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..9c54ce06f44b5b5d39489db64048486fed93572a 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) {
   __asm {
     mov        eax, [esp + 4]      /* src */
     mov        edx, [esp + 8]      /* dst */
@@ -6114,7 +6110,7 @@ void ShortToF16Row_AVX2(const uint16* src, int16* dst, float scale, int width) {
     lea         eax, [eax + 16]
     vcvtdq2ps   ymm0, ymm0        // convert 8 ints to floats
     vmulps      ymm0, ymm0, ymm4  // scale to normalized range 0 to 1
-    vcvtps2ph   xmm0, ymm0, 0     // float conver to 8 half floats round even
+    vcvtps2ph   xmm0, ymm0, 3     // float convert to 8 half floats truncate
     vmovdqu     [edx], xmm0
     lea         edx, [edx + 16]
     sub         ecx, 8
@@ -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.