45 results about "Arbitrary-precision arithmetic" patented technology

The invention relates to the technical field of digital twinning, in particular to a high-performance digital twinning system rendering method based on dual-grid and neural network mapping, and the method comprises the steps: constructing a high-precision calculation grid and a low-precision rendering grid, and building a space mapping relation between the two grids; executing multi-working-condition numerical simulation to obtain a high-dimensional simulation result snapshot matrix; the dimension reduction projection operator is used for reducing the dimension of a high-dimensional simulation result of any working condition into an r-dimensional feature vector; generating low-dimensional reference field data by using the space mapping relation; training the neural network until a mapping network parameter is obtained through convergence; obtaining current working condition parameters of the equipment in real time, and obtaining a current r-dimensional feature vector by using the dimensionality reduction projection operator; and quickly generating physical field prediction distribution on the low-precision rendering grid and performing pseudo-color rendering. According to the scheme, the problem that a traditional method is low in rendering efficiency is solved, and the method has the advantages of reducing computing resource consumption and keeping visualization precision.

The invention provides a comprehensive pipe gallery safety state evaluation method and system, and relates to the technical field of comprehensive pipe gallery safety state evaluation methods, and the method comprises the steps: employing a steady-state heat release method to calculate the simulation temperature of each three-dimensional grid node; a finite difference method is adopted to calculate a temperature field distribution coefficient of each three-dimensional grid node, and a temperature distribution coefficient of a three-dimensional grid center point is obtained through an arithmetic average method; a coupling damage method is adopted to construct a coupling coefficient matrix of any grid position, equivalent damage energy of any grid position is calculated through an inner product method, and the equivalent damage energy is mapped into a safety coefficient of any grid center point through normalization. By adopting three-dimensional grid discretization, complex geometric characteristics of the pipe gallery structure can be more accurately captured, the position of a potential fire source is clearly marked, the hidden danger recognition efficiency is improved through targeted monitoring, high-precision calculation of a simulated temperature field under the complex boundary condition is achieved, equivalent damage energy is mapped into a dimensionless safety coefficient, and cross-regional risk comparability is achieved.

The invention belongs to the technical field of in-memory computing, and discloses a multi-precision in-memory computing circuit with balanced output bit width, a memory and a chip, which are characterized in that through the collaborative change of a broadcast mode of an input side and an accumulation mode of an output side, under different computing precisions, the multi-precision in-memory computing circuit with balanced output bit width is obtained; the external interface bandwidth (the number of ADCs / the number of adders) of the physical array is kept constant. According to the invention, the contradiction between high bandwidth demand in a low-precision mode and low bandwidth demand in a high-precision mode in a multi-precision in-memory computing architecture is solved, the idle waste of the output bandwidth during high-precision computing is avoided, the relatively balanced bit width of an output interface is realized, and the area cost of a chip can be greatly reduced.

The application discloses a lightweight method of river surface flow velocity field depth reconstruction model, relates to the computer science and fluid mechanics cross technical field, collects river surface PTV and PIV data, and constructs a flow velocity field matrix of the target river surface; based on the matrix block calculation strategy and the flow velocity field matrix, the target river surface is divided into a plurality of sub-matrix regions; based on the pattern recognition technology and the flow velocity gradient characteristics, each sub-matrix region is distinguished. The application divides the global flow velocity field into a plurality of sub-matrix regions through the matrix block strategy, adaptively determines the optimal block size according to the GPU parallel thread number, memory bandwidth and other constraints, maximizes the hardware utilization rate, and in the calculation process, adopts a differentiated resource allocation mechanism, concentrates high-precision calculation in a high-error clustering area, and adopts a simplified model in a low-error clustering area, significantly reduces the overall calculation load, and realizes efficient balance between calculation efficiency improvement and resource dynamic scheduling.

