Batched implementation of drawLattice() for GPU

Bechmarks (Nexus 6P):

Src=100x100, Dst=250x250, NumRects=9
Android              77.7us
Skia (without patch) 57.2us
Skia (with patch)    30.9us

Src=100x100, Dst=500x500, NumRects=9
Android              77.0us
Skia (without patch) 56.9us
Skia (with patch)    31.8us

Src=100x100, Dst=1000x1000, NumRects=9
Android              180us
Skia (without patch) 96.8us
Skia (with patch)    70.5us

Src=100x100, Dst=250x250, NumRects=15
Android              208us
Skia (without patch) 155us
Skia (with patch)    38.2us

Src=100x100, Dst=500x500, NumRects=15
Android              207us
Skia (without patch) 152us
Skia (with patch)    38.4us

Src=100x100, Dst=1000x1000, NumRects=15
Android              233us
Skia (without patch) 156us
Skia (with patch)    99.9us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2255963002

Committed: https://skia.googlesource.com/skia/+/93242c4ae50dfcc0d922cdb3ba80bbc7b4bbe93d
Review-Url: https://codereview.chromium.org/2255963002

10 files changed, 183 insertions, 57 deletions

diff --git a/bench/DrawLatticeBench.cpp b/bench/DrawLatticeBench.cpp
index 4c03ca1dfa..9c4d31a099 100644
--- a/bench/DrawLatticeBench.cpp
+++ b/bench/DrawLatticeBench.cpp
@@ -29,6 +29,10 @@ public:
         return fName.c_str();
     }
 
+    SkIPoint onGetSize() override {
+        return SkIPoint::Make(1000, 1000);
+    }
+
     bool isSuitableFor(Backend backend) override {
         return kRaster_Backend == backend || kGPU_Backend == backend;
     }
@@ -54,6 +58,17 @@ private:
     typedef Benchmark INHERITED;
 };
 
-static int gDivs[2] = { 250, 750, };
-DEF_BENCH(return new DrawLatticeBench(gDivs, 2, gDivs, 2, SkISize::Make(1000, 1000),
-                                      SkRect::MakeWH(4000.0f, 4000.0f), "StandardNine");)
+static int gDivs9[2] = { 25, 75, };
+DEF_BENCH(return new DrawLatticeBench(gDivs9, 2, gDivs9, 2, SkISize::Make(100, 100),
+                                      SkRect::MakeWH(250.0f, 250.0f), "Src100_Dst250_Rects9");)
+DEF_BENCH(return new DrawLatticeBench(gDivs9, 2, gDivs9, 2, SkISize::Make(100, 100),
+                                      SkRect::MakeWH(500.0f, 500.0f), "Src100_Dst500_Rects9");)
+DEF_BENCH(return new DrawLatticeBench(gDivs9, 2, gDivs9, 2, SkISize::Make(100, 100),
+                                      SkRect::MakeWH(1000.0f, 1000.0f), "Src100_Dst1000_Rects9");)
+static int gDivs15[4] = { 15, 45, 55, 85, };
+DEF_BENCH(return new DrawLatticeBench(gDivs15, 4, gDivs15, 4, SkISize::Make(100, 100),
+                                      SkRect::MakeWH(250.0f, 250.0f), "Src100_Dst250_Rects15");)
+DEF_BENCH(return new DrawLatticeBench(gDivs15, 4, gDivs15, 4, SkISize::Make(100, 100),
+                                      SkRect::MakeWH(500.0f, 500.0f), "Src100_Dst500_Rects15");)
+DEF_BENCH(return new DrawLatticeBench(gDivs15, 4, gDivs15, 4, SkISize::Make(100, 100),
+                                      SkRect::MakeWH(1000.0f, 1000.0f), "Src100_Dst1000_Rects15");)
diff --git a/include/gpu/GrDrawContext.h b/include/gpu/GrDrawContext.h
index 6e98404a44..77a3443f9e 100644
--- a/include/gpu/GrDrawContext.h
+++ b/include/gpu/GrDrawContext.h
@@ -35,6 +35,7 @@ struct GrUserStencilSettings;
 class SkDrawFilter;
 struct SkIPoint;
 struct SkIRect;
+class SkLatticeIter;
 class SkMatrix;
 class SkPaint;
 class SkPath;
@@ -231,26 +232,15 @@ public:
                   const GrStyle& style);
 
