michael@0: 
michael@0: /*
michael@0:  * Copyright 2012 Google Inc.
michael@0:  *
michael@0:  * Use of this source code is governed by a BSD-style license that can be
michael@0:  * found in the LICENSE file.
michael@0:  */
michael@0: 
michael@0: #ifndef SkRTree_DEFINED
michael@0: #define SkRTree_DEFINED
michael@0: 
michael@0: #include "SkRect.h"
michael@0: #include "SkTDArray.h"
michael@0: #include "SkChunkAlloc.h"
michael@0: #include "SkBBoxHierarchy.h"
michael@0: 
michael@0: /**
michael@0:  * An R-Tree implementation. In short, it is a balanced n-ary tree containing a hierarchy of
michael@0:  * bounding rectangles.
michael@0:  *
michael@0:  * Much like a B-Tree it maintains balance by enforcing minimum and maximum child counts, and
michael@0:  * splitting nodes when they become overfull. Unlike B-trees, however, we're using spatial data; so
michael@0:  * there isn't a canonical ordering to use when choosing insertion locations and splitting
michael@0:  * distributions. A variety of heuristics have been proposed for these problems; here, we're using
michael@0:  * something resembling an R*-tree, which attempts to minimize area and overlap during insertion,
michael@0:  * and aims to minimize a combination of margin, overlap, and area when splitting.
michael@0:  *
michael@0:  * One detail that is thus far unimplemented that may improve tree quality is attempting to remove
michael@0:  * and reinsert nodes when they become full, instead of immediately splitting (nodes that may have
michael@0:  * been placed well early on may hurt the tree later when more nodes have been added; removing
michael@0:  * and reinserting nodes generally helps reduce overlap and make a better tree). Deletion of nodes
michael@0:  * is also unimplemented.
michael@0:  *
michael@0:  * For more details see:
michael@0:  *
michael@0:  *  Beckmann, N.; Kriegel, H. P.; Schneider, R.; Seeger, B. (1990). "The R*-tree:
michael@0:  *      an efficient and robust access method for points and rectangles"
michael@0:  *
michael@0:  * It also supports bulk-loading from a batch of bounds and values; if you don't require the tree
michael@0:  * to be usable in its intermediate states while it is being constructed, this is significantly
michael@0:  * quicker than individual insertions and produces more consistent trees.
michael@0:  */
michael@0: class SkRTree : public SkBBoxHierarchy {
michael@0: public:
michael@0:     SK_DECLARE_INST_COUNT(SkRTree)
michael@0: 
michael@0:     /**
michael@0:      * Create a new R-Tree with specified min/max child counts.
michael@0:      * The child counts are valid iff:
michael@0:      * - (max + 1) / 2 >= min (splitting an overfull node must be enough to populate 2 nodes)
michael@0:      * - min < max
michael@0:      * - min > 0
michael@0:      * - max < SK_MaxU16
michael@0:      * If you have some prior information about the distribution of bounds you're expecting, you
michael@0:      * can provide an optional aspect ratio parameter. This allows the bulk-load algorithm to create
michael@0:      * better proportioned tiles of rectangles.
michael@0:      */
michael@0:     static SkRTree* Create(int minChildren, int maxChildren, SkScalar aspectRatio = 1,
michael@0:             bool orderWhenBulkLoading = true);
michael@0:     virtual ~SkRTree();
michael@0: 
michael@0:     /**
michael@0:      * Insert a node, consisting of bounds and a data value into the tree, if we don't immediately
michael@0:      * need to use the tree; we may allow the insert to be deferred (this can allow us to bulk-load
michael@0:      * a large batch of nodes at once, which tends to be faster and produce a better tree).