The invention provides a slow action generation method based on motion focus area analysis. The method comprises the following steps: determining a video of a slow action to be generated; detecting a motion focus area in the video, and grading the motion focus area; and for different grades of motion focus areas, different strategies are adopted to carry out motion modeling and differential interpolation processing, and a slow motion frame sequence is generated. According to the invention, through identification and grading processing of the motion focus area, high-precision calculation of the core motion object, simplified calculation of the secondary area and simplest processing of the static background, compared with a traditional scheme of uniform and dense calculation of the whole frame, the method has the advantages of high calculation complexity and lower calculation amount, and guarantees the visual effect of the core area.

The invention discloses a lightweight method for a river surface flow velocity field depth reconstruction model, and relates to the technical field of crossing of computer science and fluid mechanics, and the method comprises the steps: collecting the PTV and PIV data of a river surface, and constructing a flow velocity field matrix of a target river surface; dividing the surface of the target river into a plurality of sub-matrix areas based on a matrix partitioning calculation strategy and the flow velocity field matrix; and based on a mode recognition technology and flow velocity gradient features, distinguishing each sub-matrix region. According to the method, a global flow velocity field is divided into a plurality of sub-matrix areas through a matrix partitioning strategy, the optimal partitioning size is adaptively determined according to constraints such as GPU parallel thread count and memory bandwidth, the hardware utilization rate is maximized, and in the calculation process, a differentiated resource allocation mechanism is adopted, high-precision calculation is concentrated in a high-error clustering area, and the calculation efficiency is improved. A simplified model is adopted in the low-error clustering area, the overall calculation load is remarkably reduced, and efficient balance between calculation efficiency improvement and resource dynamic scheduling is achieved.

The invention provides a multi-level virtual-real coordinate mapping and drift correction method, system and platform, a dynamic floating original point mechanism is established, the original point position of a virtual coordinate system is adjusted in real time, so that a scene around a user is always in a high-precision calculation range of a single-precision floating-point number, and the user experience is improved. The dynamic floating origin mechanism comprises a coordinate mapping engine, an origin decision module and a seamless switching manager. A multi-level real world-virtual-real space-virtual environment conversion link is directly opened, large-scale development of augmented reality XR contents of cities / parks / outdoor / indoor / temporary spaces, automatic driving, unmanned aerial vehicles, intelligent robot navigation and the like is promoted, and the problems of large-space rendering jitter, low development efficiency and the like in the prior art are solved. In order to achieve the purpose, the invention provides an underlying OS operating system, and provides a cross-multi-level real world coordinate system and virtual environment coordinate system mapping method, a drift correction system and a developer operating platform.

The invention discloses a multi-scale space target relative pose estimation method and system using parallel double encoders, and belongs to the technical field of space target pose estimation. The invention aims to solve the problems caused by distance change, illumination difference and insufficient multi-scale feature expression of a space target in a dynamic environment. According to the technical key points, a parallel double-encoder architecture is adopted, and local and global features of a space target are extracted; constructing a feature fusion module to fuse features extracted by the CNN and the Transform; multi-scale features are fused through FPN, and angular points are predicted; training and optimizing the network; and solving the pose by using a random consistency perspective n-point method. A plurality of non-coplanar control points are selected, other points are expressed as a linear combination of the control points, a linear equation set is constructed, and a rotation matrix and a translation vector between a target and a camera coordinate system are solved by using a singular value decomposition method. The deep learning network algorithm provides an efficient and accurate solution for space target pose estimation, the algorithm can deeply mine rich information of image data by adopting an encoder-decoder structure, and finally high-precision calculation of the six-degree-of-freedom pose of the target is achieved.

The application discloses a kind of quantization bit width dynamic selection neural network acceleration hardware architecture and method, hardware architecture includes global storage module, data scheduling module, local storage module, dynamic quantization prediction controller and computing unit array;The feature map in network is divided into neurons with block as unit, in each block, continue to be divided into group with space dimension as unit, define group as the minimum unit of executing dynamic quantization operation;Each block in all feature maps of target network is configured with a trainable threshold parameter, and according to the given basic quantization bit width and target bit width total quantity constraint, the upper and lower bounds of the sparsity selectable by each block are determined;Establish the training and inference model of dynamic quantization neural network, divide inference into high-precision and low-precision calculation two parts, whether the low-precision calculation is executed is judged by the result of high-precision calculation;On the premise of ensuring that the accuracy is not lost, the inference time in actual hardware is reduced as much as possible.

