The relevance of heterogeneous architectures has significantly increased over the last decade due to stagnation of performance scaling. Concurrently, increased performance-energy tradeoff requirements driven by the growth of edge computing, with a large focus on Artificial Intelligence (AI) inference, further motivates efforts towards hardware customization. In this context, the open RISC-V Instruction Set Architecture (ISA) and its custom extension oriented paradigm are a relevant technology towards this specialization. However, customizing a processor is a lengthy process requiring Hardware Description Language (HDL) expertise. Furthermore, for validation and simulation purposes, implementing an Instruction Set Simulator (ISS) of the modified core may also be a necessity. This introduces the need for development of two unrelated codebases, increasing development time and effort. In this paper, we explore High-Level-Synthesis (HLS) to realize both the hardware and the respective simulator through a single codebase, which reduces design effort and simplifies specialization of a RISC-V through specification of custom instructions at high level. We present a C++ based design of a RISC-V core, and validate it as an ISS, as well as a hardware module synthesized for an AMD Zynq UltraScale+ Field Programmable Gate Array (FPGA) through HLS, which we integrated in a System-on-Chip (SoC), and functionally validated through a state-of-the-art set of unit tests.

Advancements in semiconductor technology no longer occur at the pace the industry had been accustomed to. We have entered what is considered by many to be the post-Moore era. In order to continue scaling performance, increasingly heterogeneous architectures are being developed and the use of special purpose accelerators is on the rise. One notable example are Field-Programmable-Gate-Arrays (FPGAs), both in the data-center and embedded spaces. Advances in FPGA features and tools is allowing for critical kernels to be accelerated on specialized hardware without fabrication costs. However, re-targeting code to such heterogeneous platforms still requires significant refactoring of the compute intensive kernels, as well as knowledge of parallel compute and hardware design concepts for maximization of performance. We present Tribble, a source-to-source framework under active development, capable of transforming regular C/C++ programs for execution on heterogeneous architectures. This includes transforming the target kernel source code so that it is amenable for circuit generation while keeping the original version for software execution, inserting code for task and memory management and injecting a scheduler algorithm.

By using Dynamic Binary Translation, instruction traces from pre-compiled applications can be offloaded, at runtime, to FPGA-based accelerators, such as Coarse-Grained Loop Accelerators, in a transparent way. However, scheduling onto coarse-grain accelerators is challenging, with two of current known issues being the density of computations that can be mapped, and the effects of memory accesses on performance. Using an in-house framework for analysis of instruction traces, we explore the effect of different window sizes when applying list scheduling, to map the window operations to a coarse-grain loop accelerator model that has been previously experimentally validated. For all window sizes, we vary the number of ALUs and memory ports available in the model, and comment how these parameters affect the resulting latency. For a set of benchmarks taken from the PolyBench suite, compiled for the 32-bit MicroBlaze softcore, we have achieved an average iteration speedup of 5.10x for a basic block repeated 5 times and scheduled with 8 ALUs and memory ports, and an average speedup of 5.46x when not considering resource constraints. We also identify which benchmarks contribute to the difference between these two speedups, and breakdown their limiting factors. Finally, we reflect on the impact memory dependencies have on scheduling.

We discuss the concept of FPGA-unfriendliness, the property of certain algorithms, programs, or domains which may limit their applicability to FPGAs. Specifically, we look at graph analysis, which has recently seen increased interest in combination with High-Level Synthesis, but has yet to find great success compared to established acceleration mechanisms. To this end, we make use of Xilinx's Vitis Graph Library to implement Single-Source Shortest Paths (SSSP) and PageRank (PR), and present a custom kernel written from the ground up for Distinctiveness Centrality (DC, a novel graph centrality measure). We use public datasets to test these implementations, and analyse power consumption and execution time. Our comparisons against published data for GPU and CPU execution show FPGA slowdowns in execution time between around 18.5x and 328x for SSSP, and around 1.8x and 195x for PR, respectively. In some instances, we obtained FPGA speedups versus CPU of up to 2.5x for PR. Regarding DC, results show speedups from 0.1x to 3.5x, and energy efficiency increases from 0.8x to 6x. Lastly, we provide some insights regarding the applicability of FPGAs in FPGA-unfriendly domains, and comment on the future as FPGA and HLS technology advances.

Performance and power efficiency in edge and embedded systems can benefit from specialized hardware. To avoid the effort of manual hardware design, we explore the generation of accelerator circuits from binary instruction traces for several Instruction Set Architectures.

