Modern High Level Synthesis (HLS) tools succeed well in their engineering productivity goal, but still require toolset and target technology specific modifications to the source code to guide the process towards an efficient implementation. Furthermore, their end result is a fixed function accelerator with limited field and runtime flexibility. In this paper we describe the status of AEx, a novel work-in-progress HLS tool developed in the FitOptiVis ECSEL JU project. AEx is based on automated exploration of architectures using a flexible and lightweight parallel co-processor template. We compare its current performance in CHStone C-language benchmarks to the state of the art FPGA HLS tool Vitis, provide ASIC implementation numbers, and identify the main remaining toolset features that are expected to dramatically further improve the performance. The potential is explored with a hand-optimized case study that shows only 1.64x performance slowdown with the programmable co-processor in comparison to the fixed function Vitis HLS result.

@en

Many applications make extensive use of various forms of compression techniques for storing and communicating data. As decompression is highly regular and repetitive, it is a suitable candidate for acceleration. Examples are offloading (de)compression to a dedicated circuit on a heterogeneous System-on-Chip, or attaching FPGAs or ASICs directly to storage so they can perform these tasks on-the-fly and transparently to the application. ASIC or FPGA implementations will usually result in higher energy-efficiency compared to CPUs. Various ASIC and FPGA accelerators have been developed, but they typically target a single algorithm. However, supporting different compression algorithms could be desirable in many situations. For example, the Apache Parquet file format popular in Big Data analytics supports using different compression standards, even between blocks in a single file. This calls for a more flexible software based co-processor approach. To this end, we propose a compiler-supported Application-Specific Instruction-set Processor (ASIP) design that is able to decompress a range of lossless compression standard without FPGA reconfiguration. We perform a case study of searching a compressed database dump of the entire English Wikipedia.

@en

The big data revolution has ushered an era with ever increasing volumes and complexity of data requiring ever faster computational analysis. During this very same era, CPU performance growth has been stagnating, pushing the industry to either scale their computation horizontally using multiple nodes in datacenters, or to scale vertically using heterogeneous components to reduce compute time. However, networking and storage continue to provide both higher throughput and lower latency, which allows for leveraging heterogeneous components, deployed in data centers around the world. Still, the integration of big data analytics frameworks with heterogeneous hardware components such as GPGPUs and FPGAs is challenging, because there is an increasing gap in the level of abstraction between analytics solutions developed with big data analytics frameworks, and accelerated kernels developed with heterogeneous components. In this article, we focus on FPGA accelerators that have seen wide-scale deployment in large cloud infrastructures. FPGAs allow the implementation of highly optimized hardware architectures, tailored exactly to an application, and unburdened by the overhead associated with traditional general-purpose computer architectures. FPGAs implementing dataflow-oriented architectures with high levels of (pipeline) parallelism can provide high application throughput, often providing high energy efficiency. Latency-sensitive applications can leverage FPGA accelerators by directly connecting to the physical layer of a network, and perform data transformations without going through the software stacks of the host system. While these advantages of FPGA accelerators hold promise, difficulties associated with programming and integration limit their use. This article explores the existing practices in big data analytics frameworks, discusses the aforementioned gap in development abstractions, and provides some perspectives on how to address these challenges in the future. @en

In the domain of big data analytics, the bottleneck of converting storage-focused file formats to in-memory data structures has shifted from the bandwidth of storage to the performance of decoding and decompression software. Two widely used formats for big data storage and in-memory data are Apache Parquet and Apache Arrow, respectively. In order to improve the speed at which data can be loaded from disk to memory, we propose an FPGA accelerator design that converts Parquet files to Arrow in-memory data structures. We describe an extensible, publicly available, free and open-source implementation of the proposed converter that supports various Parquet file configurations. The performance of the converter is measured on an AWS EC2 F1 system and on a POWER9 system using the recently released OpenCAPI interface. A single instance of the converter can reach between 6 and 12 GB/s of end-to-end throughput, and shows up to a threefold improvement over the fastest single-thread CPU implementation. It has a low resource utilization (less than 5% for all types of FPGA resources). This allows scaling out the design to match the bandwidth of the coming generation of accelerator interfaces. The proposed design and implementation can be extended to support more of the many possible Parquet file configurations.@en

We propose a novel reconfigurable hardware architecture to implement Monte Carlo based simulation of physical dose accumulation for intensity-modulated adaptive radiotherapy. The long term goal of our effort is to provide accurate dose calculation in real-time during patient treatment. This will allow wider adoption of personalised patient therapies which has the potential to significantly reduce dose exposure to the patient as well as shorten treatment and greatly reduce costs. The proposed architecture exploits the inherent parallelism of Monte Carlo simulations to perform domain decomposition and provide high resolution simulation without being limited by on-chip memory capacity. We present our architecture in detail and provide a performance model to estimate execution time, hardware area and bandwidth utilisation. Finally, we evaluate our architecture on a Xilinx VU9P platform as well as the Xilinx Alveo U250 and show that three VU9P based cards or two Alevo U250s are sufficient to meet our real time target of 100 million randomly generated particle histories per second.

