Neural Networks (NN) are often trained offline on large datasets and deployed on specialised hardware for inference, with a strict separation between training and inference. However, in many realistic applications the training environment differs from the real world, or data arrives in a streaming fashion and is continuously changing. In these scenarios, the ability to continuously train and update NN models is desirable. Continual learning (CL) algorithms allow training of models on a stream of data. CL algorithms are often designed to work in constrained settings, such as limited memory and computational power, or limitations on the ability to store past data (e.g, due to privacy concerns or memory requirements). High-energy physics experiments are developing intelligent detectors, with algorithms running on computer systems located close to the detector to meet the challenges of increased data rates and occupancies. The use of NN algorithms in this context is limited by changing detector conditions, such as degradation over time or failure of an input signal which might cause the NNs to lose accuracy leading, in the worst case to the loss of interesting events. CL has the potential to solve this issue, using large amounts of continuously streaming data to allow the network to recognise changes, and to learn and adapt to detector conditions. It has the potential to outperform traditional NN training techniques as not all possible scenarios can be predicted and modelled in static training data samples. However, NN training is computationally expensive and when combined with the strict timing requirements of embedded processors deployed close to the detector, current state-of-the-art offline approaches cannot be directly applied to the real-time systems. Alternatives to typical backpropagation-based training that can be deployed on FPGAs for real-time data processing are presented, and their computational and accuracy characteristics are discussed in the context of High-Luminosity LHC.

Random number generation is key to many applications in a wide variety of disciplines. Depending on the application, the quality of the random numbers from a particular generator can directly impact both computational performance and critically the outcome of the calculation. High-energy physics applications use Monte Carlo simulations and machine learning widely, which both require high-quality random numbers. In recent years, to meet increasing performance requirements, many high-energy physics workloads leverage GPU acceleration. While on a CPU, there exist a wide variety of generators with different performance and quality characteristics, the same cannot be stated for GPU and FPGA accelerators. On GPUs, the most common implementation is provided by cuRAND - an NVIDIA library that is not open source or peer reviewed by the scientific community. The highest-quality generator implemented in cuRAND is a version of the Mersenne Twister. Given the availability of better and faster random number generators, high-energy physics moved away from Mersenne Twister several years ago and nowadays MIXMAX is the standard generator in Geant4 via CLHEP. The MIXMAX original design supports parallel streams with a seeding algorithm that makes it especially suited for GPU and FPGA where extreme parallelism is a key factor. In this study we implement the MIXMAX generator on both architectures and analyze its suitability and applicability for accelerator implementations. We evaluated the results against “Mersenne Twister for a Graphic Processor” (MTGP32) on GPUs which resulted in 5, 13 and 14 times higher throughput when a 240, 17 and 8 sized vector space was used respectively. The MIXMAX generator coded in VHDL and implemented on Xilinx Ultrascale+ FPGAs, requires 50% fewer total Look Up Tables (LUTs) compared to a 32-bit Mersenne Twister (MT-19337), or 75% fewer LUTs per output bit. In summary, the state-of-the art MIXMAX pseudo random number generator has been implemented on GPU and FPGA platforms and the performance benchmarked.

Processing graphs on a large scale presents a range of difficulties, including irregular memory access patterns, device memory limitations, and the need for effective partitioning in distributed systems, all of which can lead to performance problems on traditional architectures such as CPUs and GPUs. To address these challenges, recent research emphasizes the use of Field-Programmable Gate Arrays (FPGAs) within distributed frameworks, harnessing the power of FPGAs in a distributed environment for accelerated graph processing. This paper examines the effectiveness of a multi-FPGA distributed architecture in combination with a partitioning system to improve data locality and reduce inter-partition communication. Utilizing Hadoop at a higher level, the framework maps the graph to the hardware, efficiently distributing pre-processed data to FPGAs. The FPGA processing engine, integrated into a cluster framework, optimizes data transfers, using offline partitioning for large-scale graph distribution. A first evaluation of the framework is based on the popular PageRank algorithm, which assigns a value to each node in a graph based on its importance. In the realm of large-scale graphs, the single FPGA solution outperformed the GPU solution that were restricted by memory capacity and surpassing CPU speedup by 26x compared to 12x. Moreover, when a single FPGA device was limited due to the size of the graph, our performance model showed that a distributed system with multiple FPGAs could increase performance by around 12x. This highlights the effectiveness of our solution for handling large datasets that surpass on-chip memory restrictions.

