Work

Synthesis and Exploration of Loop Accelerators for Systems-on-a-Chip

Year: 2011
Type: dissertation
Abstract: The compelling next generation streaming applications containing several computationally intensive nested loop programs are the driving force for the System-on-chip (SoC) architectures. SoC platforms are characterized by the presence of diverse components like traditional processors augmented with programmable or non-programmable accelerators. The accelerators implement computationally intensive loop programs of a given application. An SoC platform can improve the performance, optimize power, and reduce cost by orders of magnitude due to specialized execution on these accelerators. The effort for programming and synthesizing the loop accelerators for streaming applications is still enormous. Therefore, the PARO design methodology is presented in this thesis, which enables automated synthesis and exploration of accelerators for streaming applications. In this dissertation, we present a source-to-source transformation called hierarchical tiling, which kneads the loop descriptions in a high-level language using a polyhedral framework with multiple hierarchies of tiles. This important extension of well known transformation called loop tiling, enables utilization of multiple levels of parallelism and memories in an accelerator architecture. With hierarchical tiling, the system designer is thus able to specify the degree of parallelism (number of PEs), local memory usage, and requisite communication bandwidth for the accelerator architecture. Other design flows contain simple tiling or loop unrolling, which can only specify only one or the other criteria but not all of them. The transformation, however, may create a loop code characterized by the presence of a lot more control conditions, which could become a performance bottleneck. Therefore, a novel back-end methodology for control generation was presented. I/O communication can also become a performance bottleneck. Therefore, depending on the tiling parameters and schedule, a custom memory architecture with multiple banks, address generators, and I/O controllers is automatically generated. The consideration of front-end transformation like hierarchical tiling and novel back-end synthesis of control and I/O communication unit, leads to highly efficient accelerator implementations. We used several loop benchmarks from different class of loop algorithms ranging from linear algebra, image processing, to networking. The synthesized loop accelerators show an average gain of around 2.5x, 4.5x, and 50x in terms of area, power, and performance over embedded processors. Streaming applications are characterized by the presence of multiple communicating loops. With state-of-the-art polyhedral techniques, one can generate fast individual loop accelerators, but there exists no methodology for automated generation of intermediate communication hardware which is often a major bottleneck. In this context, a novel intermediate dependence graph called mapped loop graph was introduced for the modular representation of task level parallelism of communicating loops and their mapping information in the polyhedral model. A methodology is proposed for the projection of a mapped loop graph in the polyhedral model onto model parameters of a data flow model of computation called windowed synchronous data flow (WSDF). Subsequently, the generation of an efficient dedicated communication primitive called multi-dimensional FIFO from the windowed synchronous data flow model is undertaken. The communication synthesis is able to generate communication hardware for data transfer and synchronization between the loop accelerators, which is able to handle parallel access and out-of-order communication in entirety, unlike the existing approaches. In order to ease the integration of accelerators in an SoC, a methodology for automated generation of a memory map, a software driver, and a hardware wrapper was proposed. The hardware/software codesign approach may offer an advantage of 2-20x over pure software solutions. The selection of an optimal architecture can be daunting due to a plethora of architecture and compiler design decisions. Exhaustive exploration of the design space is prohibitive due to absurdly large execution time. Therefore, we propose a method using modern search heuristics based on evolutionary algorithms and estimation of objectives to identify Pareto-optimal designs (i.e., non-dominated designs with best trade-offs) in terms of area cost, power consumption, and performance. This not only reduces the exploration time to a matter of around an hour for relevant accelerator benchmarks, but also delivers better design solutions than random search techniques. For a given workload scenario, a best-fit accelerator can be chosen from the Pareto-optimal set of designs. The proposed analytical method for finding the best-fit accelerator is based on real-time calculus and uses system performance models.
Language: English
Open Access status: green
OpenAlex ID: W2404022787
Cites: 173
Cited by: 5
Related to: 10