@en

We propose a novel reconfigurable hardware architecture to implement Monte Carlo based simulation of physical dose accumulation for intensity-modulated adaptive radiotherapy. The long term goal of our effort is to provide accurate online dose calculation in real-time during patient treatment. This will allow wider adoption of personalised patient therapies which has the potential to significantly reduce dose exposure to the patient as well as shorten treatment and greatly reduce costs. The proposed architecture exploits the inherent parallelism of Monte Carlo simulations to perform domain decomposition and provide high resolution simulation without being limited by on-chip memory capacity. We present our architecture in detail and provide a performance model to estimate execution time, hardware area and bandwidth utilisation. Finally, we evaluate our architecture on a Xilinx VU9P platform and show that three cards are sufficient to meet our real time target of 100 million randomly generated particle histories per second.

@en

The proliferation of processing hardware alternatives allows developers to use various customized computing platforms to run their applications in an optimal way. However, porting application code on custom hardware requires a lot of development and porting effort. This paper describes a heterogeneous computational platform (the ALMARVI execution platform) comprising of multiple communicating processors that allow easy programmability through an interface to OpenCL. The ALMARVI platform uses processing elements based on both VLIW and Transport Triggered Architectures (ρ-VEX and TCE cores, respectively). It can be implemented on Zynq devices such as the ZedBoard, and supports OpenCL by means of the pocl (Portable OpenCL) project and our ALMAIF interface specification. This allows developers to execute kernels transparently on either processing elements, thereby allowing to optimize execution time with minimal design and development effort.@en

This paper presents and evaluates an approach to deploy image and video processing pipelines that are developed frame-oriented on a hardware platform that is stream-oriented, such as an FPGA. First, this calls for a specialized streaming memory hierarchy and accompanying software framework that transparently moves image segments between stages in the image processing pipeline. Second, we use softcore VLIW processors, that are targetable by a C compiler and have hardware debugging capabilities, to evaluate and debug the software before moving to a High-Level Synthesis flow. The algorithm development phase, including debugging and optimizing on the target platform, is often a very time consuming step in the development of a new product. Our proposed platform allows both software developers and hardware designers to test iterations in a matter of seconds (compilation time) instead of hours (synthesis or circuit simulation time).

@en

In the design of modern-day processors, energy consumption and fault tolerance have gained significant importance next to performance. This is caused by battery constraints, thermal design limits, and higher susceptibility to errors as transistor feature sizes are decreasing. However, achieving the ideal balance among them is challenging due to their conflicting nature (e.g., fault-tolerance techniques usually influence execution time or increase energy consumption), and that is why current processor designs target at most two of these axes. Based on that, we propose a new VLIW-based processor design capable of adapting the execution of the application at run-time in a totally transparent fashion, considering performance, fault tolerance, and energy consumption altogether, in which the weight (priority) of each one can be defined a priori. This is achieved by a novel decision module that dynamically controls the application's ILP to increase the possibility of replicating instructions or applying power gating. For an energy-oriented configuration, it is possible, on average, to reduce energy consumption by 37.2% with an overhead of only 8.2% in performance, while maintaining low levels of failure rate, when compared to a fault-tolerant design.@en

To achieve energy savings while maintaining adequate performance, system designers and programmers wish to create the best possible match between program behavior and the underlying hardware. Well-known current approaches include DVFS and task migrations in heterogeneous platforms such as big.LITTLE processors. Additionally, processors have been proposed in literature that are able to adapt (parts of) their organization to the workload. These reconfigurations can be managed using hardware monitors, profiling and other compile-time information or a combination of both. Many current solutions are suitable for heterogeneous systems, as migration penalties pose a practical limit to the maximum adaptation frequency, but not for dynamic processors that can adapt much more fine-grained.

