GPGPU Research Articles

Variation aware power management for GPU memories

In the midst of progressively shrinking silicon technologies, enhancements in performance and area requirements come at the cost of some side effects. One such cost is the deviation of nominal parameters of process, voltage, and temperature. Considered a manufacturing defect, variations limit the maximum achievable performance. These affected areas possess lower reliability and consume more power than their theoretical counterpart.In this paper, we study and mitigate variations observed in GDDRx based GPU memories. GPU devices are exploited primarily for their massive parallelism and applications with both regular and irregular memory accesses see significant performance benefits. We also find that these consume a significant portion of the total GPU power. We propose a variation mitigating and power-saving technique for GPU memories. This accounts for a power savings of up to 44.61% (20.3% on average). Simultaneously, we also maintain the device data which prevents page faults and re-computation costs. This, however, leads to some performance overheads, which are limited to 4%. Mitigation of process variation combined with state preservation leads to more reliable GPU computations. Additionally, our model lowers access latencies by 15.7% in comparison to a variation affected baseline GPU, which ultimately helps to improve the throughput of the device.

Multi-Gbps LDPC Decoder on GPU Devices

To meet the high throughput requirement of communication systems, the design of high-throughput low-density parity-check (LDPC) decoders has attracted significant attention. This paper proposes a high-throughput GPU-based LDPC decoder, aiming at the large-scale data process scenario, which optimizes the decoder from the perspectives of the decoding parallelism and data scheduling strategy, respectively. For decoding parallelism, the intra-codeword parallelism is fully exploited by combining the characteristics of the flooding-based decoding algorithm and GPU programming model, and the inter-codeword parallelism is improved using the single-instruction multiple-data (SIMD) instructions. For the data scheduling strategy, the utilization of off-chip memory is optimized to satisfy the demands of large-scale data processing. The experimental results demonstrate that the decoder achieves 10 Gbps throughput by incorporating the early termination mechanism on general-purpose GPU (GPGPU) devices and can also achieve a high-throughput and high-power-efficiency performance on low-power embedded GPU (EGPU) devices. Compared with the state-of-the-art work, the proposed decoder had a ×1.787 normalized throughput speedup at the same error correcting performance.

Development of a three-dimensional grain-based combined finite-discrete element method to model the failure process of fine-grained sandstones

This paper introduces a novel three-dimensional grain-based hybrid finite-discrete element method parallelized based on general-purpose graphic processing units (GPGPUs) and applies it to investigate the failure process of sedimentary fine-grained sandstones. The grain-based method considers the actual microstructures of rocks with Voronoi or grain-growth tessellations to model their failures, including transgranular, intergranular and intragranular crack propagations. The novel semi-adaptive contact activation approach (semi-ACAA) proposed by the authors and the efficient tetrahedron-to-point (TtoP) contact interaction algorithm developed by HOSS are implemented to speed up grain-based modelling in addition to GPGPU parallelization. Semi-ACAA and TtoP are approximately 2–20 times and 1.5 times faster than the brute-force contact activation approach and tetrahedron-to-triangle contact interaction algorithm, respectively, which are prevalent in the FDEM community. The grain-based modelling elucidates that most intragranular cracks are observed in the medium- and high-strength grains of the sandstones, while most intergranular cracks occur at the grain boundaries between low- and high-strength grains. The transgranular cracks do not discriminate any grains on their propagating paths but propagate and coalesce with intragranular and intergranular cracks, which results in the final failure of the sandstones.

Deep Learning for Clothing Style Recognition Using YOLOv5.

With the rapid development of artificial intelligence, much more attention has been paid to deep learning. However, as the complexity of learning algorithms increases, the needs of computation power of hardware facilities become more crucial. Instead of the focus being on computing devices like GPU computers, a lightweight learning algorithm could be the answer for this problem. Cross-domain applications of deep learning have attracted great interest amongst researchers in academia and industries. For beginners who do not have enough support with software and hardware, an open-source development environment is very helpful. In this paper, a relatively lightweight algorithm YOLOv5s is addressed, and the Google Colab is used for model training and testing. Based on the developed environment, many state-of-art learning algorithms can be studied for performance comparisons. To highlight the benefits of one-stage object detection algorithms, the recognition of clothing styles is investigated. The image samples are selected from datasets of fashion clothes and the web crawling of online stores. The image data are categorized into five groups: plaid; plain; block; horizontal; and vertical. Average precison, mean average precison, recall, F1-score, model size, and frame per second are the metrics used for performance validations. From the experimental outcomes, it shows that YOLOv5s is better than other learning algorithms in the recognition accuracy and detection speed.

