Parallel Acceleration Research Articles

Edge Real-Time Object Detection and DPU-Based Hardware Implementation for Optical Remote Sensing Images

The accuracy of current deep learning algorithms has certainly increased. However, deploying deep learning networks on edge devices with limited resources is challenging due to their inherent depth and high parameter count. Here, we proposed an improved YOLO model based on an attention mechanism and receptive field (RFA-YOLO) model, applying the MobileNeXt network as the backbone to reduce parameters and complexity, adopting the Receptive Field Block (RFB) and Efficient Channel Attention (ECA) modules to improve the detection accuracy of multi-scale and small objects. Meanwhile, an FPGA-based model deployment solution was proposed to implement parallel acceleration and low-power deployment of the detection algorithm model, which achieved real-time object detection for optical remote sensing images. We implement the proposed DPU and Vitis AI-based object detection algorithms with FPGA deployment to achieve low power consumption and real-time performance requirements. Experimental results on DIOR dataset demonstrate the effectiveness and superiority of our RFA-YOLO model for object detection algorithms. Moreover, to evaluate the performance of the proposed hardware implementation, it was implemented on a Xilinx ZCU104 board. Results of the experiments for hardware and software simulation show that our DPU-based hardware implementation are more power efficient than central processing units (CPUs) and graphics processing units (GPUs), and have the potential to be applied to onboard processing systems with limited resources and power consumption.

A GPU‐based real‐time processing system for frequency division multiple‐input‐multiple‐output radar

AbstractMultiple‐Input‐Multiple‐Output (MIMO) radar has the characteristic of multiple antenna channels, bringing on a big data volume for signal processing. Therefore, the trade‐off between detection ability and computational efficiency is always considered. In this article, an optimisation system is proposed to enhance the real‐time performance of frequency division MIMO radar without compromising accuracy. For the scenario of low‐altitude small target detection, a signal processing acceleration method is proposed and a MIMO radar optimisation system based on graphics processing unit (GPU) architecture is built. The signal model of frequency division MIMO radar is first established, improving the classical signal processing flow efficiency from the perspective of reducing fast Fourier transform (FFT) times and windowing operation times. To achieve an advanced acceleration, the parallel architecture of ArrayFire‐library in GPU is then employed. Distinct minimum parallel units are extracted by analysing the principle of digital beamforming (DBF), pulse compression, moving target detection (MTD), and constant false‐alarm rate (CFAR) algorithms. And the corresponding parallel algorithms are designed to constitute a parallel acceleration system of frequency division MIMO. Simulation results indicate that the proposed method significantly improves the efficiency of MIMO system with a maximum acceleration ratio of 65 times, meeting the real‐time processing requirements.

10-channel phased-array coil for carotid wall MRI at 3T.

Accurate assessment of plaque accumulation near the carotid bifurcation is important for the effective prevention and treatment of stroke. However, vessel and plaque delineation using MRI can be limited by low contrast-to-noise ratio (CNR) and long acquisition times. In this work, a 10-channel phased-array receive coil design for bilateral imaging of the carotid bifurcation using 3T MRI is proposed. The proposed 10-channel receive coil was compared to a commercial 4-channel receive coil configuration using data acquired from phantoms and healthy volunteers (N = 9). The relative performance of the coils was assessed, by comparing signal-to-noise ratio (SNR), noise correlation, g-factor noise amplification, and the CNR between vessel wall and lumen using black-blood sequences. Patient data were acquired from 12 atherosclerotic carotid artery disease patients. The 10-channel coil consistently provided substantially increased SNR in phantoms (+77 ± 27%) and improved CNR in healthy carotid arteries (+62 ± 11%), or reduced g-factor noise amplification. Patient data showed excellent delineation of atherosclerotic plaque along the length of the carotid bifurcation using the 10-channel coil. The proposed 10-channel coil design allows for improved visualization of the carotid arteries and the carotid bifurcation and increased parallel imaging acceleration factors relative to a commercial 4-channel coil design.

