Graphics Processing Units Optimization Research Articles

GPU optimization techniques to accelerate optiGAN—a particle simulation GAN

The demand for specialized hardware to train AI models has increased in tandem with the increase in the model complexity over the recent years. Graphics processing unit (GPU) is one such hardware that is capable of parallelizing operations performed on a large chunk of data. Companies like Nvidia, AMD, and Google have been constantly scaling-up the hardware performance as fast as they can. Nevertheless, there is still a gap between the required processing power and processing capacity of the hardware. To increase the hardware utilization, the software has to be optimized too. In this paper, we present some general GPU optimization techniques we used to efficiently train the optiGAN model, a Generative Adversarial Network that is capable of generating multidimensional probability distributions of optical photons at the photodetector face in radiation detectors, on an 8GB Nvidia Quadro RTX 4000 GPU. We analyze and compare the performances of all the optimizations based on the execution time and the memory consumed using the Nvidia Nsight Systems profiler tool. The optimizations gave approximately a 4.5x increase in the runtime performance when compared to a naive training on the GPU, without compromising the model performance. Finally we discuss optiGANs future work and how we are planning to scale the model on GPUs.

GPU accelerated volumetric lattice Boltzmann model for image-based hemodynamics in portal hypertension

We are developing a new technique for monitoring portal hypertension by the pressure gradient between the portal vein and the inferior vena cava (PPG) based on non-invasive measurements (MRI images). Massive parametrization and classification are required to investigate the underlying relationship between the porosity and the stages of liver cirrhosis numerically, and the hepatic-portal venous system is a multi-scale system. Both of them need high computational costs. The suitability of the lattice Boltzmann method for GPU (Graphics Processing Unit) parallel computation provides an opportunity to overcome it. In this paper, we perform GPU parallelization and optimization for the volumetric lattice Boltzmann method with arbitrary geometry based on images. Three application cases, including pipe flow, hemodynamics in the portal venous system, and hemodynamics in a simple hepatic-portal venous system, are employed to prove the method can be applied in the hepatic-portal venous system based on accuracy and efficiency. The reliability of the model is qualitatively validated by the analytical solution of velocity and pressure difference distribution of pipe flow and quantitatively confirmed by the pulsatility of velocity and pressure difference that can be neglected in the portal venous system. The performance of the application cases is examined with Intel Broadwell E5-2683 v3@ 2.30 GHz (CPU) and NVIDIA Tesla V100 16GB (GPU). It shows the GPU algorithm for sparse geometry (SPARSE) has a similar speed to the regular GPU algorithm for dense geometry (DENSE) when the fluid volume fraction (q) is close to 1. And SPARSE speeds up to 2.2 times compared with DENSE when q is in the range of 0.19∼0.27. Meanwhile, the saving ratio of memory cost depends on q. For a numerical case in the hepatic-portal venous system, i.e., a large-scale system, parallel execution can be converged around half an hour with SPARSE, while the memory spills the limitation with DENSE with a single GPU. Hence, multi-GPU implementation is applied to release the limitation, and it can improve performance by increasing the number of GPU cards. In summary, the method presented in the paper is feasibility applied in the hepatic-portal venous system, laying the foundation of the new technique development.

Mining Cryptocurrency-Based Security Using Renewable Energy as Source

Cryptocurrency mining and blockchain technology using renewable energy as the main electricity source has gained attention for sustainable development in financial areas. However, very few studies have been reported concerning the power usage of cryptocurrencies using renewable energy. In this article, we report the effect of overclocking and undervolting on power usage and the hash rate for mining dogecoin with solar energy as renewable energy. The mining rig used in this work consists of different graphics processing units (GPUs) and non-LHR (lite hash rate) cards. The UnMineable software has been used for mining dogecoin as well as for wallet integration. The results indicate that mining dogecoin with solar energy as renewable energy consumes 2000 Watts power with overclocking and 1700 Watts power with undervolting technique. This work implicates the potential future of crypto-mining with renewable energy and the hardware configuration associated with it, which is expected to reduce e-waste and improve sustainable development. To reduce the e-waste and high electricity consumption, we have introduced two important techniques named GPU optimization and use of renewable energy for mining, which helps the miners to reduce the e-waste and electricity consumption significantly at the same time getting most out of the GPU by not having any impact on the environment.