In this paper, we present two novel concepts to aid these low-penalty reconfigurable processors - one requiring an ISA extension and one without. Our experimental results show that our approaches enable a dynamic processor to reduce the energy-delay product by up to 25% and on average 10% to 18% compared to the best performing static setups.@en

Embedded systems range from very simple devices, such as a digital watch, to highly complex systems such as smartphones. In these complex devices, an increasing number of applications need to be executed on a computing platform. Moreover, the number of applications (or programs) usually exceeds the number of processors found on such platforms. This creates the need for scheduling. Furthermore, each program exhibits different characteristics and their interaction with the (real-life) environmentment leads to real-time requirements. Consequently, the set of programs, called workload, exhibits highly dynamic behavior. Workloads can be dynamic in intensity (i.e., the number of concurrent tasks), characteristics (amount and type of parallelism), and requirements (real-time constraints, power budgets, performance). We argue that dynamic workloads require a dynamic computing platform and propose to use one that comprises the 휌-VEX reconfigurable VLIW processor. It can dynamically adapt to the workload while it is running. Adaptations can be triggered by a user, programmer, compiler, or an operating system. The latter two methods can operate fully automatic and exploring these is one of the goals of this work. Besides dynamic workloads, a number of new classes of embedded devices are running application programs that are very static, but require very high throughput. Examples are the latest generations mobile telecommunications hardware and vision-based applications (automation, surveillance, automated driving). In this case, adapting to the workload at run-time is not advantageous because there are no changes to adapt to. Optimizing for these applications is possible, but must be done before the hardware platform is manufactured (during the design phase) or by making use of Field-Programmable Gate Arrays (FPGAs). This thesis explores the use of the proposed reconfigurable processor to target the full spectrum of embedded workloads. First, design-time reconfigurability is employed to optimize a hardware platform for a static, streaming image processing workload. Second, we explore the run-time reconfigurable processor for dynamic workloads. This is achieved by adapting to a single program to optimize energy efficiency, followed by adapting to a generated set of programs optimizing for throughput. Third, the real-time characteristics of the processor are evaluated and it is shown to have better schedulability compared to static processors. The VLIW architecture results in good timing-predictability, which allows finding tight bounds on the worst-case execution time. Last, we show that the processor is able to assign more parallel execution resources to a static program that is added into the workload, while still guaranteeing time-safety for critical tasks. @en

As embedded systems are faced with ever more demanding workloads and more tasks are being consolidated onto a smaller number of microcontrollers, system designers are faced with opposing requirements of increasing performance while retaining real-time analyzability. For example, one can think of the following performance-enhancing techniques: caches, branch prediction, out-of-order (OoO) superscalar processing, simultaneous multi-threading (SMT). Clearly, there is a need for a platform that can deliver high performance for non-critical tasks and full analyzability and predictability for critical tasks (mixed-criticality systems). In this paper, we demonstrate how a polymorphic VLIW processor can satisfy these seemingly contradicting goals by allowing a schedule to dynamically, i.e., at run-time, distribute the computing resources to one or multiple threads. The core provides full performance isolation between threads and can keep multiple task contexts in hardware (virtual processors, similar to SMT) between which it can switch with minimal penalty (a pipeline flush). In this work, we show that this dynamic platform can improve performance over current predictable processors (by a factor of 5 on average using the highest performing configuration), and provides schedulability that is on par with an earlier study that explored the concept of creating a dynamic processor based on a superscalar architecture. Furthermore, we measured a 15% improvement in schedulability over a heterogeneous multi-core platform with an equal number of datapaths. Finally, our VHDL design and tools (including compiler, simulator, libraries etc.) are available for download for the academic community.@en

Abstract—Multi-threaded applications execute their threads on different cores with their own local caches and need to share data among the threads. Shared caches are used to avoid lengthy and costly main memory accesses. The degree of cache sharing is a balance between reducing misses and increased hit latency. Dynamic caches have been proposed to adapt this balance to the workload type. Similarly, dynamic processors aim to execute workloads as efficient as possible to being able to balance between exploiting Instruction-level parallelism (ILP) and Thread-level parallelism (TLP). To support this, they consist of multiple processing components and caches that have adaptable interconnects between them. Depending on the workload characteristics, these can connect them together to form a large core that exploits ILP, or split them up to form multiple cores that can run multiple threads (exploiting TLP). In this paper, we propose a cache system that is able to further exploit this additional connectivity of a dynamic VLIW processor by being able to forward cache accesses to multiple cache blocks while the processor is running in multi-threaded (‘split’) mode. Additionally, only requests to global data are broadcasted, while accesses to local data are kept private. This will improve the hit rates similar to existing cache sharing schemes, but reduce the penalty due to stalling the other subcores. Local accesses are recognized by distinguishing memory accesses relative to the stack frame pointer. Results show that our cache exhibits similar miss rate reductions as shared caches (up to 90% and on average 26%), and reduces the number of broadcasted accesses by 21%.@en