A Human-Detection Method Based on YOLOv5 and Transfer Learning Using Thermal Image Data from UAV Perspective for Surveillance System

At this time, many illegal activities are being been carried out, such as illegal mining, hunting, logging, and forest burning. These things can have a substantial negative impact on the environment. These illegal activities are increasingly rampant because of the limited number of officers and the high cost required to monitor them. One possible solution is to create a surveillance system that utilizes artificial intelligence to monitor the area. Unmanned aerial vehicles (UAV) and NVIDIA Jetson modules (general-purpose GPUs) can be inexpensive and efficient because they use few resources. The problem from the object-detection field utilizing the drone’s perspective is that the objects are relatively small compared to the observation space, and there are also illumination and environmental challenges. In this study, we will demonstrate the use of the state-of-the-art object-detection method you only look once (YOLO) v5 using a dataset of visual images taken from a UAV (RGB-image), along with thermal infrared information (TIR), to find poachers. There are seven scenario training methods that we have employed in this research with RGB and thermal infrared data to find the best model that we will deploy on the Jetson Nano module later. The experimental result shows that a new model with pre-trained model transfer learning from the MS COCO dataset can improve YOLOv5 to detect the human–object in the RGBT image dataset.

Probabilistically Certified Management of Data Centers Using Predictive Control

Data centers are facilities with large number of servers providing cloud services. The increasing number of data centers in the last years has generated environmental concern due to the large amount of energy consumed by them. This also includes some auxiliary services such as the cooling equipment which is known to be very costly. For that reason, efficient data center strategies are needed in order to provide an acceptable quality of service (QoS) and suitable temperature for every server while using the least amount of resources possible. This article presents some strategies to deal with the unified workload and temperature problem that appears in the data center. As the system is modeled as a queue and the control variables have an hybrid nature, some highly parallelizable particle-based optimization algorithms are proposed to solve the optimization problem. Numerical simulations are provided in order to illustrate the effectiveness of the strategy. These simulations also show the improvements obtained from the GPU computing. Finally, a probabilistic evaluation approach is developed in order to provide certificates on the probability of constraint satisfaction without increasing the computational burden of the online problem. Note to Practitioners—This article addresses the problem of deciding in real-time the number of active servers in a data center that is required to meet the quality of service (QoS) demands while keeping energy consumption at a minimum. The temperature set point of the cooling equipment must also be taken into account, as it is advisable to use the minimum cooling that keeps the servers running in safe conditions. The management strategy proposed is based on predictive control. In this way, the number of active servers and temperature set point will be chosen so that the future energy consumption is minimized while guaranteeing that QoS and safety demands are met under different possible operating conditions. Furthermore, the proposed management strategy can be tuned depending on the QoS that it is desirable to provide. The proposed strategy will lead to energy-consumption improvements while having guarantees on the data center performance.

Toward Energy-Efficient Federated Learning Over 5G+ Mobile Devices

The continuous convergence of machine learning algorithms, 5G and beyond (5G+) wireless communications, and artificial intelligence (AI) hardware implementation hastens the birth of federated learning (FL) over 5G+ mobile devices, which pushes AI functions to mobile devices and initiates a new era of on-device AI applications. Despite the remarkable progress made in FL, huge energy consumption is one of the most significant obstacles restricting the development of FL over battery-constrained 5G+ mobile devices. To address this issue, in this article, we investigate how to develop energy- efficient FL over 5G+ mobile devices by making a trade-off between energy consumption for “working” (i.e., local computing) and that for “talking” (i.e., wireless communications) in order to boost the overall energy efficiency. Specifically, we first examine energy consumption models for GPU computation and wireless transmissions. Then, we overview the state of the art of integrating FL procedure with energy-efficient learning techniques (gradient sparsification, weight quantization, pruning, etc.). Finally, we present several potential future research directions for FL over 5G+ mobile devices from the perspective of energy efficiency.

