ACM Transactions on Reconfigurable Technology and Systems最新文献

High performance stream aggregation is critical for many emerging applications that analyze massive volumes of data. Incoming data needs to be stored in a sliding window during processing, in case the aggregation functions cannot be computed incrementally. Updating the window with new incoming values and reading it to feed the aggregation functions are the two primary steps in stream aggregation. Although window updates can be supported efficiently using multi-level queues, frequent window aggregations remain a performance bottleneck as they put tremendous pressure on the memory bandwidth and capacity. This article addresses this problem by enhancing StreamZip, a dataflow stream aggregation engine that is able to compress the sliding windows. StreamZip deals with a number of data and control dependency challenges to integrate a compressor in the stream aggregation pipeline and alleviate the memory pressure posed by frequent aggregations. In addition, StreamZip incorporates a caching mechanism for dealing with skewed-key distributions in the incoming data stream. In doing so, StreamZip offers higher throughput as well as larger effective window capacity to support larger problems. StreamZip supports diverse compression algorithms offering both lossless and lossy compression to integers as well as floating-point numbers. Compared to designs without compression, StreamZip lossless and lossy designs achieve up to 7.5× and 22× higher throughput, while improving the effective memory capacity by up to 5× and 23×, respectively.

Fibonacci and Galois are two different kinds of configurations in stream ciphers. Although many transformations between two configurations have been proposed, there is no sufficient analysis of their FPGA performance. Espresso stream cipher provides an ideal sample to explore such a problem. The 128-bit secret key Espresso is designed in Galois configuration, and there is a Fibonacci-configured Espresso variant proved with the equivalent security level. To fully leverage the efficiency of two configurations, we explore the hardware optimization approaches toward area and throughput, respectively. In short, the FPGA-implemented Fibonacci cipher is more suitable for extremely resource-constrained or high-throughput applications, while the Galois cipher compromises both area and speed. To the best of our knowledge, this is the first work to systematically compare the FPGA performance of cipher configurations under relatively fair cryptographic security. We hope this work can serve as a reference for the cryptography hardware architecture research community.

Practical implementation of deep neural networks (DNNs) demands significant hardware resources, necessitating high computational power and memory bandwidth. While existing field-programmable gate array (FPGA)–based DNN accelerators are primarily optimized for fast single-task performance, cost, energy efficiency, and overall throughput are crucial considerations for their practical use in various applications. This article proposes a performance-centric pipeline Coordinate Rotation Digital Computer (CORDIC)–based MAC unit and implements a scalable CORDIC-based DNN architecture that is area- and power-efficient and has high throughput. The CORDIC-based neuron engine uses bit-rounding to maintain input-output precision and minimal hardware resource overhead. The results demonstrate the versatility of the proposed pipelined MAC, which operates at 460 MHz and allows for higher network throughput. A software-based implementation platform evaluates the proposed MAC operation’s accuracy for more extensive neural networks and complex datasets. The DNN accelerator with parameterized and modular layer-multiplexed architecture is designed. Empirical evaluation through Pareto analysis is used to improve the efficiency of DNN implementations by fixing the arithmetic precision and optimal pipeline stages. The proposed architecture utilizes layer-multiplexing, a technique that effectively reuses a single DNN layer to enhance efficiency while maintaining modularity and adaptability for integrating various network configurations. The proposed CORDIC MAC-based DNN architecture is scalable for any bit-precision network size, and the DNN accelerator is prototyped using the Xilinx Virtex-7 VC707 FPGA board, operating at 66 MHz. The proposed design does not use any Xilinx macros, making it easily adaptable for ASIC implementation. Compared with state-of-the-art designs, the proposed design reduces resource use by 45% and power consumption by 4× without sacrificing performance. The accelerator is validated using the MNIST dataset, achieving 95.06% accuracy, only 0.35% less than other cutting-edge implementations.

