Implement instanced rendering for simple shapes

src/gpu/GrInstancedRendering.cpp

Index: src/gpu/GrInstancedRendering.cpp
diff --git a/src/gpu/GrInstancedRendering.cpp b/src/gpu/GrInstancedRendering.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..75a029bc4832494582d5ef24af0fb24af28f0e74
--- /dev/null
+++ b/src/gpu/GrInstancedRendering.cpp
@@ -0,0 +1,412 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "GrInstancedRendering.h"
+
+#include "GrBatchFlushState.h"
+#include "GrPipeline.h"
+#include "GrResourceProvider.h"
+#include "effects/GrInstanceProcessor.h"
+
+GrInstancedRendering::GrInstancedRendering(GrGpu* gpu, uint32_t supportedAAModes,
+                                           size_t sizeofBatchClass)
+    : fGpu(SkRef(gpu)),
+      fSupportedAAModes(supportedAAModes),
+      fState(State::kRecordingDraws),
+      fBatchAllocator(sizeofBatchClass) {
+    SkDEBUGCODE(fInUseBatchCount = 0;)
+}
+
+GrDrawBatch* GrInstancedRendering::recordRect(const SkRect& rect, const SkMatrix& viewMatrix,
+                                              GrColor color, bool antialias, uint32_t flags,
+                                              bool* useHWAA) {
+    return this->recordShape(kRect_ShapeType, rect, viewMatrix, color, rect, antialias, flags,
+                             useHWAA);
+}
+
+GrDrawBatch* GrInstancedRendering::recordRect(const SkRect& rect, const SkMatrix& viewMatrix,
+                                              GrColor color, const SkRect& localRect,
+                                              bool antialias, uint32_t flags, bool* useHWAA) {
+    return this->recordShape(kRect_ShapeType, rect, viewMatrix, color, localRect, antialias, flags,
+                             useHWAA);
+}
+
+GrDrawBatch* GrInstancedRendering::recordRect(const SkRect& rect, const SkMatrix& viewMatrix,
+                                              GrColor color, const SkMatrix& localMatrix,
+                                              bool antialias, uint32_t flags, bool* useHWAA) {
+    if (localMatrix.hasPerspective()) {
+        return nullptr; // Perspective is not yet supported in the local matrix.
+    }
+    if (Batch* batch = this->recordShape(kRect_ShapeType, rect, viewMatrix, color, rect, antialias,
+                                         flags, useHWAA)) {
+        fInstances.back().fInfo |= kLocalMatrix_InfoFlag;
+        this->appendParamsTexel(localMatrix.getScaleX(), localMatrix.getSkewX(),
+                                localMatrix.getTranslateX());
+        this->appendParamsTexel(localMatrix.getSkewY(), localMatrix.getScaleY(),
+                                localMatrix.getTranslateY());
+        batch->fInfo.fHasLocalMatrix = true;
+        batch->fInfo.fHasParams = true;
+        return batch;
+    }
+    return nullptr;
+}
+
+GrDrawBatch* GrInstancedRendering::recordOval(const SkRect& oval, const SkMatrix& viewMatrix,
+                                              GrColor color, bool antialias, uint32_t flags,
+                                              bool* useHWAA) {
+    return this->recordShape(kOval_ShapeType, oval, viewMatrix, color, oval, antialias, flags,
+                             useHWAA);
+}
+
+GrDrawBatch* GrInstancedRendering::recordRRect(const SkRRect& rrect, const SkMatrix& viewMatrix,
+                                               GrColor color, bool antialias, uint32_t flags,
+                                               bool* useHWAA) {
+    if (Batch* batch = this->recordShape(RRectShapeType(rrect), rrect.rect(), viewMatrix, color,
+                                         rrect.rect(), antialias, flags, useHWAA)) {
+        this->appendRRectParams(rrect, &batch->fInfo);
+        return batch;
+    }
+    return nullptr;
+}
+
+GrDrawBatch* GrInstancedRendering::recordDRRect(const SkRRect& outer, const SkRRect& inner,
+                                                const SkMatrix& viewMatrix, GrColor color,
+                                                bool antialias, uint32_t flags, bool* useHWAA) {
+    if (inner.getType() > SkRRect::kSimple_Type) {
+       return nullptr; // Complex inner round rects are not yet supported.