Acceleration and implementation of convolutional neural networks based on FPGA

Target detection algorithms are an important technology in the field of computer vision and usually need to be implemented on devices such as CPU or GPU with high computational power. However, traditional CPU are unable to support real-time processing and computations; the power consumption of GPU is relatively high and the application scenarios are limited. FPGA, with their high throughput rate, high power efficiency and reconfigurability, are the best platform for research on hardware acceleration of target detection algorithms. Therefore, how to implement efficient target detection algorithms on embedded devices is an important challenge. A method for YOLOv4-Tiny object detection algorithm accelerator based on FPGA is proposed and demonstrated in this paper. In this design, an optimization strategy is proposed, including the design of IP cores for convolutional computation in HLS, the fusion of batch normalization layers with convolutional layers, dynamic fixed-point 16-bit quantization, loop unrolling double-buffered storage and channel parallel acceleration, which can improve the computational performance and efficiency of the target detection algorithm. The YOLOv4-Tiny target detection algorithm was deployed on an FPGA development board, and experimental results showed that it achieved a computational performance of 18.32 GOP/s and energy efficiency of 6.66 GOP-W/s on the FPGA, an increase of 1.81 times in computational performance and 1.86 times in energy efficiency compared to the same platform. Compared to unquantized CPU platforms, the computational performance is 0.51 times better, the energy efficiency is 34 times better, and the latency of accelerating a single image is reduced to 383 ms, which allows for better detection performance in the field of target detection and tracking.

Efficient SNN multi-cores MAC array acceleration on SpiNNaker 2.

The potential low-energy feature of the spiking neural network (SNN) engages the attention of the AI community. Only CPU-involved SNN processing inevitably results in an inherently long temporal span in the cases of large models and massive datasets. This study introduces the MAC array, a parallel architecture on each processing element (PE) of SpiNNaker 2, into the computational process of SNN inference. Based on the work of single-core optimization algorithms, we investigate the parallel acceleration algorithms for collaborating with multi-core MAC arrays. The proposed Echelon Reorder model information densification algorithm, along with the adapted multi-core two-stage splitting and authorization deployment strategies, achieves efficient spatio-temporal load balancing and optimization performance. We evaluate the performance by benchmarking a wide range of constructed SNN models to research on the influence degree of different factors. We also benchmark with two actual SNN models (the gesture recognition model of the real-world application and balanced random cortex-like network from neuroscience) on the neuromorphic multi-core hardware SpiNNaker 2. The echelon optimization algorithm with mixed processors realizes 74.28% and 85.78% memory footprint of the original MAC calculation on these two models, respectively. The execution time of echelon algorithms using only MAC or mixed processors accounts for ≤ 24.56% of the serial ARM baseline. Accelerating SNN inference with algorithms in this study is essentially the general sparse matrix-matrix multiplication (SpGEMM) problem. This article explicitly expands the application field of the SpGEMM issue to SNN, developing novel SpGEMM optimization algorithms fitting the SNN feature and MAC array.

Creative destruction in academia: a time to reimagine practices in alignment with sustainability values

AbstractAcademia has experienced acceleration and expansion in parallel with the Great Acceleration, which has shaped the Anthropocene. Among other pressures, the expectation to be internationally mobile conflicts with many values held by sustainability scholars and results in disillusionment. The changes in the academic system can be seen through the framework of the adaptive cycle, which can help us understand historical parallels and shape the system to better align with sustainability values in future. We hope this piece can contribute to the discussion of the next steps forward to reimagine academia.

Parallel photonic acceleration processor for matrix-matrix multiplication.

We propose and experimentally demonstrate a highly parallel photonic acceleration processor based on a wavelength division multiplexing (WDM) system and a non-coherent Mach-Zehnder interferometer (MZI) array for matrix-matrix multiplication. The dimensional expansion is achieved by WDM devices, which play a crucial role in realizing matrix-matrix multiplication together with the broadband characteristics of an MZI. We implemented a 2 × 2 arbitrary nonnegative valued matrix using a reconfigurable 8 × 8 MZI array structure. Through experimentation, we verified that this structure could achieve 90.5% inference accuracy in a classification task for the Modified National Institute of Standards and Technology (MNIST) handwritten dataset. This provides a new effective solution for large-scale integrated optical computing systems based on convolution acceleration processors.

Design of GPU Network-on-Chip for Real-Time Video Super-Resolution Reconstruction.

Deep learning has a better output quality compared with traditional algorithms for video super-resolution (SR), but the network model needs large resources and has poor real-time performance. This paper focuses on solving the speed problem of SR; it achieves real-time SR by the collaborative design of a deep learning video SR algorithm and GPU parallel acceleration. An algorithm combining deep learning networks with a lookup table (LUT) is proposed for the video SR, which ensures both the SR effect and ease of GPU parallel acceleration. The computational efficiency of the GPU network-on-chip algorithm is improved to ensure real-time performance by three major GPU optimization strategies: storage access optimization, conditional branching function optimization, and threading optimization. Finally, the network-on-chip was implemented on a RTX 3090 GPU, and the validity of the algorithm was demonstrated through ablation experiments. In addition, SR performance is compared with existing classical algorithms based on standard datasets. The new algorithm was found to be more efficient than the SR-LUT algorithm. The average PSNR was 0.61 dB higher than the SR-LUT-V algorithm and 0.24 dB higher than the SR-LUT-S algorithm. At the same time, the speed of real video SR was tested. For a real video with a resolution of 540×540, the proposed GPU network-on-chip achieved a speed of 42 FPS. The new method is 9.1 times faster than the original SR-LUT-S fast method, which was directly imported into the GPU for processing.

