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

source/row_common.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 2314 matching lines @@
 
     dst_argb[0] = Clamp((int32)(db));
     dst_argb[1] = Clamp((int32)(dg));
     dst_argb[2] = Clamp((int32)(dr));
     dst_argb[3] = Clamp((int32)(da));
     src_argb += 4;
     dst_argb += 4;
   }
 }
 
+// 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

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

Should be 511 I think.
(and 1023 below.)

[fbarchard1, 2016/09/28 00:01:31]

Depends on the exact range you want to map across.
If the '9 bit' value is 0..511 and you want the half float to be 0..1, you are
right - scale should be 1/511.
If the '9 bit' value is 0..512 inclusive and you want the half float to be 0..1,
scale should be 1/512.
If the '9 bit' value is 0..511 and you want the half float to be 0..0.999.
(almost 1), scale should be 1/512.

This function doesnt care... its a scale factor to adjust the range from the int
to float.

+// e.g. 10 bit scale is 1.0f / 1024.0f
+void HalfFloatRow_C(const uint16* src, int16* dst, float scale, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+      float value = src[i] * scale;
+      uint32 f = *(uint32*)&value;

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

Why not use the 2^-112 trick?
AFAIk it produces the same result.

[fbarchard1, 2016/09/28 00:01:31]

Acknowledged.

C code is for reference and I wanted something known to be correct.

[hubbe, 2016/09/28 00:41:18]

Will there also be an optimized C code version for odd platforms? 

[fbarchard1, 2016/09/28 00:53:40]

Acknowledged.

In this version I've removed the sign bit, which is a bit
fasteTestHalfFloatPlane (1167 ms)

+      dst[i] = ((f >> 16) & 0x8000) |
+               ((((f & 0x7f800000) - 0x38000000) >> 13) & 0x7c00) |
+               ((f >> 13) & 0x03ff);
+  }
+}
+
 void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
                              const uint8* luma, uint32 lumacoeff) {
   uint32 bc = lumacoeff & 0xff;
   uint32 gc = (lumacoeff >> 8) & 0xff;
   uint32 rc = (lumacoeff >> 16) & 0xff;
 
   int i;
   for (i = 0; i < width - 1; i += 2) {
     // Luminance in rows, color values in columns.
     const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
@@ skipping 273 matching lines @@
     dst_rgb565 += twidth * 2;
     width -= twidth;
   }
 }
 #endif
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif