Speed up integral images

[minnerbe]

While investigating a performance issue (which turned out to be unrelated to this library) in some downstream code, I found that some minimal changes could provide significant speedup: Integral images are accumulated row-wise and column-wise, where column-wise accumulation took significantly longer than the row-wise one.

This PR changes the memory access pattern of the column-wise accumulation to row-wise access, thereby significantly reducing the run time; see the changes in IntIntegralImage as an example. These are the min/max run times for 10 runs of computing the integral image for a 20000x20000 image before and after the changes:

	Before	After
Row-wise	214ms - 267ms	222ms - 269ms
Column-wise	627ms - 716ms	131ms - 141ms
Total	841ms - 983ms	353ms - 410ms

All changes either change the accumulation in exactly the same manner, or are IDE warnings which are automatically fixed. For the former category of changes, I manually checked correctness by using images with elements i -> 1 and i -> i + 1.

Let me know what you think, @axtimwalde

[axtimwalde]

Beautiful! Have you benchmarked small images? 1024^2 or something?

[minnerbe]

Good point @axtimwalde ! Our use case was large images (slices of compute blocks of large 3D stacks), so I focused on that.

I rerun a small benchmark for n x n images, where n varied between 10000-100 and I took the average run time over 100-10000 runs. As expected, the difference is not as stark for medium-sized images but still noticeable, and at least there's no degradation in run time for small-sized images.

Size	Before	After
10000	240ms	97ms
1000	1.23ms	0.97ms
100	0.012ms	0.011ms

I haven't looked at the assembly instructions generated by the JIT compiler, but I suspect that all of this is 'just' more efficient memory access, since Java doesn't reliably compile SIMD instructions. Should this change in the future, the current way of doing column-wise summation is very amenable to this kind of automatic further optimizations.

[axtimwalde]

Thanks a lot!