diff options
-rw-r--r-- | bench/QuickRejectBench.cpp | 61 | ||||
-rw-r--r-- | include/core/SkCanvas.h | 21 | ||||
-rw-r--r-- | include/core/SkPostConfig.h | 12 | ||||
-rw-r--r-- | src/core/SkCanvas.cpp | 125 | ||||
-rw-r--r-- | tests/QuickRejectTest.cpp | 38 |
5 files changed, 210 insertions, 47 deletions
diff --git a/bench/QuickRejectBench.cpp b/bench/QuickRejectBench.cpp new file mode 100644 index 0000000000..1feade41f5 --- /dev/null +++ b/bench/QuickRejectBench.cpp @@ -0,0 +1,61 @@ +/* + * 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 "Benchmark.h" +#include "SkCanvas.h" +#include "SkRandom.h" + +class QuickRejectBench : public Benchmark { + enum { N = 1000000 }; + float fFloats[N]; + int fInts [N]; + + const char* onGetName() override { return "quick_reject"; } + bool isSuitableFor(Backend backend) override { return backend != kNonRendering_Backend; } + + void onDelayedSetup() override { + SkRandom rand; + for (int i = 0; i < N; ++i) { + fFloats[i] = 300.0f * (rand.nextSScalar1() + 0.5f); + } + } + + void onDraw(int loops, SkCanvas* canvas) override { + while (loops --> 0) { + for (int i = 0; i < N - 4; i++) { + if (canvas->quickReject(*(SkRect*)(fFloats+i))) { + fInts[i] = 11; + } else { + fInts[i] = 24; + } + } + } + } +}; +DEF_BENCH( return new QuickRejectBench; ) + +class ConcatBench : public Benchmark { + SkMatrix fMatrix; + + const char* onGetName() override { return "concat"; } + bool isSuitableFor(Backend backend) override { return backend != kNonRendering_Backend; } + + void onDelayedSetup() override { + SkRandom r; + fMatrix.setScale(5.0f, 5.0f); + fMatrix.setTranslateX(10.0f); + fMatrix.setTranslateY(10.0f); + } + + void onDraw(int loops, SkCanvas* canvas) override { + while (loops --> 0) { + canvas->setMatrix(SkMatrix::MakeScale(3.0f)); + canvas->concat(fMatrix); + } + } +}; +DEF_BENCH( return new ConcatBench; ) diff --git a/include/core/SkCanvas.h b/include/core/SkCanvas.h index ef2d7dad70..73fcf43a5f 100644 --- a/include/core/SkCanvas.h +++ b/include/core/SkCanvas.h @@ -1609,25 +1609,18 @@ private: */ bool canDrawBitmapAsSprite(SkScalar x, SkScalar y, int w, int h, const SkPaint&); - /* These maintain a cache of the clip bounds in local coordinates, - (converted to 2s-compliment if floats are slow). + + /** + * Keep track of the device clip bounds and if the matrix is scale-translate. This allows + * us to do a fast quick reject in the common case. */ - mutable SkRect fCachedLocalClipBounds; - mutable bool fCachedLocalClipBoundsDirty; + bool fConservativeIsScaleTranslate; + SkRect fDeviceClipBounds; + bool fAllowSoftClip; bool fAllowSimplifyClip; const bool fConservativeRasterClip; - const SkRect& getLocalClipBounds() const { - if (fCachedLocalClipBoundsDirty) { - if (!this->getClipBounds(&fCachedLocalClipBounds)) { - fCachedLocalClipBounds.setEmpty(); - } - fCachedLocalClipBoundsDirty = false; - } - return fCachedLocalClipBounds; - } - class AutoValidateClip : ::SkNoncopyable { public: explicit AutoValidateClip(SkCanvas* canvas) : fCanvas(canvas) { diff --git a/include/core/SkPostConfig.h b/include/core/SkPostConfig.h index b67c863887..97c868f30f 100644 --- a/include/core/SkPostConfig.h +++ b/include/core/SkPostConfig.h @@ -275,6 +275,18 @@ # endif #endif +/** + * If your judgment is better than the