use conservative bounds to disable clipping

Bug: skia:6754
Change-Id: I658e7e7a176395eff7e239093fbad8038a4141ad
Reviewed-on: https://skia-review.googlesource.com/107884
Reviewed-by: Cary Clark <caryclark@google.com>
Commit-Queue: Mike Reed <reed@google.com>
This commit is contained in:
Mike Reed 2018-02-15 12:55:01 -05:00 committed by Skia Commit-Bot
parent 0ff79b2ed0
commit 3708f024b1

View File

@ -561,71 +561,49 @@ static bool clip_to_limit(const SkRegion& orig, SkRegion* reduced) {
return true;
}
/**
* Variants of SkScalarRoundToInt, identical to SkDScalarRoundToInt except when the input fraction
* is 0.5. When SK_RASTERIZE_EVEN_ROUNDING is enabled, we must bias the result before rounding to
* account for potential FDot6 rounding edge-cases.
*/
#ifdef SK_RASTERIZE_EVEN_ROUNDING
static const double kRoundBias = 0.5 / SK_FDot6One;
#else
static const double kRoundBias = 0.0;
#endif
// Bias used for conservative rounding of float rects to int rects, to nudge the irects a little
// larger, so we don't "think" a path's bounds are inside a clip, when (due to numeric drift in
// the scan-converter) we might walk beyond the predicted limits.
static const double kConservativeRoundBias = 0.5 + 0.5 / SK_FDot6One;
/**
* Round the value down. This is used to round the top and left of a rectangle,
* and corresponds to the way the scan converter treats the top and left edges.
*/
* Round the value down. This is used to round the top and left of a rectangle,
* and corresponds to the way the scan converter treats the top and left edges.
* It has a slight bias to make the "rounded" int smaller than a normal round, to create a more
* conservative int-bounds (larger) from a float rect.
*/
static inline int round_down_to_int(SkScalar x) {
double xx = x;
xx -= 0.5 + kRoundBias;
xx -= kConservativeRoundBias;
return sk_double_saturate2int(ceil(xx));
}
/**
* Round the value up. This is used to round the bottom and right of a rectangle,
* and corresponds to the way the scan converter treats the bottom and right edges.
* Round the value up. This is used to round the right and bottom of a rectangle.
* It has a slight bias to make the "rounded" int smaller than a normal round, to create a more
* conservative int-bounds (larger) from a float rect.
*/
static inline int round_up_to_int(SkScalar x) {
double xx = x;
xx += 0.5 + kRoundBias;
xx += kConservativeRoundBias;
return sk_double_saturate2int(floor(xx));
}
/**
* Variant of SkRect::round() that explicitly performs the rounding step (i.e. floor(x + 0.5))
* using double instead of SkScalar (float). It does this by calling SkDScalarRoundToInt(),
* which may be slower than calling SkScalarRountToInt(), but gives slightly more accurate
* results. Also rounds top and left using double, flooring when the fraction is exactly 0.5f.
*
* e.g.
* SkScalar left = 0.5f;
* int ileft = SkScalarRoundToInt(left);
* SkASSERT(0 == ileft); // <--- fails
* int ileft = round_down_to_int(left);
* SkASSERT(0 == ileft); // <--- succeeds
* SkScalar right = 0.49999997f;
* int iright = SkScalarRoundToInt(right);
* SkASSERT(0 == iright); // <--- fails
* iright = SkDScalarRoundToInt(right);
* SkASSERT(0 == iright); // <--- succeeds
*
*
* If using SK_RASTERIZE_EVEN_ROUNDING, we need to ensure we account for edges bounded by this
* rect being rounded to FDot6 format before being later rounded to an integer. For example, a
* value like 0.499 can be below 0.5, but round to 0.5 as FDot6, which would finally round to
* the integer 1, instead of just rounding to 0.
*
* To handle this, a small bias of half an FDot6 increment is added before actually rounding to
* an integer value. This simulates the rounding of SkScalarRoundToFDot6 without incurring the
* range loss of converting to FDot6 format first, preserving the integer range for the SkIRect.
* Thus, bottom and right are rounded in this manner (biased up), ensuring the rect is large
* enough.
/*
* Conservative rounding function, which effectively nudges the int-rect to be slightly larger
* than SkRect::round() might have produced. This is a safety-net for the scan-converter, which
* inspects the returned int-rect, and may disable clipping (for speed) if it thinks all of the
* edges will fit inside the clip's bounds. The scan-converter introduces slight numeric errors
* due to accumulated += of the slope, so this function is used to return a conservatively large
* int-bounds, and thus we will only disable clipping if we're sure the edges will stay in-bounds.
*/
static void round_asymmetric_to_int(const SkRect& src, SkIRect* dst) {
SkASSERT(dst);
dst->set(round_down_to_int(src.fLeft), round_down_to_int(src.fTop),
round_up_to_int(src.fRight), round_up_to_int(src.fBottom));
static SkIRect conservative_round_to_int(const SkRect& src) {
return {
round_down_to_int(src.fLeft),
round_down_to_int(src.fTop),
round_up_to_int(src.fRight),
round_up_to_int(src.fBottom),
};
}
void SkScan::FillPath(const SkPath& path, const SkRegion& origClip,
@ -644,7 +622,7 @@ void SkScan::FillPath(const SkPath& path, const SkRegion& origClip,
}
clipPtr = &finiteClip;
}
// don't reference "origClip" any more, just use clipPtr
// don't reference "origClip" any more, just use clipPtr
SkRect bounds = path.getBounds();
@ -655,12 +633,8 @@ void SkScan::FillPath(const SkPath& path, const SkRegion& origClip,
}
irPreClipped = true;
}
SkIRect ir;
// We deliberately call round_asymmetric_to_int() instead of round(), since we can't afford
// to generate a bounds that is tighter than the corresponding SkEdges. The edge code basically
// converts the floats to fixed, and then "rounds". If we called round() instead of
// round_asymmetric_to_int() here, we could generate the wrong ir for values like 0.4999997.
round_asymmetric_to_int(bounds, &ir);
SkIRect ir = conservative_round_to_int(bounds);
if (ir.isEmpty()) {
if (path.isInverseFillType()) {
blitter->blitRegion(*clipPtr);