Background and purpose: Physiological motion impacts the dose delivered to tumours and vital organs in external beam radiotherapy and particularly in particle therapy. The excellent soft-tissue demarcation of 4D magnetic resonance imaging (4D-MRI) could inform on intra-fractional motion, but long image reconstruction times hinder its use in online treatment adaptation. Here we employ techniques from high-performance computing to reduce 4D-MRI reconstruction times below two minutes to facilitate their use in MR-guided radiotherapy. Material and methods: Four patients with pancreatic adenocarcinoma were scanned with a radial stack-of-stars gradient echo sequence on a 1.5T MR-Linac. Fast parallelised open-source implementations of the extra-dimensional golden-angle radial sparse parallel algorithm were developed for central processing unit (CPU) and graphics processing unit (GPU) architectures. We assessed the impact of architecture, oversampling and respiratory binning strategy on 4D-MRI reconstruction time and compared images using the structural similarity (SSIM) index against a MATLAB reference implementation. Scaling and bottlenecks for the different architectures were studied using multi-GPU systems. Results: All reconstructed 4D-MRI were identical to the reference implementation (SSIM > 0.99). Images reconstructed with overlapping respiratory bins were sharper at the cost of longer reconstruction times. The CPU + GPU implementation was over 17 times faster than the reference implementation, reconstructing images in 60 ± 1 s and hyper-scaled using multiple GPUs. Conclusion: Respiratory-resolved 4D-MRI reconstruction times can be reduced using high-performance computing methods for online workflows in MR-guided radiotherapy with potential applications in particle therapy.

Processing large-scale graphs is challenging due to the nature of the computation that causes irregular memory access patterns. Managing such irregular accesses may cause significant performance degradation on both CPUs and GPUs. Thus, recent research trends propose graph processing acceleration with Field-Programmable Gate Arrays (FPGA). FPGAs are programmable hardware devices that can be fully customised to perform specific tasks in a highly parallel and efficient manner. However, FPGAs have a limited amount of on-chip memory that cannot fit the entire graph. Due to the limited device memory size, data needs to be repeatedly transferred to and from the FPGA on-chip memory, which makes data transfer time dominate over the computation time. A possible way to overcome the FPGA accelerators’ resource limitation is to engage a multi-FPGA distributed architecture and use an efficient partitioning scheme. Such a scheme aims to increase data locality and minimise communication between different partitions. This work proposes an FPGA processing engine that overlaps, hides and customises all data transfers so that the FPGA accelerator is fully utilised. This engine is integrated into a framework for using FPGA clusters and is able to use an offline partitioning method to facilitate the distribution of large-scale graphs. The proposed framework uses Hadoop at a higher level to map a graph to the underlying hardware platform. The higher layer of computation is responsible for gathering the blocks of data that have been pre-processed and stored on the host’s file system and distribute to a lower layer of computation made of FPGAs. We show how graph partitioning combined with an FPGA architecture will lead to high performance, even when the graph has Millions of vertices and Billions of edges. In the case of the PageRank algorithm, widely used for ranking the importance of nodes in a graph, compared to state-of-the-art CPU and GPU solutions, our implementation is the fastest, achieving a speedup of 13 compared to 8 and 3 respectively. Moreover, in the case of the large-scale graphs, the GPU solution fails due to memory limitations while the CPU solution achieves a speedup of 12 compared to the 26x achieved by our FPGA solution. Other state-of-the-art FPGA solutions are 28 times slower than our proposed solution. When the size of a graph limits the performance of a single FPGA device, our performance model shows that using multi-FPGAs in a distributed system can further improve the performance by about 12x. This highlights our implementation efficiency for large datasets not fitting in the on-chip memory of a hardware device.

We propose a design methodology to facilitate rigorous development of complex applications targeting reconfigurable hardware. Our methodology relies on analytical estimation of system performance and area utilisation for a given specific application and a particular system instance consisting of a controlflow machine working in conjunction with one or more reconfigurable dataflow accelerators. The targeted application is carefully analyzed, and the parts identified for hardware acceleration are reimplemented as a set of representative software models. Next, with the results of the application analysis, a suitable system architecture is devised and its performance is evaluated to determine bottlenecks, allowing predictable design. The architecture is iteratively refined, until the final version satisfying the specification requirements in terms of performance and required hardware area is obtained. We validate the presented methodology using a widely accepted convolutional neural network (VGG-16) and an important HPC application (BQCD). In both cases, our methodology relieved and alleviated all system bottlenecks before the hardware implementation was started. As a result the architectures were implemented first time right, achieving state-of-the-art performance within 15% of our modelling estimations.