The use of specialized accelerator circuits is a feasible solution to address performance and energy issues in embedded systems. This paper extends a previous field-programmable gate array-based approach that automatically generates pipelined customized loop accelerators (CLAs) from runtime instruction traces. Despite efficient acceleration, the approach suffered from high area and resource requirements when offloading a large number of kernels from the target application. This paper addresses this by enhancing the CLA with dynamic partial reconfiguration (DPR) support. Each kernel to accelerate is implemented as a variant of a reconfigurable area of the CLA which hosts all functional units and configuration memory. Evaluation of the proposed system is performed on a Virtex-7 device. We show, for a set of 21 kernels, that when comparing two CLAs capable of accelerating the same subset of kernels, the one which benefits from DPR can be up to$4.3\times $smaller. Resorting to DPR allows for the implementation of CLAs which support numerous kernels without a significant decrease in operating frequency and does not affect the initiation intervals at which kernels are scheduled. Finally, the area required by a CLA instance can be further reduced by increasing the IIs of the scheduled kernels.

Many embedded applications process large amounts of data using regular computational kernels, amenable to acceleration by specialized hardware coprocessors. To reduce the significant design effort, the dedicated hardware may be automatically generated, usually starting from the application’s source or binary code. This paper presents a moduloscheduled loop accelerator capable of executing multiple loops and a supporting toolchain. A generation/scheduling procedure, which fully relies on MicroBlaze instruction traces, produces accelerator instances, customized in terms of functional units and interconnections. The accelerators support integer and single-precision floating-point arithmetic, and exploit instruction-level parallelism, loop pipelining, and memory access parallelism via two read/write ports. A complete implementation of the proposed architecture is evaluated in a Virtex-7 device. Augmenting a MicroBlaze processor with a tailored accelerator achieves a geometric mean speedup, over software-only execution, of 6.61$\times $for 13 floating-point kernels from the Livermore Loops set, and of 4.08$\times $for 11 integer kernels from Texas Instruments’ IMGLIB. The proposed customized accelerators are compared with ALU-based ones. The average specialized accelerator requires only 0.47$\times $the number of field-programmable gate array slices of an accelerator with four ALUs. A geometric mean speedup of 1.78$\times $over a four-issue very long instruction word (without floating-point support) was obtained for the integer kernels.

This article presents a reconfigurable hardware/software architecture for binary acceleration of embedded applications. A Reconfigurable Processing Unit (RPU) is used as a coprocessor of the General Purpose Processor (GPP) to accelerate the execution of repetitive instruction sequences called Megablocks . A toolchain detects Megablocks from instruction traces and generates customized RPU implementations. The implementation of Megablocks with memory accesses uses a memory-sharing mechanism to support concurrent accesses to the entire address space of the GPP’s data memory. The scheduling of load/store operations and memory access handling have been optimized to minimize the latency introduced by memory accesses. The system is able to dynamically switch the execution between the GPP and the RPU when executing the original binaries of the input application. Our proof-of-concept prototype achieved geometric mean speedups of 1.60× and 1.18× for, respectively, a set of 37 benchmarks and a subset considering the 9 most complex benchmarks. With respect to a previous version of our approach, we achieved geometric mean speedup improvements from 1.22 to 1.53 for the 10 benchmarks previously used.

This paper presents a binary acceleration approach based on extending a General Purpose Processor (GPP) with a Reconfigurable Processing Unit (RPU), both sharing an external data memory. In this approach repeating sequences of GPP instructions are migrated to the RPU. The RPU resources are selected and organized off-line using execution trace information. The RPU core is composed of Functional Units (FUs) that correspond to single CPU instructions. The FUs are arranged in stages of mutually independent operations. The RPU can enable several stages in tandem, depending on the data dependencies. External data memory accesses are handled by a configurable dual-port cache. A prototype implementation of the architecture on a Spartan-6 FPGA was validated with 12 benchmarks and achieved an overall geometric mean speedup of 1.91x.

This paper presents a novel approach to accelerate program execution by mapping repetitive traces of executed instructions, called Megablocks, to a runtime reconfigurable array of functional units. An offline tool suite extracts Megablocks from microprocessor instruction traces and generates a Reconfigurable Processing Unit (RPU) tailored for the execution of those Megablocks. The system is able to transparently movebcomputations from the microprocessor to the RPU at runtime. A prototype implementation of the system using a cacheless MicroBlaze microprocessor running code located in external memory reaches speedups from$2.2\times$to$18.2 \times$for a set of 14 benchmark kernels. For a system setup which maximizes microprocessor performance by having the application code located in internal block RAMs, speedups from$1.4 \times$to$2.8 \times$were estimated.