Block RAMs (BRAMs) are the storage houses of FPGAs, providing extensive on-chip memory bandwidth to the compute units implemented using Logic Blocks (LBs) and Digital Signal Processing (DSP) slices. We propose modifying BRAMs to convert them to CoMeFa (Compute-In-Memory Blocks for FPGAs) RAMs. These RAMs provide highly parallel compute-in-memory by combining computation and storage capabilities in one block. CoMeFa RAMs utilize the true dual-port nature of FPGA BRAMs and contain multiple configurable single-bit bit-serial processing elements. CoMeFa RAMs can be used to compute with any precision, which is extremely important for applications like Deep Learning (DL). Adding CoMeFa RAMs to FPGAs significantly increases their compute density, while also reducing data movement. We explore and propose two architectures of these RAMs: CoMeFa-D (optimized for delay) and CoMeFa-A (optimized for area). Compared to existing proposals, CoMeFa RAMs do not require changing the underlying SRAM technology like simultaneously activating multiple wordlines on the same port, and are practical to implement. CoMeFa RAMs are especially suitable for parallel and compute-intensive applications like DL, but these versatile blocks find applications in diverse applications like signal processing, databases, etc. By augmenting an Intel Arria-10-like FPGA with CoMeFa-D (CoMeFa-A) RAMs at the cost of 3.8% (1.2%) area, and with algorithmic improvements and efficient mapping, we observe a geomean speedup of 2.55x (1.85x) across microbenchmarks from various applications and a geomean speedup of up to 2.5x across multiple Deep Neural Networks. Replacing all or some BRAMs with CoMeFa RAMs in FPGAs can make them better accelerators of DL workloads.

To achieve resource-efficient hardware designs, high-level synthesis tools share (i.e., time-multiplex) functional units among operations of the same type. This optimization is typically performed in conjunction with operation scheduling to ensure the best possible unit usage at each point in time. Dataflow circuits have emerged as an alternative HLS approach to efficiently handle irregular and control-dominated code. However, these circuits do not have a predetermined schedule—in its absence, it is challenging to determine which operations can share a functional unit without a performance penalty. More critically, although sharing seems to imply only some trivial circuitry, time-multiplexing units in dataflow circuits may cause deadlock by blocking certain data transfers and preventing operations from executing. In this paper, we present a technique to automatically identify performance-acceptable resource sharing opportunities in dataflow circuits. More importantly, we describe a sharing mechanism which achieves functionally correct and deadlock-free dataflow designs. On a set of benchmarks obtained from C code, we show that our approach effectively implements resource sharing. It results in significant area savings at a minor performance penalty compared to dataflow circuits which do not support this feature (i.e., it achieves a 64%, 2%, and 18% average reduction in DSPs, LUTs, and FFs, respectively, with an average increase in total execution time of only 2%) and matches the sharing capabilities of a state-of-the-art HLS tool.

Recently, there is a trend to use high-level synthesis (HLS) tools to generate dynamically scheduled hardware. The generated hardware is made up of components connected using handshake signals. These handshake signals schedule the components at run time when inputs become available. Such approaches promise superior performance on ‘irregular’ source programs, such as those whose control flow depends on input data. This is at the cost of additional area. Current dynamic scheduling techniques are well able to exploit parallelism among instructions within each basic block (BB) of the source program, but parallelism between BBs is under-explored, due to the complexity in run-time control flows and memory dependencies. Existing tools allow some of the operations of different BBs to overlap, but in order to simplify the analysis required at compile time they require the BBs to start in strict program order, thus limiting the achievable parallelism and overall performance.

We formulate a general dependency model suitable for comparing the ability of different dynamic scheduling approaches to extract maximal parallelism at run-time. Using this model, we explore a variety of mechanisms for run-time scheduling, incorporating and generalising existing approaches. In particular, we precisely identify the restrictions in existing scheduling implementation and define possible optimisation solutions. We identify two particularly promising examples where the compile-time overhead is small and the area overhead is minimal and yet we are able to significantly speed-up execution time: (1) parallelising consecutive independent loops; and (2) parallelising independent inner-loop instances in a nested loop as individual threads. Using benchmark sets from related works, we compare our proposed toolflow against a state-of-the-art dynamic-scheduling HLS tool called Dynamatic. Our results show that on average, our toolflow yields a 4 × speedup from (1) and a 2.9 × speedup from (2), with a negligible area overhead. This increases to a 14.3 × average speedup when combining (1) and (2).