diff vp8data.h @ 11989:176c5deb6756 libavcodec

Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but we apply them as 16x8/8x16/8x8 subblocks where possible. Since this allows us to use width=8/16 instead of width=4 MC functions, we can now take more advantage of SSE2/SSSE3 optimizations, leading to a total speedup for splitMV filter of about 10%.

--- a/vp8data.h	Mon Jun 28 10:56:16 2010 +0000
+++ b/vp8data.h	Mon Jun 28 13:50:55 2010 +0000
@@ -62,6 +62,13 @@
     VP8_SUBMVMODE_NEW4X4
 };
 
+enum inter_splitmvmode {
+    VP8_SPLITMVMODE_16x8 = 0,    ///< 2 16x8 blocks (vertical)
+    VP8_SPLITMVMODE_8x16,        ///< 2 8x16 blocks (horizontal)
+    VP8_SPLITMVMODE_8x8,         ///< 2x2 blocks of 8x8px each
+    VP8_SPLITMVMODE_4x4,         ///< 4x4 blocks of 4x4px each
+};
+
 static const uint8_t vp8_pred4x4_mode[] =
 {
     [DC_PRED8x8]    = DC_PRED,
@@ -130,10 +137,10 @@
 };
 
 static const int8_t vp8_mbsplit_tree[3][2] = {
-    { -3,  1 },                             // '0' - 16 individual MVs
-     { -2,  2 },                            // '10' - quarter-based MVs
-      { -0, -1 }                            // '110' - top/bottom MVs,
-                                            // '111' - left/right MVs
+    { -VP8_SPLITMVMODE_4x4,  1 },           // '0' - 16 individual MVs
+     { -VP8_SPLITMVMODE_8x8,  2 },          // '10' - quarter-based MVs
+      { -VP8_SPLITMVMODE_16x8,              // '110' - top/bottom MVs
+        -VP8_SPLITMVMODE_8x16 }             // '111' - left/right MVs
 };
 static const uint8_t vp8_mbsplit_count[4] = {   2,   2,   4,  16 };
 static const uint8_t vp8_mbsplit_prob[3]  = { 110, 111, 150 };