Optimized implementation of quickReject()
Impl Overview
(1) Keep the device clip bounds up to date. This
requires minimal additional work in a few places
throughout canvas.
(2) Keep track of if the ctm isScaleTranslate. Yes,
there's a function that does this, but it's slow
to call.
(3) Perform the src->device transform in quick reject,
then check intersection/nan.
Other Notes:
(1) NaN and intersection checks are performed
simultaneously.
(2) We no longer quick reject infinity.
(3) Affine and perspective are both handled in the slow
case.
(4) SkRasterClip::isEmpty() is handled by the intersection
check.
Performance on Nexus 6P:
93.2ms -> 59.8ms
Overall Android Jank Tests Performance Impact:
Should gain us a ms or two on some tests.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2225393002
Committed: https://skia.googlesource.com/skia/+/d22a817ff57986407facd16af36320fc86ce02da
Review-Url: https://codereview.chromium.org/2225393002
diff --git a/src/core/SkCanvas.cpp b/src/core/SkCanvas.cpp
index 24d9506..4160df7 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 +633,28 @@
////////////////////////////////////////////////////////////////////////////
+static inline SkRect qr_clip_bounds(const SkIRect& bounds) {
+ if (bounds.isEmpty()) {
+ return SkRect::MakeEmpty();
+ }
+
+ // 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 +665,6 @@
// const-cast.
*const_cast<bool*>(&fConservativeRasterClip) = SkToBool(flags & kConservativeRasterClip_InitFlag);
- fCachedLocalClipBounds.setEmpty();
- fCachedLocalClipBoundsDirty = true;
fAllowSoftClip = true;
fAllowSimplifyClip = false;
fDeviceCMDirty = true;
@@ -682,7 +693,10 @@
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 +1111,8 @@
// 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 +1123,9 @@
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 +1317,6 @@
SkASSERT(fMCStack.count() != 0);
fDeviceCMDirty = true;
- fCachedLocalClipBoundsDirty = true;
fClipStack->restore();
@@ -1337,6 +1350,11 @@
// 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 +1509,20 @@
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 +1594,7 @@
fClipStack->clipEmpty();
(void)fMCRec->fRasterClip.setEmpty();
+ fDeviceClipBounds.setEmpty();
return;
}
}
@@ -1606,7 +1624,6 @@
AutoValidateClip avc(this);
fDeviceCMDirty = true;
- fCachedLocalClipBoundsDirty = true;
if (isScaleTrans) {
const bool isAA = kSoft_ClipEdgeStyle == edgeStyle;
@@ -1622,6 +1639,8 @@
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 +1659,6 @@
AutoValidateClip avc(this);
fDeviceCMDirty = true;
- fCachedLocalClipBoundsDirty = true;
if (!fAllowSoftClip) {
edgeStyle = kHard_ClipEdgeStyle;
}
@@ -1649,6 +1667,7 @@
fMCRec->fRasterClip.op(transformedRRect, this->getTopLayerBounds(), op,
kSoft_ClipEdgeStyle == edgeStyle);
+ fDeviceClipBounds = qr_clip_bounds(fMCRec->fRasterClip.getBounds());
return;
}
@@ -1696,6 +1715,7 @@
fClipStack->clipEmpty();
(void)fMCRec->fRasterClip.setEmpty();
+ fDeviceClipBounds.setEmpty();
return;
}
}
@@ -1704,7 +1724,6 @@
AutoValidateClip avc(this);
fDeviceCMDirty = true;
- fCachedLocalClipBoundsDirty = true;
if (!fAllowSoftClip) {
edgeStyle = kHard_ClipEdgeStyle;
}
@@ -1735,6 +1754,7 @@
}
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 +1766,13 @@
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 +1829,74 @@
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 !defined(SKNX_NO_SIMD) && 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(SKNX_NO_SIMD) && 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;
- }
-
- if (fMCRec->fMatrix.hasPerspective()) {
- SkRect dst;
- fMCRec->fMatrix.mapRect(&dst, rect);
- return !SkIRect::Intersects(dst.roundOut(), fMCRec->fRasterClip.getBounds());
+ SkASSERT(fDeviceClipBounds.isEmpty() || tmp == fDeviceClipBounds);
} else {
- const SkRect& clipR = this->getLocalClipBounds();
-
- // 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;
+ SkASSERT(tmp == fDeviceClipBounds);
}
+
+ // Verify that fConservativeIsScaleTranslate is set properly.
+ SkASSERT(!fConservativeIsScaleTranslate || fMCRec->fMatrix.isScaleTranslate());
+#endif
+
+ 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 {
@@ -2457,7 +2520,7 @@
iter.fDevice->drawBitmap(iter, bitmap, matrix, looper.paint());
}
}
-
+
LOOPER_END
}