59 results about "Parallel computing architecture" patented technology

A method and an apparatus that execute a parallel computing program in a programming language for a parallel computing architecture are described. The parallel computing program is stored in memory in a system with parallel processors. The system includes a host processor, a graphics processing unit (GPU) coupled to the host processor and a memory coupled to at least one of the host processor and the GPU. The parallel computing program is stored in the memory to allocate threads between the host processor and the GPU. The programming language includes an API to allow an application to make calls using the API to allocate execution of the threads between the host processor and the GPU. The programming language includes host function data tokens for host functions performed in the host processor and kernel function data tokens for compute kernel functions performed in one or more compute processors, e.g. GPUs or CPUs, separate from the host processor. Standard data tokens in the programming language schedule a plurality of threads for execution on a plurality of processors, such as CPUs or GPUs in parallel. Extended data tokens in the programming language implement executables for the plurality of threads according to the schedules from the standard data tokens.

The invention relates to a heating pipe network hydraulic simulation model identification correction method and a system. The heating pipe network hydraulic simulation model identification correction method comprises the steps of S1, selecting multiple groups of actually measured data of a heating system under steady work conditions, S2, calculating theoretical values of the drag coefficient of all pipe section, and S3, establishing a hydraulic simulation model of heating pipe network, using particle swarm optimization method accompanied by multiple groups of actually measured data, identifying and calculating various resistance coefficient correction amount vectors, and through the various resistance coefficient correction amount vectors in the heating pipe network hydraulic simulation model, making the comparison between the flow conditions and the actual measured data under multiple working conditions, identifying the optimal resistance coefficient correction amount. The corrected model can better describe the actual physical system, and increase the calculation precision of the hydraulic calculation of heating pipe network. The application of a parallel computing architecture for the implementation for the identification of the particle swarm algorithm drag coefficient correction with high identification efficiency.

The invention discloses a video monitoring equipment fault automatic detection method which includes the following steps: a convolutional auto-encoder uses different numbers of network layers and different numbers of hidden nodes to extract the features of a monitoring image, image feature extraction is carried out in a pooling layer, and the convolutional auto-encoder is iterated until the accuracy of a classification model converges; a convolutional neural network uses the features of the monitoring image extracted by the convolutional auto-encoder as the basis of image classification to realize supervised learning; and the convolutional neural network uses a parallel computing architecture to train a network model, and after training of the network model, the model is applied to the classification process of a test image, and the damage condition of video monitoring equipment is judged according to the classification result. According to the video monitoring equipment fault automatic detection method of the invention, a corresponding improved algorithm is put forward for video monitoring equipment fault detection, and an automatic and fast fault detection method is put forward.

The invention provides a high-efficiency image segmentation method aiming at remote-sensing image multi-scale segmentation. The method comprises the following steps: reading original remote-sensing image information to be segmented, setting segmentation parameters and formulating an optimized segmentation strategy in a CPU segmentation start control stage; segmenting by adopting a two-layer parallel computing architecture, wherein a first layer is used for solving limitation of a storage space and guaranteeing load balancing between the CPU and GPU based on a coarse-grained parallel architecture between image blocks, and a second layer is used for guaranteeing load balancing inside the CPU and GPU based on fine-grained parallel architecture of pixels; and finally, carrying out a CPU result saving stage. The method has the advantages that the segmentation speed is high, the segmentation precision is high, setting of multiple segmentation parameters is supported, segmentation of ultra-large remote-sensing images is supported, export of multiple segmentation results is supported, and the like.

The invention provides a river network water flow quality real-time prediction method and device based on data assimilation. According to the technical scheme, on the basis of real-time monitoring data such as the water level, the flow and the water quality of the river network, real-time monitoring data are assimilated into the river network water flow and quality model through the ensemble Kalman filter or the improved algorithm of the ensemble Kalman filter, the river network water flow and quality data assimilation model is constructed, and the computing efficiency of the data assimilationmodel is improved through a parallel computing architecture. Due to the data assimilation technology, water flow and quality model structure errors, inlet and outlet boundary errors and observation value errors are comprehensively considered; real-time observation data are fused to dynamically correct water level, flow and water quality concentration state variables and roughness and water quality parameters of the water quality model, so that the water level, flow and water quality concentration of a complex river network system can be dynamically predicted by adopting a river network waterquality data assimilation model and a parallel computing architecture, and the model prediction precision can be improved.

The invention discloses a method utilizing a graphic processing unit (GPU) for achieving rapid wavelet transformation through segmentation. The method utilizing the GPU for achieving rapid wavelet transformation through segmentation mainly solves the problem that the speed of wavelet transformation in the prior art is low. Directing at the features of a parallel computing framework of the GPU, the method which segments data and conducts parallel processing to the data is designed. The method includes the following steps: copying an original image, segmenting the data, expanding the data, copying the data to a shared memory, transforming rows, transforming lines, canceling data expansion and copying the transformed data to a host machine memory. According to the method utilizing the GPU for achieving rapid wavelet transformation through segmentation, accuracy of a wavelet transformation result is guaranteed through the expansion of the segmented data, the visiting speed of the data is improved through the transformation in the shared memory, interaction between the data and a global memory is avoided, and the processing speed of the whole image is improved through the paralleling of image blocks and the paralleling among pixel points of image block branches.