GPU-DAEMON: GPU algorithm design, data management & optimization template for array based big omics data

In the age of ever increasing data, faster and more efficient data processing algorithms are needed. Graphics Processing Units (GPU) are emerging as a cost-effective alternative architecture for high-end computing. The optimal design of GPU algorithms is a challenging task which requires thorough understanding of the high performance computing architecture as well as the algorithmic design. The steep learning curve needed for effective GPU-centric algorithm design and implementation requires considerable expertise, time, and resources. In this paper, we present GPU-DAEMON, a GPU Data Management, Algorithm Design and Optimization technique suitable for processing array based big omics data. Our proposed GPU algorithm design template outlines and provides generic methods to tackle critical bottlenecks which can be followed to implement high performance, scalable GPU algorithms for given big data problem. We study the capability of GPU-DAEMON by reviewing the implementation of GPU-DAEMON based algorithms for three different big data problems. Speed up of as large as 386x (over the sequential version) and 50x (over naive GPU design methods) are observed using the proposed GPU-DAEMON. GPU-DAEMON template is available at https://github.com/pcdslab/GPU-DAEMON and the source codes for GPU-ArraySort, G-MSR and GPU-PCC are available at https://github.com/pcdslab.

Computational Benefit of GPU Optimization for the Atmospheric Chemistry Modeling

AbstractGlobal chemistry‐climate models are computationally burdened as the chemical mechanisms become more complex and realistic. Optimization for graphics processing units (GPU) may make longer global simulation with regional detail possible, but limited study has been done to explore the potential benefit for the atmospheric chemistry modeling. Hence, in this study, the second‐order Rosenbrock solver of the chemistry module of CAM4‐Chem is ported to the GPU to gauge potential speed‐up. We find that on the CPU, the fastest performance is achieved using the Intel compiler with a block interleaved memory layout. Different combinations of compiler and memory layout lead to ~11.02× difference in the computational time. In contrast, the GPU version performs the best when using a combination of fully interleaved memory layout with block size equal to the warp size, CUDA streams for independent kernels, and constant memory. Moreover, the most efficient data transfer between CPU and GPU is gained by allocating the memory contiguously during the data initialization on the GPU. Compared to one CPU core, the speed‐up of using one GPU alone reaches a factor of ~11.7× for the computation alone and ~3.82× when the data transfer between CPU and GPU is considered. Using one GPU alone is also generally faster than the multithreaded implementation for 16 CPU cores in a compute node and the single‐source solution (OpenACC). The best performance is achieved by the implementation of the hybrid CPU/GPU version, but rescheduling the workload among the CPU cores is required before the practical CAM4‐Chem simulation.

GPU Accelerated Multilevel Lagrangian Carotid Strain Imaging.

A multilevel Lagrangian carotid strain imaging algorithm is analyzed to identify computational bottlenecks for implementation on a graphics processing unit (GPU). Displacement tracking including regularization was found to be the most computationally expensive aspect of this strain imaging algorithm taking about 2.2 h for an entire cardiac cycle. This intensive displacement tracking was essential to obtain Lagrangian strain tensors. However, most of the computational techniques used for displacement tracking are parallelizable, and hence GPU implementation is expected to be beneficial. A new scheme for subsample displacement estimation referred to as a multilevel global peak finder was also developed since the Nelder-Mead simplex optimization technique used in the CPU implementation was not suitable for GPU implementation. GPU optimizations to minimize thread divergence and utilization of shared and texture memories were also implemented. This enables efficient use of the GPU computational hardware and memory bandwidth. Overall, an application speedup of was obtained enabling the algorithm to finish in about 50 s for a cardiac cycle. Last, comparison of GPU and CPU implementations demonstrated no significant difference in the quality of displacement vector and strain tensor estimation with the two implementations up to a 5% interframe deformation. Hence, a GPU implementation is feasible for clinical adoption and opens opportunity for other computationally intensive techniques.

Caffe CNN-based classification of hyperspectral images on GPU

Deep learning techniques based on Convolutional Neural Networks (CNNs) are extensively used for the classification of hyperspectral images. These techniques present high computational cost. In this paper, a GPU (Graphics Processing Unit) implementation of a spatial-spectral supervised classification scheme based on CNNs and applied to remote sensing datasets is presented. In particular, two deep learning libraries, Caffe and CuDNN, are used and compared. In order to achieve an efficient GPU projection, different techniques and optimizations have been applied. The implemented scheme comprises Principal Component Analysis (PCA) to extract the main features, a patch extraction around each pixel to take the spatial information into account, one convolutional layer for processing the spectral information, and fully connected layers to perform the classification. To improve the initial GPU implementation accuracy, a second convolutional layer has been added. High speedups are obtained together with competitive classification accuracies.

