267 results about "Parallel optimization" patented technology

The invention discloses a regional energy internet hierarchical control method considering flexible loads, which comprises the following steps of: constructing a global layer, dividing a regional energy internet into a plurality of electrically strongly coupled sub-regions by adopting a dynamic partitioning algorithm, aggregating equivalent adjustable capacities of the flexible loads in the sub-regions, and performing rolling optimization in an hour-level time scale, an improved Benders decomposition algorithm is adopted to accelerate solving and generate a partition-level power regulation instruction; constructing a regional layer, designing a thermodynamic-economic hybrid constraint model according to flexible load physical characteristics and user behavior elasticity, introducing a user comfort recovery time constraint and elastic electricity price response function, and decomposing an adjustment instruction issued by the global layer to each flexible load cluster in a minute-level time scale by adopting a parallel ADMM optimization algorithm; and constructing a load layer, constructing a millisecond-level adjustable load priority switching mechanism based on FPGA hardware acceleration and finite-state machine logic, calculating a load adjustable power boundary in real time, and dynamically adjusting a response queue in combination with system frequency deviation.

The invention relates to the technical field of control, and particularly discloses a multivariable energy efficiency optimization control system for a heating furnace, which is used for solving the problems of local overheating, non-uniform temperature and difficulty in accurate positioning and compensation of heat loss in the operation of the existing cracking heating furnace. Comprising a parameter detection module, a multivariable coupling modeling and simulation module, an optimization control module and an execution and feedback module. According to the method, dynamic digital twinning is constructed through multi-modal online sensing and data assimilation, a Pareto frontier solution is generated based on model prediction control and improved NSGA-II parallel optimization, and the weight is adaptively adjusted; and when the hot spot / cold spot is triggered, a quadric surface fitting compensation strategy is implemented and issued for execution, so that high-precision simulation prediction, precise closed-loop control and real-time online energy efficiency optimization are realized.

The invention provides a satellite networking communication method, device and equipment, and the method comprises the steps: carrying out the space-time window decomposition processing of a satellite constellation orbit parameter based on orbit phase synchronization, and decomposing a long-period scheduling problem into a plurality of short-period sub-problems; performing state transition conflict detection processing on the double-feed antenna system according to the space-time window decomposition result, and determining the working state of the feed antenna; parallel neighborhood search processing based on reinforcement learning is carried out according to the feed antenna resource configuration scheme, and a communication link scheduling strategy is generated through multi-thread collaborative optimization; and performing dynamic fusion processing based on state prediction according to the parallel optimization scheduling scheme, and adjusting a satellite-ground station communication link establishment time sequence through learning type template matching. Through space-time decomposition, intelligent conflict detection, reinforcement learning optimization and dynamic fusion, the problems of complexity and real-time performance of large-scale satellite networking communication scheduling are solved, and the scheduling efficiency and the communication quality are remarkably improved.

The invention relates to a load matching optimization method of a multi-channel gate driver and related equipment thereof. The method comprises the following steps of: acquiring grid current, grid voltage and switching frequency of the multi-channel parallel grid driver to obtain driver load operation data; performing region division processing on the operation space to obtain a safe operation region and a transition region; inputting a first adaptive prediction model for feature extraction to obtain a load balancing prediction result; parallel optimization calculation is carried out on the control strategy to obtain each driving channel control parameter, and each driving channel control parameter comprises a grid driving voltage regulating variable and a driving time sequence regulating variable; performing real-time adjustment on each driving channel control parameter to obtain a load balance adjustment instruction; and performing performance index evaluation calculation to obtain a performance optimization parameter, and updating the first adaptive prediction model to obtain a second adaptive prediction model. According to the invention, through cooperative control of voltage regulation and time sequence compensation, dynamic balance of each channel load is realized.

The invention relates to the technical field of electrical property detection, and discloses a graphene film surface property detection device, which comprises a data acquisition module used for acquiring electrical property data of the surface of a graphene film through a multi-point array electrode probe; the vector grid modeling module is used for constructing a vector grid model for representing the surface performance of the graphene film according to the electrical characteristic data; the iterative inversion calculation module is used for performing iterative inversion calculation on the vector grid model to determine the surface performance distribution of the graphene film; the parallel optimization fusion module is used for performing parallel processing and optimization fusion on the multi-scale performance data through a distributed computing framework; the three-dimensional visualization module is used for generating a three-dimensional visualization model of the surface performance of the graphene film according to the optimized and fused performance data; the problems of insufficient precision and low efficiency of a traditional detection method in graphene film performance detection are solved.