Magnetic Resonance (MR)-guided online Adaptive RadioTherapy (MRgoART) utilises the excellent soft-tissue contrast of MR images taken just before the patient's treatment to quickly update and personalise radiotherapy treatment plans. Four-dimensional (4D) MR Imaging (MRI) can resolve variations in respiratory motion patterns. 4D MRI data can be used to adapt the radiation beams to maximally target the tumour while sparing as much healthy tissue as possible. 4D MRI reconstruction, however, is computationally challenging and current state-of-the-art implementations are unable to meet MRgoART time requirements. This study bridges the gap between high-performance computing and medical applications by developing and implementing a parallel, heterogeneous architecture for the XD-GRASP algorithm capable of meeting the MRgoART time requirements. Our architecture exploits long-vector instructions and utilises all available resources, while minimising and hiding the communication overhead when external GPUs are used. As a result, the reconstruction time was reduced from 994 seconds to just 90 seconds with a speedup of more than 11x. In addition, we evaluated the impact of the emerging Processing-in-Memory (PIM) technology. Our simulation results show that 16 low power, in-order PIM cores with no SIMD unit are 2.7x faster than an Intel Core™ i7-9700 8-core CPU equipped with AVX512 SIMD units. Additionally, 40 PIM cores match the performance of two AMD EPYC 7551 CPUs, with 32 cores each and just 87 PIM cores will match the performance of an NVIDIA Tesla V100 GPU equipped with 5,120 CUDA cores.

Field Programmable Gate Arrays (FPGAs) are gaining popularity in the context of scientific computing due to the recent advances of High-Level Synthesis (HLS) toolchains for customised hardware implementations combined with the increase in computing capabilities of modern FPGAs. As a result, developers are able to implement more complex scientific workloads which often require the evaluation of univariate numerical functions. In this study, we propose a methodology for table-based polynomial interpolation aiming at producing area-efficient implementations of such functions on FPGAs achieving the same accuracy and at similar performance as direct implementations. We also provide a rigorous error analysis to guarantee the correctness of the results. Our methodology covers the forecast of resource utilisation of the polynomial interpolator and, based on the characteristics of the function, guides the developer to the most area-efficient FPGA implementation. Our experiments show that in the case of a radiation spectrum of a Black Body application based on evaluating Planck's Law, it is possible to reduce resource utilisation by up to 90% when compared to direct implementations not using table-based methods. Moreover, when only the kernels are considered, our method uses up to two orders of magnitude fewer resources with no performance penalties. Based on previous more theoretical works, our study investigates practical applications of table-based methods in the context of high performance and scientific computing where it is used to implement common but more complex functions than the elementary functions widely studied in the related literature.

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.

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.

This paper proposes an algorithm for mapping logical to physical memory resources on FPGAs. Our greedy strategy based algorithm is specifically designed to facilitate timing closure on modern multi-die FPGAs for static-dataflow accelerators utilising most of the on-chip resources. The main objective of the proposed algorithm is to ensure that specific sub-parts of the design under consideration can fully reside within a single die to limit inter-die communication. The above is achieved by performing the memory mapping for each sub-part of the design separately while keeping allocation of the available physical resources balanced. As a result the number of inter-die connections is reduced on average by 50% compared to an algorithm targeting minimal area usage for real, complex applications using most of the on-chip's resources. Additionally, our algorithm is the only one out of the four evaluated approaches which successfully produces place and route results for all 33 applications and benchmarks.

In this paper we propose a technique to minimise the area overhead of a double buffered implementation of Radix-4 Fast Fourier Transformation (FFT). Our proposal circumvents the need for double buffering by exploiting opportunities in the specific data reordering of the buffers that are needed when implementing a fully pipelined FFT. By using the same read and write pattern, a single buffer is sufficient to perform data reordering while maintaining data integrity without degrading performance. We demonstrate this approach in an FPGA implementation. As a result of our optimisation, the memory depth can be reduced by a factor of two with very small overhead in control logic complexity.

In this paper we discuss a high performance implementation for Convolutional Neural Networks (CNNs) inference on the latest generation of Dataflow Engines (DFEs). We discuss the architectural choices made during the design phase taking into account the DFE chip properties. We then perform design space exploration, considering the memory bandwidth and resources utilisation constraints derived from the used DFE and the chosen architecture. Finally, we discuss the high performance implementation and compare the obtained performance against other implementations, showing that our proposed design reaches 2,450 GOPS when running VGG16 as a test case.

We demonstrate the use of dataflow technology in the computation of the correlation energy in molecules at the Møller-Plesset perturbation theory (MP2) level. Specifically, we benchmark density fitting (DF)-MP2 for as many as 168 atoms (in valinomycin) and show that speed-ups between 3 and 3.8 times can be achieved when compared to the MOLPRO package run on a single CPU. Acceleration is achieved by offloading the matrix multiplications steps in DF-MP2 to Dataflow Engines (DFEs). We project that the acceleration factor could be as much as 24 with the next generation of DFEs.