The invention discloses a symmetric magnetic field calculation method, which comprises the following steps of: analyzing the symmetry of magnetic field source distribution, and selecting a reference source in a group of symmetric sources; calculating the magnetic field intensity of the reference source at any field point in the space; based on symmetry, mapping the field points to equivalent field points corresponding to the symmetric sources through a rotation matrix, and calculating magnetic fields of the symmetric sources at the equivalent field points; restoring the magnetic field at the equivalent field point to an original field point direction through an inverse matrix of the rotating matrix; and finally, superposing the magnetic field contributions of all symmetrical sources at the original field point to obtain the total magnetic field of the point. According to the method, the symmetry of a magnetic field source is fully utilized, a rotating group processing method in the group theory is adopted, accurate calculation is performed in combination with the Bios-Savart law, the symmetry and direction correctness of the magnetic field can be strictly ensured, and high-precision calculation of the extremely low-intensity magnetic field is realized; the problems of errors and direction errors existing in symmetrical field and weak field calculation of a traditional finite element method are solved.

The invention belongs to the technical field of hydraulic structure numerical analysis, and particularly relates to a two-dimensional temperature field adaptive grid merging method. The method is based on an independent coverage manifold method, an error estimator is constructed through the heat flow continuity degree between adjacent grids to replace an empirical formula, full-domain automatic error estimation is achieved, the accuracy of judging whether grid merging is carried out or not is enhanced, the method comprises the steps of error estimation, grid marking, grid shape judgment, grid merging and the like, and the method is suitable for large-scale popularization and application. And by merging the local grids, the calculation precision and the calculation efficiency are both considered, and local high-precision calculation is realized.

The invention relates to the technical field of artificial intelligence accelerators, in particular to a configurable mixing precision calculation array circuit based on an outer product. According to the method, the basic int4 multiplier is taken as a reference, the results of the basic multiplier are fused according to different configuration modes to support high-precision calculation, calculation of 12 different specifications can be supported at most, the resource utilization rate of the multiplier is increased, and the resource utilization rate of the multiplier reaches 100%; by expanding a 1-bit sign bit, the array supports multiplying a signed number by a signed number, multiplying an unsigned number by a signed number and multiplying an unsigned number by an unsigned number, and various calculation requirements of the neural network are flexibly adapted; a matrix transposition switch is built in, so that the overhead of external transposition is reduced; the bit width of output data is limited through functions such as saturation truncation, and storage overhead is reduced. According to the method, the calculation efficiency is improved, at most four kinds of calculation precision and 12 kinds of calculation modes are supported, and the resource utilization rate is improved by 2-4 times; and application scene extension, perfect signed / unsigned calculation support and dynamic reconfiguration during operation are supported.

The application discloses a single-precision hardware implementation double-precision simulation method based on floating-point decomposition and reconstruction, including a 32-bit single-precision server; firstly, a prediction model is acquired to train floating-point 64-bit double-precision scalar or tensor original data, and the implementation double-precision simulation method comprises the following steps: step S1, data preprocessing; step S2, computation graph reconstruction; two floating-point 32-bit single-precision data are rewritten into a double-channel computation form by using a traditional computation flow; step S3, dynamic precision recovery; step S4, adaptive quantization control; an error feedback mechanism is introduced, when the error exceeds a threshold value, a dynamic decomposition result of computation is triggered, and the decomposition result is iterated to two floating-point 32-bit single-precision data, the application meets the high-precision computation demand of the single-precision server; the application is completely based on a single-precision instruction set and does not need special hardware support, and is suitable for an NPU cluster; compared with original double-precision storage, the memory occupation is reduced by 25%.

