Vectorization realization method for deconvolution matrix of GPDSP

Abstract

The invention discloses a vectorization realization method for a deconvolution matrix of a GPDSP. A CPU core of the GPDSP distributes corresponding scalar storage space and vector storage space for a weight matrix generated at a forward propagation stage and a residual matrix at a reverse calculation stage in a convolutional neural network. The method comprises the steps that S1, a residual matrix A(m,m), a weight matrix B(n,n) and a deconvolution result matrix C(m+n-1,m+n-1) are set, wherein m is greater than n; S2, by controlling the number of cycles, elements in the first (n-1) rows of the deconvolution result matrix C are calculated first; S3, the number of cycles is fixed, and elements from the n(th) row to the m(th) row of the deconvolution result matrix C are calculated; and S4, by controlling the number of cycles, elements from the reciprocal (n-1)(th) row to the reciprocal first row of the deconvolution result matrix C are calculated. The method has the advantages that the principle is simple, operation is convenient, a vector processor can be fully utilized to complete special data calculation, total algorithm running time is shortened, and algorithm execution efficiency is improved.

Description

technical field [0001] The invention mainly relates to the fields of vector processors and machine learning, in particular to a vectorization implementation method of a GPDSP-oriented deconvolution matrix. Background technique [0002] Deep learning (DL) is an important research direction in the field of machine learning. DL simulates the hierarchical perception of the human brain by constructing a multilayer perceptron (MLP). MLP can express attribute categories or high-level abstract features by combining low-level features, thus becoming the focus of current research in the field of object recognition. [0003] The classic DL models mainly include Auto Encoder (AE), Deep Belief Networks (DBNs) and Convolutional Neural Networks (CNN). Generally speaking, the above model mainly extracts features from the input image through the encoder, and transforms the image from the bottom layer to the high-level feature space. Correspondingly, the decoder is used to reconstruct the fe...

AI Technical Summary

Problems solved by technology

[0004] Since the CNN model involves a large number of matrix operations, such as matrix-matrix multiplication, matrix-vector multiplication, vector-vector multiplication, matrix-matrix convolution, matrix expansion, matrix deconvolution, and various transcendental function calculations, CNN The model requires a lot of computing resources

Images