     /**
-     *  Draw the image stretched differentially to fit into dst.
-     *  center is a rect within the image, and logically divides the image
-     *  into 9 sections (3x3). For example, if the middle pixel of a [5x5]
-     *  image is the "center", then the center-rect should be [2, 2, 3, 3].
-     *
-     *  If the dst is >= the image size, then...
-     *  - The 4 corners are not stretched at all.
-     *  - The sides are stretched in only one axis.
-     *  - The center is stretched in both axes.
-     * Else, for each axis where dst < image,
-     *  - The corners shrink proportionally
-     *  - The sides (along the shrink axis) and center are not drawn
+     *  Draw the image as a set of rects, specified by |iter|.
      */
-    void drawImageNine(const GrClip&,
-                       const GrPaint& paint,
-                       const SkMatrix& viewMatrix,
-                       int imageWidth,
-                       int imageHeight,
-                       const SkIRect& center,
-                       const SkRect& dst);
+    void drawImageLattice(const GrClip&,
+                          const GrPaint& paint,
+                          const SkMatrix& viewMatrix,
+                          int imageWidth,
+                          int imageHeight,
+                          std::unique_ptr<SkLatticeIter> iter,
+                          const SkRect& dst);
 
     /**
      * After this returns any pending surface IO will be issued to the backend 3D API and
diff --git a/include/private/SkTArray.h b/include/private/SkTArray.h
index 1fe2c3857c..1c06bf331a 100644
--- a/include/private/SkTArray.h
+++ b/include/private/SkTArray.h
@@ -238,6 +238,19 @@ public:
     }
 
     /**
+     * Version of above that uses the move constructor to set n items.
+     */
+    T* move_back_n(int n, T* t) {
+        SkASSERT(n >= 0);
+        this->checkRealloc(n);
+        for (int i = 0; i < n; ++i) {
+            new (fItemArray + fCount + i) T(std::move(t[i]));
+        }
+        fCount += n;
+        return fItemArray + fCount - n;
+    }
+
+    /**
      * Removes the last element. Not safe to call when count() == 0.
      */
     void pop_back() {
diff --git a/src/core/SkLatticeIter.cpp b/src/core/SkLatticeIter.cpp
index 24ab3f1cbc..ba6ac97a5b 100644
--- a/src/core/SkLatticeIter.cpp
+++ b/src/core/SkLatticeIter.cpp
@@ -159,6 +159,7 @@ SkLatticeIter::SkLatticeIter(int srcWidth, int srcHeight, const SkCanvas::Lattic
 
     fCurrX = fCurrY = 0;
     fDone = false;
+    fNumRects = (xCount + 1) * (yCount + 1);
 }
 
 bool SkLatticeIter::Valid(int width, int height, const SkIRect& center) {
@@ -205,6 +206,7 @@ SkLatticeIter::SkLatticeIter(int w, int h, const SkIRect& c, const SkRect& dst)
 
     fCurrX = fCurrY = 0;
     fDone = false;
+    fNumRects = 9;
 }
 
 bool SkLatticeIter::next(SkRect* src, SkRect* dst) {
@@ -228,3 +230,18 @@ bool SkLatticeIter::next(SkRect* src, SkRect* dst) {
     }
     return true;
 }
+
+void SkLatticeIter::mapDstScaleTranslate(const SkMatrix& matrix) {
+    SkASSERT(matrix.isScaleTranslate());
+    SkScalar tx = matrix.getTranslateX();
+    SkScalar sx = matrix.getScaleX();
+    for (int i = 0; i < fDstX.count(); i++) {
+        fDstX[i] = fDstX[i] * sx + tx;
+    }
+
+    SkScalar ty = matrix.getTranslateY();
+    SkScalar sy = matrix.getScaleY();
+    for (int i = 0; i < fDstY.count(); i++) {
+        fDstY[i] = fDstY[i] * sy + ty;
+    }
+}
diff --git a/src/core/SkLatticeIter.h b/src/core/SkLatticeIter.h
index e2d767f1c8..192b6c5b54 100644
--- a/src/core/SkLatticeIter.h
+++ b/src/core/SkLatticeIter.h
@@ -35,15 +35,28 @@ public:
      */
     bool next(SkRect* src, SkRect* dst);
 
