The detailed algorithm implementation is illustrated below:
In the calculation of dense similarity, most of the internal loop size is set by the config variables, so that the reference vertex is aligned with others. The source vertex is initialized by multiplying the value of coefficient. Only integer value can be processed in the kernel, and all the calculation is using LUT arithmetics. In the integer version, the 32-bit input is accumulated by 64-bit registers, and the output float similarity is the division result of two 64-bit integers. The overall diagram of dense similarity kernel has a insert sort module, which returns the top K number of similarity values. The maximum number of K is a template number that can be changed by rebuilding the xclbin. The default value of top K is 32.