Add post on custom array indices

[timholy]

This isn't ready yet, but I'm opening as a PR now to make it easier for anyone to comment or ask questions at specific points in the text.

[KristofferC]

Perhaps use v0.6 stacktraces in the final version of the post?

[timholy]

Good idea. Esp. since 0.5.0 (rather than release-0.5) has some significant bugs when it comes to indices support.

[ViralBShah]

Bump. In case this is ready to merge.

[rsrock]

Question for you, Tim (related to what follows). Let's say I wanted to extract a region from an image into a buffer. I'm going to do a lot of expensive calculations on that region, so it makes sense to copy that data into a buffer instead of using a view into the original image. Let's also say that the buffer will be a fixed size, but the target location will change once I'm done processing one object. For example, I grab the left eye, calculate, then grab the right eye and calculate.

Can I mutate the ranges used to define the offset array (keeping the array size fixed) so I don't have to fill up the garbage collector with a whole bunch of buffers?

[timholy]

You can definitely re-use the buffer. You have to create a new OffsetArray wrapper. Something like this:

buf = Matrix{T}(30, 30)
chunk = OffsetArray(buf, 201:230, 201:230)
R = CartesianRange(indices(chunk))
copy!(chunk, R, img, R)
# do something useful
chunk = OffsetArray(buf, 201:230, 401:430)
R = CartesianRange(indices(chunk))
copy!(chunk, R, img, R)
# do something useful

Does that answer your question?

You may know this already, but for reasons of safety we don't allow turning off bounds-checking on OffsetArrays; you'll want the @unsafe macro if you need high performance. By Julia 1.0 I think we should turn on the ability to remove bounds-checks. This blog post will hopefully get more people testing non-1 arrays so more of our code is ready.

[rsrock]

Perfect, thanks for the tip.

[timholy]

OK, I reworked the introductory example to exploit the work that @Evizero and I have been doing in ImageTransformations. I think that makes a more compelling first example.

Is there a good way to preview the result? The image links are specific for the paths here, and there's math formatting. If not, let's just merge it and I can fix the issues before I advertise the new post.

[rfourquet]

given -> give (?)

[timholy]

If anyone besides me ends up merging this, my intention is to squash into a single commit.

[ViralBShah]

Best if you merge it whenever you are ready.

[timholy]

That makes sense; that way I can be on the lookout for formatting errors and fix them quickly.

I might wait until I tag JuliaIO/ImageMagick.jl#85 (hopefully in a day or two).

[KristofferC]

The way I chose to put the paths for the images, they are at least showed in the markdown preview: https://github.com/JuliaLang/julialang.github.com/blob/kc/repl0.6/blog/_posts/2017-02-22-repl-0.6-highlights.md

[ararslan]

Woot! Great to see this merged. Excellent work here, Tim!

[timholy]

Well, the images are all borked.

[timholy]

OK, fixed now (see #556). I'll wait a few hours before I advertise this on Discourse, in case anyone sees problems with the "published" product that didn't show up during review.