+    /**
+     *  Apply a matrix to the dst points.
+     */
+    void mapDstScaleTranslate(const SkMatrix& matrix);
+
+    /**
+     *  Returns the total number of rects that will be drawn.
+     */
+    int numRects() const {
+        return fNumRects;
+    }
+
 private:
     SkTArray<SkScalar> fSrcX;
     SkTArray<SkScalar> fSrcY;
     SkTArray<SkScalar> fDstX;
     SkTArray<SkScalar> fDstY;
 
-    int fCurrX;
-    int fCurrY;
+    int  fCurrX;
+    int  fCurrY;
     bool fDone;
+    int  fNumRects;
 };
 
 #endif
diff --git a/src/gpu/GrDrawContext.cpp b/src/gpu/GrDrawContext.cpp
index 202be23590..db14498962 100644
--- a/src/gpu/GrDrawContext.cpp
+++ b/src/gpu/GrDrawContext.cpp
@@ -35,6 +35,8 @@
 
 #include "../private/GrAuditTrail.h"
 
+#include "SkLatticeIter.h"
+
 #define ASSERT_OWNED_RESOURCE(R) SkASSERT(!(R) || (R)->getContext() == fDrawingManager->getContext())
 #define ASSERT_SINGLE_OWNER \
     SkDEBUGCODE(GrSingleOwner::AutoEnforce debug_SingleOwner(fSingleOwner);)
@@ -999,23 +1001,23 @@ void GrDrawContext::drawOval(const GrClip& clip,
     this->internalDrawPath(clip, paint, viewMatrix, path, style);
 }
 
-void GrDrawContext::drawImageNine(const GrClip& clip,
-                                  const GrPaint& paint,
-                                  const SkMatrix& viewMatrix,
-                                  int imageWidth,
-                                  int imageHeight,
-                                  const SkIRect& center,
-                                  const SkRect& dst) {
+void GrDrawContext::drawImageLattice(const GrClip& clip,
+                                     const GrPaint& paint,
+                                     const SkMatrix& viewMatrix,
+                                     int imageWidth,
+                                     int imageHeight,
+                                     std::unique_ptr<SkLatticeIter> iter,
+                                     const SkRect& dst) {
     ASSERT_SINGLE_OWNER
     RETURN_IF_ABANDONED
     SkDEBUGCODE(this->validate();)
-    GR_AUDIT_TRAIL_AUTO_FRAME(fAuditTrail, "GrDrawContext::drawImageNine");
+    GR_AUDIT_TRAIL_AUTO_FRAME(fAuditTrail, "GrDrawContext::drawImageLattice");
 
     AutoCheckFlush acf(fDrawingManager);
 
     SkAutoTUnref<GrDrawBatch> batch(GrNinePatch::CreateNonAA(paint.getColor(), viewMatrix,
                                                              imageWidth, imageHeight,
-                                                             center, dst));
+                                                             std::move(iter), dst));
 
     GrPipelineBuilder pipelineBuilder(paint, this->mustUseHWAA(paint));
     this->getDrawTarget()->drawBatch(pipelineBuilder, this, clip, batch);
diff --git a/src/gpu/SkGpuDevice.cpp b/src/gpu/SkGpuDevice.cpp
index 376d5e1b57..040833b731 100644
--- a/src/gpu/SkGpuDevice.cpp
+++ b/src/gpu/SkGpuDevice.cpp
@@ -1409,8 +1409,10 @@ void SkGpuDevice::drawProducerNine(const SkDraw& draw, GrTextureProducer* produc
         return;
     }
 
