35results about How to "Reduce multiplication" patented technology

The invention discloses a neural network convolution operation device and method. The device is used for realizing convolution operation of a weight value matrix and a neural element matrix in a neural network via a matrix multiplication manner, and comprises a displacement operation unit, a matrix multiplication operation unit and a controller, wherein the displacement operation unit is used forrespectively carrying out winograd transformation on the neural element matrix and weight value matrix so as to obtain a transformed neural element matrix and a transformed weight value matrix; the matrix multiplication operation unit is used for carrying out a matrix multiplication operation such as counterpoint multiplication on the transformed neural element matrix and the transformed weight value matrix so as to obtain a multiplication matrix; the displacement operation unit is furthermore used for carrying out winograd inverse transformation on the multiplication matrix so as to obtain aconvolution operation result; and the controller is used for controlling the displacement operation unit to carry out the winograd transformation or the winograd inverse transformation, and controlling the matrix multiplication operation unit to carry out the matrix multiplication operation.

The invention discloses an elliptic curve-based efficient and anonymous certificateless multi-receiver signcryption method. The method is used for solving the technical problem in the prior art that an existing certificateless multi-receiver signcryption method is low in operation efficiency. According to the technical scheme of the invention, firstly, a user public key and a partial private key, generated by a private key generation center PKG, are sent to a user. Secondly, after the user receives the user public key and the partial private key, the user calculates its own private key. Meanwhile, the user verifies whether the received user public key and the received partial private key are valid or not. If valid, the operation is continued. Otherwise, the operation is stopped. Thirdly, the signcryption algorithm is adopted to design an elliptic curve and the ciphertext information does not include the identity information of a sender and a recipient. Finally, the ciphertext information is broadcasted. Only authorized recipients can fairly decrypt and verify the ciphertext information. According to the technical scheme of the invention, the signcryption process is designed to conduct based on the elliptic curve. Therefore, the multiplication operation on the elliptic curve is reduced, and the calculation efficiency is improved.

The invention discloses a neural-network operation device. The device includes: a sparseness processing unit, which is used for carrying out sparseness processing on a weight matrix to generate a binary sparse sequence, wherein "0" corresponds to an element, of which a numerical value is "0", in the weight matrix, and "1" corresponds to an element, of which a numerical value is not 0, in the transformed weight matrix; a mapping unit, which is used for generating a mapping relationship table of the sparse sequence and element positions in a neuron matrix, wherein a K-th bit of the sparse sequence corresponds to an element of an i-th row and a j-th column in the neuron matrix of M rows x N columns, and it is met that (i-1) x N + j = K or (j-1) x M + i = K; and a controller, which is used forcontrolling the sparseness processing unit to generate the sparse sequence, and controlling the mapping unit to generate the mapping relationship table.

The invention discloses a convolutional filtering optimizing method based on linear texture filtering. The method comprises the following steps of: dividing a convolutional filter into two one-dimensional filters if the convolutional filter can be separated; inspecting the weights in a convolution kernel and pairing two adjacent weights of which the values are not zero and the sizes are same so as to ensure that paired weights in the whole convolution kernel are as many as possible; calculating the paired weights to acquire coordinate offset amount required by the linear texture filtering and a new weight; performing texture searching according to the coordinate offset amount in a mode of linear texture filtering; multiplying the acquired texture value by the corresponding new weight to acquire an action result of the paired weight in the convolutional filtering; performing normal convolutional filtering on the weights which are not paired; and adding the filtering results with the results of the paired weight to acquire the whole convolutional filtering result. By the convolutional filtering optimizing method, a linear texture filtering function of a modern ground power unit (GPU) is utilized, and the times for performing texture searching, multiplication and addition which are required for convolutional filtering in the real-time image rendering process are reduced, so that the aim of improving the convolutional filtering efficiency is fulfilled.

The invention relates to a method of realizing bilinear interpolation in video encoding and decoding, belonging to the video encoding and decoding field in signal processing. The method computes and obtains the value of one pixel S in m*n reference image block through bilinear interpolation, comprising: computing the value of ac, and the value is ac=((Dy-dy)*A+dy*C)/Dy; computing the value of bd, and the value is bd=((Dy-dy)*B+dy*D)/Dy; computing the value of s, and the value is s=((Dx-dx)*ac+dx*bd)/Dx; computing S=floor(s+0.5); the invention avoids the use of multiplication and a multiplier in the process of realizing bilinear interpolation, meanwhile reduces the use of addition and an adder, thereby reducing realization cost and improving interpolation efficiency.

The invention discloses a simplified weighted repeat pseudo-median filtering method with negative coefficients. The method comprises the following steps: taking the frequency response characteristics of an approximation same-order finite impulse response (FIR) filter as an objective, designing weighting coefficients containing negative numbers for weighted pseudo-median computation of filtering window data by an MALLOWS comprehensive method, adjusting the window data by the weighting coefficients; simultaneously, computing weighting coefficients based on the slopes of a group of special structures designed for the filter, computing the overall slope of the window data repeatedly by the weighted pseudo median, carrying out linear transformation on the window data by the slope; and finally carrying out weighted computation on the transformed data by the pseudo-median weighting coefficients to obtain the output value of weighted repeat pseudo-median filtering. The method realizes a better frequency selection characteristic than that of the existing negative coefficient weighted median or pseudo-median filtering method, and simplifies computation in the algorithm to meet the requirement of systems with finite resources, such as a wireless sensor and the like for low power consumption.

