libyuv::YUY2ToY for isolating Y channel of YUY2.

libyuv::YUY2ToY for isolating Y channel of YUY2.

This function is the first step of YUY2 To I420.
Provided primarily for diagnostics.

TBR=wangcheng@google.com
BUG=libyuv:647
TESTED=LibYUVConvertTest.YUY2ToY_Opt

Committed: https://chromium.googlesource.com/libyuv/libyuv/+/edd3a84d0524a6ad2d9f9bc3c19742e79648354d

[fbarchard1, 2016-10-07 01:36:29]

fbarchard@google.com changed reviewers:
+ goldbergs@google.com

[fbarchard1, 2016-10-07 01:36:29]

Qualcomm 820:

C
YUY2ToY_Opt (363 ms)
YUY2ToY_Invert (1572 ms)

NEON
YUY2ToY_Opt (248 ms)
YUY2ToY_Invert (767 ms)

Intel Ivybridge
C 
YUY2ToY_Opt (736 ms)
YUY2ToY_Invert (804 ms)

SSE2
YUY2ToY_Opt (167 ms)
YUY2ToY_Invert (179 ms)

[fbarchard1, 2016-10-07 17:53:10]

YUY2ToI422_Opt (289 ms)
YUY2ToY_Opt (167 ms)

I422 to YUV 422, which is a Y plane and a UV of half width, full height.
YUY2ToY is the first step of YUY2ToI422 and should be faster.
Currently using a split function that outputs UV, intended for YUY2ToNV12

YUY2ToNV12_Opt (155 ms)

 71.53%  libyuv_unittest  libyuv_unittest      [.] SplitUVRow_SSE2
 17.45%  libyuv_unittest  libc-2.19.so         [.] __memmove_ssse3_back

[fbarchard1, 2016-10-07 18:01:55]

On 2016/10/07 17:53:10, fbarchard1 wrote:
> YUY2ToI422_Opt (289 ms)
> YUY2ToY_Opt (167 ms)
> 
> I422 to YUV 422, which is a Y plane and a UV of half width, full height.
> YUY2ToY is the first step of YUY2ToI422 and should be faster.
> Currently using a split function that outputs UV, intended for YUY2ToNV12
> 
> YUY2ToNV12_Opt (155 ms)
> 
>  71.53%  libyuv_unittest  libyuv_unittest      [.] SplitUVRow_SSE2
>  17.45%  libyuv_unittest  libc-2.19.so         [.] __memmove_ssse3_back

Neon
YUY2ToI420_Opt (522 ms)
YUY2ToI422_Opt (889 ms)
YUY2ToNV12_Opt (568 ms)
YUY2ToY_Opt (254 ms)

[fbarchard1, 2016-10-07 18:50:27]

C
I420ToYUY2_Opt (397 ms)
I422ToYUY2_Opt (382 ms)
YUY2ToI420_Opt (1758 ms)
YUY2ToI422_Opt (716 ms)
YUY2ToNV12_Opt (560 ms)
ARGBToYUY2_Opt (5256 ms)
YUY2ToARGB_Opt (8361 ms)
YUY2ToY_Opt (365 ms)
I420ToYUY2_ARGB_Opt (427 ms)
I422ToYUY2_ARGB_Opt (462 ms)

NEON
I420ToYUY2_Opt (304 ms)
I422ToYUY2_Opt (277 ms)
YUY2ToI420_Opt (523 ms)
YUY2ToI422_Opt (955 ms)
YUY2ToNV12_Opt (598 ms)
ARGBToYUY2_Opt (3407 ms)
YUY2ToARGB_Opt (1402 ms)
YUY2ToY_Opt (249 ms)
I420ToYUY2_ARGB_Opt (319 ms)
I422ToYUY2_ARGB_Opt (342 ms)

Theres a few things that stick out as wrong here.
1. neon is actually C for YUY2ToY and YUY2ToUV422, so performance should be same
on YUY2ToI422.
2. YUY2ToY should be the fastest
interesting that I420toYUY2 is so much faster than YUY2ToI420.
Consider doing 'split' functions to temporary buffer for 2nd write, then memcpy
temp to memory.

