Image segmentation processing device based on SGDMA (scatter gather direct memory access)

Abstract

The invention discloses an image segmentation processing device based on SGDMA (scatter gather direct memory access). The image segmentation processing device comprises a DMA (direct memory access) controller, an external memory connected to the DMA controller, a video memory / an internal memory connected to the DMA controller through a bus, and GPUs (graphic processing units) connected to the video memory / the internal memory; and in addition, the image segmentation processing device further comprises an image segmentation controller, which is connected between the DMA controller and the external memory and used for segmenting data of an entire image in the external memory into blocks in advance; then the blocks are transmitted to the video memory / the internal memory by the DMA controller and further undertake the follow-up processing by the GPUs. According to the image segmentation processing device disclosed by the invention, the data of the image is segmented before DMA transmission, and the special data is only transmitted to each GPU, so that the bandwidth utilization rates of the bus and the external memory can be improved.

Description

technical field [0001] The invention relates to digital image processing technology, in particular to an image segmentation processing device based on SGDMA, which is especially suitable for the field of large-scale digital splicing display walls. Background technique [0002] DMA (Direct Memory Access) is an ideal way of high-speed data transmission. DMA directly transmits data between memory and I / O devices, and its data operations are completed by the DMA controller without changing the CPU. Participation, thus greatly improving the CPU utilization. [0003] There are two implementations of DMA: one is block transfer DMA (block DMA); the other is SGDMA (Scatter Gather DMA, disperse aggregation DMA), which can move discontinuously stored data to continuous address space, The reverse is also fine. [0004] In the process of DMA data transfer, the source physical address and the target physical address are usually required to be continuous, but in some applications, the so...

AI Technical Summary

Problems solved by technology

The disadvantages of this image segmentation method are:

[0012] (1) This segmentation method completely transmits a pair of image data to each GPU, but in fact each GPU needs only a part of the data, which causes the data to occupy the bandwidth of the bus during the transmission process. Therefore, this method results in lower utilization of bandwidth on the bus;

[0013] (2) The image data is taken out from the external memory, and too much invalid data will also reduce the bandwidth utilization of the external memory;

[0014] (3) The image segmentation function is realized by GPU, which increases the complexity of GPU programming

Images