The invention provides a fixed-point acceleration method and device for convolution calculation, and relates to the technical field of convolution calculation. The fixed point acceleration method forconvolution calculation comprises the following steps: carrying out first conversion operation on a floating point matrix in a preset convolution algorithm to obtain a converted fixed point matrix; Replacing the floating point matrix in the preset convolution algorithm with a fixed point matrix to obtain a replaced convolution algorithm; performing convolution calculation according to the replacedconvolution algorithm and the input data to obtain a convolution calculation result; and performing a second conversion operation on the convolution calculation result to obtain a convolution calculation result of the input data. In the embodiment of the invention, the method comprises the following steps: obtaining a target object; replacing a floating point matrix in a preset convolution algorithm by a fixed point matrix to obtain a replaced convolution algorithm; the method does not comprise a floating point matrix any more, can be directly applied to fixed-point calculation, reduces multiplication in the convolution process, can effectively reduce access and calculation amount of data volume in the fixed-point convolution algorithm, and can improve the performance of convolution calculation.

The invention discloses a fast Montgomery modular multiplier operation optimization method suitable for a national cryptographic sm2p256v1 algorithm. The method is realized through a Montgomery modular multiplier in an information security chip. After the Montgomery modular multiplier carries out initialization operation, a large number multiplication operation is completed, then one time of cyclic accumulation calculation is carried out, and finally one time of unsigned large number subtraction operation is carried out to obtain a result. Time-consuming multiplication operation in Montgomerymodular multiplication is greatly reduced, so that the operation efficiency of the Montgomery modular multiplier when the information security chip operates the sm2p256v1 curve parameters is improved,and the power consumption in the same operation process is reduced.

The invention discloses a multi-carrier signal-based low-operand underwater acoustic wake-up method, and the method comprises the following steps: constructing a corresponding multi-carrier signal for a communication node, performing absolute value summation on signals in a window at a receiving end to perform signal arrival detection, performing multiple times of frequency point detection on the signals by using real number fast Fourier transform, and realizing the wake-up detection. The method is suitable for accurate awakening of any distance within the maximum communication distance of the two underwater acoustic nodes, is small in operand, and is suitable for a low-power-consumption single chip microcomputer. The communication machine can be in a low-power-consumption dormant state underwater for a long time.

The invention discloses a neural network convolution calculation acceleration method based on a Kmeans algorithm, and the method comprises the steps: adding a Kmeans clustering layer in neural networkconvolution, and enabling input data passing through the Kmeans clustering layer to be clustered into fixed k types; and calculating the multiplication of the k types of data and the convolution kernel in the convolution layer, making the result into a lookup table, and searching and obtaining the result of the multiplication of all the original input data and the convolution kernel through the lookup table. The method has the advantages that the implementation method is simple, multiplication operation of the convolutional layer of the neural network can be effectively reduced, and the like.

The invention belongs to a video and image processing technology, in particular relates to a method for convolving a symmetrical convolution kernel of a video/audio signal by applying an SSE (Streaming SIMD Extension) instruction set. The method comprises the following steps of: dividing a floating-point signal sequence S(n) of an audio input signal into a left interval, a middle interval and a right interval and carrying out convolution processing on the left interval and the right interval by using a common C language algorithm; and carrying out convolution processing on a signal in the middle interval by utilizing the SSE instruction set according to one group of every four adjacent points. In the method, the symmetrical characteristic of the convolution kernel is utilized, and half multiplication calculation is effectively reduced; in addition, because the SSE instruction set can calculate four points once, the number of the instructions is effectively reduced, no any conditional transfer instruction exists in concrete processing, and the production line of a CPU (Central Processing Unit) can be more fully utilized.

Embodiments of the invention provide a data transmission method and a data transmission device based on user scheduling, wherein the method comprises the following steps of: determining, according tothe acquired information of the user terminal equipment, a combination of corresponding data layers from an alternative multi-user pairing set; iteratively determining a pre-coding matrix of the combination of the data layers; determining a user scheduling result according to the pre-coding matrix, and performing data transmission to the corresponding user terminal equipment according to the userscheduling result. In the embodiments of the invention, the inverse operation of the matrix is saved, the computational complexity is greatly reduced, the efficiency and speed of determining the pre-coding matrix are improved, and the pre-coding matrix of the combination of the plurality of data layers of multi-user pairing may be determined quickly enough, without a reduction in system capacity.Therefore, the invention is capable of improving the efficiency of multi-user scheduling without affecting the system capacity, and accelerating the response to the user terminal device as a whole, thereby improving the user experience.