Iterative hybrid regularization for extremely noisy full models in single particle analysis

Cryo-electron microscopy single particle analysis (SPA) represents a vital tool for structure determination of macromolecules. Discrete inverse problems arising in this field are extremely large and seriously contaminated by noise. The model matrix is highly structured and can not be stored explicitly. Iterative regularization methods are used here only rarely, since it is believed that they are computationally expensive and require complicated stopping criteria. In this paper, we overcome these difficulties and demonstrate that Hybrid Krylov subspace methods can be used to solve SPA inverse problems efficiently. We propose an application-driven regularization parameter selection approach and present a matrix-free implementation of the hybrid solver for GPU computations in single precision arithmetic. In comparison to Fourier-based techniques, the hybrid approach allows to compute reconstructions from full (nonreduced) SPA models even for highly noisy data sets.

Herbal Compound Screening with GPU Computation on ZINC Database through Similarity Comparison Approach

Covid-19 is a global pandemic that drives many researchers to strive to look for its solution, especially in the field of health, medicine, and total countermeasures. Early screening with in-silico processes is crucial to minimize the search space of the potential drugs to cure a disease. This research aims to find potential drugs of covid-19 disease in the ZINC database to be further investigated through the in-vitro method. About 997.402.117 chemical compounds are searched about their similarity to some of the confirmed drugs to combat coronavirus. Sequential computation would take months to accomplish this task. The general programming graphic processing unit approach is used to implement a similarity comparison algorithm in parallel, in order to speed up the process. The result of this study shows the parallel algorithm implementation can speed up the computation process up to 55 times faster, and also that some of the chemical compounds have high similarity scores and can be found in nature

Differentiable Learning of Sequence-Specific Minimizer Schemes with DeepMinimizer.

Minimizers are widely used to sample representative k-mers from biological sequences in many applications, such as read mapping and taxonomy prediction. In most scenarios, having the minimizer scheme select as few k-mer positions as possible (i.e., having a low density) is desirable to reduce computation and memory cost. Despite the growing interest in minimizers, learning an effective scheme with optimal density is still an open question, as it requires solving an apparently challenging discrete optimization problem on the permutation space of k-mer orderings. Most existing schemes are designed to work well in expectation over random sequences, which have limited applicability to many practical tools. On the other hand, several methods have been proposed to construct minimizer schemes for a specific target sequence. These methods, however, only approximate the original objective with likewise discrete surrogate tasks that are not able to significantly improve the density performance. This article introduces the first continuous relaxation of the density minimizing objective, DeepMinimizer, which employs a novel Deep Learning twin architecture to simultaneously ensure both validity and performance of the minimizer scheme. Our surrogate objective is fully differentiable and, therefore, amenable to efficient gradient-based optimization using GPU computing. Finally, we demonstrate that DeepMinimizer discovers minimizer schemes that significantly outperform state-of-the-art constructions on human genomic sequences.

An LED-Based structured illumination microscope using a digital micromirror device and GPU accelerated image reconstruction.

When combined with computational approaches, fluorescence imaging becomes one of the most powerful tools in biomedical research. It is possible to achieve resolution figures beyond the diffraction limit, and improve the performance and flexibility of high-resolution imaging systems with techniques such as structured illumination microscopy (SIM) reconstruction. In this study, the hardware and software implementation of an LED-based super-resolution imaging system using SIM employing GPU accelerated parallel image reconstruction is presented. The sample is illuminated with two-dimensional sinusoidal patterns with various orientations and lateral phase shifts generated using a digital micromirror device (DMD). SIM reconstruction is carried out in frequency space using parallel CUDA kernel functions. Furthermore, a general purpose toolbox for the parallel image reconstruction algorithm and an infrastructure that allows all users to perform parallel operations on images without developing any CUDA kernel code is presented. The developed image reconstruction algorithm was run separately on a CPU and a GPU. Two different SIM reconstruction algorithms have been developed for the CPU as mono-thread CPU algorithm and multi-thread OpenMP CPU algorithm. SIM reconstruction of 1024 × 1024 px images was achieved in 1.49 s using GPU computation, indicating an enhancement by ∼28 and ∼20 in computation time when compared with mono-thread CPU computation and multi-thread OpenMP CPU computation, respectively.

