Exploring the design space in high-level synthesis

Scheduling and variable binding for improved testability in high level synthesis

In Static List Scheduling (SLS), initially the ALAP (As Late As Possible) algorithm is performed, followed by ASAP (As Soon As Possible) algorithm. The time interval for control steps and successor selection are the vital drawback of ASAP algorithm. Hence, in this paper, the optimized scheduling algorithm is proposed for High Level Synthesis (HLS), where the ALAP in SLS is followed by Force Directed Scheduling (FDS). The proposed scheduling algorithm is implemented in Icarus Verilog and targeted in ZC7020 FPGA board. The proposed optimized scheduling algorithm is evaluated in terms of control steps and resource utilization. The evaluation is carried out for various benchmarks such as, EW Filter, IIR filter, FIR filter, AR filter and Discrete Cosine Transform (DCT). The achieved results in terms of the control steps and resource utilization (i.e. LUT and Flip-flop) shows the proposed optimized scheduling algorithm outperforms the conventional Static List Scheduling (SLS) algorithm.

With the development of fabrication, the capacities of application-specific integrated circuits (ASICs) that implement data-intensive applications are increasing as well. in high-level view, the results of functional units (FUs) and register allocation have a significant impact on power, thermal, area and timing etc., especially for data-dominated behaviors. Techniques for power minimization might be applied in different levels of the design hierarchy, but high-level synthesis (HLS) has attracted a special attention due to its inherent ability of better design space exploration. All kinds of methods for register optimization are presented in the recent past. This paper presents a new simultaneous FUs and register allocation method, which combining heuristic list scheduling algorithm and left-edge algorithm to optimize register number and power. From the experimental results, it shows that different FUs allocation results deliver an average 1.83 reduction in register number. And compared with traditional allocation method, we can obtain a 5.81% reduction in power consumption. In order to optimize area with the design demand, it will be at the expense of power dissipation, our design method also can achieve power-area balance.

FIDER: A Force-Balance-Based Interconnect Delay Driven Re-Synthesis Algorithm for Data-Path Optimization After Floorplan

Reconfigurable computing refers to the notions of configware and flowware. Configware means structural programming, or programming in space to execute computation in space. Flowware means dataflow programming that schedules the data flow for output from or input to the configware architecture. In this paper, data flows of a synthesized computation are formalized. This means that data flow is specified as a behavioral stream function in stream calculus, which is used to underpin the semantics for Register Transfer Level (RTL) synthesis. A stream representation allows the use of coinductive principles in stream calculus. In particular, using the coinductive proof principle, we show that behavioral stream functions in the three-stage synthesis process (scheduling, register allocation and binding, allocation and binding of functional units) are always bisimilar regardless of changes in a scheduling, allocation or binding procedure. Our formalization makes pipelining possible, in which all functional units as well as registers of hardware resources are reused during different control steps (C-steps). Moreover, a coinductive approach to verifying flowware synthesis, which is independent of the heuristic during register allocating and binding step, is proposed as a practical technique.

Approximate computing has been recognized as an effective low power technique for applications with intrinsic error tolerance, such as image processing and machine learning. Existing efforts on this front are mostly focused on approximate circuit design, approximate logic synthesis or processor architecture approximation techniques. This work aims at how to make good use of approximate circuits at system and block level. In particular, approximation aware scheduling, functional unit allocation and binding algorithms are developed for data intensive applications. Simple yet credible error models, which are essential for precision control in the optimizations, are investigated. The algorithms are further extended to include bitwidth optimization in fixed point computations. Experimental results, including those from Verilog simulations, indicate that the proposed techniques facilitate desired energy savings under latency and accuracy constraints.