ERStruct: a fast Python package for inferring the number of top principal components from whole genome sequencing data

BackgroundLarge-scale multi-ethnic DNA sequencing data is increasingly available owing to decreasing cost of modern sequencing technologies. Inference of the population structure with such sequencing data is fundamentally important. However, the ultra-dimensionality and complicated linkage disequilibrium patterns across the whole genome make it challenging to infer population structure using traditional principal component analysis based methods and software.ResultsWe present the ERStruct Python Package, which enables the inference of population structure using whole-genome sequencing data. By leveraging parallel computing and GPU acceleration, our package achieves significant improvements in the speed of matrix operations for large-scale data. Additionally, our package features adaptive data splitting capabilities to facilitate computation on GPUs with limited memory.ConclusionOur Python package ERStruct is an efficient and user-friendly tool for estimating the number of top informative principal components that capture population structure from whole genome sequencing data.

5T magnetic resonance imaging: radio frequency hardware and initial brain imaging.

We aimed to demonstrate the feasibility of generating high-resolution human brain magnetic resonance imaging (MRI) at 5 Tesla (T) using a quadrature birdcage transmit/48-channel receiver coil assembly. A quadrature birdcage transmit/48-channel receiver coil assembly was designed for human brain imaging at 5T. The radio frequency (RF) coil assembly was validated by electromagnetic (EM) simulations and phantom imaging experimental studies. The simulated B1+ field inside a human head phantom and inside a human head model generated by the birdcage coils driven in circularly polarized (CP) mode at 3T, 5T and 7T was compared. Signal-to-noise ratio (SNR) maps, the inverse g-factor maps for evaluation of parallel imaging performance, anatomic images, angiography images, vessel wall images and susceptibility weighted images (SWI) were acquired using the RF coil assembly at 5T and compared to those acquired using a 32-channel head coil on a 3T MRI scanner. For the EM simulations, 5T MRI provided less RF inhomogeneity compared to that of 7T. In the phantom imaging study, the distributions of the measured B1+ field were consistent with the distributions of the simulated B1+ field. In the human brain imaging study, the average SNR value of the brain in the transversal plane at 5T was 1.6 times of that at 3T. The 48-channel head coil at 5T had higher parallel acceleration capability than the 32-channel head coil at 3T. The anatomic images at 5T also showed higher SNR than those at 3T. Improved delineation of the hippocampus, lenticulostriate arteries, and basilar arteries was observed at 5T compared to 3T. SWI with a higher resolution of 0.3 mm ×0.3 mm ×1.2 mm could be acquired at 5T, which enabled better visualization of small blood vessels compared to that at 3T. 5T MRI can provide significant SNR improvement compared to that of 3T with less RF inhomogeneity than that of 7T. The ability to obtain high quality in vivo human brain images at 5T using the quadrature birdcage transmit/48-channel receiver coil assembly has significant in clinical and scientific research applications.

FPGA-based parallel implementation to classify Hyperspectral images by using a Convolutional Neural Network

Thanks to its richness in extractable features, Hyperspectral images (HSI) find an accelerated use in medical, industrial, agricultural, and environmental fields. In this paper, we present a wavelet-based reduction technique that creates a Hypercube containing the most significant features extracted from the original HSI and representing a multi-dimensional array that is utilized for training a Convolutional Neural Network (CNN), which is designed here to classify different types of surfaces or materials. The performance of this approach is tested and proved using two distinct datasets. Then, we compare the same approach with the PCA, a widely used reduction method. The most important contribution of this paper is the implementation of an FPGA-based parallel accelerator to train the same suggested CNN in only 10% of the computational time compared to the classical CPU-based techniques. The Microblaze will be explained and exploited here to play the role of an embedded microprocessor.

Dammann gratings-based truly parallel optical matrix multiplication accelerator.

