Improve nvpr glyph batching

src/gpu/batches/GrDrawPathBatch.cpp

Index: src/gpu/batches/GrDrawPathBatch.cpp
diff --git a/src/gpu/batches/GrDrawPathBatch.cpp b/src/gpu/batches/GrDrawPathBatch.cpp
index c2007ce933b7399cf4e9ab55307d7615e87da0cd..3264c0fbafe49f84a56dfc8d6fdb89082c6d11a6 100644
--- a/src/gpu/batches/GrDrawPathBatch.cpp
+++ b/src/gpu/batches/GrDrawPathBatch.cpp
@@ -7,6 +7,10 @@
 
 #include "GrDrawPathBatch.h"
 
+static void pre_translate_transform_values(const float* xforms,
+                                           GrPathRendering::PathTransformType type, int count,
+                                           SkScalar x, SkScalar y, float* dst);
+
 SkString GrDrawPathBatch::dumpInfo() const {
     SkString string;
     string.printf("PATH: 0x%p", fPath.get());
@@ -25,51 +29,59 @@ void GrDrawPathBatch::onDraw(GrBatchFlushState* state) {
     state->gpu()->pathRendering()->drawPath(args, fPath.get());
 }
 
-GrDrawPathRangeBatch::~GrDrawPathRangeBatch() {
-    for (DrawList::Iter iter(fDraws); iter.get(); iter.next()) {
-        (*iter.get())->unref();
-    }
-}
-
 SkString GrDrawPathRangeBatch::dumpInfo() const {
     SkString string;
-    string.printf("RANGE: 0x%p COUNTS: [", *fDraws.head());
+    string.printf("RANGE: 0x%p COUNTS: [", fPathRange.get());
     for (DrawList::Iter iter(fDraws); iter.get(); iter.next()) {
-        string.appendf("%d ,", (*iter.get())->count());
+        string.appendf("%d, ", iter.get()->fInstanceData->count());
     }
     string.remove(string.size() - 2, 2);
     string.append("]");
     return string;
 }
 
-GrDrawPathRangeBatch::GrDrawPathRangeBatch(const SkMatrix& viewMatrix, const SkMatrix& localMatrix,
-                                           GrColor color, GrPathRendering::FillType fill,
-                                           GrPathRange* range, GrPathRangeDraw* draw,
-                                           const SkRect& bounds)
+GrDrawPathRangeBatch::GrDrawPathRangeBatch(const SkMatrix& viewMatrix, SkScalar scale, SkScalar x,
+                                           SkScalar y, GrColor color,
+                                           GrPathRendering::FillType fill, GrPathRange* range,
+                                           const InstanceData* instanceData, const SkRect& bounds)
     : INHERITED(ClassID(), viewMatrix, color, fill)
     , fPathRange(range)
-    , fLocalMatrix(localMatrix) {
-    SkDEBUGCODE(draw->fUsedInBatch = true;)
-    fDraws.addToHead(SkRef(draw));
-    fTotalPathCount = draw->count();
+    , fTotalPathCount(instanceData->count())
+    , fScale(scale) {
+    fDraws.addToHead()->set(instanceData, x, y);
     fBounds = bounds;
 }
 
 bool GrDrawPathRangeBatch::onCombineIfPossible(GrBatch* t, const GrCaps& caps) {
     GrDrawPathRangeBatch* that = t->cast<GrDrawPathRangeBatch>();
-    if (this->fPathRange.get() != that->fPathRange.get()) {
-        return false;
-    }
-    if (!GrPathRangeDraw::CanMerge(**this->fDraws.head(), **that->fDraws.head())) {
+    if (this->fPathRange.get() != that->fPathRange.get() ||
+        this->transformType() != that->transformType() ||
+        this->fScale != that->fScale ||
+        this->color() != that->color() ||
+        !this->viewMatrix().cheapEqualTo(that->viewMatrix())) {
         return false;
     }
     if (!GrPipeline::AreEqual(*this->pipeline(), *that->pipeline(), false)) {
         return false;
     }
-    if (this->color() != that->color() ||
-        !this->viewMatrix().cheapEqualTo(that->viewMatrix()) ||
-        !fLocalMatrix.cheapEqualTo(that->fLocalMatrix)) {
-        return false;
+    switch (fDraws.head()->fInstanceData->transformType()) {
+        case GrPathRendering::kNone_PathTransformType:
+            if (this->fDraws.head()->fX != that->fDraws.head()->fX ||
+                this->fDraws.head()->fY != that->fDraws.head()->fY) {
+                return false;
+            }
+            break;
+        case GrPathRendering::kTranslateX_PathTransformType:
+            if (this->fDraws.head()->fY != that->fDraws.head()->fY) {
+                return false;
+            }
+            break;
+        case GrPathRendering::kTranslateY_PathTransformType:
+            if (this->fDraws.head()->fX != that->fDraws.head()->fX) {
+                return false;
+            }
+            break;
+        default: break;
     }
     // TODO: Check some other things here. (winding, opaque, pathProc color, vm, ...)
     // Try to combine this call with the previous DrawPaths. We do this by stenciling all the
@@ -85,47 +97,117 @@ bool GrDrawPathRangeBatch::onCombineIfPossible(GrBatch* t, const GrCaps& caps) {
     }
     SkASSERT(!that->overrides().willColorBlendWithDst());
     fTotalPathCount += that->fTotalPathCount;
-    while (GrPathRangeDraw** head = that->fDraws.head()) {
-        fDraws.addToTail(*head);
-        // We're stealing that's refs, so pop without unreffing.
+    while (Draw* head = that->fDraws.head()) {
+        Draw* draw = fDraws.addToTail();
+        draw->fInstanceData.reset(head->fInstanceData.detach());
+        draw->fX = head->fX;
+        draw->fY = head->fY;
         that->fDraws.popHead();
     }
     return true;
 }
 
 void GrDrawPathRangeBatch::onDraw(GrBatchFlushState* state) {
-    GrProgramDesc  desc;
+    const Draw& head = *fDraws.head();
+
+    SkMatrix drawMatrix(this->viewMatrix());
+    drawMatrix.preScale(fScale, fScale);
+    drawMatrix.preTranslate(head.fX, head.fY);
+
+    SkMatrix localMatrix;
+    localMatrix.setScale(fScale, fScale);
+    localMatrix.preTranslate(head.fX, head.fY);
+
     SkAutoTUnref<GrPathProcessor> pathProc(GrPathProcessor::Create(this->color(),
                                                                    this->overrides(),
-                                                                   this->viewMatrix(),
-                                                                   fLocalMatrix));
+                                                                   drawMatrix,
+                                                                   localMatrix));
+
+    GrProgramDesc  desc;
     state->gpu()->buildProgramDesc(&desc, *pathProc, *this->pipeline());
     GrPathRendering::DrawPathArgs args(pathProc, this->pipeline(),
-                                        &desc, &this->stencilSettings());
+                                       &desc, &this->stencilSettings());
+
     if (fDraws.count() == 1) {
-        const GrPathRangeDraw& draw = **fDraws.head();
-        state->gpu()->pathRendering()->drawPaths(args, fPathRange.get(), draw.indices(),
-            GrPathRange::kU16_PathIndexType, draw.transforms(), draw.transformType(),
-            draw.count());
-        return;
+        const InstanceData& instances = *head.fInstanceData;
+        state->gpu()->pathRendering()->drawPaths(args, fPathRange.get(), instances.indices(),
+                                                 GrPathRange::kU16_PathIndexType,
+                                                 instances.transformValues(),
+                                                 instances.transformType(),
+                                                 instances.count());
+    } else {
+        int floatsPerTransform = GrPathRendering::PathTransformSize(this->transformType());
+#if defined(GOOGLE3)
+        //Stack frame size is limited in GOOGLE3.
+        SkAutoSTMalloc<512, float> transformStorage(transformSize * fTotalPathCount);
+        SkAutoSTMalloc<256, uint16_t> indexStorage(fTotalPathCount);
+#else
+        SkAutoSTMalloc<4096, float> transformStorage(floatsPerTransform * fTotalPathCount);
+        SkAutoSTMalloc<2048, uint16_t> indexStorage(fTotalPathCount);
+#endif
+        int idx = 0;
+        for (DrawList::Iter iter(fDraws); iter.get(); iter.next()) {
+            const Draw& draw = *iter.get();
+            const InstanceData& instances = *draw.fInstanceData;
+            memcpy(&indexStorage[idx], instances.indices(), instances.count() * sizeof(uint16_t));
+            pre_translate_transform_values(instances.transformValues(), this->transformType(),
+                                           instances.count(),
+                                           draw.fX - head.fX, draw.fY - head.fY,
+                                           &transformStorage[floatsPerTransform * idx]);
+            idx += instances.count();
+
+            // TODO: Support mismatched transform types if we start using more types other than 2D.
+            SkASSERT(instances.transformType() == this->transformType());
+        }
+        SkASSERT(idx == fTotalPathCount);
+
+        state->gpu()->pathRendering()->drawPaths(args, fPathRange.get(), indexStorage,
+                                                 GrPathRange::kU16_PathIndexType, transformStorage,
+                                                 this->transformType(), fTotalPathCount);
     }
+}
 
