Make the fallback loop very accurate.

The summation xs = xs + dx was causing xs to overflow the bounds of the source.
Change the algorithm to eliminate error accumulation. Performance is not a
concern, because the fallback should only be used in the rare cases.

R=mtklein@google.com

BUG=skia:6216

Change-Id: Iff8e55af5eb79606c7b1e693ae825ceaeda44afd
Reviewed-on: https://skia-review.googlesource.com/8975
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Mike Klein <mtklein@google.com>

1 files changed, 11 insertions, 6 deletions

diff --git a/src/core/SkLinearBitmapPipeline_core.h b/src/core/SkLinearBitmapPipeline_core.h
index 5ef6fcab5b..ce6c05b752 100644
--- a/src/core/SkLinearBitmapPipeline_core.h
+++ b/src/core/SkLinearBitmapPipeline_core.h
@@ -157,23 +157,28 @@ void span_fallback(Span span, Stage* stage) {
     SkScalar length;
     int count;
     std::tie(start, length, count) = span;
-    Sk4f xs{X(start)};
+    Sk4f startXs{X(start)};
     Sk4f ys{Y(start)};
+    Sk4f mults = {0.0f, 1.0f, 2.0f, 3.0f};
 
     // Initializing this is not needed, but some compilers can't figure this out.
-    Sk4s fourDx{0.0f};
+    Sk4s dXs{0.0f};
     if (count > 1) {
         SkScalar dx = length / (count - 1);
-        xs = xs + Sk4f{0.0f, 1.0f, 2.0f, 3.0f} * dx;
-        // Only used if count is >= 4.
-        fourDx = Sk4f{4.0f * dx};
+        dXs = Sk4f{dx};
     }
 
+    // Instead of using xs = xs + dx every round, this uses xs = i * dx + X(start). This
+    // eliminates the rounding error for the sum.
+    Sk4f xs = startXs + mults * dXs;
     while (count >= 4) {
         stage->pointList4(xs, ys);
-        xs = xs + fourDx;
+
+        mults += Sk4f{4.0f};
+        xs = mults * dXs + startXs;
         count -= 4;
     }
+
     if (count > 0) {
         stage->pointListFew(count, xs, ys);
     }