This paper addresses the problem of true delay estimation during high level design. The true delay is the delay of the longest sensitizable path in the resulting circuit, as opposed to the topological delay which is the delay of the longest path in the circuit. The existing delay estimation techniques either estimate the topological delay, which may be pessimistic if the longest path is unsensitizable or false, or estimate the true delay using gate-level timing analysis which may be prohibitively expensive. Resource sharing in high level synthesis can create false paths in the circuit implementation. Hence, determining the clock period using topological delay can be unduly conservative, resulting in excessive hardware to meet tight timing specifications. In this paper, we introduce an efficient technique to compute an estimate of the true delay. The proposed technique relies on partitioning the paths in the circuit and topological delay computation, and not on path sensitization. The paths in the implementation are partitioned into two sets given the high level information on scheduling and resource sharing: the complete determining path set (CDP/sub R/) and the nondetermining path set (NDP/sub R/). We prove that the delay of the longest path in CDP/sub R/ is lower bounded by the true delay and upper bounded by the topological delay of the circuit. Consequently, an estimate of the true delay of the resulting circuit can be computed by measuring the topological delay of the longest path in CDP/sub R/. We have developed a Functional delay ESTimation tool (FEST). Experimental results on a set of benchmarks reveal the following: approximately 50% of all paths are in NDP/sub R/ and can be ignored for true delay estimation, and the true delay estimates are on the average 15% less than the topological delay. The high level true delay estimates are accurate, as verified by comparing with the true delays obtained by gate-level timing analysis on actual implementations. Furthermore, results reveal that high level true delay estimation can be done very fast, even when gate-level true delay estimation becomes infeasible.

Introduction of self-checking capacity in arithmetic systems since the initial steps of high-level synthesis is taken into account, as an alternative to conventional solutions that adopt ad-hoc coding or comparable techniques after the register-level architecture has been fully defined. A technique based on initial partitioning of the Data Flow Graph into detectable subgraphs is proposed, by which all single errors appearing within one subgraph are detected; an algorithm leading to optimize resource sharing (allocation and binding of both functional units and registers) while keeping minimum latency and optimum number of checkers is discussed.

Java bytecodes can be executed with the following three methods: a Java interpreter running on a particular machine interprets bytecodes; a Just-in-Time (JIT) compiler translates bytecodes to the native primitives of the particular machine and the machine executes the translated codes; and a Java processor executes bytecodes directly. The first two methods require no special hardware support for the execution of Java bytecodes and are widely used currently. The last method requires an embedded Java processor, picoJavaI or picoJavaII for instance. The picoJavaI and picoJavaII are simple pipelined processors with no ILP (instruction level parallelism) and TLP (thread level parallelism) supports. A so-called MAJC (microprocessor architecture for Java computing) design can exploit ILP and TLP by using a modified VLIW (very long instruction word) architecture and vertical multithreading technique, but it has its own instruction set and cannot execute Java bytecodes directly. In this paper, we investigate a processor architecture which can directly execute Java bytecodes meanwhile can exploit Java ILP and TLP simultaneously. The proposed processor consists of multiple slots implementing horizontal multithreading and multiple functional units shared by all threads executed in parallel. Our architectural simulation results show that the Java processor could achieve an average 20 IPC (instructions per cycle), or 7.33 EIPC (effective IPC), with 8 slots and a 4-instruction scheduling window for each slot. We also check other configurations and give the utilization of functional units as well as the performance improvement with various kinds of working loads.

This paper proposes a novel behavioral synthesis method that improves performance of synthesized circuits utilizing specialized functional units effectively. Specialized functional units (e.g. multiply-accumulator) are designed for specific operation patterns to achieve shorter delay and/or smaller area than cascaded basic functional units. Almost all conventional methods cannot use specialized functional units effectively under a total area constraint because of their less flexibility for resource sharing. The proposed method makes it possible to solve module selection, scheduling, and functional unit allocation problems utilizing specialized functional units in practical time with some heuristics, and to reduce the number of clock cycles under total area and clock cycle time constraints. Experimental results show that the proposed method has achieved up to 35% and on average 14% reduction of the number of cycles with specialized functional units in practical time.