A Fast Discrete Wavelet Transform Using Hybrid Parallelism on GPUs

Wavelet transform has been widely used in many signal and image processing applications. Due to its wide adoption for time-critical applications, such as streaming and real-time signal processing, many acceleration techniques were developed during the past decade. Recently, the graphics processing unit (GPU) has gained much attention for accelerating computationally-intensive problems and many solutions of GPU-based discrete wavelet transform (DWT) have been introduced, but most of them did not fully leverage the potential of the GPU. In this paper, we present various state-of-the-art GPU optimization strategies in DWT implementation, such as leveraging shared memory, registers, warp shuffling instructions, and thread- and instruction-level parallelism (TLP, ILP), and finally elaborate our hybrid approach to further boost up its performance. In addition, we introduce a novel mixed-band memory layout for Haar DWT, where multi-level transform can be carried out in a single fused kernel launch. As a result, unlike recent GPU DWT methods that focus mainly on maximizing ILP, we show that the optimal GPU DWT performance can be achieved by hybrid parallelism combining both TLP and ILP together in a mixed-band approach. We demonstrate the performance of our proposed method by comparison with other CPU and GPU DWT methods.

Solving optimization problems using a hybrid systolic search on GPU plus CPU

In recent years, graphics processing units (GPUs) have emerged as a powerful architecture for solving a broad spectrum of applications in very short periods of time. However, most existing GPU optimization approaches do not exploit the full power available in a CPU–GPU platform. They have a tendency to leave one of them partially unused (usually the CPU) and fail to establish an accurate exchange of information that could help solve the target problem efficiently. Thus, better performance is expected from devising a hybrid CPU–GPU parallel algorithm that combines the highly parallel stream processing power of GPUs with the higher power of multi-core architectures. We have developed a hybrid methodology to efficiently solve optimization problems. We use a hybrid CPU–GPU architecture, to benefit from running it, in parallel, on both the CPU and the GPU. Our experiments over a heterogeneous set of combinatorial optimization problems with increasing dimensionality show a time gain of up to \(365\times \) in our proposal, while demonstrating high numerical accuracy. This work is intended to open up a new line of research that matches both architectures with new algorithms and cooperation techniques.

High performance GPU based optimized feature matching for computer vision applications

In this paper, a graphics processing unit (GPU) based matching technique is proposed to perform fast feature matching between different images under various image conditions with viewpoint changes. Most recently, general-purpose graphics processing units (GPGPUs or GPUs) have become commonplace within high performance supercomputers. GPUs allow developers to effectively exploit the computational power for high performance computing. This paper focuses on improving the performance of feature matching based on self-organizing map by porting it onto the GPUs. GPU optimization has been applied for the fast computation of keypoints to make the system fast and efficient. This scheme has enhanced the overall performance and is much more efficient compared to other methods without degradation of detection results. The proposed matching algorithm is partitioned between the CPU and GPU in a way that best exploits the parallelism and perform matching between the different images. Experimental results demonstrate that fast feature matching is achieved using the graphics processing units, and its computational efficiency is checked by comparing its results and run times with those of some standard software (MATLAB) run on central processing unit (CPU).

Parallel Implementation of Sparse Representation Classifiers for Hyperspectral Imagery on GPUs

Classification is one of the most important analysis techniques for hyperspectral image analysis. Sparse representation is an extremely powerful tool for this purpose, but the high computational complexity of sparse representation-based classification techniques limits their application in time-critical scenarios. To improve the efficiency and performance of sparse representation classification techniques for hyperspectral image analysis, this paper develops a new parallel implementation on graphics processing units (GPUs). First, an optimized sparse representation model based on spatial correlation regularization and a spectral fidelity term is introduced. Then, we use this approach as a case study to illustrate the advantages and potential challenges of applying GPU parallel optimization principles to the considered problem. The first GPU optimization algorithm for sparse representation classification (SRCSC_P) of hyperspectral images is proposed in this paper, and a parallel implementation of the proposed method is developed using compute unified device architecture (CUDA) on GPUs. The GPU parallel implementation is compared with the serial and multicore implementations on CPUs. Experimental results based on real hyperspectral datasets show that the average speedup of SRCSC_P is more than $\mathbf{130} \times$ , and the proposed approach is able to provide results accurately and fast, which is appealing for computationally efficient hyperspectral data processing.