The invention discloses a method for implementing the SHA-1 algorithm by using a GPU, which includes the following steps: (1) implementing the SHA-1 algorithm on a general-purpose CPU; (2) transplanting the program to the GPU, and performing preprocessing according to the following method Processing: a. Create an empty function body; B. Copy the parameters from the main memory to the global matching storage space of the GPU through the parameter pointer, and copy the result from the global matching storage space of the GPU back to the main memory; C. Add device configuration modifier; D. Synchronize. The present invention provides a method for implementing the SHA-1 algorithm by using the GPU, utilizing the rapidly developing graphics hardware technology, adopting the CUDA parallel computing architecture of NVIDIA Corporation for parallel implementation, and using the GPU as a parallel computing device for program release and management operations, with flexible It has good performance, reduces CPU load, improves the processing performance of information security equipment, improves computing efficiency, and has the advantages of high security in the computing process.

The invention discloses a network intrusion detection method, device and system and a storage medium, and the method comprises the steps: initializing a distributed computing system, and enabling a detection node to report node information to a management and control node; distributing a feature rule base to detection nodes of corresponding detection groups, and copying the same service data mirror image traffic to different distributed detection groups by using a stream copying technology; according to the flow copied to the detection group by the flow copying technology, controlling the service data to be shunted to a single detection node or a detection node of the detection group of the same feature rule base; and enabling the detection node to receive the to-be-detected message, perform matching detection according to the feature rule base and output a detection result to the management and control node. Through a technical means of adopting a distributed and parallel computing architecture and shunting a network data stream to a plurality of detection nodes by utilizing distributed and sdn switch stream replication, the technical problem of low detection efficiency during large-flow data in the prior art is solved, and the effect of improving the system detection efficiency is achieved.

The invention discloses a test method for achieving electromagnetic transient real-time simulation of an electrical power system based on CUDA parallel computing. An electromagnetic transient simulation test host computer adopts a CPU (central processing unit) and GPU (ground power unit) mode; the GPU uses a CUDA platform parallel computing framework; the simulation test host computer is connected with collection devices of all channels of signals through the gigabit Ethernet; a model can be greatly simplified mainly based on of the digital substation simulation instead of the whole-network or whole system simulation, the traditional simulation speed can be increased by dozens of times or hundreds of times, and requirements of transformer substation electromagnetic transient real-time simulation computing are met; and compared with the cost of a large parallel computer or a professional parallel computer, the cost of the CUDA parallel computing platform is greatly reduced. According to the method, in the aspect of electromagnetic transient real-time simulation of the electrical power system, the performance is greatly improved, the cost is substantially reduced, and the method has a wide popularization value.

The embodiment of the invention provides a method and device for achieving accelerated processing of a computing task. The method comprises the steps: receiving a calculation task and calculation datafrom an external device; determining a parallel computing scheme of the computing task according to the operation type of the computing task and the data size of the computing data, wherein the parallel computing scheme comprises the number of parallel computing blocks and a block thread resource allocation scheme of each parallel computing block, realizing the parallel computing in each parallelcomputing block by utilizing the plurality of distributed threads. calling a parallel computing architecture to execute parallel computing according to the partitioning thread resource allocation scheme to obtain a computing result; and providing the calculation result to the external device. By means of the method, a double-layer parallel processing scheme based on block parallelism and thread parallelism can be used for efficiently achieving computing task parallel processing.

A water surface garbage detection method based on a deep learning algorithm comprises the steps that hardware equipment is arranged and installed, front-end cameras are arranged in a plurality of important sections of water areas such as an inland river and a lake to be detected, and the front-end cameras are finally connected to a core switch through an access layer switch; a deep learning algorithm and a feedback mechanism are arranged in the video analysis server based on the deep learning algorithm, and the feedback mechanism performs secondary training on pictures which are misreported ormissed in deep learning training, so that the model characteristics of the continuously trained deep learning algorithm are more comprehensively covered, and the false report rate and the missed report rate of the pictures are reduced; the video analysis server adopts a multi-GPU parallel computing architecture, hardware server resources are utilized to the maximum extent, and the investment costis saved; a CAFFE-based SSD deep learning algorithm is outstanding in calculation speed, and the detection real-time performance is effectively ensured.

The invention discloses a satellite-borne target fusion detection method based on a rotating region convolutional neural network, and relates to the field of satellite-borne target detection. The invention is oriented to intelligent computing platforms such as a universal parallel computing architecture processor, a field programmable gate array, a system-on-chip (SOC) and the like. The method comprises the following steps: firstly, fusing a panchromatic (PAN) image and a multispectral (MS) image by adopting a panchromatic sharpening method, and then detecting a target in the fused spaceborneimage by adopting a target detection framework based on a rotating region convolutional neural network, wherein the target detection framework is based on a Faster R-CNN structure. According to the invention, the detection accuracy of the small target in the multi-source satellite-borne image is improved, the target detection framework has good generalization ability, and the method can be widelyapplied to the fields of target detection, safety monitoring and the like.