-    fDrawContext->drawImageNine(fClip, grPaint, *draw.fMatrix, producer->width(),
-                                producer->height(), center, dst);
+    std::unique_ptr<SkLatticeIter> iter(
+            new SkLatticeIter(producer->width(), producer->height(), center, dst));
+    fDrawContext->drawImageLattice(fClip, grPaint, *draw.fMatrix, producer->width(),
+                                   producer->height(), std::move(iter), dst);
 }
 
 void SkGpuDevice::drawImageNine(const SkDraw& draw, const SkImage* image,
@@ -1440,6 +1442,61 @@ void SkGpuDevice::drawBitmapNine(const SkDraw& draw, const SkBitmap& bitmap, con
     this->drawProducerNine(draw, &maker, center, dst, paint);
 }
 
+void SkGpuDevice::drawProducerLattice(const SkDraw& draw, GrTextureProducer* producer,
+                                      const SkCanvas::Lattice& lattice, const SkRect& dst,
+                                      const SkPaint& paint) {
+    GR_CREATE_TRACE_MARKER_CONTEXT("SkGpuDevice", "drawProducerLattice", fContext);
+
+    CHECK_SHOULD_DRAW(draw);
+
+    static const GrTextureParams::FilterMode kMode = GrTextureParams::kNone_FilterMode;
+    sk_sp<GrFragmentProcessor> fp(
+        producer->createFragmentProcessor(SkMatrix::I(),
+                                          SkRect::MakeIWH(producer->width(), producer->height()),
+                                          GrTextureProducer::kNo_FilterConstraint, true,
+                                          &kMode, fDrawContext->getColorSpace(),
+                                          fDrawContext->sourceGammaTreatment()));
+    GrPaint grPaint;
+    if (!SkPaintToGrPaintWithTexture(this->context(), fDrawContext.get(), paint, *draw.fMatrix,
+                                     std::move(fp), producer->isAlphaOnly(), &grPaint)) {
+        return;
+    }
+
+    std::unique_ptr<SkLatticeIter> iter(
+            new SkLatticeIter(producer->width(), producer->height(), lattice, dst));
+    fDrawContext->drawImageLattice(fClip, grPaint, *draw.fMatrix, producer->width(),
+                                   producer->height(), std::move(iter), dst);
+}
+
+void SkGpuDevice::drawImageLattice(const SkDraw& draw, const SkImage* image,
+                                   const SkCanvas::Lattice& lattice, const SkRect& dst,
+                                   const SkPaint& paint) {
+    ASSERT_SINGLE_OWNER
+    uint32_t pinnedUniqueID;
+    if (sk_sp<GrTexture> tex = as_IB(image)->refPinnedTexture(&pinnedUniqueID)) {
+        CHECK_SHOULD_DRAW(draw);
+        GrTextureAdjuster adjuster(tex.get(), image->alphaType(), image->bounds(), pinnedUniqueID,
+                                   as_IB(image)->onImageInfo().colorSpace());
+        this->drawProducerLattice(draw, &adjuster, lattice, dst, paint);
+    } else {
+        SkBitmap bm;
+        if (SkImageCacherator* cacher = as_IB(image)->peekCacherator()) {
+            GrImageTextureMaker maker(fContext, cacher, image, SkImage::kAllow_CachingHint);
+            this->drawProducerLattice(draw, &maker, lattice, dst, paint);
+        } else if (as_IB(image)->getROPixels(&bm)) {
+            this->drawBitmapLattice(draw, bm, lattice, dst, paint);
+        }
+    }
+}
+
+void SkGpuDevice::drawBitmapLattice(const SkDraw& draw, const SkBitmap& bitmap,
+                                    const SkCanvas::Lattice& lattice, const SkRect& dst,
+                                    const SkPaint& paint) {
+    ASSERT_SINGLE_OWNER
+    GrBitmapTextureMaker maker(fContext, bitmap);
+    this->drawProducerLattice(draw, &maker, lattice, dst, paint);
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 
 // must be in SkCanvas::VertexMode order
diff --git a/src/gpu/SkGpuDevice.h b/src/gpu/SkGpuDevice.h
index 2185f162d7..d5fe04febf 100644
--- a/src/gpu/SkGpuDevice.h
+++ b/src/gpu/SkGpuDevice.h
@@ -122,6 +122,11 @@ public:
     void drawBitmapNine(const SkDraw& draw, const SkBitmap& bitmap, const SkIRect& center,
                         const SkRect& dst, const SkPaint& paint) override;
 
