The Window2D block is essentially built from two basic blocks: the first is called a “line buffer”, and the second is called a “Window”.
The line buffer is used to buffer multiple lines of a full image, and specifically, here it is designed to buffer FILTER_V_SIZE - 1 image lines, where FILTER_V_SIZE is the height of the convolution filter. The total number of pixels held by the line buffer is (FILTER_V_SIZE-1) * MAX_IMAGE_WIDTH.
The “Window” block holds FILTER_V_SIZE * FILTER_H_SIZE pixels. The 2-D convolution filtering operation consists of centering the filtering mask (filter coefficients) on the index of output pixel and calculating the SOP as described in the previous lab. The following figure shows how these centering and SOP operations are carried.
The figure above shows SOP carried out for a full image being processed. If you look carefully when output pixels are produced line by line, it is not required to have all the image pixels in memory. Only the lines where the filtering mask overlaps are required which is essentially FILTER_V_SIZE lines, which can even be reduced to FILTER_V_SIZE-1. Essentially, that is the amount of data that needs to be on-chip or housed by a data mover at any given time.
The figure above illustrates the operation and requirements for a line and Window buffer. The image size is assumed 8x8, and the filter size is 3x3. For this example, you are generating the filtered output of pixel number 10. In this case, you need a 3x3 block of input pixels centered around pixel 10, as shown in step A.
Step B in the figure highlights what is required for producing pixel number 11. Another 3x3 block, but it has a significant overlap with the previous input block. Essentially a column moves out from the left, and a column moves in from the right. One important thing to notice in steps A, B, and C, is that from the input side; it only needs one new pixel to produce one output pixel (ignoring the initial latency of filling the line buffer with multiple pixels, which is one-time only).
The line buffer holds FILTER_V_SIZE-1 lines. In general, it requires FILTER_V_SIZE lines, but a line is reduced by using the line buffer in a circular fashion and by exploiting the fact that pixels at the start of the first line buffer can be used to write new incoming pixels since they are no longer needed. The window buffer is implemented as FILTER_V_SIZE * FILTER_H_SIZE storage fully partitioned, giving parallel access to all elements inside the window. The data moves as a column vector of size FILTER_V_SIZE from line buffer to window buffer, and then this whole widow is passed through a stream to the Filter2D function for processing.
The overall scheme (data mover) is built to maximize the data reuse providing maximum parallel data to the processing element. For a deeper understanding of the modeling style and minute details of the data mover, examine the Window2D function details in the source code. The function can be found in the src/filter2d_hw.cpp source file in the convolutioanl tutorial directory.