-    GrPathRendering::PathTransformType transformType = (*fDraws.head())->transformType();
-    int floatsPerTransform = GrPathRendering::PathTransformSize(transformType);
-    SkAutoSTMalloc<512, float> transformStorage(floatsPerTransform * fTotalPathCount);
-    SkAutoSTMalloc<256, uint16_t> indexStorage(fTotalPathCount);
-    uint16_t* indices = indexStorage.get();
-    float* transforms = transformStorage.get();
-    for (DrawList::Iter iter(fDraws); iter.get(); iter.next()) {
-        SkASSERT((*iter.get())->transformType() == transformType);
-        int cnt = (*iter.get())->count();
-        memcpy(indices, (*iter.get())->indices(), cnt * sizeof(uint16_t));
-        indices += cnt;
-        memcpy(transforms, (*iter.get())->transforms(), cnt * floatsPerTransform * sizeof(float));
-        transforms += cnt * floatsPerTransform;
+inline void pre_translate_transform_values(const float* xforms,
+                                           GrPathRendering::PathTransformType type, int count,
+                                           SkScalar x, SkScalar y, float* dst) {
+    if (0 == x && 0 == y) {
+        memcpy(dst, xforms, count * GrPathRendering::PathTransformSize(type) * sizeof(float));
+        return;
+    }
+    switch (type) {
+        case GrPathRendering::kNone_PathTransformType:
+            SkFAIL("Cannot pre-translate kNone_PathTransformType.");
+            break;
+        case GrPathRendering::kTranslateX_PathTransformType:
+            SkASSERT(0 == y);
+            for (int i = 0; i < count; i++) {
+                dst[i] = xforms[i] + x;
+            }
+            break;
+        case GrPathRendering::kTranslateY_PathTransformType:
+            SkASSERT(0 == x);
+            for (int i = 0; i < count; i++) {
+                dst[i] = xforms[i] + y;
+            }
+            break;
+        case GrPathRendering::kTranslate_PathTransformType:
+            for (int i = 0; i < 2 * count; i += 2) {
+                dst[i] = xforms[i] + x;
+                dst[i + 1] = xforms[i + 1] + y;
+            }
+            break;
+        case GrPathRendering::kAffine_PathTransformType:
+            for (int i = 0; i < 6 * count; i += 6) {
+                dst[i] = xforms[i];
+                dst[i + 1] = xforms[i + 1];
+                dst[i + 2] = xforms[i] * x + xforms[i + 1] * y + xforms[i + 2];
+                dst[i + 3] = xforms[i + 3];
+                dst[i + 4] = xforms[i + 4];
+                dst[i + 5] = xforms[i + 3] * x + xforms[i + 4] * y + xforms[i + 5];
+            }
+            break;
+        default:
+            SkFAIL("Unknown transform type.");
+            break;
     }
-    SkASSERT(indices - indexStorage.get() == fTotalPathCount);
-    state->gpu()->pathRendering()->drawPaths(args, fPathRange.get(), indexStorage.get(),
-        GrPathRange::kU16_PathIndexType, transformStorage.get(), transformType,
-        fTotalPathCount);
 }