Convolution calculation data reuse method based on heterogeneous many-core processor

Abstract

The invention discloses a convolution calculation data reuse method based on a heterogeneous many-core processor. A CPU completes convolution calculation of a data block C through a data block A and adata block B. The method comprises the following steps: S1, according to the number of cores of the heterogeneous many-core processor, performing two-dimensional mapping into N * N, dividing the datablock A, the data block B and the data block C into N * N blocks, enabling the (i, j)th kernel to read the (j, i)th block data from the memory into the own on-chip memory, wherein convolution calculation of the data block C (i, j) needs a data block A (i, k) and a data block B (k, j), and k = 1, 2,..., N; and S2, entering a cycle k for N cycles from 1 to N, and completing the Kth convolution calculation of the data block C by using the obtained data block A and data block B. The memory access requirement of convolution calculation on the heterogeneous many-core processor is remarkably reduced, and the many-core calculation capability is fully exerted, so that the high performance of convolution calculation is realized, and the calculation performance of convolution calculation on the heterogeneous many-core processor is improved.

Description

technical field [0001] The invention relates to a method for reusing convolution calculation data based on heterogeneous many-core processors, and belongs to the technical field of deep learning. Background technique [0002] Convolution is one of the most important concepts in deep learning. During the training and reasoning process of the convolutional neural network, the convolution operation occupies most of the calculation. High-performance computing platforms usually provide specialized solutions for such core operations. For calculation-intensive functions, such as convolution in deep learning, how to provide enough data to the powerful calculation kernel in a timely manner is a problem that needs to be solved. Heterogeneous many-core processors have super computing power, multi-level storage levels, and efficient on-chip communication methods, making it possible to achieve efficient data reuse. [0003] At present, the commonly used convolution calculation optimiz...

AI Technical Summary

Problems solved by technology

For some heterogeneous many-core processors, the maximum memory access speed of the memory does not match the powerful computing capabilities of the many cores, resulting in convolution calculations that can only play 10% to 20% of the CPU's computing performance

Images