pdf page flipper
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
.settings
.worksheet/src
src/main
.cache
.gitignore
README.md
build.sbt

README.md

flip

This is dual-purpose experiment, using both senses of the word functional:

  • Make a better pdf reader.

    One reason thing I don't like about ebook readers is that you can't flip through the pages properly. You can search for things. You can use internal hyperlinks. But, with a real book, you often flip quickly through the pages, usually recognizing what you were looking for slightly after it's past, and then backtracking more slowly.

    Since physical pages can flip by at hundreds of pages per second, it isn't feasible to display successive pdf pages quickly enough to duplicate the physical experience. However, you can precalculate a rolling average of pages and, when the scroll widget is moved quickly, display them blurred together - effectively duplicating your eye's latent image retention.

    It works. The current implementation hasn't been much optimized, so it's both slow and leaves something to be desired in terms of ergonomics. For example, the averaging window is always 8 pages, and the view is always centered in it.

  • Use functional programming concepts to solve a real problem in concurrent application development.

    Specifically, the computation of blurred pages takes some time, and since it's a rolling average, it needs to proceed sequentially. At the same time, I would like the application to come up immediately and to behave as well as it can, i.e. if the blurred image hasn't yet been calculated for this region, display an unprocessed image, while a progress counter ticks up.

    A boring,procedural way to handle this is to start a thread that processes the pages in a big for loop and have some sort of progress variable, e.g. an AtomicInteger so we know what to show. But that's, well, boring, and also potentially abusable. Instead:

    1. We want to express the algorithm as a composition of list operations, i.e.

          val pagesView = pdfdoc.view.
            map(imageToPixels).
            map(unpackIntsToLongs).
            scanLeftLazy(new Rolling(navg,npix))((r,ll)=>r.push(ll)).
            map(r => r.avg).
            map(packLongsToInts).
            map(pixelsToBufferedImage).
            zipWithIndex.
            map( p => new Page(p._1,p._2))
      

      Without knowing exactly what this does, you can kind of get the idea immediately.

    2. We don't want the steps (lines in the above code) to run completely before the next one runs. That would be bad since all pages would complete at the same time, and because we'd have to stage each intermediate result in a possibly enormous list. The view method converts the Iterable to an IterableView, which has the property of calculating elements on demand. Furthermore, chained list methods compose their respective functions, e.g.

         iterableThing.view.map(x=>f(x)).map(x=>g(x))
      

      is implemented as

         iterableThing.view.map(x=>g(f(x)))
      

      So each page is processed completely, and separately from the others.

    3. We want to get intermediate results as Future objects, which represent a computation proceeding in another thread and can tell us whether they've finished or not. That's accomplished in the code with the line:

      val pages : Seq[Future[Page]] = viewToSequentialFutures(pagesView)

      (where we wrote the viewToSequentialFutures function) instead of

      val pages : Seq[Page] = pagesView.force

      (where force is already a method of the various View classes). The latter would block until all the pages are finished. The former returns immediately.

    Getting this to work properly took surprisingly little code, but that code took me a long time to write. I think it's a nice paradigm.