compiler's (i.e. you've profiled it), + * you can use SK_NEVER_INLINE to prevent inlining. + */ +#if !defined(SK_NEVER_INLINE) +# if defined(SK_BUILD_FOR_WIN) +# define SK_NEVER_INLINE __declspec(noinline) +# else +# define SK_NEVER_INLINE SK_ATTRIBUTE(noinline) +# endif +#endif + ////////////////////////////////////////////////////////////////////// #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1 diff --git a/src/core/SkCanvas.cpp b/src/core/SkCanvas.cpp index 24d9506bd7..ce786bfe8a 100644 --- a/src/core/SkCanvas.cpp +++ b/src/core/SkCanvas.cpp @@ -22,6 +22,7 @@ #include "SkLatticeIter.h" #include "SkMatrixUtils.h" #include "SkMetaData.h" +#include "SkNx.h" #include "SkPaintPriv.h" #include "SkPatchUtils.h" #include "SkPicture.h" @@ -632,16 +633,24 @@ bool AutoDrawLooper::doNext(SkDrawFilter::Type drawType) { //////////////////////////////////////////////////////////////////////////// +static inline SkRect qr_clip_bounds(const SkIRect& bounds) { + // Expand bounds out by 1 in case we are anti-aliasing. We store the + // bounds as floats to enable a faster quick reject implementation. + SkRect dst; + SkNx_cast<float>(Sk4i::Load(&bounds.fLeft) + Sk4i(-1,-1,1,1)).store(&dst.fLeft); + return dst; +} + void SkCanvas::resetForNextPicture(const SkIRect& bounds) { this->restoreToCount(1); - fCachedLocalClipBounds.setEmpty(); - fCachedLocalClipBoundsDirty = true; fClipStack->reset(); fMCRec->reset(bounds); // We're peering through a lot of structs here. Only at this scope do we // know that the device is an SkBitmapDevice (really an SkNoPixelsBitmapDevice). static_cast<SkBitmapDevice*>(fMCRec->fLayer->fDevice)->setNewSize(bounds.size()); + fDeviceClipBounds = qr_clip_bounds(bounds); + fConservativeIsScaleTranslate = true; } SkBaseDevice* SkCanvas::init(SkBaseDevice* device, InitFlags flags) { @@ -652,8 +661,6 @@ SkBaseDevice* SkCanvas::init(SkBaseDevice* device, InitFlags flags) { // const-cast. *const_cast<bool*>(&fConservativeRasterClip) = SkToBool(flags & kConservativeRasterClip_InitFlag); - fCachedLocalClipBounds.setEmpty(); - fCachedLocalClipBoundsDirty = true; fAllowSoftClip = true; fAllowSimplifyClip = false; fDeviceCMDirty = true; @@ -682,7 +689,10 @@ SkBaseDevice* SkCanvas::init(SkBaseDevice* device, InitFlags flags) { SkASSERT(fProps.pixelGeometry() == device->surfaceProps().pixelGeometry()); fMCRec->fLayer->fDevice = SkRef(device); fMCRec->fRasterClip.setRect(device->getGlobalBounds()); + fDeviceClipBounds = qr_clip_bounds(device->getGlobalBounds()); + fConservativeIsScaleTranslate = true; } + return device; } @@ -1097,8 +1107,8 @@ bool SkCanvas::clipRectBounds(const SkRect* bounds, SaveLayerFlags saveLayerFlag // early exit if the layer's bounds are clipped out if (!ir.intersect(clipBounds)) { if (BoundsAffectsClip(saveLayerFlags)) { - fCachedLocalClipBoundsDirty = true; fMCRec->fRasterClip.setEmpty(); + fDeviceClipBounds.setEmpty(); } return false; } @@ -1109,9 +1119,9 @@ bool SkCanvas::clipRectBounds(const SkRect* bounds, SaveLayerFlags saveLayerFlag if (BoundsAffectsClip(saveLayerFlags)) { // Simplify the current clips since they will be applied properly during restore() - fCachedLocalClipBoundsDirty = true; fClipStack->clipDevRect(ir, SkRegion::kReplace_Op); fMCRec->fRasterClip.setRect(ir); + fDeviceClipBounds = qr_clip_bounds(ir); } if (intersection) { @@ -1303,7 +1313,6 @@ void SkCanvas::internalRestore() { SkASSERT(fMCStack.count() != 0); fDeviceCMDirty = true; - fCachedLocalClipBoundsDirty = true; fClipStack->restore(); @@ -1337,6 +1346,11 @@ void SkCanvas::internalRestore() { // no need to update fMCRec, 'cause we're killing the canvas } } + + if (fMCRec) { + fConservativeIsScaleTranslate = fMCRec->fMatrix.isScaleTranslate(); + fDeviceClipBounds = qr_clip_bounds(fMCRec->fRasterClip.getBounds()); + } } sk_sp<SkSurface> SkCanvas::makeSurface(const SkImageInfo& info, const SkSurfaceProps* props) { @@ -1491,21 +1505,20 @@ void SkCanvas::concat(const SkMatrix& matrix) { this->checkForDeferredSave(); fDeviceCMDirty = true; - fCachedLocalClipBoundsDirty = true; fMCRec->fMatrix.preConcat(matrix); - + fConservativeIsScaleTranslate = fMCRec->fMatrix.isScaleTranslate(); this->didConcat(matrix); } void SkCanvas::internalSetMatrix(const SkMatrix& matrix) { fDeviceCMDirty = true; - fCachedLocalClipBoundsDirty = true; fMCRec->fMatrix = matrix; } void SkCanvas::setMatrix(const SkMatrix& matrix) { this->checkForDeferredSave(); this->internalSetMatrix(matrix); + fConservativeIsScaleTranslate = matrix.isScaleTranslate(); this->didSetMatrix(matrix); } @@ -1577,6 +1590,7 @@ void SkCanvas::onClipRect(const SkRect& rect, SkRegion::Op op, ClipEdgeStyle edg fClipStack->clipEmpty(); (void)fMCRec->fRasterClip.setEmpty(); + fDeviceClipBounds.setEmpty(); return; } } @@ -1606,7 +1620,6 @@ void SkCanvas::onClipRect(const SkRect& rect, SkRegion::Op op, ClipEdgeStyle edg AutoValidateClip avc(this); fDeviceCMDirty = true; - fCachedLocalClipBoundsDirty = true; if (isScaleTrans) { const bool isAA = kSoft_ClipEdgeStyle == edgeStyle; @@ -1622,6 +1635,8 @@ void SkCanvas::onClipRect(const SkRect& rect, SkRegion::Op op, ClipEdgeStyle edg path.addRect(rect); this->SkCanvas::onClipPath(path, op, edgeStyle); } + + fDeviceClipBounds = qr_clip_bounds(fMCRec->fRasterClip.getBounds()); } void SkCanvas::clipRRect(const SkRRect& rrect, SkRegion::Op op, bool doAA) { @@ -1640,7 +1655,6 @@ void SkCanvas::onClipRRect(const SkRRect& rrect, SkRegion::Op op, ClipEdgeStyle AutoValidateClip avc(this); fDeviceCMDirty = true; - fCachedLocalClipBoundsDirty = true; if (!fAllowSoftClip) { edgeStyle = kHard_ClipEdgeStyle; } @@ -1649,6 +1663,7 @@ void SkCanvas::onClipRRect(const SkRRect& rrect, SkRegion::Op op, ClipEdgeStyle fMCRec->fRasterClip.op(transformedRRect, this->getTopLayerBounds(), op, kSoft_ClipEdgeStyle == edgeStyle); + fDeviceClipBounds = qr_clip_bounds(fMCRec->fRasterClip.getBounds()); return; } @@ -1696,6 +1711,7 @@ void SkCanvas::onClipPath(const SkPath& path, SkRegion::Op op, ClipEdgeStyle edg fClipStack->clipEmpty(); (void)fMCRec->fRasterClip.setEmpty(); + fDeviceClipBounds.setEmpty(); return; } } @@ -1704,7 +1720,6 @@ void SkCanvas::onClipPath(const SkPath& path, SkRegion::Op op, ClipEdgeStyle edg AutoValidateClip avc(this); fDeviceCMDirty = true; - fCachedLocalClipBoundsDirty = true; if (!fAllowSoftClip) { edgeStyle = kHard_ClipEdgeStyle; } @@ -1735,6 +1750,7 @@ void SkCanvas::onClipPath(const SkPath& path, SkRegion::Op op, ClipEdgeStyle edg } fMCRec->fRasterClip.op(devPath, this->getTopLayerBounds(), op, edgeStyle); + fDeviceClipBounds = qr_clip_bounds(fMCRec->fRasterClip.getBounds()); } void