Search

Channel pruning has recently become a widely used model compression method. However, most existing channel pruning methods only prune to decrease the model size, such as the number of parameters or FLOPs, and hence the decrease in model size does not effectively lead to an improvement in inference performance. To address this problem, this paper proposes a latency-optimized channel pruning method for CNN inference acceleration on GPU platforms by latency stair-step discrimination, two-stage benefit assessment and latency-sharing channel pruning. Compared with recent state-of-the-art model compression methods, it can achieve significant improvements in inference performance with comparable compression rates and model accuracy. The contributions of this paper include the following: first, a three-point latency stair-step discrimination method is proposed for determining the candidate prunable coordinates with the best latency performance adapted to the current hardware. Then, a two-stage benefit assessment method based on interlayer dependencies is proposed for determining the optimal channel pruning rate of each layer in the network. Finally, a latency-sharing channel pruning framework is proposed to accelerate the model pruning adaptation process. The method proposed in this paper can significantly reduce the model inference latency on multiple types of GPU platforms. To verify the effectiveness, we use three general-purpose GPU platforms and two embedded GPU platforms to evaluate the algorithm performance. The experimental results show that for recent state-of-the-art CNNs, the proposed method can achieve a 22.0–6.6% latency reduction and a 1.3 –3.0 inference performance improvement as well as a 1.2–4.3 pruning adaptation speedup with high model accuracy. [ABSTRACT FROM AUTHOR]

The Vehicle Routing Problem (VRP) is an NP-hard scheduling problem for goods transportation with vehicle capacity and transportation cost constraints. This paper presents GPU-based parallel strategies for the Local Search Heuristic (LSH) algorithm to solve the large-scale Capacitated Vehicle Routing Problem (CVRP) instances. This work employs a combination of five improvement heuristic approaches to improve the constructed feasible solution. Typically, 99% of CPU execution time is spent in the feasible solution improvement phase. Two GPU-based parallel strategies, namely route level and customer level parallel designs, have been developed to reduce the execution time of the solution improvement phase. The proposed parallel LSH version has been tested on large-scale instances with up to 30,000 customers. When compared to the corresponding sequential version, the customer level parallel design offers a speedup of up to 147.19 times. [ABSTRACT FROM AUTHOR]

With the development of image processing technology, pencil drawing has been widely used in video games and mobile phone applications. However, the existing pencil drawing algorithms require a large amount of time to convert a real picture into a pencil drawing; hence, it is difficult to apply them to real-time systems. This paper proposes a parallel fast pencil drawing generation algorithm based on the graphics processing unit (GPU) to accelerate the real-time rendering process of sketch painting. The parallelism of the pencil drawing generation algorithm is identified via a theoretical analysis at first. Then, sub-algorithms of the sequential algorithm are designed in parallel using the compute unified device architecture (CUDA) programming model and executed via thread-level parallel techniques. Furthermore, an optimal cache pattern of data that reduce the access time of the most frequently used data is structured using shared memory and constant memory. Finally, task-level parallelism is achieved by the CUDA stream technology, which overlaps independent sub-tasks for further acceleration. On the CUDA platform, the experimental results demonstrate that the proposed parallel algorithm can achieve a significant increase in speedup. The proposed algorithm achieves a performance improvement of 448.59 times compared with the sequential algorithm, on 2560×1920-resolution images, and maintains a high degree of similarity with the real pencil paintings. Hence, the proposed algorithm is suitable for real-time pencil drawing rendering and has promising application prospects in non-photorealistic rendering.

The efficient operation and management of a geothermal project can be largely affected by geological, physical, operational and economic uncertainties. Systematic uncertainty quantification (UQ) involving these parameters helps to determine the probability of the focused outputs, e.g., energy production, Net Present Value (NPV), etc. However, how to efficiently assess the specific impacts of different uncertain parameters on the outputs of a geothermal project is still not clear. In this study, we performed a comprehensive UQ to a low-enthalpy geothermal reservoir using the GPU implementation of the Delft Advanced Research Terra Simulator (DARTS) framework with stochastic Monte Carlo samplings of uncertain parameters. With processing the simulation results, large uncertainties have been found in the production temperature, pressure drop, produced energy and NPV. It is also clear from the analysis that salinity influences the producing energy and NPV via changing the amount of energy carried in the fluid. Our work shows that the uncertainty in NPV is much larger than that in produced energy, as more uncertain factors were encompassed in NPV evaluation. An attempt to substitute original 3D models with upscaled 2D models in UQ demonstrates significant differences in the stochastic response of these two approaches in representation of realistic heterogeneity. The GPU version of DARTS significantly improved the simulation performance, which guarantees the full set (10,000 times) UQ with a large model (circa 3.2 million cells) finished within a day. With this study, the importance of UQ to geothermal field development is comprehensively addressed. This work provides a framework for assessing the impacts of uncertain parameters on the concerning system output of a geothermal project and will facilitate analyses with similar procedures. • A high-fidelity model is needed to characterize heterogeneous geothermal reservoirs. • Uncertain geological, physical, operational and economic parameters are integrated. • Stochastic MC simulations assist the comprehensive uncertainty quantification. • GPU version of DARTS significantly improves the performance of geothermal modeling. • Upscaled models are not capable to capture the performance of geothermal reservoir. [ABSTRACT FROM AUTHOR]

Nowadays, similarity search on metric spaces is becoming a research field of interest due to the fact of its application to different scientific areas. However, when these applications produce a huge amount of data, it is necessary to accelerate the searching process by means of parallel architectures such as clusters, multiprocessors (multicores) or clusters of multiprocessors (multicores). Currently, graphic accelerators have emerged as a technology that allows for good performance at a low cost. In order to exploit the underlined architecture formed by multicores and a graphic accelerator it is nedeed to carry out heterogeneous programming, where CPUs and GPU are working at the same time taking benefits of the complete architecture. In this paper, a suitable generic structure adapted to the multicore/GPU system is presented and the experimental results, obtained in terms of execution time and speed-up, show the advantages of using this heterogeneous system, as well as a study of power consumption. [ABSTRACT FROM AUTHOR]