Monte Carlo Simulation of Stochastic Differential Equation to Study Information Geometry.

Information Geometry is a useful tool to study and compare the solutions of a Stochastic Differential Equations (SDEs) for non-equilibrium systems. As an alternative method to solving the Fokker–Planck equation, we propose a new method to calculate time-dependent probability density functions (PDFs) and to study Information Geometry using Monte Carlo (MC) simulation of SDEs. Specifically, we develop a new MC SDE method to overcome the challenges in calculating a time-dependent PDF and information geometric diagnostics and to speed up simulations by utilizing GPU computing. Using MC SDE simulations, we reproduce Information Geometric scaling relations found from the Fokker–Planck method for the case of a stochastic process with linear and cubic damping terms. We showcase the advantage of MC SDE simulation over FPE solvers by calculating unequal time joint PDFs. For the linear process with a linear damping force, joint PDF is found to be a Gaussian. In contrast, for the cubic process with a cubic damping force, joint PDF exhibits a bimodal structure, even in a stationary state. This suggests a finite memory time induced by a nonlinear force. Furthermore, several power-law scalings in the characteristics of bimodal PDFs are identified and investigated.

Remote sensing image fusion on 3D scenarios: A review of applications for agriculture and forestry

Three-dimensional (3D) image mapping of real-world scenarios has a great potential to provide the user with a more accurate scene understanding. This will enable, among others, unsupervised automatic sampling of meaningful material classes from the target area for adaptive semi-supervised deep learning techniques. This path is already being taken by the recent and fast-developing research in computational fields, however, some issues related to computationally expensive processes in the integration of multi-source sensing data remain. Recent studies focused on Earth observation and characterization are enhanced by the proliferation of Unmanned Aerial Vehicles (UAV) and sensors able to capture massive datasets with a high spatial resolution. In this scope, many approaches have been presented for 3D modeling, remote sensing, image processing and mapping, and multi-source data fusion. This survey aims to present a summary of previous work according to the most relevant contributions for the reconstruction and analysis of 3D models of real scenarios using multispectral, thermal and hyperspectral imagery. Surveyed applications are focused on agriculture and forestry since these fields concentrate most applications and are widely studied. Many challenges are currently being overcome by recent methods based on the reconstruction of multi-sensorial 3D scenarios. In parallel, the processing of large image datasets has recently been accelerated by General-Purpose Graphics Processing Unit (GPGPU) approaches that are also summarized in this work. Finally, as a conclusion, some open issues and future research directions are presented.

GPU Acceleration of Large-Scale Full-Frequency GW Calculations.

Many-body perturbation theory is a powerful method to simulate electronic excitations in molecules and materials starting from the output of density functional theory calculations. By implementing the theory efficiently so as to run at scale on the latest leadership high-performance computing systems it is possible to extend the scope of GW calculations. We present a GPU acceleration study of the full-frequency GW method as implemented in the WEST code. Excellent performance is achieved through the use of (i) optimized GPU libraries, e.g., cuFFT and cuBLAS, (ii) a hierarchical parallelization strategy that minimizes CPU-CPU, CPU-GPU, and GPU-GPU data transfer operations, (iii) nonblocking MPI communications that overlap with GPU computations, and (iv) mixed precision in selected portions of the code. A series of performance benchmarks has been carried out on leadership high-performance computing systems, showing a substantial speedup of the GPU-accelerated version of WEST with respect to its CPU version. Good strong and weak scaling is demonstrated using up to 25 920 GPUs. Finally, we showcase the capability of the GPU version of WEST for large-scale, full-frequency GW calculations of realistic systems, e.g., a nanostructure, an interface, and a defect, comprising up to 10 368 valence electrons.