+    }
+    if (SkRRect::kEmpty_Type == inner.getType()) {
+        return this->recordRRect(outer, viewMatrix, color, antialias, flags, useHWAA);
+    }
+    if (Batch* batch = this->recordShape(RRectShapeType(outer), outer.rect(), viewMatrix, color,
+                                         outer.rect(), antialias, flags, useHWAA)) {
+        this->appendRRectParams(outer, &batch->fInfo);
+        ShapeType innerShapeType = RRectShapeType(inner);
+        batch->fInfo.fInnerShapeTypes |= (1 << innerShapeType);
+        fInstances.back().fInfo |= (innerShapeType << kInnerShapeType_InfoBit);
+        this->appendParamsTexel(inner.rect().asScalars(), 4);
+        this->appendRRectParams(inner, &batch->fInfo);
+        batch->fInfo.fHasParams = true;
+        return batch;
+    }
+    return nullptr;
+}
+
+GrInstancedRendering::Batch* GrInstancedRendering::recordShape(ShapeType type, const SkRect& bounds,
+                                                               const SkMatrix& viewMatrix,
+                                                               GrColor color,
+                                                               const SkRect& localRect,
+                                                               bool antialias, uint32_t flags,
+                                                               bool* useHWAA) {
+    SkASSERT(State::kRecordingDraws == fState);
+
+    uint32_t paramsIdx = fParams.count();
+    if (paramsIdx > kParamsIdx_InfoMask) {
+        return nullptr; // paramsIdx is too large for its allotted space.
+    }
+
+    AntialiasMode antialiasMode;
+    if (!this->selectAntialiasMode(viewMatrix, antialias, flags, &antialiasMode, useHWAA)) {
+        return nullptr;
+    }
+
+    Batch* batch = this->constructBatch(fBatchAllocator.push_back(), fInstances.count());
+    SkASSERT(batch == fBatchAllocator.back()); // We rely on batch ptr equality with the allocator.
+    SkDEBUGCODE(++fInUseBatchCount;)
+    batch->fInfo.fAntialiasMode = antialiasMode;
+    batch->fInfo.fShapeTypes = (1 << type);
+    batch->fInfo.fCannotDiscard = !(flags & kUseDiscard_Flag);
+
+    Instance& instance = fInstances.push_back();
+    instance.fInfo = (type << kShapeType_InfoBit) | paramsIdx;
+
+    // The instanced shape renderer draws rectangles of [-1, -1, +1, +1], so we find the matrix that
+    // will map this rectangle to the same device coordinates as "viewMatrix * bounds".
+    float sx = 0.5f * bounds.width();
+    float sy = 0.5f * bounds.height();
+    float tx = sx + bounds.fLeft;
+    float ty = sy + bounds.fTop;
+    if (!viewMatrix.hasPerspective()) {
+        float* m = instance.fShapeMatrix2x3;
+        m[0] = viewMatrix.getScaleX() * sx;
+        m[1] = viewMatrix.getSkewX() * sy;
+        m[2] = viewMatrix.getTranslateX() +
+               viewMatrix.getScaleX() * tx + viewMatrix.getSkewX() * ty;
+
+        m[3] = viewMatrix.getSkewY() * sx;
+        m[4] = viewMatrix.getScaleY() * sy;
+        m[5] = viewMatrix.getTranslateY() +
+               viewMatrix.getSkewY() * tx + viewMatrix.getScaleY() * ty;
+
+        // Since 'm' is a 2x3 matrix that maps the rect [-1, +1] into the shape's device-space quad,
+        // it's quite simple to find the bounding rectangle:
+        float devBoundsHalfWidth = fabsf(m[0]) + fabsf(m[1]);
+        float devBoundsHalfHeight = fabsf(m[3]) + fabsf(m[4]);
+        batch->fBounds.fLeft = m[2] - devBoundsHalfWidth;
+        batch->fBounds.fRight = m[2] + devBoundsHalfWidth;
+        batch->fBounds.fTop = m[5] - devBoundsHalfHeight;
+        batch->fBounds.fBottom = m[5] + devBoundsHalfHeight;
+
+        // TODO: Is this worth the CPU overhead?
