Search

This paper describes the final stage of the FPGA prototype development of a recurrent signal processor. During the development of this prototype, a set of tools was created, based on which design verification was carried out. We describe the development process and the prototype validation methodology on a class of DSP tasks using a demo task of isolated word recognition. Taking the previously developed tools and methods for verifying software and hardware models, we have developed a specialized design validation tool. This solution made it possible to ensure the uniformity of the validation process for various types of architecture implementation and to establish the correctness of their operation.

The paper examines the self-timed (ST) pipeline register’s tolerance to soft errors in the pipeline stage’s combinational part and in itself. It aims to analyze the known storage register bit’s circuit cases and improve its soft error tolerance. The use of failure tolerant ST coding, which treats the anti-spacer state as a spacer, increases the ST pipeline’s failure tolerance level. Layout techniques of spacing the dual-rail signal component sources at a distance of more than 2 µm from each other reduce the number of failure types in the ST circuits. In particular, switching the dual-rail signal from a correct working state to an inverse one becomes unrealizable. Circuitry techniques, including cross-connections and local feedback, prevent the bit of the ST-pipeline register from sticking in the anti-spacer state and significantly increase the register’s insensitivity to single soft errors. The use of a DICE-like C-element with two in-phase outputs instead of a known circuitry solution ensures the immunity of the register bit for soft errors inside it. All proposed techniques improve the ST circuit’s soft error tolerance level from 76% to 95%.

The paper presents the results of a study of self-timed (ST) digital circuits' soft-error tolerance. Practical ST circuits have a pipeline structure. The combinational parts of the ST pipeline are naturally immune to 72% of short-term soft errors. The proposed circuitry and layout methods increase the ST pipeline combinational part's failure tolerance to 98% and higher. ST pipeline stage register is the most susceptible to soft errors. A typical variant of the ST pipeline register bit unit based on C-elements has a failure tolerance of 83%. The proposed register bit implementation cases increase the failure tolerance of the ST pipeline up to 98%.

The article discusses the results of a preliminary assessment of the data-flow recurrent architecture's performance on a subset of digital signal processing key tasks. Various implementations of vector convolution, FIR filters, IIR filters, adaptive filter, and 256-Point-In-Place FFT are evaluated. The implementation of the listed algorithms is based on the TMS320C55x DSP Library. The preliminary results showed that the performance of the recurrent architecture based on data-flow principles is not inferior to the TMS320C55x, based on von Neumann principles, in terms of the number of computation cycles. Architecture improvement suggestions are presented.

This paper covers the optimization research for the novel data-flow computational architecture called Hybrid Architecture of Recurrent Signal Processor. The testing methodology, based on the shift towards Test-Driven Development of architecture models, is provided. We cover the toolset developed to unify the methodology for both software and hardware models. The testing results are analyzed, and the issues are formulated. We propose the pipeline extension by splitting our largest component down. We show the new pipeline utilization ratio this solution provides.

Paper presents the results of a study of the self-timed fused multiply-add-subtract unit (FMAS) alternates. All FMAS alternates comply with the IEEE 754 standard and use the modified Booth algorithm to multiply two input 64-bit operands, followed by the addition and subtraction of the third operand. They differ from each other by internal signals self-timed coding: dual-rail, redundant ternary, or redundant quaternary code. The paper analyzes and compares their features, offers their optimized pipeline implementations and recommendations for their use. FMAS alternates have approximately the same performance but different hardware costs and layout sizes.

Paper discusses a problem of the CMOS self-timed circuits' indication. Large number of indicating signals in the multi-bit computational devices and registers requires an additional hardware and time for their combining and forming a single control signal that provides a request-acknowledge interaction between interconnected self-timed functional blocks. Indication subcircuit performs this. Multi-input hysteretic triggers allows for accelerating indication subcircuit by factor of 1.1 − 1.6 and reducing its complexity in several times in comparison to standard implementation basis on static and semi-static Muller's elements. A penalty for this is some short-circuit current in the worst case.

An unconventional multi-core recurrent data-flow architecture, that is being developed at Federal Research Center “Computer Science and Control” of the Russian Academy of Sciences (FRS CSC RAS)was successfully tested on digital signal processing domain both at the model level and on a hardware sample. Based on the test results, several mechanisms had been identified that required improvement and a decision was made to investigate the architecture on other subject domains. Software and main architectural blocks debugging are carried out with the specially developed hardware and software modeling tools. The active extension and debugging of the architecture by using these tools revealed a number of shortcomings of the existing software. To eliminate these shortcomings, two problems have to be solved: to provide a high degree of reconfigurability of the architecture's imitational model (to debug its mechanisms)and implement a symbolic modeling mode (to debug its software). The redesigning results of modeling and debugging tools for recurrent data-flow architecture are discussed in the article.

The results of a Speed-Independent Fused Multiply-Add (SIFMA) unit pipeline implementation research are presented. SIFMA is compliant with IEEE 754 Standard. A criterion of the SIFMA pipeline's maximum performance is formulated. A method of reducing hardware costs of SIFMA multiplier by 1.5-2 times depending on its features is offered. The multiplier utilizes a modified Booth algorithm using self-timed redundant code. A new energy efficient self-timed organization of an input and output FIFO was developed. It provides less complexity versus a previous SIFMA implementation on base of semi-dense register.

This paper provides approbation results of the multi-core hybrid architecture for recurrent signal processing (HARSP) as a hardware sample. The prototype has been designed primarily to check architecture's integrity and universality on digital signal processing domain and to verify the hardware implementation of its imitational model, while operational frequency has not been as relevant. Hardware sample has been implemented on FPGA basis with Cyclone V GT Development Kit. Every data-flow processor implements fixed-point 16-bit processing core while the control level is implemented via generated NIOS II processor. The isolated word recognition with a high confidence threshold (at least 95% with a noise level of 15 dB) has been selected as the subject area. We compare HARSP efficiency against specialized TI C66x DSP by implementing the subset of BTDImark2000 algorithms, using computational steps amount the comparison metric.

This paper covers non-conventional recurrent data-flow architecture, its features, and implementation aspects. Recurrence — the main feature of the new architecture efficiently solves data redundancy problem, typical for data-flow architectures while increasing performance. Conventional recurrence implementation has an overhead of configuration operand insertion that provides required functional fields (tags). Functional capabilities expansion of the architecture mechanism implementing this feature resulted in further efficiency by eliminating said overhead in some instances. We cover enhancements implemented in multicore recurrent data-flow architecture, designed to increase the versatility of recurrent computational process utilization. We compare Viterbi algorithm implementations with and without enhancements.

This paper covers one of the main disadvantages of data-flow computational model — high overhead costs associated with storing and processing large volumes of tag data. Overcoming this disadvantage using various ways of data compression comes with a number of problems described in this article. To solve given problems new recurrent data-flow computational model was created as well as architecture based on that model. This paper describes key features and mechanics of the new model and architecture allowing us to reduce data redundancy in memory storage. Efficiency of developed mechanics is shown in implementation of fast Fourier transform algorithm.