This invention discloses a high-precision approximation calculation device for the softmax function, comprising a first module, a second module, and a third module. The first module includes a first zero-value prediction unit, a first softmax function calculation unit, and a first selector unit. The second module includes a second softmax function calculation unit. The third module includes a second zero-value prediction unit, a third softmax function calculation unit, and a second selector unit. The first and second zero-value prediction units are used to predict whether the output data of the device is zero based on their corresponding input data, so as to skip subsequent calculations. The first, second, and third softmax function calculation units are used to implement high-precision calculation of the softmax function, and the first and second selector units are used to select the final output data. This invention can greatly reduce the power consumption and area overhead of the computing device, and improve the energy efficiency of the device.

The invention discloses a high-precision calculation method and device for a normalized arc tangent function. The method comprises the steps that S1, input floating point data is decomposed and preprocessed; the symbol, the order code and the mantissa of the input floating point data are obtained through data conversion, and the order code is compared with the single-precision offset code to obtain a shift code; obtaining input data for second-order polynomial calculation by using the shift code; s2, simplifying a calculation interval by using the characteristics of an odd function; s3, layering and segmenting the effective interval and selecting a polynomial coefficient; s4, carrying out second-order polynomial and fifth-order Taylor dual-path parallel calculation; step S5: result formatting; and S6, dynamically selecting and outputting. The device is used for executing the method. The invention has the advantages of high precision, low time delay, small area, low cost and the like.

The invention belongs to the field of numerical calculation of structural dynamics, and discloses a high-precision calculation method, application and equipment for structural dynamic response considering a load derivative. According to the equation, loads and load derivatives at the current moment and the next moment are considered, a load matrix at the current moment and a load matrix at the next moment are constructed respectively, and the load action part of a traditional discrete state equation is improved. The algorithm provided by the invention gives consideration to the requirements of rapid calculation and high precision, can be widely applied to the fields of structural vibration analysis, vibration control, health monitoring, response prediction and the like, and has a good application prospect.

The invention discloses an automatic CFD solving method for an internal flow field of a bimodal scramjet engine, and the method comprises the steps: constructing an AnsysFluent automatic solving environment through employing a Python script, loading a grid file of a bimodal scramjet engine model, configuring a physical model and boundary conditions of the bimodal scramjet engine through employing the Python script, and carrying out the automatic solving of the internal flow field of the bimodal scramjet engine. A Python script is utilized to complete setting of a numerical discrete format and a solving method and construction of a flow field initial value of a computational domain; taking the flow field initial value as an iteration starting point, and executing dynamic climbing calculation under first-order precision by utilizing a Python script to obtain stable initial flow field distribution; the stable initial flow field distribution is used as an initial calculation value, and a Python script is used for switching second-order high-precision calculation and carrying out convergence judgment. According to the method, the problems of tedious manual intervention, difficulty in shock wave capture and subjective convergence judgment in high-Mach-number strong compressible flow calculation are effectively solved.

The present disclosure relates to an information processing apparatus, an information processing method, and a computer-readable storage medium. The information processing apparatus according to the present disclosure includes processing circuitry configured to: obtain a sound field in a virtual space, by using a sound field synthesis model, based on sound source information about a sound source in a real space and virtual space information indicating an object present in a virtual space. The sound field synthesis model is pre-obtained through machine learning by using a sound field in a virtual space obtained from high-precision calculation as learning data.

The application provides a construction safety visual calculation and resource scheduling method, system and electronic equipment, which comprises the following steps: deploying and initializing the calculation and resource scheduling system; the edge end collects video streams in real time and performs preprocessing, light-weight model inference and local alarm; the edge end performs three-level screening on the video data, intelligent compression of the region of interest (ROI) and end-to-end encryption before uploading to the cloud; a multi-objective optimization decision model is constructed based on FAHP, and the allocation strategy of the task in the edge end and the cloud is dynamically adjusted according to the real-time collected network state parameters, computing power state parameters and event characteristic parameters; the cloud adopts a high-precision model for inference, links a knowledge base to generate a rectification scheme, and realizes full-process closed-loop management; an incremental data set is constructed to fine-tune the light-weight model and the high-precision model. The scheme provided by the application can balance the contradiction between the real-time alarm demand of the construction site and the high-precision calculation demand of the complex model.

