Improve the performance of SkConvertConfig8888Pixels using Array lookup

src/core/SkConfig8888.cpp

Index: src/core/SkConfig8888.cpp
diff --git a/src/core/SkConfig8888.cpp b/src/core/SkConfig8888.cpp
index f889f20744fa3b3f3c20189ae9ea56f31ac96a78..3364f445a62bc6baa8672cd11d1370556d941214 100644
--- a/src/core/SkConfig8888.cpp
+++ b/src/core/SkConfig8888.cpp
@@ -1,5 +1,6 @@
 #include "SkConfig8888.h"
 #include "SkMathPriv.h"
+#include "SkUnPreMultiply.h"
 
 namespace {
 
@@ -57,23 +58,21 @@ inline uint32_t convert_pixel(uint32_t pixel) {
     uint32_t a, r, g, b;
     unpack_config8888<IN_A_IDX, IN_R_IDX, IN_G_IDX, IN_B_IDX>(pixel, &a, &r, &g, &b);
     if (IN_PM && !OUT_PM) {
-        // We're doing the explicit divide to match WebKit layout
-        // test expectations. We can modify and rebaseline if there
-        // it can be shown that there is a more performant way to
-        // unpremul.
+        // It is verified that using SkUnPreMultiply::ApplyScale is a more
+        // performant way to unpremul than (value * 0xff) / a.

[Noel Gordon, 2013/06/04 05:55:18]

Does (v * 255 / a) equal your SkUnPreMultiply::ApplyScale approach for all v and
a in [0-255]?

[Jun Jiang, 2013/06/04 06:35:10]

Hi, Noel. (v * 255 / a) doesn't equal SkUnPreMultiply::ApplyScale() for all v
and a in [0-255]. In fact, the difference can be 1 between the two values for
some v and a in [0-255] and 0 for all other pairs. You can see the test in
SkUnPreMultiply_BuildTable() in src/core/SkUnPremultiply.cpp when building the
divide table.

[Tom Hudson, 2013/06/10 09:35:18]

Nit: better grammar: "Using SkUnPreMultiply::ApplyScale is faster than (value *
0xff) / a"

         if (a) {
-            r = r * 0xff / a;
-            g = g * 0xff / a;
-            b = b * 0xff / a;
+            SkUnPreMultiply::Scale scale = SkUnPreMultiply::GetScale(a);
+            r = SkUnPreMultiply::ApplyScale(scale, r);
+            g = SkUnPreMultiply::ApplyScale(scale, g);
+            b = SkUnPreMultiply::ApplyScale(scale, b);
         } else {
             return 0;
         }
     } else if (!IN_PM && OUT_PM) {
-        // This matches WebKit's conversion which we are replacing.
-        // We can consider alternative rounding rules for performance.
-        r = SkMulDiv255Ceiling(r, a);
-        g = SkMulDiv255Ceiling(g, a);
-        b = SkMulDiv255Ceiling(b, a);
+        // This matches SkUnPreMultiply conversion which we are replacing.
+        r = SkDiv255Round(r * a);

[Noel Gordon, 2013/06/04 07:16:06]

Could you test using the following here and tell me if the skia tests pass?

r = SkDiv255Round(SkMulS16(r, a));
g = SkDiv255Round(SkMulS16(g, a));
b = SkDiv255Round(SkMulS16(b, a));

+        g = SkDiv255Round(g * a);
+        b = SkDiv255Round(b * a);
     }
     return pack_config8888<OUT_A_IDX, OUT_R_IDX, OUT_G_IDX, OUT_B_IDX>(a, r, g, b);
 }