Analysis and Implementation of the Semi-Global Matching 3D Vision Algorithm Using Code Transformations and High-Level Synthesis

High-level synthesis (HLS) offers several advantages, such as faster simulation run-time and better design re-use, thanks to the higher level of abstraction. This work uses HLS to implement the Semi-Global Matching (SGM) algorithm, which is frequently used in stereo vision systems, e.g. for automotive applications. The hardware implementation is based on a Xilinx^® Virtex 7 FPGA. The initial algorithmic “golden” model used very large arrays, which had to be mapped to an external DRAM and brought into the on-chip RAM of the FPGA on demand. This required both adding the memory transfer loops and inserting calls to the AXI transactors that access the DRAM through the on-chip DDR slave. Moreover, the initial single-threaded algorithm had to be parallelized, by converting the top-level sweeps of the image in eight directions into as many threads. The access to the DRAM was then managed with a centralized controller. This modified SystemC design proved to be suitable to achieve the target real-time performance. The design space was thus explored by making several fairly different micro-architectural choices. In the end, it was possible to obtain an implementation which is comparable to a very efficient (and hence very inflexible) manual RTL design that had been previously developed, including a very sophisticated fine-grained management of data and computation.