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

source/row_common.cc

Index: source/row_common.cc
diff --git a/source/row_common.cc b/source/row_common.cc
index 099ab600dbcfcdcbfce4b9bf540653606763a2ac..14f910c77f3997fb5a05f2556470baadb7e22e73 100644
--- a/source/row_common.cc
+++ b/source/row_common.cc
@@ -2333,6 +2333,25 @@ void ARGBPolynomialRow_C(const uint8* src_argb,
   }
 }
 
+// Samples assumed to be unsigned in low 9, 10 or 12 bits.  Scale factor
+// adjust the source integer range to the half float range desired.
+
+// This magic constant is 2^-112. Multiplying by this
+// is the same as subtracting 112 from the exponent, which
+// is the difference in exponent bias between 32-bit and
+// 16-bit floats. Once we've done this subtraction, we can
+// simply extract the low bits of the exponent and the high
+// bits of the mantissa from our float and we're done.
+

[wangcheng, 2016/09/29 19:12:14]

maybe uint16* for dst

[fbarchard1, 2016/09/29 21:56:02]

Done.

Still does signed extend:
 124:	f3 0f 6f 18          	movdqu (%rax),%xmm3
 128:	66 0f 6f e3          	movdqa %xmm3,%xmm4
 12c:	66 0f 61 e2          	punpcklwd %xmm2,%xmm4
 130:	0f 5b e4             	cvtdq2ps %xmm4,%xmm4
 133:	66 0f 69 da          	punpckhwd %xmm2,%xmm3
 137:	0f 5b db             	cvtdq2ps %xmm3,%xmm3
 13a:	0f 59 d9             	mulps  %xmm1,%xmm3
 13d:	0f 59 e1             	mulps  %xmm1,%xmm4
 140:	66 0f 72 d4 0d       	psrld  $0xd,%xmm4
 145:	66 0f 72 d3 0d       	psrld  $0xd,%xmm3
 14a:	66 0f 72 f3 10       	pslld  $0x10,%xmm3
 14f:	66 0f 72 e3 10       	psrad  $0x10,%xmm3
 154:	66 0f 72 f4 10       	pslld  $0x10,%xmm4
 159:	66 0f 72 e4 10       	psrad  $0x10,%xmm4
 15e:	66 0f 6b e3          	packssdw %xmm3,%xmm4
 162:	f3 0f 7f 21          	movdqu %xmm4,(%rcx)

+void HalfFloatRow_C(const uint16* src, int16* dst, float scale, int width) {
+  int i;
+  float mult = 1.9259299444e-34f * scale;
+  for (i = 0; i < width; ++i) {
+    float value = src[i] * mult;
+    dst[i] = (*(uint32_t*)&value) >> 13;
+  }
+}
+
 void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
                              const uint8* luma, uint32 lumacoeff) {
   uint32 bc = lumacoeff & 0xff;