On 11 Jun 2009, at 12:40, Vincent Hennebert wrote:
Hi Vincent
<snip />
I spent some time looking at the current code and it seems to me that
a hack could be implemented without too many difficulties. It
basically
consists in 2 steps:
1. in the Knuth breaking algorithm, when creating a new active node,
look whether the IPD for the following page is the same or not. If
not, deactivate the node. Once we run out of active nodes, select
the
best of those deactivated nodes and treat it as if it were the
regular final node. Add areas for content up to that node.
2. re-create Knuth elements, starting from the index corresponding to
that node. Re-launch the breaking algorithm, starting from there.
Then back to step 1, until the end of the document is reached.
Step 1 should be doable without turning everything upside down.
Step 2 implies to change the signature of the
LayoutManager.getNextKnuthElements method, adding a parameter that
corresponds to the index from where to start the generation of Knuth
elements. We could largely ignore it, except in BlockLayoutManager
where
we would re-launch the line-breaking algorithm, taking the new IPD
into
account.
Obviously this is a limited approach. There is likely to be
a (potentially huge) waste of CPU cycles due to the re-creation of
Knuth
elements. There may be side effects that I’ve missed so far. But I
think
it’s worth giving it a try.
The only thing I'm slightly concerned about, is the case where there
would be multiple IPD changes for subsequent pages. I'm assuming that
is the waste of CPU cycles you're referring to (?)
If I interpret correctly, we would (supposing a page-sequence without
forced breaks and/or span changes):
a) generate the complete block list (effectively computing the line-
breaks for the whole page-sequence)
b) when computing the page-breaks, and encountering a new page with
different available IPD, re-generate the remaining elements and
recompute the line-breaks after that position
b) would occur as many times as we have IPD changes.
I'm wondering whether it would not be equally feasible to have the
FlowLM check the total height of the block-boxes up to a point
(estimated). Do this after every call to
childLM.getNextKnuthElements(). Then, if the total height exceeds the
BPD for the current page, call back to check if the next page has the
same IPD (or the same amount of columns). If so, then we happily
continue, just as we do now. If not, then we hand the list off to the
PageBreaker, run it through a PageBreakingAlgorithm to compute the
page-breaks up to that point, add the areas, and resume later, passing
the LineBreakingAlgorithm an updated LayoutContext corresponding to
the next page.
If this functionality could somehow be factored in to BlockStackingLM,
all the better, since we're definitely going to need it to avoid
BlockLMs and TableLMs from accumulating the element-lists for all
their content. They too will need to be able to stop when their
accumulated content-height exceeds the threshold (available BPD + a
percentage?)
Step 2) would still be necessary, since we need to know at what point
to resume the line-breaking later on.
The benefit being that we would catch the IPD-change long before the
end of the page-sequence is reached by the line-breaking loop. The
amount of elements that need to be re-generated would always remain as
small as possible.
If I judge correctly, and it is feasible, then this may present an
opportunity for a pure FO hack to reduce the memory consumption for
arbitrarily sized page-sequences: use alternating page-masters with a
different IPD. Only make the difference practically invisible to naked
eye. FOP would still detect it, and flush the list up to that point.
Thoughts, you asked? ;-)
Andreas Delmelle