The invention discloses a training and reasoning system, a training method and a reasoning method, and relates to the technical field of artificial intelligence, the training and reasoning system comprises a resistive random access memory storage layer, a dynamic random access memory storage layer and a data scheduling module; the resistive random access memory storage layer and the dynamic random access memory storage layer are heterogeneously integrated; the resistive random access memory storage layer is used for storing the weight after training stabilization and executing multiplication and addition operation; the dynamic random access memory storage layer is used for storing intermediate results obtained in training and reasoning processes; and the data scheduling module is used for storing the weight to be updated in the storage layer of the dynamic random access memory, storing the weight after training stabilization in the storage layer of the resistive random access memory, and storing an intermediate result obtained based on multiplication and addition operation in the training and reasoning process in the storage layer of the dynamic random access memory. The system can meet the requirements of high bandwidth, low power consumption, high capacity and high-precision calculation at the same time.

The application relates to a high-precision calculation method of an edge-smoothed edge element static magnetic field based on an MVP formula, which divides grid data into a plurality of polyhedral smoothing domains with corresponding edges as centers according to each edge in the grid data, and the polyhedral smoothing domain is formed by nodes, body centers and face centers of tetrahedral units in the grid data around the corresponding edge; the vector function gradient of the corresponding edge is calculated based on the vector function of each tetrahedral unit in the polyhedral smoothing domain corresponding to the edge; the static magnetic field equation is constructed based on the MVP formula and the Coulomb norm, and the static magnetic field equation is rewritten according to the vector function gradient of each edge and the Galerkin method, so that the system equation of the static magnetic field is obtained; the system equation of the static magnetic field is further rewritten into Newton iteration linear equations; the Newton iteration method is used to iteratively solve the Newton iteration linear equations until convergence, so that the magnetic vector potential in the static magnetic field is obtained. The method realizes the high calculation precision and stability requirement of an engineering static magnetic field problem.

Techniques for replacing input values being loaded into a computational circuit are described. Small input values such as denormal numbers can be replaced with alternative values such as zeros to reduce switching activity in the computational circuit, and thus reduce power consumption. In applications such as most neural networks, the impact on the prediction results when replacing small numbers with zeros can be negligible. In applications where high precision computations may be desirable, the input values can be loaded into the computation circuit without modification.

Apparatuses, systesms, and techniques to simulate high-precision calculations with a series expansion of lower precision tensor cores. In at least one embodiment, one or more multiplication operands of a first precision are represented by a sum of two or more operands of a different precision.

This application relates to the field of intelligent transportation technology, and discloses a method, device, system, and storage medium for real-time collaborative optimization of traffic emissions. The method includes collecting vehicle operation data to calculate vehicle power, constructing a two-dimensional bin-structured index of the instantaneous emission rate as a baseline, and correcting it in real-time using pre-fixed rules to output single-vehicle, single-step emissions at the microscopic level; aggregating emission data according to spatial hierarchical logic and simultaneously calculating performance indicators such as average commuting time; calling templates to generate strategies when indicators exceed limits; compiling the strategies into a set of six-tuple standard actions and performing fairness and conflict checks; using an exactly-once mechanism and consistency window to ensure atomic execution across devices; and evaluating strategies based on in-loop monitoring, combined with a third-generation non-dominated sorting genetic algorithm and a conditional risk-value risk term, outputting a Pareto optimal solution set that is continuously updated. This invention balances traffic efficiency, environmental benefits, and system execution stability through a high-precision calculation model and robust optimization algorithm.

A computer-implemented method and device performing digital post-processing of an in-memory computing crossbar array. The computer-implemented method includes providing a digital computing block positioned at a periphery of the in-memory computing crossbar array. The digital computing block is configured to perform fixed-point computations of an input, compression on the fixed-point computations of the input; and a nonlinear activation function.

