a regression model to perform a fast and quite accurate evaluation of the candidate solutions; (iii) to our knowledge, this is the first work that applied a fitness inheritance scheme to HLS in order to reduce the number of evaluations. We validated our approach on several benchmark problems.…”

“…To support scalability and to explore a larger set of alternative designs, several non-deterministic approaches (e.g., [20]), and in particular GAs [12,14,21,24,25], have been efficiently applied to HLS. Most of them focused on only one of the HLS sub-task.…”

“…Most of them focused on only one of the HLS sub-task. In [24], GAs are used to schedule the operation, while in [25] they are used to allocate and bind a scheduled graph. Grewal et al [14] implemented a hierarchical genetic algorithm, where genetic module allocation is followed by a genetic scheduling.…”

“…However, they performed a single-objective optimization using a weighted average of the design objectives, that has been proved to be not effective [39]. Several works [12,25] introduced a binding-based encoding, also for system-level synthesis [27,36], where solutions are represented as the binding between operations and functional units where they will be executed. In some of these approaches the exploration can generate unfeasible solutions that have to be recovered or discarded, wasting time and computation resources.…”

A Novel Framework for Applying Multiobjective GA and PSO Based Approaches for Simultaneous Area, Delay, and Power Optimization in High Level Synthesis of Datapaths

“…The authors have used two different encoding schemes. The priority based scheme [2] arranges nodes in the DFG in the order in which they have to be scheduled by a list scheduler whereas the binding based scheme [23] incorporates binding information pertaining to each DFG node. The area and delay costs are extracted from a regression model built using actual characterizations.…”