Matrix multiplication (MM) is a fundamental operation in various scientific and engineering computations, as well as in artificial intelligence algorithms. Efficient implementation of MM is crucial for speeding up numerous applications. Photonics presents an opportunity for efficient acceleration of dense matrix computation, owing to its intrinsic advantages, such as huge parallelism, low latency, and low power consumption. However, most optical matrix computing architectures have been limited to realizing single-channel vector-matrix multiplication or using complex configurations to expand the number of channels, which does not fully exploit the parallelism of optics. In this study, we propose a novel, to the best of our knowledge, scheme for the implementation of large-scale two-dimensional optical MM with truly massive parallelism based on a specially designed Dammann grating. We demonstrate a sequence of MMs of 50 pairs of randomly generated 4 × 8 and 8 × 4 matrices in our proof-of-principle experiment. The results indicate that the mean relative error is approximately 0.048, thereby demonstrating optical robustness and high accuracy.

ACCELERATED PARALLEL MRI USING MEMORY EFFICIENT AND ROBUST MONOTONE OPERATOR LEARNING (MOL).

Model-based deep learning methods that combine imaging physics with learned regularization priors have been emerging as powerful tools for parallel MRI acceleration. The main focus of this paper is to determine the utility of the monotone operator learning (MOL) framework in the parallel MRI setting. The MOL algorithm alternates between a gradient descent step using a monotone convolutional neural network (CNN) and a conjugate gradient algorithm to encourage data consistency. The benefits of this approach include similar guarantees as compressive sensing algorithms including uniqueness, convergence, and stability, while being significantly more memory efficient than unrolled methods. We validate the proposed scheme by comparing it with different unrolled algorithms in the context of accelerated parallel MRI for static and dynamic settings.

Power System Stability Estimation Scheme Based on Big Data

Aiming at the poor convergence of power flow calculation and the difficulty in location of the power system operation, this paper proposes a fast parallel calculation scheme for power system stability assessment based on big data technology. Through the analysis and research of Newton method which is sensitive to the initial value selection in the actual calculation of regional power flow, an approach for selecting the optimal initial value is proposed by genetic algorithm. Then, by the state sampling method and Monte Carlo parallel simulation, the coarse granularity algorithm under the message passing programming model is designed. The experimental analysis adopts the improved method to test the data in IEEE 9 system, actual environment, the automatic power flow adjustment. The results show that our scheme can achieve high parallel efficiency and acceleration ratio, which improves the calculation ability and accuracy of power flow and reliability evaluation of large-scale power systems.

Synchronization of parallel intense electron beam accelerators based on single trigger generator

In this study, the authors provide results of the precisely synchronized triggering of an intense electron beam accelerator (IEBA). The trigger generator was composed of a fractional-turn ratio saturable-pulse transformer and a compact six-stage Marx generator. The main switch of the IEBA was a corona-stabilized triggered switch (CSTS) based on the stabilized corona mechanism. The output voltage of a single IEBA exceeded 500 kV with less than 4 ns of jitter on a 50 Ω dummy load. We also conducted an experiment on the synchronous triggering of two IEBAs by using two independent trigger generators. The synchronization-related jitter was 6.1 ns, while the average time difference was 1.3 ns. We used this to attempt to trigger two parallel IEBAs by using a single trigger generator. The results showed a reduction in the synchronization-related jitter of 31% to 4.2 ns. The designs of the CSTS and the trigger generator guaranteed a precisely synchronous trigger by using a single trigger generator. Thus, the proposed method appears to be promising for accurately and synchronously triggering multiple IEBAs by using a single trigger generator. This provides an effective method to generate pulses with high power.

3D least-squares reverse time migration in VTI media based on pseudoacoustic wave equation and multi-GPU parallel acceleration

As the high-density seismic acquisition is progressively becoming an essential mode of seismic exploration, developing high-precision imaging methods for 3D datasets has become an urgent industry requirement. Unfortunately, the two-dimensional least-squares reverse time migration (LSRTM) can hardly deal with the influence of the third dimension because the structure of reflectors is constantly changing in nearly all directions. Furthermore, anisotropy is very common in underground media, which makes the LSRTM based on isotropic assumption fail to locate the reflectors accurately and seriously limits the identification of geological stratification. Considering the above two problems, this paper derives the least-squares reverse time migration in three-dimensional vertical transversely isotropic media (3D-VTI-LSRTM) and develops a workflow which can realize 3D-LSRTM in VTI media based on this theory. Our workflow uses multi-GPU acceleration and heterogeneous CPU/GPU parallelism. Compared with the general parallel CPU workflow or single GPU workflow, it is feasible that the GPU application significantly improves computing efficiency in practical applications. In addition, this paper extends VTI-LSRTM from 2D to 3D, which solves the problem that the 2D algorithm cannot eliminate the lateral direction reflection artifacts. For the self-made 3D-VTI depression model and the 3D-VTI-SEG/EAGE salt model, we use four NVIDIA Tesla k40c graphics cards for testing. The results of the tests are correct and significantly improved compared to the 3D-VTI-RTM method.