This paper proposes a novel behavioral synthesis method that improves performance of synthesized circuits utilizing specialized functional units effectively. Specialized functional units (e.g. multiply-accumulator) are designed for specific operation patterns to achieve shorter delay and/or smaller area than cascaded basic functional units. Almost all conventional methods cannot use specialized functional units effectively under a total area constraint because of their less flexibility for resource sharing. The proposed method makes it possible to solve module selection, scheduling, and functional unit allocation problems utilizing specialized functional units in practical time with some heuristics, and to reduce the number of clock cycles under total area and clock cycle time constraints. Experimental results show that the proposed method has achieved up to 35% and on average 14% reduction of the number of cycles with specialized functional units in practical time.

Speculative Functional Units (SFUs) enable a new execution paradigm for High Level Synthesis (HLS). SFUs are arithmetic functional units that operate using a predictor for the carry signal, which reduces the critical path delay. The performance of these units is determined by the success in the prediction of the carry value, i.e. the hit rate of the prediction. Hence SFUs reduce critical path at a low cost, but they cannot be used in HLS with the current techniques. In order to use them, it is necessary to include hardware support to recover from mispredictions of the carry signals. In this paper, we present techniques for designing a datapath controller for seamless deployment of SFUs in HLS. We have developed two techniques for this goal. The first approach stops the execution of the entire datapath for each misprediction and resumes execution once the correct value of the carry is known. The second approach decouples the functional unit suffering from the misprediction from the rest of the datapath. Hence, it allows the rest of the SFUs to carry on execution and be at different scheduling states at different times. Experiments show that it is possible to reduce execution time by as much as 38% and by 33% on average.

A new method that integrated high level synthesis and floorplan are proposed to reduce the total power dissipation. Function units are sorted into several clusters according to their working voltages. Floorplan are executed with the information of cluster gained from high level synthesis, and then the high level synthesis resolutions are adjusted iteratively with the interconnect information and switching activity gained from floorplan. Not only function unit power and interconnect power, but also the voltage converting power can be reduced in our method. Experimental results indicate that, under the given resource and latency constrains, the interconnect power, level convert power and total power of our method can be reduced by 30.4%, 35.1% and 22.9% respectively.

Reducing resource usage is one of the most important optimization objectives in behavioral synthesis due to its direct impact on power, performance and cost. The datapath in a typical design is composed of different kinds of components, including functional units, registers and multiplexers. To optimize the overall resource usage, a behavioral synthesis tool should consider all kinds of components at the same time. However, most previous work on behavioral synthesis has the limitations of (i) not being able to consider all kinds of resources globally, and/or (ii) separating the synthesis process into a sequence of optimization steps without a consistent optimization objective. In this paper we present a behavioral synthesis flow in which all types of components in the datapath are modeled and optimized consistently. The key idea is to feed to the scheduler the intentions for sharing functional units and registers in favor of the global optimization goal (such as total area), so that the scheduler could generate a schedule that makes the sharing intentions feasible. Experiments show that compared to the solution of minimizing functional unit requirements in scheduling and using the least number of functional units and registers in binding, our solution achieves a 24% reduction in total area; compared to the online tool provided by c-to-verilog.com, our solution achieves a 30% reduction on average.

In this paper, we present a new scheduling algorithm that generates area-efficient register transfer level datapaths with multiport memories. The proposed scheduling algorithm assigns an operation to a specific control step such that maximal sharing of functional units can be achieved with minimal number of memory ports, while satisfying given constraints. We propose a measure of multiport memory cost, MAV (Multiple Access Variable) which is defined as a variable accessed at several control steps, and overall memory cost is reduced by equally distributing the MAVs throughout all the control steps. When compared with previous approaches for several benchmarks available from the literature, the proposed algorithm generates the datapaths with less memory modules and interconnection structures by reflecting the memory cost in the scheduling process.