Memory-Centric Hardware Synthesis from Dataflow Models

Generation of hardware architectures directly from dataflow representations is increasingly being considered as research moves toward system level design methodologies. Creation of networks of IP cores to implement actor functionality is a common approach to the problem, but often the memory sub-systems produced using these techniques are inefficiently utilised. This paper explores some of the issues in terms of memory organisation and accesses when developing systems from these high level representations. Using a template matching design study, challenges such as modelling memory reuse and minimising buffer requirements are examined, yielding results with significantly less memory requirements and costly off-chip memory accesses.

SoC memory hierarchy derivation from dataflow graphs

Hardware synthesis from dataflow graphs of signal processing systems is a growing research area as focus shifts to high level design methodologies. For data intensive systems, dataflow based synthesis can lead to an inefficient usage of memory due to the restrictive nature of synchronous dataflow and its inability to easily model data reuse. This paper explores how dataflow graph changes can be used to drive both the on-chip and off-chip memory organisation and how these memory architectures can be mapped to a hardware implementation. By exploiting the data reuse inherent to many image processing algorithms and by creating memory hierarchies, off-chip memory bandwidth can be reduced by a factor of a thousand from the original dataflow graph level specification of a motion estimation algorithm, with a minimal increase in memory size. This analysis is verified using results gathered from implementation of the motion estimation algorithm on a Xilinx Virtex-4 FPGA, where the delay between the memories and processing elements drops from 14.2 ns down to 1.878 ns through the refinement of the memory architecture. Care must be taken when modeling these algorithms however, as inefficiencies in these models can be easily translated into overuse of hardware resources.

Evolutionary Requirements for Next-Generation Dataflow-Based FPGA System Design

The use of dataflow digital signal processing system modelling
and synthesis techniques has been a fruitful research theme for many years and has yielded many powerful rapid system synthesis and optimisation capabilities. However, recent years have seen the spectrum of languages and techniques splinter in an application specific manner, resulting in an ad-hoc design process which is increasingly dependent on the particular application under development. This poses a major problem for automated toolflows attempting to provide rapid system synthesis for a wide ranges of applications. By analysing a number of dataflow FPGA implementation case studies, this paper shows that despit ethis common traits may be found in current techniques, which fall largely into three classes. Further, it exposes limitations pertaining to their ability to adapt algorith models to implementations for different operating environments and target platforms.

Valved Dataflow For FPGA Memory Hierarchy Synthesis

For modern FPGA, implementation of memory intensive processing applications such as high end image and video processing systems necessitates manual design of complex multilevel memory hierarchies incorporating off-chip DDR and onchip BRAM and LUT RAM. In fact, automated synthesis of multi-level memory hierarchies is an open problem facing high level synthesis technologies for FPGA devices. In this paper we describe the first automated solution to this problem.
By exploiting a novel dataflow application modelling dialect, known as Valved Dataflow, we show for the first time how, not only can such architectures be automatically derived, but also that the resulting implementations support real-time processing for current image processing application standards such as H.264. We demonstrate the viability of this approach by reporting the performance and cost of hierarchies automatically generated for Motion Estimation, Matrix Multiplication and Sobel Edge Detection applications on Virtex-5 FPGA.

A Unified Scheduler for Recursive and Task Dataflow Parallelism

Task dataflow languages simplify the specification of parallel programs by dynamically detecting and enforcing dependencies between tasks. These languages are, however, often restricted to a single level of parallelism. This language design is reflected in the runtime system, where a master thread explicitly generates a task graph and worker threads execute ready tasks and wake-up their dependents. Such an approach is incompatible with state-of-the-art schedulers such as the Cilk scheduler, that minimize the creation of idle tasks (work-first principle) and place all task creation and scheduling off the critical path. This paper proposes an extension to the Cilk scheduler in order to reconcile task dependencies with the work-first principle. We discuss the impact of task dependencies on the properties of the Cilk scheduler. Furthermore, we propose a low-overhead ticket-based technique for dependency tracking and enforcement at the object level. Our scheduler also supports renaming of objects in order to increase task-level parallelism. Renaming is implemented using versioned objects, a new type of hyper object. Experimental evaluation shows that the unified scheduler is as efficient as the Cilk scheduler when tasks have no dependencies. Moreover, the unified scheduler is more efficient than SMPSS, a particular implementation of a task dataflow language.

Analysis of Dependence Tracking Algorithms for Task Dataflow Execution

Processor architectures has taken a turn towards many-core processors, which integrate multiple processing cores on a single chip to increase overall performance, and there are no signs that this trend will stop in the near future. Many-core processors are harder to program than multi-core and single-core processors due to the need of writing parallel or concurrent programs with high degrees of parallelism. Moreover, many-cores have to operate in a mode of strong scaling because of memory bandwidth constraints. In strong scaling increasingly finer-grain parallelism must be extracted in order to keep all processing cores busy.

Task dataflow programming models have a high potential to simplify parallel program- ming because they alleviate the programmer from identifying precisely all inter-task de- pendences when writing programs. Instead, the task dataflow runtime system detects and enforces inter-task dependences during execution based on the description of memory each task accesses. The runtime constructs a task dataflow graph that captures all tasks and their dependences. Tasks are scheduled to execute in parallel taking into account dependences specified in the task graph.