In this paper, we present and evaluate an FPGA acceleration fabric that uses VLIW softcores as processing elements, combined with a
memory hierarchy that is designed to stream data between intermediate stages of an image processing pipeline. These pipelines are commonplace in medical applications such as X-ray imagers. By using a streaming memory hierarchy, performance is increased by a factor that depends on the number of stages (7.5× when using 4 consecutive filters). Using a Xilinx VC707 board, we are able to place up to 75 cores. A platform of 64 cores can be routed at 193 MHz, achieving real-time performance, while keeping 20% resources available for off-board interfacing. Our VHDL implementation and associated tools (compiler, simulator, etc.) are available for download for the academic community.@en

In today’s computing environments, the concurrent execution of multiple applications/threads is common and multi-cores are very
well-suited to handle such workloads. However, they suffer from the fact that any mismatch between the application’s inherent instruction-level parallelism (ILP) and the core’s parallelism leads to unused resources or loss in performance. An accepted solution is to include several types of cores and match them dynamically depending on the performance needs of the application. This approach becomes less efficient when the number of cores does not match the number of parallel threads. Furthermore,
the heterogeneity of (fixed) cores cannot be increased indefinitely as it would result in even higher degrees of mismatching and increased movement of instruction and data streams. In this paper, we are proposing a polymorphic processor, based on VLIW architectures, that can adapt its issue-width during runtime. By design, the processor can be perceived as a single wide core (8-issue VLIW) or two medium-wide cores (4-issue) or four small cores (2-issue) that can run high-ILP/low DLP, medium-ILP/medium DLP, and low-ILP/high-DLP applications, respectively. Furthermore, we are executing one single generic binary while performing these reconfigurations. In order to show the effectiveness of our approach, we synthesized different versions of the core to represent fixed heterogeneous cores and compared them to the dynamic implementation of the core. Our experiments show that the dynamically adaptive solution performs on average 7% faster and uses 5% less area than a platform which consists of fixed cores with 1.5× as many datapaths.@en

The register file is an expensive component in the design of any processor, especially, when considering the additional ports that are needed to support multiple datapaths within a wide-issue VLIW processor. In a recent work, these additional resources were used to dynamically reconfigure the register file to support a dynamically reconfigurable VLIW core. The design can be perceived as a single 8-issue, two 4-issue, or four 2-issue VLIW cores. Consequently, the multi-ported design can operate in different modes, namely as one, two, or four register files, respectively, corresponding to the active number of cores. The implementation of the register file design on FPGAs using Block RAMs still results in unused resources due to the coarseness of the Block RAMs. In this paper, we propose to re-purpose these unused BRAM resources to additionally support multiple contexts next to earliermentioned modes. In this manner, the 8-issue, 4-issue, and 2- issue cores have access to 4, 2, and 1 contexts, respectively. Consequently, we can avoid saving and restoring of the task states in a multi-task environment, turning context switching from a traditionally time-consuming event to an almost instantaneous event. The advantage of this is the reduction of interrupt latency and task switching latency, which are important in real-time and embedded systems. Our results show that our technique can improve interrupt latency by a factor of 17.4x compared to using a software register spill routine, depending on the behavior of the memory system. Likewise, the task switching time can be improved by 6.7x.

@en

Very Long Instruction Word (VLIW) processors are commonplace in embedded systems due to their inherent lowpower consumption as the instruction scheduling is performed by the compiler instead by sophisticated and power-hungry hardware instruction schedulers used in their RISC counterparts. This is achieved by maximizing resource utilization by only targeting a certain application domain. However, when the inherent application ILP (instruction-level parallelism) is low, resources are under-utilized/wasted and the encoding of NOPs results in large code sizes and consequently additional pressure on the memory subsystem to store these NOPs. To address the resource-utilization issue, we proposed a dynamic VLIW processor design that can merge unused resources to form additional cores to execute more threads. Therefore, the formation of cores can result in issue widths of 2, 4, and 8. Without sacrificing the possibility of code interruptability and resumption, we proposed a generic binary scheme that allows a single binary to be executed on these different issue-width cores. However, the code size issue remains as the generic binary scheme even slightly further increases the number NOPS. Therefore, in this paper, we propose to apply a well-known stop-bit code compression technique to the generic binaries that, most importantly, maintains its code compatibility characteristic allowing it to be executed on different cores. In addition, we present the hardware designs to support this technique in our dynamic core. For prototyping purposes, we implemented our design on a Xilinx Virtex-6 FPGA device and executed 14 embedded benchmarks. For comparison, we selected a nondynamic/ static VLIW core that incorporates a similar stop-bit technique for its code compression. We demonstrate, while maintaining code compatibility on top of a flexible dynamic VLIW processor, that the code size can be significantly reduced (up to 80%) resulting in energy savings, and that the performance can be increased (up to a factor of three). Finally, our experimental results show that we can use smaller caches (2 to 4 times as small), which will further help in decreasing energy consumption.

@en