Automatic Creation of High-bandwidth Memory Architectures from Domain-specific Languages: The Case of Computational Fluid Dynamics

Numerical simulations can help solve complex problems. Most of these algorithms are massively parallel and thus good candidates for FPGA acceleration thanks to spatial parallelism. Modern FPGA devices can leverage high-bandwidth memory technologies, but when applications are memory-bound designers must craft advanced communication and memory architectures for efficient data movement and on-chip storage. This development process requires hardware design skills that are uncommon in domain-specific experts. In this article, we propose an automated tool flow from a domain-specific language for tensor expressions to generate massively parallel accelerators on high-bandwidth-memory-equipped FPGAs. Designers can use this flow to integrate and evaluate various compiler or hardware optimizations. We use computational fluid dynamics (CFD) as a paradigmatic example. Our flow starts from the high-level specification of tensor operations and combines a multi-level intermediate representation–based compiler with an in-house hardware generation flow to generate systems with parallel accelerators and a specialized memory architecture that moves data efficiently, aiming at fully exploiting the available CPU-FPGA bandwidth. We simulated applications with millions of elements, achieving up to 103 GFLOPS with one compute unit and custom precision when targeting a Xilinx Alveo U280. Our FPGA implementation is up to 25× more energy efficient than expert-crafted Intel CPU implementations.

Hybrid quantum–classical convolutional neural networks with privacy quantum computing

Machine learning algorithms help us discover knowledge from big data. Data used for training or prediction often contain private information about users. Discovering knowledge while protecting data or user privacy is the way machine learning is expected, especially in the cloud environment. Quantum machine learning is a kind of machine learning that realizes parallel acceleration by quantum superposition. Quantum computing power for quantum machine learning is typically provided by quantum cloud computing services. Existing quantum machine learning algorithms hardly consider privacy protection. This paper presents an encryption method for image data which can effectively protect the input data privacy in hybrid quantum–classical convolutional neural networks algorithm. The user’s original image data is first encrypted, and then sent to the quantum cloud to calculate the image convolution. By doing so, the feature map of the ciphertext image is obtained by the user. The result obtained by decrypting the feature map is the same as that obtained by using the original image as the input of convolution calculation. Experiments show that our privacy protection scheme can protect the privacy of input image data in the hybrid quantum–classical neural networks algorithm, but does not affect the accuracy of the algorithm. In addition to image encryption and feature map decryption, the proposed scheme does not bring additional computational complexity.

Research on elaborate image simulation method for close-range space target

In order to meet the urgent needs of target image data in the testing of close-range space target situational awareness algorithm, an elaborate image simulation method that fully considers the target surface texture structure is proposed. First, the physical mechanism is analysed by constructing a radiation transmission model. Then, Phong and five-parameters BRDF models are adopted to characterize the optical properties of the material. The geometric characteristic of texture structure is also characterized through displacement mapping. Finally, the sequence images of the space target are obtained by simulation based on the designed simulation process and given imaging conditions. The CPU multi-core parallel acceleration based on the Embree engine is taken to compensate for the speed reduction caused by the increase in the number of model faces. Compared with the measured image in orbit for verification, the image simulation similarity is above 80%. The results show that the simulation method in this article can better describe the characteristics of the target in the actual on-orbit environment and meet the near real-time simulation requirements of the system, providing a more realistic image data source for the ground test of the on-orbit service technology.

Programming for High-Performance Computing on Edge Accelerators

The field of edge computing has grown considerably over the past few years, with applications in artificial intelligence and big data processing, particularly due to its powerful accelerators offering a large amount of hardware parallelism. As the computing power of the latest edge systems increases, applications of edge computing are being expanded to areas that have traditionally required substantially high-performant computing resources such as scientific computing. In this paper, we review the latest literature and present the current status of research for implementing high-performance computing (HPC) on edge devices equipped with parallel accelerators, focusing on software environments including programming models and benchmark methods. We also examine the applicability of existing approaches and discuss possible improvements necessary towards realizing HPC on modern edge systems.