speedup exp() in gaussian colorfilter

~20% faster
differs in low 1 from prev impl

Bug: skia:
Change-Id: If7aa7f5523f15a5424ebf5a09313c5e1d43e1a1f
Reviewed-on: https://skia-review.googlesource.com/15105
Commit-Queue: Mike Reed <reed@google.com>
Reviewed-by: Jim Van Verth <jvanverth@google.com>

1 files changed, 54 insertions, 5 deletions

diff --git a/src/utils/SkShadowUtils.cpp b/src/utils/SkShadowUtils.cpp
index 928eb66bcb..f3311cd268 100644
--- a/src/utils/SkShadowUtils.cpp
+++ b/src/utils/SkShadowUtils.cpp
@@ -52,14 +52,63 @@ private:
     typedef SkColorFilter INHERITED;
 };
 
+static void build_table() {
+    SkDebugf("const uint16_t gByteExpU16Table[256] = {");
+    for (int i = 0; i <= 255; ++i) {
+        if (!(i % 8)) {
+            SkDebugf("\n");
+        }
+        SkScalar factor = SK_Scalar1 - i / 255.f;
+        factor = SkScalarExp(-factor * factor * 4) - 0.018f;
+        int v = (int)(factor * 65536);
+        SkDebugf(" 0x%04X,", v);
+    }
+    SkDebugf("\n};\n");
+}
+
+const uint16_t gByteExpU16Table[256] = {
+    0x0014, 0x003A, 0x0062, 0x008A, 0x00B3, 0x00DE, 0x010A, 0x0136,
+    0x0165, 0x0194, 0x01C4, 0x01F6, 0x0229, 0x025E, 0x0294, 0x02CB,
+    0x0304, 0x033E, 0x0379, 0x03B7, 0x03F5, 0x0435, 0x0477, 0x04BB,
+    0x0500, 0x0546, 0x058F, 0x05D9, 0x0625, 0x0673, 0x06C3, 0x0714,
+    0x0768, 0x07BD, 0x0814, 0x086E, 0x08C9, 0x0926, 0x0986, 0x09E8,
+    0x0A4B, 0x0AB1, 0x0B1A, 0x0B84, 0x0BF1, 0x0C60, 0x0CD2, 0x0D46,
+    0x0DBC, 0x0E35, 0x0EB0, 0x0F2E, 0x0FAF, 0x1032, 0x10B7, 0x1140,
+    0x11CB, 0x1258, 0x12E9, 0x137C, 0x1412, 0x14AB, 0x1547, 0x15E6,
+    0x1688, 0x172D, 0x17D5, 0x187F, 0x192D, 0x19DE, 0x1A92, 0x1B4A,
+    0x1C04, 0x1CC2, 0x1D83, 0x1E47, 0x1F0E, 0x1FD9, 0x20A7, 0x2178,
+    0x224D, 0x2325, 0x2401, 0x24E0, 0x25C2, 0x26A8, 0x2792, 0x287F,
+    0x296F, 0x2A63, 0x2B5A, 0x2C56, 0x2D54, 0x2E56, 0x2F5C, 0x3065,
+    0x3172, 0x3283, 0x3397, 0x34AE, 0x35CA, 0x36E9, 0x380B, 0x3931,
+    0x3A5B, 0x3B88, 0x3CB9, 0x3DED, 0x3F25, 0x4061, 0x41A0, 0x42E2,
+    0x4428, 0x4572, 0x46BF, 0x480F, 0x4963, 0x4ABA, 0x4C14, 0x4D72,
+    0x4ED3, 0x5038, 0x519F, 0x530A, 0x5478, 0x55E9, 0x575D, 0x58D4,
+    0x5A4F, 0x5BCC, 0x5D4C, 0x5ECF, 0x6054, 0x61DD, 0x6368, 0x64F6,
+    0x6686, 0x6819, 0x69AE, 0x6B45, 0x6CDF, 0x6E7B, 0x701A, 0x71BA,
+    0x735D, 0x7501, 0x76A7, 0x784F, 0x79F9, 0x7BA4, 0x7D51, 0x7F00,
+    0x80AF, 0x8260, 0x8413, 0x85C6, 0x877A, 0x8930, 0x8AE6, 0x8C9C,
+    0x8E54, 0x900C, 0x91C4, 0x937D, 0x9535, 0x96EE, 0x98A7, 0x9A60,
+    0x9C18, 0x9DD1, 0x9F88, 0xA13F, 0xA2F6, 0xA4AB, 0xA660, 0xA814,
+    0xA9C6, 0xAB78, 0xAD27, 0xAED6, 0xB082, 0xB22D, 0xB3D6, 0xB57D,
+    0xB722, 0xB8C5, 0xBA65, 0xBC03, 0xBD9E, 0xBF37, 0xC0CD, 0xC25F,
+    0xC3EF, 0xC57B, 0xC704, 0xC889, 0xCA0B, 0xCB89, 0xCD04, 0xCE7A,
+    0xCFEC, 0xD15A, 0xD2C4, 0xD429, 0xD58A, 0xD6E6, 0xD83D, 0xD990,
+    0xDADD, 0xDC25, 0xDD68, 0xDEA6, 0xDFDE, 0xE111, 0xE23E, 0xE365,
+    0xE486, 0xE5A2, 0xE6B7, 0xE7C6, 0xE8CF, 0xE9D1, 0xEACD, 0xEBC3,
+    0xECB2, 0xED9A, 0xEE7C, 0xEF56, 0xF02A, 0xF0F6, 0xF1BC, 0xF27A,
+    0xF332, 0xF3E1, 0xF48A, 0xF52B, 0xF5C5, 0xF657, 0xF6E1, 0xF764,
+    0xF7DF, 0xF852, 0xF8BE, 0xF922, 0xF97E, 0xF9D2, 0xFA1E, 0xFA62,
+    0xFA9F, 0xFAD3, 0xFAFF, 0xFB23, 0xFB40, 0xFB54, 0xFB60, 0xFB64,
+};
+
 void SkGaussianColorFilter::filterSpan(const SkPMColor src[], int count, SkPMColor dst[]) const {
+    // to re-build the table, call build_table() which will dump it out using SkDebugf.
+    if (false) {
+        build_table();
+    }
     for (int i = 0; i < count; ++i) {
         SkPMColor c = src[i];
-
-        SkScalar factor = SK_Scalar1 - SkGetPackedB32(c) / 255.f;
-        factor = SkScalarExp(-factor * factor * 4) - 0.018f;
-
-        SkScalar a = factor * SkGetPackedG32(c);
+        uint8_t a = gByteExpU16Table[SkGetPackedB32(c)] * SkGetPackedG32(c) >> 16;
         dst[i] = SkPackARGB32(a, a, a, a);
     }
 }