The invention relates to the technical field of artificial intelligence, in particular to a large model and multi-agent collaborative decision-making method based on dynamic knowledge flow, which comprises the following steps of: analyzing a static knowledge and dynamic information fusion relationship through joint modeling, extracting a hierarchical structure of equipment constraints and environment variables, identifying constraint conflicts and deviations in task decomposition, and obtaining a multi-agent collaborative decision-making result; screening a consistency decomposition direction, extracting a task constraint parallel optimization theory, correcting constraint conflicts, evaluating consistency changes, and outputting a collaborative task convergence robust state identifier. According to the method, by integrating static knowledge and dynamic information and optimizing understanding of task decomposition and equipment constraints, the collaborative decision-making capacity of multiple agents in a complex environment is enhanced, the conflict and deviation processing capacity in the task execution process is improved, the accuracy and consistency of tasks are enhanced, and the convergence and stability of task targets are improved; the robustness of multi-agent cooperative work is promoted, and finally more efficient resource utilization and task completion effects are achieved.

The invention discloses a parallel task scheduling method based on node dynamic load and task relevance measurement. The parallel task scheduling method comprises the following steps: constructing a global scheduler for computing resource virtualization, sample allocation based on reinforcement learning and a local scheduler based on hybrid drive model scheduling on a server; in the preprocessing stage, computing resources are counted, and computing resource virtualization, GPU virtualization multiplexing technology, GPU acceleration and the like are carried out; a global scheduling stage oriented to test task parallel optimization: performing sample scheduling feasibility analysis containing fuzzy time, computational node selection based on an EXP3 algorithm, and sample scheduling oriented to node load; and a multilevel model parallelization stage based on domain knowledge and data hybrid driving, wherein a model cohesion coupling degree measurement mechanism, a model coupling mode knowledge base and a multilevel model parallelization method are included. According to the method, the global scheduler and the local scheduler are comprehensively utilized, the parallelism degree between samples and parallelism between models in the samples are achieved, and the method has wide application value and use prospects in all fields.

The invention provides a sparse matrix vector multiplication parallel optimization method and system based on a CSR storage format, and relates to the technical field of high-performance parallel computing, and the method comprises the steps: obtaining sparse matrix data stored in the CSR format; splitting and integrating the sparse matrix by adopting a row grouping method, dividing rows with similar non-zero element numbers into the same group, and configuring computing resources for each group; the method comprises the following steps: firstly, pre-loading sparse matrix data in each group, defining a texture memory, and pre-loading the sparse matrix data into the texture memory; an independent kernel is set for each group to execute SpMV calculation of all rows in each group, and non-zero elements of sparse matrix data and corresponding vector data are preloaded into a shared memory in the calculation execution process of the kernel; and the kernel executes SpMV calculation by adopting a hybrid static-dynamic thread scheduling method until all rows are calculated, and a final result is obtained.

The invention discloses an efficient knowledge distillation method for improving rotation interference resistance of a three-dimensional point cloud model, and the method comprises the following steps: S1, determining a basic teacher network and a student network, and carrying out the parallel optimization of the two networks; s2, fusing an online distillation strategy, and accurately aligning the rotation robustness features acquired by the teacher network and the features extracted by the student network in each layer of network structure through a specially designed alignment mechanism; s3, integrating difference information between the teacher network and the student network into a loss function, and constraining the network through distillation loss and cross entropy loss; and S4, removing the teacher network, and directly processing the original three-dimensional coordinate data by the student network to complete a prediction task. According to the method, the rotation interference resistance of the network can be improved, the noise margin and abnormal value interference resistance of the network can be improved, the balance of the speed and the precision can be realized, complex coordinate transformation can be avoided, extra calculation overhead can be saved, and the method has remarkable technical advantages and good industrial application prospects.

The invention belongs to the field of electronic design automation and static time sequence analysis, particularly relates to a parallel optimization method for FPGA static time sequence analysis, and aims to solve the problems of low calculation efficiency and long compiling time of static time sequence analysis in the prior art. The method comprises the following steps: performing memory layout optimization on a time sequence diagram and time sequence constraints according to a built FPGA-oriented STA data structure; task decomposition is carried out, calculation tasks of nodes and edges of each hierarchy are allocated to a plurality of parallel kernels to be executed in parallel, and single CPU threads are allocated to different hierarchies to be executed in sequence; during execution, traversing the time sequence diagram, and marking to-be-updated nodes and edges; and obtaining effective arrival time and effective demand time of the nodes through bidirectional traversal calculation, determining a key path in FPGA (Field Programmable Gate Array) time sequence static analysis, and generating a parallel optimization report of the FPGA static time sequence analysis. According to the method, the calculation efficiency in static time sequence analysis is improved, and the compiling time is shortened.

