Krita/Filter API Discussion Notes
Appearance
< Krita
Cyrille and Dmitry were discussing changes for filter and data manager API at the February 2010 sprint.
Here are some conclusions:
Filter API changes
- Filters should be applied to paint devices in place. It means that we will have something like KisFilter::process(KisPaintDeviceSP device, QRect applyRect, KisFilterConfiguration* kfc) instead of KisFilter::process(KisPaintDeviceSP src, KisPaintDeviceSP dst, QPoint srcOrigin, QPoint dstOrigin, QSize size, KisFilterConfiguration* kfc) (the last line is a bit simplified, just to show the meaning)
Why this change?
+ it allows us to use KisRectIterator inside filters that is much faster than any other iterators + we need one iterator only instead of two: one for destination and another one for source devices - it needs some special api from KisDataManager (see below)
- We don't need KisProcessingInformation anymore as we eliminated the use of two devices
- Filter implementation do not need to worry about selections. KisFilter::process do not have selections anymore. All the selection painting should be done at the higher level. As a solution we decided to use some wrapper above KisFilter::process, e.g. KisFilter::process(KisPaintDeviceSP device, QRect applyRect, KisFilterConfiguration* kfc, KisSelectionSP selection) This function will be part of KisFilter api and need not to be overriden by implementation.
- The wrapper above will rely on a special api from a KisDataManager. The latter will return the region of non-default pixels (actually this will be the list of non-default tiles). The filter will be applied to the rects from this list.
Still to be discussed: Usage of default pixel in KisSelection. Atm it is quite weird, so it can lower the use of regions to zero.
KisDataManager API changes
- we need an opposite method to getMemento(). It will be called commit(). This method will fix the changes made during transaction onto the device.
- Why this change?
- KisTileDataPooler should know when to start preallocating tiles. Atm it works quite inefficient because new commits come right before getting new memento (and creating new transaction)
- How this change?
- KisTransaction will not inherit QUndoCommand anymore. KisTransactionData will do. KisTransaction will be only a wrapper for KisTransactionData and be stored in the stack (not heap) and will work like QMutexLocker.
- QRegion KisDataManager::region() returns a region of non-default tiles. Works like extent() but returns QRegion instead of QRect.
- void KisDataManager::bitBlt(KisDataManagerSP src, QPoint srcOrigin, QPoint dstOrigin, QSize size) This is a fast version of bitBlt that will explicitly share the tiles between data managers. Should be used by the KisPainter for cases when both color spaces coincide
- KisDataManagerSP KisDataManager::fastClone(QRect regionOfInterest) This method will create a new data manager that resebles a source data manager on a rect AT LEAST as big as regionOfInteres. What does it mean? It means that the clone can be a bit larger than we asked in regionOfInterest. Why it is needed? For creating temporary devices during filtering or some other stuff.
The plan
Status | Description | Contact | |
---|---|---|---|
DONE | KisDataManager::region() | Dmitry | |
TO DO | A wrapper for selections based on region() functionality | Cyrille | <{{{3}}}> |
DONE | Filters in-place application | Cyrille | |
IN PROGRESS | Remove KisProcessingInformation | Cyrille | |
DONE | Commit'able transactions | Dmitry | |
DONE | KisDataManager::bitBlt | Dmitry | |
DONE | KisDataManager::fastClone | Dmitry | |
TO DO | Connect KisPaintDevice::fastBitBlt to KisPainter | Cyrille | <{{{3}}}> |