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

source/row_win.cc

 /*
  *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
  *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
@@ skipping 6077 matching lines @@
     vmovq       qword ptr [edx], xmm0
     lea         edx, [edx + 8]
     sub         ecx, 2
     jg          convertloop
     vzeroupper
     ret
   }
 }
 #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, uint16* dst, float scale, int width) {
   __asm {
     mov        eax, [esp + 4]      /* src */
     mov        edx, [esp + 8]      /* dst */
     vbroadcastss ymm4, [esp + 12]  /* scale */
     mov        ecx, [esp + 16]     /* width */
 
     // 8 pixel loop.
  convertloop:
     vpmovzxwd   ymm0, xmmword ptr [eax]  // 8 shorts -> 8 ints
-    lea         eax, [eax + 16]
+    vpmovzxwd   ymm1, xmmword ptr [eax + 16]  // 8 more shorts
+    lea         eax, [eax + 32]
     vcvtdq2ps   ymm0, ymm0        // convert 8 ints to floats
+    vcvtdq2ps   ymm1, ymm1
     vmulps      ymm0, ymm0, ymm4  // scale to normalized range 0 to 1
-    vcvtps2ph   xmm0, ymm0, 0     // float conver to 8 half floats round even
+    vmulps      ymm1, ymm1, ymm4
+    vcvtps2ph   xmm0, ymm0, 3     // float convert to 8 half floats truncate
+    vcvtps2ph   xmm1, ymm1, 3
     vmovdqu     [edx], xmm0
-    lea         edx, [edx + 16]
-    sub         ecx, 8
+    vmovdqu     [edx + 16], xmm1
+    lea         edx, [edx + 32]
+    sub         ecx, 16
     jg          convertloop
     vzeroupper
     ret
   }
 }
-#endif  // HAS_SHORTTOHALFFLOAT_AVX2
+#endif  // HAS_HALFFLOATROW_AVX2
 
 #ifdef HAS_ARGBCOLORTABLEROW_X86
 // Tranform ARGB pixels with color table.
 __declspec(naked)
 void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
                            int width) {
   __asm {
     push       esi
     mov        eax, [esp + 4 + 4]   /* dst_argb */
     mov        esi, [esp + 4 + 8]   /* table_argb */
@@ skipping 153 matching lines @@
 #endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
 
 #endif  // defined(_M_X64)
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
 #endif  // !defined(LIBYUV_DISABLE_X86) && (defined(_M_IX86) || defined(_M_X64))