RUnning a smaller test

128x72 is 9216 bytes, fitting cache.

util/android/test_runner.py gtest -s libyuv_unittest -t 7200 --verbose --release
--gtest_filter=*YUY2*Opt -a "--libyuv_width=128 --libyuv_height=72
--libyuv_repeat=99999 --libyuv_flags=-1  --libyuv_cpu_info=-1" | grep ms

C
I420ToYUY2_Opt (344 ms)
I422ToYUY2_Opt (300 ms)
YUY2ToI420_Opt (966 ms)
YUY2ToI422_Opt (385 ms)
YUY2ToNV12_Opt (657 ms)
ARGBToYUY2_Opt (5380 ms)
YUY2ToARGB_Opt (10617 ms)
YUY2ToY_Opt (410 ms)
I420ToYUY2_ARGB_Opt (363 ms)
I422ToYUY2_ARGB_Opt (301 ms)

NEON
I420ToYUY2_Opt (113 ms)
I422ToYUY2_Opt (109 ms)
YUY2ToI420_Opt (140 ms)
YUY2ToI422_Opt (138 ms)
YUY2ToNV12_Opt (224 ms)
ARGBToYUY2_Opt (1443 ms)
YUY2ToARGB_Opt (1312 ms)
YUY2ToY_Opt (180 ms)
I420ToYUY2_ARGB_Opt (114 ms)
I422ToYUY2_ARGB_Opt (110 ms)

The C code is hurt by high overhead, but the performance numbers make more sense
in terms of simple conversions go faster.
YUY2ToI422 writes more memory than YUY2ToI420 but doesnt require 2 reads and an
average, and since the buffers are cached the memory write is less expensive
than the 2 reads and average.
YUY2ToY isnt the fastest, but currently uses a split to a temporary buffer.

[fbarchard1, 2016-10-08 00:17:57]

1280x720 *1000 iteations
I420ToYUY2_Opt (252 ms)
I422ToYUY2_Opt (269 ms)
YUY2ToI420_Opt (521 ms)
YUY2ToI422_Opt (887 ms)
YUY2ToNV12_Opt (314 ms)
ARGBToYUY2_Opt (3392 ms)
YUY2ToARGB_Opt (1419 ms)
YUY2ToY_Opt (568 ms)
I420ToYUY2_ARGB_Opt (331 ms)
I422ToYUY2_ARGB_Opt (334 ms)


128x72 *100000 iteations
I420ToYUY2_Opt (115 ms)
I422ToYUY2_Opt (107 ms)
YUY2ToI420_Opt (141 ms)
YUY2ToI422_Opt (144 ms)
YUY2ToNV12_Opt (223 ms)
ARGBToYUY2_Opt (1285 ms)
YUY2ToARGB_Opt (1314 ms)
YUY2ToY_Opt (50 ms)
I420ToYUY2_ARGB_Opt (114 ms)
I422ToYUY2_ARGB_Opt (111 ms)

[fbarchard1, 2016-10-08 00:19:31]

Description was changed from

==========
libyuv::YUY2ToY for isolating Y channel of YUY2.

This function is the first step of YUY2 To I420.
Provided primarily for diagnostics.

BUG=libyuv:647
TESTED=LibYUVConvertTest.YUY2ToY_Opt
==========

to

==========
libyuv::YUY2ToY for isolating Y channel of YUY2.

This function is the first step of YUY2 To I420.
Provided primarily for diagnostics.

TBR=wangcheng@google.com
BUG=libyuv:647
TESTED=LibYUVConvertTest.YUY2ToY_Opt
==========

[fbarchard1, 2016-10-08 00:19:31]

fbarchard@google.com changed reviewers:
+ wangcheng@google.com

[fbarchard1, 2016-10-08 00:20:34]

Description was changed from

==========
libyuv::YUY2ToY for isolating Y channel of YUY2.

This function is the first step of YUY2 To I420.
Provided primarily for diagnostics.

TBR=wangcheng@google.com
BUG=libyuv:647
TESTED=LibYUVConvertTest.YUY2ToY_Opt
==========

to

==========
libyuv::YUY2ToY for isolating Y channel of YUY2.

This function is the first step of YUY2 To I420.
Provided primarily for diagnostics.

TBR=wangcheng@google.com
BUG=libyuv:647
TESTED=LibYUVConvertTest.YUY2ToY_Opt

Committed:
https://chromium.googlesource.com/libyuv/libyuv/+/edd3a84d0524a6ad2d9f9bc3c19...
==========

[fbarchard1, 2016-10-08 00:20:35]

Committed patchset #6 (id:100001) manually as
edd3a84d0524a6ad2d9f9bc3c19742e79648354d (presubmit successful).