Parallel calculation method and apparatus used for astronomical software Griddding

Abstract

The invention discloses a parallel calculation method and apparatus used for astronomical software Griddding. The method comprises the steps of mapping a gind value corresponding to each piece of astronomical sample data in an astronomical sample array to a grid point of a standard two-dimensional grid to serve as a data point, thereby obtaining a mapping graph; performing data division by a preset step length along horizontal and longitudinal coordinates of the mapping graph to obtain a plurality of memory grids taking the preset step length as the length and the width; allocating a memory grid to each thread, wherein an index range of a starting index dind of each thread is partial astronomical sample data corresponding to the memory grid; taking the data point corresponding to the sample data currently corresponding to the starting index dind of each thread as a starting data point, performing convolutional calculation according to the starting data point and the data point in the memory grid allocated to each thread to obtain a calculation result, and writing the calculation result in a memory. According to the method and the apparatus, the access conflict of each thread to the memory can be avoided as far as possible, so that the parallel processing capability and the memory access efficiency of a Gridding program are improved.

Description

technical field [0001] The invention relates to the technical field of memory access optimization, in particular to a parallel computing method and device for astronomical software Gridding. Background technique [0002] Gridding software is a very important software in astronomy. Taking the upcoming world's largest astronomical radio telescope - the Square Kilometer Array (SKA) project as an example, Gridding's calculations account for about 30% of the total calculations. [0003] There are three layers of loops in the calculation process of Gridding. The index of the outermost loop is dind, which is used to index the astronomical sample data in the astronomical sample array, the index of the middle loop is suppv, and the index of the innermost loop is suppu. In the outermost loop, the values ​​of gind and cind are calculated based on the astronomical sample data, where the samples array stores the astronomical sample data to be processed. [0004] Since samples[dind].iu ...

AI Technical Summary

Problems solved by technology

[0004] Since samples[dind].iu and samples[dind].iv are generated by random numbers, resulting in different astronomical sample data, their values ​​may be the same, so for different cycle indexes dind (or for different astronomical samples data), the calculated gind values ​​may be the same, and each gind element corresponds to a unique memory address. If the data blocks are divided according to astronomical data (or dind), although different data blocks contain different astronomical samples data, but may contain the same gind element. In this case, when different threads perform convolution calculations based on the gind elements in each data block, a memory address may be accessed by multiple threads at the same time. This leads to inconsistency in reading and writing of memory, resulting in calculation failure

Images