Fix narrowing conversion errors from r11986

src/utils/SkMatrix44.cpp

Index: src/utils/SkMatrix44.cpp
diff --git a/src/utils/SkMatrix44.cpp b/src/utils/SkMatrix44.cpp
index a598e523f14de83307f586733fe5d17895106e7e..37ab8db27f08ee89a71a9376e0b5acb6caf5cb7a 100644
--- a/src/utils/SkMatrix44.cpp
+++ b/src/utils/SkMatrix44.cpp
@@ -819,7 +819,10 @@ static void map2_pf(const SkMScalar mat[][4], const float* SK_RESTRICT src2,
         float sx = src2[0];
         float sy = src2[1];
         for (int i = 0; i < 4; i++) {
-            dst4[i] = mat[0][i] * sx + mat[1][i] * sy + mat[3][i];
+            dst4[i] =
+                SkMScalarToFloat(mat[0][i]) * sx +

[reed1, 2013/10/29 14:58:52]

This technique throws away all of the precision benefits of having doubles in
MScalar. Look at map2_af for an alternate technique. It keeps all intermediates
in doubles (MScalar), and just pays the (perf) cost of the conversion when
writing the final answer.

[epoger, 2013/10/29 15:18:30]

Thanks for catching this, Mike.

Do you think it's worth having some unit tests to make sure the precision of
these operations is maintained?

[enne (OOO), 2013/10/29 16:56:37]

When SkMScalar is a float, map2_af's technique is slow because of all the float
to double conversions.  I think there either need to be #ifdefs or some use of
SkMScalar as intermediates to keep floats as floats and doubles as doubles
without the conversion cost.

+                SkMScalarToFloat(mat[1][i]) * sy +
+                SkMScalarToFloat(mat[3][i]);
         }
         src2 += 2;
         dst4 += 4;