The invention provides a parallel optimization method and system for a GPGPU (General Purpose Graphics Processing Unit) instruction execution stage, and the method comprises the steps: splitting an input instruction into a plurality of sub-instructions which can be independently scheduled based on an instruction type, operand dependency analysis and a hardware resource real-time state, and generating a dependency table to record the input and output association of the sub-instructions; acquiring real-time load data of the calculation unit, the storage unit and the communication unit, and distributing the sub-instructions to the calculation unit, the storage unit or the communication unit by adopting a priority dynamic adjustment algorithm and combining instruction key path analysis; according to the dependency relationship of the sub-instructions and the availability of hardware resources, the execution sequence of the sub-instructions is adjusted to maximize the degree of parallelism; and analyzing the dependency relationship among the sub-instructions in real time to maximize the degree of parallelism to dynamically adjust a parallel execution strategy so as to realize serial execution of dependent instructions and parallel execution of non-dependent instructions. According to the method, the complexity of instruction compiling is reduced, the efficiency of instruction compiling is improved, and meanwhile extra consumed time is reduced.

The invention discloses a cloud edge end heterogeneous resource cooperative training strategy based on hierarchical hybrid parallel, and aims to research how a cross-domain distributed computing framework fully utilizes various cloud edge end heterogeneous computing resources, research an equipment graph and computing graph construction mechanism, construct a simplified and merged computing graph and equipment graph, and reduce the search space of the hierarchical hybrid parallel strategy. A search algorithm of the hierarchical hybrid parallel algorithm is researched, a possible hierarchical hybrid parallel algorithm is searched through a dynamic programming algorithm and a resource constraint mechanism, and efficient hierarchical hybrid parallel in a cloud edge-end collaborative cross-domain scene is realized; a cross-domain hierarchical hybrid parallel optimization strategy during operation is studied, the calculation speed between nodes is adjusted through calculation of data loads in clusters and secondary balanced distribution of the calculation loads, and slow nodes during operation are avoided. By means of the mechanism, efficient adaptation of the cloud edge heterogeneous computing resources and the hybrid parallel strategy is achieved, the utilization efficiency of the cloud edge heterogeneous computing resources is improved, and the cross-domain distributed training speed of the model is increased.

The invention belongs to the technical field of additive manufacturing, and particularly relates to a multi-scale parallel optimization method for a lightweight structure based on lattice discrete optimization. According to the method, through multi-scale parallel optimization design, a macro structure and a micro lattice structure are described through geometric parameters of a movable deformation rod piece, an improved two-value coding parameterization method, called a BCP method for short, is adopted to solve discrete optimization of the micro lattice structure, and geometric parameters of an optimization result are easy to extract.

The invention provides a parallel iteration solving method and system for an electromagnetic finite element equation set, and relates to the technical field of electromagnetic finite elements, and the method comprises the steps: constructing a to-be-solved electromagnetic finite element equation set # imgabs0 #; dividing the electromagnetic finite element equation set into a plurality of calculation sub-tasks, and distributing a sub-matrix block # imgabs 1 # and a sub-right-end item vector # imgabs 2 # of the sub-tasks to each process in the super computer; by means of the initialized Hoseholder-GMRES algorithm, the processes are subjected to subtask calculation in parallel, and the local optimal solution # imgabs3 # of each process is obtained; summarizing the local optimal solution # imgabs4 # of each process to obtain an optimal solution # imgabs5 # of the electromagnetic finite element equation set; according to the method, the electromagnetic finite element, the Householder-GMRES algorithm and parallel computing are combined, and the solving efficiency of the linear equation set of the electromagnetic finite element is improved by utilizing a parallel optimization technology.