Several papers report important overheads for task dataflow systems, which severely limits the scalability and usability of such systems. In this paper we study efficient schemes to manage task graphs and analyze their scalability. We assume a programming model that supports input, output and in/out annotations on task arguments, as well as commutative in/out and reductions. We analyze the structure of task graphs and identify versions and generations as key concepts for efficient management of task graphs. Then, we present three schemes to manage task graphs building on graph representations, hypergraphs and lists. We also consider a fourth edge-less scheme that synchronizes tasks using integers. Analysis using micro-benchmarks shows that the graph representation is not always scalable and that the edge-less scheme introduces least overhead in nearly all situations.

Dataflow toolset for soft-core processors on FPGA for image processing applications

With security and surveillance, there is an increasing need to be able to process image data efficiently and effectively either at source or in a large data networks. Whilst Field Programmable Gate Arrays have been seen as a key technology for enabling this, they typically use high level and/or hardware description language synthesis approaches; this provides a major disadvantage in terms of the time needed to design or program them and to verify correct operation; it considerably reduces the programmability capability of any technique based on this technology. The work here proposes a different approach of using optimised soft-core processors which can be programmed in software. In particular, the paper proposes a design tool chain for programming such processors that uses the CAL Actor Language as a starting point for describing an image processing algorithm and targets its implementation to these custom designed, soft-core processors on FPGA. The main purpose is to exploit the task and data parallelism in order to achieve the same parallelism as a previous HDL implementation but avoiding the design time, verification and debugging steps associated with such approaches.

Efficiently Scheduling Task Dataflow Parallelism: A Comparison Between Swan and QUARK

Increased system variability and irregularity of parallelism in applications put increasing demands on the ef- ficiency of dynamic task schedulers. This paper presents a new design for a work-stealing scheduler supporting both Cilk- style recursively parallel code and parallelism deduced from dataflow dependences. Initial evaluation on a set of linear algebra kernels demonstrates that our scheduler outperforms PLASMA’s QUARK scheduler by up to 12% on a 16-thread Intel Xeon and by up to 50% on a 32-thread AMD Bulldozer.

Automatic FPGA Synthesis of Memory Intensive C-based Kernels

Realising high performance image and signal processing
applications on modern FPGA presents a challenging implementation problem due to the large data frames streaming through these systems. Specifically, to meet the high bandwidth and data storage demands of these applications, complex hierarchical memory architectures must be manually specified
at the Register Transfer Level (RTL). Automated approaches which convert high-level operation descriptions, for instance in the form of C programs, to an FPGA architecture, are unable to automatically realise such architectures. This paper
presents a solution to this problem. It presents a compiler to automatically derive such memory architectures from a C program. By transforming the input C program to a unique dataflow modelling dialect, known as Valved Dataflow (VDF), a mapping and synthesis approach developed for this dialect can
be exploited to automatically create high performance image and video processing architectures. Memory intensive C kernels for Motion Estimation (CIF Frames at 30 fps), Matrix Multiplication (128x128 @ 500 iter/sec) and Sobel Edge Detection (720p @ 30 fps), which are unrealisable by current state-of-the-art C-based synthesis tools, are automatically derived from a C description of the algorithm.

Prefetching and Cache Management using Task Lifetimes

Task-based dataflow programming models and runtimes emerge as promising candidates for programming multicore and manycore architectures. These programming models analyze dynamically task dependencies at runtime and schedule independent tasks concurrently to the processing elements. In such models, cache locality, which is critical for performance, becomes more challenging in the presence of fine-grain tasks, and in architectures with many simple cores.

This paper presents a combined hardware-software approach to improve cache locality and offer better performance is terms of execution time and energy in the memory system. We propose the explicit bulk prefetcher (EBP) and epoch-based cache management (ECM) to help runtimes prefetch task data and guide the replacement decisions in caches. The runtimem software can use this hardware support to expose its internal knowledge about the tasks to the architecture and achieve more efficient task-based execution. Our combined scheme outperforms HW-only prefetchers and state-of-the-art replacement policies, improves performance by an average of 17%, generates on average 26% fewer L2 misses, and consumes on average 28% less energy in the components of the memory system.

Constructive Synthesis of Memory-Intensive Accelerators for FPGA From Nested Loop Kernels

Field-programmable gate arrays are ideal hosts to custom accelerators for signal, image, and data processing but de- mand manual register transfer level design if high performance and low cost are desired. High-level synthesis reduces this design burden but requires manual design of complex on-chip and off-chip memory architectures, a major limitation in applications such as video processing. This paper presents an approach to resolve this shortcoming. A constructive process is described that can derive such accelerators, including on- and off-chip memory storage from a C description such that a user-defined throughput constraint is met. By employing a novel statement-oriented approach, dataflow intermediate models are derived and used to support simple ap- proaches for on-/off-chip buffer partitioning, derivation of custom on-chip memory hierarchies and architecture transformation to ensure user-defined throughput constraints are met with minimum cost. When applied to accelerators for full search motion estima- tion, matrix multiplication, Sobel edge detection, and fast Fourier transform, it is shown how real-time performance up to an order of magnitude in advance of existing commercial HLS tools is enabled whilst including all requisite memory infrastructure. Further, op- timizations are presented that reduce the on-chip buffer capacity and physical resource cost by up to 96% and 75%, respectively, whilst maintaining real-time performance.