+        batch->fInfo.fNonSquare =
+            fabsf(devBoundsHalfHeight - devBoundsHalfWidth) > 0.5f || // Early out.
+            fabs(m[0] * m[3] + m[1] * m[4]) > 1e-3f || // Skew?
+            fabs(m[0] * m[0] + m[1] * m[1] - m[3] * m[3] - m[4] * m[4]) > 1e-2f; // Diff. lengths?
+    } else {
+        SkMatrix shapeMatrix(viewMatrix);
+        shapeMatrix.preTranslate(tx, ty);
+        shapeMatrix.preScale(sx, sy);
+        instance.fInfo |= kPerspective_InfoFlag;
+
+        float* m = instance.fShapeMatrix2x3;
+        m[0] = SkScalarToFloat(shapeMatrix.getScaleX());
+        m[1] = SkScalarToFloat(shapeMatrix.getSkewX());
+        m[2] = SkScalarToFloat(shapeMatrix.getTranslateX());
+        m[3] = SkScalarToFloat(shapeMatrix.getSkewY());
+        m[4] = SkScalarToFloat(shapeMatrix.getScaleY());
+        m[5] = SkScalarToFloat(shapeMatrix.getTranslateY());
+
+        // Send the perspective column as a param.
+        this->appendParamsTexel(shapeMatrix[SkMatrix::kMPersp0], shapeMatrix[SkMatrix::kMPersp1],
+                                shapeMatrix[SkMatrix::kMPersp2]);
+        batch->fInfo.fHasPerspective = true;
+        batch->fInfo.fHasParams = true;
+
+        viewMatrix.mapRect(&batch->fBounds, bounds);
+
+        batch->fInfo.fNonSquare = true;
+    }
+
+    instance.fColor = color;
+
+    const float* rectAsFloats = localRect.asScalars(); // Ensure SkScalar == float.
+    memcpy(&instance.fLocalRect, rectAsFloats, 4 * sizeof(float));
+
+    return batch;
+}
+
+inline bool GrInstancedRendering::selectAntialiasMode(const SkMatrix& viewMatrix, bool antialias,
+                                                      uint32_t flags, AntialiasMode* antialiasMode,
+                                                      bool* useHWAA) {
+    SkASSERT(flags & (kColorWrite_Flag | kStencilWrite_Flag));
+    SkASSERT((flags & (kColorBufferMSAA_Flag | kStencilBufferMSAA_Flag)) != kColorBufferMSAA_Flag);
+
+    if (!antialias) {
+        if (!(fSupportedAAModes & kNone_AntialiasFlag)) {
+            return false;
+        }
+        if ((flags & (kStencilWrite_Flag | kUseDiscard_Flag)) == kStencilWrite_Flag) {
+            // We can only subtract coverage from the stencil output via discard when no MSAA.
+            return false;
+        }
+        *antialiasMode = kNone_AntialiasMode;
+        *useHWAA = false;
+        return true;
+    }
+
+    if (!(flags & (kColorBufferMSAA_Flag | kStencilWrite_Flag)) &&
+        viewMatrix.preservesRightAngles()) {
+        SkASSERT(fSupportedAAModes & kCoverage_AntialiasFlag);
+        *antialiasMode = kCoverage_AntialiasMode;
+        *useHWAA = false;
+        return true;
+    }
+
+    if ((fSupportedAAModes & kMSAA_AntialiasFlag) && (flags & kStencilBufferMSAA_Flag)) {
+        if ((flags ^ kColorWrite_Flag) & (kColorWrite_Flag | kColorBufferMSAA_Flag)) {
+            // We either do not write color, or the color buffer is multisampled.