SkCanvas::clipRegion(const SkRegion& rgn, SkRegion::Op op) { @@ -1746,13 +1762,13 @@ void SkCanvas::onClipRegion(const SkRegion& rgn, SkRegion::Op op) { AutoValidateClip avc(this); fDeviceCMDirty = true; - fCachedLocalClipBoundsDirty = true; // todo: signal fClipStack that we have a region, and therefore (I guess) // we have to ignore it, and use the region directly? fClipStack->clipDevRect(rgn.getBounds(), op); fMCRec->fRasterClip.op(rgn, op); + fDeviceClipBounds = qr_clip_bounds(fMCRec->fRasterClip.getBounds()); } #ifdef SK_DEBUG @@ -1809,31 +1825,74 @@ bool SkCanvas::isClipRect() const { return fMCRec->fRasterClip.isRect(); } -bool SkCanvas::quickReject(const SkRect& rect) const { - if (!rect.isFinite()) - return true; +static inline bool is_nan_or_clipped(const Sk4f& devRect, const Sk4f& devClip) { +#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 + __m128 lLtT = _mm_unpacklo_ps(devRect.fVec, devClip.fVec); + __m128 RrBb = _mm_unpackhi_ps(devClip.fVec, devRect.fVec); + __m128 mask = _mm_cmplt_ps(lLtT, RrBb); + return 0xF != _mm_movemask_ps(mask); +#elif defined(SK_ARM_HAS_NEON) + float32x4_t lLtT = vzipq_f32(devRect.fVec, devClip.fVec).val[0]; + float32x4_t RrBb = vzipq_f32(devClip.fVec, devRect.fVec).val[1]; + uint32x4_t mask = vcltq_f32(lLtT, RrBb); + return 0xFFFFFFFFFFFFFFFF != (uint64_t) vmovn_u32(mask); +#else + SkRect devRectAsRect; + SkRect devClipAsRect; + devRect.store(&devRectAsRect.fLeft); + devClip.store(&devClipAsRect.fLeft); + return !devRectAsRect.isFinite() || !devRectAsRect.intersect(devClipAsRect); +#endif +} +// It's important for this function to not be inlined. Otherwise the compiler will share code +// between the fast path and the slow path, resulting in two slow paths. +static SK_NEVER_INLINE bool quick_reject_slow_path(const SkRect& src, const SkRect& deviceClip, + const SkMatrix& matrix) { + SkRect deviceRect; + matrix.mapRect(&deviceRect, src); + return !deviceRect.isFinite() || !deviceRect.intersect(deviceClip); +} + +bool SkCanvas::quickReject(const SkRect& src) const { +#ifdef SK_DEBUG + // Verify that fDeviceClipBounds are set properly. + SkRect tmp = qr_clip_bounds(fMCRec->fRasterClip.getBounds()); if (fMCRec->fRasterClip.isEmpty()) { - return true; + SkASSERT(fDeviceClipBounds.isEmpty() || tmp == fDeviceClipBounds); + } else { + SkASSERT(tmp == fDeviceClipBounds); } - if (fMCRec->fMatrix.hasPerspective()) { - SkRect dst; - fMCRec->fMatrix.mapRect(&dst, rect); - return !SkIRect::Intersects(dst.roundOut(), fMCRec->fRasterClip.getBounds()); - } else { - const SkRect& clipR = this->getLocalClipBounds(); + // Verify that fConservativeIsScaleTranslate is set properly. + SkASSERT(!fConservativeIsScaleTranslate || fMCRec->fMatrix.isScaleTranslate()); +#endif - // for speed, do the most likely reject compares first - // TODO: should we use | instead, or compare all 4 at once? - if (rect.fTop >= clipR.fBottom || rect.fBottom <= clipR.fTop) { - return true; - } - if (rect.fLeft >= clipR.fRight || rect.fRight <= clipR.fLeft) { - return true; - } - return false; + if (!fConservativeIsScaleTranslate) { + return quick_reject_slow_path(src, fDeviceClipBounds, fMCRec->fMatrix); } + + // We inline the implementation of mapScaleTranslate() for the fast path. + float sx = fMCRec->fMatrix.getScaleX(); + float sy = fMCRec->fMatrix.getScaleY(); + float tx = fMCRec->fMatrix.getTranslateX(); + float ty = fMCRec->fMatrix.getTranslateY(); + Sk4f scale(sx, sy, sx, sy); + Sk4f trans(tx, ty, tx, ty); + + // Apply matrix. + Sk4f ltrb = Sk4f::Load(&src.fLeft) * scale + trans; + + // Make sure left < right, top < bottom. + Sk4f rblt(ltrb[2], ltrb[3], ltrb[0], ltrb[1]); + Sk4f min = Sk4f::Min(ltrb, rblt); + Sk4f max = Sk4f::Max(ltrb, rblt); + // We can extract either pair [0,1] or [2,3] from min and max and be correct, but on + // ARM this sequence generates the fastest (a single instruction). + Sk4f devRect = Sk4f(min[2], min[3], max[0], max[1]); + + // Check if the device rect is NaN or outside the clip. + return is_nan_or_clipped(devRect, Sk4f::Load(&fDeviceClipBounds.fLeft)); } bool SkCanvas::quickReject(const SkPath& path) const { diff --git a/tests/QuickRejectTest.cpp b/tests/QuickRejectTest.cpp index 6a978743f6..b39484bd89 100644 --- a/tests/QuickRejectTest.cpp +++ b/tests/QuickRejectTest.cpp @@ -103,7 +103,45 @@ static void test_layers(skiatest::Reporter* reporter) { REPORTER_ASSERT(reporter, false == canvas.quickReject(SkRect::MakeWH(60, 60))); } +static void test_quick_reject(skiatest::Reporter* reporter) { + SkCanvas canvas(100, 100); + SkRect r0 = SkRect::MakeLTRB(-50.0f, -50.0f, 50.0f, 50.0f); + SkRect r1 = SkRect::MakeLTRB(-50.0f, 110.0f, 50.0f, 120.0f); + SkRect r2 = SkRect::MakeLTRB(110.0f, -50.0f, 120.0f, 50.0f); + SkRect r3 = SkRect::MakeLTRB(-120.0f, -50.0f, 120.0f, 50.0f); + SkRect r4 = SkRect::MakeLTRB(-50.0f, -120.0f, 50.0f, 120.0f); + SkRect r5 = SkRect::MakeLTRB(-120.0f, -120.0f, 120.0f, 120.0f); + SkRect r6 = SkRect::MakeLTRB(-120.0f, -120.0f, -110.0f, -110.0f); + SkRect r7 = SkRect::MakeLTRB(SK_ScalarNaN, -50.0f, 50.0f, 50.0f); + SkRect r8 = SkRect::MakeLTRB(-50.0f, SK_ScalarNaN, 50.0f, 50.0f); + SkRect r9 = SkRect::MakeLTRB(-50.0f, -50.0f, SK_ScalarNaN, 50.0f); + SkRect r10 = SkRect::MakeLTRB(-50.0f, -50.0f, 50.0f, SK_ScalarNaN); + REPORTER_ASSERT(reporter, false == canvas.quickReject(r0)); + REPORTER_ASSERT(reporter, true == canvas.quickReject(r1)); + REPORTER_ASSERT(reporter, true == canvas.quickReject(r2)); + REPORTER_ASSERT(reporter, false == canvas.quickReject(r3)); + REPORTER_ASSERT(reporter, false == canvas.quickReject(r4)); + REPORTER_ASSERT(reporter, false == canvas.quickReject(r5)); + REPORTER_ASSERT(reporter, true == canvas.quickReject(r6)); + REPORTER_ASSERT(reporter, true == canvas.quickReject(r7)); + REPORTER_ASSERT(reporter, true == canvas.quickReject(r8)); + REPORTER_ASSERT(reporter, true == canvas.quickReject(r9)); + REPORTER_ASSERT(reporter, true == canvas.quickReject(r10)); + + SkMatrix m = SkMatrix::MakeScale(2.0f); + m.setTranslateX(10.0f); + m.setTranslateY(10.0f); + canvas.setMatrix(m); + SkRect r11 = SkRect::MakeLTRB(5.0f, 5.0f, 100.0f, 100.0f); + SkRect r12 = SkRect::MakeLTRB(5.0f, 50.0f, 100.0f, 100.0f); + SkRect r13 = SkRect::MakeLTRB(50.0f, 5.0f, 100.0f, 100.0f); + REPORTER_ASSERT(reporter, false == canvas.quickReject(r11)); + REPORTER_ASSERT(reporter, true == canvas.quickReject(r12)); + REPORTER_ASSERT(reporter, true == canvas.quickReject(r13)); +} + DEF_TEST(QuickReject, reporter) { test_drawBitmap(reporter); test_layers(reporter); + test_quick_reject(reporter); } |