To handle the stringent performance requirements of future exascale-class applications, High Performance Computing (HPC) systems need ultra-efficient heterogeneous compute nodes. To reduce power and increase performance, such compute nodes will require hardware accelerators with a high degree of specialization. Ideally, dynamic reconfiguration will be an intrinsic feature, so that specific HPC application features can be optimally accelerated, even if they regularly change over time. In the EXTRA project, we create a new and flexible exploration platform for developing reconfigurable architectures, design tools and HPC applications with run-time reconfiguration built-in as a core fundamental feature instead of an add-on. EXTRA covers the entire stack from architecture up to the application, focusing on the fundamental building blocks for run-time reconfigurable exascale HPC systems: new chip architectures with very low reconfiguration overhead, new tools that truly take reconfiguration as a central design concept, and applications that are tuned to maximally benefit from the proposed run-time reconfiguration techniques. Ultimately, this open platform will improve Europe's competitive advantage and leadership in the field.

To handle the stringent performance requirements of future exascale-class applications, High Performance Computing (HPC) systems need ultra-efficient heterogeneous compute nodes. To reduce power and increase performance, such compute nodes will require hardware accelerators with a high degree of specialization. Ideally, dynamic reconfiguration will be an intrinsic feature, so that specific HPC application features can be optimally accelerated, even if they regularly change over time. In the EXTRA project, we create a new and flexible exploration platform for developing reconfigurable architectures, design tools and HPC applications with run-time reconfiguration built-in as a core fundamental feature instead of an add-on. EXTRA covers the entire stack from architecture up to the application, focusing on the fundamental building blocks for run-time reconfigurable exascale HPC systems: new chip architectures with very low reconfiguration overhead, new tools that truly take reconfiguration as a central design concept, and applications that are tuned to maximally benefit from the proposed run-time reconfiguration techniques. Ultimately, this open platform will improve Europe's competitive advantage and leadership in the field.

Even though FPGAs are becoming more and more popular as they are used in many different scenarios like communications and HPC, the steep learning curve needed to work with this technology is still the major limiting factor to their full success. Many works proposed to mitigate this problem by creating a companion of tools to support the designer during the development phase for this technology. The EU FASTER Project aims at realizing an integrated toolchain that assists the designer in the steps of the design flow that are necessary to port a given application onto an FPGA device. The novelty of the framework relies in the fact that the partial dynamic reconfiguration, which FPGA devices can exploit, is seen as a first class citizen throughout the whole design flow. This work reports a case study in which the FASTER toolchain has been used to port a raytracer application onto the STM Spear prototyping embedded platform. The paper discusses the steps done for the realization of the prototype and the results obtained on the target device. It finally reports some improvements that can be exploited to improve the performance of the hardware implementation that has been realized.

While fine-grain, reconfigurable devices have been available for years, they are mostly used in a fixed functionality, "asic-replacement" manner. To exploit opportunities for flexible and adaptable run-time exploitation of fine grain reconfigurable resources (as implemented currently in dynamic, partial reconfiguration), better tool support is needed. The FASTER project aims to provide a methodology and a tool-chain that will enable designers to efficiently implement a reconfigurable system on a platform combining software and reconfigurable resources. Starting from a high-level application description and a target platform, our tools analyse the application, evaluate reconfiguration options, and implement the designer choices on underlying vendor tools. In addition, FASTER addresses micro-reconfiguration, verification, and the run-time management of system resources. We use industrial applications to demonstrate the effectiveness of the proposed framework and identify new opportunities for reconfigurable technologies.

The FASTER project aims to ease the definition, implementation and use of dynamically changing hardware systems. Our motivation stems from the promise reconfigurable systems hold for achieving better performance and extending product functionality and lifetime via the addition of new features that work at hardware speed. This is a clear advantage over the more straightforward software component adaptivity. However, designing a changing hardware system is both challenging and time consuming. The FASTER project will facilitate the use of reconfigurable technology by providing a complete methodology that enables designers to easily specify, analyse, implement and verify applications on platforms with general-purpose processors and acceleration modules implemented in the latest reconfigurable technology. To better adapt to different application requirements, the tool-chain will support both region-based and micro-reconfiguration and provide a flexible run-time system that will efficiently manage the reconfigurable resources. We will use applications from the embedded, high performance computing, and desktop domains to demonstrate the potential benefits of the FASTER tools on metrics such as performance, power consumption and total ownership cost.