+            *antialiasMode = kMSAA_AntialiasMode;
+            *useHWAA = true;
+            return true;
+        }
+        if (fSupportedAAModes & kMixedSamples_AntialiasFlag) {
+            *antialiasMode = kMixedSamples_AntialiasMode;
+            *useHWAA = true;
+            return true;
+        }
+    }
+
+    return false;
+}
+
+void GrInstancedRendering::appendRRectParams(const SkRRect& rrect, BatchInfo* batchInfo) {
+    switch (rrect.getType()) {
+        case SkRRect::kSimple_Type: {
+            const SkVector& radii = rrect.getSimpleRadii();
+            this->appendParamsTexel(radii.x(), radii.y(), rrect.width(), rrect.height());
+            batchInfo->fHasParams = true;
+            return;
+        }
+        case SkRRect::kNinePatch_Type: {
+            float twoOverW = 2 / rrect.width();
+            float twoOverH = 2 / rrect.height();
+            const SkVector& radiiTL = rrect.radii(SkRRect::kUpperLeft_Corner);
+            const SkVector& radiiBR = rrect.radii(SkRRect::kLowerRight_Corner);
+            this->appendParamsTexel(radiiTL.x() * twoOverW, radiiBR.x() * twoOverW,
+                                    radiiTL.y() * twoOverH, radiiBR.y() * twoOverH);
+            batchInfo->fHasParams = true;
+            return;
+        }
+        case SkRRect::kComplex_Type: {
+            /**
+             * The x and y radii of each arc are stored in separate vectors,
+             * in the following order:
+             *
+             *        __x1 _ _ _ x3__
+             *    y1 |               | y2
+             *
+             *       |               |
+             *
+             *    y3 |__   _ _ _   __| y4
+             *          x2       x4
+             *
+             */
+            float twoOverW = 2 / rrect.width();
+            float twoOverH = 2 / rrect.height();
+            const SkVector& radiiTL = rrect.radii(SkRRect::kUpperLeft_Corner);
+            const SkVector& radiiTR = rrect.radii(SkRRect::kUpperRight_Corner);
+            const SkVector& radiiBR = rrect.radii(SkRRect::kLowerRight_Corner);
+            const SkVector& radiiBL = rrect.radii(SkRRect::kLowerLeft_Corner);
+            this->appendParamsTexel(radiiTL.x() * twoOverW, radiiBL.x() * twoOverW,
+                                    radiiTR.x() * twoOverW, radiiBR.x() * twoOverW);
+            this->appendParamsTexel(radiiTL.y() * twoOverH, radiiTR.y() * twoOverH,
+                                    radiiBL.y() * twoOverH, radiiBR.y() * twoOverH);
+            batchInfo->fHasParams = true;
+            return;
+        }
+        default: return;
+    }
+}
+
+void GrInstancedRendering::appendParamsTexel(const SkScalar* vals, int count) {
+    SkASSERT(count <= 4 && count >= 0);
+    const float* valsAsFloats = vals; // Ensure SkScalar == float.
+    memcpy(&fParams.push_back(), valsAsFloats, count * sizeof(float));
+}
+
+void GrInstancedRendering::appendParamsTexel(SkScalar x, SkScalar y, SkScalar z, SkScalar w) {
+    ParamsTexel& texel = fParams.push_back();
+    texel.fX = SkScalarToFloat(x);
+    texel.fY = SkScalarToFloat(y);
+    texel.fZ = SkScalarToFloat(z);
+    texel.fW = SkScalarToFloat(w);
+}
+
+void GrInstancedRendering::appendParamsTexel(SkScalar x, SkScalar y, SkScalar z) {
+    ParamsTexel& texel = fParams.push_back();
+    texel.fX = SkScalarToFloat(x);
+    texel.fY = SkScalarToFloat(y);
+    texel.fZ = SkScalarToFloat(z);
+}
+
+void GrInstancedRendering::Batch::computePipelineOptimizations(GrInitInvariantOutput* color,
+                                                            GrInitInvariantOutput* coverage,
+                                                            GrBatchToXPOverrides* overrides) const {
+    // We need to be careful about fInfo here and consider how it might change as batches combine.
+    // e.g. We can't make an assumption based on fInfo.isSimpleRects() because the batch might
+    // later combine with an oval.
