Apply linear gradient premul in 4f

src/effects/gradients/SkLinearGradient.cpp

Index: src/effects/gradients/SkLinearGradient.cpp
diff --git a/src/effects/gradients/SkLinearGradient.cpp b/src/effects/gradients/SkLinearGradient.cpp
index a249ae4ac82c2195ce2e559978cf057ce0a8377a..b6bafbf6b0dae40d440a27fcd0d1911ee26c6ec6 100644
--- a/src/effects/gradients/SkLinearGradient.cpp
+++ b/src/effects/gradients/SkLinearGradient.cpp
@@ -527,11 +527,25 @@ find_backward(const SkLinearGradient::LinearGradientContext::Rec rec[], float ti
 
 template <bool apply_alpha> SkPMColor trunc_from_255(const Sk4f& x) {
     SkPMColor c;
+
+#ifdef SK_SUPPORT_LEGACY_GRADIENT_PREMUL
     SkNx_cast<uint8_t>(x).store(&c);
     if (apply_alpha) {
         c = SkPreMultiplyARGB(SkGetPackedA32(c), SkGetPackedR32(c),
                               SkGetPackedG32(c), SkGetPackedB32(c));
     }
+#else
+    Sk4f c4f255;
+    if (!apply_alpha) {
+        c4f255 = x;
+    } else {
+        // We have to pin, due to dithering bias.
+        c4f255 = Sk4f::Min(Sk4f::Max(x, Sk4f(0)), Sk4f(255));

[reed1, 2016/10/10 18:20:17]

does the store() call also pin? Why don't we have to pin in the non-alpha case?

[f(malita), 2016/10/10 18:33:18]

Good question.

Doesn't look like it pins, just truncates - but somehow we don't run into
trouble in that case (maybe because all values are biased the same way).

When we apply the premul in 4f though we seem to sometimes end up breaking the
premul invariant (rgb > a) without a pin, presumably because after premul the
channels truncate differently.  This manifests as dithery-looking artifacts in
the output.

Also, I seem to recall we pre-bias all values by 0.5 to prepare for truncate --
so maybe we only need an upper clamp on alpha?  Let me see if I tweak this some
more.

+        const float scale = c4f255[SkPM4f::A] * 1.f / 255;
+        c4f255 *= Sk4f(scale, scale, scale, 1);
+    }
+    SkNx_cast<uint8_t>(c4f255).store(&c);
+#endif
     return c;
 }