Add low level support for 12 bit 420, 422 and 444 YUV video frame conversion.

source/row_win.cc

Index: source/row_win.cc
diff --git a/source/row_win.cc b/source/row_win.cc
index d54f05e29e6b3e4a33918293d2f442d1aacc65b3..7822f750b8d5dc28ce2b52599d995abfe423d8d6 100644
--- a/source/row_win.cc
+++ b/source/row_win.cc
@@ -6095,13 +6095,9 @@ void ARGBPolynomialRow_AVX2(const uint8* src_argb,
 }
 #endif  // HAS_ARGBPOLYNOMIALROW_AVX2
 
-// Samples assumed to be unsigned in low 9, 10 or 12 bits.  Scale factor
-// adjust the sample range to 0 to 1 using a float multiply.
-// e.g. 9 bit scale is 1.0f / 512.0f
-// e.g. 10 bit scale is 1.0f / 1024.0f
-#ifdef HAS_SHORTTOHALFFLOAT_AVX2
+#ifdef HAS_HALFFLOATROW_AVX2
 __declspec(naked)
-void ShortToF16Row_AVX2(const uint16* src, int16* dst, float scale, int width) {
+void HalfFloatRow_AVX2(const uint16* src, int16* dst, float scale, int width) {

[hubbe, 2016/09/27 23:25:56]

I'm not really qualified to review this part properly.

   __asm {
     mov        eax, [esp + 4]      /* src */
     mov        edx, [esp + 8]      /* dst */
@@ -6123,7 +6119,7 @@ void ShortToF16Row_AVX2(const uint16* src, int16* dst, float scale, int width) {
     ret
   }
 }
-#endif  // HAS_SHORTTOHALFFLOAT_AVX2
+#endif  // HAS_HALFFLOATROW_AVX2
 
 #ifdef HAS_ARGBCOLORTABLEROW_X86
 // Tranform ARGB pixels with color table.