GPU Optimized Stereo Image Matching Technique for Computer Vision Applications

In this paper, we propose a graphics processing unit (GPU) based matching technique to perform fast feature matching between different images. Lowe proposed a scale invariant feature transform algorithm that has been successfully used in various feature matching applications such as stereo vision, object recognition, and many others, but this algorithm is computationally intensive. In order to solve this problem, we propose a matching technique optimized for graphics processing units to perform computation with less time. We have applied GPU optimization for the fast computation of keypoints to make our system fast and efficient. The proposed method used self-organizing map feature matching technique to perform efficient matching between different images. The experiments are performed on various images to examine the performance of the system in diverse conditions such as image rotation, scaling, and blurring conditions. The experimental results reveal that the proposed algorithm outperforms the existing feature matching methods resulting into fast feature matching with the optimization of graphics processing unit. Index Terms—Feature matching, stereo vision, self- organizing map, graphics processing unit

Efficient Parallel GPU Design on WRF Five-Layer Thermal Diffusion Scheme

Satellite remote-sensing observations and ground-based radar can detect the weather conditions from a distance and are widely used to monitor the weather all around the globe. The assimilated satellite/radar data are passed through the weather models for weather forecasting. The five-layer thermal diffusion scheme is one of the weather models, handling with an energy budget made up of sensible, latent, and radiative heat fluxes. The model feature of no interactions among horizontal grid points makes this scheme very favorable for parallel processing. This study demonstrates implementation of this scheme using graphics processing unit (GPU) massively parallel architecture. By employing one NVIDIA Tesla K40 GPU, our GPU optimization effort on this scheme achieves a speedup of $311 \times$ with respect to its CPU counterpart Fortran code running on one CPU core of Intel Xeon E5-2603, whereas the speedup for one CPU socket (four cores) with respect to one CPU core is only $3.1 \times$ . We can even boost the speedup of this scheme to $398 \times$ with respect to one CPU core when two NVIDIA Tesla K40 GPUs are applied.

GPU parallelization of unstructured/hybrid grid ALE multigrid unsteady solver for moving body problems

Graphics Processing Units (GPUs) are currently being used to accelerate Computational Fluid Dynamics (CFD) codes and many GPU based CFD solvers have already showed that GPUs have great capabilities in numerical simulations. The development and validation of an unstructured/hybrid grid ALE (Arbitrary Lagrangian–Eulerian) multigrid unsteady solver on the GPU for moving body problems are present. This GPU based unsteady solver consists of two modules. A BICGSTAB based mesh deformation module which updates the mesh nodes to new positions as the interfaces of bodies move, and a geometrical multigrid (GMG) RANS (Reynolds-Averaged Navier–Stokes) solver which manages the flow computation using the dual time stepping approach. The GPU implementation and optimization for this two modules are discussed, respectively. Both 2D and 3D validation cases are carried out to analyze the accuracy and efficiency of this solver. The GPU results obtained agree well with the experimental measurements, as well as the numerical solutions presented by other researchers. A mixed MPI-CUDA implementation makes the unsteady flow solver capable of running on a multi-GPU cluster that consists of a set of GPUs and CPUs. Both the strong and weak scalability of the code are investigated on a heterogeneous computing platform.

Sailfish: A flexible multi-GPU implementation of the lattice Boltzmann method

We present Sailfish, an open source fluid simulation package implementing the lattice Boltzmann method (LBM) on modern Graphics Processing Units (GPUs) using CUDA/OpenCL. We take a novel approach to GPU code implementation and use run-time code generation techniques and a high level programming language (Python) to achieve state of the art performance, while allowing easy experimentation with different LBM models and tuning for various types of hardware. We discuss the general design principles of the code, scaling to multiple GPUs in a distributed environment, as well as the GPU implementation and optimization of many different LBM models, both single component (BGK, MRT, ELBM) and multicomponent (Shan–Chen, free energy). The paper also presents results of performance benchmarks spanning the last three NVIDIA GPU generations (Tesla, Fermi, Kepler), which we hope will be useful for researchers working with this type of hardware and similar codes. Program SummaryProgram title: SailfishCatalogue identifier: AETA_v1_0Program summary URL:http://cpc.cs.qub.ac.uk/summaries/AETA_v1_0.htmlProgram obtainable from: CPC Program Library, Queen’s University, Belfast, N. IrelandLicensing provisions: GNU Lesser General Public License, version 3No. of lines in distributed program, including test data, etc.: 225864No. of bytes in distributed program, including test data, etc.: 46861049Distribution format: tar.gzProgramming language: Python, CUDA C, OpenCL.Computer: Any with an OpenCL or CUDA-compliant GPU.Operating system: No limits (tested on Linux and Mac OS X).RAM: Hundreds of megabytes to tens of gigabytes for typical cases.Classification: 12, 6.5.External routines: PyCUDA/PyOpenCL, Numpy, Mako, ZeroMQ (for multi-GPU simulations), scipy, sympyNature of problem:GPU-accelerated simulation of single- and multi-component fluid flows.Solution method:A wide range of relaxation models (LBGK, MRT, regularized LB, ELBM, Shan–Chen, free energy, free surface) and boundary conditions within the lattice Boltzmann method framework. Simulations can be run in single or double precision using one or more GPUs.Restrictions:The lattice Boltzmann method works for low Mach number flows only.Unusual features:The actual numerical calculations run exclusively on GPUs. The numerical code is built dynamically at run-time in CUDA C or OpenCL, using templates and symbolic formulas. The high-level control of the simulation is maintained by a Python process.Additional comments:!!!!! The distribution file for this program is over 45 Mbytes and therefore is not delivered directly when Download or Email is requested. Instead a html file giving details of how the program can be obtained is sent. !!!!!Running time:Problem-dependent, typically minutes (for small cases or short simulations) to hours (large cases or long simulations).