The invention provides a Cholesky decomposition heterogeneous parallel optimization method and system based on an SW architecture, and relates to the technical field of high-performance computing. The method comprises the following steps: performing sub-block division on a symmetric positive definite matrix based on a distributed parallel distribution scheme, and performing iteration to complete matrix decomposition; each sub-block is distributed to different processes through an MPI programming model, data exchange is carried out between the processes through asynchronous communication, and coarse-grained task-level parallel acceleration is carried out; performing two-stage parallel acceleration on four operations in Cholesky decomposition by utilizing the acceleration parallel characteristic of a master core and a slave core of the SW architecture; wherein for GEMM and SYRK operations, column vectors of a matrix are mapped to a slave core array, and columns are divided according to the number of slave cores; the calculation process is optimized through a double-buffering mechanism, vectorization operation and a loop expansion technology, and the parallel efficiency is improved; and for the TRSM operation, the TRSM operation is decomposed into a plurality of TRSV operations, the TRSV operations are allocated to the slave cores for parallel execution, and a circular reading and data broadcasting mode is adopted to reduce data dependence and realize efficient parallel calculation.

The invention relates to the technical field of airspace flight plan management, and provides a flight plan conflict detection method based on a three-dimensional subdivision grid and a related device thereof. The method comprises the following steps: discretizing flight plan track data and limited flight area data to obtain three-dimensional subdivision grid data of the flight plan track data and the limited flight area data; a grid buffer area algorithm and the distance of the grid buffer area are used for generating a flight plan track and a grid buffer area for limiting the three-dimensional subdivision grid data of the flight area; dividing the three-dimensional subdivision grid data and the gridding buffer area of the flight plan track and the three-dimensional subdivision grid data and the gridding buffer area of the limited flight area into a plurality of detection blocks based on the three-dimensional subdivision grid hierarchy; and for each detection block, traversing the three-dimensional subdivision grids of the flight plan track data at each moment to obtain conflict detection results in the detection blocks, and outputting attribute information of the three-dimensional subdivision grids with conflicts. The problems that space-time relation calculation is dense, invalid redundant calculation is more, and the parallel optimization effect is poor can be solved.

The invention discloses a power distribution network photovoltaic openable capacity dynamic evaluation method fusing voltage stability margin and neural network optimization, and aims to solve the problems that a traditional method does not fully consider dynamic stability and is low in calculation efficiency. Firstly, a prediction method combining kernel density estimation and quantile regression is adopted to accurately quantify the uncertainty of distributed photovoltaic output. The core innovation of the method is that on the basis of traditional security constraints, a static voltage stability margin (VDSM) is introduced as a key constraint condition, and a double-layer interval analysis model capable of guaranteeing the dynamic stability of a power grid is constructed. Secondly, in order to efficiently solve, the invention provides a framework of'neural network pre-screening + parallel optimization ': after a model is decomposed into optimistic sub-problems and pessimistic sub-problems through an interval decoupling technology, massive candidate solutions are quickly screened by utilizing a neural network model, so that a feasible solution space is greatly reduced, and the solution efficiency is improved; and carrying out parallel optimization solution on the sub-models in combination with an improved particle swarm optimization algorithm.

The invention relates to the technical field of model training, and particularly provides a joint optimization distributed training method based on heterogeneous hybrid parallel and communication. The method comprises the following steps: constructing a heterogeneous parallel optimization strategy through a heterogeneous parallel optimization algorithm HPO; constructing a dynamic dual-heterogeneous synchronous communication mechanism through a dynamic dual-heterogeneous synchronous strategy DHS; according to the heterogeneous parallel optimization strategy and the dynamic dual-heterogeneous synchronous communication mechanism, a hybrid parallel and communication distributed training framework in the dual-heterogeneous environment is constructed to perform joint optimization distributed training, the method optimizes computing resource scheduling, reasonably and dynamically solves the problem of communication delay, and the efficiency of distributed training is improved.

The invention provides a sparse triangle solution parallel optimization method and system based on a GPU, and relates to the technical field of high-performance parallel computing. Obtaining a to-be-solved linear equation set; decomposing a sparse coefficient matrix in the system of linear equations to obtain a triangular sparse matrix; based on the triangular sparse matrix, sparse triangular parallel solving is carried out, and a final solution is obtained; wherein the sparse triangle parallel solution is characterized in that the hierarchy of each non-zero element is determined according to the dependency relationship among the non-zero elements, sub-tasks executed in sequence are divided based on the hierarchy, parallel solution is carried out by adopting the non-zero elements in the same hierarchy in a single sub-task, and synchronous solution is carried out among blocks among different hierarchies; obtaining a final solution through the kernel of the GPU; according to the method, the multi-thread parallelism of sparse triangle solution on the GPU can be further improved, and the problems of load imbalance and low resource utilization rate can be effectively avoided.

