weights[] defines the value and the size of the filter.

I think for computation such as convolution we can change the "for" loop into matrix multiplication. By using parallel computing on the sparse matrix, it seems that the performance sometimes might be better?

The number of total math ops: width * height * 9, or 9 math ops per pixel. Arithmetic intensity is very high, but also this function is very costly, which leads us to the next slide.