Multi-dimensional characterization of electrostatic surface potential computation on graphics processors.

BackgroundCalculating the electrostatic surface potential (ESP) of a biomolecule is critical towards understanding biomolecular function. Because of its quadratic computational complexity (as a function of the number of atoms in a molecule), there have been continual efforts to reduce its complexity either by improving the algorithm or the underlying hardware on which the calculations are performed.ResultsWe present the combined effect of (i) a multi-scale approximation algorithm, known as hierarchical charge partitioning (HCP), when applied to the calculation of ESP and (ii) its mapping onto a graphics processing unit (GPU). To date, most molecular modeling algorithms perform an artificial partitioning of biomolecules into a grid/lattice on the GPU. In contrast, HCP takes advantage of the natural partitioning in biomolecules, which in turn, better facilitates its mapping onto the GPU. Specifically, we characterize the effect of known GPU optimization techniques like use of shared memory. In addition, we demonstrate how the cost of divergent branching on a GPU can be amortized across algorithms like HCP in order to deliver a massive performance boon.ConclusionsWe accelerated the calculation of ESP by 25-fold solely by parallelization on the GPU. Combining GPU and HCP, resulted in a speedup of at most 1,860-fold for our largest molecular structure. The baseline for these speedups is an implementation that has been hand-tuned SSE-optimized and parallelized across 16 cores on the CPU. The use of GPU does not deteriorate the accuracy of our results.

Robust Parallel Preconditioned Power Grid Simulation on GPU With Adaptive Runtime Performance Modeling and Optimization

Leveraging the power of nowadays graphics processing units for robust power grid simulation remains a challenging task. Existing preconditioned iterative methods that require incomplete matrix factorizations cannot be effectively accelerated on graphics processing unit (GPU) due to its limited hardware resource as well as data parallel computing. This paper presents an efficient GPU-based multigrid preconditioning algorithm for robust power grid analysis. By combining the fast geometric multigrid solver with the robust Krylov-subspace iterative solver, power grid DC and transient analysis can be performed efficiently on GPU without loss of accuracy (largest errors <;0.5 mV). Unlike previous GPU-based algorithms that rely on good power grid regularities, the proposed algorithm can be applied for more general power grid structures. Additionally, we also propose an accuracy-aware GPU performance modeling and optimization framework to automatically obtain the best power grid simulation configurations. Experimental results show that the DC and transient analysis on GPU can achieve more than 25X speedups over the best available CPU-based solvers.

Fast convolution‐superposition dose calculation on graphics hardware

The numerical calculation of dose is central to treatment planning in radiation therapy and is at the core of optimization strategies for modern delivery techniques. In a clinical environment, dose calculation algorithms are required to be accurate and fast. The accuracy is typically achieved through the integration of patient-specific data and extensive beam modeling, which generally results in slower algorithms. In order to alleviate execution speed problems, the authors have implemented a modern dose calculation algorithm on a massively parallel hardware architecture. More specifically, they have implemented a convolution-superposition photon beam dose calculation algorithm on a commodity graphics processing unit (GPU). They have investigated a simple porting scenario as well as slightly more complex GPU optimization strategies. They have achieved speed improvement factors ranging from 10 to 20 times with GPU implementations compared to central processing unit (CPU) implementations, with higher values corresponding to larger kernel and calculation grid sizes. In all cases, they preserved the numerical accuracy of the GPU calculations with respect to the CPU calculations. These results show that streaming architectures such as GPUs can significantly accelerate dose calculation algorithms and let envision benefits for numerically intensive processes such as optimizing strategies, in particular, for complex delivery techniques such as IMRT and are therapy.