A method for topological design optimization in which a number of objective scattering(S) parameters are initially decomposed into a product of smaller devices (primitives). Each of the primitives are optimized in parallel using inverse design; and the topologies of the smaller devices are then composited into a final device design region, in which the final device design region is an initialization of a final optimization, advantageously saving computation time. According to at least one embodiment, the herein described optimization procedure can be utilized in order to design a composite power combiner that is essentially phase-insensitive.

The present invention discloses a large model prompt word engineering method and system, the method comprising: receiving a task requirement input by a user for interacting with a large model, generating a set of initial prompt word sets based on the task requirement; performing feature mapping on the initial prompt word set, constructing a prompt word feature vector set, decomposing the feature vectors in the prompt word feature vector set into a number of sub-feature vectors to capture the different characteristics of the prompt words; performing multi-objective optimization on the sub-feature vectors using a parallel optimization algorithm to generate an optimized prompt word set; screening the optimized prompt word set for semantic consistency and generation diversity, and determining a number of final prompt words as the large model input corresponding to the task requirement. By using the embodiments of the present invention, it is possible to realize the intelligent generation and optimization of the prompt word set according to the user's task requirement, improve the efficiency and quality of prompt word construction, and provide more flexible support for the practical application of the large model.

The invention provides a CNN-oriented batch matrix multiplication parallel optimization method and system on a SW architecture, and belongs to the technical field of artificial intelligence parallel optimization. Comprising the following steps: respectively converting an input feature map and a convolution kernel in a convolution layer into an input matrix and a weight matrix, and processing the input matrix and the weight matrix into a plurality of groups of independent matrix multiplication tasks in batches; the main core encapsulates a matrix multiplication task into a parameter structure array, the parameter structure array is transmitted to the slave core through single DMA, and the slave core divides rows of an input matrix into row block tasks by adopting a dynamic row block division algorithm according to the total number of threads and the height of the matrix; and executing sub-matrix multiplication calculation on the distributed independent row blocks, asynchronously prefetching matrix sub-blocks by adopting a double-buffer DMA (Direct Memory Access), and executing matrix multiply-accumulate calculation. The parallel processing efficiency of batch matrix multiplication between the master core and the slave core of the SW processor can be improved, and the algorithm performance is optimized.

The invention discloses a photovoltaic optimization regulation and control method and device based on an MPPT technology, and particularly relates to the technical field of photovoltaic power generation control, and the method specifically comprises the following steps: S1, full-link operation state synchronous collection, S2, DC bus ripple feature extraction, S3, multi-source disturbance quantitative evaluation, S4, cooperative control parameter decision making, and S5, grid-connected signal synthesis modulation. A multi-source disturbance quantitative evaluation system is adopted, the irradiance change gradient, the frequency ratio and the switching noise contribution ratio are subjected to multi-dimensional fusion, a dynamic weight distribution mechanism is established, and parallel optimization of MPPT mode selection, disturbance step length adjustment and PWM phase shift angle is achieved through a cooperative control parameter decision architecture. A response delay bottleneck existing in a traditional serial decision mode is broken through, and a ripple compensation component is dynamically injected while stable output of fundamental current is maintained based on a grid-connected signal modulation strategy of vector synthesis.

The invention discloses a non-uniform grid data calculation method based on a hierarchy calling rule and a related device, and the method comprises the steps: firstly analyzing an input parameter of a calculation task, and selecting corresponding calculation models according to a parameter type, the calculation models comprise a C-type model, a B-type model and an A-type model, and the models are divided according to a dependency relationship; thirdly, constructing a directed acyclic graph according to a hierarchical calling rule, representing a dependency relationship between tasks and excluding cyclic calling; then, a task execution sequence is generated based on the graph, and is ranked according to a hierarchical call rule. And performing parallel optimization on the task execution sequence according to the computing resource state, and allocating the task execution sequence to a distributed computing engine for execution. And finally, storing the result in a database, and supporting a subsequent task to directly call a historical result. The method aims at solving the problems of disordered dependency, high redundancy, low efficiency and the like of calculation tasks, and the organization efficiency and performance of building energy-saving calculation are improved.

The present application belongs to the technical field of additive manufacturing, and particularly relates to a lightweight structure multi-scale parallel optimization method based on lattice discrete optimization. The present application uses the geometric parameters of movable and deformable rods to describe macroscopic structures and microscopic lattice structures through multi-scale parallel optimization design, and uses an improved double-value coding parameterization method, referred to as BCP method, to solve the discrete optimization of the microscopic lattice structure and easily extract the geometric parameters of the optimization results.

