Use BBH reserve hook to preallocate space for tiles.

Before getting too far into changing how SkTileGrid stores its tiles, I figured I'd
better see how much I can tweak out the existing format.  Cleverly, that way
any improvements I make by changing the format will look that much less
impressive.

This CL looks like it will be a 5-15% win in time spent recording, with no effect
on playback.

This CL also shrinks the tiles to fit exactly when we're done inserting,
using newly added SkTDArray::shrinkToFit().  It's quite cheap to run (maybe
taking back 1-2% from those 5-15% wins), and means we'll lug around about 15%
fewer bytes in the tile grids.  Note though this strategy temporarily uses up to
30% more memory while building the tile grid.  For our largest SKPs, that's
maybe 75-100K extra.

BUG=skia:

Committed: https://skia.googlesource.com/skia/+/52455cbc02d7f480d988ae7cdacc11ad69078c2c

CQ_EXTRA_TRYBOTS=tryserver.skia:Canary-Chrome-Ubuntu13.10-Ninja-x86_64-ToT-Trybot

Review URL: https://codereview.chromium.org/639823005

2 files changed, 30 insertions, 0 deletions

diff --git a/src/core/SkTileGrid.cpp b/src/core/SkTileGrid.cpp
index 8b96b73fa1..30ca4b9179 100644
--- a/src/core/SkTileGrid.cpp
+++ b/src/core/SkTileGrid.cpp
@@ -23,6 +23,33 @@ SkTileGrid::~SkTileGrid() {
     SkDELETE_ARRAY(fTiles);
 }
 
+void SkTileGrid::reserve(unsigned opCount) {
+    if (fXTiles * fYTiles == 0) {
+        return;  // A tileless tile grid is nonsensical, but happens in at least cc_unittests.
+    }
+
+    // If we assume every op we're about to try to insert() falls within our grid bounds,
+    // then every op has to hit at least one tile.  In fact, a quick scan over our small
+    // SKP set shows that in the average SKP, each op hits two 256x256 tiles.
+
+    // If we take those observations and further assume the ops are distributed evenly
+    // across the picture, we get this guess for number of ops per tile:
+    const int opsPerTileGuess = (2 * opCount) / (fXTiles * fYTiles);
+
+    for (SkTDArray<unsigned>* tile = fTiles; tile != fTiles + (fXTiles * fYTiles); tile++) {
+        tile->setReserve(opsPerTileGuess);
+    }
+
+    // In practice, this heuristic means we'll temporarily allocate about 30% more bytes
+    // than if we made no setReserve() calls, but time spent in insert() drops by about 50%.
+}
+
+void SkTileGrid::flushDeferredInserts() {
+    for (SkTDArray<unsigned>* tile = fTiles; tile != fTiles + (fXTiles * fYTiles); tile++) {
+        tile->shrinkToFit();
+    }
+}
+
 // Adjustments to user-provided bounds common to both insert() and search().
 // Call this after making insert- or search- specific adjustments.
 void SkTileGrid::commonAdjust(SkRect* rect) const {
diff --git a/src/core/SkTileGrid.h b/src/core/SkTileGrid.h
index e8a0d96a65..fd7584fd9c 100644
--- a/src/core/SkTileGrid.h
+++ b/src/core/SkTileGrid.h
@@ -39,6 +39,9 @@ public:
     // For testing.
     int tileCount(int x, int y) { return fTiles[y * fXTiles + x].count(); }
 
+    virtual void reserve(unsigned opCount) SK_OVERRIDE;
+    virtual void flushDeferredInserts() SK_OVERRIDE;
+
 private:
     void commonAdjust(SkRect*) const;
     void userToGrid(const SkRect&, SkIRect* grid) const;