Convolutional neural networks (CNNs) gained great success in machine learning applications and much attention was paid to their acceleration on field programmable gate arrays (FPGAs). The most demanding computational complexity of CNNs is found in the convolutional layers, which account for 90% of the total operations. The fact that parameters in convolutional layers do not change over a long time interval in weight stationary CNNs allows the use of reconfiguration to reduce the resource requirements. This work proposes several alternative reconfiguration schemes that significantly reduce the complexity of sum-of-products operations. The proposed direct configuration schemes provide the least resource requirements and fast reconfiguration times of 32 clock cycles but require additional memory for the pre-computed configurations. The proposed online reconfiguration scheme uses an online computation of the LUT contents to avoid this memory overhead. Finally, a scheme that duplicates the reconfigurable LUTs is proposed for which the reconfiguration time can be completely hidden in the computation time. Combined with a few online reconfiguration circuits, this provides the same configuration memory and configuration time as a conventional parallel kernel but offers large resource reductions of up to 80% of the LUTs.

A key element for achieving high throughput, e.g. circuits generated with high-level synthesis (HLS) methods and model-based hardware design, is the use of modulo scheduling. Integer linear programming (ILP)-based modulo schedulers are capable of computing schedules that are optimal regarding throughput and latency, while keeping run times to practically usable lengths. However, the generated schedules may lead to an excessive number of registers for storing intermediate values. We propose extensions for ILP-based modulo scheduling that minimizes these registers. The ILP formulation incorporates the elimination of redundant registers by post binding optimization. Extensive experiments on different benchmark sets show average register reductions of 30.4% compared to commonly used minimum lifetime approaches that reduce register requirements. This comes without any loss in throughput or latency and with less than 4% additional scheduling run time compared to state-of-the-art ILP-based modulo schedulers.

This paper presents a new method called optimal shift reassignment (OSR), used for reconfigurable multiplication circuits. These circuits consist of adders, subtractors, shifts, and multiplexers (MUXs). They calculate the multiplication of an input number by one out of several constants which can be selected dynamically during run-time. The OSR method is based on the idea that shifts can be placed at different positions along the circuit, while the calculated output constant stays the same. This differs from previous approaches, which were limited by the fact that all constants within the constant multiplier were forced to be odd. The OSR method subsequently releases this restriction. As a result, the number of required MUXs in the circuit can be reduced. This happens when the shift reassignment aligns the shift values of different inputs of an MUX. Experimental results show MUX savings of up to 50% and average savings between 11% and 16% using the OSR method compared to previous approaches.

Hardware applications in an industrial context often have tight area, latency and throughput requirements or a specific combination thereof. This paper presents a method to improve area and throughput figures for folded circuits generated during a model-based hardware design process. The method targets FPGA implementations and is based on the automatic combination of isomorphic subgraphs and the detailed consideration of pipelined primitive operations for folding core scheduling. In the course of a design space exploration, the user is provided with fine-grain control over the area/throughput trade-off.

This paper introduces a new heuristic to generate pipelined run-time reconfigurable constant multipliers for field-programmable gate arrays (FPGAs). It produces results close to the optimum. It is based on an optimal algorithm which fuses already optimized pipelined constant multipliers generated by an existing heuristic called reduced pipelined adder graph (RPAG). Switching between different single or multiple constant outputs is realized by the insertion of multiplexers. The heuristic searches for a solution that results in minimal multiplexer overhead. Using the proposed heuristic reduces the run-time of the fusion process, which raises the usability and application domain of the proposed method of run-time reconfiguration. An extensive evaluation of the proposed method confirms a 9%-26% FPGA resource reduction on average compared to previous work. For reconfigurable multiple constant multiplication, resource savings of up to 75% can be shown compared to a standard generic lookup table based multiplier. Two low level optimizations are presented, which further reduce resource consumption and are included into an automatic VHDL code generation based on the FloPoCo library.

This paper presents a new algorithm to automatically create pipelined run-time reconfigurable constant multipliers. Reconfiguration between several constants is achieved by merging optimized pipelined adder graphs using multiplexers. The adder graphs perform the required multiplications by using additions, subtractions and bit-shifts only. They are generated by an existing heuristic called RPAG. The resulting reconfigurable pipelined single and multiple constant multipliers can be used for time-multiplexed multiplication reducing the required FPGA logic resources. In contrast to earlier approaches aiming at application-specific integrated circuits (ASICs) we introduce pipelining and take special care of the size of the added multiplexers to obtain FPGA-optimized solutions. We can show that the achieved pipelined run-time reconfigurable constant multipliers on average only need about 77% of the slices compared to the best solutions based on the merging of adder graphs published so far.

Distributed arithmetic is a popular method for implementing digital FIR filters on FPGAs. One essential optimization method is the division of large look-up tables (LUTs) into smaller partial LUTs by using additional adders. Previous work indicates, that the size of these partial LUTs should be chosen to the LUT input size of the FPGA which was 4 for a long time. Nowadays, modern FPGAs offer 6-input LUTs which can be configured to two 5-input LUTs with shared inputs. This paper investigates the optimal input size of partial LUTs on FPGAs with 4-input and 5/6-input LUTs. On FPGAs with 4-input LUTs, it turnes out that only in 62% of the cases (out of 220), a LUT input size of 4 leads to the best implementation. However, the slice overhead is 6.3% on average for the other cases. On FPGAs with 5/6-input LUTs, the least slice overhead (10% on average) is paid when the LUT input size is chosen to 6. However, it was shown that a resource reduction of up to 32% can be achieved when all input sizes in the range 4...7 are evaluated. Using the best partial LUT size, slice reductions of over 50% on average compared to Xilinx Coregen could be achieved for Virtex 6 FPGAs.

An architecture for a dynamically run-time reconfigurable finite impulse response (FIR) filter is presented in this work. It is based on distributed arithmetic (DA) combined with a look-up table (LUT) reduction technique which allows the direct mapping to reconfigurable LUTs (CFGLUT) of the latest Xilinx FPGAs. The resulting FIR filter can be reconfigured with arbitrary coefficients which are only limited by their length and word size. The number of filter instances for reconfiguration is only limited by the block memory of the FPGA which typically allows hundreds of different configurations. The proposed reconfigurable architecture consumes 16% less slices on average than a fixed coefficient DA filter generated by Xilinx Coregen. As the direct mapping to CFGLUTs leads to invalid filter output during reconfiguration, an alternative architecture is proposed which avoids this limitation at the cost of 19% more slice resources on average. Using a parallel reconfiguration scheme, reconfiguration times of about 100ns could be achieved.