DPU: DAG Processing Unit for Irregular Graphs With Precision-Scalable Posit Arithmetic in 28 nm

Computation in several real-world applications like probabilistic machine learning, sparse linear algebra, and robotic navigation, can be modeled as irregular directed acyclic graphs (DAGs). The irregular data dependencies in DAGs pose challenges to parallel execution on general-purpose CPUs and GPUs, resulting in severe under-utilization of the hardware. This paper proposes DPU, a specialized processor designed for the efficient execution of irregular DAGs. The DPU is equipped with parallel compute units that execute different subgraphs of a DAG independently. The compute units can synchronize within a cycle using a hardware-supported synchronization primitive, and communicate via an efficient interconnect to a global banked scratchpad. Furthermore, a precision-scalable posit arithmetic unit is developed to enable application-dependent precision. The DPU is taped-out in 28nm CMOS, achieving a speedup of 5.1$\times$ and 20.6$\times$ over state-of-the-art CPU and GPU implementations on DAGs of sparse linear algebra and probabilistic machine learning workloads. This performance is achieved while operating at a power budget of 0.23W, as opposed to 55W and 98W of the CPU and GPU, resulting in a peak efficiency of 538 GOPS/W with DPU, which is 1350$\times$ and 9000$\times$ higher than the CPU and GPU, respectively. Thus, with specialized architecture, DPU enables low-power execution of irregular DAG workloads.

A scalable algorithm for many-body dissipative particle dynamics using multiple general purpose graphic processing units

We present a novel algorithm for the many-body Dissipative Particle Dynamics (DPD) forces calculation which allows to efficiently scale the DL_MESO software package on Multiple General Purpose Graphic Processing Units. Together with the extension to 64-bit integer arrays and addition of hard surface boundary conditions, the proposed algorithm allows to simulate very large complex mesoscale systems up to 14 billion beads. The implementation takes advantages of the CUDA language stream features to overlap the exchange of particle positions and local densities and the computation of the short range forces. We tested a water drop between two plates system using tree of the main European supercomputers: Piz Daint, Marconi and JUWELS. Results shows an improvement on the speedup compared to a naive implementation up to 1.5x when using 1024 GPUs.

DeepRS: A Library of Recommendation Algorithms Based on Deep Learning

In recent years, recommendation systems have become more complex with increasing research on user preferences. Recommendation algorithm based on deep learning has attracted a lot of attention from researchers in academia and industry, and many new algorithm models are proposed every year. Researchers often need to implement the proposed model to compare the results, which is a great challenge. Even if some papers provide source code, there are a variety of programming languages or deep learning frameworks, and it is not easy to compare the results in the different frameworks. In view of the lack of easily extensible deep learning-based recommendation algorithm libraries, based on the common analysis of deep learning algorithms in attention factorization machine (AFM), neural factorization machine (NFM), deep factorization machine (DeepFM) and deep cross-network (DCN), a recommendation algorithm library based on deep learning (DeepRS for short) is designed and implemented. It consists of three levels: framework level, abstract level and algorithm level. The framework level adopts the Tensorflow open source framework, which provides interfaces, such as automatic differentiation, tensor computing, GPU computing, and numerical optimization algorithms. The abstraction level uses the interface of the framework level to realize the embedding layer (EL), the full connection layer (FCL), the multi-layer perceptron layer (MLPL), the prediction layer (PL), the factorization machine layer (FML), the attention network layer (ANL), the cross-layer (CL) and the cross-network layer (CNL). The algorithm level implements the deep learning-based recommendation algorithms, such as AFM, NFM, DeepFM and DCN, on the basis of the abstraction level and the framework level. Experiments show that the proposed algorithm library has good scalability, ease of use and correctness.