The invention provides a traffic flow simulation method based on routing optimization and parallel computing architecture. The traffic flow simulation method comprises the following steps: performingscene three-dimensional modeling in combination with a building information model; performing intelligent agent processing on all the movable units in the scene to obtain a plurality of intelligent agents; setting boundary conditions of the intelligent agent; setting initial coordinates of the intelligent agent in the scene according to the road information in the scene; using a routing algorithmaccording to the road information in the scene, and calculating behavior strategies of a plurality of intelligent agents at the same time under a parallel computing architecture; updating real-time coordinates of the plurality of agents; displaying the motion trails of the plurality of intelligent agents in the scene, and completing the simulation of the traffic flow; the invention further provides a traffic flow simulation system based on the routing optimization and parallel computing architecture. According to the invention, the technical problems that the traditional offline simulation mode through the following flow model cannot be linked with the traffic scheme change in real time, and the difference between the calculated traffic flow simulation result and the real traffic conditionof the road is large can be solved.

The invention discloses a deep learning parallel computing architecture method and hyper-parameter automatic configuration optimization thereof, and particularly relates to the field of deep learning.Firstly, a CNN is used for capturing spatial features of all places, and then an SRU is used for capturing sequential features of spatiotemporal data on the basis of the spatial features and used forregression prediction of the spatiotemporal data and hyper-parameter automatic configuration optimization of the spatiotemporal data; the invention has the beneficial effects that 1, regression prediction of spatio-temporal data is realized based on the SRU, so that parallel acceleration of the model to a certain extent is realized, and the time consumed by training and reasoning is reduced; 2, automatic configuration of the hyper-parameters of the model is realized based on a parallel genetic algorithm, so that manpower, energy and time consumed by hyper-parameter configuration are reduced,the hyper-parameters are more reasonable, and the prediction performance of the model is better.

The invention discloses a regional electromagnetic environment ray propagation path parallel search method. The method comprises the steps that an electromagnetic environment region to be calculated is rasterized, parallel search is conducted on all ray propagation paths starting from a radiation source, statistics is conducted on all paths reaching grids, and the field intensity distribution of the whole region is obtained. The regional electromagnetic environment ray propagation path parallel search method has the advantages that as field intensity calculation of a small grid point in the region is completed, calculation of the field intensity distribution of the whole region is completed, and therefore the calculation quantity is greatly reduced; all the ray paths in the region can be searched for concurrently, and parallel computing architectures including cuda, OpenCL, open MP, Mpi, MapReduce and the like are quite suitable for accelerating the search process; the calculation quantity of the field intensity distribution of the whole region is reduced into the calculation quantity of the small grid point in the region, the additional calculation quantity caused by the size increase of the region is greatly reduced, and the bottleneck limiting the regional electromagnetic calculation is broken.

The invention discloses a method for improving operating efficiency of a parallel architecture. A data flow based flow equalizing algorithm is provided in the parallel computing architecture. The method comprises the steps of firstly determining channel priority levels according to processing capacities of channels, and determining the processing capacities required by data messages according to feature fields in data heads after receiving data; sending the data needing good processing capacities to high-priority-level channels to queue up; sequentially sending out the data according to data bulks in buffer areas of the channels different in priority level from large to small; firstly sending out the data in the high-priority-level channels under the same conditions. The requirement of the multichannel input and multichannel output computing architecture having priority level relation existing among the output channels can be met, and the processing efficiency of the parallel computing architecture is remarkably improved.

The invention discloses a method for partitioning a large-scale static network based on a graphics processor, which mainly aims to solve the problem that the existing network partitioning method is not suitable for partitioning a large-scale network and low in efficiency. The method comprises the following steps: by defining a result matrix S of network partitioning, transforming the operation ofrespectively calculating the characteristic vector and characteristic value of each sub-module in network partitioning into the operation of calculating the characteristic vector and characteristic value of the whole network; and according to the parallel computing architecture of the graphics processor, decomposing the calculation of the characteristic vector and the characteristic value in network partitioning into a plurality of basic calculations, so that the original huge dense matrix calculation is simplified to the computations between sparse matrixes and vectors. In the invention, through reasonably combining the basic calculations, memory space is saved, the consumption of data transmission is decreased, and the efficiency of network partitioning is effectively improved, so that the partitioning of an extra-large scale static network which is difficult to implement becomes possible.

The invention provides a remote sensing image preprocessing method. The method comprises a step of carrying out orthographic correction on input remote sensing images and a step of carrying out Gram-Schmidt fusion on orthographically corrected images. For each waveband of each image, a plurality of computing threads are executed in parallel on a GPU; and the images comprise single-scene low-resolution multispectral images MSS and high-resolution single-waveband images PAN. The method is based on a CUDA (Computed Unified Device Architecture) universal parallel computing architecture. Accordingto the method, the remote sensing images are preprocessed by utilizing the parallel processing capability of the GPU; and the orthographic correction and the Gram-Schmidt fusion algorithms in processing are improved; and therefore, the parallel processing of the remote sensing images on thousands of threads on the GPU is realized, and processing speed is greatly increased.