One of the key advantages of custom design hardware accelerators, for which FPGAs are well suited, is the choice and architecture of custom data movers. These customized data movers facilitate efficient access to global device memory and optimize bandwidth utilization by reusing data. Specialized data movers at the interface with the main memory can be built at the input and output of the data processing engine or processing elements. The convolution filter is an excellent example of this. Looking from a pure software implementation point of view, it seems that to produce a single sample at the output side requires 450 memory accesses at the input side and one write access to the output.
Memory Accesses to Read filter Co-efficients = 15x15 = 225
Memory Accesses to Read Neighbouring Pixels = 15x15 = 225
Memory Accesses to Write to Output = 1
Total Memory Accesses = 451
For a pure software implementation, even though many of these accesses can become fast because of caching, a large number of memory accesses will be a performance bottleneck. But designing on FPGA allows efficient data movement and access schemes to be easily built. One of the key and major advantages is the availability of substantial on-chip memory bandwidth (distributed and block memory) and the choice of a custom configuration of this bandwidth. This custom configuration choice essentially allows you to create an on-demand cache architecture tailored explicitly for the given algorithm. The next section elaborates on the design of the Window2D block.