Automated segmentation of Acute Lymphocytic Leukemia (ALL) subtypes by the combination of color space conversion and K-means cluster

Leukemia is blood cancer, and it is one of the most common and deadly causes of death in the world. Morphologically, Leukemia cells are classified into three types of L1, L2, and L3 by the French‑American‑British (FAB) classification. A new method of automatically segmenting blast cells from microscopic blood smear images proposed in this research. This study proposes significant pre-processing to obtain high segmentation performance and presents a new combination of image processing approaches. specially, the five color spaces selected with K-means cluster to segment subtypes of Acute Lymphocytic Leukemia. The majority of the components of the performance color space were chosen based on their similarity with ground truth image through using five evaluation parameters. The proposed codes for Acute Lymphocytic Leukemia subtype accurate segmentation applied on local and public datasets (427 images). The best color space type was YIQ which had 87% performance for the public dataset with segmentation evaluation 96.24% for dice parameter References AL-JABORIY, S. S., SJARIF, N. N. A., CHUPRAT, S. & ABDUALLAH, W. M. 2019. Acute lymphoblastic leukemia segmentation using local pixel information. Pattern Recognition Letters, 125, 85-90. AL HAMAD, H. A. Use an efficient neural network to improve the Arabic handwriting recognition. 2013 IEEE International Conference on Signal and Image Processing Applications, 2013. IEEE, 269-274. AMIN, M. M., KERMANI, S., TALEBI, A. & OGHLI, M. G. 2015. Recognition of acute lymphoblastic leukemia cells in microscopic images using k-means clustering and support vector machine classifier. Journal of medical signals and sensors, 5, 49. ASHOUR, A. S., WAHBA, M. A. & GHANNAM, R. 2021. A Cascaded Classification-Segmentation Reversible System for Computer-aided Detection and Cells Counting in Microscopic Peripheral Blood Smear Basophils and Eosinophils Images. IEEE Access. ASLAN, M. S., MOSTAFA, E., ABDELMUNIM, H., SHALABY, A., FARAG, A. A. & ARNOLD, B. A novel probabilistic simultaneous segmentation and registration using level set. 2011 18th IEEE International Conference on Image Processing, 2011. IEEE, 2161-2164. BHIMANI, J., LEESER, M. & MI, N. Accelerating K-Means clustering with parallel implementations and GPU computing. 2015 IEEE High Performance Extreme Computing Conference (HPEC), 2015. IEEE, 1-6. GHANE, N., VARD, A., TALEBI, A. & NEMATOLLAHY, P. 2017. Segmentation of white blood cells from microscopic images using a novel combination of K-means clustering and modified watershed algorithm. Journal of medical signals and sensors, 7, 92. HEGDE, R. B., PRASAD, K., HEBBAR, H. & SINGH, B. M. K. 2019. Image processing approach for detection of leukocytes in peripheral blood smears. Journal of medical systems, 43, 1-11. KADRY, S., RAJINIKANTH, V., TANIAR, D., DAMAŠEVIČIUS, R. & VALENCIA, X. P. B. 2021. Automated segmentation of leukocyte from hematological images—a study using various CNN schemes. The Journal of Supercomputing, 1-21. KARWAN, M., ABDULLAH, O., AMIN, A., HASAN, B., MOHAMED, Z., SULAIMAN, L., SHEKHA, M., NAJMULDEEN, H., BARZINGI, B. & SALIH, A. 2021. Cancer Statistics in Kurdistan Region of Iraq: A Tale of Two Cities. LABATI, R. D., PIURI, V. & SCOTTI, F. 2011. The Acute Lymphoblastic Leukemia Image Database for Image Processing. Universita Degli Studi Di Milano, 10. MILLER, D. R., LEIKIN, S., ALBO, V., SATHER, H. & HAMMOND, D. 1981. Prognostic importance of morphology (FAB classification) in childhood acute lymphoblastic leukaemia (ALL). British Journal of Haematology, 48, 199-206. PHILIP, A. T., SHIFAANA, S., SUNNY, S. & MANIMEGALAI, P. Detection of Acute Lymphoblastic Leukemia in Microscopic images using Image Processing Techniques. Journal of Physics: Conference Series, 2021. IOP Publishing, 012022. PRABHA, D. S. & KUMAR, J. S. 2016. Performance evaluation of image segmentation using objective methods. Indian J. Sci. Technol, 9, 1-8. SARRAFZADEH, O., DEHNAVI, A. M., RABBANI, H. & TALEBI, A. A simple and accurate method for white blood cells segmentation using K-means algorithm. 2015 IEEE Workshop on Signal Processing Systems (SiPS), 2015. IEEE, 1-6. SHAFIQUE, S. & TEHSIN, S. 2018. Acute lymphoblastic leukemia detection and classification of its subtypes using pretrained deep convolutional neural networks. Technology in cancer research & treatment, 17, 1533033818802789. SIEGEL, R. L., MILLER, K. D., GODING SAUER, A., FEDEWA, S. A., BUTTERLY, L. F., ANDERSON, J. C., CERCEK, A., SMITH, R. A. & JEMAL, A. 2020. Colorectal cancer statistics, 2020. CA: a cancer journal for clinicians, 70, 145-164. TAVAKOLI, E., GHAFFARI, A., KOUZEHKANAN, Z. M. & HOSSEINI, R. 2021. New Segmentation and Feature Extraction Algorithm for Classification of White Blood Cells in Peripheral Smear Images. bioRxiv. TSAI, C.-M. & LEE, H.-J. 2002. Binarization of color document images via luminance and saturation color features. IEEE Transactions on Image Processing, 11, 434-451. ZOU, K. H., WARFIELD, S. K., BHARATHA, A., TEMPANY, C. M., KAUS, M. R., HAKER, S. J., WELLS III, W. M., JOLESZ, F. A. & KIKINIS, R. 2004. Statistical validation of image segmentation quality based on a spatial overlap index1: scientific reports. Academic radiology, 11, 178-189.