The invention discloses a rapid high-precision calculation method and system for a variable bottom exponential function. The method comprises the following steps: scaling one of three initial values for subsequent calculation; the exponential function of any fixed base number meeting the convergence range is calculated through multiple iterations, the number of iterations is reduced based on the optimization of the CORDIC exponential function of the radix-8 hyperbolic rotation mode, and the efficiency is improved; and obtaining a final result from the converged variable value. According to the method, on the basis of keeping the flexibility of the generalized CORDIC, any fixed base number exponential function meeting the convergence condition can be directly calculated; besides, single-iteration multi-bit parallel processing is realized by adopting a radix-8 basis decomposition technology, so that the total number of iterations is reduced by more than 50%, meanwhile, the calculation precision is kept unchanged, and the hardware implementation efficiency is effectively optimized.

The present invention discloses a method and system for implementing convolutional neural networks in parallel based on computational precision. This method belongs to the field of convolutional neural network technology and includes the following steps: S1: finding substructures; S2: operator conversion; S3: input and output precision conversion; and S4: low-precision calculation. The present invention divides and fuses the computational logic of a convolutional neural network into multiple independent modules for pipeline parallelization. It also quantizes modules suitable for quantization, thereby reducing throughput. The modules are then placed on high-precision computing cards or, after quantization, on low-precision cards for calculation. This method fully utilizes computing resources of different precisions to jointly complete convolutional neural network calculations, making it worthy of widespread use.

The invention discloses a high and low bit weight splicing method based on a storage and calculation integrated framework, which is used for solving the problem of calculation precision reduction caused by non-ideal characteristics of a memory. Neural network weights of offline training are stored through a multi-bit memristor device, a block storage and error compensation mechanism of high and low weights is combined, the high weights are used for storing main information, the low weights are used for error compensation, and high-precision matrix vector multiplication is achieved. The storage and calculation integrated framework comprises a storage and calculation array, an analog input and output module, a digital shift addition module, an arithmetic logic unit, a control unit and a register. Through hardware optimization and module cooperation, high-precision calculation under a low-power-consumption condition is realized. The method is suitable for various memory devices such as flash memories, resistive random access memories and phase change memories, and has wide application scenes including the fields of speech recognition, image processing, deep learning reasoning and the like.

The invention discloses a method for realizing double-precision simulation by single-precision hardware based on floating point decomposition and reconstruction. The method comprises a 32-bit single-precision server; firstly, a prediction model is obtained and used for training floating point 64-bit double-precision scalar or tensor original data, and the method for achieving double-precision simulation comprises the steps of S1, data preprocessing, S2, calculation graph reconstruction, specifically, two floating point 32-bit single-precision data are rewritten into a double-channel calculation form through a traditional calculation process; an error feedback mechanism is introduced, when an error exceeds a threshold value, a calculation result is triggered to be subjected to dynamic re-decomposition, two floating point 32-bit single-precision data are iterated through a re-decomposition result, and the high-precision calculation requirement of a single-precision server is met; the method is realized completely based on a single-precision instruction set, does not need special hardware support, and is suitable for an NPU cluster; and compared with native double-precision storage, the memory occupation is reduced by 25%.

The application discloses a kind of dual-mode scramjet internal flow field automation CFD solving method, comprising: using Python script to build AnsysFluent automation solving environment, load the grid file of dual-mode scramjet model, using Python script to configure the physical model and boundary condition of dual-mode scramjet, using Python script to complete the setting of numerical discrete format, solving method and the flow field initial value construction of calculation domain;Flow field initial value is used as the starting point of iteration, using Python script to execute the dynamic climbing calculation under first-order accuracy, obtain stable initial flow field distribution;Stable initial flow field distribution is used as the initial value of calculation, using Python script to switch second-order high-precision calculation and carry out convergence determination.This method effectively solves the problems of artificial intervention tedious, shock capturing difficulty and convergence determination subject in high Mach number strong compressible flow calculation.