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

unit_test/planar_test.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 2063 matching lines @@
   }
 
   for (int i = 0; i < 1280; ++i) {
     EXPECT_EQ(dst_pixels_c[i][0], dst_pixels_opt[i][0]);
     EXPECT_EQ(dst_pixels_c[i][1], dst_pixels_opt[i][1]);
     EXPECT_EQ(dst_pixels_c[i][2], dst_pixels_opt[i][2]);
     EXPECT_EQ(dst_pixels_c[i][3], dst_pixels_opt[i][3]);
   }
 }
 
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane) {
+  int i, j;
+  const int y_plane_size = benchmark_width_ * benchmark_height_ * 2;
+
+  align_buffer_page_end(orig_y, y_plane_size);
+  align_buffer_page_end(dst_c, y_plane_size);
+  align_buffer_page_end(dst_opt, y_plane_size);
+  MemRandomize(orig_y, y_plane_size);
+  memset(dst_c, 0, y_plane_size);
+  memset(dst_opt, 1, y_plane_size);
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+  double c_time = get_time();
+  for (j = 0; j < benchmark_iterations_; j++) {
+    HalfFloatPlane((uint16*)orig_y, benchmark_width_ * 2,
+                   (int16*)dst_c, benchmark_width_ * 2,
+                   1.0f / 4096.0f, benchmark_width_, benchmark_height_);
+  }
+  c_time = (get_time() - c_time) / benchmark_iterations_;
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+  double opt_time = get_time();
+  for (j = 0; j < benchmark_iterations_; j++) {
+    HalfFloatPlane((uint16*)orig_y, benchmark_width_ * 2,
+                   (int16*)dst_opt, benchmark_width_ * 2,
+                   1.0f / 4096.0f, benchmark_width_, benchmark_height_);
+  }
+  opt_time = (get_time() - opt_time) / benchmark_iterations_;
+
+  for (i = 0; i < y_plane_size; ++i) {
+    EXPECT_EQ(dst_c[i], dst_opt[i]);

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

Might not be the best way to write this test.
Half-float rounding could be different with different optimizations, which I
don't think is a problem as long as the rounding behavior is consistent across
the range of numbers.

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

My passes for the float scale value of 1/4096.
I've tested the full range of uint16, which scaled to 4.12 fixed point... values
up to 16.0f roughly.  Some of the larger values will loose lower bits, but the C
and AVX2 function appear to match, so as long as we can make all platforms
exactly match, this is a good test.

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

AVX has the option of rounding to "nearest" though, which is probably better if
it is available....

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

The current AVX2 enables rounding.
http://www.felixcloutier.com/x86/VCVTPS2PH.html


Rounding is not applicable to small values:
integers between 0 and 2048 can be exactly represented
See also https://en.wikipedia.org/wiki/Half-precision_floating-point_format

+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  free_aligned_buffer_page_end(dst_c);
+  free_aligned_buffer_page_end(dst_opt);
+}
+
 TEST_F(LibYUVPlanarTest, TestARGBLumaColorTable) {
   SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
   SIMD_ALIGNED(uint8 dst_pixels_opt[1280][4]);
   SIMD_ALIGNED(uint8 dst_pixels_c[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
   align_buffer_page_end(lumacolortable, 32768);
   int v = 0;
   for (int i = 0; i < 32768; ++i) {
     lumacolortable[i] = v;
@@ skipping 353 matching lines @@
   }
 
   free_aligned_buffer_page_end(src_pixels);
   free_aligned_buffer_page_end(tmp_pixels_u);
   free_aligned_buffer_page_end(tmp_pixels_v);
   free_aligned_buffer_page_end(dst_pixels_opt);
   free_aligned_buffer_page_end(dst_pixels_c);
 }
 
 }  // namespace libyuv