+    color->setUnknownFourComponents();
+    if (fInfo.fAntialiasMode >= kMSAA_AntialiasMode) {
+        coverage->setKnownSingleComponent(255);
+    } else if (kNone_AntialiasMode == fInfo.fAntialiasMode && !fInfo.fCannotDiscard) {
+        // We can rely on fCannotDiscard because this particular field does not change.
+        coverage->setKnownSingleComponent(255);
+    } else {
+        coverage->setUnknownSingleComponent();
+    }
+}
+
+void GrInstancedRendering::Batch::initBatchTracker(const GrXPOverridesForBatch& overrides) {
+    fInfo.fUsesLocalCoords = overrides.readsLocalCoords();
+    fInfo.fCannotTweakAlphaForCoverage = !overrides.canTweakAlphaForCoverage();
+
+    GrColor overrideColor;
+    if (overrides.getOverrideColorIfSet(&overrideColor)) {
+        SkASSERT(State::kRecordingDraws == fInstancedRendering->fState);
+        SkASSERT(!fIsCombined);
+        fInstancedRendering->fInstances[fFirstInstanceIdx].fColor = overrideColor;
+    }
+}
+
+void GrInstancedRendering::beginFlush(GrResourceProvider* rp) {
+    SkASSERT(State::kRecordingDraws == fState);
+    fState = State::kFlushing;
+
+    if (fInstances.empty()) {
+        return;
+    }
+
+    if (!fVertexBuffer) {
+        fVertexBuffer.reset(GrInstanceProcessor::FindOrCreateVertexBuffer(fGpu));
+        if (!fVertexBuffer) {
+            return;
+        }
+    }
+
+    if (!fIndexBuffer) {
+        fIndexBuffer.reset(GrInstanceProcessor::FindOrCreateIndex8Buffer(fGpu));
+        if (!fIndexBuffer) {
+            return;
+        }
+    }
+
+    fInstanceBuffer.reset(rp->createBuffer(fInstances.count() * sizeof(Instance),
+                                           kVertex_GrBufferType, kDynamic_GrAccessPattern,
+                                           GrResourceProvider::kNoPendingIO_Flag,
+                                           fInstances.begin()));
+    if (!fInstanceBuffer) {
+        return;
+    }
+
+    if (!fParams.empty()) {
+        fParamsBuffer.reset(rp->createBuffer(fParams.count() * sizeof(ParamsTexel),
+                                             kTexel_GrBufferType, kDynamic_GrAccessPattern,
+                                             GrResourceProvider::kNoPendingIO_Flag,
+                                             fParams.begin()));
+        if (!fParamsBuffer) {
+            return;
+        }
+    }
+
+    this->onBeginFlush(rp);
+}
+
+void GrInstancedRendering::Batch::onDraw(GrBatchFlushState* state) {
+    SkASSERT(State::kFlushing == fInstancedRendering->fState);
+    SkASSERT(state->gpu() == fInstancedRendering->gpu());
+
+    GrInstanceProcessor instProc(fInfo, fInstancedRendering->fParamsBuffer);
+    fInstancedRendering->onDraw(*this->pipeline(), instProc, this);
+}
+
+void GrInstancedRendering::Batch::onDelete() const {
+    this->~Batch();
+    SkDEBUGCODE(--fInstancedRendering->fInUseBatchCount);
+    if (this == fInstancedRendering->fBatchAllocator.back()) {
+        fInstancedRendering->fBatchAllocator.pop_back();
+    }
+}
+
+void GrInstancedRendering::endFlush() {
+    SkASSERT(0 == fInUseBatchCount);
+    fBatchAllocator.reset();
+    // Hold on to the shape coords and index buffers.
+    fInstances.reset();
+    fParams.reset();
+    fInstanceBuffer.reset();
+    fParamsBuffer.reset();
+    this->onEndFlush();
+    fState = State::kRecordingDraws;
+}
+
+void GrInstancedRendering::resetGpuResources(ResetType resetType) {
+    fVertexBuffer.reset();
+    fIndexBuffer.reset();
+    fInstanceBuffer.reset();
+    fParamsBuffer.reset();
+    this->onResetGpuResources(resetType);
+}