+    void drawImageLattice(const SkDraw&, const SkImage*, const SkCanvas::Lattice&,
+                          const SkRect& dst, const SkPaint&) override;
+    void drawBitmapLattice(const SkDraw&, const SkBitmap&, const SkCanvas::Lattice&,
+                           const SkRect& dst, const SkPaint&) override;
+
     void drawSpecial(const SkDraw&, SkSpecialImage*,
                      int left, int top, const SkPaint& paint) override;
     sk_sp<SkSpecialImage> makeSpecial(const SkBitmap&) override;
@@ -243,6 +248,9 @@ private:
     void drawProducerNine(const SkDraw&, GrTextureProducer*, const SkIRect& center,
                           const SkRect& dst, const SkPaint&);
 
+    void drawProducerLattice(const SkDraw&, GrTextureProducer*, const SkCanvas::Lattice& lattice,
+                             const SkRect& dst, const SkPaint&);
+
     bool drawDashLine(const SkPoint pts[2], const SkPaint& paint);
     void drawStrokedLine(const SkPoint pts[2], const SkDraw&, const SkPaint&);
 
diff --git a/src/gpu/batches/GrNinePatch.cpp b/src/gpu/batches/GrNinePatch.cpp
index cde3d266ca..3417c5c095 100644
--- a/src/gpu/batches/GrNinePatch.cpp
+++ b/src/gpu/batches/GrNinePatch.cpp
@@ -29,15 +29,14 @@ public:
 
     static const int kVertsPerRect = 4;
     static const int kIndicesPerRect = 6;
-    static const int kRectsPerInstance = 9; // We could skip empty rects
 
     GrNonAANinePatchBatch(GrColor color, const SkMatrix& viewMatrix, int imageWidth,
-                          int imageHeight, const SkIRect& center, const SkRect &dst)
+                          int imageHeight, std::unique_ptr<SkLatticeIter> iter, const SkRect &dst)
         : INHERITED(ClassID()) {
         Patch& patch = fPatches.push_back();
         patch.fViewMatrix = viewMatrix;
         patch.fColor = color;
-        patch.fCenter = center;
+        patch.fIter = std::move(iter);
         patch.fDst = dst;
 
         fImageWidth = imageWidth;
@@ -53,12 +52,9 @@ public:
         SkString str;
 
         for (int i = 0; i < fPatches.count(); ++i) {
-            str.appendf("%d: Color: 0x%08x Center [L: %d, T: %d, R: %d, B: %d], "
-                        "Dst [L: %.2f, T: %.2f, R: %.2f, B: %.2f]\n",
+            str.appendf("%d: Color: 0x%08x Dst [L: %.2f, T: %.2f, R: %.2f, B: %.2f]\n",
                         i,
                         fPatches[i].fColor,
-                        fPatches[i].fCenter.fLeft, fPatches[i].fCenter.fTop,
-                        fPatches[i].fCenter.fRight, fPatches[i].fCenter.fBottom,
                         fPatches[i].fDst.fLeft, fPatches[i].fDst.fTop,
                         fPatches[i].fDst.fRight, fPatches[i].fDst.fBottom);
         }
@@ -84,35 +80,40 @@ private:
 
         size_t vertexStride = gp->getVertexStride();
         int patchCnt = fPatches.count();
+        int numRects = 0;
+        for (int i = 0; i < patchCnt; i++) {
+            numRects += fPatches[i].fIter->numRects();
+        }
 
         SkAutoTUnref<const GrBuffer> indexBuffer(
                 target->resourceProvider()->refQuadIndexBuffer());
         InstancedHelper helper;
         void* vertices = helper.init(target, kTriangles_GrPrimitiveType, vertexStride,
                                      indexBuffer, kVertsPerRect,
-                                     kIndicesPerRect, patchCnt * kRectsPerInstance);
+                                     kIndicesPerRect, numRects);
         if (!vertices || !indexBuffer) {
             SkDebugf("Could not allocate vertices\n");
             return;
         }
 
