I want to apply an operation on a sliding window over an image.
Yet it is Spatially Variant hence cannot be executed ad a convolution.
Those are needed in order to Vectorize the operation and keep the highest performance.
Unless there are better more optimized way to engage vectorization when dealing with Spatially Variant filters.
it is possible, but I think will be too expensive, as requires several steps:
1) transpose an image - so columns become rows
2) create additional IppiSize structure with your sliding window width and height
3) define the new image (sliding window) step that is == your window width*sizeof(your data type)
4) one limitation of such approach - your image should be dense (step == width*sizeof(data type) or width*height of window < image width
for col2im you can use the same approach - transpose, etc. but I think that simple C-loop will be more efficient and doesn't have any limitations.