University of Warsaw Lagrangian Cloud Model (UWLCM) 2.0: adaptation of a mixed Eulerian–Lagrangian numerical model for heterogeneous computing clusters

Abstract. A numerical cloud model with Lagrangian particles coupled to an Eulerian flow is adapted for distributed memory systems. Eulerian and Lagrangian calculations can be done in parallel on CPUs and GPUs, respectively. The fraction of time when CPUs and GPUs work simultaneously is maximized at around 80 % for an optimal ratio of CPU and GPU workloads. The optimal ratio of workloads is different for different systems because it depends on the relation between computing performance of CPUs and GPUs. GPU workload can be adjusted by changing the number of Lagrangian particles, which is limited by device memory. Lagrangian computations scale with the number of nodes better than Eulerian computations because the former do not require collective communications. This means that the ratio of CPU and GPU computation times also depends on the number of nodes. Therefore, for a fixed number of Lagrangian particles, there is an optimal number of nodes, for which the time CPUs and GPUs work simultaneously is maximized. Scaling efficiency up to this optimal number of nodes is close to 100 %. Simulations that use both CPUs and GPUs take between 10 and 120 times less time and use between 10 to 60 times less energy than simulations run on CPUs only. Simulations with Lagrangian microphysics take up to 8 times longer to finish than simulations with Eulerian bulk microphysics, but the difference decreases as more nodes are used. The presented method of adaptation for computing clusters can be used in any numerical model with Lagrangian particles coupled to an Eulerian fluid flow.

DARSegNet: A Real-Time Semantic Segmentation Method Based on Dual Attention Fusion Module and Encoder-Decoder Network

The convolutional neural network achieves excellent semantic segmentation results in artificially annotated datasets with complex scenes. However, semantic segmentation methods still suffer from several problems such as low use rate of the features, high computational complexity, and being far from practical real-time application, which bring about challenges for the image semantic segmentation. Two factors are very critical to semantic segmentation task: global context and multilevel semantics. However, generating these two factors will always lead to high complexity. In order to solve this, we propose a novel structure, dual attention fusion module (DAFM), by eliminating structural redundancy. Unlike most of the existing algorithms, we combine the attention mechanism with the depth pyramid pool module (DPPM) to extract accurate dense features for pixel labeling rather than complex expansion convolution. Specifically, we introduce a DPPM to execute the spatial pyramid structure in output and combine the global pool method. The DAFM is introduced in each decoder layer. Finally, the low-level features and high-level features are fused to obtain semantic segmentation result. The experiments and visualization results on Cityscapes and CamVid datasets show that, in real-time semantic segmentation, we have achieved a satisfactory balance between accuracy and speed, which proves the effectiveness of the proposed algorithm. In particular, on a single 1080ti GPU computer, ResNet-18 produces 75.53% MIoU at 70 FPS on Cityscapes and 73.96% MIoU at 109 FPS on CamVid.