+        intptr_t verts = reinterpret_cast<intptr_t>(vertices);
         for (int i = 0; i < patchCnt; i++) {
-            intptr_t verts = reinterpret_cast<intptr_t>(vertices) +
-                             i * kRectsPerInstance * kVertsPerRect * vertexStride;
-
             const Patch& patch = fPatches[i];
-            SkLatticeIter iter(fImageWidth, fImageHeight, patch.fCenter, patch.fDst);
+
+            // Apply the view matrix here if it is scale-translate.  Otherwise, we need to
+            // wait until we've created the dst rects.
+            bool isScaleTranslate = patch.fViewMatrix.isScaleTranslate();
+            if (isScaleTranslate) {
+                patch.fIter->mapDstScaleTranslate(patch.fViewMatrix);
+            }
 
             SkRect srcR, dstR;
-            while (iter.next(&srcR, &dstR)) {
+            intptr_t patchVerts = verts;
+            while (patch.fIter->next(&srcR, &dstR)) {
                 SkPoint* positions = reinterpret_cast<SkPoint*>(verts);
-
                 positions->setRectFan(dstR.fLeft, dstR.fTop,
                                       dstR.fRight, dstR.fBottom, vertexStride);
 
-                SkASSERT(!patch.fViewMatrix.hasPerspective());
-                patch.fViewMatrix.mapPointsWithStride(positions, vertexStride, kVertsPerRect);
-
                 // Setup local coords
                 static const int kLocalOffset = sizeof(SkPoint) + sizeof(GrColor);
                 SkPoint* coords = reinterpret_cast<SkPoint*>(verts + kLocalOffset);
@@ -126,6 +127,13 @@ private:
                 }
                 verts += kVertsPerRect * vertexStride;
             }
+
+            // If we didn't handle it above, apply the matrix here.
+            if (!isScaleTranslate) {
+                SkPoint* positions = reinterpret_cast<SkPoint*>(patchVerts);
+                patch.fViewMatrix.mapPointsWithStride(positions, vertexStride,
+                                                      kVertsPerRect * patch.fIter->numRects());
+            }
         }
         helper.recordDraw(target, gp.get());
     }
@@ -151,14 +159,14 @@ private:
             fOverrides = that->fOverrides;
         }
 
-        fPatches.push_back_n(that->fPatches.count(), that->fPatches.begin());
+        fPatches.move_back_n(that->fPatches.count(), that->fPatches.begin());
         this->joinBounds(*that);
         return true;
     }
 
     struct Patch {
         SkMatrix fViewMatrix;
-        SkIRect fCenter;
+        std::unique_ptr<SkLatticeIter> fIter;
         SkRect fDst;
         GrColor fColor;
     };
@@ -173,7 +181,8 @@ private:
 
 namespace GrNinePatch {
 GrDrawBatch* CreateNonAA(GrColor color, const SkMatrix& viewMatrix, int imageWidth, int imageHeight,
-                         const SkIRect& center, const SkRect& dst) {
-    return new GrNonAANinePatchBatch(color, viewMatrix, imageWidth, imageHeight, center, dst);
+                         std::unique_ptr<SkLatticeIter> iter, const SkRect& dst) {
+    return new GrNonAANinePatchBatch(color, viewMatrix, imageWidth, imageHeight, std::move(iter),
+                                     dst);
 }
 };
diff --git a/src/gpu/batches/GrNinePatch.h b/src/gpu/batches/GrNinePatch.h
index 0a4ffd6e03..02664c6449 100644
--- a/src/gpu/batches/GrNinePatch.h
+++ b/src/gpu/batches/GrNinePatch.h
@@ -9,16 +9,18 @@
 #define GrNinePatch_DEFINED
 
 #include "GrColor.h"
+#include "SkCanvas.h"
 
 class GrDrawBatch;
 class SkBitmap;
+class SkLatticeIter;
 class SkMatrix;
 struct SkIRect;
 struct SkRect;
 
 namespace GrNinePatch {
 GrDrawBatch* CreateNonAA(GrColor color, const SkMatrix& viewMatrix, int imageWidth, int imageHeight,
-                         const SkIRect& center, const SkRect& dst);
+                         std::unique_ptr<SkLatticeIter> iter, const SkRect& dst);
 };
 
 #endif