Rapid imaging processing system for multi-GPU parallel computing

Abstract

The invention provides a rapid imaging processing system for multi-GPU parallel computing. The rapid imaging processing system comprises an acquisition module, a receiving module, a GPU group and a CPU. The acquisition module is used for acquiring echo data and sequentially processing the echo data to obtain sequential data; and the receiving module is used for receiving the sequence data and directly loading the sequential data. According to the invention, echo data are acquired through the acquisition module; echo data are processed in sequence and the processed data are sent to the receiving module; after receiving echo data, the receiving module directly loads the echo data into the GPU1 through a PCIe interface and an RDMA technology, so that echo data loading time is greatly saved and a plurality of GPUs in the GPU group are subjected to parallel processing; the topological structure among the GPUs and the parallel computing method are optimized, so that imaging computing time and the number of the GPUs are basically in a linear relationship, and the computing efficiency of the GPU group is greatly improved.

Description

technical field [0001] The invention belongs to the technical field of fast imaging processing, in particular to a fast imaging processing system for multi-GPU parallel computing. Background technique [0002] The two-dimensional multiple input multiple output (Multiple input multiple output, MIMO) array can obtain high-resolution three-dimensional images of the target, and with its fast data acquisition capability, it can be applied to rapid human body security inspections. The detection system using MIMO array, its transmitting end and receiving end both adopt multi-element antenna structure, and use multiple mutually orthogonal transmitting signals to detect the target at the same time, and the multi-channel orthogonal signals are scattered by all arrays at the receiving end after being scattered by the target Yuan received. Each received signal passes through a set of matched filters to sort the echo signals corresponding to each transmitted waveform, thus introducing m...

AI Technical Summary

Problems solved by technology

A common method is to connect the data acquisition board and the GPU processing system through Ethernet. There are two obvious disadvantages of this method. One is that when the number of antenna arrays in the imaging system is large, the network transmission time may be as long as tens of Milliseconds or even longer, this is a large delay that cannot be ignored; second, the data transmitted by the network will first be loaded into the CPU, and then loaded into the GPU from the CPU. In particular, when there are more than one GPU, the data needs to be loaded separately. In these GPUs, this data path is lengthy and inefficient

Images