The invention provides an attitude control method, system and equipment of an underwater leveling machine and the underwater leveling machine, and relates to the technical field of ocean engineering. The method comprises the following steps: establishing a dynamic model; constructing a physical information neural network based on the dynamic model; constructing a digital twinborn body of the leveling machine, and performing simulation verification on the attitude error prediction result in the digital twinborn body; based on a simulation verification result, taking control parameters of the underwater leveling machine as optimization variables, performing parallel optimization on a plurality of control targets of the underwater leveling machine, and generating a control solution set; and determining a target control parameter combination from the control solution set, and generating a control instruction for controlling the underwater leveling machine to perform attitude adjustment based on the target control parameter combination. According to the method, the reliability of attitude control of the underwater leveling machine can be improved through prediction of the integrated physical information neural network, simulation verification of the digital twin and multi-objective optimization.

The invention relates to a deep learning model reasoning method and device, computer equipment and a storage medium. The method comprises the following steps: calling an inference framework matched with a domestic accelerator, and performing compilation optimization, parallel optimization, memory hierarchical optimization and deep fusion of calculation acceleration components on a deep learning model to obtain a target optimization model; compiling the target optimization model into an executable machine code of the domestic accelerator; and loading the executable machine code on the domestic accelerator, and executing a data reasoning process of the executable machine code on to-be-reasoned data to obtain a model reasoning result. By adopting the method, the problem that an existing general reasoning framework is difficult to directly adapt to the domestic accelerator can be solved, the calculation potential of the domestic accelerator is completely released, and the reasoning efficiency and the resource utilization rate are remarkably improved.

The invention discloses a numerical control machine tool synchronous evolution control method and system associated with dynamic and static errors, relates to the technical field of information processing management, and is used for solving the problem that a machine tool structure optimization method is poor in structural integrity, structural connectivity and dynamic response characteristic analysis. By monitoring and analyzing geometric structure parameter changes of all parts in the machine tool operation process, effective geometric structure parameters with the high influence degree are screened out, preliminary optimization is carried out, a geometric structure optimization evaluation model is established, and an optimization result is evaluated; the method comprises the following steps of: acquiring local rigidity and weight, establishing an optimization evaluation model, defining geometric structure parameters of a plurality of components as an integral assembly structure, constructing a parallel optimization model by utilizing a grid division and rigidity matrix calculation method, and finally obtaining optimization results of a macroscopic unit and a connecting unit by taking minimization of structural flexibility as a target. Key parameters can be accurately monitored and optimized, the structural connectivity is considered, and the overall performance and stability are improved.

The invention discloses a distributed power storage system Praft parallel optimization method and system, and relates to the technical field of computer storage, and the method comprises the following steps: electing a leader, receiving a state machine command request of a client, converting a command into a log item, copying a log, submitting the log, copying the log in a cluster, and carrying out parallel optimization on the Praft in the cluster. Log items sent by the client are copied, logs are executed on all the nodes according to the same sequence, whether connection between the client and the leader node is abnormal or not is monitored in real time, the leader node needs to be selected again when the connection is abnormal, request instruction processing is carried out, and correctness analysis is carried out on a consensus algorithm. According to the method, the optimization of log processing in the distributed power storage system is realized by electing the leader, receiving the command request of the state machine of the client, converting the command into the log item, copying the log and submitting the log, the average delay of log processing is shortened, and the efficiency and consistency of log processing are ensured.

The invention discloses an adaptive two-stage matrix division method based on GEMM (General Empirical Matrix Modulation). The method comprises the following steps: determining the size of a block for the first time; carrying out first blocking processing on the input matrix from a global memory to a shared memory; loading the matrix partitioned for the first time to double buffer areas of a shared memory; determining the size of a secondary block; partitioning the partitioned matrix blocks from the shared memory to the register again; and loading the matrix blocks which are partitioned again to a buffer area of the register for calculation. The invention aims to provide a parallel optimization method based on hardware-software collaboration aiming at the bottlenecks of the current GPU-based GEMM in the aspects of data transmission efficiency, computing resource utilization rate, load balance and the like. By introducing an optimization scheme of a self-adaptive two-stage matrix partitioning mechanism in a matrix calculation process, the method aims to improve data locality, reduce memory bandwidth pressure and realize dynamic balance of loads among threads. According to the method, by reducing data transmission delay and optimizing allocation of computing resources, the efficiency of GEMM operation is improved from the source, and the efficient requirements of deep learning and large-scale scientific computing are met.