michael@0:      *  @param data The data value
michael@0:      *  @param bounds The corresponding bounding box
michael@0:      *  @param defer Can this insert be deferred? (this may be ignored)
michael@0:      */
michael@0:     virtual void insert(void* data, const SkIRect& bounds, bool defer = false) SK_OVERRIDE;
michael@0: 
michael@0:     /**
michael@0:      * If any inserts have been deferred, this will add them into the tree
michael@0:      */
michael@0:     virtual void flushDeferredInserts() SK_OVERRIDE;
michael@0: 
michael@0:     /**
michael@0:      * Given a query rectangle, populates the passed-in array with the elements it intersects
michael@0:      */
michael@0:     virtual void search(const SkIRect& query, SkTDArray<void*>* results) SK_OVERRIDE;
michael@0: 
michael@0:     virtual void clear() SK_OVERRIDE;
michael@0:     bool isEmpty() const { return 0 == fCount; }
michael@0: 
michael@0:     /**
michael@0:      * Gets the depth of the tree structure
michael@0:      */
michael@0:     virtual int getDepth() const SK_OVERRIDE {
michael@0:         return this->isEmpty() ? 0 : fRoot.fChild.subtree->fLevel + 1;
michael@0:     }
michael@0: 
michael@0:     /**
michael@0:      * This gets the insertion count (rather than the node count)
michael@0:      */
michael@0:     virtual int getCount() const SK_OVERRIDE { return fCount; }
michael@0: 
michael@0:     virtual void rewindInserts() SK_OVERRIDE;
michael@0: 
michael@0: private:
michael@0: 
michael@0:     struct Node;
michael@0: 
michael@0:     /**
michael@0:      * A branch of the tree, this may contain a pointer to another interior node, or a data value
michael@0:      */
michael@0:     struct Branch {
michael@0:         union {
michael@0:             Node* subtree;
michael@0:             void* data;
michael@0:         } fChild;
michael@0:         SkIRect fBounds;
michael@0:     };
michael@0: 
michael@0:     /**
michael@0:      * A node in the tree, has between fMinChildren and fMaxChildren (the root is a special case)
michael@0:      */
michael@0:     struct Node {
michael@0:         uint16_t fNumChildren;
michael@0:         uint16_t fLevel;
michael@0:         bool isLeaf() { return 0 == fLevel; }
michael@0:         // Since we want to be able to pick min/max child counts at runtime, we assume the creator
michael@0:         // has allocated sufficient space directly after us in memory, and index into that space
michael@0:         Branch* child(size_t index) {
michael@0:             return reinterpret_cast<Branch*>(this + 1) + index;
michael@0:         }
michael@0:     };
michael@0: 
michael@0:     typedef int32_t SkIRect::*SortSide;
michael@0: 
michael@0:     // Helper for sorting our children arrays by sides of their rects
michael@0:     struct RectLessThan {
michael@0:         RectLessThan(SkRTree::SortSide side) : fSide(side) { }
michael@0:         bool operator()(const SkRTree::Branch lhs, const SkRTree::Branch rhs) const {
michael@0:             return lhs.fBounds.*fSide < rhs.fBounds.*fSide;
michael@0:         }
michael@0:     private:
michael@0:         const SkRTree::SortSide fSide;
michael@0:     };
michael@0: 
michael@0:     struct RectLessX {
michael@0:         bool operator()(const SkRTree::Branch lhs, const SkRTree::Branch rhs) {
michael@0:             return ((lhs.fBounds.fRight - lhs.fBounds.fLeft) >> 1) <
michael@0:                    ((rhs.fBounds.fRight - lhs.fBounds.fLeft) >> 1);
michael@0:         }
michael@0:     };
michael@0: 
michael@0:     struct RectLessY {
michael@0:         bool operator()(const SkRTree::Branch lhs, const SkRTree::Branch rhs) {
michael@0:             return ((lhs.fBounds.fBottom - lhs.fBounds.fTop) >> 1) <
michael@0:                    ((rhs.fBounds.fBottom - lhs.fBounds.fTop) >> 1);
michael@0:         }
michael@0:     };
michael@0: 
michael@0:     SkRTree(int minChildren, int maxChildren, SkScalar aspectRatio, bool orderWhenBulkLoading);
michael@0: 
michael@0:     /**
michael@0:      * Recursively descend the tree to find an insertion position for 'branch', updates
michael@0:      * bounding boxes on the way up.
michael@0:      */
michael@0:     Branch* insert(Node* root, Branch* branch, uint16_t level = 0);
michael@0: 
michael@0:     int chooseSubtree(Node* root, Branch* branch);
michael@0:     SkIRect computeBounds(Node* n);
michael@0:     int distributeChildren(Branch* children);
michael@0:     void search(Node* root, const SkIRect query, SkTDArray<void*>* results) const;
michael@0: 
michael@0:     /**
michael@0:      * This performs a bottom-up bulk load using the STR (sort-tile-recursive) algorithm, this
michael@0:      * seems to generally produce better, more consistent trees at significantly lower cost than
michael@0:      * repeated insertions.
michael@0:      *
michael@0:      * This consumes the input array.
michael@0:      *
michael@0:      * TODO: Experiment with other bulk-load algorithms (in particular the Hilbert pack variant,
michael@0:      * which groups rects by position on the Hilbert curve, is probably worth a look). There also
michael@0:      * exist top-down bulk load variants (VAMSplit, TopDownGreedy, etc).
michael@0:      */
michael@0:     Branch bulkLoad(SkTDArray<Branch>* branches, int level = 0);
michael@0: 
michael@0:     void validate();
michael@0:     int validateSubtree(Node* root, SkIRect bounds, bool isRoot = false);
michael@0: 
michael@0:     const int fMinChildren;
michael@0:     const int fMaxChildren;
michael@0:     const size_t fNodeSize;
michael@0: 
michael@0:     // This is the count of data elements (rather than total nodes in the tree)
michael@0:     int fCount;
michael@0: 
michael@0:     Branch fRoot;
michael@0:     SkChunkAlloc fNodes;
michael@0:     SkTDArray<Branch> fDeferredInserts;
michael@0:     SkScalar fAspectRatio;
michael@0:     bool fSortWhenBulkLoading;
michael@0: 
michael@0:     Node* allocateNode(uint16_t level);
michael@0: 
michael@0:     typedef SkBBoxHierarchy INHERITED;
michael@0: };
michael@0: 
michael@0: #endif